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 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; /* 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 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_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
304static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
305
306static LIST_HEAD(workqueues); /* PR: list of all workqueues */
307static bool workqueue_freezing; /* PL: have wqs started freezing? */
308
309/* PL: allowable cpus for unbound wqs and work items */
310static cpumask_var_t wq_unbound_cpumask;
311
312/* CPU where unbound work was last round robin scheduled from this CPU */
313static DEFINE_PER_CPU(int, wq_rr_cpu_last);
314
315/*
316 * Local execution of unbound work items is no longer guaranteed. The
317 * following always forces round-robin CPU selection on unbound work items
318 * to uncover usages which depend on it.
319 */
320#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
321static bool wq_debug_force_rr_cpu = true;
322#else
323static bool wq_debug_force_rr_cpu = false;
324#endif
325module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
326
327/* the per-cpu worker pools */
328static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
329
330static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
331
332/* PL: hash of all unbound pools keyed by pool->attrs */
333static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
334
335/* I: attributes used when instantiating standard unbound pools on demand */
336static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
337
338/* I: attributes used when instantiating ordered pools on demand */
339static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
340
341struct workqueue_struct *system_wq __read_mostly;
342EXPORT_SYMBOL(system_wq);
343struct workqueue_struct *system_highpri_wq __read_mostly;
344EXPORT_SYMBOL_GPL(system_highpri_wq);
345struct workqueue_struct *system_long_wq __read_mostly;
346EXPORT_SYMBOL_GPL(system_long_wq);
347struct workqueue_struct *system_unbound_wq __read_mostly;
348EXPORT_SYMBOL_GPL(system_unbound_wq);
349struct workqueue_struct *system_freezable_wq __read_mostly;
350EXPORT_SYMBOL_GPL(system_freezable_wq);
351struct workqueue_struct *system_power_efficient_wq __read_mostly;
352EXPORT_SYMBOL_GPL(system_power_efficient_wq);
353struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
354EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
355
356static int worker_thread(void *__worker);
357static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
358
359#define CREATE_TRACE_POINTS
360#include <trace/events/workqueue.h>
361
362#define assert_rcu_or_pool_mutex() \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
364 !lockdep_is_held(&wq_pool_mutex), \
365 "RCU or wq_pool_mutex should be held")
366
367#define assert_rcu_or_wq_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
369 !lockdep_is_held(&wq->mutex), \
370 "RCU or wq->mutex should be held")
371
372#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
373 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
374 !lockdep_is_held(&wq->mutex) && \
375 !lockdep_is_held(&wq_pool_mutex), \
376 "RCU, wq->mutex or wq_pool_mutex should be held")
377
378#define for_each_cpu_worker_pool(pool, cpu) \
379 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
380 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
381 (pool)++)
382
383/**
384 * for_each_pool - iterate through all worker_pools in the system
385 * @pool: iteration cursor
386 * @pi: integer used for iteration
387 *
388 * This must be called either with wq_pool_mutex held or RCU read
389 * locked. If the pool needs to be used beyond the locking in effect, the
390 * caller is responsible for guaranteeing that the pool stays online.
391 *
392 * The if/else clause exists only for the lockdep assertion and can be
393 * ignored.
394 */
395#define for_each_pool(pool, pi) \
396 idr_for_each_entry(&worker_pool_idr, pool, pi) \
397 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
398 else
399
400/**
401 * for_each_pool_worker - iterate through all workers of a worker_pool
402 * @worker: iteration cursor
403 * @pool: worker_pool to iterate workers of
404 *
405 * This must be called with wq_pool_attach_mutex.
406 *
407 * The if/else clause exists only for the lockdep assertion and can be
408 * ignored.
409 */
410#define for_each_pool_worker(worker, pool) \
411 list_for_each_entry((worker), &(pool)->workers, node) \
412 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
413 else
414
415/**
416 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
417 * @pwq: iteration cursor
418 * @wq: the target workqueue
419 *
420 * This must be called either with wq->mutex held or RCU read locked.
421 * If the pwq needs to be used beyond the locking in effect, the caller is
422 * responsible for guaranteeing that the pwq stays online.
423 *
424 * The if/else clause exists only for the lockdep assertion and can be
425 * ignored.
426 */
427#define for_each_pwq(pwq, wq) \
428 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
429 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
430 else
431
432#ifdef CONFIG_DEBUG_OBJECTS_WORK
433
434static struct debug_obj_descr work_debug_descr;
435
436static void *work_debug_hint(void *addr)
437{
438 return ((struct work_struct *) addr)->func;
439}
440
441static bool work_is_static_object(void *addr)
442{
443 struct work_struct *work = addr;
444
445 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
446}
447
448/*
449 * fixup_init is called when:
450 * - an active object is initialized
451 */
452static bool work_fixup_init(void *addr, enum debug_obj_state state)
453{
454 struct work_struct *work = addr;
455
456 switch (state) {
457 case ODEBUG_STATE_ACTIVE:
458 cancel_work_sync(work);
459 debug_object_init(work, &work_debug_descr);
460 return true;
461 default:
462 return false;
463 }
464}
465
466/*
467 * fixup_free is called when:
468 * - an active object is freed
469 */
470static bool work_fixup_free(void *addr, enum debug_obj_state state)
471{
472 struct work_struct *work = addr;
473
474 switch (state) {
475 case ODEBUG_STATE_ACTIVE:
476 cancel_work_sync(work);
477 debug_object_free(work, &work_debug_descr);
478 return true;
479 default:
480 return false;
481 }
482}
483
484static struct debug_obj_descr work_debug_descr = {
485 .name = "work_struct",
486 .debug_hint = work_debug_hint,
487 .is_static_object = work_is_static_object,
488 .fixup_init = work_fixup_init,
489 .fixup_free = work_fixup_free,
490};
491
492static inline void debug_work_activate(struct work_struct *work)
493{
494 debug_object_activate(work, &work_debug_descr);
495}
496
497static inline void debug_work_deactivate(struct work_struct *work)
498{
499 debug_object_deactivate(work, &work_debug_descr);
500}
501
502void __init_work(struct work_struct *work, int onstack)
503{
504 if (onstack)
505 debug_object_init_on_stack(work, &work_debug_descr);
506 else
507 debug_object_init(work, &work_debug_descr);
508}
509EXPORT_SYMBOL_GPL(__init_work);
510
511void destroy_work_on_stack(struct work_struct *work)
512{
513 debug_object_free(work, &work_debug_descr);
514}
515EXPORT_SYMBOL_GPL(destroy_work_on_stack);
516
517void destroy_delayed_work_on_stack(struct delayed_work *work)
518{
519 destroy_timer_on_stack(&work->timer);
520 debug_object_free(&work->work, &work_debug_descr);
521}
522EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
523
524#else
525static inline void debug_work_activate(struct work_struct *work) { }
526static inline void debug_work_deactivate(struct work_struct *work) { }
527#endif
528
529/**
530 * worker_pool_assign_id - allocate ID and assing it to @pool
531 * @pool: the pool pointer of interest
532 *
533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
534 * successfully, -errno on failure.
535 */
536static int worker_pool_assign_id(struct worker_pool *pool)
537{
538 int ret;
539
540 lockdep_assert_held(&wq_pool_mutex);
541
542 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
543 GFP_KERNEL);
544 if (ret >= 0) {
545 pool->id = ret;
546 return 0;
547 }
548 return ret;
549}
550
551/**
552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
553 * @wq: the target workqueue
554 * @node: the node ID
555 *
556 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
557 * read locked.
558 * If the pwq needs to be used beyond the locking in effect, the caller is
559 * responsible for guaranteeing that the pwq stays online.
560 *
561 * Return: The unbound pool_workqueue for @node.
562 */
563static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
564 int node)
565{
566 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
567
568 /*
569 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
570 * delayed item is pending. The plan is to keep CPU -> NODE
571 * mapping valid and stable across CPU on/offlines. Once that
572 * happens, this workaround can be removed.
573 */
574 if (unlikely(node == NUMA_NO_NODE))
575 return wq->dfl_pwq;
576
577 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
578}
579
580static unsigned int work_color_to_flags(int color)
581{
582 return color << WORK_STRUCT_COLOR_SHIFT;
583}
584
585static int get_work_color(struct work_struct *work)
586{
587 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
588 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
589}
590
591static int work_next_color(int color)
592{
593 return (color + 1) % WORK_NR_COLORS;
594}
595
596/*
597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
598 * contain the pointer to the queued pwq. Once execution starts, the flag
599 * is cleared and the high bits contain OFFQ flags and pool ID.
600 *
601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
602 * and clear_work_data() can be used to set the pwq, pool or clear
603 * work->data. These functions should only be called while the work is
604 * owned - ie. while the PENDING bit is set.
605 *
606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
607 * corresponding to a work. Pool is available once the work has been
608 * queued anywhere after initialization until it is sync canceled. pwq is
609 * available only while the work item is queued.
610 *
611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
612 * canceled. While being canceled, a work item may have its PENDING set
613 * but stay off timer and worklist for arbitrarily long and nobody should
614 * try to steal the PENDING bit.
615 */
616static inline void set_work_data(struct work_struct *work, unsigned long data,
617 unsigned long flags)
618{
619 WARN_ON_ONCE(!work_pending(work));
620 atomic_long_set(&work->data, data | flags | work_static(work));
621}
622
623static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
624 unsigned long extra_flags)
625{
626 set_work_data(work, (unsigned long)pwq,
627 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
628}
629
630static void set_work_pool_and_keep_pending(struct work_struct *work,
631 int pool_id)
632{
633 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
634 WORK_STRUCT_PENDING);
635}
636
637static void set_work_pool_and_clear_pending(struct work_struct *work,
638 int pool_id)
639{
640 /*
641 * The following wmb is paired with the implied mb in
642 * test_and_set_bit(PENDING) and ensures all updates to @work made
643 * here are visible to and precede any updates by the next PENDING
644 * owner.
645 */
646 smp_wmb();
647 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
648 /*
649 * The following mb guarantees that previous clear of a PENDING bit
650 * will not be reordered with any speculative LOADS or STORES from
651 * work->current_func, which is executed afterwards. This possible
652 * reordering can lead to a missed execution on attempt to queue
653 * the same @work. E.g. consider this case:
654 *
655 * CPU#0 CPU#1
656 * ---------------------------- --------------------------------
657 *
658 * 1 STORE event_indicated
659 * 2 queue_work_on() {
660 * 3 test_and_set_bit(PENDING)
661 * 4 } set_..._and_clear_pending() {
662 * 5 set_work_data() # clear bit
663 * 6 smp_mb()
664 * 7 work->current_func() {
665 * 8 LOAD event_indicated
666 * }
667 *
668 * Without an explicit full barrier speculative LOAD on line 8 can
669 * be executed before CPU#0 does STORE on line 1. If that happens,
670 * CPU#0 observes the PENDING bit is still set and new execution of
671 * a @work is not queued in a hope, that CPU#1 will eventually
672 * finish the queued @work. Meanwhile CPU#1 does not see
673 * event_indicated is set, because speculative LOAD was executed
674 * before actual STORE.
675 */
676 smp_mb();
677}
678
679static void clear_work_data(struct work_struct *work)
680{
681 smp_wmb(); /* see set_work_pool_and_clear_pending() */
682 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
683}
684
685static struct pool_workqueue *get_work_pwq(struct work_struct *work)
686{
687 unsigned long data = atomic_long_read(&work->data);
688
689 if (data & WORK_STRUCT_PWQ)
690 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
691 else
692 return NULL;
693}
694
695/**
696 * get_work_pool - return the worker_pool a given work was associated with
697 * @work: the work item of interest
698 *
699 * Pools are created and destroyed under wq_pool_mutex, and allows read
700 * access under RCU read lock. As such, this function should be
701 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
702 *
703 * All fields of the returned pool are accessible as long as the above
704 * mentioned locking is in effect. If the returned pool needs to be used
705 * beyond the critical section, the caller is responsible for ensuring the
706 * returned pool is and stays online.
707 *
708 * Return: The worker_pool @work was last associated with. %NULL if none.
709 */
710static struct worker_pool *get_work_pool(struct work_struct *work)
711{
712 unsigned long data = atomic_long_read(&work->data);
713 int pool_id;
714
715 assert_rcu_or_pool_mutex();
716
717 if (data & WORK_STRUCT_PWQ)
718 return ((struct pool_workqueue *)
719 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
720
721 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
722 if (pool_id == WORK_OFFQ_POOL_NONE)
723 return NULL;
724
725 return idr_find(&worker_pool_idr, pool_id);
726}
727
728/**
729 * get_work_pool_id - return the worker pool ID a given work is associated with
730 * @work: the work item of interest
731 *
732 * Return: The worker_pool ID @work was last associated with.
733 * %WORK_OFFQ_POOL_NONE if none.
734 */
735static int get_work_pool_id(struct work_struct *work)
736{
737 unsigned long data = atomic_long_read(&work->data);
738
739 if (data & WORK_STRUCT_PWQ)
740 return ((struct pool_workqueue *)
741 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
742
743 return data >> WORK_OFFQ_POOL_SHIFT;
744}
745
746static void mark_work_canceling(struct work_struct *work)
747{
748 unsigned long pool_id = get_work_pool_id(work);
749
750 pool_id <<= WORK_OFFQ_POOL_SHIFT;
751 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
752}
753
754static bool work_is_canceling(struct work_struct *work)
755{
756 unsigned long data = atomic_long_read(&work->data);
757
758 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
759}
760
761/*
762 * Policy functions. These define the policies on how the global worker
763 * pools are managed. Unless noted otherwise, these functions assume that
764 * they're being called with pool->lock held.
765 */
766
767static bool __need_more_worker(struct worker_pool *pool)
768{
769 return !atomic_read(&pool->nr_running);
770}
771
772/*
773 * Need to wake up a worker? Called from anything but currently
774 * running workers.
775 *
776 * Note that, because unbound workers never contribute to nr_running, this
777 * function will always return %true for unbound pools as long as the
778 * worklist isn't empty.
779 */
780static bool need_more_worker(struct worker_pool *pool)
781{
782 return !list_empty(&pool->worklist) && __need_more_worker(pool);
783}
784
785/* Can I start working? Called from busy but !running workers. */
786static bool may_start_working(struct worker_pool *pool)
787{
788 return pool->nr_idle;
789}
790
791/* Do I need to keep working? Called from currently running workers. */
792static bool keep_working(struct worker_pool *pool)
793{
794 return !list_empty(&pool->worklist) &&
795 atomic_read(&pool->nr_running) <= 1;
796}
797
798/* Do we need a new worker? Called from manager. */
799static bool need_to_create_worker(struct worker_pool *pool)
800{
801 return need_more_worker(pool) && !may_start_working(pool);
802}
803
804/* Do we have too many workers and should some go away? */
805static bool too_many_workers(struct worker_pool *pool)
806{
807 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
808 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
809 int nr_busy = pool->nr_workers - nr_idle;
810
811 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
812}
813
814/*
815 * Wake up functions.
816 */
817
818/* Return the first idle worker. Safe with preemption disabled */
819static struct worker *first_idle_worker(struct worker_pool *pool)
820{
821 if (unlikely(list_empty(&pool->idle_list)))
822 return NULL;
823
824 return list_first_entry(&pool->idle_list, struct worker, entry);
825}
826
827/**
828 * wake_up_worker - wake up an idle worker
829 * @pool: worker pool to wake worker from
830 *
831 * Wake up the first idle worker of @pool.
832 *
833 * CONTEXT:
834 * spin_lock_irq(pool->lock).
835 */
836static void wake_up_worker(struct worker_pool *pool)
837{
838 struct worker *worker = first_idle_worker(pool);
839
840 if (likely(worker))
841 wake_up_process(worker->task);
842}
843
844/**
845 * wq_worker_running - a worker is running again
846 * @task: task waking up
847 *
848 * This function is called when a worker returns from schedule()
849 */
850void wq_worker_running(struct task_struct *task)
851{
852 struct worker *worker = kthread_data(task);
853
854 if (!worker->sleeping)
855 return;
856 if (!(worker->flags & WORKER_NOT_RUNNING))
857 atomic_inc(&worker->pool->nr_running);
858 worker->sleeping = 0;
859}
860
861/**
862 * wq_worker_sleeping - a worker is going to sleep
863 * @task: task going to sleep
864 *
865 * This function is called from schedule() when a busy worker is
866 * going to sleep.
867 */
868void wq_worker_sleeping(struct task_struct *task)
869{
870 struct worker *next, *worker = kthread_data(task);
871 struct worker_pool *pool;
872
873 /*
874 * Rescuers, which may not have all the fields set up like normal
875 * workers, also reach here, let's not access anything before
876 * checking NOT_RUNNING.
877 */
878 if (worker->flags & WORKER_NOT_RUNNING)
879 return;
880
881 pool = worker->pool;
882
883 if (WARN_ON_ONCE(worker->sleeping))
884 return;
885
886 worker->sleeping = 1;
887 spin_lock_irq(&pool->lock);
888
889 /*
890 * The counterpart of the following dec_and_test, implied mb,
891 * worklist not empty test sequence is in insert_work().
892 * Please read comment there.
893 *
894 * NOT_RUNNING is clear. This means that we're bound to and
895 * running on the local cpu w/ rq lock held and preemption
896 * disabled, which in turn means that none else could be
897 * manipulating idle_list, so dereferencing idle_list without pool
898 * lock is safe.
899 */
900 if (atomic_dec_and_test(&pool->nr_running) &&
901 !list_empty(&pool->worklist)) {
902 next = first_idle_worker(pool);
903 if (next)
904 wake_up_process(next->task);
905 }
906 spin_unlock_irq(&pool->lock);
907}
908
909/**
910 * wq_worker_last_func - retrieve worker's last work function
911 * @task: Task to retrieve last work function of.
912 *
913 * Determine the last function a worker executed. This is called from
914 * the scheduler to get a worker's last known identity.
915 *
916 * CONTEXT:
917 * spin_lock_irq(rq->lock)
918 *
919 * This function is called during schedule() when a kworker is going
920 * to sleep. It's used by psi to identify aggregation workers during
921 * dequeuing, to allow periodic aggregation to shut-off when that
922 * worker is the last task in the system or cgroup to go to sleep.
923 *
924 * As this function doesn't involve any workqueue-related locking, it
925 * only returns stable values when called from inside the scheduler's
926 * queuing and dequeuing paths, when @task, which must be a kworker,
927 * is guaranteed to not be processing any works.
928 *
929 * Return:
930 * The last work function %current executed as a worker, NULL if it
931 * hasn't executed any work yet.
932 */
933work_func_t wq_worker_last_func(struct task_struct *task)
934{
935 struct worker *worker = kthread_data(task);
936
937 return worker->last_func;
938}
939
940/**
941 * worker_set_flags - set worker flags and adjust nr_running accordingly
942 * @worker: self
943 * @flags: flags to set
944 *
945 * Set @flags in @worker->flags and adjust nr_running accordingly.
946 *
947 * CONTEXT:
948 * spin_lock_irq(pool->lock)
949 */
950static inline void worker_set_flags(struct worker *worker, unsigned int flags)
951{
952 struct worker_pool *pool = worker->pool;
953
954 WARN_ON_ONCE(worker->task != current);
955
956 /* If transitioning into NOT_RUNNING, adjust nr_running. */
957 if ((flags & WORKER_NOT_RUNNING) &&
958 !(worker->flags & WORKER_NOT_RUNNING)) {
959 atomic_dec(&pool->nr_running);
960 }
961
962 worker->flags |= flags;
963}
964
965/**
966 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
967 * @worker: self
968 * @flags: flags to clear
969 *
970 * Clear @flags in @worker->flags and adjust nr_running accordingly.
971 *
972 * CONTEXT:
973 * spin_lock_irq(pool->lock)
974 */
975static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
976{
977 struct worker_pool *pool = worker->pool;
978 unsigned int oflags = worker->flags;
979
980 WARN_ON_ONCE(worker->task != current);
981
982 worker->flags &= ~flags;
983
984 /*
985 * If transitioning out of NOT_RUNNING, increment nr_running. Note
986 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
987 * of multiple flags, not a single flag.
988 */
989 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
990 if (!(worker->flags & WORKER_NOT_RUNNING))
991 atomic_inc(&pool->nr_running);
992}
993
994/**
995 * find_worker_executing_work - find worker which is executing a work
996 * @pool: pool of interest
997 * @work: work to find worker for
998 *
999 * Find a worker which is executing @work on @pool by searching
1000 * @pool->busy_hash which is keyed by the address of @work. For a worker
1001 * to match, its current execution should match the address of @work and
1002 * its work function. This is to avoid unwanted dependency between
1003 * unrelated work executions through a work item being recycled while still
1004 * being executed.
1005 *
1006 * This is a bit tricky. A work item may be freed once its execution
1007 * starts and nothing prevents the freed area from being recycled for
1008 * another work item. If the same work item address ends up being reused
1009 * before the original execution finishes, workqueue will identify the
1010 * recycled work item as currently executing and make it wait until the
1011 * current execution finishes, introducing an unwanted dependency.
1012 *
1013 * This function checks the work item address and work function to avoid
1014 * false positives. Note that this isn't complete as one may construct a
1015 * work function which can introduce dependency onto itself through a
1016 * recycled work item. Well, if somebody wants to shoot oneself in the
1017 * foot that badly, there's only so much we can do, and if such deadlock
1018 * actually occurs, it should be easy to locate the culprit work function.
1019 *
1020 * CONTEXT:
1021 * spin_lock_irq(pool->lock).
1022 *
1023 * Return:
1024 * Pointer to worker which is executing @work if found, %NULL
1025 * otherwise.
1026 */
1027static struct worker *find_worker_executing_work(struct worker_pool *pool,
1028 struct work_struct *work)
1029{
1030 struct worker *worker;
1031
1032 hash_for_each_possible(pool->busy_hash, worker, hentry,
1033 (unsigned long)work)
1034 if (worker->current_work == work &&
1035 worker->current_func == work->func)
1036 return worker;
1037
1038 return NULL;
1039}
1040
1041/**
1042 * move_linked_works - move linked works to a list
1043 * @work: start of series of works to be scheduled
1044 * @head: target list to append @work to
1045 * @nextp: out parameter for nested worklist walking
1046 *
1047 * Schedule linked works starting from @work to @head. Work series to
1048 * be scheduled starts at @work and includes any consecutive work with
1049 * WORK_STRUCT_LINKED set in its predecessor.
1050 *
1051 * If @nextp is not NULL, it's updated to point to the next work of
1052 * the last scheduled work. This allows move_linked_works() to be
1053 * nested inside outer list_for_each_entry_safe().
1054 *
1055 * CONTEXT:
1056 * spin_lock_irq(pool->lock).
1057 */
1058static void move_linked_works(struct work_struct *work, struct list_head *head,
1059 struct work_struct **nextp)
1060{
1061 struct work_struct *n;
1062
1063 /*
1064 * Linked worklist will always end before the end of the list,
1065 * use NULL for list head.
1066 */
1067 list_for_each_entry_safe_from(work, n, NULL, entry) {
1068 list_move_tail(&work->entry, head);
1069 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1070 break;
1071 }
1072
1073 /*
1074 * If we're already inside safe list traversal and have moved
1075 * multiple works to the scheduled queue, the next position
1076 * needs to be updated.
1077 */
1078 if (nextp)
1079 *nextp = n;
1080}
1081
1082/**
1083 * get_pwq - get an extra reference on the specified pool_workqueue
1084 * @pwq: pool_workqueue to get
1085 *
1086 * Obtain an extra reference on @pwq. The caller should guarantee that
1087 * @pwq has positive refcnt and be holding the matching pool->lock.
1088 */
1089static void get_pwq(struct pool_workqueue *pwq)
1090{
1091 lockdep_assert_held(&pwq->pool->lock);
1092 WARN_ON_ONCE(pwq->refcnt <= 0);
1093 pwq->refcnt++;
1094}
1095
1096/**
1097 * put_pwq - put a pool_workqueue reference
1098 * @pwq: pool_workqueue to put
1099 *
1100 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1101 * destruction. The caller should be holding the matching pool->lock.
1102 */
1103static void put_pwq(struct pool_workqueue *pwq)
1104{
1105 lockdep_assert_held(&pwq->pool->lock);
1106 if (likely(--pwq->refcnt))
1107 return;
1108 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1109 return;
1110 /*
1111 * @pwq can't be released under pool->lock, bounce to
1112 * pwq_unbound_release_workfn(). This never recurses on the same
1113 * pool->lock as this path is taken only for unbound workqueues and
1114 * the release work item is scheduled on a per-cpu workqueue. To
1115 * avoid lockdep warning, unbound pool->locks are given lockdep
1116 * subclass of 1 in get_unbound_pool().
1117 */
1118 schedule_work(&pwq->unbound_release_work);
1119}
1120
1121/**
1122 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1123 * @pwq: pool_workqueue to put (can be %NULL)
1124 *
1125 * put_pwq() with locking. This function also allows %NULL @pwq.
1126 */
1127static void put_pwq_unlocked(struct pool_workqueue *pwq)
1128{
1129 if (pwq) {
1130 /*
1131 * As both pwqs and pools are RCU protected, the
1132 * following lock operations are safe.
1133 */
1134 spin_lock_irq(&pwq->pool->lock);
1135 put_pwq(pwq);
1136 spin_unlock_irq(&pwq->pool->lock);
1137 }
1138}
1139
1140static void pwq_activate_delayed_work(struct work_struct *work)
1141{
1142 struct pool_workqueue *pwq = get_work_pwq(work);
1143
1144 trace_workqueue_activate_work(work);
1145 if (list_empty(&pwq->pool->worklist))
1146 pwq->pool->watchdog_ts = jiffies;
1147 move_linked_works(work, &pwq->pool->worklist, NULL);
1148 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1149 pwq->nr_active++;
1150}
1151
1152static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1153{
1154 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1155 struct work_struct, entry);
1156
1157 pwq_activate_delayed_work(work);
1158}
1159
1160/**
1161 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1162 * @pwq: pwq of interest
1163 * @color: color of work which left the queue
1164 *
1165 * A work either has completed or is removed from pending queue,
1166 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1167 *
1168 * CONTEXT:
1169 * spin_lock_irq(pool->lock).
1170 */
1171static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1172{
1173 /* uncolored work items don't participate in flushing or nr_active */
1174 if (color == WORK_NO_COLOR)
1175 goto out_put;
1176
1177 pwq->nr_in_flight[color]--;
1178
1179 pwq->nr_active--;
1180 if (!list_empty(&pwq->delayed_works)) {
1181 /* one down, submit a delayed one */
1182 if (pwq->nr_active < pwq->max_active)
1183 pwq_activate_first_delayed(pwq);
1184 }
1185
1186 /* is flush in progress and are we at the flushing tip? */
1187 if (likely(pwq->flush_color != color))
1188 goto out_put;
1189
1190 /* are there still in-flight works? */
1191 if (pwq->nr_in_flight[color])
1192 goto out_put;
1193
1194 /* this pwq is done, clear flush_color */
1195 pwq->flush_color = -1;
1196
1197 /*
1198 * If this was the last pwq, wake up the first flusher. It
1199 * will handle the rest.
1200 */
1201 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1202 complete(&pwq->wq->first_flusher->done);
1203out_put:
1204 put_pwq(pwq);
1205}
1206
1207/**
1208 * try_to_grab_pending - steal work item from worklist and disable irq
1209 * @work: work item to steal
1210 * @is_dwork: @work is a delayed_work
1211 * @flags: place to store irq state
1212 *
1213 * Try to grab PENDING bit of @work. This function can handle @work in any
1214 * stable state - idle, on timer or on worklist.
1215 *
1216 * Return:
1217 * 1 if @work was pending and we successfully stole PENDING
1218 * 0 if @work was idle and we claimed PENDING
1219 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1220 * -ENOENT if someone else is canceling @work, this state may persist
1221 * for arbitrarily long
1222 *
1223 * Note:
1224 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1225 * interrupted while holding PENDING and @work off queue, irq must be
1226 * disabled on entry. This, combined with delayed_work->timer being
1227 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1228 *
1229 * On successful return, >= 0, irq is disabled and the caller is
1230 * responsible for releasing it using local_irq_restore(*@flags).
1231 *
1232 * This function is safe to call from any context including IRQ handler.
1233 */
1234static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1235 unsigned long *flags)
1236{
1237 struct worker_pool *pool;
1238 struct pool_workqueue *pwq;
1239
1240 local_irq_save(*flags);
1241
1242 /* try to steal the timer if it exists */
1243 if (is_dwork) {
1244 struct delayed_work *dwork = to_delayed_work(work);
1245
1246 /*
1247 * dwork->timer is irqsafe. If del_timer() fails, it's
1248 * guaranteed that the timer is not queued anywhere and not
1249 * running on the local CPU.
1250 */
1251 if (likely(del_timer(&dwork->timer)))
1252 return 1;
1253 }
1254
1255 /* try to claim PENDING the normal way */
1256 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1257 return 0;
1258
1259 rcu_read_lock();
1260 /*
1261 * The queueing is in progress, or it is already queued. Try to
1262 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1263 */
1264 pool = get_work_pool(work);
1265 if (!pool)
1266 goto fail;
1267
1268 spin_lock(&pool->lock);
1269 /*
1270 * work->data is guaranteed to point to pwq only while the work
1271 * item is queued on pwq->wq, and both updating work->data to point
1272 * to pwq on queueing and to pool on dequeueing are done under
1273 * pwq->pool->lock. This in turn guarantees that, if work->data
1274 * points to pwq which is associated with a locked pool, the work
1275 * item is currently queued on that pool.
1276 */
1277 pwq = get_work_pwq(work);
1278 if (pwq && pwq->pool == pool) {
1279 debug_work_deactivate(work);
1280
1281 /*
1282 * A delayed work item cannot be grabbed directly because
1283 * it might have linked NO_COLOR work items which, if left
1284 * on the delayed_list, will confuse pwq->nr_active
1285 * management later on and cause stall. Make sure the work
1286 * item is activated before grabbing.
1287 */
1288 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1289 pwq_activate_delayed_work(work);
1290
1291 list_del_init(&work->entry);
1292 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1293
1294 /* work->data points to pwq iff queued, point to pool */
1295 set_work_pool_and_keep_pending(work, pool->id);
1296
1297 spin_unlock(&pool->lock);
1298 rcu_read_unlock();
1299 return 1;
1300 }
1301 spin_unlock(&pool->lock);
1302fail:
1303 rcu_read_unlock();
1304 local_irq_restore(*flags);
1305 if (work_is_canceling(work))
1306 return -ENOENT;
1307 cpu_relax();
1308 return -EAGAIN;
1309}
1310
1311/**
1312 * insert_work - insert a work into a pool
1313 * @pwq: pwq @work belongs to
1314 * @work: work to insert
1315 * @head: insertion point
1316 * @extra_flags: extra WORK_STRUCT_* flags to set
1317 *
1318 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1319 * work_struct flags.
1320 *
1321 * CONTEXT:
1322 * spin_lock_irq(pool->lock).
1323 */
1324static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1325 struct list_head *head, unsigned int extra_flags)
1326{
1327 struct worker_pool *pool = pwq->pool;
1328
1329 /* we own @work, set data and link */
1330 set_work_pwq(work, pwq, extra_flags);
1331 list_add_tail(&work->entry, head);
1332 get_pwq(pwq);
1333
1334 /*
1335 * Ensure either wq_worker_sleeping() sees the above
1336 * list_add_tail() or we see zero nr_running to avoid workers lying
1337 * around lazily while there are works to be processed.
1338 */
1339 smp_mb();
1340
1341 if (__need_more_worker(pool))
1342 wake_up_worker(pool);
1343}
1344
1345/*
1346 * Test whether @work is being queued from another work executing on the
1347 * same workqueue.
1348 */
1349static bool is_chained_work(struct workqueue_struct *wq)
1350{
1351 struct worker *worker;
1352
1353 worker = current_wq_worker();
1354 /*
1355 * Return %true iff I'm a worker executing a work item on @wq. If
1356 * I'm @worker, it's safe to dereference it without locking.
1357 */
1358 return worker && worker->current_pwq->wq == wq;
1359}
1360
1361/*
1362 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1363 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1364 * avoid perturbing sensitive tasks.
1365 */
1366static int wq_select_unbound_cpu(int cpu)
1367{
1368 static bool printed_dbg_warning;
1369 int new_cpu;
1370
1371 if (likely(!wq_debug_force_rr_cpu)) {
1372 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1373 return cpu;
1374 } else if (!printed_dbg_warning) {
1375 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1376 printed_dbg_warning = true;
1377 }
1378
1379 if (cpumask_empty(wq_unbound_cpumask))
1380 return cpu;
1381
1382 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1383 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1384 if (unlikely(new_cpu >= nr_cpu_ids)) {
1385 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1386 if (unlikely(new_cpu >= nr_cpu_ids))
1387 return cpu;
1388 }
1389 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1390
1391 return new_cpu;
1392}
1393
1394static void __queue_work(int cpu, struct workqueue_struct *wq,
1395 struct work_struct *work)
1396{
1397 struct pool_workqueue *pwq;
1398 struct worker_pool *last_pool;
1399 struct list_head *worklist;
1400 unsigned int work_flags;
1401 unsigned int req_cpu = cpu;
1402
1403 /*
1404 * While a work item is PENDING && off queue, a task trying to
1405 * steal the PENDING will busy-loop waiting for it to either get
1406 * queued or lose PENDING. Grabbing PENDING and queueing should
1407 * happen with IRQ disabled.
1408 */
1409 lockdep_assert_irqs_disabled();
1410
1411 debug_work_activate(work);
1412
1413 /* if draining, only works from the same workqueue are allowed */
1414 if (unlikely(wq->flags & __WQ_DRAINING) &&
1415 WARN_ON_ONCE(!is_chained_work(wq)))
1416 return;
1417 rcu_read_lock();
1418retry:
1419 if (req_cpu == WORK_CPU_UNBOUND)
1420 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1421
1422 /* pwq which will be used unless @work is executing elsewhere */
1423 if (!(wq->flags & WQ_UNBOUND))
1424 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1425 else
1426 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1427
1428 /*
1429 * If @work was previously on a different pool, it might still be
1430 * running there, in which case the work needs to be queued on that
1431 * pool to guarantee non-reentrancy.
1432 */
1433 last_pool = get_work_pool(work);
1434 if (last_pool && last_pool != pwq->pool) {
1435 struct worker *worker;
1436
1437 spin_lock(&last_pool->lock);
1438
1439 worker = find_worker_executing_work(last_pool, work);
1440
1441 if (worker && worker->current_pwq->wq == wq) {
1442 pwq = worker->current_pwq;
1443 } else {
1444 /* meh... not running there, queue here */
1445 spin_unlock(&last_pool->lock);
1446 spin_lock(&pwq->pool->lock);
1447 }
1448 } else {
1449 spin_lock(&pwq->pool->lock);
1450 }
1451
1452 /*
1453 * pwq is determined and locked. For unbound pools, we could have
1454 * raced with pwq release and it could already be dead. If its
1455 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1456 * without another pwq replacing it in the numa_pwq_tbl or while
1457 * work items are executing on it, so the retrying is guaranteed to
1458 * make forward-progress.
1459 */
1460 if (unlikely(!pwq->refcnt)) {
1461 if (wq->flags & WQ_UNBOUND) {
1462 spin_unlock(&pwq->pool->lock);
1463 cpu_relax();
1464 goto retry;
1465 }
1466 /* oops */
1467 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1468 wq->name, cpu);
1469 }
1470
1471 /* pwq determined, queue */
1472 trace_workqueue_queue_work(req_cpu, pwq, work);
1473
1474 if (WARN_ON(!list_empty(&work->entry)))
1475 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 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 * spin_lock_irq(pool->lock).
1764 */
1765static void worker_enter_idle(struct worker *worker)
1766{
1767 struct worker_pool *pool = worker->pool;
1768
1769 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1770 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1771 (worker->hentry.next || worker->hentry.pprev)))
1772 return;
1773
1774 /* can't use worker_set_flags(), also called from create_worker() */
1775 worker->flags |= WORKER_IDLE;
1776 pool->nr_idle++;
1777 worker->last_active = jiffies;
1778
1779 /* idle_list is LIFO */
1780 list_add(&worker->entry, &pool->idle_list);
1781
1782 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1783 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1784
1785 /*
1786 * Sanity check nr_running. Because unbind_workers() releases
1787 * pool->lock between setting %WORKER_UNBOUND and zapping
1788 * nr_running, the warning may trigger spuriously. Check iff
1789 * unbind is not in progress.
1790 */
1791 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1792 pool->nr_workers == pool->nr_idle &&
1793 atomic_read(&pool->nr_running));
1794}
1795
1796/**
1797 * worker_leave_idle - leave idle state
1798 * @worker: worker which is leaving idle state
1799 *
1800 * @worker is leaving idle state. Update stats.
1801 *
1802 * LOCKING:
1803 * spin_lock_irq(pool->lock).
1804 */
1805static void worker_leave_idle(struct worker *worker)
1806{
1807 struct worker_pool *pool = worker->pool;
1808
1809 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1810 return;
1811 worker_clr_flags(worker, WORKER_IDLE);
1812 pool->nr_idle--;
1813 list_del_init(&worker->entry);
1814}
1815
1816static struct worker *alloc_worker(int node)
1817{
1818 struct worker *worker;
1819
1820 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1821 if (worker) {
1822 INIT_LIST_HEAD(&worker->entry);
1823 INIT_LIST_HEAD(&worker->scheduled);
1824 INIT_LIST_HEAD(&worker->node);
1825 /* on creation a worker is in !idle && prep state */
1826 worker->flags = WORKER_PREP;
1827 }
1828 return worker;
1829}
1830
1831/**
1832 * worker_attach_to_pool() - attach a worker to a pool
1833 * @worker: worker to be attached
1834 * @pool: the target pool
1835 *
1836 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1837 * cpu-binding of @worker are kept coordinated with the pool across
1838 * cpu-[un]hotplugs.
1839 */
1840static void worker_attach_to_pool(struct worker *worker,
1841 struct worker_pool *pool)
1842{
1843 mutex_lock(&wq_pool_attach_mutex);
1844
1845 /*
1846 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1847 * online CPUs. It'll be re-applied when any of the CPUs come up.
1848 */
1849 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1850
1851 /*
1852 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1853 * stable across this function. See the comments above the flag
1854 * definition for details.
1855 */
1856 if (pool->flags & POOL_DISASSOCIATED)
1857 worker->flags |= WORKER_UNBOUND;
1858
1859 list_add_tail(&worker->node, &pool->workers);
1860 worker->pool = pool;
1861
1862 mutex_unlock(&wq_pool_attach_mutex);
1863}
1864
1865/**
1866 * worker_detach_from_pool() - detach a worker from its pool
1867 * @worker: worker which is attached to its pool
1868 *
1869 * Undo the attaching which had been done in worker_attach_to_pool(). The
1870 * caller worker shouldn't access to the pool after detached except it has
1871 * other reference to the pool.
1872 */
1873static void worker_detach_from_pool(struct worker *worker)
1874{
1875 struct worker_pool *pool = worker->pool;
1876 struct completion *detach_completion = NULL;
1877
1878 mutex_lock(&wq_pool_attach_mutex);
1879
1880 list_del(&worker->node);
1881 worker->pool = NULL;
1882
1883 if (list_empty(&pool->workers))
1884 detach_completion = pool->detach_completion;
1885 mutex_unlock(&wq_pool_attach_mutex);
1886
1887 /* clear leftover flags without pool->lock after it is detached */
1888 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1889
1890 if (detach_completion)
1891 complete(detach_completion);
1892}
1893
1894/**
1895 * create_worker - create a new workqueue worker
1896 * @pool: pool the new worker will belong to
1897 *
1898 * Create and start a new worker which is attached to @pool.
1899 *
1900 * CONTEXT:
1901 * Might sleep. Does GFP_KERNEL allocations.
1902 *
1903 * Return:
1904 * Pointer to the newly created worker.
1905 */
1906static struct worker *create_worker(struct worker_pool *pool)
1907{
1908 struct worker *worker = NULL;
1909 int id = -1;
1910 char id_buf[16];
1911
1912 /* ID is needed to determine kthread name */
1913 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1914 if (id < 0)
1915 goto fail;
1916
1917 worker = alloc_worker(pool->node);
1918 if (!worker)
1919 goto fail;
1920
1921 worker->id = id;
1922
1923 if (pool->cpu >= 0)
1924 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1925 pool->attrs->nice < 0 ? "H" : "");
1926 else
1927 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1928
1929 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1930 "kworker/%s", id_buf);
1931 if (IS_ERR(worker->task))
1932 goto fail;
1933
1934 set_user_nice(worker->task, pool->attrs->nice);
1935 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1936
1937 /* successful, attach the worker to the pool */
1938 worker_attach_to_pool(worker, pool);
1939
1940 /* start the newly created worker */
1941 spin_lock_irq(&pool->lock);
1942 worker->pool->nr_workers++;
1943 worker_enter_idle(worker);
1944 wake_up_process(worker->task);
1945 spin_unlock_irq(&pool->lock);
1946
1947 return worker;
1948
1949fail:
1950 if (id >= 0)
1951 ida_simple_remove(&pool->worker_ida, id);
1952 kfree(worker);
1953 return NULL;
1954}
1955
1956/**
1957 * destroy_worker - destroy a workqueue worker
1958 * @worker: worker to be destroyed
1959 *
1960 * Destroy @worker and adjust @pool stats accordingly. The worker should
1961 * be idle.
1962 *
1963 * CONTEXT:
1964 * spin_lock_irq(pool->lock).
1965 */
1966static void destroy_worker(struct worker *worker)
1967{
1968 struct worker_pool *pool = worker->pool;
1969
1970 lockdep_assert_held(&pool->lock);
1971
1972 /* sanity check frenzy */
1973 if (WARN_ON(worker->current_work) ||
1974 WARN_ON(!list_empty(&worker->scheduled)) ||
1975 WARN_ON(!(worker->flags & WORKER_IDLE)))
1976 return;
1977
1978 pool->nr_workers--;
1979 pool->nr_idle--;
1980
1981 list_del_init(&worker->entry);
1982 worker->flags |= WORKER_DIE;
1983 wake_up_process(worker->task);
1984}
1985
1986static void idle_worker_timeout(struct timer_list *t)
1987{
1988 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1989
1990 spin_lock_irq(&pool->lock);
1991
1992 while (too_many_workers(pool)) {
1993 struct worker *worker;
1994 unsigned long expires;
1995
1996 /* idle_list is kept in LIFO order, check the last one */
1997 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1998 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1999
2000 if (time_before(jiffies, expires)) {
2001 mod_timer(&pool->idle_timer, expires);
2002 break;
2003 }
2004
2005 destroy_worker(worker);
2006 }
2007
2008 spin_unlock_irq(&pool->lock);
2009}
2010
2011static void send_mayday(struct work_struct *work)
2012{
2013 struct pool_workqueue *pwq = get_work_pwq(work);
2014 struct workqueue_struct *wq = pwq->wq;
2015
2016 lockdep_assert_held(&wq_mayday_lock);
2017
2018 if (!wq->rescuer)
2019 return;
2020
2021 /* mayday mayday mayday */
2022 if (list_empty(&pwq->mayday_node)) {
2023 /*
2024 * If @pwq is for an unbound wq, its base ref may be put at
2025 * any time due to an attribute change. Pin @pwq until the
2026 * rescuer is done with it.
2027 */
2028 get_pwq(pwq);
2029 list_add_tail(&pwq->mayday_node, &wq->maydays);
2030 wake_up_process(wq->rescuer->task);
2031 }
2032}
2033
2034static void pool_mayday_timeout(struct timer_list *t)
2035{
2036 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2037 struct work_struct *work;
2038
2039 spin_lock_irq(&pool->lock);
2040 spin_lock(&wq_mayday_lock); /* for wq->maydays */
2041
2042 if (need_to_create_worker(pool)) {
2043 /*
2044 * We've been trying to create a new worker but
2045 * haven't been successful. We might be hitting an
2046 * allocation deadlock. Send distress signals to
2047 * rescuers.
2048 */
2049 list_for_each_entry(work, &pool->worklist, entry)
2050 send_mayday(work);
2051 }
2052
2053 spin_unlock(&wq_mayday_lock);
2054 spin_unlock_irq(&pool->lock);
2055
2056 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2057}
2058
2059/**
2060 * maybe_create_worker - create a new worker if necessary
2061 * @pool: pool to create a new worker for
2062 *
2063 * Create a new worker for @pool if necessary. @pool is guaranteed to
2064 * have at least one idle worker on return from this function. If
2065 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2066 * sent to all rescuers with works scheduled on @pool to resolve
2067 * possible allocation deadlock.
2068 *
2069 * On return, need_to_create_worker() is guaranteed to be %false and
2070 * may_start_working() %true.
2071 *
2072 * LOCKING:
2073 * spin_lock_irq(pool->lock) which may be released and regrabbed
2074 * multiple times. Does GFP_KERNEL allocations. Called only from
2075 * manager.
2076 */
2077static void maybe_create_worker(struct worker_pool *pool)
2078__releases(&pool->lock)
2079__acquires(&pool->lock)
2080{
2081restart:
2082 spin_unlock_irq(&pool->lock);
2083
2084 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2085 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2086
2087 while (true) {
2088 if (create_worker(pool) || !need_to_create_worker(pool))
2089 break;
2090
2091 schedule_timeout_interruptible(CREATE_COOLDOWN);
2092
2093 if (!need_to_create_worker(pool))
2094 break;
2095 }
2096
2097 del_timer_sync(&pool->mayday_timer);
2098 spin_lock_irq(&pool->lock);
2099 /*
2100 * This is necessary even after a new worker was just successfully
2101 * created as @pool->lock was dropped and the new worker might have
2102 * already become busy.
2103 */
2104 if (need_to_create_worker(pool))
2105 goto restart;
2106}
2107
2108/**
2109 * manage_workers - manage worker pool
2110 * @worker: self
2111 *
2112 * Assume the manager role and manage the worker pool @worker belongs
2113 * to. At any given time, there can be only zero or one manager per
2114 * pool. The exclusion is handled automatically by this function.
2115 *
2116 * The caller can safely start processing works on false return. On
2117 * true return, it's guaranteed that need_to_create_worker() is false
2118 * and may_start_working() is true.
2119 *
2120 * CONTEXT:
2121 * spin_lock_irq(pool->lock) which may be released and regrabbed
2122 * multiple times. Does GFP_KERNEL allocations.
2123 *
2124 * Return:
2125 * %false if the pool doesn't need management and the caller can safely
2126 * start processing works, %true if management function was performed and
2127 * the conditions that the caller verified before calling the function may
2128 * no longer be true.
2129 */
2130static bool manage_workers(struct worker *worker)
2131{
2132 struct worker_pool *pool = worker->pool;
2133
2134 if (pool->flags & POOL_MANAGER_ACTIVE)
2135 return false;
2136
2137 pool->flags |= POOL_MANAGER_ACTIVE;
2138 pool->manager = worker;
2139
2140 maybe_create_worker(pool);
2141
2142 pool->manager = NULL;
2143 pool->flags &= ~POOL_MANAGER_ACTIVE;
2144 wake_up(&wq_manager_wait);
2145 return true;
2146}
2147
2148/**
2149 * process_one_work - process single work
2150 * @worker: self
2151 * @work: work to process
2152 *
2153 * Process @work. This function contains all the logics necessary to
2154 * process a single work including synchronization against and
2155 * interaction with other workers on the same cpu, queueing and
2156 * flushing. As long as context requirement is met, any worker can
2157 * call this function to process a work.
2158 *
2159 * CONTEXT:
2160 * spin_lock_irq(pool->lock) which is released and regrabbed.
2161 */
2162static void process_one_work(struct worker *worker, struct work_struct *work)
2163__releases(&pool->lock)
2164__acquires(&pool->lock)
2165{
2166 struct pool_workqueue *pwq = get_work_pwq(work);
2167 struct worker_pool *pool = worker->pool;
2168 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2169 int work_color;
2170 struct worker *collision;
2171#ifdef CONFIG_LOCKDEP
2172 /*
2173 * It is permissible to free the struct work_struct from
2174 * inside the function that is called from it, this we need to
2175 * take into account for lockdep too. To avoid bogus "held
2176 * lock freed" warnings as well as problems when looking into
2177 * work->lockdep_map, make a copy and use that here.
2178 */
2179 struct lockdep_map lockdep_map;
2180
2181 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2182#endif
2183 /* ensure we're on the correct CPU */
2184 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2185 raw_smp_processor_id() != pool->cpu);
2186
2187 /*
2188 * A single work shouldn't be executed concurrently by
2189 * multiple workers on a single cpu. Check whether anyone is
2190 * already processing the work. If so, defer the work to the
2191 * currently executing one.
2192 */
2193 collision = find_worker_executing_work(pool, work);
2194 if (unlikely(collision)) {
2195 move_linked_works(work, &collision->scheduled, NULL);
2196 return;
2197 }
2198
2199 /* claim and dequeue */
2200 debug_work_deactivate(work);
2201 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2202 worker->current_work = work;
2203 worker->current_func = work->func;
2204 worker->current_pwq = pwq;
2205 work_color = get_work_color(work);
2206
2207 /*
2208 * Record wq name for cmdline and debug reporting, may get
2209 * overridden through set_worker_desc().
2210 */
2211 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2212
2213 list_del_init(&work->entry);
2214
2215 /*
2216 * CPU intensive works don't participate in concurrency management.
2217 * They're the scheduler's responsibility. This takes @worker out
2218 * of concurrency management and the next code block will chain
2219 * execution of the pending work items.
2220 */
2221 if (unlikely(cpu_intensive))
2222 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2223
2224 /*
2225 * Wake up another worker if necessary. The condition is always
2226 * false for normal per-cpu workers since nr_running would always
2227 * be >= 1 at this point. This is used to chain execution of the
2228 * pending work items for WORKER_NOT_RUNNING workers such as the
2229 * UNBOUND and CPU_INTENSIVE ones.
2230 */
2231 if (need_more_worker(pool))
2232 wake_up_worker(pool);
2233
2234 /*
2235 * Record the last pool and clear PENDING which should be the last
2236 * update to @work. Also, do this inside @pool->lock so that
2237 * PENDING and queued state changes happen together while IRQ is
2238 * disabled.
2239 */
2240 set_work_pool_and_clear_pending(work, pool->id);
2241
2242 spin_unlock_irq(&pool->lock);
2243
2244 lock_map_acquire(&pwq->wq->lockdep_map);
2245 lock_map_acquire(&lockdep_map);
2246 /*
2247 * Strictly speaking we should mark the invariant state without holding
2248 * any locks, that is, before these two lock_map_acquire()'s.
2249 *
2250 * However, that would result in:
2251 *
2252 * A(W1)
2253 * WFC(C)
2254 * A(W1)
2255 * C(C)
2256 *
2257 * Which would create W1->C->W1 dependencies, even though there is no
2258 * actual deadlock possible. There are two solutions, using a
2259 * read-recursive acquire on the work(queue) 'locks', but this will then
2260 * hit the lockdep limitation on recursive locks, or simply discard
2261 * these locks.
2262 *
2263 * AFAICT there is no possible deadlock scenario between the
2264 * flush_work() and complete() primitives (except for single-threaded
2265 * workqueues), so hiding them isn't a problem.
2266 */
2267 lockdep_invariant_state(true);
2268 trace_workqueue_execute_start(work);
2269 worker->current_func(work);
2270 /*
2271 * While we must be careful to not use "work" after this, the trace
2272 * point will only record its address.
2273 */
2274 trace_workqueue_execute_end(work);
2275 lock_map_release(&lockdep_map);
2276 lock_map_release(&pwq->wq->lockdep_map);
2277
2278 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2279 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2280 " last function: %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 !PREEMPT
2289 * kernels, where a requeueing work item waiting for something to
2290 * happen could deadlock with stop_machine as such work item could
2291 * indefinitely requeue itself while all other CPUs are trapped in
2292 * stop_machine. At the same time, report a quiescent RCU state so
2293 * the same condition doesn't freeze RCU.
2294 */
2295 cond_resched();
2296
2297 spin_lock_irq(&pool->lock);
2298
2299 /* clear cpu intensive status */
2300 if (unlikely(cpu_intensive))
2301 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2302
2303 /* tag the worker for identification in schedule() */
2304 worker->last_func = worker->current_func;
2305
2306 /* we're done with it, release */
2307 hash_del(&worker->hentry);
2308 worker->current_work = NULL;
2309 worker->current_func = NULL;
2310 worker->current_pwq = NULL;
2311 pwq_dec_nr_in_flight(pwq, work_color);
2312}
2313
2314/**
2315 * process_scheduled_works - process scheduled works
2316 * @worker: self
2317 *
2318 * Process all scheduled works. Please note that the scheduled list
2319 * may change while processing a work, so this function repeatedly
2320 * fetches a work from the top and executes it.
2321 *
2322 * CONTEXT:
2323 * spin_lock_irq(pool->lock) which may be released and regrabbed
2324 * multiple times.
2325 */
2326static void process_scheduled_works(struct worker *worker)
2327{
2328 while (!list_empty(&worker->scheduled)) {
2329 struct work_struct *work = list_first_entry(&worker->scheduled,
2330 struct work_struct, entry);
2331 process_one_work(worker, work);
2332 }
2333}
2334
2335static void set_pf_worker(bool val)
2336{
2337 mutex_lock(&wq_pool_attach_mutex);
2338 if (val)
2339 current->flags |= PF_WQ_WORKER;
2340 else
2341 current->flags &= ~PF_WQ_WORKER;
2342 mutex_unlock(&wq_pool_attach_mutex);
2343}
2344
2345/**
2346 * worker_thread - the worker thread function
2347 * @__worker: self
2348 *
2349 * The worker thread function. All workers belong to a worker_pool -
2350 * either a per-cpu one or dynamic unbound one. These workers process all
2351 * work items regardless of their specific target workqueue. The only
2352 * exception is work items which belong to workqueues with a rescuer which
2353 * will be explained in rescuer_thread().
2354 *
2355 * Return: 0
2356 */
2357static int worker_thread(void *__worker)
2358{
2359 struct worker *worker = __worker;
2360 struct worker_pool *pool = worker->pool;
2361
2362 /* tell the scheduler that this is a workqueue worker */
2363 set_pf_worker(true);
2364woke_up:
2365 spin_lock_irq(&pool->lock);
2366
2367 /* am I supposed to die? */
2368 if (unlikely(worker->flags & WORKER_DIE)) {
2369 spin_unlock_irq(&pool->lock);
2370 WARN_ON_ONCE(!list_empty(&worker->entry));
2371 set_pf_worker(false);
2372
2373 set_task_comm(worker->task, "kworker/dying");
2374 ida_simple_remove(&pool->worker_ida, worker->id);
2375 worker_detach_from_pool(worker);
2376 kfree(worker);
2377 return 0;
2378 }
2379
2380 worker_leave_idle(worker);
2381recheck:
2382 /* no more worker necessary? */
2383 if (!need_more_worker(pool))
2384 goto sleep;
2385
2386 /* do we need to manage? */
2387 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2388 goto recheck;
2389
2390 /*
2391 * ->scheduled list can only be filled while a worker is
2392 * preparing to process a work or actually processing it.
2393 * Make sure nobody diddled with it while I was sleeping.
2394 */
2395 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2396
2397 /*
2398 * Finish PREP stage. We're guaranteed to have at least one idle
2399 * worker or that someone else has already assumed the manager
2400 * role. This is where @worker starts participating in concurrency
2401 * management if applicable and concurrency management is restored
2402 * after being rebound. See rebind_workers() for details.
2403 */
2404 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2405
2406 do {
2407 struct work_struct *work =
2408 list_first_entry(&pool->worklist,
2409 struct work_struct, entry);
2410
2411 pool->watchdog_ts = jiffies;
2412
2413 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2414 /* optimization path, not strictly necessary */
2415 process_one_work(worker, work);
2416 if (unlikely(!list_empty(&worker->scheduled)))
2417 process_scheduled_works(worker);
2418 } else {
2419 move_linked_works(work, &worker->scheduled, NULL);
2420 process_scheduled_works(worker);
2421 }
2422 } while (keep_working(pool));
2423
2424 worker_set_flags(worker, WORKER_PREP);
2425sleep:
2426 /*
2427 * pool->lock is held and there's no work to process and no need to
2428 * manage, sleep. Workers are woken up only while holding
2429 * pool->lock or from local cpu, so setting the current state
2430 * before releasing pool->lock is enough to prevent losing any
2431 * event.
2432 */
2433 worker_enter_idle(worker);
2434 __set_current_state(TASK_IDLE);
2435 spin_unlock_irq(&pool->lock);
2436 schedule();
2437 goto woke_up;
2438}
2439
2440/**
2441 * rescuer_thread - the rescuer thread function
2442 * @__rescuer: self
2443 *
2444 * Workqueue rescuer thread function. There's one rescuer for each
2445 * workqueue which has WQ_MEM_RECLAIM set.
2446 *
2447 * Regular work processing on a pool may block trying to create a new
2448 * worker which uses GFP_KERNEL allocation which has slight chance of
2449 * developing into deadlock if some works currently on the same queue
2450 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2451 * the problem rescuer solves.
2452 *
2453 * When such condition is possible, the pool summons rescuers of all
2454 * workqueues which have works queued on the pool and let them process
2455 * those works so that forward progress can be guaranteed.
2456 *
2457 * This should happen rarely.
2458 *
2459 * Return: 0
2460 */
2461static int rescuer_thread(void *__rescuer)
2462{
2463 struct worker *rescuer = __rescuer;
2464 struct workqueue_struct *wq = rescuer->rescue_wq;
2465 struct list_head *scheduled = &rescuer->scheduled;
2466 bool should_stop;
2467
2468 set_user_nice(current, RESCUER_NICE_LEVEL);
2469
2470 /*
2471 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2472 * doesn't participate in concurrency management.
2473 */
2474 set_pf_worker(true);
2475repeat:
2476 set_current_state(TASK_IDLE);
2477
2478 /*
2479 * By the time the rescuer is requested to stop, the workqueue
2480 * shouldn't have any work pending, but @wq->maydays may still have
2481 * pwq(s) queued. This can happen by non-rescuer workers consuming
2482 * all the work items before the rescuer got to them. Go through
2483 * @wq->maydays processing before acting on should_stop so that the
2484 * list is always empty on exit.
2485 */
2486 should_stop = kthread_should_stop();
2487
2488 /* see whether any pwq is asking for help */
2489 spin_lock_irq(&wq_mayday_lock);
2490
2491 while (!list_empty(&wq->maydays)) {
2492 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2493 struct pool_workqueue, mayday_node);
2494 struct worker_pool *pool = pwq->pool;
2495 struct work_struct *work, *n;
2496 bool first = true;
2497
2498 __set_current_state(TASK_RUNNING);
2499 list_del_init(&pwq->mayday_node);
2500
2501 spin_unlock_irq(&wq_mayday_lock);
2502
2503 worker_attach_to_pool(rescuer, pool);
2504
2505 spin_lock_irq(&pool->lock);
2506
2507 /*
2508 * Slurp in all works issued via this workqueue and
2509 * process'em.
2510 */
2511 WARN_ON_ONCE(!list_empty(scheduled));
2512 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2513 if (get_work_pwq(work) == pwq) {
2514 if (first)
2515 pool->watchdog_ts = jiffies;
2516 move_linked_works(work, scheduled, &n);
2517 }
2518 first = false;
2519 }
2520
2521 if (!list_empty(scheduled)) {
2522 process_scheduled_works(rescuer);
2523
2524 /*
2525 * The above execution of rescued work items could
2526 * have created more to rescue through
2527 * pwq_activate_first_delayed() or chained
2528 * queueing. Let's put @pwq back on mayday list so
2529 * that such back-to-back work items, which may be
2530 * being used to relieve memory pressure, don't
2531 * incur MAYDAY_INTERVAL delay inbetween.
2532 */
2533 if (need_to_create_worker(pool)) {
2534 spin_lock(&wq_mayday_lock);
2535 get_pwq(pwq);
2536 list_move_tail(&pwq->mayday_node, &wq->maydays);
2537 spin_unlock(&wq_mayday_lock);
2538 }
2539 }
2540
2541 /*
2542 * Put the reference grabbed by send_mayday(). @pool won't
2543 * go away while we're still attached to it.
2544 */
2545 put_pwq(pwq);
2546
2547 /*
2548 * Leave this pool. If need_more_worker() is %true, notify a
2549 * regular worker; otherwise, we end up with 0 concurrency
2550 * and stalling the execution.
2551 */
2552 if (need_more_worker(pool))
2553 wake_up_worker(pool);
2554
2555 spin_unlock_irq(&pool->lock);
2556
2557 worker_detach_from_pool(rescuer);
2558
2559 spin_lock_irq(&wq_mayday_lock);
2560 }
2561
2562 spin_unlock_irq(&wq_mayday_lock);
2563
2564 if (should_stop) {
2565 __set_current_state(TASK_RUNNING);
2566 set_pf_worker(false);
2567 return 0;
2568 }
2569
2570 /* rescuers should never participate in concurrency management */
2571 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2572 schedule();
2573 goto repeat;
2574}
2575
2576/**
2577 * check_flush_dependency - check for flush dependency sanity
2578 * @target_wq: workqueue being flushed
2579 * @target_work: work item being flushed (NULL for workqueue flushes)
2580 *
2581 * %current is trying to flush the whole @target_wq or @target_work on it.
2582 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2583 * reclaiming memory or running on a workqueue which doesn't have
2584 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2585 * a deadlock.
2586 */
2587static void check_flush_dependency(struct workqueue_struct *target_wq,
2588 struct work_struct *target_work)
2589{
2590 work_func_t target_func = target_work ? target_work->func : NULL;
2591 struct worker *worker;
2592
2593 if (target_wq->flags & WQ_MEM_RECLAIM)
2594 return;
2595
2596 worker = current_wq_worker();
2597
2598 WARN_ONCE(current->flags & PF_MEMALLOC,
2599 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2600 current->pid, current->comm, target_wq->name, target_func);
2601 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2602 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2603 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2604 worker->current_pwq->wq->name, worker->current_func,
2605 target_wq->name, target_func);
2606}
2607
2608struct wq_barrier {
2609 struct work_struct work;
2610 struct completion done;
2611 struct task_struct *task; /* purely informational */
2612};
2613
2614static void wq_barrier_func(struct work_struct *work)
2615{
2616 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2617 complete(&barr->done);
2618}
2619
2620/**
2621 * insert_wq_barrier - insert a barrier work
2622 * @pwq: pwq to insert barrier into
2623 * @barr: wq_barrier to insert
2624 * @target: target work to attach @barr to
2625 * @worker: worker currently executing @target, NULL if @target is not executing
2626 *
2627 * @barr is linked to @target such that @barr is completed only after
2628 * @target finishes execution. Please note that the ordering
2629 * guarantee is observed only with respect to @target and on the local
2630 * cpu.
2631 *
2632 * Currently, a queued barrier can't be canceled. This is because
2633 * try_to_grab_pending() can't determine whether the work to be
2634 * grabbed is at the head of the queue and thus can't clear LINKED
2635 * flag of the previous work while there must be a valid next work
2636 * after a work with LINKED flag set.
2637 *
2638 * Note that when @worker is non-NULL, @target may be modified
2639 * underneath us, so we can't reliably determine pwq from @target.
2640 *
2641 * CONTEXT:
2642 * spin_lock_irq(pool->lock).
2643 */
2644static void insert_wq_barrier(struct pool_workqueue *pwq,
2645 struct wq_barrier *barr,
2646 struct work_struct *target, struct worker *worker)
2647{
2648 struct list_head *head;
2649 unsigned int linked = 0;
2650
2651 /*
2652 * debugobject calls are safe here even with pool->lock locked
2653 * as we know for sure that this will not trigger any of the
2654 * checks and call back into the fixup functions where we
2655 * might deadlock.
2656 */
2657 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2658 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2659
2660 init_completion_map(&barr->done, &target->lockdep_map);
2661
2662 barr->task = current;
2663
2664 /*
2665 * If @target is currently being executed, schedule the
2666 * barrier to the worker; otherwise, put it after @target.
2667 */
2668 if (worker)
2669 head = worker->scheduled.next;
2670 else {
2671 unsigned long *bits = work_data_bits(target);
2672
2673 head = target->entry.next;
2674 /* there can already be other linked works, inherit and set */
2675 linked = *bits & WORK_STRUCT_LINKED;
2676 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2677 }
2678
2679 debug_work_activate(&barr->work);
2680 insert_work(pwq, &barr->work, head,
2681 work_color_to_flags(WORK_NO_COLOR) | linked);
2682}
2683
2684/**
2685 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2686 * @wq: workqueue being flushed
2687 * @flush_color: new flush color, < 0 for no-op
2688 * @work_color: new work color, < 0 for no-op
2689 *
2690 * Prepare pwqs for workqueue flushing.
2691 *
2692 * If @flush_color is non-negative, flush_color on all pwqs should be
2693 * -1. If no pwq has in-flight commands at the specified color, all
2694 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2695 * has in flight commands, its pwq->flush_color is set to
2696 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2697 * wakeup logic is armed and %true is returned.
2698 *
2699 * The caller should have initialized @wq->first_flusher prior to
2700 * calling this function with non-negative @flush_color. If
2701 * @flush_color is negative, no flush color update is done and %false
2702 * is returned.
2703 *
2704 * If @work_color is non-negative, all pwqs should have the same
2705 * work_color which is previous to @work_color and all will be
2706 * advanced to @work_color.
2707 *
2708 * CONTEXT:
2709 * mutex_lock(wq->mutex).
2710 *
2711 * Return:
2712 * %true if @flush_color >= 0 and there's something to flush. %false
2713 * otherwise.
2714 */
2715static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2716 int flush_color, int work_color)
2717{
2718 bool wait = false;
2719 struct pool_workqueue *pwq;
2720
2721 if (flush_color >= 0) {
2722 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2723 atomic_set(&wq->nr_pwqs_to_flush, 1);
2724 }
2725
2726 for_each_pwq(pwq, wq) {
2727 struct worker_pool *pool = pwq->pool;
2728
2729 spin_lock_irq(&pool->lock);
2730
2731 if (flush_color >= 0) {
2732 WARN_ON_ONCE(pwq->flush_color != -1);
2733
2734 if (pwq->nr_in_flight[flush_color]) {
2735 pwq->flush_color = flush_color;
2736 atomic_inc(&wq->nr_pwqs_to_flush);
2737 wait = true;
2738 }
2739 }
2740
2741 if (work_color >= 0) {
2742 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2743 pwq->work_color = work_color;
2744 }
2745
2746 spin_unlock_irq(&pool->lock);
2747 }
2748
2749 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2750 complete(&wq->first_flusher->done);
2751
2752 return wait;
2753}
2754
2755/**
2756 * flush_workqueue - ensure that any scheduled work has run to completion.
2757 * @wq: workqueue to flush
2758 *
2759 * This function sleeps until all work items which were queued on entry
2760 * have finished execution, but it is not livelocked by new incoming ones.
2761 */
2762void flush_workqueue(struct workqueue_struct *wq)
2763{
2764 struct wq_flusher this_flusher = {
2765 .list = LIST_HEAD_INIT(this_flusher.list),
2766 .flush_color = -1,
2767 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2768 };
2769 int next_color;
2770
2771 if (WARN_ON(!wq_online))
2772 return;
2773
2774 lock_map_acquire(&wq->lockdep_map);
2775 lock_map_release(&wq->lockdep_map);
2776
2777 mutex_lock(&wq->mutex);
2778
2779 /*
2780 * Start-to-wait phase
2781 */
2782 next_color = work_next_color(wq->work_color);
2783
2784 if (next_color != wq->flush_color) {
2785 /*
2786 * Color space is not full. The current work_color
2787 * becomes our flush_color and work_color is advanced
2788 * by one.
2789 */
2790 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2791 this_flusher.flush_color = wq->work_color;
2792 wq->work_color = next_color;
2793
2794 if (!wq->first_flusher) {
2795 /* no flush in progress, become the first flusher */
2796 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2797
2798 wq->first_flusher = &this_flusher;
2799
2800 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2801 wq->work_color)) {
2802 /* nothing to flush, done */
2803 wq->flush_color = next_color;
2804 wq->first_flusher = NULL;
2805 goto out_unlock;
2806 }
2807 } else {
2808 /* wait in queue */
2809 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2810 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2811 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2812 }
2813 } else {
2814 /*
2815 * Oops, color space is full, wait on overflow queue.
2816 * The next flush completion will assign us
2817 * flush_color and transfer to flusher_queue.
2818 */
2819 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2820 }
2821
2822 check_flush_dependency(wq, NULL);
2823
2824 mutex_unlock(&wq->mutex);
2825
2826 wait_for_completion(&this_flusher.done);
2827
2828 /*
2829 * Wake-up-and-cascade phase
2830 *
2831 * First flushers are responsible for cascading flushes and
2832 * handling overflow. Non-first flushers can simply return.
2833 */
2834 if (wq->first_flusher != &this_flusher)
2835 return;
2836
2837 mutex_lock(&wq->mutex);
2838
2839 /* we might have raced, check again with mutex held */
2840 if (wq->first_flusher != &this_flusher)
2841 goto out_unlock;
2842
2843 wq->first_flusher = NULL;
2844
2845 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2846 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2847
2848 while (true) {
2849 struct wq_flusher *next, *tmp;
2850
2851 /* complete all the flushers sharing the current flush color */
2852 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2853 if (next->flush_color != wq->flush_color)
2854 break;
2855 list_del_init(&next->list);
2856 complete(&next->done);
2857 }
2858
2859 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2860 wq->flush_color != work_next_color(wq->work_color));
2861
2862 /* this flush_color is finished, advance by one */
2863 wq->flush_color = work_next_color(wq->flush_color);
2864
2865 /* one color has been freed, handle overflow queue */
2866 if (!list_empty(&wq->flusher_overflow)) {
2867 /*
2868 * Assign the same color to all overflowed
2869 * flushers, advance work_color and append to
2870 * flusher_queue. This is the start-to-wait
2871 * phase for these overflowed flushers.
2872 */
2873 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2874 tmp->flush_color = wq->work_color;
2875
2876 wq->work_color = work_next_color(wq->work_color);
2877
2878 list_splice_tail_init(&wq->flusher_overflow,
2879 &wq->flusher_queue);
2880 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2881 }
2882
2883 if (list_empty(&wq->flusher_queue)) {
2884 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2885 break;
2886 }
2887
2888 /*
2889 * Need to flush more colors. Make the next flusher
2890 * the new first flusher and arm pwqs.
2891 */
2892 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2893 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2894
2895 list_del_init(&next->list);
2896 wq->first_flusher = next;
2897
2898 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2899 break;
2900
2901 /*
2902 * Meh... this color is already done, clear first
2903 * flusher and repeat cascading.
2904 */
2905 wq->first_flusher = NULL;
2906 }
2907
2908out_unlock:
2909 mutex_unlock(&wq->mutex);
2910}
2911EXPORT_SYMBOL(flush_workqueue);
2912
2913/**
2914 * drain_workqueue - drain a workqueue
2915 * @wq: workqueue to drain
2916 *
2917 * Wait until the workqueue becomes empty. While draining is in progress,
2918 * only chain queueing is allowed. IOW, only currently pending or running
2919 * work items on @wq can queue further work items on it. @wq is flushed
2920 * repeatedly until it becomes empty. The number of flushing is determined
2921 * by the depth of chaining and should be relatively short. Whine if it
2922 * takes too long.
2923 */
2924void drain_workqueue(struct workqueue_struct *wq)
2925{
2926 unsigned int flush_cnt = 0;
2927 struct pool_workqueue *pwq;
2928
2929 /*
2930 * __queue_work() needs to test whether there are drainers, is much
2931 * hotter than drain_workqueue() and already looks at @wq->flags.
2932 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2933 */
2934 mutex_lock(&wq->mutex);
2935 if (!wq->nr_drainers++)
2936 wq->flags |= __WQ_DRAINING;
2937 mutex_unlock(&wq->mutex);
2938reflush:
2939 flush_workqueue(wq);
2940
2941 mutex_lock(&wq->mutex);
2942
2943 for_each_pwq(pwq, wq) {
2944 bool drained;
2945
2946 spin_lock_irq(&pwq->pool->lock);
2947 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2948 spin_unlock_irq(&pwq->pool->lock);
2949
2950 if (drained)
2951 continue;
2952
2953 if (++flush_cnt == 10 ||
2954 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2955 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2956 wq->name, flush_cnt);
2957
2958 mutex_unlock(&wq->mutex);
2959 goto reflush;
2960 }
2961
2962 if (!--wq->nr_drainers)
2963 wq->flags &= ~__WQ_DRAINING;
2964 mutex_unlock(&wq->mutex);
2965}
2966EXPORT_SYMBOL_GPL(drain_workqueue);
2967
2968static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2969 bool from_cancel)
2970{
2971 struct worker *worker = NULL;
2972 struct worker_pool *pool;
2973 struct pool_workqueue *pwq;
2974
2975 might_sleep();
2976
2977 rcu_read_lock();
2978 pool = get_work_pool(work);
2979 if (!pool) {
2980 rcu_read_unlock();
2981 return false;
2982 }
2983
2984 spin_lock_irq(&pool->lock);
2985 /* see the comment in try_to_grab_pending() with the same code */
2986 pwq = get_work_pwq(work);
2987 if (pwq) {
2988 if (unlikely(pwq->pool != pool))
2989 goto already_gone;
2990 } else {
2991 worker = find_worker_executing_work(pool, work);
2992 if (!worker)
2993 goto already_gone;
2994 pwq = worker->current_pwq;
2995 }
2996
2997 check_flush_dependency(pwq->wq, work);
2998
2999 insert_wq_barrier(pwq, barr, work, worker);
3000 spin_unlock_irq(&pool->lock);
3001
3002 /*
3003 * Force a lock recursion deadlock when using flush_work() inside a
3004 * single-threaded or rescuer equipped workqueue.
3005 *
3006 * For single threaded workqueues the deadlock happens when the work
3007 * is after the work issuing the flush_work(). For rescuer equipped
3008 * workqueues the deadlock happens when the rescuer stalls, blocking
3009 * forward progress.
3010 */
3011 if (!from_cancel &&
3012 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3013 lock_map_acquire(&pwq->wq->lockdep_map);
3014 lock_map_release(&pwq->wq->lockdep_map);
3015 }
3016 rcu_read_unlock();
3017 return true;
3018already_gone:
3019 spin_unlock_irq(&pool->lock);
3020 rcu_read_unlock();
3021 return false;
3022}
3023
3024static bool __flush_work(struct work_struct *work, bool from_cancel)
3025{
3026 struct wq_barrier barr;
3027
3028 if (WARN_ON(!wq_online))
3029 return false;
3030
3031 if (WARN_ON(!work->func))
3032 return false;
3033
3034 if (!from_cancel) {
3035 lock_map_acquire(&work->lockdep_map);
3036 lock_map_release(&work->lockdep_map);
3037 }
3038
3039 if (start_flush_work(work, &barr, from_cancel)) {
3040 wait_for_completion(&barr.done);
3041 destroy_work_on_stack(&barr.work);
3042 return true;
3043 } else {
3044 return false;
3045 }
3046}
3047
3048/**
3049 * flush_work - wait for a work to finish executing the last queueing instance
3050 * @work: the work to flush
3051 *
3052 * Wait until @work has finished execution. @work is guaranteed to be idle
3053 * on return if it hasn't been requeued since flush started.
3054 *
3055 * Return:
3056 * %true if flush_work() waited for the work to finish execution,
3057 * %false if it was already idle.
3058 */
3059bool flush_work(struct work_struct *work)
3060{
3061 return __flush_work(work, false);
3062}
3063EXPORT_SYMBOL_GPL(flush_work);
3064
3065struct cwt_wait {
3066 wait_queue_entry_t wait;
3067 struct work_struct *work;
3068};
3069
3070static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3071{
3072 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3073
3074 if (cwait->work != key)
3075 return 0;
3076 return autoremove_wake_function(wait, mode, sync, key);
3077}
3078
3079static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3080{
3081 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3082 unsigned long flags;
3083 int ret;
3084
3085 do {
3086 ret = try_to_grab_pending(work, is_dwork, &flags);
3087 /*
3088 * If someone else is already canceling, wait for it to
3089 * finish. flush_work() doesn't work for PREEMPT_NONE
3090 * because we may get scheduled between @work's completion
3091 * and the other canceling task resuming and clearing
3092 * CANCELING - flush_work() will return false immediately
3093 * as @work is no longer busy, try_to_grab_pending() will
3094 * return -ENOENT as @work is still being canceled and the
3095 * other canceling task won't be able to clear CANCELING as
3096 * we're hogging the CPU.
3097 *
3098 * Let's wait for completion using a waitqueue. As this
3099 * may lead to the thundering herd problem, use a custom
3100 * wake function which matches @work along with exclusive
3101 * wait and wakeup.
3102 */
3103 if (unlikely(ret == -ENOENT)) {
3104 struct cwt_wait cwait;
3105
3106 init_wait(&cwait.wait);
3107 cwait.wait.func = cwt_wakefn;
3108 cwait.work = work;
3109
3110 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3111 TASK_UNINTERRUPTIBLE);
3112 if (work_is_canceling(work))
3113 schedule();
3114 finish_wait(&cancel_waitq, &cwait.wait);
3115 }
3116 } while (unlikely(ret < 0));
3117
3118 /* tell other tasks trying to grab @work to back off */
3119 mark_work_canceling(work);
3120 local_irq_restore(flags);
3121
3122 /*
3123 * This allows canceling during early boot. We know that @work
3124 * isn't executing.
3125 */
3126 if (wq_online)
3127 __flush_work(work, true);
3128
3129 clear_work_data(work);
3130
3131 /*
3132 * Paired with prepare_to_wait() above so that either
3133 * waitqueue_active() is visible here or !work_is_canceling() is
3134 * visible there.
3135 */
3136 smp_mb();
3137 if (waitqueue_active(&cancel_waitq))
3138 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3139
3140 return ret;
3141}
3142
3143/**
3144 * cancel_work_sync - cancel a work and wait for it to finish
3145 * @work: the work to cancel
3146 *
3147 * Cancel @work and wait for its execution to finish. This function
3148 * can be used even if the work re-queues itself or migrates to
3149 * another workqueue. On return from this function, @work is
3150 * guaranteed to be not pending or executing on any CPU.
3151 *
3152 * cancel_work_sync(&delayed_work->work) must not be used for
3153 * delayed_work's. Use cancel_delayed_work_sync() instead.
3154 *
3155 * The caller must ensure that the workqueue on which @work was last
3156 * queued can't be destroyed before this function returns.
3157 *
3158 * Return:
3159 * %true if @work was pending, %false otherwise.
3160 */
3161bool cancel_work_sync(struct work_struct *work)
3162{
3163 return __cancel_work_timer(work, false);
3164}
3165EXPORT_SYMBOL_GPL(cancel_work_sync);
3166
3167/**
3168 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3169 * @dwork: the delayed work to flush
3170 *
3171 * Delayed timer is cancelled and the pending work is queued for
3172 * immediate execution. Like flush_work(), this function only
3173 * considers the last queueing instance of @dwork.
3174 *
3175 * Return:
3176 * %true if flush_work() waited for the work to finish execution,
3177 * %false if it was already idle.
3178 */
3179bool flush_delayed_work(struct delayed_work *dwork)
3180{
3181 local_irq_disable();
3182 if (del_timer_sync(&dwork->timer))
3183 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3184 local_irq_enable();
3185 return flush_work(&dwork->work);
3186}
3187EXPORT_SYMBOL(flush_delayed_work);
3188
3189/**
3190 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3191 * @rwork: the rcu work to flush
3192 *
3193 * Return:
3194 * %true if flush_rcu_work() waited for the work to finish execution,
3195 * %false if it was already idle.
3196 */
3197bool flush_rcu_work(struct rcu_work *rwork)
3198{
3199 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3200 rcu_barrier();
3201 flush_work(&rwork->work);
3202 return true;
3203 } else {
3204 return flush_work(&rwork->work);
3205 }
3206}
3207EXPORT_SYMBOL(flush_rcu_work);
3208
3209static bool __cancel_work(struct work_struct *work, bool is_dwork)
3210{
3211 unsigned long flags;
3212 int ret;
3213
3214 do {
3215 ret = try_to_grab_pending(work, is_dwork, &flags);
3216 } while (unlikely(ret == -EAGAIN));
3217
3218 if (unlikely(ret < 0))
3219 return false;
3220
3221 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3222 local_irq_restore(flags);
3223 return ret;
3224}
3225
3226/**
3227 * cancel_delayed_work - cancel a delayed work
3228 * @dwork: delayed_work to cancel
3229 *
3230 * Kill off a pending delayed_work.
3231 *
3232 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3233 * pending.
3234 *
3235 * Note:
3236 * The work callback function may still be running on return, unless
3237 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3238 * use cancel_delayed_work_sync() to wait on it.
3239 *
3240 * This function is safe to call from any context including IRQ handler.
3241 */
3242bool cancel_delayed_work(struct delayed_work *dwork)
3243{
3244 return __cancel_work(&dwork->work, true);
3245}
3246EXPORT_SYMBOL(cancel_delayed_work);
3247
3248/**
3249 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3250 * @dwork: the delayed work cancel
3251 *
3252 * This is cancel_work_sync() for delayed works.
3253 *
3254 * Return:
3255 * %true if @dwork was pending, %false otherwise.
3256 */
3257bool cancel_delayed_work_sync(struct delayed_work *dwork)
3258{
3259 return __cancel_work_timer(&dwork->work, true);
3260}
3261EXPORT_SYMBOL(cancel_delayed_work_sync);
3262
3263/**
3264 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3265 * @func: the function to call
3266 *
3267 * schedule_on_each_cpu() executes @func on each online CPU using the
3268 * system workqueue and blocks until all CPUs have completed.
3269 * schedule_on_each_cpu() is very slow.
3270 *
3271 * Return:
3272 * 0 on success, -errno on failure.
3273 */
3274int schedule_on_each_cpu(work_func_t func)
3275{
3276 int cpu;
3277 struct work_struct __percpu *works;
3278
3279 works = alloc_percpu(struct work_struct);
3280 if (!works)
3281 return -ENOMEM;
3282
3283 get_online_cpus();
3284
3285 for_each_online_cpu(cpu) {
3286 struct work_struct *work = per_cpu_ptr(works, cpu);
3287
3288 INIT_WORK(work, func);
3289 schedule_work_on(cpu, work);
3290 }
3291
3292 for_each_online_cpu(cpu)
3293 flush_work(per_cpu_ptr(works, cpu));
3294
3295 put_online_cpus();
3296 free_percpu(works);
3297 return 0;
3298}
3299
3300/**
3301 * execute_in_process_context - reliably execute the routine with user context
3302 * @fn: the function to execute
3303 * @ew: guaranteed storage for the execute work structure (must
3304 * be available when the work executes)
3305 *
3306 * Executes the function immediately if process context is available,
3307 * otherwise schedules the function for delayed execution.
3308 *
3309 * Return: 0 - function was executed
3310 * 1 - function was scheduled for execution
3311 */
3312int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3313{
3314 if (!in_interrupt()) {
3315 fn(&ew->work);
3316 return 0;
3317 }
3318
3319 INIT_WORK(&ew->work, fn);
3320 schedule_work(&ew->work);
3321
3322 return 1;
3323}
3324EXPORT_SYMBOL_GPL(execute_in_process_context);
3325
3326/**
3327 * free_workqueue_attrs - free a workqueue_attrs
3328 * @attrs: workqueue_attrs to free
3329 *
3330 * Undo alloc_workqueue_attrs().
3331 */
3332void free_workqueue_attrs(struct workqueue_attrs *attrs)
3333{
3334 if (attrs) {
3335 free_cpumask_var(attrs->cpumask);
3336 kfree(attrs);
3337 }
3338}
3339
3340/**
3341 * alloc_workqueue_attrs - allocate a workqueue_attrs
3342 *
3343 * Allocate a new workqueue_attrs, initialize with default settings and
3344 * return it.
3345 *
3346 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3347 */
3348struct workqueue_attrs *alloc_workqueue_attrs(void)
3349{
3350 struct workqueue_attrs *attrs;
3351
3352 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3353 if (!attrs)
3354 goto fail;
3355 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3356 goto fail;
3357
3358 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3359 return attrs;
3360fail:
3361 free_workqueue_attrs(attrs);
3362 return NULL;
3363}
3364
3365static void copy_workqueue_attrs(struct workqueue_attrs *to,
3366 const struct workqueue_attrs *from)
3367{
3368 to->nice = from->nice;
3369 cpumask_copy(to->cpumask, from->cpumask);
3370 /*
3371 * Unlike hash and equality test, this function doesn't ignore
3372 * ->no_numa as it is used for both pool and wq attrs. Instead,
3373 * get_unbound_pool() explicitly clears ->no_numa after copying.
3374 */
3375 to->no_numa = from->no_numa;
3376}
3377
3378/* hash value of the content of @attr */
3379static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3380{
3381 u32 hash = 0;
3382
3383 hash = jhash_1word(attrs->nice, hash);
3384 hash = jhash(cpumask_bits(attrs->cpumask),
3385 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3386 return hash;
3387}
3388
3389/* content equality test */
3390static bool wqattrs_equal(const struct workqueue_attrs *a,
3391 const struct workqueue_attrs *b)
3392{
3393 if (a->nice != b->nice)
3394 return false;
3395 if (!cpumask_equal(a->cpumask, b->cpumask))
3396 return false;
3397 return true;
3398}
3399
3400/**
3401 * init_worker_pool - initialize a newly zalloc'd worker_pool
3402 * @pool: worker_pool to initialize
3403 *
3404 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3405 *
3406 * Return: 0 on success, -errno on failure. Even on failure, all fields
3407 * inside @pool proper are initialized and put_unbound_pool() can be called
3408 * on @pool safely to release it.
3409 */
3410static int init_worker_pool(struct worker_pool *pool)
3411{
3412 spin_lock_init(&pool->lock);
3413 pool->id = -1;
3414 pool->cpu = -1;
3415 pool->node = NUMA_NO_NODE;
3416 pool->flags |= POOL_DISASSOCIATED;
3417 pool->watchdog_ts = jiffies;
3418 INIT_LIST_HEAD(&pool->worklist);
3419 INIT_LIST_HEAD(&pool->idle_list);
3420 hash_init(pool->busy_hash);
3421
3422 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3423
3424 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3425
3426 INIT_LIST_HEAD(&pool->workers);
3427
3428 ida_init(&pool->worker_ida);
3429 INIT_HLIST_NODE(&pool->hash_node);
3430 pool->refcnt = 1;
3431
3432 /* shouldn't fail above this point */
3433 pool->attrs = alloc_workqueue_attrs();
3434 if (!pool->attrs)
3435 return -ENOMEM;
3436 return 0;
3437}
3438
3439#ifdef CONFIG_LOCKDEP
3440static void wq_init_lockdep(struct workqueue_struct *wq)
3441{
3442 char *lock_name;
3443
3444 lockdep_register_key(&wq->key);
3445 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3446 if (!lock_name)
3447 lock_name = wq->name;
3448
3449 wq->lock_name = lock_name;
3450 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3451}
3452
3453static void wq_unregister_lockdep(struct workqueue_struct *wq)
3454{
3455 lockdep_unregister_key(&wq->key);
3456}
3457
3458static void wq_free_lockdep(struct workqueue_struct *wq)
3459{
3460 if (wq->lock_name != wq->name)
3461 kfree(wq->lock_name);
3462}
3463#else
3464static void wq_init_lockdep(struct workqueue_struct *wq)
3465{
3466}
3467
3468static void wq_unregister_lockdep(struct workqueue_struct *wq)
3469{
3470}
3471
3472static void wq_free_lockdep(struct workqueue_struct *wq)
3473{
3474}
3475#endif
3476
3477static void rcu_free_wq(struct rcu_head *rcu)
3478{
3479 struct workqueue_struct *wq =
3480 container_of(rcu, struct workqueue_struct, rcu);
3481
3482 wq_free_lockdep(wq);
3483
3484 if (!(wq->flags & WQ_UNBOUND))
3485 free_percpu(wq->cpu_pwqs);
3486 else
3487 free_workqueue_attrs(wq->unbound_attrs);
3488
3489 kfree(wq->rescuer);
3490 kfree(wq);
3491}
3492
3493static void rcu_free_pool(struct rcu_head *rcu)
3494{
3495 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3496
3497 ida_destroy(&pool->worker_ida);
3498 free_workqueue_attrs(pool->attrs);
3499 kfree(pool);
3500}
3501
3502/**
3503 * put_unbound_pool - put a worker_pool
3504 * @pool: worker_pool to put
3505 *
3506 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3507 * safe manner. get_unbound_pool() calls this function on its failure path
3508 * and this function should be able to release pools which went through,
3509 * successfully or not, init_worker_pool().
3510 *
3511 * Should be called with wq_pool_mutex held.
3512 */
3513static void put_unbound_pool(struct worker_pool *pool)
3514{
3515 DECLARE_COMPLETION_ONSTACK(detach_completion);
3516 struct worker *worker;
3517
3518 lockdep_assert_held(&wq_pool_mutex);
3519
3520 if (--pool->refcnt)
3521 return;
3522
3523 /* sanity checks */
3524 if (WARN_ON(!(pool->cpu < 0)) ||
3525 WARN_ON(!list_empty(&pool->worklist)))
3526 return;
3527
3528 /* release id and unhash */
3529 if (pool->id >= 0)
3530 idr_remove(&worker_pool_idr, pool->id);
3531 hash_del(&pool->hash_node);
3532
3533 /*
3534 * Become the manager and destroy all workers. This prevents
3535 * @pool's workers from blocking on attach_mutex. We're the last
3536 * manager and @pool gets freed with the flag set.
3537 */
3538 spin_lock_irq(&pool->lock);
3539 wait_event_lock_irq(wq_manager_wait,
3540 !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3541 pool->flags |= POOL_MANAGER_ACTIVE;
3542
3543 while ((worker = first_idle_worker(pool)))
3544 destroy_worker(worker);
3545 WARN_ON(pool->nr_workers || pool->nr_idle);
3546 spin_unlock_irq(&pool->lock);
3547
3548 mutex_lock(&wq_pool_attach_mutex);
3549 if (!list_empty(&pool->workers))
3550 pool->detach_completion = &detach_completion;
3551 mutex_unlock(&wq_pool_attach_mutex);
3552
3553 if (pool->detach_completion)
3554 wait_for_completion(pool->detach_completion);
3555
3556 /* shut down the timers */
3557 del_timer_sync(&pool->idle_timer);
3558 del_timer_sync(&pool->mayday_timer);
3559
3560 /* RCU protected to allow dereferences from get_work_pool() */
3561 call_rcu(&pool->rcu, rcu_free_pool);
3562}
3563
3564/**
3565 * get_unbound_pool - get a worker_pool with the specified attributes
3566 * @attrs: the attributes of the worker_pool to get
3567 *
3568 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3569 * reference count and return it. If there already is a matching
3570 * worker_pool, it will be used; otherwise, this function attempts to
3571 * create a new one.
3572 *
3573 * Should be called with wq_pool_mutex held.
3574 *
3575 * Return: On success, a worker_pool with the same attributes as @attrs.
3576 * On failure, %NULL.
3577 */
3578static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3579{
3580 u32 hash = wqattrs_hash(attrs);
3581 struct worker_pool *pool;
3582 int node;
3583 int target_node = NUMA_NO_NODE;
3584
3585 lockdep_assert_held(&wq_pool_mutex);
3586
3587 /* do we already have a matching pool? */
3588 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3589 if (wqattrs_equal(pool->attrs, attrs)) {
3590 pool->refcnt++;
3591 return pool;
3592 }
3593 }
3594
3595 /* if cpumask is contained inside a NUMA node, we belong to that node */
3596 if (wq_numa_enabled) {
3597 for_each_node(node) {
3598 if (cpumask_subset(attrs->cpumask,
3599 wq_numa_possible_cpumask[node])) {
3600 target_node = node;
3601 break;
3602 }
3603 }
3604 }
3605
3606 /* nope, create a new one */
3607 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3608 if (!pool || init_worker_pool(pool) < 0)
3609 goto fail;
3610
3611 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3612 copy_workqueue_attrs(pool->attrs, attrs);
3613 pool->node = target_node;
3614
3615 /*
3616 * no_numa isn't a worker_pool attribute, always clear it. See
3617 * 'struct workqueue_attrs' comments for detail.
3618 */
3619 pool->attrs->no_numa = false;
3620
3621 if (worker_pool_assign_id(pool) < 0)
3622 goto fail;
3623
3624 /* create and start the initial worker */
3625 if (wq_online && !create_worker(pool))
3626 goto fail;
3627
3628 /* install */
3629 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3630
3631 return pool;
3632fail:
3633 if (pool)
3634 put_unbound_pool(pool);
3635 return NULL;
3636}
3637
3638static void rcu_free_pwq(struct rcu_head *rcu)
3639{
3640 kmem_cache_free(pwq_cache,
3641 container_of(rcu, struct pool_workqueue, rcu));
3642}
3643
3644/*
3645 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3646 * and needs to be destroyed.
3647 */
3648static void pwq_unbound_release_workfn(struct work_struct *work)
3649{
3650 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3651 unbound_release_work);
3652 struct workqueue_struct *wq = pwq->wq;
3653 struct worker_pool *pool = pwq->pool;
3654 bool is_last;
3655
3656 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3657 return;
3658
3659 mutex_lock(&wq->mutex);
3660 list_del_rcu(&pwq->pwqs_node);
3661 is_last = list_empty(&wq->pwqs);
3662 mutex_unlock(&wq->mutex);
3663
3664 mutex_lock(&wq_pool_mutex);
3665 put_unbound_pool(pool);
3666 mutex_unlock(&wq_pool_mutex);
3667
3668 call_rcu(&pwq->rcu, rcu_free_pwq);
3669
3670 /*
3671 * If we're the last pwq going away, @wq is already dead and no one
3672 * is gonna access it anymore. Schedule RCU free.
3673 */
3674 if (is_last) {
3675 wq_unregister_lockdep(wq);
3676 call_rcu(&wq->rcu, rcu_free_wq);
3677 }
3678}
3679
3680/**
3681 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3682 * @pwq: target pool_workqueue
3683 *
3684 * If @pwq isn't freezing, set @pwq->max_active to the associated
3685 * workqueue's saved_max_active and activate delayed work items
3686 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3687 */
3688static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3689{
3690 struct workqueue_struct *wq = pwq->wq;
3691 bool freezable = wq->flags & WQ_FREEZABLE;
3692 unsigned long flags;
3693
3694 /* for @wq->saved_max_active */
3695 lockdep_assert_held(&wq->mutex);
3696
3697 /* fast exit for non-freezable wqs */
3698 if (!freezable && pwq->max_active == wq->saved_max_active)
3699 return;
3700
3701 /* this function can be called during early boot w/ irq disabled */
3702 spin_lock_irqsave(&pwq->pool->lock, flags);
3703
3704 /*
3705 * During [un]freezing, the caller is responsible for ensuring that
3706 * this function is called at least once after @workqueue_freezing
3707 * is updated and visible.
3708 */
3709 if (!freezable || !workqueue_freezing) {
3710 pwq->max_active = wq->saved_max_active;
3711
3712 while (!list_empty(&pwq->delayed_works) &&
3713 pwq->nr_active < pwq->max_active)
3714 pwq_activate_first_delayed(pwq);
3715
3716 /*
3717 * Need to kick a worker after thawed or an unbound wq's
3718 * max_active is bumped. It's a slow path. Do it always.
3719 */
3720 wake_up_worker(pwq->pool);
3721 } else {
3722 pwq->max_active = 0;
3723 }
3724
3725 spin_unlock_irqrestore(&pwq->pool->lock, flags);
3726}
3727
3728/* initialize newly alloced @pwq which is associated with @wq and @pool */
3729static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3730 struct worker_pool *pool)
3731{
3732 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3733
3734 memset(pwq, 0, sizeof(*pwq));
3735
3736 pwq->pool = pool;
3737 pwq->wq = wq;
3738 pwq->flush_color = -1;
3739 pwq->refcnt = 1;
3740 INIT_LIST_HEAD(&pwq->delayed_works);
3741 INIT_LIST_HEAD(&pwq->pwqs_node);
3742 INIT_LIST_HEAD(&pwq->mayday_node);
3743 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3744}
3745
3746/* sync @pwq with the current state of its associated wq and link it */
3747static void link_pwq(struct pool_workqueue *pwq)
3748{
3749 struct workqueue_struct *wq = pwq->wq;
3750
3751 lockdep_assert_held(&wq->mutex);
3752
3753 /* may be called multiple times, ignore if already linked */
3754 if (!list_empty(&pwq->pwqs_node))
3755 return;
3756
3757 /* set the matching work_color */
3758 pwq->work_color = wq->work_color;
3759
3760 /* sync max_active to the current setting */
3761 pwq_adjust_max_active(pwq);
3762
3763 /* link in @pwq */
3764 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3765}
3766
3767/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3768static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3769 const struct workqueue_attrs *attrs)
3770{
3771 struct worker_pool *pool;
3772 struct pool_workqueue *pwq;
3773
3774 lockdep_assert_held(&wq_pool_mutex);
3775
3776 pool = get_unbound_pool(attrs);
3777 if (!pool)
3778 return NULL;
3779
3780 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3781 if (!pwq) {
3782 put_unbound_pool(pool);
3783 return NULL;
3784 }
3785
3786 init_pwq(pwq, wq, pool);
3787 return pwq;
3788}
3789
3790/**
3791 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3792 * @attrs: the wq_attrs of the default pwq of the target workqueue
3793 * @node: the target NUMA node
3794 * @cpu_going_down: if >= 0, the CPU to consider as offline
3795 * @cpumask: outarg, the resulting cpumask
3796 *
3797 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3798 * @cpu_going_down is >= 0, that cpu is considered offline during
3799 * calculation. The result is stored in @cpumask.
3800 *
3801 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3802 * enabled and @node has online CPUs requested by @attrs, the returned
3803 * cpumask is the intersection of the possible CPUs of @node and
3804 * @attrs->cpumask.
3805 *
3806 * The caller is responsible for ensuring that the cpumask of @node stays
3807 * stable.
3808 *
3809 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3810 * %false if equal.
3811 */
3812static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3813 int cpu_going_down, cpumask_t *cpumask)
3814{
3815 if (!wq_numa_enabled || attrs->no_numa)
3816 goto use_dfl;
3817
3818 /* does @node have any online CPUs @attrs wants? */
3819 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3820 if (cpu_going_down >= 0)
3821 cpumask_clear_cpu(cpu_going_down, cpumask);
3822
3823 if (cpumask_empty(cpumask))
3824 goto use_dfl;
3825
3826 /* yeap, return possible CPUs in @node that @attrs wants */
3827 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3828
3829 if (cpumask_empty(cpumask)) {
3830 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3831 "possible intersect\n");
3832 return false;
3833 }
3834
3835 return !cpumask_equal(cpumask, attrs->cpumask);
3836
3837use_dfl:
3838 cpumask_copy(cpumask, attrs->cpumask);
3839 return false;
3840}
3841
3842/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3843static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3844 int node,
3845 struct pool_workqueue *pwq)
3846{
3847 struct pool_workqueue *old_pwq;
3848
3849 lockdep_assert_held(&wq_pool_mutex);
3850 lockdep_assert_held(&wq->mutex);
3851
3852 /* link_pwq() can handle duplicate calls */
3853 link_pwq(pwq);
3854
3855 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3856 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3857 return old_pwq;
3858}
3859
3860/* context to store the prepared attrs & pwqs before applying */
3861struct apply_wqattrs_ctx {
3862 struct workqueue_struct *wq; /* target workqueue */
3863 struct workqueue_attrs *attrs; /* attrs to apply */
3864 struct list_head list; /* queued for batching commit */
3865 struct pool_workqueue *dfl_pwq;
3866 struct pool_workqueue *pwq_tbl[];
3867};
3868
3869/* free the resources after success or abort */
3870static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3871{
3872 if (ctx) {
3873 int node;
3874
3875 for_each_node(node)
3876 put_pwq_unlocked(ctx->pwq_tbl[node]);
3877 put_pwq_unlocked(ctx->dfl_pwq);
3878
3879 free_workqueue_attrs(ctx->attrs);
3880
3881 kfree(ctx);
3882 }
3883}
3884
3885/* allocate the attrs and pwqs for later installation */
3886static struct apply_wqattrs_ctx *
3887apply_wqattrs_prepare(struct workqueue_struct *wq,
3888 const struct workqueue_attrs *attrs)
3889{
3890 struct apply_wqattrs_ctx *ctx;
3891 struct workqueue_attrs *new_attrs, *tmp_attrs;
3892 int node;
3893
3894 lockdep_assert_held(&wq_pool_mutex);
3895
3896 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3897
3898 new_attrs = alloc_workqueue_attrs();
3899 tmp_attrs = alloc_workqueue_attrs();
3900 if (!ctx || !new_attrs || !tmp_attrs)
3901 goto out_free;
3902
3903 /*
3904 * Calculate the attrs of the default pwq.
3905 * If the user configured cpumask doesn't overlap with the
3906 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3907 */
3908 copy_workqueue_attrs(new_attrs, attrs);
3909 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3910 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3911 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3912
3913 /*
3914 * We may create multiple pwqs with differing cpumasks. Make a
3915 * copy of @new_attrs which will be modified and used to obtain
3916 * pools.
3917 */
3918 copy_workqueue_attrs(tmp_attrs, new_attrs);
3919
3920 /*
3921 * If something goes wrong during CPU up/down, we'll fall back to
3922 * the default pwq covering whole @attrs->cpumask. Always create
3923 * it even if we don't use it immediately.
3924 */
3925 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3926 if (!ctx->dfl_pwq)
3927 goto out_free;
3928
3929 for_each_node(node) {
3930 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3931 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3932 if (!ctx->pwq_tbl[node])
3933 goto out_free;
3934 } else {
3935 ctx->dfl_pwq->refcnt++;
3936 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3937 }
3938 }
3939
3940 /* save the user configured attrs and sanitize it. */
3941 copy_workqueue_attrs(new_attrs, attrs);
3942 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3943 ctx->attrs = new_attrs;
3944
3945 ctx->wq = wq;
3946 free_workqueue_attrs(tmp_attrs);
3947 return ctx;
3948
3949out_free:
3950 free_workqueue_attrs(tmp_attrs);
3951 free_workqueue_attrs(new_attrs);
3952 apply_wqattrs_cleanup(ctx);
3953 return NULL;
3954}
3955
3956/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3957static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3958{
3959 int node;
3960
3961 /* all pwqs have been created successfully, let's install'em */
3962 mutex_lock(&ctx->wq->mutex);
3963
3964 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3965
3966 /* save the previous pwq and install the new one */
3967 for_each_node(node)
3968 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3969 ctx->pwq_tbl[node]);
3970
3971 /* @dfl_pwq might not have been used, ensure it's linked */
3972 link_pwq(ctx->dfl_pwq);
3973 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3974
3975 mutex_unlock(&ctx->wq->mutex);
3976}
3977
3978static void apply_wqattrs_lock(void)
3979{
3980 /* CPUs should stay stable across pwq creations and installations */
3981 get_online_cpus();
3982 mutex_lock(&wq_pool_mutex);
3983}
3984
3985static void apply_wqattrs_unlock(void)
3986{
3987 mutex_unlock(&wq_pool_mutex);
3988 put_online_cpus();
3989}
3990
3991static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3992 const struct workqueue_attrs *attrs)
3993{
3994 struct apply_wqattrs_ctx *ctx;
3995
3996 /* only unbound workqueues can change attributes */
3997 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3998 return -EINVAL;
3999
4000 /* creating multiple pwqs breaks ordering guarantee */
4001 if (!list_empty(&wq->pwqs)) {
4002 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4003 return -EINVAL;
4004
4005 wq->flags &= ~__WQ_ORDERED;
4006 }
4007
4008 ctx = apply_wqattrs_prepare(wq, attrs);
4009 if (!ctx)
4010 return -ENOMEM;
4011
4012 /* the ctx has been prepared successfully, let's commit it */
4013 apply_wqattrs_commit(ctx);
4014 apply_wqattrs_cleanup(ctx);
4015
4016 return 0;
4017}
4018
4019/**
4020 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4021 * @wq: the target workqueue
4022 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4023 *
4024 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4025 * machines, this function maps a separate pwq to each NUMA node with
4026 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4027 * NUMA node it was issued on. Older pwqs are released as in-flight work
4028 * items finish. Note that a work item which repeatedly requeues itself
4029 * back-to-back will stay on its current pwq.
4030 *
4031 * Performs GFP_KERNEL allocations.
4032 *
4033 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4034 *
4035 * Return: 0 on success and -errno on failure.
4036 */
4037int apply_workqueue_attrs(struct workqueue_struct *wq,
4038 const struct workqueue_attrs *attrs)
4039{
4040 int ret;
4041
4042 lockdep_assert_cpus_held();
4043
4044 mutex_lock(&wq_pool_mutex);
4045 ret = apply_workqueue_attrs_locked(wq, attrs);
4046 mutex_unlock(&wq_pool_mutex);
4047
4048 return ret;
4049}
4050
4051/**
4052 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4053 * @wq: the target workqueue
4054 * @cpu: the CPU coming up or going down
4055 * @online: whether @cpu is coming up or going down
4056 *
4057 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4058 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4059 * @wq accordingly.
4060 *
4061 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4062 * falls back to @wq->dfl_pwq which may not be optimal but is always
4063 * correct.
4064 *
4065 * Note that when the last allowed CPU of a NUMA node goes offline for a
4066 * workqueue with a cpumask spanning multiple nodes, the workers which were
4067 * already executing the work items for the workqueue will lose their CPU
4068 * affinity and may execute on any CPU. This is similar to how per-cpu
4069 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4070 * affinity, it's the user's responsibility to flush the work item from
4071 * CPU_DOWN_PREPARE.
4072 */
4073static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4074 bool online)
4075{
4076 int node = cpu_to_node(cpu);
4077 int cpu_off = online ? -1 : cpu;
4078 struct pool_workqueue *old_pwq = NULL, *pwq;
4079 struct workqueue_attrs *target_attrs;
4080 cpumask_t *cpumask;
4081
4082 lockdep_assert_held(&wq_pool_mutex);
4083
4084 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4085 wq->unbound_attrs->no_numa)
4086 return;
4087
4088 /*
4089 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4090 * Let's use a preallocated one. The following buf is protected by
4091 * CPU hotplug exclusion.
4092 */
4093 target_attrs = wq_update_unbound_numa_attrs_buf;
4094 cpumask = target_attrs->cpumask;
4095
4096 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4097 pwq = unbound_pwq_by_node(wq, node);
4098
4099 /*
4100 * Let's determine what needs to be done. If the target cpumask is
4101 * different from the default pwq's, we need to compare it to @pwq's
4102 * and create a new one if they don't match. If the target cpumask
4103 * equals the default pwq's, the default pwq should be used.
4104 */
4105 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4106 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4107 return;
4108 } else {
4109 goto use_dfl_pwq;
4110 }
4111
4112 /* create a new pwq */
4113 pwq = alloc_unbound_pwq(wq, target_attrs);
4114 if (!pwq) {
4115 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4116 wq->name);
4117 goto use_dfl_pwq;
4118 }
4119
4120 /* Install the new pwq. */
4121 mutex_lock(&wq->mutex);
4122 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4123 goto out_unlock;
4124
4125use_dfl_pwq:
4126 mutex_lock(&wq->mutex);
4127 spin_lock_irq(&wq->dfl_pwq->pool->lock);
4128 get_pwq(wq->dfl_pwq);
4129 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4130 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4131out_unlock:
4132 mutex_unlock(&wq->mutex);
4133 put_pwq_unlocked(old_pwq);
4134}
4135
4136static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4137{
4138 bool highpri = wq->flags & WQ_HIGHPRI;
4139 int cpu, ret;
4140
4141 if (!(wq->flags & WQ_UNBOUND)) {
4142 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4143 if (!wq->cpu_pwqs)
4144 return -ENOMEM;
4145
4146 for_each_possible_cpu(cpu) {
4147 struct pool_workqueue *pwq =
4148 per_cpu_ptr(wq->cpu_pwqs, cpu);
4149 struct worker_pool *cpu_pools =
4150 per_cpu(cpu_worker_pools, cpu);
4151
4152 init_pwq(pwq, wq, &cpu_pools[highpri]);
4153
4154 mutex_lock(&wq->mutex);
4155 link_pwq(pwq);
4156 mutex_unlock(&wq->mutex);
4157 }
4158 return 0;
4159 }
4160
4161 get_online_cpus();
4162 if (wq->flags & __WQ_ORDERED) {
4163 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4164 /* there should only be single pwq for ordering guarantee */
4165 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4166 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4167 "ordering guarantee broken for workqueue %s\n", wq->name);
4168 } else {
4169 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4170 }
4171 put_online_cpus();
4172
4173 return ret;
4174}
4175
4176static int wq_clamp_max_active(int max_active, unsigned int flags,
4177 const char *name)
4178{
4179 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4180
4181 if (max_active < 1 || max_active > lim)
4182 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4183 max_active, name, 1, lim);
4184
4185 return clamp_val(max_active, 1, lim);
4186}
4187
4188/*
4189 * Workqueues which may be used during memory reclaim should have a rescuer
4190 * to guarantee forward progress.
4191 */
4192static int init_rescuer(struct workqueue_struct *wq)
4193{
4194 struct worker *rescuer;
4195 int ret;
4196
4197 if (!(wq->flags & WQ_MEM_RECLAIM))
4198 return 0;
4199
4200 rescuer = alloc_worker(NUMA_NO_NODE);
4201 if (!rescuer)
4202 return -ENOMEM;
4203
4204 rescuer->rescue_wq = wq;
4205 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4206 ret = PTR_ERR_OR_ZERO(rescuer->task);
4207 if (ret) {
4208 kfree(rescuer);
4209 return ret;
4210 }
4211
4212 wq->rescuer = rescuer;
4213 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4214 wake_up_process(rescuer->task);
4215
4216 return 0;
4217}
4218
4219__printf(1, 4)
4220struct workqueue_struct *alloc_workqueue(const char *fmt,
4221 unsigned int flags,
4222 int max_active, ...)
4223{
4224 size_t tbl_size = 0;
4225 va_list args;
4226 struct workqueue_struct *wq;
4227 struct pool_workqueue *pwq;
4228
4229 /*
4230 * Unbound && max_active == 1 used to imply ordered, which is no
4231 * longer the case on NUMA machines due to per-node pools. While
4232 * alloc_ordered_workqueue() is the right way to create an ordered
4233 * workqueue, keep the previous behavior to avoid subtle breakages
4234 * on NUMA.
4235 */
4236 if ((flags & WQ_UNBOUND) && max_active == 1)
4237 flags |= __WQ_ORDERED;
4238
4239 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4240 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4241 flags |= WQ_UNBOUND;
4242
4243 /* allocate wq and format name */
4244 if (flags & WQ_UNBOUND)
4245 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4246
4247 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4248 if (!wq)
4249 return NULL;
4250
4251 if (flags & WQ_UNBOUND) {
4252 wq->unbound_attrs = alloc_workqueue_attrs();
4253 if (!wq->unbound_attrs)
4254 goto err_free_wq;
4255 }
4256
4257 va_start(args, max_active);
4258 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4259 va_end(args);
4260
4261 max_active = max_active ?: WQ_DFL_ACTIVE;
4262 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4263
4264 /* init wq */
4265 wq->flags = flags;
4266 wq->saved_max_active = max_active;
4267 mutex_init(&wq->mutex);
4268 atomic_set(&wq->nr_pwqs_to_flush, 0);
4269 INIT_LIST_HEAD(&wq->pwqs);
4270 INIT_LIST_HEAD(&wq->flusher_queue);
4271 INIT_LIST_HEAD(&wq->flusher_overflow);
4272 INIT_LIST_HEAD(&wq->maydays);
4273
4274 wq_init_lockdep(wq);
4275 INIT_LIST_HEAD(&wq->list);
4276
4277 if (alloc_and_link_pwqs(wq) < 0)
4278 goto err_unreg_lockdep;
4279
4280 if (wq_online && init_rescuer(wq) < 0)
4281 goto err_destroy;
4282
4283 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4284 goto err_destroy;
4285
4286 /*
4287 * wq_pool_mutex protects global freeze state and workqueues list.
4288 * Grab it, adjust max_active and add the new @wq to workqueues
4289 * list.
4290 */
4291 mutex_lock(&wq_pool_mutex);
4292
4293 mutex_lock(&wq->mutex);
4294 for_each_pwq(pwq, wq)
4295 pwq_adjust_max_active(pwq);
4296 mutex_unlock(&wq->mutex);
4297
4298 list_add_tail_rcu(&wq->list, &workqueues);
4299
4300 mutex_unlock(&wq_pool_mutex);
4301
4302 return wq;
4303
4304err_unreg_lockdep:
4305 wq_unregister_lockdep(wq);
4306 wq_free_lockdep(wq);
4307err_free_wq:
4308 free_workqueue_attrs(wq->unbound_attrs);
4309 kfree(wq);
4310 return NULL;
4311err_destroy:
4312 destroy_workqueue(wq);
4313 return NULL;
4314}
4315EXPORT_SYMBOL_GPL(alloc_workqueue);
4316
4317/**
4318 * destroy_workqueue - safely terminate a workqueue
4319 * @wq: target workqueue
4320 *
4321 * Safely destroy a workqueue. All work currently pending will be done first.
4322 */
4323void destroy_workqueue(struct workqueue_struct *wq)
4324{
4325 struct pool_workqueue *pwq;
4326 int node;
4327
4328 /* drain it before proceeding with destruction */
4329 drain_workqueue(wq);
4330
4331 /* sanity checks */
4332 mutex_lock(&wq->mutex);
4333 for_each_pwq(pwq, wq) {
4334 int i;
4335
4336 for (i = 0; i < WORK_NR_COLORS; i++) {
4337 if (WARN_ON(pwq->nr_in_flight[i])) {
4338 mutex_unlock(&wq->mutex);
4339 show_workqueue_state();
4340 return;
4341 }
4342 }
4343
4344 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4345 WARN_ON(pwq->nr_active) ||
4346 WARN_ON(!list_empty(&pwq->delayed_works))) {
4347 mutex_unlock(&wq->mutex);
4348 show_workqueue_state();
4349 return;
4350 }
4351 }
4352 mutex_unlock(&wq->mutex);
4353
4354 /*
4355 * wq list is used to freeze wq, remove from list after
4356 * flushing is complete in case freeze races us.
4357 */
4358 mutex_lock(&wq_pool_mutex);
4359 list_del_rcu(&wq->list);
4360 mutex_unlock(&wq_pool_mutex);
4361
4362 workqueue_sysfs_unregister(wq);
4363
4364 if (wq->rescuer)
4365 kthread_stop(wq->rescuer->task);
4366
4367 if (!(wq->flags & WQ_UNBOUND)) {
4368 wq_unregister_lockdep(wq);
4369 /*
4370 * The base ref is never dropped on per-cpu pwqs. Directly
4371 * schedule RCU free.
4372 */
4373 call_rcu(&wq->rcu, rcu_free_wq);
4374 } else {
4375 /*
4376 * We're the sole accessor of @wq at this point. Directly
4377 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4378 * @wq will be freed when the last pwq is released.
4379 */
4380 for_each_node(node) {
4381 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4382 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4383 put_pwq_unlocked(pwq);
4384 }
4385
4386 /*
4387 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4388 * put. Don't access it afterwards.
4389 */
4390 pwq = wq->dfl_pwq;
4391 wq->dfl_pwq = NULL;
4392 put_pwq_unlocked(pwq);
4393 }
4394}
4395EXPORT_SYMBOL_GPL(destroy_workqueue);
4396
4397/**
4398 * workqueue_set_max_active - adjust max_active of a workqueue
4399 * @wq: target workqueue
4400 * @max_active: new max_active value.
4401 *
4402 * Set max_active of @wq to @max_active.
4403 *
4404 * CONTEXT:
4405 * Don't call from IRQ context.
4406 */
4407void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4408{
4409 struct pool_workqueue *pwq;
4410
4411 /* disallow meddling with max_active for ordered workqueues */
4412 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4413 return;
4414
4415 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4416
4417 mutex_lock(&wq->mutex);
4418
4419 wq->flags &= ~__WQ_ORDERED;
4420 wq->saved_max_active = max_active;
4421
4422 for_each_pwq(pwq, wq)
4423 pwq_adjust_max_active(pwq);
4424
4425 mutex_unlock(&wq->mutex);
4426}
4427EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4428
4429/**
4430 * current_work - retrieve %current task's work struct
4431 *
4432 * Determine if %current task is a workqueue worker and what it's working on.
4433 * Useful to find out the context that the %current task is running in.
4434 *
4435 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4436 */
4437struct work_struct *current_work(void)
4438{
4439 struct worker *worker = current_wq_worker();
4440
4441 return worker ? worker->current_work : NULL;
4442}
4443EXPORT_SYMBOL(current_work);
4444
4445/**
4446 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4447 *
4448 * Determine whether %current is a workqueue rescuer. Can be used from
4449 * work functions to determine whether it's being run off the rescuer task.
4450 *
4451 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4452 */
4453bool current_is_workqueue_rescuer(void)
4454{
4455 struct worker *worker = current_wq_worker();
4456
4457 return worker && worker->rescue_wq;
4458}
4459
4460/**
4461 * workqueue_congested - test whether a workqueue is congested
4462 * @cpu: CPU in question
4463 * @wq: target workqueue
4464 *
4465 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4466 * no synchronization around this function and the test result is
4467 * unreliable and only useful as advisory hints or for debugging.
4468 *
4469 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4470 * Note that both per-cpu and unbound workqueues may be associated with
4471 * multiple pool_workqueues which have separate congested states. A
4472 * workqueue being congested on one CPU doesn't mean the workqueue is also
4473 * contested on other CPUs / NUMA nodes.
4474 *
4475 * Return:
4476 * %true if congested, %false otherwise.
4477 */
4478bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4479{
4480 struct pool_workqueue *pwq;
4481 bool ret;
4482
4483 rcu_read_lock();
4484 preempt_disable();
4485
4486 if (cpu == WORK_CPU_UNBOUND)
4487 cpu = smp_processor_id();
4488
4489 if (!(wq->flags & WQ_UNBOUND))
4490 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4491 else
4492 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4493
4494 ret = !list_empty(&pwq->delayed_works);
4495 preempt_enable();
4496 rcu_read_unlock();
4497
4498 return ret;
4499}
4500EXPORT_SYMBOL_GPL(workqueue_congested);
4501
4502/**
4503 * work_busy - test whether a work is currently pending or running
4504 * @work: the work to be tested
4505 *
4506 * Test whether @work is currently pending or running. There is no
4507 * synchronization around this function and the test result is
4508 * unreliable and only useful as advisory hints or for debugging.
4509 *
4510 * Return:
4511 * OR'd bitmask of WORK_BUSY_* bits.
4512 */
4513unsigned int work_busy(struct work_struct *work)
4514{
4515 struct worker_pool *pool;
4516 unsigned long flags;
4517 unsigned int ret = 0;
4518
4519 if (work_pending(work))
4520 ret |= WORK_BUSY_PENDING;
4521
4522 rcu_read_lock();
4523 pool = get_work_pool(work);
4524 if (pool) {
4525 spin_lock_irqsave(&pool->lock, flags);
4526 if (find_worker_executing_work(pool, work))
4527 ret |= WORK_BUSY_RUNNING;
4528 spin_unlock_irqrestore(&pool->lock, flags);
4529 }
4530 rcu_read_unlock();
4531
4532 return ret;
4533}
4534EXPORT_SYMBOL_GPL(work_busy);
4535
4536/**
4537 * set_worker_desc - set description for the current work item
4538 * @fmt: printf-style format string
4539 * @...: arguments for the format string
4540 *
4541 * This function can be called by a running work function to describe what
4542 * the work item is about. If the worker task gets dumped, this
4543 * information will be printed out together to help debugging. The
4544 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4545 */
4546void set_worker_desc(const char *fmt, ...)
4547{
4548 struct worker *worker = current_wq_worker();
4549 va_list args;
4550
4551 if (worker) {
4552 va_start(args, fmt);
4553 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4554 va_end(args);
4555 }
4556}
4557EXPORT_SYMBOL_GPL(set_worker_desc);
4558
4559/**
4560 * print_worker_info - print out worker information and description
4561 * @log_lvl: the log level to use when printing
4562 * @task: target task
4563 *
4564 * If @task is a worker and currently executing a work item, print out the
4565 * name of the workqueue being serviced and worker description set with
4566 * set_worker_desc() by the currently executing work item.
4567 *
4568 * This function can be safely called on any task as long as the
4569 * task_struct itself is accessible. While safe, this function isn't
4570 * synchronized and may print out mixups or garbages of limited length.
4571 */
4572void print_worker_info(const char *log_lvl, struct task_struct *task)
4573{
4574 work_func_t *fn = NULL;
4575 char name[WQ_NAME_LEN] = { };
4576 char desc[WORKER_DESC_LEN] = { };
4577 struct pool_workqueue *pwq = NULL;
4578 struct workqueue_struct *wq = NULL;
4579 struct worker *worker;
4580
4581 if (!(task->flags & PF_WQ_WORKER))
4582 return;
4583
4584 /*
4585 * This function is called without any synchronization and @task
4586 * could be in any state. Be careful with dereferences.
4587 */
4588 worker = kthread_probe_data(task);
4589
4590 /*
4591 * Carefully copy the associated workqueue's workfn, name and desc.
4592 * Keep the original last '\0' in case the original is garbage.
4593 */
4594 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4595 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4596 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4597 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4598 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4599
4600 if (fn || name[0] || desc[0]) {
4601 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4602 if (strcmp(name, desc))
4603 pr_cont(" (%s)", desc);
4604 pr_cont("\n");
4605 }
4606}
4607
4608static void pr_cont_pool_info(struct worker_pool *pool)
4609{
4610 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4611 if (pool->node != NUMA_NO_NODE)
4612 pr_cont(" node=%d", pool->node);
4613 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4614}
4615
4616static void pr_cont_work(bool comma, struct work_struct *work)
4617{
4618 if (work->func == wq_barrier_func) {
4619 struct wq_barrier *barr;
4620
4621 barr = container_of(work, struct wq_barrier, work);
4622
4623 pr_cont("%s BAR(%d)", comma ? "," : "",
4624 task_pid_nr(barr->task));
4625 } else {
4626 pr_cont("%s %ps", comma ? "," : "", work->func);
4627 }
4628}
4629
4630static void show_pwq(struct pool_workqueue *pwq)
4631{
4632 struct worker_pool *pool = pwq->pool;
4633 struct work_struct *work;
4634 struct worker *worker;
4635 bool has_in_flight = false, has_pending = false;
4636 int bkt;
4637
4638 pr_info(" pwq %d:", pool->id);
4639 pr_cont_pool_info(pool);
4640
4641 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4642 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4643
4644 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4645 if (worker->current_pwq == pwq) {
4646 has_in_flight = true;
4647 break;
4648 }
4649 }
4650 if (has_in_flight) {
4651 bool comma = false;
4652
4653 pr_info(" in-flight:");
4654 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4655 if (worker->current_pwq != pwq)
4656 continue;
4657
4658 pr_cont("%s %d%s:%ps", comma ? "," : "",
4659 task_pid_nr(worker->task),
4660 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4661 worker->current_func);
4662 list_for_each_entry(work, &worker->scheduled, entry)
4663 pr_cont_work(false, work);
4664 comma = true;
4665 }
4666 pr_cont("\n");
4667 }
4668
4669 list_for_each_entry(work, &pool->worklist, entry) {
4670 if (get_work_pwq(work) == pwq) {
4671 has_pending = true;
4672 break;
4673 }
4674 }
4675 if (has_pending) {
4676 bool comma = false;
4677
4678 pr_info(" pending:");
4679 list_for_each_entry(work, &pool->worklist, entry) {
4680 if (get_work_pwq(work) != pwq)
4681 continue;
4682
4683 pr_cont_work(comma, work);
4684 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4685 }
4686 pr_cont("\n");
4687 }
4688
4689 if (!list_empty(&pwq->delayed_works)) {
4690 bool comma = false;
4691
4692 pr_info(" delayed:");
4693 list_for_each_entry(work, &pwq->delayed_works, entry) {
4694 pr_cont_work(comma, work);
4695 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4696 }
4697 pr_cont("\n");
4698 }
4699}
4700
4701/**
4702 * show_workqueue_state - dump workqueue state
4703 *
4704 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4705 * all busy workqueues and pools.
4706 */
4707void show_workqueue_state(void)
4708{
4709 struct workqueue_struct *wq;
4710 struct worker_pool *pool;
4711 unsigned long flags;
4712 int pi;
4713
4714 rcu_read_lock();
4715
4716 pr_info("Showing busy workqueues and worker pools:\n");
4717
4718 list_for_each_entry_rcu(wq, &workqueues, list) {
4719 struct pool_workqueue *pwq;
4720 bool idle = true;
4721
4722 for_each_pwq(pwq, wq) {
4723 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4724 idle = false;
4725 break;
4726 }
4727 }
4728 if (idle)
4729 continue;
4730
4731 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4732
4733 for_each_pwq(pwq, wq) {
4734 spin_lock_irqsave(&pwq->pool->lock, flags);
4735 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4736 show_pwq(pwq);
4737 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4738 /*
4739 * We could be printing a lot from atomic context, e.g.
4740 * sysrq-t -> show_workqueue_state(). Avoid triggering
4741 * hard lockup.
4742 */
4743 touch_nmi_watchdog();
4744 }
4745 }
4746
4747 for_each_pool(pool, pi) {
4748 struct worker *worker;
4749 bool first = true;
4750
4751 spin_lock_irqsave(&pool->lock, flags);
4752 if (pool->nr_workers == pool->nr_idle)
4753 goto next_pool;
4754
4755 pr_info("pool %d:", pool->id);
4756 pr_cont_pool_info(pool);
4757 pr_cont(" hung=%us workers=%d",
4758 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4759 pool->nr_workers);
4760 if (pool->manager)
4761 pr_cont(" manager: %d",
4762 task_pid_nr(pool->manager->task));
4763 list_for_each_entry(worker, &pool->idle_list, entry) {
4764 pr_cont(" %s%d", first ? "idle: " : "",
4765 task_pid_nr(worker->task));
4766 first = false;
4767 }
4768 pr_cont("\n");
4769 next_pool:
4770 spin_unlock_irqrestore(&pool->lock, flags);
4771 /*
4772 * We could be printing a lot from atomic context, e.g.
4773 * sysrq-t -> show_workqueue_state(). Avoid triggering
4774 * hard lockup.
4775 */
4776 touch_nmi_watchdog();
4777 }
4778
4779 rcu_read_unlock();
4780}
4781
4782/* used to show worker information through /proc/PID/{comm,stat,status} */
4783void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4784{
4785 int off;
4786
4787 /* always show the actual comm */
4788 off = strscpy(buf, task->comm, size);
4789 if (off < 0)
4790 return;
4791
4792 /* stabilize PF_WQ_WORKER and worker pool association */
4793 mutex_lock(&wq_pool_attach_mutex);
4794
4795 if (task->flags & PF_WQ_WORKER) {
4796 struct worker *worker = kthread_data(task);
4797 struct worker_pool *pool = worker->pool;
4798
4799 if (pool) {
4800 spin_lock_irq(&pool->lock);
4801 /*
4802 * ->desc tracks information (wq name or
4803 * set_worker_desc()) for the latest execution. If
4804 * current, prepend '+', otherwise '-'.
4805 */
4806 if (worker->desc[0] != '\0') {
4807 if (worker->current_work)
4808 scnprintf(buf + off, size - off, "+%s",
4809 worker->desc);
4810 else
4811 scnprintf(buf + off, size - off, "-%s",
4812 worker->desc);
4813 }
4814 spin_unlock_irq(&pool->lock);
4815 }
4816 }
4817
4818 mutex_unlock(&wq_pool_attach_mutex);
4819}
4820
4821#ifdef CONFIG_SMP
4822
4823/*
4824 * CPU hotplug.
4825 *
4826 * There are two challenges in supporting CPU hotplug. Firstly, there
4827 * are a lot of assumptions on strong associations among work, pwq and
4828 * pool which make migrating pending and scheduled works very
4829 * difficult to implement without impacting hot paths. Secondly,
4830 * worker pools serve mix of short, long and very long running works making
4831 * blocked draining impractical.
4832 *
4833 * This is solved by allowing the pools to be disassociated from the CPU
4834 * running as an unbound one and allowing it to be reattached later if the
4835 * cpu comes back online.
4836 */
4837
4838static void unbind_workers(int cpu)
4839{
4840 struct worker_pool *pool;
4841 struct worker *worker;
4842
4843 for_each_cpu_worker_pool(pool, cpu) {
4844 mutex_lock(&wq_pool_attach_mutex);
4845 spin_lock_irq(&pool->lock);
4846
4847 /*
4848 * We've blocked all attach/detach operations. Make all workers
4849 * unbound and set DISASSOCIATED. Before this, all workers
4850 * except for the ones which are still executing works from
4851 * before the last CPU down must be on the cpu. After
4852 * this, they may become diasporas.
4853 */
4854 for_each_pool_worker(worker, pool)
4855 worker->flags |= WORKER_UNBOUND;
4856
4857 pool->flags |= POOL_DISASSOCIATED;
4858
4859 spin_unlock_irq(&pool->lock);
4860 mutex_unlock(&wq_pool_attach_mutex);
4861
4862 /*
4863 * Call schedule() so that we cross rq->lock and thus can
4864 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4865 * This is necessary as scheduler callbacks may be invoked
4866 * from other cpus.
4867 */
4868 schedule();
4869
4870 /*
4871 * Sched callbacks are disabled now. Zap nr_running.
4872 * After this, nr_running stays zero and need_more_worker()
4873 * and keep_working() are always true as long as the
4874 * worklist is not empty. This pool now behaves as an
4875 * unbound (in terms of concurrency management) pool which
4876 * are served by workers tied to the pool.
4877 */
4878 atomic_set(&pool->nr_running, 0);
4879
4880 /*
4881 * With concurrency management just turned off, a busy
4882 * worker blocking could lead to lengthy stalls. Kick off
4883 * unbound chain execution of currently pending work items.
4884 */
4885 spin_lock_irq(&pool->lock);
4886 wake_up_worker(pool);
4887 spin_unlock_irq(&pool->lock);
4888 }
4889}
4890
4891/**
4892 * rebind_workers - rebind all workers of a pool to the associated CPU
4893 * @pool: pool of interest
4894 *
4895 * @pool->cpu is coming online. Rebind all workers to the CPU.
4896 */
4897static void rebind_workers(struct worker_pool *pool)
4898{
4899 struct worker *worker;
4900
4901 lockdep_assert_held(&wq_pool_attach_mutex);
4902
4903 /*
4904 * Restore CPU affinity of all workers. As all idle workers should
4905 * be on the run-queue of the associated CPU before any local
4906 * wake-ups for concurrency management happen, restore CPU affinity
4907 * of all workers first and then clear UNBOUND. As we're called
4908 * from CPU_ONLINE, the following shouldn't fail.
4909 */
4910 for_each_pool_worker(worker, pool)
4911 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4912 pool->attrs->cpumask) < 0);
4913
4914 spin_lock_irq(&pool->lock);
4915
4916 pool->flags &= ~POOL_DISASSOCIATED;
4917
4918 for_each_pool_worker(worker, pool) {
4919 unsigned int worker_flags = worker->flags;
4920
4921 /*
4922 * A bound idle worker should actually be on the runqueue
4923 * of the associated CPU for local wake-ups targeting it to
4924 * work. Kick all idle workers so that they migrate to the
4925 * associated CPU. Doing this in the same loop as
4926 * replacing UNBOUND with REBOUND is safe as no worker will
4927 * be bound before @pool->lock is released.
4928 */
4929 if (worker_flags & WORKER_IDLE)
4930 wake_up_process(worker->task);
4931
4932 /*
4933 * We want to clear UNBOUND but can't directly call
4934 * worker_clr_flags() or adjust nr_running. Atomically
4935 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4936 * @worker will clear REBOUND using worker_clr_flags() when
4937 * it initiates the next execution cycle thus restoring
4938 * concurrency management. Note that when or whether
4939 * @worker clears REBOUND doesn't affect correctness.
4940 *
4941 * WRITE_ONCE() is necessary because @worker->flags may be
4942 * tested without holding any lock in
4943 * wq_worker_running(). Without it, NOT_RUNNING test may
4944 * fail incorrectly leading to premature concurrency
4945 * management operations.
4946 */
4947 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4948 worker_flags |= WORKER_REBOUND;
4949 worker_flags &= ~WORKER_UNBOUND;
4950 WRITE_ONCE(worker->flags, worker_flags);
4951 }
4952
4953 spin_unlock_irq(&pool->lock);
4954}
4955
4956/**
4957 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4958 * @pool: unbound pool of interest
4959 * @cpu: the CPU which is coming up
4960 *
4961 * An unbound pool may end up with a cpumask which doesn't have any online
4962 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4963 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4964 * online CPU before, cpus_allowed of all its workers should be restored.
4965 */
4966static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4967{
4968 static cpumask_t cpumask;
4969 struct worker *worker;
4970
4971 lockdep_assert_held(&wq_pool_attach_mutex);
4972
4973 /* is @cpu allowed for @pool? */
4974 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4975 return;
4976
4977 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4978
4979 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4980 for_each_pool_worker(worker, pool)
4981 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4982}
4983
4984int workqueue_prepare_cpu(unsigned int cpu)
4985{
4986 struct worker_pool *pool;
4987
4988 for_each_cpu_worker_pool(pool, cpu) {
4989 if (pool->nr_workers)
4990 continue;
4991 if (!create_worker(pool))
4992 return -ENOMEM;
4993 }
4994 return 0;
4995}
4996
4997int workqueue_online_cpu(unsigned int cpu)
4998{
4999 struct worker_pool *pool;
5000 struct workqueue_struct *wq;
5001 int pi;
5002
5003 mutex_lock(&wq_pool_mutex);
5004
5005 for_each_pool(pool, pi) {
5006 mutex_lock(&wq_pool_attach_mutex);
5007
5008 if (pool->cpu == cpu)
5009 rebind_workers(pool);
5010 else if (pool->cpu < 0)
5011 restore_unbound_workers_cpumask(pool, cpu);
5012
5013 mutex_unlock(&wq_pool_attach_mutex);
5014 }
5015
5016 /* update NUMA affinity of unbound workqueues */
5017 list_for_each_entry(wq, &workqueues, list)
5018 wq_update_unbound_numa(wq, cpu, true);
5019
5020 mutex_unlock(&wq_pool_mutex);
5021 return 0;
5022}
5023
5024int workqueue_offline_cpu(unsigned int cpu)
5025{
5026 struct workqueue_struct *wq;
5027
5028 /* unbinding per-cpu workers should happen on the local CPU */
5029 if (WARN_ON(cpu != smp_processor_id()))
5030 return -1;
5031
5032 unbind_workers(cpu);
5033
5034 /* update NUMA affinity of unbound workqueues */
5035 mutex_lock(&wq_pool_mutex);
5036 list_for_each_entry(wq, &workqueues, list)
5037 wq_update_unbound_numa(wq, cpu, false);
5038 mutex_unlock(&wq_pool_mutex);
5039
5040 return 0;
5041}
5042
5043struct work_for_cpu {
5044 struct work_struct work;
5045 long (*fn)(void *);
5046 void *arg;
5047 long ret;
5048};
5049
5050static void work_for_cpu_fn(struct work_struct *work)
5051{
5052 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5053
5054 wfc->ret = wfc->fn(wfc->arg);
5055}
5056
5057/**
5058 * work_on_cpu - run a function in thread context on a particular cpu
5059 * @cpu: the cpu to run on
5060 * @fn: the function to run
5061 * @arg: the function arg
5062 *
5063 * It is up to the caller to ensure that the cpu doesn't go offline.
5064 * The caller must not hold any locks which would prevent @fn from completing.
5065 *
5066 * Return: The value @fn returns.
5067 */
5068long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5069{
5070 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5071
5072 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5073 schedule_work_on(cpu, &wfc.work);
5074 flush_work(&wfc.work);
5075 destroy_work_on_stack(&wfc.work);
5076 return wfc.ret;
5077}
5078EXPORT_SYMBOL_GPL(work_on_cpu);
5079
5080/**
5081 * work_on_cpu_safe - run a function in thread context on a particular cpu
5082 * @cpu: the cpu to run on
5083 * @fn: the function to run
5084 * @arg: the function argument
5085 *
5086 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5087 * any locks which would prevent @fn from completing.
5088 *
5089 * Return: The value @fn returns.
5090 */
5091long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5092{
5093 long ret = -ENODEV;
5094
5095 get_online_cpus();
5096 if (cpu_online(cpu))
5097 ret = work_on_cpu(cpu, fn, arg);
5098 put_online_cpus();
5099 return ret;
5100}
5101EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5102#endif /* CONFIG_SMP */
5103
5104#ifdef CONFIG_FREEZER
5105
5106/**
5107 * freeze_workqueues_begin - begin freezing workqueues
5108 *
5109 * Start freezing workqueues. After this function returns, all freezable
5110 * workqueues will queue new works to their delayed_works list instead of
5111 * pool->worklist.
5112 *
5113 * CONTEXT:
5114 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5115 */
5116void freeze_workqueues_begin(void)
5117{
5118 struct workqueue_struct *wq;
5119 struct pool_workqueue *pwq;
5120
5121 mutex_lock(&wq_pool_mutex);
5122
5123 WARN_ON_ONCE(workqueue_freezing);
5124 workqueue_freezing = true;
5125
5126 list_for_each_entry(wq, &workqueues, list) {
5127 mutex_lock(&wq->mutex);
5128 for_each_pwq(pwq, wq)
5129 pwq_adjust_max_active(pwq);
5130 mutex_unlock(&wq->mutex);
5131 }
5132
5133 mutex_unlock(&wq_pool_mutex);
5134}
5135
5136/**
5137 * freeze_workqueues_busy - are freezable workqueues still busy?
5138 *
5139 * Check whether freezing is complete. This function must be called
5140 * between freeze_workqueues_begin() and thaw_workqueues().
5141 *
5142 * CONTEXT:
5143 * Grabs and releases wq_pool_mutex.
5144 *
5145 * Return:
5146 * %true if some freezable workqueues are still busy. %false if freezing
5147 * is complete.
5148 */
5149bool freeze_workqueues_busy(void)
5150{
5151 bool busy = false;
5152 struct workqueue_struct *wq;
5153 struct pool_workqueue *pwq;
5154
5155 mutex_lock(&wq_pool_mutex);
5156
5157 WARN_ON_ONCE(!workqueue_freezing);
5158
5159 list_for_each_entry(wq, &workqueues, list) {
5160 if (!(wq->flags & WQ_FREEZABLE))
5161 continue;
5162 /*
5163 * nr_active is monotonically decreasing. It's safe
5164 * to peek without lock.
5165 */
5166 rcu_read_lock();
5167 for_each_pwq(pwq, wq) {
5168 WARN_ON_ONCE(pwq->nr_active < 0);
5169 if (pwq->nr_active) {
5170 busy = true;
5171 rcu_read_unlock();
5172 goto out_unlock;
5173 }
5174 }
5175 rcu_read_unlock();
5176 }
5177out_unlock:
5178 mutex_unlock(&wq_pool_mutex);
5179 return busy;
5180}
5181
5182/**
5183 * thaw_workqueues - thaw workqueues
5184 *
5185 * Thaw workqueues. Normal queueing is restored and all collected
5186 * frozen works are transferred to their respective pool worklists.
5187 *
5188 * CONTEXT:
5189 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5190 */
5191void thaw_workqueues(void)
5192{
5193 struct workqueue_struct *wq;
5194 struct pool_workqueue *pwq;
5195
5196 mutex_lock(&wq_pool_mutex);
5197
5198 if (!workqueue_freezing)
5199 goto out_unlock;
5200
5201 workqueue_freezing = false;
5202
5203 /* restore max_active and repopulate worklist */
5204 list_for_each_entry(wq, &workqueues, list) {
5205 mutex_lock(&wq->mutex);
5206 for_each_pwq(pwq, wq)
5207 pwq_adjust_max_active(pwq);
5208 mutex_unlock(&wq->mutex);
5209 }
5210
5211out_unlock:
5212 mutex_unlock(&wq_pool_mutex);
5213}
5214#endif /* CONFIG_FREEZER */
5215
5216static int workqueue_apply_unbound_cpumask(void)
5217{
5218 LIST_HEAD(ctxs);
5219 int ret = 0;
5220 struct workqueue_struct *wq;
5221 struct apply_wqattrs_ctx *ctx, *n;
5222
5223 lockdep_assert_held(&wq_pool_mutex);
5224
5225 list_for_each_entry(wq, &workqueues, list) {
5226 if (!(wq->flags & WQ_UNBOUND))
5227 continue;
5228 /* creating multiple pwqs breaks ordering guarantee */
5229 if (wq->flags & __WQ_ORDERED)
5230 continue;
5231
5232 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5233 if (!ctx) {
5234 ret = -ENOMEM;
5235 break;
5236 }
5237
5238 list_add_tail(&ctx->list, &ctxs);
5239 }
5240
5241 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5242 if (!ret)
5243 apply_wqattrs_commit(ctx);
5244 apply_wqattrs_cleanup(ctx);
5245 }
5246
5247 return ret;
5248}
5249
5250/**
5251 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5252 * @cpumask: the cpumask to set
5253 *
5254 * The low-level workqueues cpumask is a global cpumask that limits
5255 * the affinity of all unbound workqueues. This function check the @cpumask
5256 * and apply it to all unbound workqueues and updates all pwqs of them.
5257 *
5258 * Retun: 0 - Success
5259 * -EINVAL - Invalid @cpumask
5260 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5261 */
5262int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5263{
5264 int ret = -EINVAL;
5265 cpumask_var_t saved_cpumask;
5266
5267 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5268 return -ENOMEM;
5269
5270 /*
5271 * Not excluding isolated cpus on purpose.
5272 * If the user wishes to include them, we allow that.
5273 */
5274 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5275 if (!cpumask_empty(cpumask)) {
5276 apply_wqattrs_lock();
5277
5278 /* save the old wq_unbound_cpumask. */
5279 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5280
5281 /* update wq_unbound_cpumask at first and apply it to wqs. */
5282 cpumask_copy(wq_unbound_cpumask, cpumask);
5283 ret = workqueue_apply_unbound_cpumask();
5284
5285 /* restore the wq_unbound_cpumask when failed. */
5286 if (ret < 0)
5287 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5288
5289 apply_wqattrs_unlock();
5290 }
5291
5292 free_cpumask_var(saved_cpumask);
5293 return ret;
5294}
5295
5296#ifdef CONFIG_SYSFS
5297/*
5298 * Workqueues with WQ_SYSFS flag set is visible to userland via
5299 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5300 * following attributes.
5301 *
5302 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5303 * max_active RW int : maximum number of in-flight work items
5304 *
5305 * Unbound workqueues have the following extra attributes.
5306 *
5307 * pool_ids RO int : the associated pool IDs for each node
5308 * nice RW int : nice value of the workers
5309 * cpumask RW mask : bitmask of allowed CPUs for the workers
5310 * numa RW bool : whether enable NUMA affinity
5311 */
5312struct wq_device {
5313 struct workqueue_struct *wq;
5314 struct device dev;
5315};
5316
5317static struct workqueue_struct *dev_to_wq(struct device *dev)
5318{
5319 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5320
5321 return wq_dev->wq;
5322}
5323
5324static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5325 char *buf)
5326{
5327 struct workqueue_struct *wq = dev_to_wq(dev);
5328
5329 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5330}
5331static DEVICE_ATTR_RO(per_cpu);
5332
5333static ssize_t max_active_show(struct device *dev,
5334 struct device_attribute *attr, char *buf)
5335{
5336 struct workqueue_struct *wq = dev_to_wq(dev);
5337
5338 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5339}
5340
5341static ssize_t max_active_store(struct device *dev,
5342 struct device_attribute *attr, const char *buf,
5343 size_t count)
5344{
5345 struct workqueue_struct *wq = dev_to_wq(dev);
5346 int val;
5347
5348 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5349 return -EINVAL;
5350
5351 workqueue_set_max_active(wq, val);
5352 return count;
5353}
5354static DEVICE_ATTR_RW(max_active);
5355
5356static struct attribute *wq_sysfs_attrs[] = {
5357 &dev_attr_per_cpu.attr,
5358 &dev_attr_max_active.attr,
5359 NULL,
5360};
5361ATTRIBUTE_GROUPS(wq_sysfs);
5362
5363static ssize_t wq_pool_ids_show(struct device *dev,
5364 struct device_attribute *attr, char *buf)
5365{
5366 struct workqueue_struct *wq = dev_to_wq(dev);
5367 const char *delim = "";
5368 int node, written = 0;
5369
5370 get_online_cpus();
5371 rcu_read_lock();
5372 for_each_node(node) {
5373 written += scnprintf(buf + written, PAGE_SIZE - written,
5374 "%s%d:%d", delim, node,
5375 unbound_pwq_by_node(wq, node)->pool->id);
5376 delim = " ";
5377 }
5378 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5379 rcu_read_unlock();
5380 put_online_cpus();
5381
5382 return written;
5383}
5384
5385static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5386 char *buf)
5387{
5388 struct workqueue_struct *wq = dev_to_wq(dev);
5389 int written;
5390
5391 mutex_lock(&wq->mutex);
5392 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5393 mutex_unlock(&wq->mutex);
5394
5395 return written;
5396}
5397
5398/* prepare workqueue_attrs for sysfs store operations */
5399static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5400{
5401 struct workqueue_attrs *attrs;
5402
5403 lockdep_assert_held(&wq_pool_mutex);
5404
5405 attrs = alloc_workqueue_attrs();
5406 if (!attrs)
5407 return NULL;
5408
5409 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5410 return attrs;
5411}
5412
5413static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5414 const char *buf, size_t count)
5415{
5416 struct workqueue_struct *wq = dev_to_wq(dev);
5417 struct workqueue_attrs *attrs;
5418 int ret = -ENOMEM;
5419
5420 apply_wqattrs_lock();
5421
5422 attrs = wq_sysfs_prep_attrs(wq);
5423 if (!attrs)
5424 goto out_unlock;
5425
5426 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5427 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5428 ret = apply_workqueue_attrs_locked(wq, attrs);
5429 else
5430 ret = -EINVAL;
5431
5432out_unlock:
5433 apply_wqattrs_unlock();
5434 free_workqueue_attrs(attrs);
5435 return ret ?: count;
5436}
5437
5438static ssize_t wq_cpumask_show(struct device *dev,
5439 struct device_attribute *attr, char *buf)
5440{
5441 struct workqueue_struct *wq = dev_to_wq(dev);
5442 int written;
5443
5444 mutex_lock(&wq->mutex);
5445 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5446 cpumask_pr_args(wq->unbound_attrs->cpumask));
5447 mutex_unlock(&wq->mutex);
5448 return written;
5449}
5450
5451static ssize_t wq_cpumask_store(struct device *dev,
5452 struct device_attribute *attr,
5453 const char *buf, size_t count)
5454{
5455 struct workqueue_struct *wq = dev_to_wq(dev);
5456 struct workqueue_attrs *attrs;
5457 int ret = -ENOMEM;
5458
5459 apply_wqattrs_lock();
5460
5461 attrs = wq_sysfs_prep_attrs(wq);
5462 if (!attrs)
5463 goto out_unlock;
5464
5465 ret = cpumask_parse(buf, attrs->cpumask);
5466 if (!ret)
5467 ret = apply_workqueue_attrs_locked(wq, attrs);
5468
5469out_unlock:
5470 apply_wqattrs_unlock();
5471 free_workqueue_attrs(attrs);
5472 return ret ?: count;
5473}
5474
5475static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5476 char *buf)
5477{
5478 struct workqueue_struct *wq = dev_to_wq(dev);
5479 int written;
5480
5481 mutex_lock(&wq->mutex);
5482 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5483 !wq->unbound_attrs->no_numa);
5484 mutex_unlock(&wq->mutex);
5485
5486 return written;
5487}
5488
5489static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5490 const char *buf, size_t count)
5491{
5492 struct workqueue_struct *wq = dev_to_wq(dev);
5493 struct workqueue_attrs *attrs;
5494 int v, ret = -ENOMEM;
5495
5496 apply_wqattrs_lock();
5497
5498 attrs = wq_sysfs_prep_attrs(wq);
5499 if (!attrs)
5500 goto out_unlock;
5501
5502 ret = -EINVAL;
5503 if (sscanf(buf, "%d", &v) == 1) {
5504 attrs->no_numa = !v;
5505 ret = apply_workqueue_attrs_locked(wq, attrs);
5506 }
5507
5508out_unlock:
5509 apply_wqattrs_unlock();
5510 free_workqueue_attrs(attrs);
5511 return ret ?: count;
5512}
5513
5514static struct device_attribute wq_sysfs_unbound_attrs[] = {
5515 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5516 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5517 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5518 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5519 __ATTR_NULL,
5520};
5521
5522static struct bus_type wq_subsys = {
5523 .name = "workqueue",
5524 .dev_groups = wq_sysfs_groups,
5525};
5526
5527static ssize_t wq_unbound_cpumask_show(struct device *dev,
5528 struct device_attribute *attr, char *buf)
5529{
5530 int written;
5531
5532 mutex_lock(&wq_pool_mutex);
5533 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5534 cpumask_pr_args(wq_unbound_cpumask));
5535 mutex_unlock(&wq_pool_mutex);
5536
5537 return written;
5538}
5539
5540static ssize_t wq_unbound_cpumask_store(struct device *dev,
5541 struct device_attribute *attr, const char *buf, size_t count)
5542{
5543 cpumask_var_t cpumask;
5544 int ret;
5545
5546 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5547 return -ENOMEM;
5548
5549 ret = cpumask_parse(buf, cpumask);
5550 if (!ret)
5551 ret = workqueue_set_unbound_cpumask(cpumask);
5552
5553 free_cpumask_var(cpumask);
5554 return ret ? ret : count;
5555}
5556
5557static struct device_attribute wq_sysfs_cpumask_attr =
5558 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5559 wq_unbound_cpumask_store);
5560
5561static int __init wq_sysfs_init(void)
5562{
5563 int err;
5564
5565 err = subsys_virtual_register(&wq_subsys, NULL);
5566 if (err)
5567 return err;
5568
5569 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5570}
5571core_initcall(wq_sysfs_init);
5572
5573static void wq_device_release(struct device *dev)
5574{
5575 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5576
5577 kfree(wq_dev);
5578}
5579
5580/**
5581 * workqueue_sysfs_register - make a workqueue visible in sysfs
5582 * @wq: the workqueue to register
5583 *
5584 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5585 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5586 * which is the preferred method.
5587 *
5588 * Workqueue user should use this function directly iff it wants to apply
5589 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5590 * apply_workqueue_attrs() may race against userland updating the
5591 * attributes.
5592 *
5593 * Return: 0 on success, -errno on failure.
5594 */
5595int workqueue_sysfs_register(struct workqueue_struct *wq)
5596{
5597 struct wq_device *wq_dev;
5598 int ret;
5599
5600 /*
5601 * Adjusting max_active or creating new pwqs by applying
5602 * attributes breaks ordering guarantee. Disallow exposing ordered
5603 * workqueues.
5604 */
5605 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5606 return -EINVAL;
5607
5608 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5609 if (!wq_dev)
5610 return -ENOMEM;
5611
5612 wq_dev->wq = wq;
5613 wq_dev->dev.bus = &wq_subsys;
5614 wq_dev->dev.release = wq_device_release;
5615 dev_set_name(&wq_dev->dev, "%s", wq->name);
5616
5617 /*
5618 * unbound_attrs are created separately. Suppress uevent until
5619 * everything is ready.
5620 */
5621 dev_set_uevent_suppress(&wq_dev->dev, true);
5622
5623 ret = device_register(&wq_dev->dev);
5624 if (ret) {
5625 put_device(&wq_dev->dev);
5626 wq->wq_dev = NULL;
5627 return ret;
5628 }
5629
5630 if (wq->flags & WQ_UNBOUND) {
5631 struct device_attribute *attr;
5632
5633 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5634 ret = device_create_file(&wq_dev->dev, attr);
5635 if (ret) {
5636 device_unregister(&wq_dev->dev);
5637 wq->wq_dev = NULL;
5638 return ret;
5639 }
5640 }
5641 }
5642
5643 dev_set_uevent_suppress(&wq_dev->dev, false);
5644 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5645 return 0;
5646}
5647
5648/**
5649 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5650 * @wq: the workqueue to unregister
5651 *
5652 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5653 */
5654static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5655{
5656 struct wq_device *wq_dev = wq->wq_dev;
5657
5658 if (!wq->wq_dev)
5659 return;
5660
5661 wq->wq_dev = NULL;
5662 device_unregister(&wq_dev->dev);
5663}
5664#else /* CONFIG_SYSFS */
5665static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5666#endif /* CONFIG_SYSFS */
5667
5668/*
5669 * Workqueue watchdog.
5670 *
5671 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5672 * flush dependency, a concurrency managed work item which stays RUNNING
5673 * indefinitely. Workqueue stalls can be very difficult to debug as the
5674 * usual warning mechanisms don't trigger and internal workqueue state is
5675 * largely opaque.
5676 *
5677 * Workqueue watchdog monitors all worker pools periodically and dumps
5678 * state if some pools failed to make forward progress for a while where
5679 * forward progress is defined as the first item on ->worklist changing.
5680 *
5681 * This mechanism is controlled through the kernel parameter
5682 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5683 * corresponding sysfs parameter file.
5684 */
5685#ifdef CONFIG_WQ_WATCHDOG
5686
5687static unsigned long wq_watchdog_thresh = 30;
5688static struct timer_list wq_watchdog_timer;
5689
5690static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5691static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5692
5693static void wq_watchdog_reset_touched(void)
5694{
5695 int cpu;
5696
5697 wq_watchdog_touched = jiffies;
5698 for_each_possible_cpu(cpu)
5699 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5700}
5701
5702static void wq_watchdog_timer_fn(struct timer_list *unused)
5703{
5704 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5705 bool lockup_detected = false;
5706 struct worker_pool *pool;
5707 int pi;
5708
5709 if (!thresh)
5710 return;
5711
5712 rcu_read_lock();
5713
5714 for_each_pool(pool, pi) {
5715 unsigned long pool_ts, touched, ts;
5716
5717 if (list_empty(&pool->worklist))
5718 continue;
5719
5720 /* get the latest of pool and touched timestamps */
5721 pool_ts = READ_ONCE(pool->watchdog_ts);
5722 touched = READ_ONCE(wq_watchdog_touched);
5723
5724 if (time_after(pool_ts, touched))
5725 ts = pool_ts;
5726 else
5727 ts = touched;
5728
5729 if (pool->cpu >= 0) {
5730 unsigned long cpu_touched =
5731 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5732 pool->cpu));
5733 if (time_after(cpu_touched, ts))
5734 ts = cpu_touched;
5735 }
5736
5737 /* did we stall? */
5738 if (time_after(jiffies, ts + thresh)) {
5739 lockup_detected = true;
5740 pr_emerg("BUG: workqueue lockup - pool");
5741 pr_cont_pool_info(pool);
5742 pr_cont(" stuck for %us!\n",
5743 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5744 }
5745 }
5746
5747 rcu_read_unlock();
5748
5749 if (lockup_detected)
5750 show_workqueue_state();
5751
5752 wq_watchdog_reset_touched();
5753 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5754}
5755
5756notrace void wq_watchdog_touch(int cpu)
5757{
5758 if (cpu >= 0)
5759 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5760 else
5761 wq_watchdog_touched = jiffies;
5762}
5763
5764static void wq_watchdog_set_thresh(unsigned long thresh)
5765{
5766 wq_watchdog_thresh = 0;
5767 del_timer_sync(&wq_watchdog_timer);
5768
5769 if (thresh) {
5770 wq_watchdog_thresh = thresh;
5771 wq_watchdog_reset_touched();
5772 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5773 }
5774}
5775
5776static int wq_watchdog_param_set_thresh(const char *val,
5777 const struct kernel_param *kp)
5778{
5779 unsigned long thresh;
5780 int ret;
5781
5782 ret = kstrtoul(val, 0, &thresh);
5783 if (ret)
5784 return ret;
5785
5786 if (system_wq)
5787 wq_watchdog_set_thresh(thresh);
5788 else
5789 wq_watchdog_thresh = thresh;
5790
5791 return 0;
5792}
5793
5794static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5795 .set = wq_watchdog_param_set_thresh,
5796 .get = param_get_ulong,
5797};
5798
5799module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5800 0644);
5801
5802static void wq_watchdog_init(void)
5803{
5804 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5805 wq_watchdog_set_thresh(wq_watchdog_thresh);
5806}
5807
5808#else /* CONFIG_WQ_WATCHDOG */
5809
5810static inline void wq_watchdog_init(void) { }
5811
5812#endif /* CONFIG_WQ_WATCHDOG */
5813
5814static void __init wq_numa_init(void)
5815{
5816 cpumask_var_t *tbl;
5817 int node, cpu;
5818
5819 if (num_possible_nodes() <= 1)
5820 return;
5821
5822 if (wq_disable_numa) {
5823 pr_info("workqueue: NUMA affinity support disabled\n");
5824 return;
5825 }
5826
5827 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5828 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5829
5830 /*
5831 * We want masks of possible CPUs of each node which isn't readily
5832 * available. Build one from cpu_to_node() which should have been
5833 * fully initialized by now.
5834 */
5835 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5836 BUG_ON(!tbl);
5837
5838 for_each_node(node)
5839 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5840 node_online(node) ? node : NUMA_NO_NODE));
5841
5842 for_each_possible_cpu(cpu) {
5843 node = cpu_to_node(cpu);
5844 if (WARN_ON(node == NUMA_NO_NODE)) {
5845 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5846 /* happens iff arch is bonkers, let's just proceed */
5847 return;
5848 }
5849 cpumask_set_cpu(cpu, tbl[node]);
5850 }
5851
5852 wq_numa_possible_cpumask = tbl;
5853 wq_numa_enabled = true;
5854}
5855
5856/**
5857 * workqueue_init_early - early init for workqueue subsystem
5858 *
5859 * This is the first half of two-staged workqueue subsystem initialization
5860 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5861 * idr are up. It sets up all the data structures and system workqueues
5862 * and allows early boot code to create workqueues and queue/cancel work
5863 * items. Actual work item execution starts only after kthreads can be
5864 * created and scheduled right before early initcalls.
5865 */
5866int __init workqueue_init_early(void)
5867{
5868 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5869 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5870 int i, cpu;
5871
5872 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5873
5874 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5875 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5876
5877 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5878
5879 /* initialize CPU pools */
5880 for_each_possible_cpu(cpu) {
5881 struct worker_pool *pool;
5882
5883 i = 0;
5884 for_each_cpu_worker_pool(pool, cpu) {
5885 BUG_ON(init_worker_pool(pool));
5886 pool->cpu = cpu;
5887 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5888 pool->attrs->nice = std_nice[i++];
5889 pool->node = cpu_to_node(cpu);
5890
5891 /* alloc pool ID */
5892 mutex_lock(&wq_pool_mutex);
5893 BUG_ON(worker_pool_assign_id(pool));
5894 mutex_unlock(&wq_pool_mutex);
5895 }
5896 }
5897
5898 /* create default unbound and ordered wq attrs */
5899 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5900 struct workqueue_attrs *attrs;
5901
5902 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5903 attrs->nice = std_nice[i];
5904 unbound_std_wq_attrs[i] = attrs;
5905
5906 /*
5907 * An ordered wq should have only one pwq as ordering is
5908 * guaranteed by max_active which is enforced by pwqs.
5909 * Turn off NUMA so that dfl_pwq is used for all nodes.
5910 */
5911 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5912 attrs->nice = std_nice[i];
5913 attrs->no_numa = true;
5914 ordered_wq_attrs[i] = attrs;
5915 }
5916
5917 system_wq = alloc_workqueue("events", 0, 0);
5918 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5919 system_long_wq = alloc_workqueue("events_long", 0, 0);
5920 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5921 WQ_UNBOUND_MAX_ACTIVE);
5922 system_freezable_wq = alloc_workqueue("events_freezable",
5923 WQ_FREEZABLE, 0);
5924 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5925 WQ_POWER_EFFICIENT, 0);
5926 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5927 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5928 0);
5929 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5930 !system_unbound_wq || !system_freezable_wq ||
5931 !system_power_efficient_wq ||
5932 !system_freezable_power_efficient_wq);
5933
5934 return 0;
5935}
5936
5937/**
5938 * workqueue_init - bring workqueue subsystem fully online
5939 *
5940 * This is the latter half of two-staged workqueue subsystem initialization
5941 * and invoked as soon as kthreads can be created and scheduled.
5942 * Workqueues have been created and work items queued on them, but there
5943 * are no kworkers executing the work items yet. Populate the worker pools
5944 * with the initial workers and enable future kworker creations.
5945 */
5946int __init workqueue_init(void)
5947{
5948 struct workqueue_struct *wq;
5949 struct worker_pool *pool;
5950 int cpu, bkt;
5951
5952 /*
5953 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5954 * CPU to node mapping may not be available that early on some
5955 * archs such as power and arm64. As per-cpu pools created
5956 * previously could be missing node hint and unbound pools NUMA
5957 * affinity, fix them up.
5958 *
5959 * Also, while iterating workqueues, create rescuers if requested.
5960 */
5961 wq_numa_init();
5962
5963 mutex_lock(&wq_pool_mutex);
5964
5965 for_each_possible_cpu(cpu) {
5966 for_each_cpu_worker_pool(pool, cpu) {
5967 pool->node = cpu_to_node(cpu);
5968 }
5969 }
5970
5971 list_for_each_entry(wq, &workqueues, list) {
5972 wq_update_unbound_numa(wq, smp_processor_id(), true);
5973 WARN(init_rescuer(wq),
5974 "workqueue: failed to create early rescuer for %s",
5975 wq->name);
5976 }
5977
5978 mutex_unlock(&wq_pool_mutex);
5979
5980 /* create the initial workers */
5981 for_each_online_cpu(cpu) {
5982 for_each_cpu_worker_pool(pool, cpu) {
5983 pool->flags &= ~POOL_DISASSOCIATED;
5984 BUG_ON(!create_worker(pool));
5985 }
5986 }
5987
5988 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5989 BUG_ON(!create_worker(pool));
5990
5991 wq_online = true;
5992 wq_watchdog_init();
5993
5994 return 0;
5995}
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#include <linux/kvm_para.h>
54
55#include "workqueue_internal.h"
56
57enum {
58 /*
59 * worker_pool flags
60 *
61 * A bound pool is either associated or disassociated with its CPU.
62 * While associated (!DISASSOCIATED), all workers are bound to the
63 * CPU and none has %WORKER_UNBOUND set and concurrency management
64 * is in effect.
65 *
66 * While DISASSOCIATED, the cpu may be offline and all workers have
67 * %WORKER_UNBOUND set and concurrency management disabled, and may
68 * be executing on any CPU. The pool behaves as an unbound one.
69 *
70 * Note that DISASSOCIATED should be flipped only while holding
71 * wq_pool_attach_mutex to avoid changing binding state while
72 * worker_attach_to_pool() is in progress.
73 */
74 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
75 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
76
77 /* worker flags */
78 WORKER_DIE = 1 << 1, /* die die die */
79 WORKER_IDLE = 1 << 2, /* is idle */
80 WORKER_PREP = 1 << 3, /* preparing to run works */
81 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
82 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
83 WORKER_REBOUND = 1 << 8, /* worker was rebound */
84
85 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
86 WORKER_UNBOUND | WORKER_REBOUND,
87
88 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
89
90 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
91 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
92
93 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
94 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
95
96 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
97 /* call for help after 10ms
98 (min two ticks) */
99 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
100 CREATE_COOLDOWN = HZ, /* time to breath after fail */
101
102 /*
103 * Rescue workers are used only on emergencies and shared by
104 * all cpus. Give MIN_NICE.
105 */
106 RESCUER_NICE_LEVEL = MIN_NICE,
107 HIGHPRI_NICE_LEVEL = MIN_NICE,
108
109 WQ_NAME_LEN = 24,
110};
111
112/*
113 * Structure fields follow one of the following exclusion rules.
114 *
115 * I: Modifiable by initialization/destruction paths and read-only for
116 * everyone else.
117 *
118 * P: Preemption protected. Disabling preemption is enough and should
119 * only be modified and accessed from the local cpu.
120 *
121 * L: pool->lock protected. Access with pool->lock held.
122 *
123 * X: During normal operation, modification requires pool->lock and should
124 * be done only from local cpu. Either disabling preemption on local
125 * cpu or grabbing pool->lock is enough for read access. If
126 * POOL_DISASSOCIATED is set, it's identical to L.
127 *
128 * A: wq_pool_attach_mutex protected.
129 *
130 * PL: wq_pool_mutex protected.
131 *
132 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
133 *
134 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
135 *
136 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
137 * RCU for reads.
138 *
139 * WQ: wq->mutex protected.
140 *
141 * WR: wq->mutex protected for writes. RCU protected for reads.
142 *
143 * MD: wq_mayday_lock protected.
144 */
145
146/* struct worker is defined in workqueue_internal.h */
147
148struct worker_pool {
149 raw_spinlock_t lock; /* the pool lock */
150 int cpu; /* I: the associated cpu */
151 int node; /* I: the associated node ID */
152 int id; /* I: pool ID */
153 unsigned int flags; /* X: flags */
154
155 unsigned long watchdog_ts; /* L: watchdog timestamp */
156
157 struct list_head worklist; /* L: list of pending works */
158
159 int nr_workers; /* L: total number of workers */
160 int nr_idle; /* L: currently idle workers */
161
162 struct list_head idle_list; /* X: list of idle workers */
163 struct timer_list idle_timer; /* L: worker idle timeout */
164 struct timer_list mayday_timer; /* L: SOS timer for workers */
165
166 /* a workers is either on busy_hash or idle_list, or the manager */
167 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
168 /* L: hash of busy workers */
169
170 struct worker *manager; /* L: purely informational */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
173
174 struct ida worker_ida; /* worker IDs for task name */
175
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
179
180 /*
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
183 * cacheline.
184 */
185 atomic_t nr_running ____cacheline_aligned_in_smp;
186
187 /*
188 * Destruction of pool is RCU protected to allow dereferences
189 * from get_work_pool().
190 */
191 struct rcu_head rcu;
192} ____cacheline_aligned_in_smp;
193
194/*
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
199 */
200struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
213
214 /*
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
219 */
220 struct work_struct unbound_release_work;
221 struct rcu_head rcu;
222} __aligned(1 << WORK_STRUCT_FLAG_BITS);
223
224/*
225 * Structure used to wait for workqueue flush.
226 */
227struct wq_flusher {
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
231};
232
233struct wq_device;
234
235/*
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
238 */
239struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
242
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
250
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* MD: rescue worker */
253
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
256
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259
260#ifdef CONFIG_SYSFS
261 struct wq_device *wq_dev; /* I: for sysfs interface */
262#endif
263#ifdef CONFIG_LOCKDEP
264 char *lock_name;
265 struct lock_class_key key;
266 struct lockdep_map lockdep_map;
267#endif
268 char name[WQ_NAME_LEN]; /* I: workqueue name */
269
270 /*
271 * Destruction of workqueue_struct is RCU protected to allow walking
272 * the workqueues list without grabbing wq_pool_mutex.
273 * This is used to dump all workqueues from sysrq.
274 */
275 struct rcu_head rcu;
276
277 /* hot fields used during command issue, aligned to cacheline */
278 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
279 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
280 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
281};
282
283static struct kmem_cache *pwq_cache;
284
285static cpumask_var_t *wq_numa_possible_cpumask;
286 /* possible CPUs of each node */
287
288static bool wq_disable_numa;
289module_param_named(disable_numa, wq_disable_numa, bool, 0444);
290
291/* see the comment above the definition of WQ_POWER_EFFICIENT */
292static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
293module_param_named(power_efficient, wq_power_efficient, bool, 0444);
294
295static bool wq_online; /* can kworkers be created yet? */
296
297static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
298
299/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
300static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
301
302static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
303static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
304static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
305/* wait for manager to go away */
306static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
307
308static LIST_HEAD(workqueues); /* PR: list of all workqueues */
309static bool workqueue_freezing; /* PL: have wqs started freezing? */
310
311/* PL: allowable cpus for unbound wqs and work items */
312static cpumask_var_t wq_unbound_cpumask;
313
314/* CPU where unbound work was last round robin scheduled from this CPU */
315static DEFINE_PER_CPU(int, wq_rr_cpu_last);
316
317/*
318 * Local execution of unbound work items is no longer guaranteed. The
319 * following always forces round-robin CPU selection on unbound work items
320 * to uncover usages which depend on it.
321 */
322#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
323static bool wq_debug_force_rr_cpu = true;
324#else
325static bool wq_debug_force_rr_cpu = false;
326#endif
327module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
328
329/* the per-cpu worker pools */
330static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
331
332static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
333
334/* PL: hash of all unbound pools keyed by pool->attrs */
335static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
336
337/* I: attributes used when instantiating standard unbound pools on demand */
338static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
339
340/* I: attributes used when instantiating ordered pools on demand */
341static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
342
343struct workqueue_struct *system_wq __read_mostly;
344EXPORT_SYMBOL(system_wq);
345struct workqueue_struct *system_highpri_wq __read_mostly;
346EXPORT_SYMBOL_GPL(system_highpri_wq);
347struct workqueue_struct *system_long_wq __read_mostly;
348EXPORT_SYMBOL_GPL(system_long_wq);
349struct workqueue_struct *system_unbound_wq __read_mostly;
350EXPORT_SYMBOL_GPL(system_unbound_wq);
351struct workqueue_struct *system_freezable_wq __read_mostly;
352EXPORT_SYMBOL_GPL(system_freezable_wq);
353struct workqueue_struct *system_power_efficient_wq __read_mostly;
354EXPORT_SYMBOL_GPL(system_power_efficient_wq);
355struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
356EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
357
358static int worker_thread(void *__worker);
359static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
360static void show_pwq(struct pool_workqueue *pwq);
361
362#define CREATE_TRACE_POINTS
363#include <trace/events/workqueue.h>
364
365#define assert_rcu_or_pool_mutex() \
366 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
367 !lockdep_is_held(&wq_pool_mutex), \
368 "RCU or wq_pool_mutex should be held")
369
370#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
371 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
372 !lockdep_is_held(&wq->mutex) && \
373 !lockdep_is_held(&wq_pool_mutex), \
374 "RCU, wq->mutex or wq_pool_mutex should be held")
375
376#define for_each_cpu_worker_pool(pool, cpu) \
377 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
378 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
379 (pool)++)
380
381/**
382 * for_each_pool - iterate through all worker_pools in the system
383 * @pool: iteration cursor
384 * @pi: integer used for iteration
385 *
386 * This must be called either with wq_pool_mutex held or RCU read
387 * locked. If the pool needs to be used beyond the locking in effect, the
388 * caller is responsible for guaranteeing that the pool stays online.
389 *
390 * The if/else clause exists only for the lockdep assertion and can be
391 * ignored.
392 */
393#define for_each_pool(pool, pi) \
394 idr_for_each_entry(&worker_pool_idr, pool, pi) \
395 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
396 else
397
398/**
399 * for_each_pool_worker - iterate through all workers of a worker_pool
400 * @worker: iteration cursor
401 * @pool: worker_pool to iterate workers of
402 *
403 * This must be called with wq_pool_attach_mutex.
404 *
405 * The if/else clause exists only for the lockdep assertion and can be
406 * ignored.
407 */
408#define for_each_pool_worker(worker, pool) \
409 list_for_each_entry((worker), &(pool)->workers, node) \
410 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
411 else
412
413/**
414 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
415 * @pwq: iteration cursor
416 * @wq: the target workqueue
417 *
418 * This must be called either with wq->mutex held or RCU read locked.
419 * If the pwq needs to be used beyond the locking in effect, the caller is
420 * responsible for guaranteeing that the pwq stays online.
421 *
422 * The if/else clause exists only for the lockdep assertion and can be
423 * ignored.
424 */
425#define for_each_pwq(pwq, wq) \
426 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
427 lockdep_is_held(&(wq->mutex)))
428
429#ifdef CONFIG_DEBUG_OBJECTS_WORK
430
431static const 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 const 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 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 queue
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 RCU read lock. As such, this function should be
698 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
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 * raw_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_running - a worker is running again
843 * @task: task waking up
844 *
845 * This function is called when a worker returns from schedule()
846 */
847void wq_worker_running(struct task_struct *task)
848{
849 struct worker *worker = kthread_data(task);
850
851 if (!worker->sleeping)
852 return;
853 if (!(worker->flags & WORKER_NOT_RUNNING))
854 atomic_inc(&worker->pool->nr_running);
855 worker->sleeping = 0;
856}
857
858/**
859 * wq_worker_sleeping - a worker is going to sleep
860 * @task: task going to sleep
861 *
862 * This function is called from schedule() when a busy worker is
863 * going to sleep. Preemption needs to be disabled to protect ->sleeping
864 * assignment.
865 */
866void wq_worker_sleeping(struct task_struct *task)
867{
868 struct worker *next, *worker = kthread_data(task);
869 struct worker_pool *pool;
870
871 /*
872 * Rescuers, which may not have all the fields set up like normal
873 * workers, also reach here, let's not access anything before
874 * checking NOT_RUNNING.
875 */
876 if (worker->flags & WORKER_NOT_RUNNING)
877 return;
878
879 pool = worker->pool;
880
881 /* Return if preempted before wq_worker_running() was reached */
882 if (worker->sleeping)
883 return;
884
885 worker->sleeping = 1;
886 raw_spin_lock_irq(&pool->lock);
887
888 /*
889 * The counterpart of the following dec_and_test, implied mb,
890 * worklist not empty test sequence is in insert_work().
891 * Please read comment there.
892 *
893 * NOT_RUNNING is clear. This means that we're bound to and
894 * running on the local cpu w/ rq lock held and preemption
895 * disabled, which in turn means that none else could be
896 * manipulating idle_list, so dereferencing idle_list without pool
897 * lock is safe.
898 */
899 if (atomic_dec_and_test(&pool->nr_running) &&
900 !list_empty(&pool->worklist)) {
901 next = first_idle_worker(pool);
902 if (next)
903 wake_up_process(next->task);
904 }
905 raw_spin_unlock_irq(&pool->lock);
906}
907
908/**
909 * wq_worker_last_func - retrieve worker's last work function
910 * @task: Task to retrieve last work function of.
911 *
912 * Determine the last function a worker executed. This is called from
913 * the scheduler to get a worker's last known identity.
914 *
915 * CONTEXT:
916 * raw_spin_lock_irq(rq->lock)
917 *
918 * This function is called during schedule() when a kworker is going
919 * to sleep. It's used by psi to identify aggregation workers during
920 * dequeuing, to allow periodic aggregation to shut-off when that
921 * worker is the last task in the system or cgroup to go to sleep.
922 *
923 * As this function doesn't involve any workqueue-related locking, it
924 * only returns stable values when called from inside the scheduler's
925 * queuing and dequeuing paths, when @task, which must be a kworker,
926 * is guaranteed to not be processing any works.
927 *
928 * Return:
929 * The last work function %current executed as a worker, NULL if it
930 * hasn't executed any work yet.
931 */
932work_func_t wq_worker_last_func(struct task_struct *task)
933{
934 struct worker *worker = kthread_data(task);
935
936 return worker->last_func;
937}
938
939/**
940 * worker_set_flags - set worker flags and adjust nr_running accordingly
941 * @worker: self
942 * @flags: flags to set
943 *
944 * Set @flags in @worker->flags and adjust nr_running accordingly.
945 *
946 * CONTEXT:
947 * raw_spin_lock_irq(pool->lock)
948 */
949static inline void worker_set_flags(struct worker *worker, unsigned int flags)
950{
951 struct worker_pool *pool = worker->pool;
952
953 WARN_ON_ONCE(worker->task != current);
954
955 /* If transitioning into NOT_RUNNING, adjust nr_running. */
956 if ((flags & WORKER_NOT_RUNNING) &&
957 !(worker->flags & WORKER_NOT_RUNNING)) {
958 atomic_dec(&pool->nr_running);
959 }
960
961 worker->flags |= flags;
962}
963
964/**
965 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
966 * @worker: self
967 * @flags: flags to clear
968 *
969 * Clear @flags in @worker->flags and adjust nr_running accordingly.
970 *
971 * CONTEXT:
972 * raw_spin_lock_irq(pool->lock)
973 */
974static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
975{
976 struct worker_pool *pool = worker->pool;
977 unsigned int oflags = worker->flags;
978
979 WARN_ON_ONCE(worker->task != current);
980
981 worker->flags &= ~flags;
982
983 /*
984 * If transitioning out of NOT_RUNNING, increment nr_running. Note
985 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
986 * of multiple flags, not a single flag.
987 */
988 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
989 if (!(worker->flags & WORKER_NOT_RUNNING))
990 atomic_inc(&pool->nr_running);
991}
992
993/**
994 * find_worker_executing_work - find worker which is executing a work
995 * @pool: pool of interest
996 * @work: work to find worker for
997 *
998 * Find a worker which is executing @work on @pool by searching
999 * @pool->busy_hash which is keyed by the address of @work. For a worker
1000 * to match, its current execution should match the address of @work and
1001 * its work function. This is to avoid unwanted dependency between
1002 * unrelated work executions through a work item being recycled while still
1003 * being executed.
1004 *
1005 * This is a bit tricky. A work item may be freed once its execution
1006 * starts and nothing prevents the freed area from being recycled for
1007 * another work item. If the same work item address ends up being reused
1008 * before the original execution finishes, workqueue will identify the
1009 * recycled work item as currently executing and make it wait until the
1010 * current execution finishes, introducing an unwanted dependency.
1011 *
1012 * This function checks the work item address and work function to avoid
1013 * false positives. Note that this isn't complete as one may construct a
1014 * work function which can introduce dependency onto itself through a
1015 * recycled work item. Well, if somebody wants to shoot oneself in the
1016 * foot that badly, there's only so much we can do, and if such deadlock
1017 * actually occurs, it should be easy to locate the culprit work function.
1018 *
1019 * CONTEXT:
1020 * raw_spin_lock_irq(pool->lock).
1021 *
1022 * Return:
1023 * Pointer to worker which is executing @work if found, %NULL
1024 * otherwise.
1025 */
1026static struct worker *find_worker_executing_work(struct worker_pool *pool,
1027 struct work_struct *work)
1028{
1029 struct worker *worker;
1030
1031 hash_for_each_possible(pool->busy_hash, worker, hentry,
1032 (unsigned long)work)
1033 if (worker->current_work == work &&
1034 worker->current_func == work->func)
1035 return worker;
1036
1037 return NULL;
1038}
1039
1040/**
1041 * move_linked_works - move linked works to a list
1042 * @work: start of series of works to be scheduled
1043 * @head: target list to append @work to
1044 * @nextp: out parameter for nested worklist walking
1045 *
1046 * Schedule linked works starting from @work to @head. Work series to
1047 * be scheduled starts at @work and includes any consecutive work with
1048 * WORK_STRUCT_LINKED set in its predecessor.
1049 *
1050 * If @nextp is not NULL, it's updated to point to the next work of
1051 * the last scheduled work. This allows move_linked_works() to be
1052 * nested inside outer list_for_each_entry_safe().
1053 *
1054 * CONTEXT:
1055 * raw_spin_lock_irq(pool->lock).
1056 */
1057static void move_linked_works(struct work_struct *work, struct list_head *head,
1058 struct work_struct **nextp)
1059{
1060 struct work_struct *n;
1061
1062 /*
1063 * Linked worklist will always end before the end of the list,
1064 * use NULL for list head.
1065 */
1066 list_for_each_entry_safe_from(work, n, NULL, entry) {
1067 list_move_tail(&work->entry, head);
1068 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1069 break;
1070 }
1071
1072 /*
1073 * If we're already inside safe list traversal and have moved
1074 * multiple works to the scheduled queue, the next position
1075 * needs to be updated.
1076 */
1077 if (nextp)
1078 *nextp = n;
1079}
1080
1081/**
1082 * get_pwq - get an extra reference on the specified pool_workqueue
1083 * @pwq: pool_workqueue to get
1084 *
1085 * Obtain an extra reference on @pwq. The caller should guarantee that
1086 * @pwq has positive refcnt and be holding the matching pool->lock.
1087 */
1088static void get_pwq(struct pool_workqueue *pwq)
1089{
1090 lockdep_assert_held(&pwq->pool->lock);
1091 WARN_ON_ONCE(pwq->refcnt <= 0);
1092 pwq->refcnt++;
1093}
1094
1095/**
1096 * put_pwq - put a pool_workqueue reference
1097 * @pwq: pool_workqueue to put
1098 *
1099 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1100 * destruction. The caller should be holding the matching pool->lock.
1101 */
1102static void put_pwq(struct pool_workqueue *pwq)
1103{
1104 lockdep_assert_held(&pwq->pool->lock);
1105 if (likely(--pwq->refcnt))
1106 return;
1107 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1108 return;
1109 /*
1110 * @pwq can't be released under pool->lock, bounce to
1111 * pwq_unbound_release_workfn(). This never recurses on the same
1112 * pool->lock as this path is taken only for unbound workqueues and
1113 * the release work item is scheduled on a per-cpu workqueue. To
1114 * avoid lockdep warning, unbound pool->locks are given lockdep
1115 * subclass of 1 in get_unbound_pool().
1116 */
1117 schedule_work(&pwq->unbound_release_work);
1118}
1119
1120/**
1121 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1122 * @pwq: pool_workqueue to put (can be %NULL)
1123 *
1124 * put_pwq() with locking. This function also allows %NULL @pwq.
1125 */
1126static void put_pwq_unlocked(struct pool_workqueue *pwq)
1127{
1128 if (pwq) {
1129 /*
1130 * As both pwqs and pools are RCU protected, the
1131 * following lock operations are safe.
1132 */
1133 raw_spin_lock_irq(&pwq->pool->lock);
1134 put_pwq(pwq);
1135 raw_spin_unlock_irq(&pwq->pool->lock);
1136 }
1137}
1138
1139static void pwq_activate_delayed_work(struct work_struct *work)
1140{
1141 struct pool_workqueue *pwq = get_work_pwq(work);
1142
1143 trace_workqueue_activate_work(work);
1144 if (list_empty(&pwq->pool->worklist))
1145 pwq->pool->watchdog_ts = jiffies;
1146 move_linked_works(work, &pwq->pool->worklist, NULL);
1147 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1148 pwq->nr_active++;
1149}
1150
1151static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1152{
1153 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1154 struct work_struct, entry);
1155
1156 pwq_activate_delayed_work(work);
1157}
1158
1159/**
1160 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1161 * @pwq: pwq of interest
1162 * @color: color of work which left the queue
1163 *
1164 * A work either has completed or is removed from pending queue,
1165 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1166 *
1167 * CONTEXT:
1168 * raw_spin_lock_irq(pool->lock).
1169 */
1170static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1171{
1172 /* uncolored work items don't participate in flushing or nr_active */
1173 if (color == WORK_NO_COLOR)
1174 goto out_put;
1175
1176 pwq->nr_in_flight[color]--;
1177
1178 pwq->nr_active--;
1179 if (!list_empty(&pwq->delayed_works)) {
1180 /* one down, submit a delayed one */
1181 if (pwq->nr_active < pwq->max_active)
1182 pwq_activate_first_delayed(pwq);
1183 }
1184
1185 /* is flush in progress and are we at the flushing tip? */
1186 if (likely(pwq->flush_color != color))
1187 goto out_put;
1188
1189 /* are there still in-flight works? */
1190 if (pwq->nr_in_flight[color])
1191 goto out_put;
1192
1193 /* this pwq is done, clear flush_color */
1194 pwq->flush_color = -1;
1195
1196 /*
1197 * If this was the last pwq, wake up the first flusher. It
1198 * will handle the rest.
1199 */
1200 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1201 complete(&pwq->wq->first_flusher->done);
1202out_put:
1203 put_pwq(pwq);
1204}
1205
1206/**
1207 * try_to_grab_pending - steal work item from worklist and disable irq
1208 * @work: work item to steal
1209 * @is_dwork: @work is a delayed_work
1210 * @flags: place to store irq state
1211 *
1212 * Try to grab PENDING bit of @work. This function can handle @work in any
1213 * stable state - idle, on timer or on worklist.
1214 *
1215 * Return:
1216 *
1217 * ======== ================================================================
1218 * 1 if @work was pending and we successfully stole PENDING
1219 * 0 if @work was idle and we claimed PENDING
1220 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1221 * -ENOENT if someone else is canceling @work, this state may persist
1222 * for arbitrarily long
1223 * ======== ================================================================
1224 *
1225 * Note:
1226 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1227 * interrupted while holding PENDING and @work off queue, irq must be
1228 * disabled on entry. This, combined with delayed_work->timer being
1229 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1230 *
1231 * On successful return, >= 0, irq is disabled and the caller is
1232 * responsible for releasing it using local_irq_restore(*@flags).
1233 *
1234 * This function is safe to call from any context including IRQ handler.
1235 */
1236static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1237 unsigned long *flags)
1238{
1239 struct worker_pool *pool;
1240 struct pool_workqueue *pwq;
1241
1242 local_irq_save(*flags);
1243
1244 /* try to steal the timer if it exists */
1245 if (is_dwork) {
1246 struct delayed_work *dwork = to_delayed_work(work);
1247
1248 /*
1249 * dwork->timer is irqsafe. If del_timer() fails, it's
1250 * guaranteed that the timer is not queued anywhere and not
1251 * running on the local CPU.
1252 */
1253 if (likely(del_timer(&dwork->timer)))
1254 return 1;
1255 }
1256
1257 /* try to claim PENDING the normal way */
1258 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1259 return 0;
1260
1261 rcu_read_lock();
1262 /*
1263 * The queueing is in progress, or it is already queued. Try to
1264 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1265 */
1266 pool = get_work_pool(work);
1267 if (!pool)
1268 goto fail;
1269
1270 raw_spin_lock(&pool->lock);
1271 /*
1272 * work->data is guaranteed to point to pwq only while the work
1273 * item is queued on pwq->wq, and both updating work->data to point
1274 * to pwq on queueing and to pool on dequeueing are done under
1275 * pwq->pool->lock. This in turn guarantees that, if work->data
1276 * points to pwq which is associated with a locked pool, the work
1277 * item is currently queued on that pool.
1278 */
1279 pwq = get_work_pwq(work);
1280 if (pwq && pwq->pool == pool) {
1281 debug_work_deactivate(work);
1282
1283 /*
1284 * A delayed work item cannot be grabbed directly because
1285 * it might have linked NO_COLOR work items which, if left
1286 * on the delayed_list, will confuse pwq->nr_active
1287 * management later on and cause stall. Make sure the work
1288 * item is activated before grabbing.
1289 */
1290 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1291 pwq_activate_delayed_work(work);
1292
1293 list_del_init(&work->entry);
1294 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1295
1296 /* work->data points to pwq iff queued, point to pool */
1297 set_work_pool_and_keep_pending(work, pool->id);
1298
1299 raw_spin_unlock(&pool->lock);
1300 rcu_read_unlock();
1301 return 1;
1302 }
1303 raw_spin_unlock(&pool->lock);
1304fail:
1305 rcu_read_unlock();
1306 local_irq_restore(*flags);
1307 if (work_is_canceling(work))
1308 return -ENOENT;
1309 cpu_relax();
1310 return -EAGAIN;
1311}
1312
1313/**
1314 * insert_work - insert a work into a pool
1315 * @pwq: pwq @work belongs to
1316 * @work: work to insert
1317 * @head: insertion point
1318 * @extra_flags: extra WORK_STRUCT_* flags to set
1319 *
1320 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1321 * work_struct flags.
1322 *
1323 * CONTEXT:
1324 * raw_spin_lock_irq(pool->lock).
1325 */
1326static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1327 struct list_head *head, unsigned int extra_flags)
1328{
1329 struct worker_pool *pool = pwq->pool;
1330
1331 /* record the work call stack in order to print it in KASAN reports */
1332 kasan_record_aux_stack(work);
1333
1334 /* we own @work, set data and link */
1335 set_work_pwq(work, pwq, extra_flags);
1336 list_add_tail(&work->entry, head);
1337 get_pwq(pwq);
1338
1339 /*
1340 * Ensure either wq_worker_sleeping() sees the above
1341 * list_add_tail() or we see zero nr_running to avoid workers lying
1342 * around lazily while there are works to be processed.
1343 */
1344 smp_mb();
1345
1346 if (__need_more_worker(pool))
1347 wake_up_worker(pool);
1348}
1349
1350/*
1351 * Test whether @work is being queued from another work executing on the
1352 * same workqueue.
1353 */
1354static bool is_chained_work(struct workqueue_struct *wq)
1355{
1356 struct worker *worker;
1357
1358 worker = current_wq_worker();
1359 /*
1360 * Return %true iff I'm a worker executing a work item on @wq. If
1361 * I'm @worker, it's safe to dereference it without locking.
1362 */
1363 return worker && worker->current_pwq->wq == wq;
1364}
1365
1366/*
1367 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1368 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1369 * avoid perturbing sensitive tasks.
1370 */
1371static int wq_select_unbound_cpu(int cpu)
1372{
1373 static bool printed_dbg_warning;
1374 int new_cpu;
1375
1376 if (likely(!wq_debug_force_rr_cpu)) {
1377 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1378 return cpu;
1379 } else if (!printed_dbg_warning) {
1380 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1381 printed_dbg_warning = true;
1382 }
1383
1384 if (cpumask_empty(wq_unbound_cpumask))
1385 return cpu;
1386
1387 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1388 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1389 if (unlikely(new_cpu >= nr_cpu_ids)) {
1390 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1391 if (unlikely(new_cpu >= nr_cpu_ids))
1392 return cpu;
1393 }
1394 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1395
1396 return new_cpu;
1397}
1398
1399static void __queue_work(int cpu, struct workqueue_struct *wq,
1400 struct work_struct *work)
1401{
1402 struct pool_workqueue *pwq;
1403 struct worker_pool *last_pool;
1404 struct list_head *worklist;
1405 unsigned int work_flags;
1406 unsigned int req_cpu = cpu;
1407
1408 /*
1409 * While a work item is PENDING && off queue, a task trying to
1410 * steal the PENDING will busy-loop waiting for it to either get
1411 * queued or lose PENDING. Grabbing PENDING and queueing should
1412 * happen with IRQ disabled.
1413 */
1414 lockdep_assert_irqs_disabled();
1415
1416
1417 /* if draining, only works from the same workqueue are allowed */
1418 if (unlikely(wq->flags & __WQ_DRAINING) &&
1419 WARN_ON_ONCE(!is_chained_work(wq)))
1420 return;
1421 rcu_read_lock();
1422retry:
1423 /* pwq which will be used unless @work is executing elsewhere */
1424 if (wq->flags & WQ_UNBOUND) {
1425 if (req_cpu == WORK_CPU_UNBOUND)
1426 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1427 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1428 } else {
1429 if (req_cpu == WORK_CPU_UNBOUND)
1430 cpu = raw_smp_processor_id();
1431 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1432 }
1433
1434 /*
1435 * If @work was previously on a different pool, it might still be
1436 * running there, in which case the work needs to be queued on that
1437 * pool to guarantee non-reentrancy.
1438 */
1439 last_pool = get_work_pool(work);
1440 if (last_pool && last_pool != pwq->pool) {
1441 struct worker *worker;
1442
1443 raw_spin_lock(&last_pool->lock);
1444
1445 worker = find_worker_executing_work(last_pool, work);
1446
1447 if (worker && worker->current_pwq->wq == wq) {
1448 pwq = worker->current_pwq;
1449 } else {
1450 /* meh... not running there, queue here */
1451 raw_spin_unlock(&last_pool->lock);
1452 raw_spin_lock(&pwq->pool->lock);
1453 }
1454 } else {
1455 raw_spin_lock(&pwq->pool->lock);
1456 }
1457
1458 /*
1459 * pwq is determined and locked. For unbound pools, we could have
1460 * raced with pwq release and it could already be dead. If its
1461 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1462 * without another pwq replacing it in the numa_pwq_tbl or while
1463 * work items are executing on it, so the retrying is guaranteed to
1464 * make forward-progress.
1465 */
1466 if (unlikely(!pwq->refcnt)) {
1467 if (wq->flags & WQ_UNBOUND) {
1468 raw_spin_unlock(&pwq->pool->lock);
1469 cpu_relax();
1470 goto retry;
1471 }
1472 /* oops */
1473 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1474 wq->name, cpu);
1475 }
1476
1477 /* pwq determined, queue */
1478 trace_workqueue_queue_work(req_cpu, pwq, work);
1479
1480 if (WARN_ON(!list_empty(&work->entry)))
1481 goto out;
1482
1483 pwq->nr_in_flight[pwq->work_color]++;
1484 work_flags = work_color_to_flags(pwq->work_color);
1485
1486 if (likely(pwq->nr_active < pwq->max_active)) {
1487 trace_workqueue_activate_work(work);
1488 pwq->nr_active++;
1489 worklist = &pwq->pool->worklist;
1490 if (list_empty(worklist))
1491 pwq->pool->watchdog_ts = jiffies;
1492 } else {
1493 work_flags |= WORK_STRUCT_DELAYED;
1494 worklist = &pwq->delayed_works;
1495 }
1496
1497 debug_work_activate(work);
1498 insert_work(pwq, work, worklist, work_flags);
1499
1500out:
1501 raw_spin_unlock(&pwq->pool->lock);
1502 rcu_read_unlock();
1503}
1504
1505/**
1506 * queue_work_on - queue work on specific cpu
1507 * @cpu: CPU number to execute work on
1508 * @wq: workqueue to use
1509 * @work: work to queue
1510 *
1511 * We queue the work to a specific CPU, the caller must ensure it
1512 * can't go away.
1513 *
1514 * Return: %false if @work was already on a queue, %true otherwise.
1515 */
1516bool queue_work_on(int cpu, struct workqueue_struct *wq,
1517 struct work_struct *work)
1518{
1519 bool ret = false;
1520 unsigned long flags;
1521
1522 local_irq_save(flags);
1523
1524 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1525 __queue_work(cpu, wq, work);
1526 ret = true;
1527 }
1528
1529 local_irq_restore(flags);
1530 return ret;
1531}
1532EXPORT_SYMBOL(queue_work_on);
1533
1534/**
1535 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1536 * @node: NUMA node ID that we want to select a CPU from
1537 *
1538 * This function will attempt to find a "random" cpu available on a given
1539 * node. If there are no CPUs available on the given node it will return
1540 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1541 * available CPU if we need to schedule this work.
1542 */
1543static int workqueue_select_cpu_near(int node)
1544{
1545 int cpu;
1546
1547 /* No point in doing this if NUMA isn't enabled for workqueues */
1548 if (!wq_numa_enabled)
1549 return WORK_CPU_UNBOUND;
1550
1551 /* Delay binding to CPU if node is not valid or online */
1552 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1553 return WORK_CPU_UNBOUND;
1554
1555 /* Use local node/cpu if we are already there */
1556 cpu = raw_smp_processor_id();
1557 if (node == cpu_to_node(cpu))
1558 return cpu;
1559
1560 /* Use "random" otherwise know as "first" online CPU of node */
1561 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1562
1563 /* If CPU is valid return that, otherwise just defer */
1564 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1565}
1566
1567/**
1568 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1569 * @node: NUMA node that we are targeting the work for
1570 * @wq: workqueue to use
1571 * @work: work to queue
1572 *
1573 * We queue the work to a "random" CPU within a given NUMA node. The basic
1574 * idea here is to provide a way to somehow associate work with a given
1575 * NUMA node.
1576 *
1577 * This function will only make a best effort attempt at getting this onto
1578 * the right NUMA node. If no node is requested or the requested node is
1579 * offline then we just fall back to standard queue_work behavior.
1580 *
1581 * Currently the "random" CPU ends up being the first available CPU in the
1582 * intersection of cpu_online_mask and the cpumask of the node, unless we
1583 * are running on the node. In that case we just use the current CPU.
1584 *
1585 * Return: %false if @work was already on a queue, %true otherwise.
1586 */
1587bool queue_work_node(int node, struct workqueue_struct *wq,
1588 struct work_struct *work)
1589{
1590 unsigned long flags;
1591 bool ret = false;
1592
1593 /*
1594 * This current implementation is specific to unbound workqueues.
1595 * Specifically we only return the first available CPU for a given
1596 * node instead of cycling through individual CPUs within the node.
1597 *
1598 * If this is used with a per-cpu workqueue then the logic in
1599 * workqueue_select_cpu_near would need to be updated to allow for
1600 * some round robin type logic.
1601 */
1602 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1603
1604 local_irq_save(flags);
1605
1606 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1607 int cpu = workqueue_select_cpu_near(node);
1608
1609 __queue_work(cpu, wq, work);
1610 ret = true;
1611 }
1612
1613 local_irq_restore(flags);
1614 return ret;
1615}
1616EXPORT_SYMBOL_GPL(queue_work_node);
1617
1618void delayed_work_timer_fn(struct timer_list *t)
1619{
1620 struct delayed_work *dwork = from_timer(dwork, t, timer);
1621
1622 /* should have been called from irqsafe timer with irq already off */
1623 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1624}
1625EXPORT_SYMBOL(delayed_work_timer_fn);
1626
1627static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1628 struct delayed_work *dwork, unsigned long delay)
1629{
1630 struct timer_list *timer = &dwork->timer;
1631 struct work_struct *work = &dwork->work;
1632
1633 WARN_ON_ONCE(!wq);
1634 WARN_ON_FUNCTION_MISMATCH(timer->function, delayed_work_timer_fn);
1635 WARN_ON_ONCE(timer_pending(timer));
1636 WARN_ON_ONCE(!list_empty(&work->entry));
1637
1638 /*
1639 * If @delay is 0, queue @dwork->work immediately. This is for
1640 * both optimization and correctness. The earliest @timer can
1641 * expire is on the closest next tick and delayed_work users depend
1642 * on that there's no such delay when @delay is 0.
1643 */
1644 if (!delay) {
1645 __queue_work(cpu, wq, &dwork->work);
1646 return;
1647 }
1648
1649 dwork->wq = wq;
1650 dwork->cpu = cpu;
1651 timer->expires = jiffies + delay;
1652
1653 if (unlikely(cpu != WORK_CPU_UNBOUND))
1654 add_timer_on(timer, cpu);
1655 else
1656 add_timer(timer);
1657}
1658
1659/**
1660 * queue_delayed_work_on - queue work on specific CPU after delay
1661 * @cpu: CPU number to execute work on
1662 * @wq: workqueue to use
1663 * @dwork: work to queue
1664 * @delay: number of jiffies to wait before queueing
1665 *
1666 * Return: %false if @work was already on a queue, %true otherwise. If
1667 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1668 * execution.
1669 */
1670bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1671 struct delayed_work *dwork, unsigned long delay)
1672{
1673 struct work_struct *work = &dwork->work;
1674 bool ret = false;
1675 unsigned long flags;
1676
1677 /* read the comment in __queue_work() */
1678 local_irq_save(flags);
1679
1680 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1681 __queue_delayed_work(cpu, wq, dwork, delay);
1682 ret = true;
1683 }
1684
1685 local_irq_restore(flags);
1686 return ret;
1687}
1688EXPORT_SYMBOL(queue_delayed_work_on);
1689
1690/**
1691 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1692 * @cpu: CPU number to execute work on
1693 * @wq: workqueue to use
1694 * @dwork: work to queue
1695 * @delay: number of jiffies to wait before queueing
1696 *
1697 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1698 * modify @dwork's timer so that it expires after @delay. If @delay is
1699 * zero, @work is guaranteed to be scheduled immediately regardless of its
1700 * current state.
1701 *
1702 * Return: %false if @dwork was idle and queued, %true if @dwork was
1703 * pending and its timer was modified.
1704 *
1705 * This function is safe to call from any context including IRQ handler.
1706 * See try_to_grab_pending() for details.
1707 */
1708bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1709 struct delayed_work *dwork, unsigned long delay)
1710{
1711 unsigned long flags;
1712 int ret;
1713
1714 do {
1715 ret = try_to_grab_pending(&dwork->work, true, &flags);
1716 } while (unlikely(ret == -EAGAIN));
1717
1718 if (likely(ret >= 0)) {
1719 __queue_delayed_work(cpu, wq, dwork, delay);
1720 local_irq_restore(flags);
1721 }
1722
1723 /* -ENOENT from try_to_grab_pending() becomes %true */
1724 return ret;
1725}
1726EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1727
1728static void rcu_work_rcufn(struct rcu_head *rcu)
1729{
1730 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1731
1732 /* read the comment in __queue_work() */
1733 local_irq_disable();
1734 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1735 local_irq_enable();
1736}
1737
1738/**
1739 * queue_rcu_work - queue work after a RCU grace period
1740 * @wq: workqueue to use
1741 * @rwork: work to queue
1742 *
1743 * Return: %false if @rwork was already pending, %true otherwise. Note
1744 * that a full RCU grace period is guaranteed only after a %true return.
1745 * While @rwork is guaranteed to be executed after a %false return, the
1746 * execution may happen before a full RCU grace period has passed.
1747 */
1748bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1749{
1750 struct work_struct *work = &rwork->work;
1751
1752 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1753 rwork->wq = wq;
1754 call_rcu(&rwork->rcu, rcu_work_rcufn);
1755 return true;
1756 }
1757
1758 return false;
1759}
1760EXPORT_SYMBOL(queue_rcu_work);
1761
1762/**
1763 * worker_enter_idle - enter idle state
1764 * @worker: worker which is entering idle state
1765 *
1766 * @worker is entering idle state. Update stats and idle timer if
1767 * necessary.
1768 *
1769 * LOCKING:
1770 * raw_spin_lock_irq(pool->lock).
1771 */
1772static void worker_enter_idle(struct worker *worker)
1773{
1774 struct worker_pool *pool = worker->pool;
1775
1776 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1777 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1778 (worker->hentry.next || worker->hentry.pprev)))
1779 return;
1780
1781 /* can't use worker_set_flags(), also called from create_worker() */
1782 worker->flags |= WORKER_IDLE;
1783 pool->nr_idle++;
1784 worker->last_active = jiffies;
1785
1786 /* idle_list is LIFO */
1787 list_add(&worker->entry, &pool->idle_list);
1788
1789 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1790 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1791
1792 /*
1793 * Sanity check nr_running. Because unbind_workers() releases
1794 * pool->lock between setting %WORKER_UNBOUND and zapping
1795 * nr_running, the warning may trigger spuriously. Check iff
1796 * unbind is not in progress.
1797 */
1798 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1799 pool->nr_workers == pool->nr_idle &&
1800 atomic_read(&pool->nr_running));
1801}
1802
1803/**
1804 * worker_leave_idle - leave idle state
1805 * @worker: worker which is leaving idle state
1806 *
1807 * @worker is leaving idle state. Update stats.
1808 *
1809 * LOCKING:
1810 * raw_spin_lock_irq(pool->lock).
1811 */
1812static void worker_leave_idle(struct worker *worker)
1813{
1814 struct worker_pool *pool = worker->pool;
1815
1816 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1817 return;
1818 worker_clr_flags(worker, WORKER_IDLE);
1819 pool->nr_idle--;
1820 list_del_init(&worker->entry);
1821}
1822
1823static struct worker *alloc_worker(int node)
1824{
1825 struct worker *worker;
1826
1827 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1828 if (worker) {
1829 INIT_LIST_HEAD(&worker->entry);
1830 INIT_LIST_HEAD(&worker->scheduled);
1831 INIT_LIST_HEAD(&worker->node);
1832 /* on creation a worker is in !idle && prep state */
1833 worker->flags = WORKER_PREP;
1834 }
1835 return worker;
1836}
1837
1838/**
1839 * worker_attach_to_pool() - attach a worker to a pool
1840 * @worker: worker to be attached
1841 * @pool: the target pool
1842 *
1843 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1844 * cpu-binding of @worker are kept coordinated with the pool across
1845 * cpu-[un]hotplugs.
1846 */
1847static void worker_attach_to_pool(struct worker *worker,
1848 struct worker_pool *pool)
1849{
1850 mutex_lock(&wq_pool_attach_mutex);
1851
1852 /*
1853 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1854 * stable across this function. See the comments above the flag
1855 * definition for details.
1856 */
1857 if (pool->flags & POOL_DISASSOCIATED)
1858 worker->flags |= WORKER_UNBOUND;
1859 else
1860 kthread_set_per_cpu(worker->task, pool->cpu);
1861
1862 if (worker->rescue_wq)
1863 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1864
1865 list_add_tail(&worker->node, &pool->workers);
1866 worker->pool = pool;
1867
1868 mutex_unlock(&wq_pool_attach_mutex);
1869}
1870
1871/**
1872 * worker_detach_from_pool() - detach a worker from its pool
1873 * @worker: worker which is attached to its pool
1874 *
1875 * Undo the attaching which had been done in worker_attach_to_pool(). The
1876 * caller worker shouldn't access to the pool after detached except it has
1877 * other reference to the pool.
1878 */
1879static void worker_detach_from_pool(struct worker *worker)
1880{
1881 struct worker_pool *pool = worker->pool;
1882 struct completion *detach_completion = NULL;
1883
1884 mutex_lock(&wq_pool_attach_mutex);
1885
1886 kthread_set_per_cpu(worker->task, -1);
1887 list_del(&worker->node);
1888 worker->pool = NULL;
1889
1890 if (list_empty(&pool->workers))
1891 detach_completion = pool->detach_completion;
1892 mutex_unlock(&wq_pool_attach_mutex);
1893
1894 /* clear leftover flags without pool->lock after it is detached */
1895 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1896
1897 if (detach_completion)
1898 complete(detach_completion);
1899}
1900
1901/**
1902 * create_worker - create a new workqueue worker
1903 * @pool: pool the new worker will belong to
1904 *
1905 * Create and start a new worker which is attached to @pool.
1906 *
1907 * CONTEXT:
1908 * Might sleep. Does GFP_KERNEL allocations.
1909 *
1910 * Return:
1911 * Pointer to the newly created worker.
1912 */
1913static struct worker *create_worker(struct worker_pool *pool)
1914{
1915 struct worker *worker = NULL;
1916 int id = -1;
1917 char id_buf[16];
1918
1919 /* ID is needed to determine kthread name */
1920 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1921 if (id < 0)
1922 goto fail;
1923
1924 worker = alloc_worker(pool->node);
1925 if (!worker)
1926 goto fail;
1927
1928 worker->id = id;
1929
1930 if (pool->cpu >= 0)
1931 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1932 pool->attrs->nice < 0 ? "H" : "");
1933 else
1934 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1935
1936 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1937 "kworker/%s", id_buf);
1938 if (IS_ERR(worker->task))
1939 goto fail;
1940
1941 set_user_nice(worker->task, pool->attrs->nice);
1942 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1943
1944 /* successful, attach the worker to the pool */
1945 worker_attach_to_pool(worker, pool);
1946
1947 /* start the newly created worker */
1948 raw_spin_lock_irq(&pool->lock);
1949 worker->pool->nr_workers++;
1950 worker_enter_idle(worker);
1951 wake_up_process(worker->task);
1952 raw_spin_unlock_irq(&pool->lock);
1953
1954 return worker;
1955
1956fail:
1957 if (id >= 0)
1958 ida_simple_remove(&pool->worker_ida, id);
1959 kfree(worker);
1960 return NULL;
1961}
1962
1963/**
1964 * destroy_worker - destroy a workqueue worker
1965 * @worker: worker to be destroyed
1966 *
1967 * Destroy @worker and adjust @pool stats accordingly. The worker should
1968 * be idle.
1969 *
1970 * CONTEXT:
1971 * raw_spin_lock_irq(pool->lock).
1972 */
1973static void destroy_worker(struct worker *worker)
1974{
1975 struct worker_pool *pool = worker->pool;
1976
1977 lockdep_assert_held(&pool->lock);
1978
1979 /* sanity check frenzy */
1980 if (WARN_ON(worker->current_work) ||
1981 WARN_ON(!list_empty(&worker->scheduled)) ||
1982 WARN_ON(!(worker->flags & WORKER_IDLE)))
1983 return;
1984
1985 pool->nr_workers--;
1986 pool->nr_idle--;
1987
1988 list_del_init(&worker->entry);
1989 worker->flags |= WORKER_DIE;
1990 wake_up_process(worker->task);
1991}
1992
1993static void idle_worker_timeout(struct timer_list *t)
1994{
1995 struct worker_pool *pool = from_timer(pool, t, idle_timer);
1996
1997 raw_spin_lock_irq(&pool->lock);
1998
1999 while (too_many_workers(pool)) {
2000 struct worker *worker;
2001 unsigned long expires;
2002
2003 /* idle_list is kept in LIFO order, check the last one */
2004 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2005 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2006
2007 if (time_before(jiffies, expires)) {
2008 mod_timer(&pool->idle_timer, expires);
2009 break;
2010 }
2011
2012 destroy_worker(worker);
2013 }
2014
2015 raw_spin_unlock_irq(&pool->lock);
2016}
2017
2018static void send_mayday(struct work_struct *work)
2019{
2020 struct pool_workqueue *pwq = get_work_pwq(work);
2021 struct workqueue_struct *wq = pwq->wq;
2022
2023 lockdep_assert_held(&wq_mayday_lock);
2024
2025 if (!wq->rescuer)
2026 return;
2027
2028 /* mayday mayday mayday */
2029 if (list_empty(&pwq->mayday_node)) {
2030 /*
2031 * If @pwq is for an unbound wq, its base ref may be put at
2032 * any time due to an attribute change. Pin @pwq until the
2033 * rescuer is done with it.
2034 */
2035 get_pwq(pwq);
2036 list_add_tail(&pwq->mayday_node, &wq->maydays);
2037 wake_up_process(wq->rescuer->task);
2038 }
2039}
2040
2041static void pool_mayday_timeout(struct timer_list *t)
2042{
2043 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2044 struct work_struct *work;
2045
2046 raw_spin_lock_irq(&pool->lock);
2047 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2048
2049 if (need_to_create_worker(pool)) {
2050 /*
2051 * We've been trying to create a new worker but
2052 * haven't been successful. We might be hitting an
2053 * allocation deadlock. Send distress signals to
2054 * rescuers.
2055 */
2056 list_for_each_entry(work, &pool->worklist, entry)
2057 send_mayday(work);
2058 }
2059
2060 raw_spin_unlock(&wq_mayday_lock);
2061 raw_spin_unlock_irq(&pool->lock);
2062
2063 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2064}
2065
2066/**
2067 * maybe_create_worker - create a new worker if necessary
2068 * @pool: pool to create a new worker for
2069 *
2070 * Create a new worker for @pool if necessary. @pool is guaranteed to
2071 * have at least one idle worker on return from this function. If
2072 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2073 * sent to all rescuers with works scheduled on @pool to resolve
2074 * possible allocation deadlock.
2075 *
2076 * On return, need_to_create_worker() is guaranteed to be %false and
2077 * may_start_working() %true.
2078 *
2079 * LOCKING:
2080 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2081 * multiple times. Does GFP_KERNEL allocations. Called only from
2082 * manager.
2083 */
2084static void maybe_create_worker(struct worker_pool *pool)
2085__releases(&pool->lock)
2086__acquires(&pool->lock)
2087{
2088restart:
2089 raw_spin_unlock_irq(&pool->lock);
2090
2091 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2092 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2093
2094 while (true) {
2095 if (create_worker(pool) || !need_to_create_worker(pool))
2096 break;
2097
2098 schedule_timeout_interruptible(CREATE_COOLDOWN);
2099
2100 if (!need_to_create_worker(pool))
2101 break;
2102 }
2103
2104 del_timer_sync(&pool->mayday_timer);
2105 raw_spin_lock_irq(&pool->lock);
2106 /*
2107 * This is necessary even after a new worker was just successfully
2108 * created as @pool->lock was dropped and the new worker might have
2109 * already become busy.
2110 */
2111 if (need_to_create_worker(pool))
2112 goto restart;
2113}
2114
2115/**
2116 * manage_workers - manage worker pool
2117 * @worker: self
2118 *
2119 * Assume the manager role and manage the worker pool @worker belongs
2120 * to. At any given time, there can be only zero or one manager per
2121 * pool. The exclusion is handled automatically by this function.
2122 *
2123 * The caller can safely start processing works on false return. On
2124 * true return, it's guaranteed that need_to_create_worker() is false
2125 * and may_start_working() is true.
2126 *
2127 * CONTEXT:
2128 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2129 * multiple times. Does GFP_KERNEL allocations.
2130 *
2131 * Return:
2132 * %false if the pool doesn't need management and the caller can safely
2133 * start processing works, %true if management function was performed and
2134 * the conditions that the caller verified before calling the function may
2135 * no longer be true.
2136 */
2137static bool manage_workers(struct worker *worker)
2138{
2139 struct worker_pool *pool = worker->pool;
2140
2141 if (pool->flags & POOL_MANAGER_ACTIVE)
2142 return false;
2143
2144 pool->flags |= POOL_MANAGER_ACTIVE;
2145 pool->manager = worker;
2146
2147 maybe_create_worker(pool);
2148
2149 pool->manager = NULL;
2150 pool->flags &= ~POOL_MANAGER_ACTIVE;
2151 rcuwait_wake_up(&manager_wait);
2152 return true;
2153}
2154
2155/**
2156 * process_one_work - process single work
2157 * @worker: self
2158 * @work: work to process
2159 *
2160 * Process @work. This function contains all the logics necessary to
2161 * process a single work including synchronization against and
2162 * interaction with other workers on the same cpu, queueing and
2163 * flushing. As long as context requirement is met, any worker can
2164 * call this function to process a work.
2165 *
2166 * CONTEXT:
2167 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2168 */
2169static void process_one_work(struct worker *worker, struct work_struct *work)
2170__releases(&pool->lock)
2171__acquires(&pool->lock)
2172{
2173 struct pool_workqueue *pwq = get_work_pwq(work);
2174 struct worker_pool *pool = worker->pool;
2175 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2176 int work_color;
2177 struct worker *collision;
2178#ifdef CONFIG_LOCKDEP
2179 /*
2180 * It is permissible to free the struct work_struct from
2181 * inside the function that is called from it, this we need to
2182 * take into account for lockdep too. To avoid bogus "held
2183 * lock freed" warnings as well as problems when looking into
2184 * work->lockdep_map, make a copy and use that here.
2185 */
2186 struct lockdep_map lockdep_map;
2187
2188 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2189#endif
2190 /* ensure we're on the correct CPU */
2191 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2192 raw_smp_processor_id() != pool->cpu);
2193
2194 /*
2195 * A single work shouldn't be executed concurrently by
2196 * multiple workers on a single cpu. Check whether anyone is
2197 * already processing the work. If so, defer the work to the
2198 * currently executing one.
2199 */
2200 collision = find_worker_executing_work(pool, work);
2201 if (unlikely(collision)) {
2202 move_linked_works(work, &collision->scheduled, NULL);
2203 return;
2204 }
2205
2206 /* claim and dequeue */
2207 debug_work_deactivate(work);
2208 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2209 worker->current_work = work;
2210 worker->current_func = work->func;
2211 worker->current_pwq = pwq;
2212 work_color = get_work_color(work);
2213
2214 /*
2215 * Record wq name for cmdline and debug reporting, may get
2216 * overridden through set_worker_desc().
2217 */
2218 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2219
2220 list_del_init(&work->entry);
2221
2222 /*
2223 * CPU intensive works don't participate in concurrency management.
2224 * They're the scheduler's responsibility. This takes @worker out
2225 * of concurrency management and the next code block will chain
2226 * execution of the pending work items.
2227 */
2228 if (unlikely(cpu_intensive))
2229 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2230
2231 /*
2232 * Wake up another worker if necessary. The condition is always
2233 * false for normal per-cpu workers since nr_running would always
2234 * be >= 1 at this point. This is used to chain execution of the
2235 * pending work items for WORKER_NOT_RUNNING workers such as the
2236 * UNBOUND and CPU_INTENSIVE ones.
2237 */
2238 if (need_more_worker(pool))
2239 wake_up_worker(pool);
2240
2241 /*
2242 * Record the last pool and clear PENDING which should be the last
2243 * update to @work. Also, do this inside @pool->lock so that
2244 * PENDING and queued state changes happen together while IRQ is
2245 * disabled.
2246 */
2247 set_work_pool_and_clear_pending(work, pool->id);
2248
2249 raw_spin_unlock_irq(&pool->lock);
2250
2251 lock_map_acquire(&pwq->wq->lockdep_map);
2252 lock_map_acquire(&lockdep_map);
2253 /*
2254 * Strictly speaking we should mark the invariant state without holding
2255 * any locks, that is, before these two lock_map_acquire()'s.
2256 *
2257 * However, that would result in:
2258 *
2259 * A(W1)
2260 * WFC(C)
2261 * A(W1)
2262 * C(C)
2263 *
2264 * Which would create W1->C->W1 dependencies, even though there is no
2265 * actual deadlock possible. There are two solutions, using a
2266 * read-recursive acquire on the work(queue) 'locks', but this will then
2267 * hit the lockdep limitation on recursive locks, or simply discard
2268 * these locks.
2269 *
2270 * AFAICT there is no possible deadlock scenario between the
2271 * flush_work() and complete() primitives (except for single-threaded
2272 * workqueues), so hiding them isn't a problem.
2273 */
2274 lockdep_invariant_state(true);
2275 trace_workqueue_execute_start(work);
2276 worker->current_func(work);
2277 /*
2278 * While we must be careful to not use "work" after this, the trace
2279 * point will only record its address.
2280 */
2281 trace_workqueue_execute_end(work, worker->current_func);
2282 lock_map_release(&lockdep_map);
2283 lock_map_release(&pwq->wq->lockdep_map);
2284
2285 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2286 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2287 " last function: %ps\n",
2288 current->comm, preempt_count(), task_pid_nr(current),
2289 worker->current_func);
2290 debug_show_held_locks(current);
2291 dump_stack();
2292 }
2293
2294 /*
2295 * The following prevents a kworker from hogging CPU on !PREEMPTION
2296 * kernels, where a requeueing work item waiting for something to
2297 * happen could deadlock with stop_machine as such work item could
2298 * indefinitely requeue itself while all other CPUs are trapped in
2299 * stop_machine. At the same time, report a quiescent RCU state so
2300 * the same condition doesn't freeze RCU.
2301 */
2302 cond_resched();
2303
2304 raw_spin_lock_irq(&pool->lock);
2305
2306 /* clear cpu intensive status */
2307 if (unlikely(cpu_intensive))
2308 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2309
2310 /* tag the worker for identification in schedule() */
2311 worker->last_func = worker->current_func;
2312
2313 /* we're done with it, release */
2314 hash_del(&worker->hentry);
2315 worker->current_work = NULL;
2316 worker->current_func = NULL;
2317 worker->current_pwq = NULL;
2318 pwq_dec_nr_in_flight(pwq, work_color);
2319}
2320
2321/**
2322 * process_scheduled_works - process scheduled works
2323 * @worker: self
2324 *
2325 * Process all scheduled works. Please note that the scheduled list
2326 * may change while processing a work, so this function repeatedly
2327 * fetches a work from the top and executes it.
2328 *
2329 * CONTEXT:
2330 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2331 * multiple times.
2332 */
2333static void process_scheduled_works(struct worker *worker)
2334{
2335 while (!list_empty(&worker->scheduled)) {
2336 struct work_struct *work = list_first_entry(&worker->scheduled,
2337 struct work_struct, entry);
2338 process_one_work(worker, work);
2339 }
2340}
2341
2342static void set_pf_worker(bool val)
2343{
2344 mutex_lock(&wq_pool_attach_mutex);
2345 if (val)
2346 current->flags |= PF_WQ_WORKER;
2347 else
2348 current->flags &= ~PF_WQ_WORKER;
2349 mutex_unlock(&wq_pool_attach_mutex);
2350}
2351
2352/**
2353 * worker_thread - the worker thread function
2354 * @__worker: self
2355 *
2356 * The worker thread function. All workers belong to a worker_pool -
2357 * either a per-cpu one or dynamic unbound one. These workers process all
2358 * work items regardless of their specific target workqueue. The only
2359 * exception is work items which belong to workqueues with a rescuer which
2360 * will be explained in rescuer_thread().
2361 *
2362 * Return: 0
2363 */
2364static int worker_thread(void *__worker)
2365{
2366 struct worker *worker = __worker;
2367 struct worker_pool *pool = worker->pool;
2368
2369 /* tell the scheduler that this is a workqueue worker */
2370 set_pf_worker(true);
2371woke_up:
2372 raw_spin_lock_irq(&pool->lock);
2373
2374 /* am I supposed to die? */
2375 if (unlikely(worker->flags & WORKER_DIE)) {
2376 raw_spin_unlock_irq(&pool->lock);
2377 WARN_ON_ONCE(!list_empty(&worker->entry));
2378 set_pf_worker(false);
2379
2380 set_task_comm(worker->task, "kworker/dying");
2381 ida_simple_remove(&pool->worker_ida, worker->id);
2382 worker_detach_from_pool(worker);
2383 kfree(worker);
2384 return 0;
2385 }
2386
2387 worker_leave_idle(worker);
2388recheck:
2389 /* no more worker necessary? */
2390 if (!need_more_worker(pool))
2391 goto sleep;
2392
2393 /* do we need to manage? */
2394 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2395 goto recheck;
2396
2397 /*
2398 * ->scheduled list can only be filled while a worker is
2399 * preparing to process a work or actually processing it.
2400 * Make sure nobody diddled with it while I was sleeping.
2401 */
2402 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2403
2404 /*
2405 * Finish PREP stage. We're guaranteed to have at least one idle
2406 * worker or that someone else has already assumed the manager
2407 * role. This is where @worker starts participating in concurrency
2408 * management if applicable and concurrency management is restored
2409 * after being rebound. See rebind_workers() for details.
2410 */
2411 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2412
2413 do {
2414 struct work_struct *work =
2415 list_first_entry(&pool->worklist,
2416 struct work_struct, entry);
2417
2418 pool->watchdog_ts = jiffies;
2419
2420 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2421 /* optimization path, not strictly necessary */
2422 process_one_work(worker, work);
2423 if (unlikely(!list_empty(&worker->scheduled)))
2424 process_scheduled_works(worker);
2425 } else {
2426 move_linked_works(work, &worker->scheduled, NULL);
2427 process_scheduled_works(worker);
2428 }
2429 } while (keep_working(pool));
2430
2431 worker_set_flags(worker, WORKER_PREP);
2432sleep:
2433 /*
2434 * pool->lock is held and there's no work to process and no need to
2435 * manage, sleep. Workers are woken up only while holding
2436 * pool->lock or from local cpu, so setting the current state
2437 * before releasing pool->lock is enough to prevent losing any
2438 * event.
2439 */
2440 worker_enter_idle(worker);
2441 __set_current_state(TASK_IDLE);
2442 raw_spin_unlock_irq(&pool->lock);
2443 schedule();
2444 goto woke_up;
2445}
2446
2447/**
2448 * rescuer_thread - the rescuer thread function
2449 * @__rescuer: self
2450 *
2451 * Workqueue rescuer thread function. There's one rescuer for each
2452 * workqueue which has WQ_MEM_RECLAIM set.
2453 *
2454 * Regular work processing on a pool may block trying to create a new
2455 * worker which uses GFP_KERNEL allocation which has slight chance of
2456 * developing into deadlock if some works currently on the same queue
2457 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2458 * the problem rescuer solves.
2459 *
2460 * When such condition is possible, the pool summons rescuers of all
2461 * workqueues which have works queued on the pool and let them process
2462 * those works so that forward progress can be guaranteed.
2463 *
2464 * This should happen rarely.
2465 *
2466 * Return: 0
2467 */
2468static int rescuer_thread(void *__rescuer)
2469{
2470 struct worker *rescuer = __rescuer;
2471 struct workqueue_struct *wq = rescuer->rescue_wq;
2472 struct list_head *scheduled = &rescuer->scheduled;
2473 bool should_stop;
2474
2475 set_user_nice(current, RESCUER_NICE_LEVEL);
2476
2477 /*
2478 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2479 * doesn't participate in concurrency management.
2480 */
2481 set_pf_worker(true);
2482repeat:
2483 set_current_state(TASK_IDLE);
2484
2485 /*
2486 * By the time the rescuer is requested to stop, the workqueue
2487 * shouldn't have any work pending, but @wq->maydays may still have
2488 * pwq(s) queued. This can happen by non-rescuer workers consuming
2489 * all the work items before the rescuer got to them. Go through
2490 * @wq->maydays processing before acting on should_stop so that the
2491 * list is always empty on exit.
2492 */
2493 should_stop = kthread_should_stop();
2494
2495 /* see whether any pwq is asking for help */
2496 raw_spin_lock_irq(&wq_mayday_lock);
2497
2498 while (!list_empty(&wq->maydays)) {
2499 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2500 struct pool_workqueue, mayday_node);
2501 struct worker_pool *pool = pwq->pool;
2502 struct work_struct *work, *n;
2503 bool first = true;
2504
2505 __set_current_state(TASK_RUNNING);
2506 list_del_init(&pwq->mayday_node);
2507
2508 raw_spin_unlock_irq(&wq_mayday_lock);
2509
2510 worker_attach_to_pool(rescuer, pool);
2511
2512 raw_spin_lock_irq(&pool->lock);
2513
2514 /*
2515 * Slurp in all works issued via this workqueue and
2516 * process'em.
2517 */
2518 WARN_ON_ONCE(!list_empty(scheduled));
2519 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2520 if (get_work_pwq(work) == pwq) {
2521 if (first)
2522 pool->watchdog_ts = jiffies;
2523 move_linked_works(work, scheduled, &n);
2524 }
2525 first = false;
2526 }
2527
2528 if (!list_empty(scheduled)) {
2529 process_scheduled_works(rescuer);
2530
2531 /*
2532 * The above execution of rescued work items could
2533 * have created more to rescue through
2534 * pwq_activate_first_delayed() or chained
2535 * queueing. Let's put @pwq back on mayday list so
2536 * that such back-to-back work items, which may be
2537 * being used to relieve memory pressure, don't
2538 * incur MAYDAY_INTERVAL delay inbetween.
2539 */
2540 if (pwq->nr_active && need_to_create_worker(pool)) {
2541 raw_spin_lock(&wq_mayday_lock);
2542 /*
2543 * Queue iff we aren't racing destruction
2544 * and somebody else hasn't queued it already.
2545 */
2546 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2547 get_pwq(pwq);
2548 list_add_tail(&pwq->mayday_node, &wq->maydays);
2549 }
2550 raw_spin_unlock(&wq_mayday_lock);
2551 }
2552 }
2553
2554 /*
2555 * Put the reference grabbed by send_mayday(). @pool won't
2556 * go away while we're still attached to it.
2557 */
2558 put_pwq(pwq);
2559
2560 /*
2561 * Leave this pool. If need_more_worker() is %true, notify a
2562 * regular worker; otherwise, we end up with 0 concurrency
2563 * and stalling the execution.
2564 */
2565 if (need_more_worker(pool))
2566 wake_up_worker(pool);
2567
2568 raw_spin_unlock_irq(&pool->lock);
2569
2570 worker_detach_from_pool(rescuer);
2571
2572 raw_spin_lock_irq(&wq_mayday_lock);
2573 }
2574
2575 raw_spin_unlock_irq(&wq_mayday_lock);
2576
2577 if (should_stop) {
2578 __set_current_state(TASK_RUNNING);
2579 set_pf_worker(false);
2580 return 0;
2581 }
2582
2583 /* rescuers should never participate in concurrency management */
2584 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2585 schedule();
2586 goto repeat;
2587}
2588
2589/**
2590 * check_flush_dependency - check for flush dependency sanity
2591 * @target_wq: workqueue being flushed
2592 * @target_work: work item being flushed (NULL for workqueue flushes)
2593 *
2594 * %current is trying to flush the whole @target_wq or @target_work on it.
2595 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2596 * reclaiming memory or running on a workqueue which doesn't have
2597 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2598 * a deadlock.
2599 */
2600static void check_flush_dependency(struct workqueue_struct *target_wq,
2601 struct work_struct *target_work)
2602{
2603 work_func_t target_func = target_work ? target_work->func : NULL;
2604 struct worker *worker;
2605
2606 if (target_wq->flags & WQ_MEM_RECLAIM)
2607 return;
2608
2609 worker = current_wq_worker();
2610
2611 WARN_ONCE(current->flags & PF_MEMALLOC,
2612 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2613 current->pid, current->comm, target_wq->name, target_func);
2614 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2615 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2616 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2617 worker->current_pwq->wq->name, worker->current_func,
2618 target_wq->name, target_func);
2619}
2620
2621struct wq_barrier {
2622 struct work_struct work;
2623 struct completion done;
2624 struct task_struct *task; /* purely informational */
2625};
2626
2627static void wq_barrier_func(struct work_struct *work)
2628{
2629 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2630 complete(&barr->done);
2631}
2632
2633/**
2634 * insert_wq_barrier - insert a barrier work
2635 * @pwq: pwq to insert barrier into
2636 * @barr: wq_barrier to insert
2637 * @target: target work to attach @barr to
2638 * @worker: worker currently executing @target, NULL if @target is not executing
2639 *
2640 * @barr is linked to @target such that @barr is completed only after
2641 * @target finishes execution. Please note that the ordering
2642 * guarantee is observed only with respect to @target and on the local
2643 * cpu.
2644 *
2645 * Currently, a queued barrier can't be canceled. This is because
2646 * try_to_grab_pending() can't determine whether the work to be
2647 * grabbed is at the head of the queue and thus can't clear LINKED
2648 * flag of the previous work while there must be a valid next work
2649 * after a work with LINKED flag set.
2650 *
2651 * Note that when @worker is non-NULL, @target may be modified
2652 * underneath us, so we can't reliably determine pwq from @target.
2653 *
2654 * CONTEXT:
2655 * raw_spin_lock_irq(pool->lock).
2656 */
2657static void insert_wq_barrier(struct pool_workqueue *pwq,
2658 struct wq_barrier *barr,
2659 struct work_struct *target, struct worker *worker)
2660{
2661 struct list_head *head;
2662 unsigned int linked = 0;
2663
2664 /*
2665 * debugobject calls are safe here even with pool->lock locked
2666 * as we know for sure that this will not trigger any of the
2667 * checks and call back into the fixup functions where we
2668 * might deadlock.
2669 */
2670 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2671 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2672
2673 init_completion_map(&barr->done, &target->lockdep_map);
2674
2675 barr->task = current;
2676
2677 /*
2678 * If @target is currently being executed, schedule the
2679 * barrier to the worker; otherwise, put it after @target.
2680 */
2681 if (worker)
2682 head = worker->scheduled.next;
2683 else {
2684 unsigned long *bits = work_data_bits(target);
2685
2686 head = target->entry.next;
2687 /* there can already be other linked works, inherit and set */
2688 linked = *bits & WORK_STRUCT_LINKED;
2689 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2690 }
2691
2692 debug_work_activate(&barr->work);
2693 insert_work(pwq, &barr->work, head,
2694 work_color_to_flags(WORK_NO_COLOR) | linked);
2695}
2696
2697/**
2698 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2699 * @wq: workqueue being flushed
2700 * @flush_color: new flush color, < 0 for no-op
2701 * @work_color: new work color, < 0 for no-op
2702 *
2703 * Prepare pwqs for workqueue flushing.
2704 *
2705 * If @flush_color is non-negative, flush_color on all pwqs should be
2706 * -1. If no pwq has in-flight commands at the specified color, all
2707 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2708 * has in flight commands, its pwq->flush_color is set to
2709 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2710 * wakeup logic is armed and %true is returned.
2711 *
2712 * The caller should have initialized @wq->first_flusher prior to
2713 * calling this function with non-negative @flush_color. If
2714 * @flush_color is negative, no flush color update is done and %false
2715 * is returned.
2716 *
2717 * If @work_color is non-negative, all pwqs should have the same
2718 * work_color which is previous to @work_color and all will be
2719 * advanced to @work_color.
2720 *
2721 * CONTEXT:
2722 * mutex_lock(wq->mutex).
2723 *
2724 * Return:
2725 * %true if @flush_color >= 0 and there's something to flush. %false
2726 * otherwise.
2727 */
2728static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2729 int flush_color, int work_color)
2730{
2731 bool wait = false;
2732 struct pool_workqueue *pwq;
2733
2734 if (flush_color >= 0) {
2735 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2736 atomic_set(&wq->nr_pwqs_to_flush, 1);
2737 }
2738
2739 for_each_pwq(pwq, wq) {
2740 struct worker_pool *pool = pwq->pool;
2741
2742 raw_spin_lock_irq(&pool->lock);
2743
2744 if (flush_color >= 0) {
2745 WARN_ON_ONCE(pwq->flush_color != -1);
2746
2747 if (pwq->nr_in_flight[flush_color]) {
2748 pwq->flush_color = flush_color;
2749 atomic_inc(&wq->nr_pwqs_to_flush);
2750 wait = true;
2751 }
2752 }
2753
2754 if (work_color >= 0) {
2755 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2756 pwq->work_color = work_color;
2757 }
2758
2759 raw_spin_unlock_irq(&pool->lock);
2760 }
2761
2762 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2763 complete(&wq->first_flusher->done);
2764
2765 return wait;
2766}
2767
2768/**
2769 * flush_workqueue - ensure that any scheduled work has run to completion.
2770 * @wq: workqueue to flush
2771 *
2772 * This function sleeps until all work items which were queued on entry
2773 * have finished execution, but it is not livelocked by new incoming ones.
2774 */
2775void flush_workqueue(struct workqueue_struct *wq)
2776{
2777 struct wq_flusher this_flusher = {
2778 .list = LIST_HEAD_INIT(this_flusher.list),
2779 .flush_color = -1,
2780 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2781 };
2782 int next_color;
2783
2784 if (WARN_ON(!wq_online))
2785 return;
2786
2787 lock_map_acquire(&wq->lockdep_map);
2788 lock_map_release(&wq->lockdep_map);
2789
2790 mutex_lock(&wq->mutex);
2791
2792 /*
2793 * Start-to-wait phase
2794 */
2795 next_color = work_next_color(wq->work_color);
2796
2797 if (next_color != wq->flush_color) {
2798 /*
2799 * Color space is not full. The current work_color
2800 * becomes our flush_color and work_color is advanced
2801 * by one.
2802 */
2803 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2804 this_flusher.flush_color = wq->work_color;
2805 wq->work_color = next_color;
2806
2807 if (!wq->first_flusher) {
2808 /* no flush in progress, become the first flusher */
2809 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2810
2811 wq->first_flusher = &this_flusher;
2812
2813 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2814 wq->work_color)) {
2815 /* nothing to flush, done */
2816 wq->flush_color = next_color;
2817 wq->first_flusher = NULL;
2818 goto out_unlock;
2819 }
2820 } else {
2821 /* wait in queue */
2822 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2823 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2824 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2825 }
2826 } else {
2827 /*
2828 * Oops, color space is full, wait on overflow queue.
2829 * The next flush completion will assign us
2830 * flush_color and transfer to flusher_queue.
2831 */
2832 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2833 }
2834
2835 check_flush_dependency(wq, NULL);
2836
2837 mutex_unlock(&wq->mutex);
2838
2839 wait_for_completion(&this_flusher.done);
2840
2841 /*
2842 * Wake-up-and-cascade phase
2843 *
2844 * First flushers are responsible for cascading flushes and
2845 * handling overflow. Non-first flushers can simply return.
2846 */
2847 if (READ_ONCE(wq->first_flusher) != &this_flusher)
2848 return;
2849
2850 mutex_lock(&wq->mutex);
2851
2852 /* we might have raced, check again with mutex held */
2853 if (wq->first_flusher != &this_flusher)
2854 goto out_unlock;
2855
2856 WRITE_ONCE(wq->first_flusher, NULL);
2857
2858 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2859 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2860
2861 while (true) {
2862 struct wq_flusher *next, *tmp;
2863
2864 /* complete all the flushers sharing the current flush color */
2865 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2866 if (next->flush_color != wq->flush_color)
2867 break;
2868 list_del_init(&next->list);
2869 complete(&next->done);
2870 }
2871
2872 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2873 wq->flush_color != work_next_color(wq->work_color));
2874
2875 /* this flush_color is finished, advance by one */
2876 wq->flush_color = work_next_color(wq->flush_color);
2877
2878 /* one color has been freed, handle overflow queue */
2879 if (!list_empty(&wq->flusher_overflow)) {
2880 /*
2881 * Assign the same color to all overflowed
2882 * flushers, advance work_color and append to
2883 * flusher_queue. This is the start-to-wait
2884 * phase for these overflowed flushers.
2885 */
2886 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2887 tmp->flush_color = wq->work_color;
2888
2889 wq->work_color = work_next_color(wq->work_color);
2890
2891 list_splice_tail_init(&wq->flusher_overflow,
2892 &wq->flusher_queue);
2893 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2894 }
2895
2896 if (list_empty(&wq->flusher_queue)) {
2897 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2898 break;
2899 }
2900
2901 /*
2902 * Need to flush more colors. Make the next flusher
2903 * the new first flusher and arm pwqs.
2904 */
2905 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2906 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2907
2908 list_del_init(&next->list);
2909 wq->first_flusher = next;
2910
2911 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2912 break;
2913
2914 /*
2915 * Meh... this color is already done, clear first
2916 * flusher and repeat cascading.
2917 */
2918 wq->first_flusher = NULL;
2919 }
2920
2921out_unlock:
2922 mutex_unlock(&wq->mutex);
2923}
2924EXPORT_SYMBOL(flush_workqueue);
2925
2926/**
2927 * drain_workqueue - drain a workqueue
2928 * @wq: workqueue to drain
2929 *
2930 * Wait until the workqueue becomes empty. While draining is in progress,
2931 * only chain queueing is allowed. IOW, only currently pending or running
2932 * work items on @wq can queue further work items on it. @wq is flushed
2933 * repeatedly until it becomes empty. The number of flushing is determined
2934 * by the depth of chaining and should be relatively short. Whine if it
2935 * takes too long.
2936 */
2937void drain_workqueue(struct workqueue_struct *wq)
2938{
2939 unsigned int flush_cnt = 0;
2940 struct pool_workqueue *pwq;
2941
2942 /*
2943 * __queue_work() needs to test whether there are drainers, is much
2944 * hotter than drain_workqueue() and already looks at @wq->flags.
2945 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2946 */
2947 mutex_lock(&wq->mutex);
2948 if (!wq->nr_drainers++)
2949 wq->flags |= __WQ_DRAINING;
2950 mutex_unlock(&wq->mutex);
2951reflush:
2952 flush_workqueue(wq);
2953
2954 mutex_lock(&wq->mutex);
2955
2956 for_each_pwq(pwq, wq) {
2957 bool drained;
2958
2959 raw_spin_lock_irq(&pwq->pool->lock);
2960 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2961 raw_spin_unlock_irq(&pwq->pool->lock);
2962
2963 if (drained)
2964 continue;
2965
2966 if (++flush_cnt == 10 ||
2967 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2968 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
2969 wq->name, __func__, flush_cnt);
2970
2971 mutex_unlock(&wq->mutex);
2972 goto reflush;
2973 }
2974
2975 if (!--wq->nr_drainers)
2976 wq->flags &= ~__WQ_DRAINING;
2977 mutex_unlock(&wq->mutex);
2978}
2979EXPORT_SYMBOL_GPL(drain_workqueue);
2980
2981static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2982 bool from_cancel)
2983{
2984 struct worker *worker = NULL;
2985 struct worker_pool *pool;
2986 struct pool_workqueue *pwq;
2987
2988 might_sleep();
2989
2990 rcu_read_lock();
2991 pool = get_work_pool(work);
2992 if (!pool) {
2993 rcu_read_unlock();
2994 return false;
2995 }
2996
2997 raw_spin_lock_irq(&pool->lock);
2998 /* see the comment in try_to_grab_pending() with the same code */
2999 pwq = get_work_pwq(work);
3000 if (pwq) {
3001 if (unlikely(pwq->pool != pool))
3002 goto already_gone;
3003 } else {
3004 worker = find_worker_executing_work(pool, work);
3005 if (!worker)
3006 goto already_gone;
3007 pwq = worker->current_pwq;
3008 }
3009
3010 check_flush_dependency(pwq->wq, work);
3011
3012 insert_wq_barrier(pwq, barr, work, worker);
3013 raw_spin_unlock_irq(&pool->lock);
3014
3015 /*
3016 * Force a lock recursion deadlock when using flush_work() inside a
3017 * single-threaded or rescuer equipped workqueue.
3018 *
3019 * For single threaded workqueues the deadlock happens when the work
3020 * is after the work issuing the flush_work(). For rescuer equipped
3021 * workqueues the deadlock happens when the rescuer stalls, blocking
3022 * forward progress.
3023 */
3024 if (!from_cancel &&
3025 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3026 lock_map_acquire(&pwq->wq->lockdep_map);
3027 lock_map_release(&pwq->wq->lockdep_map);
3028 }
3029 rcu_read_unlock();
3030 return true;
3031already_gone:
3032 raw_spin_unlock_irq(&pool->lock);
3033 rcu_read_unlock();
3034 return false;
3035}
3036
3037static bool __flush_work(struct work_struct *work, bool from_cancel)
3038{
3039 struct wq_barrier barr;
3040
3041 if (WARN_ON(!wq_online))
3042 return false;
3043
3044 if (WARN_ON(!work->func))
3045 return false;
3046
3047 if (!from_cancel) {
3048 lock_map_acquire(&work->lockdep_map);
3049 lock_map_release(&work->lockdep_map);
3050 }
3051
3052 if (start_flush_work(work, &barr, from_cancel)) {
3053 wait_for_completion(&barr.done);
3054 destroy_work_on_stack(&barr.work);
3055 return true;
3056 } else {
3057 return false;
3058 }
3059}
3060
3061/**
3062 * flush_work - wait for a work to finish executing the last queueing instance
3063 * @work: the work to flush
3064 *
3065 * Wait until @work has finished execution. @work is guaranteed to be idle
3066 * on return if it hasn't been requeued since flush started.
3067 *
3068 * Return:
3069 * %true if flush_work() waited for the work to finish execution,
3070 * %false if it was already idle.
3071 */
3072bool flush_work(struct work_struct *work)
3073{
3074 return __flush_work(work, false);
3075}
3076EXPORT_SYMBOL_GPL(flush_work);
3077
3078struct cwt_wait {
3079 wait_queue_entry_t wait;
3080 struct work_struct *work;
3081};
3082
3083static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3084{
3085 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3086
3087 if (cwait->work != key)
3088 return 0;
3089 return autoremove_wake_function(wait, mode, sync, key);
3090}
3091
3092static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3093{
3094 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3095 unsigned long flags;
3096 int ret;
3097
3098 do {
3099 ret = try_to_grab_pending(work, is_dwork, &flags);
3100 /*
3101 * If someone else is already canceling, wait for it to
3102 * finish. flush_work() doesn't work for PREEMPT_NONE
3103 * because we may get scheduled between @work's completion
3104 * and the other canceling task resuming and clearing
3105 * CANCELING - flush_work() will return false immediately
3106 * as @work is no longer busy, try_to_grab_pending() will
3107 * return -ENOENT as @work is still being canceled and the
3108 * other canceling task won't be able to clear CANCELING as
3109 * we're hogging the CPU.
3110 *
3111 * Let's wait for completion using a waitqueue. As this
3112 * may lead to the thundering herd problem, use a custom
3113 * wake function which matches @work along with exclusive
3114 * wait and wakeup.
3115 */
3116 if (unlikely(ret == -ENOENT)) {
3117 struct cwt_wait cwait;
3118
3119 init_wait(&cwait.wait);
3120 cwait.wait.func = cwt_wakefn;
3121 cwait.work = work;
3122
3123 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3124 TASK_UNINTERRUPTIBLE);
3125 if (work_is_canceling(work))
3126 schedule();
3127 finish_wait(&cancel_waitq, &cwait.wait);
3128 }
3129 } while (unlikely(ret < 0));
3130
3131 /* tell other tasks trying to grab @work to back off */
3132 mark_work_canceling(work);
3133 local_irq_restore(flags);
3134
3135 /*
3136 * This allows canceling during early boot. We know that @work
3137 * isn't executing.
3138 */
3139 if (wq_online)
3140 __flush_work(work, true);
3141
3142 clear_work_data(work);
3143
3144 /*
3145 * Paired with prepare_to_wait() above so that either
3146 * waitqueue_active() is visible here or !work_is_canceling() is
3147 * visible there.
3148 */
3149 smp_mb();
3150 if (waitqueue_active(&cancel_waitq))
3151 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3152
3153 return ret;
3154}
3155
3156/**
3157 * cancel_work_sync - cancel a work and wait for it to finish
3158 * @work: the work to cancel
3159 *
3160 * Cancel @work and wait for its execution to finish. This function
3161 * can be used even if the work re-queues itself or migrates to
3162 * another workqueue. On return from this function, @work is
3163 * guaranteed to be not pending or executing on any CPU.
3164 *
3165 * cancel_work_sync(&delayed_work->work) must not be used for
3166 * delayed_work's. Use cancel_delayed_work_sync() instead.
3167 *
3168 * The caller must ensure that the workqueue on which @work was last
3169 * queued can't be destroyed before this function returns.
3170 *
3171 * Return:
3172 * %true if @work was pending, %false otherwise.
3173 */
3174bool cancel_work_sync(struct work_struct *work)
3175{
3176 return __cancel_work_timer(work, false);
3177}
3178EXPORT_SYMBOL_GPL(cancel_work_sync);
3179
3180/**
3181 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3182 * @dwork: the delayed work to flush
3183 *
3184 * Delayed timer is cancelled and the pending work is queued for
3185 * immediate execution. Like flush_work(), this function only
3186 * considers the last queueing instance of @dwork.
3187 *
3188 * Return:
3189 * %true if flush_work() waited for the work to finish execution,
3190 * %false if it was already idle.
3191 */
3192bool flush_delayed_work(struct delayed_work *dwork)
3193{
3194 local_irq_disable();
3195 if (del_timer_sync(&dwork->timer))
3196 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3197 local_irq_enable();
3198 return flush_work(&dwork->work);
3199}
3200EXPORT_SYMBOL(flush_delayed_work);
3201
3202/**
3203 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3204 * @rwork: the rcu work to flush
3205 *
3206 * Return:
3207 * %true if flush_rcu_work() waited for the work to finish execution,
3208 * %false if it was already idle.
3209 */
3210bool flush_rcu_work(struct rcu_work *rwork)
3211{
3212 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3213 rcu_barrier();
3214 flush_work(&rwork->work);
3215 return true;
3216 } else {
3217 return flush_work(&rwork->work);
3218 }
3219}
3220EXPORT_SYMBOL(flush_rcu_work);
3221
3222static bool __cancel_work(struct work_struct *work, bool is_dwork)
3223{
3224 unsigned long flags;
3225 int ret;
3226
3227 do {
3228 ret = try_to_grab_pending(work, is_dwork, &flags);
3229 } while (unlikely(ret == -EAGAIN));
3230
3231 if (unlikely(ret < 0))
3232 return false;
3233
3234 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3235 local_irq_restore(flags);
3236 return ret;
3237}
3238
3239/**
3240 * cancel_delayed_work - cancel a delayed work
3241 * @dwork: delayed_work to cancel
3242 *
3243 * Kill off a pending delayed_work.
3244 *
3245 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3246 * pending.
3247 *
3248 * Note:
3249 * The work callback function may still be running on return, unless
3250 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3251 * use cancel_delayed_work_sync() to wait on it.
3252 *
3253 * This function is safe to call from any context including IRQ handler.
3254 */
3255bool cancel_delayed_work(struct delayed_work *dwork)
3256{
3257 return __cancel_work(&dwork->work, true);
3258}
3259EXPORT_SYMBOL(cancel_delayed_work);
3260
3261/**
3262 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3263 * @dwork: the delayed work cancel
3264 *
3265 * This is cancel_work_sync() for delayed works.
3266 *
3267 * Return:
3268 * %true if @dwork was pending, %false otherwise.
3269 */
3270bool cancel_delayed_work_sync(struct delayed_work *dwork)
3271{
3272 return __cancel_work_timer(&dwork->work, true);
3273}
3274EXPORT_SYMBOL(cancel_delayed_work_sync);
3275
3276/**
3277 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3278 * @func: the function to call
3279 *
3280 * schedule_on_each_cpu() executes @func on each online CPU using the
3281 * system workqueue and blocks until all CPUs have completed.
3282 * schedule_on_each_cpu() is very slow.
3283 *
3284 * Return:
3285 * 0 on success, -errno on failure.
3286 */
3287int schedule_on_each_cpu(work_func_t func)
3288{
3289 int cpu;
3290 struct work_struct __percpu *works;
3291
3292 works = alloc_percpu(struct work_struct);
3293 if (!works)
3294 return -ENOMEM;
3295
3296 get_online_cpus();
3297
3298 for_each_online_cpu(cpu) {
3299 struct work_struct *work = per_cpu_ptr(works, cpu);
3300
3301 INIT_WORK(work, func);
3302 schedule_work_on(cpu, work);
3303 }
3304
3305 for_each_online_cpu(cpu)
3306 flush_work(per_cpu_ptr(works, cpu));
3307
3308 put_online_cpus();
3309 free_percpu(works);
3310 return 0;
3311}
3312
3313/**
3314 * execute_in_process_context - reliably execute the routine with user context
3315 * @fn: the function to execute
3316 * @ew: guaranteed storage for the execute work structure (must
3317 * be available when the work executes)
3318 *
3319 * Executes the function immediately if process context is available,
3320 * otherwise schedules the function for delayed execution.
3321 *
3322 * Return: 0 - function was executed
3323 * 1 - function was scheduled for execution
3324 */
3325int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3326{
3327 if (!in_interrupt()) {
3328 fn(&ew->work);
3329 return 0;
3330 }
3331
3332 INIT_WORK(&ew->work, fn);
3333 schedule_work(&ew->work);
3334
3335 return 1;
3336}
3337EXPORT_SYMBOL_GPL(execute_in_process_context);
3338
3339/**
3340 * free_workqueue_attrs - free a workqueue_attrs
3341 * @attrs: workqueue_attrs to free
3342 *
3343 * Undo alloc_workqueue_attrs().
3344 */
3345void free_workqueue_attrs(struct workqueue_attrs *attrs)
3346{
3347 if (attrs) {
3348 free_cpumask_var(attrs->cpumask);
3349 kfree(attrs);
3350 }
3351}
3352
3353/**
3354 * alloc_workqueue_attrs - allocate a workqueue_attrs
3355 *
3356 * Allocate a new workqueue_attrs, initialize with default settings and
3357 * return it.
3358 *
3359 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3360 */
3361struct workqueue_attrs *alloc_workqueue_attrs(void)
3362{
3363 struct workqueue_attrs *attrs;
3364
3365 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3366 if (!attrs)
3367 goto fail;
3368 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3369 goto fail;
3370
3371 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3372 return attrs;
3373fail:
3374 free_workqueue_attrs(attrs);
3375 return NULL;
3376}
3377
3378static void copy_workqueue_attrs(struct workqueue_attrs *to,
3379 const struct workqueue_attrs *from)
3380{
3381 to->nice = from->nice;
3382 cpumask_copy(to->cpumask, from->cpumask);
3383 /*
3384 * Unlike hash and equality test, this function doesn't ignore
3385 * ->no_numa as it is used for both pool and wq attrs. Instead,
3386 * get_unbound_pool() explicitly clears ->no_numa after copying.
3387 */
3388 to->no_numa = from->no_numa;
3389}
3390
3391/* hash value of the content of @attr */
3392static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3393{
3394 u32 hash = 0;
3395
3396 hash = jhash_1word(attrs->nice, hash);
3397 hash = jhash(cpumask_bits(attrs->cpumask),
3398 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3399 return hash;
3400}
3401
3402/* content equality test */
3403static bool wqattrs_equal(const struct workqueue_attrs *a,
3404 const struct workqueue_attrs *b)
3405{
3406 if (a->nice != b->nice)
3407 return false;
3408 if (!cpumask_equal(a->cpumask, b->cpumask))
3409 return false;
3410 return true;
3411}
3412
3413/**
3414 * init_worker_pool - initialize a newly zalloc'd worker_pool
3415 * @pool: worker_pool to initialize
3416 *
3417 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3418 *
3419 * Return: 0 on success, -errno on failure. Even on failure, all fields
3420 * inside @pool proper are initialized and put_unbound_pool() can be called
3421 * on @pool safely to release it.
3422 */
3423static int init_worker_pool(struct worker_pool *pool)
3424{
3425 raw_spin_lock_init(&pool->lock);
3426 pool->id = -1;
3427 pool->cpu = -1;
3428 pool->node = NUMA_NO_NODE;
3429 pool->flags |= POOL_DISASSOCIATED;
3430 pool->watchdog_ts = jiffies;
3431 INIT_LIST_HEAD(&pool->worklist);
3432 INIT_LIST_HEAD(&pool->idle_list);
3433 hash_init(pool->busy_hash);
3434
3435 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3436
3437 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3438
3439 INIT_LIST_HEAD(&pool->workers);
3440
3441 ida_init(&pool->worker_ida);
3442 INIT_HLIST_NODE(&pool->hash_node);
3443 pool->refcnt = 1;
3444
3445 /* shouldn't fail above this point */
3446 pool->attrs = alloc_workqueue_attrs();
3447 if (!pool->attrs)
3448 return -ENOMEM;
3449 return 0;
3450}
3451
3452#ifdef CONFIG_LOCKDEP
3453static void wq_init_lockdep(struct workqueue_struct *wq)
3454{
3455 char *lock_name;
3456
3457 lockdep_register_key(&wq->key);
3458 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3459 if (!lock_name)
3460 lock_name = wq->name;
3461
3462 wq->lock_name = lock_name;
3463 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3464}
3465
3466static void wq_unregister_lockdep(struct workqueue_struct *wq)
3467{
3468 lockdep_unregister_key(&wq->key);
3469}
3470
3471static void wq_free_lockdep(struct workqueue_struct *wq)
3472{
3473 if (wq->lock_name != wq->name)
3474 kfree(wq->lock_name);
3475}
3476#else
3477static void wq_init_lockdep(struct workqueue_struct *wq)
3478{
3479}
3480
3481static void wq_unregister_lockdep(struct workqueue_struct *wq)
3482{
3483}
3484
3485static void wq_free_lockdep(struct workqueue_struct *wq)
3486{
3487}
3488#endif
3489
3490static void rcu_free_wq(struct rcu_head *rcu)
3491{
3492 struct workqueue_struct *wq =
3493 container_of(rcu, struct workqueue_struct, rcu);
3494
3495 wq_free_lockdep(wq);
3496
3497 if (!(wq->flags & WQ_UNBOUND))
3498 free_percpu(wq->cpu_pwqs);
3499 else
3500 free_workqueue_attrs(wq->unbound_attrs);
3501
3502 kfree(wq);
3503}
3504
3505static void rcu_free_pool(struct rcu_head *rcu)
3506{
3507 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3508
3509 ida_destroy(&pool->worker_ida);
3510 free_workqueue_attrs(pool->attrs);
3511 kfree(pool);
3512}
3513
3514/* This returns with the lock held on success (pool manager is inactive). */
3515static bool wq_manager_inactive(struct worker_pool *pool)
3516{
3517 raw_spin_lock_irq(&pool->lock);
3518
3519 if (pool->flags & POOL_MANAGER_ACTIVE) {
3520 raw_spin_unlock_irq(&pool->lock);
3521 return false;
3522 }
3523 return true;
3524}
3525
3526/**
3527 * put_unbound_pool - put a worker_pool
3528 * @pool: worker_pool to put
3529 *
3530 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3531 * safe manner. get_unbound_pool() calls this function on its failure path
3532 * and this function should be able to release pools which went through,
3533 * successfully or not, init_worker_pool().
3534 *
3535 * Should be called with wq_pool_mutex held.
3536 */
3537static void put_unbound_pool(struct worker_pool *pool)
3538{
3539 DECLARE_COMPLETION_ONSTACK(detach_completion);
3540 struct worker *worker;
3541
3542 lockdep_assert_held(&wq_pool_mutex);
3543
3544 if (--pool->refcnt)
3545 return;
3546
3547 /* sanity checks */
3548 if (WARN_ON(!(pool->cpu < 0)) ||
3549 WARN_ON(!list_empty(&pool->worklist)))
3550 return;
3551
3552 /* release id and unhash */
3553 if (pool->id >= 0)
3554 idr_remove(&worker_pool_idr, pool->id);
3555 hash_del(&pool->hash_node);
3556
3557 /*
3558 * Become the manager and destroy all workers. This prevents
3559 * @pool's workers from blocking on attach_mutex. We're the last
3560 * manager and @pool gets freed with the flag set.
3561 * Because of how wq_manager_inactive() works, we will hold the
3562 * spinlock after a successful wait.
3563 */
3564 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3565 TASK_UNINTERRUPTIBLE);
3566 pool->flags |= POOL_MANAGER_ACTIVE;
3567
3568 while ((worker = first_idle_worker(pool)))
3569 destroy_worker(worker);
3570 WARN_ON(pool->nr_workers || pool->nr_idle);
3571 raw_spin_unlock_irq(&pool->lock);
3572
3573 mutex_lock(&wq_pool_attach_mutex);
3574 if (!list_empty(&pool->workers))
3575 pool->detach_completion = &detach_completion;
3576 mutex_unlock(&wq_pool_attach_mutex);
3577
3578 if (pool->detach_completion)
3579 wait_for_completion(pool->detach_completion);
3580
3581 /* shut down the timers */
3582 del_timer_sync(&pool->idle_timer);
3583 del_timer_sync(&pool->mayday_timer);
3584
3585 /* RCU protected to allow dereferences from get_work_pool() */
3586 call_rcu(&pool->rcu, rcu_free_pool);
3587}
3588
3589/**
3590 * get_unbound_pool - get a worker_pool with the specified attributes
3591 * @attrs: the attributes of the worker_pool to get
3592 *
3593 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3594 * reference count and return it. If there already is a matching
3595 * worker_pool, it will be used; otherwise, this function attempts to
3596 * create a new one.
3597 *
3598 * Should be called with wq_pool_mutex held.
3599 *
3600 * Return: On success, a worker_pool with the same attributes as @attrs.
3601 * On failure, %NULL.
3602 */
3603static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3604{
3605 u32 hash = wqattrs_hash(attrs);
3606 struct worker_pool *pool;
3607 int node;
3608 int target_node = NUMA_NO_NODE;
3609
3610 lockdep_assert_held(&wq_pool_mutex);
3611
3612 /* do we already have a matching pool? */
3613 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3614 if (wqattrs_equal(pool->attrs, attrs)) {
3615 pool->refcnt++;
3616 return pool;
3617 }
3618 }
3619
3620 /* if cpumask is contained inside a NUMA node, we belong to that node */
3621 if (wq_numa_enabled) {
3622 for_each_node(node) {
3623 if (cpumask_subset(attrs->cpumask,
3624 wq_numa_possible_cpumask[node])) {
3625 target_node = node;
3626 break;
3627 }
3628 }
3629 }
3630
3631 /* nope, create a new one */
3632 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3633 if (!pool || init_worker_pool(pool) < 0)
3634 goto fail;
3635
3636 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3637 copy_workqueue_attrs(pool->attrs, attrs);
3638 pool->node = target_node;
3639
3640 /*
3641 * no_numa isn't a worker_pool attribute, always clear it. See
3642 * 'struct workqueue_attrs' comments for detail.
3643 */
3644 pool->attrs->no_numa = false;
3645
3646 if (worker_pool_assign_id(pool) < 0)
3647 goto fail;
3648
3649 /* create and start the initial worker */
3650 if (wq_online && !create_worker(pool))
3651 goto fail;
3652
3653 /* install */
3654 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3655
3656 return pool;
3657fail:
3658 if (pool)
3659 put_unbound_pool(pool);
3660 return NULL;
3661}
3662
3663static void rcu_free_pwq(struct rcu_head *rcu)
3664{
3665 kmem_cache_free(pwq_cache,
3666 container_of(rcu, struct pool_workqueue, rcu));
3667}
3668
3669/*
3670 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3671 * and needs to be destroyed.
3672 */
3673static void pwq_unbound_release_workfn(struct work_struct *work)
3674{
3675 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3676 unbound_release_work);
3677 struct workqueue_struct *wq = pwq->wq;
3678 struct worker_pool *pool = pwq->pool;
3679 bool is_last = false;
3680
3681 /*
3682 * when @pwq is not linked, it doesn't hold any reference to the
3683 * @wq, and @wq is invalid to access.
3684 */
3685 if (!list_empty(&pwq->pwqs_node)) {
3686 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3687 return;
3688
3689 mutex_lock(&wq->mutex);
3690 list_del_rcu(&pwq->pwqs_node);
3691 is_last = list_empty(&wq->pwqs);
3692 mutex_unlock(&wq->mutex);
3693 }
3694
3695 mutex_lock(&wq_pool_mutex);
3696 put_unbound_pool(pool);
3697 mutex_unlock(&wq_pool_mutex);
3698
3699 call_rcu(&pwq->rcu, rcu_free_pwq);
3700
3701 /*
3702 * If we're the last pwq going away, @wq is already dead and no one
3703 * is gonna access it anymore. Schedule RCU free.
3704 */
3705 if (is_last) {
3706 wq_unregister_lockdep(wq);
3707 call_rcu(&wq->rcu, rcu_free_wq);
3708 }
3709}
3710
3711/**
3712 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3713 * @pwq: target pool_workqueue
3714 *
3715 * If @pwq isn't freezing, set @pwq->max_active to the associated
3716 * workqueue's saved_max_active and activate delayed work items
3717 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3718 */
3719static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3720{
3721 struct workqueue_struct *wq = pwq->wq;
3722 bool freezable = wq->flags & WQ_FREEZABLE;
3723 unsigned long flags;
3724
3725 /* for @wq->saved_max_active */
3726 lockdep_assert_held(&wq->mutex);
3727
3728 /* fast exit for non-freezable wqs */
3729 if (!freezable && pwq->max_active == wq->saved_max_active)
3730 return;
3731
3732 /* this function can be called during early boot w/ irq disabled */
3733 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3734
3735 /*
3736 * During [un]freezing, the caller is responsible for ensuring that
3737 * this function is called at least once after @workqueue_freezing
3738 * is updated and visible.
3739 */
3740 if (!freezable || !workqueue_freezing) {
3741 bool kick = false;
3742
3743 pwq->max_active = wq->saved_max_active;
3744
3745 while (!list_empty(&pwq->delayed_works) &&
3746 pwq->nr_active < pwq->max_active) {
3747 pwq_activate_first_delayed(pwq);
3748 kick = true;
3749 }
3750
3751 /*
3752 * Need to kick a worker after thawed or an unbound wq's
3753 * max_active is bumped. In realtime scenarios, always kicking a
3754 * worker will cause interference on the isolated cpu cores, so
3755 * let's kick iff work items were activated.
3756 */
3757 if (kick)
3758 wake_up_worker(pwq->pool);
3759 } else {
3760 pwq->max_active = 0;
3761 }
3762
3763 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3764}
3765
3766/* initialize newly alloced @pwq which is associated with @wq and @pool */
3767static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3768 struct worker_pool *pool)
3769{
3770 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3771
3772 memset(pwq, 0, sizeof(*pwq));
3773
3774 pwq->pool = pool;
3775 pwq->wq = wq;
3776 pwq->flush_color = -1;
3777 pwq->refcnt = 1;
3778 INIT_LIST_HEAD(&pwq->delayed_works);
3779 INIT_LIST_HEAD(&pwq->pwqs_node);
3780 INIT_LIST_HEAD(&pwq->mayday_node);
3781 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3782}
3783
3784/* sync @pwq with the current state of its associated wq and link it */
3785static void link_pwq(struct pool_workqueue *pwq)
3786{
3787 struct workqueue_struct *wq = pwq->wq;
3788
3789 lockdep_assert_held(&wq->mutex);
3790
3791 /* may be called multiple times, ignore if already linked */
3792 if (!list_empty(&pwq->pwqs_node))
3793 return;
3794
3795 /* set the matching work_color */
3796 pwq->work_color = wq->work_color;
3797
3798 /* sync max_active to the current setting */
3799 pwq_adjust_max_active(pwq);
3800
3801 /* link in @pwq */
3802 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3803}
3804
3805/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3806static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3807 const struct workqueue_attrs *attrs)
3808{
3809 struct worker_pool *pool;
3810 struct pool_workqueue *pwq;
3811
3812 lockdep_assert_held(&wq_pool_mutex);
3813
3814 pool = get_unbound_pool(attrs);
3815 if (!pool)
3816 return NULL;
3817
3818 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3819 if (!pwq) {
3820 put_unbound_pool(pool);
3821 return NULL;
3822 }
3823
3824 init_pwq(pwq, wq, pool);
3825 return pwq;
3826}
3827
3828/**
3829 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3830 * @attrs: the wq_attrs of the default pwq of the target workqueue
3831 * @node: the target NUMA node
3832 * @cpu_going_down: if >= 0, the CPU to consider as offline
3833 * @cpumask: outarg, the resulting cpumask
3834 *
3835 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3836 * @cpu_going_down is >= 0, that cpu is considered offline during
3837 * calculation. The result is stored in @cpumask.
3838 *
3839 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3840 * enabled and @node has online CPUs requested by @attrs, the returned
3841 * cpumask is the intersection of the possible CPUs of @node and
3842 * @attrs->cpumask.
3843 *
3844 * The caller is responsible for ensuring that the cpumask of @node stays
3845 * stable.
3846 *
3847 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3848 * %false if equal.
3849 */
3850static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3851 int cpu_going_down, cpumask_t *cpumask)
3852{
3853 if (!wq_numa_enabled || attrs->no_numa)
3854 goto use_dfl;
3855
3856 /* does @node have any online CPUs @attrs wants? */
3857 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3858 if (cpu_going_down >= 0)
3859 cpumask_clear_cpu(cpu_going_down, cpumask);
3860
3861 if (cpumask_empty(cpumask))
3862 goto use_dfl;
3863
3864 /* yeap, return possible CPUs in @node that @attrs wants */
3865 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3866
3867 if (cpumask_empty(cpumask)) {
3868 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3869 "possible intersect\n");
3870 return false;
3871 }
3872
3873 return !cpumask_equal(cpumask, attrs->cpumask);
3874
3875use_dfl:
3876 cpumask_copy(cpumask, attrs->cpumask);
3877 return false;
3878}
3879
3880/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3881static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3882 int node,
3883 struct pool_workqueue *pwq)
3884{
3885 struct pool_workqueue *old_pwq;
3886
3887 lockdep_assert_held(&wq_pool_mutex);
3888 lockdep_assert_held(&wq->mutex);
3889
3890 /* link_pwq() can handle duplicate calls */
3891 link_pwq(pwq);
3892
3893 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3894 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3895 return old_pwq;
3896}
3897
3898/* context to store the prepared attrs & pwqs before applying */
3899struct apply_wqattrs_ctx {
3900 struct workqueue_struct *wq; /* target workqueue */
3901 struct workqueue_attrs *attrs; /* attrs to apply */
3902 struct list_head list; /* queued for batching commit */
3903 struct pool_workqueue *dfl_pwq;
3904 struct pool_workqueue *pwq_tbl[];
3905};
3906
3907/* free the resources after success or abort */
3908static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3909{
3910 if (ctx) {
3911 int node;
3912
3913 for_each_node(node)
3914 put_pwq_unlocked(ctx->pwq_tbl[node]);
3915 put_pwq_unlocked(ctx->dfl_pwq);
3916
3917 free_workqueue_attrs(ctx->attrs);
3918
3919 kfree(ctx);
3920 }
3921}
3922
3923/* allocate the attrs and pwqs for later installation */
3924static struct apply_wqattrs_ctx *
3925apply_wqattrs_prepare(struct workqueue_struct *wq,
3926 const struct workqueue_attrs *attrs)
3927{
3928 struct apply_wqattrs_ctx *ctx;
3929 struct workqueue_attrs *new_attrs, *tmp_attrs;
3930 int node;
3931
3932 lockdep_assert_held(&wq_pool_mutex);
3933
3934 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3935
3936 new_attrs = alloc_workqueue_attrs();
3937 tmp_attrs = alloc_workqueue_attrs();
3938 if (!ctx || !new_attrs || !tmp_attrs)
3939 goto out_free;
3940
3941 /*
3942 * Calculate the attrs of the default pwq.
3943 * If the user configured cpumask doesn't overlap with the
3944 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3945 */
3946 copy_workqueue_attrs(new_attrs, attrs);
3947 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3948 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3949 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3950
3951 /*
3952 * We may create multiple pwqs with differing cpumasks. Make a
3953 * copy of @new_attrs which will be modified and used to obtain
3954 * pools.
3955 */
3956 copy_workqueue_attrs(tmp_attrs, new_attrs);
3957
3958 /*
3959 * If something goes wrong during CPU up/down, we'll fall back to
3960 * the default pwq covering whole @attrs->cpumask. Always create
3961 * it even if we don't use it immediately.
3962 */
3963 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3964 if (!ctx->dfl_pwq)
3965 goto out_free;
3966
3967 for_each_node(node) {
3968 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3969 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3970 if (!ctx->pwq_tbl[node])
3971 goto out_free;
3972 } else {
3973 ctx->dfl_pwq->refcnt++;
3974 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3975 }
3976 }
3977
3978 /* save the user configured attrs and sanitize it. */
3979 copy_workqueue_attrs(new_attrs, attrs);
3980 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3981 ctx->attrs = new_attrs;
3982
3983 ctx->wq = wq;
3984 free_workqueue_attrs(tmp_attrs);
3985 return ctx;
3986
3987out_free:
3988 free_workqueue_attrs(tmp_attrs);
3989 free_workqueue_attrs(new_attrs);
3990 apply_wqattrs_cleanup(ctx);
3991 return NULL;
3992}
3993
3994/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3995static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3996{
3997 int node;
3998
3999 /* all pwqs have been created successfully, let's install'em */
4000 mutex_lock(&ctx->wq->mutex);
4001
4002 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4003
4004 /* save the previous pwq and install the new one */
4005 for_each_node(node)
4006 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4007 ctx->pwq_tbl[node]);
4008
4009 /* @dfl_pwq might not have been used, ensure it's linked */
4010 link_pwq(ctx->dfl_pwq);
4011 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4012
4013 mutex_unlock(&ctx->wq->mutex);
4014}
4015
4016static void apply_wqattrs_lock(void)
4017{
4018 /* CPUs should stay stable across pwq creations and installations */
4019 get_online_cpus();
4020 mutex_lock(&wq_pool_mutex);
4021}
4022
4023static void apply_wqattrs_unlock(void)
4024{
4025 mutex_unlock(&wq_pool_mutex);
4026 put_online_cpus();
4027}
4028
4029static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4030 const struct workqueue_attrs *attrs)
4031{
4032 struct apply_wqattrs_ctx *ctx;
4033
4034 /* only unbound workqueues can change attributes */
4035 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4036 return -EINVAL;
4037
4038 /* creating multiple pwqs breaks ordering guarantee */
4039 if (!list_empty(&wq->pwqs)) {
4040 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4041 return -EINVAL;
4042
4043 wq->flags &= ~__WQ_ORDERED;
4044 }
4045
4046 ctx = apply_wqattrs_prepare(wq, attrs);
4047 if (!ctx)
4048 return -ENOMEM;
4049
4050 /* the ctx has been prepared successfully, let's commit it */
4051 apply_wqattrs_commit(ctx);
4052 apply_wqattrs_cleanup(ctx);
4053
4054 return 0;
4055}
4056
4057/**
4058 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4059 * @wq: the target workqueue
4060 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4061 *
4062 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4063 * machines, this function maps a separate pwq to each NUMA node with
4064 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4065 * NUMA node it was issued on. Older pwqs are released as in-flight work
4066 * items finish. Note that a work item which repeatedly requeues itself
4067 * back-to-back will stay on its current pwq.
4068 *
4069 * Performs GFP_KERNEL allocations.
4070 *
4071 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4072 *
4073 * Return: 0 on success and -errno on failure.
4074 */
4075int apply_workqueue_attrs(struct workqueue_struct *wq,
4076 const struct workqueue_attrs *attrs)
4077{
4078 int ret;
4079
4080 lockdep_assert_cpus_held();
4081
4082 mutex_lock(&wq_pool_mutex);
4083 ret = apply_workqueue_attrs_locked(wq, attrs);
4084 mutex_unlock(&wq_pool_mutex);
4085
4086 return ret;
4087}
4088
4089/**
4090 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4091 * @wq: the target workqueue
4092 * @cpu: the CPU coming up or going down
4093 * @online: whether @cpu is coming up or going down
4094 *
4095 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4096 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4097 * @wq accordingly.
4098 *
4099 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4100 * falls back to @wq->dfl_pwq which may not be optimal but is always
4101 * correct.
4102 *
4103 * Note that when the last allowed CPU of a NUMA node goes offline for a
4104 * workqueue with a cpumask spanning multiple nodes, the workers which were
4105 * already executing the work items for the workqueue will lose their CPU
4106 * affinity and may execute on any CPU. This is similar to how per-cpu
4107 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4108 * affinity, it's the user's responsibility to flush the work item from
4109 * CPU_DOWN_PREPARE.
4110 */
4111static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4112 bool online)
4113{
4114 int node = cpu_to_node(cpu);
4115 int cpu_off = online ? -1 : cpu;
4116 struct pool_workqueue *old_pwq = NULL, *pwq;
4117 struct workqueue_attrs *target_attrs;
4118 cpumask_t *cpumask;
4119
4120 lockdep_assert_held(&wq_pool_mutex);
4121
4122 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4123 wq->unbound_attrs->no_numa)
4124 return;
4125
4126 /*
4127 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4128 * Let's use a preallocated one. The following buf is protected by
4129 * CPU hotplug exclusion.
4130 */
4131 target_attrs = wq_update_unbound_numa_attrs_buf;
4132 cpumask = target_attrs->cpumask;
4133
4134 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4135 pwq = unbound_pwq_by_node(wq, node);
4136
4137 /*
4138 * Let's determine what needs to be done. If the target cpumask is
4139 * different from the default pwq's, we need to compare it to @pwq's
4140 * and create a new one if they don't match. If the target cpumask
4141 * equals the default pwq's, the default pwq should be used.
4142 */
4143 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4144 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4145 return;
4146 } else {
4147 goto use_dfl_pwq;
4148 }
4149
4150 /* create a new pwq */
4151 pwq = alloc_unbound_pwq(wq, target_attrs);
4152 if (!pwq) {
4153 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4154 wq->name);
4155 goto use_dfl_pwq;
4156 }
4157
4158 /* Install the new pwq. */
4159 mutex_lock(&wq->mutex);
4160 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4161 goto out_unlock;
4162
4163use_dfl_pwq:
4164 mutex_lock(&wq->mutex);
4165 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4166 get_pwq(wq->dfl_pwq);
4167 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4168 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4169out_unlock:
4170 mutex_unlock(&wq->mutex);
4171 put_pwq_unlocked(old_pwq);
4172}
4173
4174static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4175{
4176 bool highpri = wq->flags & WQ_HIGHPRI;
4177 int cpu, ret;
4178
4179 if (!(wq->flags & WQ_UNBOUND)) {
4180 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4181 if (!wq->cpu_pwqs)
4182 return -ENOMEM;
4183
4184 for_each_possible_cpu(cpu) {
4185 struct pool_workqueue *pwq =
4186 per_cpu_ptr(wq->cpu_pwqs, cpu);
4187 struct worker_pool *cpu_pools =
4188 per_cpu(cpu_worker_pools, cpu);
4189
4190 init_pwq(pwq, wq, &cpu_pools[highpri]);
4191
4192 mutex_lock(&wq->mutex);
4193 link_pwq(pwq);
4194 mutex_unlock(&wq->mutex);
4195 }
4196 return 0;
4197 }
4198
4199 get_online_cpus();
4200 if (wq->flags & __WQ_ORDERED) {
4201 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4202 /* there should only be single pwq for ordering guarantee */
4203 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4204 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4205 "ordering guarantee broken for workqueue %s\n", wq->name);
4206 } else {
4207 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4208 }
4209 put_online_cpus();
4210
4211 return ret;
4212}
4213
4214static int wq_clamp_max_active(int max_active, unsigned int flags,
4215 const char *name)
4216{
4217 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4218
4219 if (max_active < 1 || max_active > lim)
4220 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4221 max_active, name, 1, lim);
4222
4223 return clamp_val(max_active, 1, lim);
4224}
4225
4226/*
4227 * Workqueues which may be used during memory reclaim should have a rescuer
4228 * to guarantee forward progress.
4229 */
4230static int init_rescuer(struct workqueue_struct *wq)
4231{
4232 struct worker *rescuer;
4233 int ret;
4234
4235 if (!(wq->flags & WQ_MEM_RECLAIM))
4236 return 0;
4237
4238 rescuer = alloc_worker(NUMA_NO_NODE);
4239 if (!rescuer)
4240 return -ENOMEM;
4241
4242 rescuer->rescue_wq = wq;
4243 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4244 if (IS_ERR(rescuer->task)) {
4245 ret = PTR_ERR(rescuer->task);
4246 kfree(rescuer);
4247 return ret;
4248 }
4249
4250 wq->rescuer = rescuer;
4251 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4252 wake_up_process(rescuer->task);
4253
4254 return 0;
4255}
4256
4257__printf(1, 4)
4258struct workqueue_struct *alloc_workqueue(const char *fmt,
4259 unsigned int flags,
4260 int max_active, ...)
4261{
4262 size_t tbl_size = 0;
4263 va_list args;
4264 struct workqueue_struct *wq;
4265 struct pool_workqueue *pwq;
4266
4267 /*
4268 * Unbound && max_active == 1 used to imply ordered, which is no
4269 * longer the case on NUMA machines due to per-node pools. While
4270 * alloc_ordered_workqueue() is the right way to create an ordered
4271 * workqueue, keep the previous behavior to avoid subtle breakages
4272 * on NUMA.
4273 */
4274 if ((flags & WQ_UNBOUND) && max_active == 1)
4275 flags |= __WQ_ORDERED;
4276
4277 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4278 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4279 flags |= WQ_UNBOUND;
4280
4281 /* allocate wq and format name */
4282 if (flags & WQ_UNBOUND)
4283 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4284
4285 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4286 if (!wq)
4287 return NULL;
4288
4289 if (flags & WQ_UNBOUND) {
4290 wq->unbound_attrs = alloc_workqueue_attrs();
4291 if (!wq->unbound_attrs)
4292 goto err_free_wq;
4293 }
4294
4295 va_start(args, max_active);
4296 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4297 va_end(args);
4298
4299 max_active = max_active ?: WQ_DFL_ACTIVE;
4300 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4301
4302 /* init wq */
4303 wq->flags = flags;
4304 wq->saved_max_active = max_active;
4305 mutex_init(&wq->mutex);
4306 atomic_set(&wq->nr_pwqs_to_flush, 0);
4307 INIT_LIST_HEAD(&wq->pwqs);
4308 INIT_LIST_HEAD(&wq->flusher_queue);
4309 INIT_LIST_HEAD(&wq->flusher_overflow);
4310 INIT_LIST_HEAD(&wq->maydays);
4311
4312 wq_init_lockdep(wq);
4313 INIT_LIST_HEAD(&wq->list);
4314
4315 if (alloc_and_link_pwqs(wq) < 0)
4316 goto err_unreg_lockdep;
4317
4318 if (wq_online && init_rescuer(wq) < 0)
4319 goto err_destroy;
4320
4321 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4322 goto err_destroy;
4323
4324 /*
4325 * wq_pool_mutex protects global freeze state and workqueues list.
4326 * Grab it, adjust max_active and add the new @wq to workqueues
4327 * list.
4328 */
4329 mutex_lock(&wq_pool_mutex);
4330
4331 mutex_lock(&wq->mutex);
4332 for_each_pwq(pwq, wq)
4333 pwq_adjust_max_active(pwq);
4334 mutex_unlock(&wq->mutex);
4335
4336 list_add_tail_rcu(&wq->list, &workqueues);
4337
4338 mutex_unlock(&wq_pool_mutex);
4339
4340 return wq;
4341
4342err_unreg_lockdep:
4343 wq_unregister_lockdep(wq);
4344 wq_free_lockdep(wq);
4345err_free_wq:
4346 free_workqueue_attrs(wq->unbound_attrs);
4347 kfree(wq);
4348 return NULL;
4349err_destroy:
4350 destroy_workqueue(wq);
4351 return NULL;
4352}
4353EXPORT_SYMBOL_GPL(alloc_workqueue);
4354
4355static bool pwq_busy(struct pool_workqueue *pwq)
4356{
4357 int i;
4358
4359 for (i = 0; i < WORK_NR_COLORS; i++)
4360 if (pwq->nr_in_flight[i])
4361 return true;
4362
4363 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4364 return true;
4365 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4366 return true;
4367
4368 return false;
4369}
4370
4371/**
4372 * destroy_workqueue - safely terminate a workqueue
4373 * @wq: target workqueue
4374 *
4375 * Safely destroy a workqueue. All work currently pending will be done first.
4376 */
4377void destroy_workqueue(struct workqueue_struct *wq)
4378{
4379 struct pool_workqueue *pwq;
4380 int node;
4381
4382 /*
4383 * Remove it from sysfs first so that sanity check failure doesn't
4384 * lead to sysfs name conflicts.
4385 */
4386 workqueue_sysfs_unregister(wq);
4387
4388 /* drain it before proceeding with destruction */
4389 drain_workqueue(wq);
4390
4391 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4392 if (wq->rescuer) {
4393 struct worker *rescuer = wq->rescuer;
4394
4395 /* this prevents new queueing */
4396 raw_spin_lock_irq(&wq_mayday_lock);
4397 wq->rescuer = NULL;
4398 raw_spin_unlock_irq(&wq_mayday_lock);
4399
4400 /* rescuer will empty maydays list before exiting */
4401 kthread_stop(rescuer->task);
4402 kfree(rescuer);
4403 }
4404
4405 /*
4406 * Sanity checks - grab all the locks so that we wait for all
4407 * in-flight operations which may do put_pwq().
4408 */
4409 mutex_lock(&wq_pool_mutex);
4410 mutex_lock(&wq->mutex);
4411 for_each_pwq(pwq, wq) {
4412 raw_spin_lock_irq(&pwq->pool->lock);
4413 if (WARN_ON(pwq_busy(pwq))) {
4414 pr_warn("%s: %s has the following busy pwq\n",
4415 __func__, wq->name);
4416 show_pwq(pwq);
4417 raw_spin_unlock_irq(&pwq->pool->lock);
4418 mutex_unlock(&wq->mutex);
4419 mutex_unlock(&wq_pool_mutex);
4420 show_workqueue_state();
4421 return;
4422 }
4423 raw_spin_unlock_irq(&pwq->pool->lock);
4424 }
4425 mutex_unlock(&wq->mutex);
4426
4427 /*
4428 * wq list is used to freeze wq, remove from list after
4429 * flushing is complete in case freeze races us.
4430 */
4431 list_del_rcu(&wq->list);
4432 mutex_unlock(&wq_pool_mutex);
4433
4434 if (!(wq->flags & WQ_UNBOUND)) {
4435 wq_unregister_lockdep(wq);
4436 /*
4437 * The base ref is never dropped on per-cpu pwqs. Directly
4438 * schedule RCU free.
4439 */
4440 call_rcu(&wq->rcu, rcu_free_wq);
4441 } else {
4442 /*
4443 * We're the sole accessor of @wq at this point. Directly
4444 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4445 * @wq will be freed when the last pwq is released.
4446 */
4447 for_each_node(node) {
4448 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4449 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4450 put_pwq_unlocked(pwq);
4451 }
4452
4453 /*
4454 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4455 * put. Don't access it afterwards.
4456 */
4457 pwq = wq->dfl_pwq;
4458 wq->dfl_pwq = NULL;
4459 put_pwq_unlocked(pwq);
4460 }
4461}
4462EXPORT_SYMBOL_GPL(destroy_workqueue);
4463
4464/**
4465 * workqueue_set_max_active - adjust max_active of a workqueue
4466 * @wq: target workqueue
4467 * @max_active: new max_active value.
4468 *
4469 * Set max_active of @wq to @max_active.
4470 *
4471 * CONTEXT:
4472 * Don't call from IRQ context.
4473 */
4474void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4475{
4476 struct pool_workqueue *pwq;
4477
4478 /* disallow meddling with max_active for ordered workqueues */
4479 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4480 return;
4481
4482 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4483
4484 mutex_lock(&wq->mutex);
4485
4486 wq->flags &= ~__WQ_ORDERED;
4487 wq->saved_max_active = max_active;
4488
4489 for_each_pwq(pwq, wq)
4490 pwq_adjust_max_active(pwq);
4491
4492 mutex_unlock(&wq->mutex);
4493}
4494EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4495
4496/**
4497 * current_work - retrieve %current task's work struct
4498 *
4499 * Determine if %current task is a workqueue worker and what it's working on.
4500 * Useful to find out the context that the %current task is running in.
4501 *
4502 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4503 */
4504struct work_struct *current_work(void)
4505{
4506 struct worker *worker = current_wq_worker();
4507
4508 return worker ? worker->current_work : NULL;
4509}
4510EXPORT_SYMBOL(current_work);
4511
4512/**
4513 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4514 *
4515 * Determine whether %current is a workqueue rescuer. Can be used from
4516 * work functions to determine whether it's being run off the rescuer task.
4517 *
4518 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4519 */
4520bool current_is_workqueue_rescuer(void)
4521{
4522 struct worker *worker = current_wq_worker();
4523
4524 return worker && worker->rescue_wq;
4525}
4526
4527/**
4528 * workqueue_congested - test whether a workqueue is congested
4529 * @cpu: CPU in question
4530 * @wq: target workqueue
4531 *
4532 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4533 * no synchronization around this function and the test result is
4534 * unreliable and only useful as advisory hints or for debugging.
4535 *
4536 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4537 * Note that both per-cpu and unbound workqueues may be associated with
4538 * multiple pool_workqueues which have separate congested states. A
4539 * workqueue being congested on one CPU doesn't mean the workqueue is also
4540 * contested on other CPUs / NUMA nodes.
4541 *
4542 * Return:
4543 * %true if congested, %false otherwise.
4544 */
4545bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4546{
4547 struct pool_workqueue *pwq;
4548 bool ret;
4549
4550 rcu_read_lock();
4551 preempt_disable();
4552
4553 if (cpu == WORK_CPU_UNBOUND)
4554 cpu = smp_processor_id();
4555
4556 if (!(wq->flags & WQ_UNBOUND))
4557 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4558 else
4559 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4560
4561 ret = !list_empty(&pwq->delayed_works);
4562 preempt_enable();
4563 rcu_read_unlock();
4564
4565 return ret;
4566}
4567EXPORT_SYMBOL_GPL(workqueue_congested);
4568
4569/**
4570 * work_busy - test whether a work is currently pending or running
4571 * @work: the work to be tested
4572 *
4573 * Test whether @work is currently pending or running. There is no
4574 * synchronization around this function and the test result is
4575 * unreliable and only useful as advisory hints or for debugging.
4576 *
4577 * Return:
4578 * OR'd bitmask of WORK_BUSY_* bits.
4579 */
4580unsigned int work_busy(struct work_struct *work)
4581{
4582 struct worker_pool *pool;
4583 unsigned long flags;
4584 unsigned int ret = 0;
4585
4586 if (work_pending(work))
4587 ret |= WORK_BUSY_PENDING;
4588
4589 rcu_read_lock();
4590 pool = get_work_pool(work);
4591 if (pool) {
4592 raw_spin_lock_irqsave(&pool->lock, flags);
4593 if (find_worker_executing_work(pool, work))
4594 ret |= WORK_BUSY_RUNNING;
4595 raw_spin_unlock_irqrestore(&pool->lock, flags);
4596 }
4597 rcu_read_unlock();
4598
4599 return ret;
4600}
4601EXPORT_SYMBOL_GPL(work_busy);
4602
4603/**
4604 * set_worker_desc - set description for the current work item
4605 * @fmt: printf-style format string
4606 * @...: arguments for the format string
4607 *
4608 * This function can be called by a running work function to describe what
4609 * the work item is about. If the worker task gets dumped, this
4610 * information will be printed out together to help debugging. The
4611 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4612 */
4613void set_worker_desc(const char *fmt, ...)
4614{
4615 struct worker *worker = current_wq_worker();
4616 va_list args;
4617
4618 if (worker) {
4619 va_start(args, fmt);
4620 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4621 va_end(args);
4622 }
4623}
4624EXPORT_SYMBOL_GPL(set_worker_desc);
4625
4626/**
4627 * print_worker_info - print out worker information and description
4628 * @log_lvl: the log level to use when printing
4629 * @task: target task
4630 *
4631 * If @task is a worker and currently executing a work item, print out the
4632 * name of the workqueue being serviced and worker description set with
4633 * set_worker_desc() by the currently executing work item.
4634 *
4635 * This function can be safely called on any task as long as the
4636 * task_struct itself is accessible. While safe, this function isn't
4637 * synchronized and may print out mixups or garbages of limited length.
4638 */
4639void print_worker_info(const char *log_lvl, struct task_struct *task)
4640{
4641 work_func_t *fn = NULL;
4642 char name[WQ_NAME_LEN] = { };
4643 char desc[WORKER_DESC_LEN] = { };
4644 struct pool_workqueue *pwq = NULL;
4645 struct workqueue_struct *wq = NULL;
4646 struct worker *worker;
4647
4648 if (!(task->flags & PF_WQ_WORKER))
4649 return;
4650
4651 /*
4652 * This function is called without any synchronization and @task
4653 * could be in any state. Be careful with dereferences.
4654 */
4655 worker = kthread_probe_data(task);
4656
4657 /*
4658 * Carefully copy the associated workqueue's workfn, name and desc.
4659 * Keep the original last '\0' in case the original is garbage.
4660 */
4661 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4662 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4663 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4664 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4665 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4666
4667 if (fn || name[0] || desc[0]) {
4668 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4669 if (strcmp(name, desc))
4670 pr_cont(" (%s)", desc);
4671 pr_cont("\n");
4672 }
4673}
4674
4675static void pr_cont_pool_info(struct worker_pool *pool)
4676{
4677 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4678 if (pool->node != NUMA_NO_NODE)
4679 pr_cont(" node=%d", pool->node);
4680 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4681}
4682
4683static void pr_cont_work(bool comma, struct work_struct *work)
4684{
4685 if (work->func == wq_barrier_func) {
4686 struct wq_barrier *barr;
4687
4688 barr = container_of(work, struct wq_barrier, work);
4689
4690 pr_cont("%s BAR(%d)", comma ? "," : "",
4691 task_pid_nr(barr->task));
4692 } else {
4693 pr_cont("%s %ps", comma ? "," : "", work->func);
4694 }
4695}
4696
4697static void show_pwq(struct pool_workqueue *pwq)
4698{
4699 struct worker_pool *pool = pwq->pool;
4700 struct work_struct *work;
4701 struct worker *worker;
4702 bool has_in_flight = false, has_pending = false;
4703 int bkt;
4704
4705 pr_info(" pwq %d:", pool->id);
4706 pr_cont_pool_info(pool);
4707
4708 pr_cont(" active=%d/%d refcnt=%d%s\n",
4709 pwq->nr_active, pwq->max_active, pwq->refcnt,
4710 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4711
4712 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4713 if (worker->current_pwq == pwq) {
4714 has_in_flight = true;
4715 break;
4716 }
4717 }
4718 if (has_in_flight) {
4719 bool comma = false;
4720
4721 pr_info(" in-flight:");
4722 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4723 if (worker->current_pwq != pwq)
4724 continue;
4725
4726 pr_cont("%s %d%s:%ps", comma ? "," : "",
4727 task_pid_nr(worker->task),
4728 worker->rescue_wq ? "(RESCUER)" : "",
4729 worker->current_func);
4730 list_for_each_entry(work, &worker->scheduled, entry)
4731 pr_cont_work(false, work);
4732 comma = true;
4733 }
4734 pr_cont("\n");
4735 }
4736
4737 list_for_each_entry(work, &pool->worklist, entry) {
4738 if (get_work_pwq(work) == pwq) {
4739 has_pending = true;
4740 break;
4741 }
4742 }
4743 if (has_pending) {
4744 bool comma = false;
4745
4746 pr_info(" pending:");
4747 list_for_each_entry(work, &pool->worklist, entry) {
4748 if (get_work_pwq(work) != pwq)
4749 continue;
4750
4751 pr_cont_work(comma, work);
4752 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4753 }
4754 pr_cont("\n");
4755 }
4756
4757 if (!list_empty(&pwq->delayed_works)) {
4758 bool comma = false;
4759
4760 pr_info(" delayed:");
4761 list_for_each_entry(work, &pwq->delayed_works, entry) {
4762 pr_cont_work(comma, work);
4763 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4764 }
4765 pr_cont("\n");
4766 }
4767}
4768
4769/**
4770 * show_workqueue_state - dump workqueue state
4771 *
4772 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4773 * all busy workqueues and pools.
4774 */
4775void show_workqueue_state(void)
4776{
4777 struct workqueue_struct *wq;
4778 struct worker_pool *pool;
4779 unsigned long flags;
4780 int pi;
4781
4782 rcu_read_lock();
4783
4784 pr_info("Showing busy workqueues and worker pools:\n");
4785
4786 list_for_each_entry_rcu(wq, &workqueues, list) {
4787 struct pool_workqueue *pwq;
4788 bool idle = true;
4789
4790 for_each_pwq(pwq, wq) {
4791 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4792 idle = false;
4793 break;
4794 }
4795 }
4796 if (idle)
4797 continue;
4798
4799 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4800
4801 for_each_pwq(pwq, wq) {
4802 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4803 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4804 show_pwq(pwq);
4805 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4806 /*
4807 * We could be printing a lot from atomic context, e.g.
4808 * sysrq-t -> show_workqueue_state(). Avoid triggering
4809 * hard lockup.
4810 */
4811 touch_nmi_watchdog();
4812 }
4813 }
4814
4815 for_each_pool(pool, pi) {
4816 struct worker *worker;
4817 bool first = true;
4818
4819 raw_spin_lock_irqsave(&pool->lock, flags);
4820 if (pool->nr_workers == pool->nr_idle)
4821 goto next_pool;
4822
4823 pr_info("pool %d:", pool->id);
4824 pr_cont_pool_info(pool);
4825 pr_cont(" hung=%us workers=%d",
4826 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4827 pool->nr_workers);
4828 if (pool->manager)
4829 pr_cont(" manager: %d",
4830 task_pid_nr(pool->manager->task));
4831 list_for_each_entry(worker, &pool->idle_list, entry) {
4832 pr_cont(" %s%d", first ? "idle: " : "",
4833 task_pid_nr(worker->task));
4834 first = false;
4835 }
4836 pr_cont("\n");
4837 next_pool:
4838 raw_spin_unlock_irqrestore(&pool->lock, flags);
4839 /*
4840 * We could be printing a lot from atomic context, e.g.
4841 * sysrq-t -> show_workqueue_state(). Avoid triggering
4842 * hard lockup.
4843 */
4844 touch_nmi_watchdog();
4845 }
4846
4847 rcu_read_unlock();
4848}
4849
4850/* used to show worker information through /proc/PID/{comm,stat,status} */
4851void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4852{
4853 int off;
4854
4855 /* always show the actual comm */
4856 off = strscpy(buf, task->comm, size);
4857 if (off < 0)
4858 return;
4859
4860 /* stabilize PF_WQ_WORKER and worker pool association */
4861 mutex_lock(&wq_pool_attach_mutex);
4862
4863 if (task->flags & PF_WQ_WORKER) {
4864 struct worker *worker = kthread_data(task);
4865 struct worker_pool *pool = worker->pool;
4866
4867 if (pool) {
4868 raw_spin_lock_irq(&pool->lock);
4869 /*
4870 * ->desc tracks information (wq name or
4871 * set_worker_desc()) for the latest execution. If
4872 * current, prepend '+', otherwise '-'.
4873 */
4874 if (worker->desc[0] != '\0') {
4875 if (worker->current_work)
4876 scnprintf(buf + off, size - off, "+%s",
4877 worker->desc);
4878 else
4879 scnprintf(buf + off, size - off, "-%s",
4880 worker->desc);
4881 }
4882 raw_spin_unlock_irq(&pool->lock);
4883 }
4884 }
4885
4886 mutex_unlock(&wq_pool_attach_mutex);
4887}
4888
4889#ifdef CONFIG_SMP
4890
4891/*
4892 * CPU hotplug.
4893 *
4894 * There are two challenges in supporting CPU hotplug. Firstly, there
4895 * are a lot of assumptions on strong associations among work, pwq and
4896 * pool which make migrating pending and scheduled works very
4897 * difficult to implement without impacting hot paths. Secondly,
4898 * worker pools serve mix of short, long and very long running works making
4899 * blocked draining impractical.
4900 *
4901 * This is solved by allowing the pools to be disassociated from the CPU
4902 * running as an unbound one and allowing it to be reattached later if the
4903 * cpu comes back online.
4904 */
4905
4906static void unbind_workers(int cpu)
4907{
4908 struct worker_pool *pool;
4909 struct worker *worker;
4910
4911 for_each_cpu_worker_pool(pool, cpu) {
4912 mutex_lock(&wq_pool_attach_mutex);
4913 raw_spin_lock_irq(&pool->lock);
4914
4915 /*
4916 * We've blocked all attach/detach operations. Make all workers
4917 * unbound and set DISASSOCIATED. Before this, all workers
4918 * except for the ones which are still executing works from
4919 * before the last CPU down must be on the cpu. After
4920 * this, they may become diasporas.
4921 */
4922 for_each_pool_worker(worker, pool)
4923 worker->flags |= WORKER_UNBOUND;
4924
4925 pool->flags |= POOL_DISASSOCIATED;
4926
4927 raw_spin_unlock_irq(&pool->lock);
4928
4929 for_each_pool_worker(worker, pool) {
4930 kthread_set_per_cpu(worker->task, -1);
4931 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
4932 }
4933
4934 mutex_unlock(&wq_pool_attach_mutex);
4935
4936 /*
4937 * Call schedule() so that we cross rq->lock and thus can
4938 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4939 * This is necessary as scheduler callbacks may be invoked
4940 * from other cpus.
4941 */
4942 schedule();
4943
4944 /*
4945 * Sched callbacks are disabled now. Zap nr_running.
4946 * After this, nr_running stays zero and need_more_worker()
4947 * and keep_working() are always true as long as the
4948 * worklist is not empty. This pool now behaves as an
4949 * unbound (in terms of concurrency management) pool which
4950 * are served by workers tied to the pool.
4951 */
4952 atomic_set(&pool->nr_running, 0);
4953
4954 /*
4955 * With concurrency management just turned off, a busy
4956 * worker blocking could lead to lengthy stalls. Kick off
4957 * unbound chain execution of currently pending work items.
4958 */
4959 raw_spin_lock_irq(&pool->lock);
4960 wake_up_worker(pool);
4961 raw_spin_unlock_irq(&pool->lock);
4962 }
4963}
4964
4965/**
4966 * rebind_workers - rebind all workers of a pool to the associated CPU
4967 * @pool: pool of interest
4968 *
4969 * @pool->cpu is coming online. Rebind all workers to the CPU.
4970 */
4971static void rebind_workers(struct worker_pool *pool)
4972{
4973 struct worker *worker;
4974
4975 lockdep_assert_held(&wq_pool_attach_mutex);
4976
4977 /*
4978 * Restore CPU affinity of all workers. As all idle workers should
4979 * be on the run-queue of the associated CPU before any local
4980 * wake-ups for concurrency management happen, restore CPU affinity
4981 * of all workers first and then clear UNBOUND. As we're called
4982 * from CPU_ONLINE, the following shouldn't fail.
4983 */
4984 for_each_pool_worker(worker, pool) {
4985 kthread_set_per_cpu(worker->task, pool->cpu);
4986 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4987 pool->attrs->cpumask) < 0);
4988 }
4989
4990 raw_spin_lock_irq(&pool->lock);
4991
4992 pool->flags &= ~POOL_DISASSOCIATED;
4993
4994 for_each_pool_worker(worker, pool) {
4995 unsigned int worker_flags = worker->flags;
4996
4997 /*
4998 * A bound idle worker should actually be on the runqueue
4999 * of the associated CPU for local wake-ups targeting it to
5000 * work. Kick all idle workers so that they migrate to the
5001 * associated CPU. Doing this in the same loop as
5002 * replacing UNBOUND with REBOUND is safe as no worker will
5003 * be bound before @pool->lock is released.
5004 */
5005 if (worker_flags & WORKER_IDLE)
5006 wake_up_process(worker->task);
5007
5008 /*
5009 * We want to clear UNBOUND but can't directly call
5010 * worker_clr_flags() or adjust nr_running. Atomically
5011 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5012 * @worker will clear REBOUND using worker_clr_flags() when
5013 * it initiates the next execution cycle thus restoring
5014 * concurrency management. Note that when or whether
5015 * @worker clears REBOUND doesn't affect correctness.
5016 *
5017 * WRITE_ONCE() is necessary because @worker->flags may be
5018 * tested without holding any lock in
5019 * wq_worker_running(). Without it, NOT_RUNNING test may
5020 * fail incorrectly leading to premature concurrency
5021 * management operations.
5022 */
5023 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5024 worker_flags |= WORKER_REBOUND;
5025 worker_flags &= ~WORKER_UNBOUND;
5026 WRITE_ONCE(worker->flags, worker_flags);
5027 }
5028
5029 raw_spin_unlock_irq(&pool->lock);
5030}
5031
5032/**
5033 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5034 * @pool: unbound pool of interest
5035 * @cpu: the CPU which is coming up
5036 *
5037 * An unbound pool may end up with a cpumask which doesn't have any online
5038 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5039 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5040 * online CPU before, cpus_allowed of all its workers should be restored.
5041 */
5042static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5043{
5044 static cpumask_t cpumask;
5045 struct worker *worker;
5046
5047 lockdep_assert_held(&wq_pool_attach_mutex);
5048
5049 /* is @cpu allowed for @pool? */
5050 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5051 return;
5052
5053 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5054
5055 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5056 for_each_pool_worker(worker, pool)
5057 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5058}
5059
5060int workqueue_prepare_cpu(unsigned int cpu)
5061{
5062 struct worker_pool *pool;
5063
5064 for_each_cpu_worker_pool(pool, cpu) {
5065 if (pool->nr_workers)
5066 continue;
5067 if (!create_worker(pool))
5068 return -ENOMEM;
5069 }
5070 return 0;
5071}
5072
5073int workqueue_online_cpu(unsigned int cpu)
5074{
5075 struct worker_pool *pool;
5076 struct workqueue_struct *wq;
5077 int pi;
5078
5079 mutex_lock(&wq_pool_mutex);
5080
5081 for_each_pool(pool, pi) {
5082 mutex_lock(&wq_pool_attach_mutex);
5083
5084 if (pool->cpu == cpu)
5085 rebind_workers(pool);
5086 else if (pool->cpu < 0)
5087 restore_unbound_workers_cpumask(pool, cpu);
5088
5089 mutex_unlock(&wq_pool_attach_mutex);
5090 }
5091
5092 /* update NUMA affinity of unbound workqueues */
5093 list_for_each_entry(wq, &workqueues, list)
5094 wq_update_unbound_numa(wq, cpu, true);
5095
5096 mutex_unlock(&wq_pool_mutex);
5097 return 0;
5098}
5099
5100int workqueue_offline_cpu(unsigned int cpu)
5101{
5102 struct workqueue_struct *wq;
5103
5104 /* unbinding per-cpu workers should happen on the local CPU */
5105 if (WARN_ON(cpu != smp_processor_id()))
5106 return -1;
5107
5108 unbind_workers(cpu);
5109
5110 /* update NUMA affinity of unbound workqueues */
5111 mutex_lock(&wq_pool_mutex);
5112 list_for_each_entry(wq, &workqueues, list)
5113 wq_update_unbound_numa(wq, cpu, false);
5114 mutex_unlock(&wq_pool_mutex);
5115
5116 return 0;
5117}
5118
5119struct work_for_cpu {
5120 struct work_struct work;
5121 long (*fn)(void *);
5122 void *arg;
5123 long ret;
5124};
5125
5126static void work_for_cpu_fn(struct work_struct *work)
5127{
5128 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5129
5130 wfc->ret = wfc->fn(wfc->arg);
5131}
5132
5133/**
5134 * work_on_cpu - run a function in thread context on a particular cpu
5135 * @cpu: the cpu to run on
5136 * @fn: the function to run
5137 * @arg: the function arg
5138 *
5139 * It is up to the caller to ensure that the cpu doesn't go offline.
5140 * The caller must not hold any locks which would prevent @fn from completing.
5141 *
5142 * Return: The value @fn returns.
5143 */
5144long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5145{
5146 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5147
5148 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5149 schedule_work_on(cpu, &wfc.work);
5150 flush_work(&wfc.work);
5151 destroy_work_on_stack(&wfc.work);
5152 return wfc.ret;
5153}
5154EXPORT_SYMBOL_GPL(work_on_cpu);
5155
5156/**
5157 * work_on_cpu_safe - run a function in thread context on a particular cpu
5158 * @cpu: the cpu to run on
5159 * @fn: the function to run
5160 * @arg: the function argument
5161 *
5162 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5163 * any locks which would prevent @fn from completing.
5164 *
5165 * Return: The value @fn returns.
5166 */
5167long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5168{
5169 long ret = -ENODEV;
5170
5171 get_online_cpus();
5172 if (cpu_online(cpu))
5173 ret = work_on_cpu(cpu, fn, arg);
5174 put_online_cpus();
5175 return ret;
5176}
5177EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5178#endif /* CONFIG_SMP */
5179
5180#ifdef CONFIG_FREEZER
5181
5182/**
5183 * freeze_workqueues_begin - begin freezing workqueues
5184 *
5185 * Start freezing workqueues. After this function returns, all freezable
5186 * workqueues will queue new works to their delayed_works list instead of
5187 * pool->worklist.
5188 *
5189 * CONTEXT:
5190 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5191 */
5192void freeze_workqueues_begin(void)
5193{
5194 struct workqueue_struct *wq;
5195 struct pool_workqueue *pwq;
5196
5197 mutex_lock(&wq_pool_mutex);
5198
5199 WARN_ON_ONCE(workqueue_freezing);
5200 workqueue_freezing = true;
5201
5202 list_for_each_entry(wq, &workqueues, list) {
5203 mutex_lock(&wq->mutex);
5204 for_each_pwq(pwq, wq)
5205 pwq_adjust_max_active(pwq);
5206 mutex_unlock(&wq->mutex);
5207 }
5208
5209 mutex_unlock(&wq_pool_mutex);
5210}
5211
5212/**
5213 * freeze_workqueues_busy - are freezable workqueues still busy?
5214 *
5215 * Check whether freezing is complete. This function must be called
5216 * between freeze_workqueues_begin() and thaw_workqueues().
5217 *
5218 * CONTEXT:
5219 * Grabs and releases wq_pool_mutex.
5220 *
5221 * Return:
5222 * %true if some freezable workqueues are still busy. %false if freezing
5223 * is complete.
5224 */
5225bool freeze_workqueues_busy(void)
5226{
5227 bool busy = false;
5228 struct workqueue_struct *wq;
5229 struct pool_workqueue *pwq;
5230
5231 mutex_lock(&wq_pool_mutex);
5232
5233 WARN_ON_ONCE(!workqueue_freezing);
5234
5235 list_for_each_entry(wq, &workqueues, list) {
5236 if (!(wq->flags & WQ_FREEZABLE))
5237 continue;
5238 /*
5239 * nr_active is monotonically decreasing. It's safe
5240 * to peek without lock.
5241 */
5242 rcu_read_lock();
5243 for_each_pwq(pwq, wq) {
5244 WARN_ON_ONCE(pwq->nr_active < 0);
5245 if (pwq->nr_active) {
5246 busy = true;
5247 rcu_read_unlock();
5248 goto out_unlock;
5249 }
5250 }
5251 rcu_read_unlock();
5252 }
5253out_unlock:
5254 mutex_unlock(&wq_pool_mutex);
5255 return busy;
5256}
5257
5258/**
5259 * thaw_workqueues - thaw workqueues
5260 *
5261 * Thaw workqueues. Normal queueing is restored and all collected
5262 * frozen works are transferred to their respective pool worklists.
5263 *
5264 * CONTEXT:
5265 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5266 */
5267void thaw_workqueues(void)
5268{
5269 struct workqueue_struct *wq;
5270 struct pool_workqueue *pwq;
5271
5272 mutex_lock(&wq_pool_mutex);
5273
5274 if (!workqueue_freezing)
5275 goto out_unlock;
5276
5277 workqueue_freezing = false;
5278
5279 /* restore max_active and repopulate worklist */
5280 list_for_each_entry(wq, &workqueues, list) {
5281 mutex_lock(&wq->mutex);
5282 for_each_pwq(pwq, wq)
5283 pwq_adjust_max_active(pwq);
5284 mutex_unlock(&wq->mutex);
5285 }
5286
5287out_unlock:
5288 mutex_unlock(&wq_pool_mutex);
5289}
5290#endif /* CONFIG_FREEZER */
5291
5292static int workqueue_apply_unbound_cpumask(void)
5293{
5294 LIST_HEAD(ctxs);
5295 int ret = 0;
5296 struct workqueue_struct *wq;
5297 struct apply_wqattrs_ctx *ctx, *n;
5298
5299 lockdep_assert_held(&wq_pool_mutex);
5300
5301 list_for_each_entry(wq, &workqueues, list) {
5302 if (!(wq->flags & WQ_UNBOUND))
5303 continue;
5304 /* creating multiple pwqs breaks ordering guarantee */
5305 if (wq->flags & __WQ_ORDERED)
5306 continue;
5307
5308 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5309 if (!ctx) {
5310 ret = -ENOMEM;
5311 break;
5312 }
5313
5314 list_add_tail(&ctx->list, &ctxs);
5315 }
5316
5317 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5318 if (!ret)
5319 apply_wqattrs_commit(ctx);
5320 apply_wqattrs_cleanup(ctx);
5321 }
5322
5323 return ret;
5324}
5325
5326/**
5327 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5328 * @cpumask: the cpumask to set
5329 *
5330 * The low-level workqueues cpumask is a global cpumask that limits
5331 * the affinity of all unbound workqueues. This function check the @cpumask
5332 * and apply it to all unbound workqueues and updates all pwqs of them.
5333 *
5334 * Retun: 0 - Success
5335 * -EINVAL - Invalid @cpumask
5336 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5337 */
5338int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5339{
5340 int ret = -EINVAL;
5341 cpumask_var_t saved_cpumask;
5342
5343 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5344 return -ENOMEM;
5345
5346 /*
5347 * Not excluding isolated cpus on purpose.
5348 * If the user wishes to include them, we allow that.
5349 */
5350 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5351 if (!cpumask_empty(cpumask)) {
5352 apply_wqattrs_lock();
5353
5354 /* save the old wq_unbound_cpumask. */
5355 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5356
5357 /* update wq_unbound_cpumask at first and apply it to wqs. */
5358 cpumask_copy(wq_unbound_cpumask, cpumask);
5359 ret = workqueue_apply_unbound_cpumask();
5360
5361 /* restore the wq_unbound_cpumask when failed. */
5362 if (ret < 0)
5363 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5364
5365 apply_wqattrs_unlock();
5366 }
5367
5368 free_cpumask_var(saved_cpumask);
5369 return ret;
5370}
5371
5372#ifdef CONFIG_SYSFS
5373/*
5374 * Workqueues with WQ_SYSFS flag set is visible to userland via
5375 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5376 * following attributes.
5377 *
5378 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5379 * max_active RW int : maximum number of in-flight work items
5380 *
5381 * Unbound workqueues have the following extra attributes.
5382 *
5383 * pool_ids RO int : the associated pool IDs for each node
5384 * nice RW int : nice value of the workers
5385 * cpumask RW mask : bitmask of allowed CPUs for the workers
5386 * numa RW bool : whether enable NUMA affinity
5387 */
5388struct wq_device {
5389 struct workqueue_struct *wq;
5390 struct device dev;
5391};
5392
5393static struct workqueue_struct *dev_to_wq(struct device *dev)
5394{
5395 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5396
5397 return wq_dev->wq;
5398}
5399
5400static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5401 char *buf)
5402{
5403 struct workqueue_struct *wq = dev_to_wq(dev);
5404
5405 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5406}
5407static DEVICE_ATTR_RO(per_cpu);
5408
5409static ssize_t max_active_show(struct device *dev,
5410 struct device_attribute *attr, char *buf)
5411{
5412 struct workqueue_struct *wq = dev_to_wq(dev);
5413
5414 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5415}
5416
5417static ssize_t max_active_store(struct device *dev,
5418 struct device_attribute *attr, const char *buf,
5419 size_t count)
5420{
5421 struct workqueue_struct *wq = dev_to_wq(dev);
5422 int val;
5423
5424 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5425 return -EINVAL;
5426
5427 workqueue_set_max_active(wq, val);
5428 return count;
5429}
5430static DEVICE_ATTR_RW(max_active);
5431
5432static struct attribute *wq_sysfs_attrs[] = {
5433 &dev_attr_per_cpu.attr,
5434 &dev_attr_max_active.attr,
5435 NULL,
5436};
5437ATTRIBUTE_GROUPS(wq_sysfs);
5438
5439static ssize_t wq_pool_ids_show(struct device *dev,
5440 struct device_attribute *attr, char *buf)
5441{
5442 struct workqueue_struct *wq = dev_to_wq(dev);
5443 const char *delim = "";
5444 int node, written = 0;
5445
5446 get_online_cpus();
5447 rcu_read_lock();
5448 for_each_node(node) {
5449 written += scnprintf(buf + written, PAGE_SIZE - written,
5450 "%s%d:%d", delim, node,
5451 unbound_pwq_by_node(wq, node)->pool->id);
5452 delim = " ";
5453 }
5454 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5455 rcu_read_unlock();
5456 put_online_cpus();
5457
5458 return written;
5459}
5460
5461static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5462 char *buf)
5463{
5464 struct workqueue_struct *wq = dev_to_wq(dev);
5465 int written;
5466
5467 mutex_lock(&wq->mutex);
5468 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5469 mutex_unlock(&wq->mutex);
5470
5471 return written;
5472}
5473
5474/* prepare workqueue_attrs for sysfs store operations */
5475static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5476{
5477 struct workqueue_attrs *attrs;
5478
5479 lockdep_assert_held(&wq_pool_mutex);
5480
5481 attrs = alloc_workqueue_attrs();
5482 if (!attrs)
5483 return NULL;
5484
5485 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5486 return attrs;
5487}
5488
5489static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5490 const char *buf, size_t count)
5491{
5492 struct workqueue_struct *wq = dev_to_wq(dev);
5493 struct workqueue_attrs *attrs;
5494 int ret = -ENOMEM;
5495
5496 apply_wqattrs_lock();
5497
5498 attrs = wq_sysfs_prep_attrs(wq);
5499 if (!attrs)
5500 goto out_unlock;
5501
5502 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5503 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5504 ret = apply_workqueue_attrs_locked(wq, attrs);
5505 else
5506 ret = -EINVAL;
5507
5508out_unlock:
5509 apply_wqattrs_unlock();
5510 free_workqueue_attrs(attrs);
5511 return ret ?: count;
5512}
5513
5514static ssize_t wq_cpumask_show(struct device *dev,
5515 struct device_attribute *attr, char *buf)
5516{
5517 struct workqueue_struct *wq = dev_to_wq(dev);
5518 int written;
5519
5520 mutex_lock(&wq->mutex);
5521 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5522 cpumask_pr_args(wq->unbound_attrs->cpumask));
5523 mutex_unlock(&wq->mutex);
5524 return written;
5525}
5526
5527static ssize_t wq_cpumask_store(struct device *dev,
5528 struct device_attribute *attr,
5529 const char *buf, size_t count)
5530{
5531 struct workqueue_struct *wq = dev_to_wq(dev);
5532 struct workqueue_attrs *attrs;
5533 int ret = -ENOMEM;
5534
5535 apply_wqattrs_lock();
5536
5537 attrs = wq_sysfs_prep_attrs(wq);
5538 if (!attrs)
5539 goto out_unlock;
5540
5541 ret = cpumask_parse(buf, attrs->cpumask);
5542 if (!ret)
5543 ret = apply_workqueue_attrs_locked(wq, attrs);
5544
5545out_unlock:
5546 apply_wqattrs_unlock();
5547 free_workqueue_attrs(attrs);
5548 return ret ?: count;
5549}
5550
5551static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5552 char *buf)
5553{
5554 struct workqueue_struct *wq = dev_to_wq(dev);
5555 int written;
5556
5557 mutex_lock(&wq->mutex);
5558 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5559 !wq->unbound_attrs->no_numa);
5560 mutex_unlock(&wq->mutex);
5561
5562 return written;
5563}
5564
5565static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5566 const char *buf, size_t count)
5567{
5568 struct workqueue_struct *wq = dev_to_wq(dev);
5569 struct workqueue_attrs *attrs;
5570 int v, ret = -ENOMEM;
5571
5572 apply_wqattrs_lock();
5573
5574 attrs = wq_sysfs_prep_attrs(wq);
5575 if (!attrs)
5576 goto out_unlock;
5577
5578 ret = -EINVAL;
5579 if (sscanf(buf, "%d", &v) == 1) {
5580 attrs->no_numa = !v;
5581 ret = apply_workqueue_attrs_locked(wq, attrs);
5582 }
5583
5584out_unlock:
5585 apply_wqattrs_unlock();
5586 free_workqueue_attrs(attrs);
5587 return ret ?: count;
5588}
5589
5590static struct device_attribute wq_sysfs_unbound_attrs[] = {
5591 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5592 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5593 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5594 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5595 __ATTR_NULL,
5596};
5597
5598static struct bus_type wq_subsys = {
5599 .name = "workqueue",
5600 .dev_groups = wq_sysfs_groups,
5601};
5602
5603static ssize_t wq_unbound_cpumask_show(struct device *dev,
5604 struct device_attribute *attr, char *buf)
5605{
5606 int written;
5607
5608 mutex_lock(&wq_pool_mutex);
5609 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5610 cpumask_pr_args(wq_unbound_cpumask));
5611 mutex_unlock(&wq_pool_mutex);
5612
5613 return written;
5614}
5615
5616static ssize_t wq_unbound_cpumask_store(struct device *dev,
5617 struct device_attribute *attr, const char *buf, size_t count)
5618{
5619 cpumask_var_t cpumask;
5620 int ret;
5621
5622 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5623 return -ENOMEM;
5624
5625 ret = cpumask_parse(buf, cpumask);
5626 if (!ret)
5627 ret = workqueue_set_unbound_cpumask(cpumask);
5628
5629 free_cpumask_var(cpumask);
5630 return ret ? ret : count;
5631}
5632
5633static struct device_attribute wq_sysfs_cpumask_attr =
5634 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5635 wq_unbound_cpumask_store);
5636
5637static int __init wq_sysfs_init(void)
5638{
5639 int err;
5640
5641 err = subsys_virtual_register(&wq_subsys, NULL);
5642 if (err)
5643 return err;
5644
5645 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5646}
5647core_initcall(wq_sysfs_init);
5648
5649static void wq_device_release(struct device *dev)
5650{
5651 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5652
5653 kfree(wq_dev);
5654}
5655
5656/**
5657 * workqueue_sysfs_register - make a workqueue visible in sysfs
5658 * @wq: the workqueue to register
5659 *
5660 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5661 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5662 * which is the preferred method.
5663 *
5664 * Workqueue user should use this function directly iff it wants to apply
5665 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5666 * apply_workqueue_attrs() may race against userland updating the
5667 * attributes.
5668 *
5669 * Return: 0 on success, -errno on failure.
5670 */
5671int workqueue_sysfs_register(struct workqueue_struct *wq)
5672{
5673 struct wq_device *wq_dev;
5674 int ret;
5675
5676 /*
5677 * Adjusting max_active or creating new pwqs by applying
5678 * attributes breaks ordering guarantee. Disallow exposing ordered
5679 * workqueues.
5680 */
5681 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5682 return -EINVAL;
5683
5684 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5685 if (!wq_dev)
5686 return -ENOMEM;
5687
5688 wq_dev->wq = wq;
5689 wq_dev->dev.bus = &wq_subsys;
5690 wq_dev->dev.release = wq_device_release;
5691 dev_set_name(&wq_dev->dev, "%s", wq->name);
5692
5693 /*
5694 * unbound_attrs are created separately. Suppress uevent until
5695 * everything is ready.
5696 */
5697 dev_set_uevent_suppress(&wq_dev->dev, true);
5698
5699 ret = device_register(&wq_dev->dev);
5700 if (ret) {
5701 put_device(&wq_dev->dev);
5702 wq->wq_dev = NULL;
5703 return ret;
5704 }
5705
5706 if (wq->flags & WQ_UNBOUND) {
5707 struct device_attribute *attr;
5708
5709 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5710 ret = device_create_file(&wq_dev->dev, attr);
5711 if (ret) {
5712 device_unregister(&wq_dev->dev);
5713 wq->wq_dev = NULL;
5714 return ret;
5715 }
5716 }
5717 }
5718
5719 dev_set_uevent_suppress(&wq_dev->dev, false);
5720 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5721 return 0;
5722}
5723
5724/**
5725 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5726 * @wq: the workqueue to unregister
5727 *
5728 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5729 */
5730static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5731{
5732 struct wq_device *wq_dev = wq->wq_dev;
5733
5734 if (!wq->wq_dev)
5735 return;
5736
5737 wq->wq_dev = NULL;
5738 device_unregister(&wq_dev->dev);
5739}
5740#else /* CONFIG_SYSFS */
5741static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5742#endif /* CONFIG_SYSFS */
5743
5744/*
5745 * Workqueue watchdog.
5746 *
5747 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5748 * flush dependency, a concurrency managed work item which stays RUNNING
5749 * indefinitely. Workqueue stalls can be very difficult to debug as the
5750 * usual warning mechanisms don't trigger and internal workqueue state is
5751 * largely opaque.
5752 *
5753 * Workqueue watchdog monitors all worker pools periodically and dumps
5754 * state if some pools failed to make forward progress for a while where
5755 * forward progress is defined as the first item on ->worklist changing.
5756 *
5757 * This mechanism is controlled through the kernel parameter
5758 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5759 * corresponding sysfs parameter file.
5760 */
5761#ifdef CONFIG_WQ_WATCHDOG
5762
5763static unsigned long wq_watchdog_thresh = 30;
5764static struct timer_list wq_watchdog_timer;
5765
5766static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5767static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5768
5769static void wq_watchdog_reset_touched(void)
5770{
5771 int cpu;
5772
5773 wq_watchdog_touched = jiffies;
5774 for_each_possible_cpu(cpu)
5775 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5776}
5777
5778static void wq_watchdog_timer_fn(struct timer_list *unused)
5779{
5780 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5781 bool lockup_detected = false;
5782 unsigned long now = jiffies;
5783 struct worker_pool *pool;
5784 int pi;
5785
5786 if (!thresh)
5787 return;
5788
5789 rcu_read_lock();
5790
5791 for_each_pool(pool, pi) {
5792 unsigned long pool_ts, touched, ts;
5793
5794 if (list_empty(&pool->worklist))
5795 continue;
5796
5797 /*
5798 * If a virtual machine is stopped by the host it can look to
5799 * the watchdog like a stall.
5800 */
5801 kvm_check_and_clear_guest_paused();
5802
5803 /* get the latest of pool and touched timestamps */
5804 if (pool->cpu >= 0)
5805 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
5806 else
5807 touched = READ_ONCE(wq_watchdog_touched);
5808 pool_ts = READ_ONCE(pool->watchdog_ts);
5809
5810 if (time_after(pool_ts, touched))
5811 ts = pool_ts;
5812 else
5813 ts = touched;
5814
5815 /* did we stall? */
5816 if (time_after(now, ts + thresh)) {
5817 lockup_detected = true;
5818 pr_emerg("BUG: workqueue lockup - pool");
5819 pr_cont_pool_info(pool);
5820 pr_cont(" stuck for %us!\n",
5821 jiffies_to_msecs(now - pool_ts) / 1000);
5822 }
5823 }
5824
5825 rcu_read_unlock();
5826
5827 if (lockup_detected)
5828 show_workqueue_state();
5829
5830 wq_watchdog_reset_touched();
5831 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5832}
5833
5834notrace void wq_watchdog_touch(int cpu)
5835{
5836 if (cpu >= 0)
5837 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5838
5839 wq_watchdog_touched = jiffies;
5840}
5841
5842static void wq_watchdog_set_thresh(unsigned long thresh)
5843{
5844 wq_watchdog_thresh = 0;
5845 del_timer_sync(&wq_watchdog_timer);
5846
5847 if (thresh) {
5848 wq_watchdog_thresh = thresh;
5849 wq_watchdog_reset_touched();
5850 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5851 }
5852}
5853
5854static int wq_watchdog_param_set_thresh(const char *val,
5855 const struct kernel_param *kp)
5856{
5857 unsigned long thresh;
5858 int ret;
5859
5860 ret = kstrtoul(val, 0, &thresh);
5861 if (ret)
5862 return ret;
5863
5864 if (system_wq)
5865 wq_watchdog_set_thresh(thresh);
5866 else
5867 wq_watchdog_thresh = thresh;
5868
5869 return 0;
5870}
5871
5872static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5873 .set = wq_watchdog_param_set_thresh,
5874 .get = param_get_ulong,
5875};
5876
5877module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5878 0644);
5879
5880static void wq_watchdog_init(void)
5881{
5882 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5883 wq_watchdog_set_thresh(wq_watchdog_thresh);
5884}
5885
5886#else /* CONFIG_WQ_WATCHDOG */
5887
5888static inline void wq_watchdog_init(void) { }
5889
5890#endif /* CONFIG_WQ_WATCHDOG */
5891
5892static void __init wq_numa_init(void)
5893{
5894 cpumask_var_t *tbl;
5895 int node, cpu;
5896
5897 if (num_possible_nodes() <= 1)
5898 return;
5899
5900 if (wq_disable_numa) {
5901 pr_info("workqueue: NUMA affinity support disabled\n");
5902 return;
5903 }
5904
5905 for_each_possible_cpu(cpu) {
5906 if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5907 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5908 return;
5909 }
5910 }
5911
5912 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5913 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5914
5915 /*
5916 * We want masks of possible CPUs of each node which isn't readily
5917 * available. Build one from cpu_to_node() which should have been
5918 * fully initialized by now.
5919 */
5920 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5921 BUG_ON(!tbl);
5922
5923 for_each_node(node)
5924 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5925 node_online(node) ? node : NUMA_NO_NODE));
5926
5927 for_each_possible_cpu(cpu) {
5928 node = cpu_to_node(cpu);
5929 cpumask_set_cpu(cpu, tbl[node]);
5930 }
5931
5932 wq_numa_possible_cpumask = tbl;
5933 wq_numa_enabled = true;
5934}
5935
5936/**
5937 * workqueue_init_early - early init for workqueue subsystem
5938 *
5939 * This is the first half of two-staged workqueue subsystem initialization
5940 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5941 * idr are up. It sets up all the data structures and system workqueues
5942 * and allows early boot code to create workqueues and queue/cancel work
5943 * items. Actual work item execution starts only after kthreads can be
5944 * created and scheduled right before early initcalls.
5945 */
5946void __init workqueue_init_early(void)
5947{
5948 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5949 int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5950 int i, cpu;
5951
5952 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5953
5954 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5955 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5956
5957 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5958
5959 /* initialize CPU pools */
5960 for_each_possible_cpu(cpu) {
5961 struct worker_pool *pool;
5962
5963 i = 0;
5964 for_each_cpu_worker_pool(pool, cpu) {
5965 BUG_ON(init_worker_pool(pool));
5966 pool->cpu = cpu;
5967 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5968 pool->attrs->nice = std_nice[i++];
5969 pool->node = cpu_to_node(cpu);
5970
5971 /* alloc pool ID */
5972 mutex_lock(&wq_pool_mutex);
5973 BUG_ON(worker_pool_assign_id(pool));
5974 mutex_unlock(&wq_pool_mutex);
5975 }
5976 }
5977
5978 /* create default unbound and ordered wq attrs */
5979 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5980 struct workqueue_attrs *attrs;
5981
5982 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5983 attrs->nice = std_nice[i];
5984 unbound_std_wq_attrs[i] = attrs;
5985
5986 /*
5987 * An ordered wq should have only one pwq as ordering is
5988 * guaranteed by max_active which is enforced by pwqs.
5989 * Turn off NUMA so that dfl_pwq is used for all nodes.
5990 */
5991 BUG_ON(!(attrs = alloc_workqueue_attrs()));
5992 attrs->nice = std_nice[i];
5993 attrs->no_numa = true;
5994 ordered_wq_attrs[i] = attrs;
5995 }
5996
5997 system_wq = alloc_workqueue("events", 0, 0);
5998 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5999 system_long_wq = alloc_workqueue("events_long", 0, 0);
6000 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6001 WQ_UNBOUND_MAX_ACTIVE);
6002 system_freezable_wq = alloc_workqueue("events_freezable",
6003 WQ_FREEZABLE, 0);
6004 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6005 WQ_POWER_EFFICIENT, 0);
6006 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6007 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6008 0);
6009 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6010 !system_unbound_wq || !system_freezable_wq ||
6011 !system_power_efficient_wq ||
6012 !system_freezable_power_efficient_wq);
6013}
6014
6015/**
6016 * workqueue_init - bring workqueue subsystem fully online
6017 *
6018 * This is the latter half of two-staged workqueue subsystem initialization
6019 * and invoked as soon as kthreads can be created and scheduled.
6020 * Workqueues have been created and work items queued on them, but there
6021 * are no kworkers executing the work items yet. Populate the worker pools
6022 * with the initial workers and enable future kworker creations.
6023 */
6024void __init workqueue_init(void)
6025{
6026 struct workqueue_struct *wq;
6027 struct worker_pool *pool;
6028 int cpu, bkt;
6029
6030 /*
6031 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6032 * CPU to node mapping may not be available that early on some
6033 * archs such as power and arm64. As per-cpu pools created
6034 * previously could be missing node hint and unbound pools NUMA
6035 * affinity, fix them up.
6036 *
6037 * Also, while iterating workqueues, create rescuers if requested.
6038 */
6039 wq_numa_init();
6040
6041 mutex_lock(&wq_pool_mutex);
6042
6043 for_each_possible_cpu(cpu) {
6044 for_each_cpu_worker_pool(pool, cpu) {
6045 pool->node = cpu_to_node(cpu);
6046 }
6047 }
6048
6049 list_for_each_entry(wq, &workqueues, list) {
6050 wq_update_unbound_numa(wq, smp_processor_id(), true);
6051 WARN(init_rescuer(wq),
6052 "workqueue: failed to create early rescuer for %s",
6053 wq->name);
6054 }
6055
6056 mutex_unlock(&wq_pool_mutex);
6057
6058 /* create the initial workers */
6059 for_each_online_cpu(cpu) {
6060 for_each_cpu_worker_pool(pool, cpu) {
6061 pool->flags &= ~POOL_DISASSOCIATED;
6062 BUG_ON(!create_worker(pool));
6063 }
6064 }
6065
6066 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6067 BUG_ON(!create_worker(pool));
6068
6069 wq_online = true;
6070 wq_watchdog_init();
6071}