Loading...
1/*
2 * kernel/workqueue.c - generic async execution with shared worker pool
3 *
4 * Copyright (C) 2002 Ingo Molnar
5 *
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
11 *
12 * Made to use alloc_percpu by Christoph Lameter.
13 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
23 *
24 * Please read Documentation/workqueue.txt for details.
25 */
26
27#include <linux/export.h>
28#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
38#include <linux/hardirq.h>
39#include <linux/mempolicy.h>
40#include <linux/freezer.h>
41#include <linux/kallsyms.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
52#include "workqueue_internal.h"
53
54enum {
55 /*
56 * worker_pool flags
57 *
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * is in effect.
62 *
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
66 *
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
70 */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72
73 /* worker flags */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
80
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
83
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
94 (min two ticks) */
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97
98 /*
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
101 */
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
104
105 WQ_NAME_LEN = 24,
106};
107
108/*
109 * Structure fields follow one of the following exclusion rules.
110 *
111 * I: Modifiable by initialization/destruction paths and read-only for
112 * everyone else.
113 *
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
116 *
117 * L: pool->lock protected. Access with pool->lock held.
118 *
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
123 *
124 * A: pool->attach_mutex protected.
125 *
126 * PL: wq_pool_mutex protected.
127 *
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129 *
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131 *
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
134 *
135 * WQ: wq->mutex protected.
136 *
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 *
139 * MD: wq_mayday_lock protected.
140 */
141
142/* struct worker is defined in workqueue_internal.h */
143
144struct worker_pool {
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
150
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
152
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
155
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
158
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
162
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
166
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
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 sched-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 sched-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; /* I: 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 struct lockdep_map lockdep_map;
265#endif
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268 /*
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
272 */
273 struct rcu_head rcu;
274
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279};
280
281static struct kmem_cache *pwq_cache;
282
283static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
285
286static bool wq_disable_numa;
287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289/* see the comment above the definition of WQ_POWER_EFFICIENT */
290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
294
295/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
297
298static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
300
301static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302static bool workqueue_freezing; /* PL: have wqs started freezing? */
303
304/* PL: allowable cpus for unbound wqs and work items */
305static cpumask_var_t wq_unbound_cpumask;
306
307/* CPU where unbound work was last round robin scheduled from this CPU */
308static DEFINE_PER_CPU(int, wq_rr_cpu_last);
309
310/*
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
314 */
315#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316static bool wq_debug_force_rr_cpu = true;
317#else
318static bool wq_debug_force_rr_cpu = false;
319#endif
320module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
321
322/* the per-cpu worker pools */
323static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
324
325static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
326
327/* PL: hash of all unbound pools keyed by pool->attrs */
328static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
329
330/* I: attributes used when instantiating standard unbound pools on demand */
331static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
332
333/* I: attributes used when instantiating ordered pools on demand */
334static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
335
336struct workqueue_struct *system_wq __read_mostly;
337EXPORT_SYMBOL(system_wq);
338struct workqueue_struct *system_highpri_wq __read_mostly;
339EXPORT_SYMBOL_GPL(system_highpri_wq);
340struct workqueue_struct *system_long_wq __read_mostly;
341EXPORT_SYMBOL_GPL(system_long_wq);
342struct workqueue_struct *system_unbound_wq __read_mostly;
343EXPORT_SYMBOL_GPL(system_unbound_wq);
344struct workqueue_struct *system_freezable_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_freezable_wq);
346struct workqueue_struct *system_power_efficient_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_power_efficient_wq);
348struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
350
351static int worker_thread(void *__worker);
352static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
353
354#define CREATE_TRACE_POINTS
355#include <trace/events/workqueue.h>
356
357#define assert_rcu_or_pool_mutex() \
358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU or wq_pool_mutex should be held")
361
362#define assert_rcu_or_wq_mutex(wq) \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq->mutex), \
365 "sched RCU or wq->mutex should be held")
366
367#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex) && \
370 !lockdep_is_held(&wq_pool_mutex), \
371 "sched RCU, wq->mutex or wq_pool_mutex should be held")
372
373#define for_each_cpu_worker_pool(pool, cpu) \
374 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
376 (pool)++)
377
378/**
379 * for_each_pool - iterate through all worker_pools in the system
380 * @pool: iteration cursor
381 * @pi: integer used for iteration
382 *
383 * This must be called either with wq_pool_mutex held or sched RCU read
384 * locked. If the pool needs to be used beyond the locking in effect, the
385 * caller is responsible for guaranteeing that the pool stays online.
386 *
387 * The if/else clause exists only for the lockdep assertion and can be
388 * ignored.
389 */
390#define for_each_pool(pool, pi) \
391 idr_for_each_entry(&worker_pool_idr, pool, pi) \
392 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
393 else
394
395/**
396 * for_each_pool_worker - iterate through all workers of a worker_pool
397 * @worker: iteration cursor
398 * @pool: worker_pool to iterate workers of
399 *
400 * This must be called with @pool->attach_mutex.
401 *
402 * The if/else clause exists only for the lockdep assertion and can be
403 * ignored.
404 */
405#define for_each_pool_worker(worker, pool) \
406 list_for_each_entry((worker), &(pool)->workers, node) \
407 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
408 else
409
410/**
411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412 * @pwq: iteration cursor
413 * @wq: the target workqueue
414 *
415 * This must be called either with wq->mutex held or sched RCU read locked.
416 * If the pwq needs to be used beyond the locking in effect, the caller is
417 * responsible for guaranteeing that the pwq stays online.
418 *
419 * The if/else clause exists only for the lockdep assertion and can be
420 * ignored.
421 */
422#define for_each_pwq(pwq, wq) \
423 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
425 else
426
427#ifdef CONFIG_DEBUG_OBJECTS_WORK
428
429static struct debug_obj_descr work_debug_descr;
430
431static void *work_debug_hint(void *addr)
432{
433 return ((struct work_struct *) addr)->func;
434}
435
436/*
437 * fixup_init is called when:
438 * - an active object is initialized
439 */
440static int work_fixup_init(void *addr, enum debug_obj_state state)
441{
442 struct work_struct *work = addr;
443
444 switch (state) {
445 case ODEBUG_STATE_ACTIVE:
446 cancel_work_sync(work);
447 debug_object_init(work, &work_debug_descr);
448 return 1;
449 default:
450 return 0;
451 }
452}
453
454/*
455 * fixup_activate is called when:
456 * - an active object is activated
457 * - an unknown object is activated (might be a statically initialized object)
458 */
459static int work_fixup_activate(void *addr, enum debug_obj_state state)
460{
461 struct work_struct *work = addr;
462
463 switch (state) {
464
465 case ODEBUG_STATE_NOTAVAILABLE:
466 /*
467 * This is not really a fixup. The work struct was
468 * statically initialized. We just make sure that it
469 * is tracked in the object tracker.
470 */
471 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
472 debug_object_init(work, &work_debug_descr);
473 debug_object_activate(work, &work_debug_descr);
474 return 0;
475 }
476 WARN_ON_ONCE(1);
477 return 0;
478
479 case ODEBUG_STATE_ACTIVE:
480 WARN_ON(1);
481
482 default:
483 return 0;
484 }
485}
486
487/*
488 * fixup_free is called when:
489 * - an active object is freed
490 */
491static int work_fixup_free(void *addr, enum debug_obj_state state)
492{
493 struct work_struct *work = addr;
494
495 switch (state) {
496 case ODEBUG_STATE_ACTIVE:
497 cancel_work_sync(work);
498 debug_object_free(work, &work_debug_descr);
499 return 1;
500 default:
501 return 0;
502 }
503}
504
505static struct debug_obj_descr work_debug_descr = {
506 .name = "work_struct",
507 .debug_hint = work_debug_hint,
508 .fixup_init = work_fixup_init,
509 .fixup_activate = work_fixup_activate,
510 .fixup_free = work_fixup_free,
511};
512
513static inline void debug_work_activate(struct work_struct *work)
514{
515 debug_object_activate(work, &work_debug_descr);
516}
517
518static inline void debug_work_deactivate(struct work_struct *work)
519{
520 debug_object_deactivate(work, &work_debug_descr);
521}
522
523void __init_work(struct work_struct *work, int onstack)
524{
525 if (onstack)
526 debug_object_init_on_stack(work, &work_debug_descr);
527 else
528 debug_object_init(work, &work_debug_descr);
529}
530EXPORT_SYMBOL_GPL(__init_work);
531
532void destroy_work_on_stack(struct work_struct *work)
533{
534 debug_object_free(work, &work_debug_descr);
535}
536EXPORT_SYMBOL_GPL(destroy_work_on_stack);
537
538void destroy_delayed_work_on_stack(struct delayed_work *work)
539{
540 destroy_timer_on_stack(&work->timer);
541 debug_object_free(&work->work, &work_debug_descr);
542}
543EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
544
545#else
546static inline void debug_work_activate(struct work_struct *work) { }
547static inline void debug_work_deactivate(struct work_struct *work) { }
548#endif
549
550/**
551 * worker_pool_assign_id - allocate ID and assing it to @pool
552 * @pool: the pool pointer of interest
553 *
554 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
555 * successfully, -errno on failure.
556 */
557static int worker_pool_assign_id(struct worker_pool *pool)
558{
559 int ret;
560
561 lockdep_assert_held(&wq_pool_mutex);
562
563 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
564 GFP_KERNEL);
565 if (ret >= 0) {
566 pool->id = ret;
567 return 0;
568 }
569 return ret;
570}
571
572/**
573 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
574 * @wq: the target workqueue
575 * @node: the node ID
576 *
577 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
578 * read locked.
579 * If the pwq needs to be used beyond the locking in effect, the caller is
580 * responsible for guaranteeing that the pwq stays online.
581 *
582 * Return: The unbound pool_workqueue for @node.
583 */
584static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
585 int node)
586{
587 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
588
589 /*
590 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
591 * delayed item is pending. The plan is to keep CPU -> NODE
592 * mapping valid and stable across CPU on/offlines. Once that
593 * happens, this workaround can be removed.
594 */
595 if (unlikely(node == NUMA_NO_NODE))
596 return wq->dfl_pwq;
597
598 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
599}
600
601static unsigned int work_color_to_flags(int color)
602{
603 return color << WORK_STRUCT_COLOR_SHIFT;
604}
605
606static int get_work_color(struct work_struct *work)
607{
608 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
609 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
610}
611
612static int work_next_color(int color)
613{
614 return (color + 1) % WORK_NR_COLORS;
615}
616
617/*
618 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
619 * contain the pointer to the queued pwq. Once execution starts, the flag
620 * is cleared and the high bits contain OFFQ flags and pool ID.
621 *
622 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
623 * and clear_work_data() can be used to set the pwq, pool or clear
624 * work->data. These functions should only be called while the work is
625 * owned - ie. while the PENDING bit is set.
626 *
627 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
628 * corresponding to a work. Pool is available once the work has been
629 * queued anywhere after initialization until it is sync canceled. pwq is
630 * available only while the work item is queued.
631 *
632 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
633 * canceled. While being canceled, a work item may have its PENDING set
634 * but stay off timer and worklist for arbitrarily long and nobody should
635 * try to steal the PENDING bit.
636 */
637static inline void set_work_data(struct work_struct *work, unsigned long data,
638 unsigned long flags)
639{
640 WARN_ON_ONCE(!work_pending(work));
641 atomic_long_set(&work->data, data | flags | work_static(work));
642}
643
644static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
645 unsigned long extra_flags)
646{
647 set_work_data(work, (unsigned long)pwq,
648 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
649}
650
651static void set_work_pool_and_keep_pending(struct work_struct *work,
652 int pool_id)
653{
654 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
655 WORK_STRUCT_PENDING);
656}
657
658static void set_work_pool_and_clear_pending(struct work_struct *work,
659 int pool_id)
660{
661 /*
662 * The following wmb is paired with the implied mb in
663 * test_and_set_bit(PENDING) and ensures all updates to @work made
664 * here are visible to and precede any updates by the next PENDING
665 * owner.
666 */
667 smp_wmb();
668 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
669 /*
670 * The following mb guarantees that previous clear of a PENDING bit
671 * will not be reordered with any speculative LOADS or STORES from
672 * work->current_func, which is executed afterwards. This possible
673 * reordering can lead to a missed execution on attempt to qeueue
674 * the same @work. E.g. consider this case:
675 *
676 * CPU#0 CPU#1
677 * ---------------------------- --------------------------------
678 *
679 * 1 STORE event_indicated
680 * 2 queue_work_on() {
681 * 3 test_and_set_bit(PENDING)
682 * 4 } set_..._and_clear_pending() {
683 * 5 set_work_data() # clear bit
684 * 6 smp_mb()
685 * 7 work->current_func() {
686 * 8 LOAD event_indicated
687 * }
688 *
689 * Without an explicit full barrier speculative LOAD on line 8 can
690 * be executed before CPU#0 does STORE on line 1. If that happens,
691 * CPU#0 observes the PENDING bit is still set and new execution of
692 * a @work is not queued in a hope, that CPU#1 will eventually
693 * finish the queued @work. Meanwhile CPU#1 does not see
694 * event_indicated is set, because speculative LOAD was executed
695 * before actual STORE.
696 */
697 smp_mb();
698}
699
700static void clear_work_data(struct work_struct *work)
701{
702 smp_wmb(); /* see set_work_pool_and_clear_pending() */
703 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
704}
705
706static struct pool_workqueue *get_work_pwq(struct work_struct *work)
707{
708 unsigned long data = atomic_long_read(&work->data);
709
710 if (data & WORK_STRUCT_PWQ)
711 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
712 else
713 return NULL;
714}
715
716/**
717 * get_work_pool - return the worker_pool a given work was associated with
718 * @work: the work item of interest
719 *
720 * Pools are created and destroyed under wq_pool_mutex, and allows read
721 * access under sched-RCU read lock. As such, this function should be
722 * called under wq_pool_mutex or with preemption disabled.
723 *
724 * All fields of the returned pool are accessible as long as the above
725 * mentioned locking is in effect. If the returned pool needs to be used
726 * beyond the critical section, the caller is responsible for ensuring the
727 * returned pool is and stays online.
728 *
729 * Return: The worker_pool @work was last associated with. %NULL if none.
730 */
731static struct worker_pool *get_work_pool(struct work_struct *work)
732{
733 unsigned long data = atomic_long_read(&work->data);
734 int pool_id;
735
736 assert_rcu_or_pool_mutex();
737
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
741
742 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
743 if (pool_id == WORK_OFFQ_POOL_NONE)
744 return NULL;
745
746 return idr_find(&worker_pool_idr, pool_id);
747}
748
749/**
750 * get_work_pool_id - return the worker pool ID a given work is associated with
751 * @work: the work item of interest
752 *
753 * Return: The worker_pool ID @work was last associated with.
754 * %WORK_OFFQ_POOL_NONE if none.
755 */
756static int get_work_pool_id(struct work_struct *work)
757{
758 unsigned long data = atomic_long_read(&work->data);
759
760 if (data & WORK_STRUCT_PWQ)
761 return ((struct pool_workqueue *)
762 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
763
764 return data >> WORK_OFFQ_POOL_SHIFT;
765}
766
767static void mark_work_canceling(struct work_struct *work)
768{
769 unsigned long pool_id = get_work_pool_id(work);
770
771 pool_id <<= WORK_OFFQ_POOL_SHIFT;
772 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
773}
774
775static bool work_is_canceling(struct work_struct *work)
776{
777 unsigned long data = atomic_long_read(&work->data);
778
779 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
780}
781
782/*
783 * Policy functions. These define the policies on how the global worker
784 * pools are managed. Unless noted otherwise, these functions assume that
785 * they're being called with pool->lock held.
786 */
787
788static bool __need_more_worker(struct worker_pool *pool)
789{
790 return !atomic_read(&pool->nr_running);
791}
792
793/*
794 * Need to wake up a worker? Called from anything but currently
795 * running workers.
796 *
797 * Note that, because unbound workers never contribute to nr_running, this
798 * function will always return %true for unbound pools as long as the
799 * worklist isn't empty.
800 */
801static bool need_more_worker(struct worker_pool *pool)
802{
803 return !list_empty(&pool->worklist) && __need_more_worker(pool);
804}
805
806/* Can I start working? Called from busy but !running workers. */
807static bool may_start_working(struct worker_pool *pool)
808{
809 return pool->nr_idle;
810}
811
812/* Do I need to keep working? Called from currently running workers. */
813static bool keep_working(struct worker_pool *pool)
814{
815 return !list_empty(&pool->worklist) &&
816 atomic_read(&pool->nr_running) <= 1;
817}
818
819/* Do we need a new worker? Called from manager. */
820static bool need_to_create_worker(struct worker_pool *pool)
821{
822 return need_more_worker(pool) && !may_start_working(pool);
823}
824
825/* Do we have too many workers and should some go away? */
826static bool too_many_workers(struct worker_pool *pool)
827{
828 bool managing = mutex_is_locked(&pool->manager_arb);
829 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
830 int nr_busy = pool->nr_workers - nr_idle;
831
832 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
833}
834
835/*
836 * Wake up functions.
837 */
838
839/* Return the first idle worker. Safe with preemption disabled */
840static struct worker *first_idle_worker(struct worker_pool *pool)
841{
842 if (unlikely(list_empty(&pool->idle_list)))
843 return NULL;
844
845 return list_first_entry(&pool->idle_list, struct worker, entry);
846}
847
848/**
849 * wake_up_worker - wake up an idle worker
850 * @pool: worker pool to wake worker from
851 *
852 * Wake up the first idle worker of @pool.
853 *
854 * CONTEXT:
855 * spin_lock_irq(pool->lock).
856 */
857static void wake_up_worker(struct worker_pool *pool)
858{
859 struct worker *worker = first_idle_worker(pool);
860
861 if (likely(worker))
862 wake_up_process(worker->task);
863}
864
865/**
866 * wq_worker_waking_up - a worker is waking up
867 * @task: task waking up
868 * @cpu: CPU @task is waking up to
869 *
870 * This function is called during try_to_wake_up() when a worker is
871 * being awoken.
872 *
873 * CONTEXT:
874 * spin_lock_irq(rq->lock)
875 */
876void wq_worker_waking_up(struct task_struct *task, int cpu)
877{
878 struct worker *worker = kthread_data(task);
879
880 if (!(worker->flags & WORKER_NOT_RUNNING)) {
881 WARN_ON_ONCE(worker->pool->cpu != cpu);
882 atomic_inc(&worker->pool->nr_running);
883 }
884}
885
886/**
887 * wq_worker_sleeping - a worker is going to sleep
888 * @task: task going to sleep
889 *
890 * This function is called during schedule() when a busy worker is
891 * going to sleep. Worker on the same cpu can be woken up by
892 * returning pointer to its task.
893 *
894 * CONTEXT:
895 * spin_lock_irq(rq->lock)
896 *
897 * Return:
898 * Worker task on @cpu to wake up, %NULL if none.
899 */
900struct task_struct *wq_worker_sleeping(struct task_struct *task)
901{
902 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
903 struct worker_pool *pool;
904
905 /*
906 * Rescuers, which may not have all the fields set up like normal
907 * workers, also reach here, let's not access anything before
908 * checking NOT_RUNNING.
909 */
910 if (worker->flags & WORKER_NOT_RUNNING)
911 return NULL;
912
913 pool = worker->pool;
914
915 /* this can only happen on the local cpu */
916 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
917 return NULL;
918
919 /*
920 * The counterpart of the following dec_and_test, implied mb,
921 * worklist not empty test sequence is in insert_work().
922 * Please read comment there.
923 *
924 * NOT_RUNNING is clear. This means that we're bound to and
925 * running on the local cpu w/ rq lock held and preemption
926 * disabled, which in turn means that none else could be
927 * manipulating idle_list, so dereferencing idle_list without pool
928 * lock is safe.
929 */
930 if (atomic_dec_and_test(&pool->nr_running) &&
931 !list_empty(&pool->worklist))
932 to_wakeup = first_idle_worker(pool);
933 return to_wakeup ? to_wakeup->task : NULL;
934}
935
936/**
937 * worker_set_flags - set worker flags and adjust nr_running accordingly
938 * @worker: self
939 * @flags: flags to set
940 *
941 * Set @flags in @worker->flags and adjust nr_running accordingly.
942 *
943 * CONTEXT:
944 * spin_lock_irq(pool->lock)
945 */
946static inline void worker_set_flags(struct worker *worker, unsigned int flags)
947{
948 struct worker_pool *pool = worker->pool;
949
950 WARN_ON_ONCE(worker->task != current);
951
952 /* If transitioning into NOT_RUNNING, adjust nr_running. */
953 if ((flags & WORKER_NOT_RUNNING) &&
954 !(worker->flags & WORKER_NOT_RUNNING)) {
955 atomic_dec(&pool->nr_running);
956 }
957
958 worker->flags |= flags;
959}
960
961/**
962 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
963 * @worker: self
964 * @flags: flags to clear
965 *
966 * Clear @flags in @worker->flags and adjust nr_running accordingly.
967 *
968 * CONTEXT:
969 * spin_lock_irq(pool->lock)
970 */
971static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
972{
973 struct worker_pool *pool = worker->pool;
974 unsigned int oflags = worker->flags;
975
976 WARN_ON_ONCE(worker->task != current);
977
978 worker->flags &= ~flags;
979
980 /*
981 * If transitioning out of NOT_RUNNING, increment nr_running. Note
982 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
983 * of multiple flags, not a single flag.
984 */
985 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
986 if (!(worker->flags & WORKER_NOT_RUNNING))
987 atomic_inc(&pool->nr_running);
988}
989
990/**
991 * find_worker_executing_work - find worker which is executing a work
992 * @pool: pool of interest
993 * @work: work to find worker for
994 *
995 * Find a worker which is executing @work on @pool by searching
996 * @pool->busy_hash which is keyed by the address of @work. For a worker
997 * to match, its current execution should match the address of @work and
998 * its work function. This is to avoid unwanted dependency between
999 * unrelated work executions through a work item being recycled while still
1000 * being executed.
1001 *
1002 * This is a bit tricky. A work item may be freed once its execution
1003 * starts and nothing prevents the freed area from being recycled for
1004 * another work item. If the same work item address ends up being reused
1005 * before the original execution finishes, workqueue will identify the
1006 * recycled work item as currently executing and make it wait until the
1007 * current execution finishes, introducing an unwanted dependency.
1008 *
1009 * This function checks the work item address and work function to avoid
1010 * false positives. Note that this isn't complete as one may construct a
1011 * work function which can introduce dependency onto itself through a
1012 * recycled work item. Well, if somebody wants to shoot oneself in the
1013 * foot that badly, there's only so much we can do, and if such deadlock
1014 * actually occurs, it should be easy to locate the culprit work function.
1015 *
1016 * CONTEXT:
1017 * spin_lock_irq(pool->lock).
1018 *
1019 * Return:
1020 * Pointer to worker which is executing @work if found, %NULL
1021 * otherwise.
1022 */
1023static struct worker *find_worker_executing_work(struct worker_pool *pool,
1024 struct work_struct *work)
1025{
1026 struct worker *worker;
1027
1028 hash_for_each_possible(pool->busy_hash, worker, hentry,
1029 (unsigned long)work)
1030 if (worker->current_work == work &&
1031 worker->current_func == work->func)
1032 return worker;
1033
1034 return NULL;
1035}
1036
1037/**
1038 * move_linked_works - move linked works to a list
1039 * @work: start of series of works to be scheduled
1040 * @head: target list to append @work to
1041 * @nextp: out parameter for nested worklist walking
1042 *
1043 * Schedule linked works starting from @work to @head. Work series to
1044 * be scheduled starts at @work and includes any consecutive work with
1045 * WORK_STRUCT_LINKED set in its predecessor.
1046 *
1047 * If @nextp is not NULL, it's updated to point to the next work of
1048 * the last scheduled work. This allows move_linked_works() to be
1049 * nested inside outer list_for_each_entry_safe().
1050 *
1051 * CONTEXT:
1052 * spin_lock_irq(pool->lock).
1053 */
1054static void move_linked_works(struct work_struct *work, struct list_head *head,
1055 struct work_struct **nextp)
1056{
1057 struct work_struct *n;
1058
1059 /*
1060 * Linked worklist will always end before the end of the list,
1061 * use NULL for list head.
1062 */
1063 list_for_each_entry_safe_from(work, n, NULL, entry) {
1064 list_move_tail(&work->entry, head);
1065 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1066 break;
1067 }
1068
1069 /*
1070 * If we're already inside safe list traversal and have moved
1071 * multiple works to the scheduled queue, the next position
1072 * needs to be updated.
1073 */
1074 if (nextp)
1075 *nextp = n;
1076}
1077
1078/**
1079 * get_pwq - get an extra reference on the specified pool_workqueue
1080 * @pwq: pool_workqueue to get
1081 *
1082 * Obtain an extra reference on @pwq. The caller should guarantee that
1083 * @pwq has positive refcnt and be holding the matching pool->lock.
1084 */
1085static void get_pwq(struct pool_workqueue *pwq)
1086{
1087 lockdep_assert_held(&pwq->pool->lock);
1088 WARN_ON_ONCE(pwq->refcnt <= 0);
1089 pwq->refcnt++;
1090}
1091
1092/**
1093 * put_pwq - put a pool_workqueue reference
1094 * @pwq: pool_workqueue to put
1095 *
1096 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1097 * destruction. The caller should be holding the matching pool->lock.
1098 */
1099static void put_pwq(struct pool_workqueue *pwq)
1100{
1101 lockdep_assert_held(&pwq->pool->lock);
1102 if (likely(--pwq->refcnt))
1103 return;
1104 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105 return;
1106 /*
1107 * @pwq can't be released under pool->lock, bounce to
1108 * pwq_unbound_release_workfn(). This never recurses on the same
1109 * pool->lock as this path is taken only for unbound workqueues and
1110 * the release work item is scheduled on a per-cpu workqueue. To
1111 * avoid lockdep warning, unbound pool->locks are given lockdep
1112 * subclass of 1 in get_unbound_pool().
1113 */
1114 schedule_work(&pwq->unbound_release_work);
1115}
1116
1117/**
1118 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1119 * @pwq: pool_workqueue to put (can be %NULL)
1120 *
1121 * put_pwq() with locking. This function also allows %NULL @pwq.
1122 */
1123static void put_pwq_unlocked(struct pool_workqueue *pwq)
1124{
1125 if (pwq) {
1126 /*
1127 * As both pwqs and pools are sched-RCU protected, the
1128 * following lock operations are safe.
1129 */
1130 spin_lock_irq(&pwq->pool->lock);
1131 put_pwq(pwq);
1132 spin_unlock_irq(&pwq->pool->lock);
1133 }
1134}
1135
1136static void pwq_activate_delayed_work(struct work_struct *work)
1137{
1138 struct pool_workqueue *pwq = get_work_pwq(work);
1139
1140 trace_workqueue_activate_work(work);
1141 if (list_empty(&pwq->pool->worklist))
1142 pwq->pool->watchdog_ts = jiffies;
1143 move_linked_works(work, &pwq->pool->worklist, NULL);
1144 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1145 pwq->nr_active++;
1146}
1147
1148static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1149{
1150 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1151 struct work_struct, entry);
1152
1153 pwq_activate_delayed_work(work);
1154}
1155
1156/**
1157 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1158 * @pwq: pwq of interest
1159 * @color: color of work which left the queue
1160 *
1161 * A work either has completed or is removed from pending queue,
1162 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1163 *
1164 * CONTEXT:
1165 * spin_lock_irq(pool->lock).
1166 */
1167static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1168{
1169 /* uncolored work items don't participate in flushing or nr_active */
1170 if (color == WORK_NO_COLOR)
1171 goto out_put;
1172
1173 pwq->nr_in_flight[color]--;
1174
1175 pwq->nr_active--;
1176 if (!list_empty(&pwq->delayed_works)) {
1177 /* one down, submit a delayed one */
1178 if (pwq->nr_active < pwq->max_active)
1179 pwq_activate_first_delayed(pwq);
1180 }
1181
1182 /* is flush in progress and are we at the flushing tip? */
1183 if (likely(pwq->flush_color != color))
1184 goto out_put;
1185
1186 /* are there still in-flight works? */
1187 if (pwq->nr_in_flight[color])
1188 goto out_put;
1189
1190 /* this pwq is done, clear flush_color */
1191 pwq->flush_color = -1;
1192
1193 /*
1194 * If this was the last pwq, wake up the first flusher. It
1195 * will handle the rest.
1196 */
1197 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1198 complete(&pwq->wq->first_flusher->done);
1199out_put:
1200 put_pwq(pwq);
1201}
1202
1203/**
1204 * try_to_grab_pending - steal work item from worklist and disable irq
1205 * @work: work item to steal
1206 * @is_dwork: @work is a delayed_work
1207 * @flags: place to store irq state
1208 *
1209 * Try to grab PENDING bit of @work. This function can handle @work in any
1210 * stable state - idle, on timer or on worklist.
1211 *
1212 * Return:
1213 * 1 if @work was pending and we successfully stole PENDING
1214 * 0 if @work was idle and we claimed PENDING
1215 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1216 * -ENOENT if someone else is canceling @work, this state may persist
1217 * for arbitrarily long
1218 *
1219 * Note:
1220 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1221 * interrupted while holding PENDING and @work off queue, irq must be
1222 * disabled on entry. This, combined with delayed_work->timer being
1223 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1224 *
1225 * On successful return, >= 0, irq is disabled and the caller is
1226 * responsible for releasing it using local_irq_restore(*@flags).
1227 *
1228 * This function is safe to call from any context including IRQ handler.
1229 */
1230static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1231 unsigned long *flags)
1232{
1233 struct worker_pool *pool;
1234 struct pool_workqueue *pwq;
1235
1236 local_irq_save(*flags);
1237
1238 /* try to steal the timer if it exists */
1239 if (is_dwork) {
1240 struct delayed_work *dwork = to_delayed_work(work);
1241
1242 /*
1243 * dwork->timer is irqsafe. If del_timer() fails, it's
1244 * guaranteed that the timer is not queued anywhere and not
1245 * running on the local CPU.
1246 */
1247 if (likely(del_timer(&dwork->timer)))
1248 return 1;
1249 }
1250
1251 /* try to claim PENDING the normal way */
1252 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1253 return 0;
1254
1255 /*
1256 * The queueing is in progress, or it is already queued. Try to
1257 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258 */
1259 pool = get_work_pool(work);
1260 if (!pool)
1261 goto fail;
1262
1263 spin_lock(&pool->lock);
1264 /*
1265 * work->data is guaranteed to point to pwq only while the work
1266 * item is queued on pwq->wq, and both updating work->data to point
1267 * to pwq on queueing and to pool on dequeueing are done under
1268 * pwq->pool->lock. This in turn guarantees that, if work->data
1269 * points to pwq which is associated with a locked pool, the work
1270 * item is currently queued on that pool.
1271 */
1272 pwq = get_work_pwq(work);
1273 if (pwq && pwq->pool == pool) {
1274 debug_work_deactivate(work);
1275
1276 /*
1277 * A delayed work item cannot be grabbed directly because
1278 * it might have linked NO_COLOR work items which, if left
1279 * on the delayed_list, will confuse pwq->nr_active
1280 * management later on and cause stall. Make sure the work
1281 * item is activated before grabbing.
1282 */
1283 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1284 pwq_activate_delayed_work(work);
1285
1286 list_del_init(&work->entry);
1287 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1288
1289 /* work->data points to pwq iff queued, point to pool */
1290 set_work_pool_and_keep_pending(work, pool->id);
1291
1292 spin_unlock(&pool->lock);
1293 return 1;
1294 }
1295 spin_unlock(&pool->lock);
1296fail:
1297 local_irq_restore(*flags);
1298 if (work_is_canceling(work))
1299 return -ENOENT;
1300 cpu_relax();
1301 return -EAGAIN;
1302}
1303
1304/**
1305 * insert_work - insert a work into a pool
1306 * @pwq: pwq @work belongs to
1307 * @work: work to insert
1308 * @head: insertion point
1309 * @extra_flags: extra WORK_STRUCT_* flags to set
1310 *
1311 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1312 * work_struct flags.
1313 *
1314 * CONTEXT:
1315 * spin_lock_irq(pool->lock).
1316 */
1317static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1318 struct list_head *head, unsigned int extra_flags)
1319{
1320 struct worker_pool *pool = pwq->pool;
1321
1322 /* we own @work, set data and link */
1323 set_work_pwq(work, pwq, extra_flags);
1324 list_add_tail(&work->entry, head);
1325 get_pwq(pwq);
1326
1327 /*
1328 * Ensure either wq_worker_sleeping() sees the above
1329 * list_add_tail() or we see zero nr_running to avoid workers lying
1330 * around lazily while there are works to be processed.
1331 */
1332 smp_mb();
1333
1334 if (__need_more_worker(pool))
1335 wake_up_worker(pool);
1336}
1337
1338/*
1339 * Test whether @work is being queued from another work executing on the
1340 * same workqueue.
1341 */
1342static bool is_chained_work(struct workqueue_struct *wq)
1343{
1344 struct worker *worker;
1345
1346 worker = current_wq_worker();
1347 /*
1348 * Return %true iff I'm a worker execuing a work item on @wq. If
1349 * I'm @worker, it's safe to dereference it without locking.
1350 */
1351 return worker && worker->current_pwq->wq == wq;
1352}
1353
1354/*
1355 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1356 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1357 * avoid perturbing sensitive tasks.
1358 */
1359static int wq_select_unbound_cpu(int cpu)
1360{
1361 static bool printed_dbg_warning;
1362 int new_cpu;
1363
1364 if (likely(!wq_debug_force_rr_cpu)) {
1365 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1366 return cpu;
1367 } else if (!printed_dbg_warning) {
1368 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1369 printed_dbg_warning = true;
1370 }
1371
1372 if (cpumask_empty(wq_unbound_cpumask))
1373 return cpu;
1374
1375 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1376 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1377 if (unlikely(new_cpu >= nr_cpu_ids)) {
1378 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1379 if (unlikely(new_cpu >= nr_cpu_ids))
1380 return cpu;
1381 }
1382 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1383
1384 return new_cpu;
1385}
1386
1387static void __queue_work(int cpu, struct workqueue_struct *wq,
1388 struct work_struct *work)
1389{
1390 struct pool_workqueue *pwq;
1391 struct worker_pool *last_pool;
1392 struct list_head *worklist;
1393 unsigned int work_flags;
1394 unsigned int req_cpu = cpu;
1395
1396 /*
1397 * While a work item is PENDING && off queue, a task trying to
1398 * steal the PENDING will busy-loop waiting for it to either get
1399 * queued or lose PENDING. Grabbing PENDING and queueing should
1400 * happen with IRQ disabled.
1401 */
1402 WARN_ON_ONCE(!irqs_disabled());
1403
1404 debug_work_activate(work);
1405
1406 /* if draining, only works from the same workqueue are allowed */
1407 if (unlikely(wq->flags & __WQ_DRAINING) &&
1408 WARN_ON_ONCE(!is_chained_work(wq)))
1409 return;
1410retry:
1411 if (req_cpu == WORK_CPU_UNBOUND)
1412 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1413
1414 /* pwq which will be used unless @work is executing elsewhere */
1415 if (!(wq->flags & WQ_UNBOUND))
1416 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1417 else
1418 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419
1420 /*
1421 * If @work was previously on a different pool, it might still be
1422 * running there, in which case the work needs to be queued on that
1423 * pool to guarantee non-reentrancy.
1424 */
1425 last_pool = get_work_pool(work);
1426 if (last_pool && last_pool != pwq->pool) {
1427 struct worker *worker;
1428
1429 spin_lock(&last_pool->lock);
1430
1431 worker = find_worker_executing_work(last_pool, work);
1432
1433 if (worker && worker->current_pwq->wq == wq) {
1434 pwq = worker->current_pwq;
1435 } else {
1436 /* meh... not running there, queue here */
1437 spin_unlock(&last_pool->lock);
1438 spin_lock(&pwq->pool->lock);
1439 }
1440 } else {
1441 spin_lock(&pwq->pool->lock);
1442 }
1443
1444 /*
1445 * pwq is determined and locked. For unbound pools, we could have
1446 * raced with pwq release and it could already be dead. If its
1447 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1448 * without another pwq replacing it in the numa_pwq_tbl or while
1449 * work items are executing on it, so the retrying is guaranteed to
1450 * make forward-progress.
1451 */
1452 if (unlikely(!pwq->refcnt)) {
1453 if (wq->flags & WQ_UNBOUND) {
1454 spin_unlock(&pwq->pool->lock);
1455 cpu_relax();
1456 goto retry;
1457 }
1458 /* oops */
1459 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1460 wq->name, cpu);
1461 }
1462
1463 /* pwq determined, queue */
1464 trace_workqueue_queue_work(req_cpu, pwq, work);
1465
1466 if (WARN_ON(!list_empty(&work->entry))) {
1467 spin_unlock(&pwq->pool->lock);
1468 return;
1469 }
1470
1471 pwq->nr_in_flight[pwq->work_color]++;
1472 work_flags = work_color_to_flags(pwq->work_color);
1473
1474 if (likely(pwq->nr_active < pwq->max_active)) {
1475 trace_workqueue_activate_work(work);
1476 pwq->nr_active++;
1477 worklist = &pwq->pool->worklist;
1478 if (list_empty(worklist))
1479 pwq->pool->watchdog_ts = jiffies;
1480 } else {
1481 work_flags |= WORK_STRUCT_DELAYED;
1482 worklist = &pwq->delayed_works;
1483 }
1484
1485 insert_work(pwq, work, worklist, work_flags);
1486
1487 spin_unlock(&pwq->pool->lock);
1488}
1489
1490/**
1491 * queue_work_on - queue work on specific cpu
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @work: work to queue
1495 *
1496 * We queue the work to a specific CPU, the caller must ensure it
1497 * can't go away.
1498 *
1499 * Return: %false if @work was already on a queue, %true otherwise.
1500 */
1501bool queue_work_on(int cpu, struct workqueue_struct *wq,
1502 struct work_struct *work)
1503{
1504 bool ret = false;
1505 unsigned long flags;
1506
1507 local_irq_save(flags);
1508
1509 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1510 __queue_work(cpu, wq, work);
1511 ret = true;
1512 }
1513
1514 local_irq_restore(flags);
1515 return ret;
1516}
1517EXPORT_SYMBOL(queue_work_on);
1518
1519void delayed_work_timer_fn(unsigned long __data)
1520{
1521 struct delayed_work *dwork = (struct delayed_work *)__data;
1522
1523 /* should have been called from irqsafe timer with irq already off */
1524 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1525}
1526EXPORT_SYMBOL(delayed_work_timer_fn);
1527
1528static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1529 struct delayed_work *dwork, unsigned long delay)
1530{
1531 struct timer_list *timer = &dwork->timer;
1532 struct work_struct *work = &dwork->work;
1533
1534 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1535 timer->data != (unsigned long)dwork);
1536 WARN_ON_ONCE(timer_pending(timer));
1537 WARN_ON_ONCE(!list_empty(&work->entry));
1538
1539 /*
1540 * If @delay is 0, queue @dwork->work immediately. This is for
1541 * both optimization and correctness. The earliest @timer can
1542 * expire is on the closest next tick and delayed_work users depend
1543 * on that there's no such delay when @delay is 0.
1544 */
1545 if (!delay) {
1546 __queue_work(cpu, wq, &dwork->work);
1547 return;
1548 }
1549
1550 timer_stats_timer_set_start_info(&dwork->timer);
1551
1552 dwork->wq = wq;
1553 dwork->cpu = cpu;
1554 timer->expires = jiffies + delay;
1555
1556 if (unlikely(cpu != WORK_CPU_UNBOUND))
1557 add_timer_on(timer, cpu);
1558 else
1559 add_timer(timer);
1560}
1561
1562/**
1563 * queue_delayed_work_on - queue work on specific CPU after delay
1564 * @cpu: CPU number to execute work on
1565 * @wq: workqueue to use
1566 * @dwork: work to queue
1567 * @delay: number of jiffies to wait before queueing
1568 *
1569 * Return: %false if @work was already on a queue, %true otherwise. If
1570 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1571 * execution.
1572 */
1573bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1574 struct delayed_work *dwork, unsigned long delay)
1575{
1576 struct work_struct *work = &dwork->work;
1577 bool ret = false;
1578 unsigned long flags;
1579
1580 /* read the comment in __queue_work() */
1581 local_irq_save(flags);
1582
1583 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1584 __queue_delayed_work(cpu, wq, dwork, delay);
1585 ret = true;
1586 }
1587
1588 local_irq_restore(flags);
1589 return ret;
1590}
1591EXPORT_SYMBOL(queue_delayed_work_on);
1592
1593/**
1594 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1595 * @cpu: CPU number to execute work on
1596 * @wq: workqueue to use
1597 * @dwork: work to queue
1598 * @delay: number of jiffies to wait before queueing
1599 *
1600 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1601 * modify @dwork's timer so that it expires after @delay. If @delay is
1602 * zero, @work is guaranteed to be scheduled immediately regardless of its
1603 * current state.
1604 *
1605 * Return: %false if @dwork was idle and queued, %true if @dwork was
1606 * pending and its timer was modified.
1607 *
1608 * This function is safe to call from any context including IRQ handler.
1609 * See try_to_grab_pending() for details.
1610 */
1611bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1612 struct delayed_work *dwork, unsigned long delay)
1613{
1614 unsigned long flags;
1615 int ret;
1616
1617 do {
1618 ret = try_to_grab_pending(&dwork->work, true, &flags);
1619 } while (unlikely(ret == -EAGAIN));
1620
1621 if (likely(ret >= 0)) {
1622 __queue_delayed_work(cpu, wq, dwork, delay);
1623 local_irq_restore(flags);
1624 }
1625
1626 /* -ENOENT from try_to_grab_pending() becomes %true */
1627 return ret;
1628}
1629EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1630
1631/**
1632 * worker_enter_idle - enter idle state
1633 * @worker: worker which is entering idle state
1634 *
1635 * @worker is entering idle state. Update stats and idle timer if
1636 * necessary.
1637 *
1638 * LOCKING:
1639 * spin_lock_irq(pool->lock).
1640 */
1641static void worker_enter_idle(struct worker *worker)
1642{
1643 struct worker_pool *pool = worker->pool;
1644
1645 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1646 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1647 (worker->hentry.next || worker->hentry.pprev)))
1648 return;
1649
1650 /* can't use worker_set_flags(), also called from create_worker() */
1651 worker->flags |= WORKER_IDLE;
1652 pool->nr_idle++;
1653 worker->last_active = jiffies;
1654
1655 /* idle_list is LIFO */
1656 list_add(&worker->entry, &pool->idle_list);
1657
1658 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1659 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1660
1661 /*
1662 * Sanity check nr_running. Because wq_unbind_fn() releases
1663 * pool->lock between setting %WORKER_UNBOUND and zapping
1664 * nr_running, the warning may trigger spuriously. Check iff
1665 * unbind is not in progress.
1666 */
1667 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1668 pool->nr_workers == pool->nr_idle &&
1669 atomic_read(&pool->nr_running));
1670}
1671
1672/**
1673 * worker_leave_idle - leave idle state
1674 * @worker: worker which is leaving idle state
1675 *
1676 * @worker is leaving idle state. Update stats.
1677 *
1678 * LOCKING:
1679 * spin_lock_irq(pool->lock).
1680 */
1681static void worker_leave_idle(struct worker *worker)
1682{
1683 struct worker_pool *pool = worker->pool;
1684
1685 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1686 return;
1687 worker_clr_flags(worker, WORKER_IDLE);
1688 pool->nr_idle--;
1689 list_del_init(&worker->entry);
1690}
1691
1692static struct worker *alloc_worker(int node)
1693{
1694 struct worker *worker;
1695
1696 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1697 if (worker) {
1698 INIT_LIST_HEAD(&worker->entry);
1699 INIT_LIST_HEAD(&worker->scheduled);
1700 INIT_LIST_HEAD(&worker->node);
1701 /* on creation a worker is in !idle && prep state */
1702 worker->flags = WORKER_PREP;
1703 }
1704 return worker;
1705}
1706
1707/**
1708 * worker_attach_to_pool() - attach a worker to a pool
1709 * @worker: worker to be attached
1710 * @pool: the target pool
1711 *
1712 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1713 * cpu-binding of @worker are kept coordinated with the pool across
1714 * cpu-[un]hotplugs.
1715 */
1716static void worker_attach_to_pool(struct worker *worker,
1717 struct worker_pool *pool)
1718{
1719 mutex_lock(&pool->attach_mutex);
1720
1721 /*
1722 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723 * online CPUs. It'll be re-applied when any of the CPUs come up.
1724 */
1725 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1726
1727 /*
1728 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1729 * stable across this function. See the comments above the
1730 * flag definition for details.
1731 */
1732 if (pool->flags & POOL_DISASSOCIATED)
1733 worker->flags |= WORKER_UNBOUND;
1734
1735 list_add_tail(&worker->node, &pool->workers);
1736
1737 mutex_unlock(&pool->attach_mutex);
1738}
1739
1740/**
1741 * worker_detach_from_pool() - detach a worker from its pool
1742 * @worker: worker which is attached to its pool
1743 * @pool: the pool @worker is attached to
1744 *
1745 * Undo the attaching which had been done in worker_attach_to_pool(). The
1746 * caller worker shouldn't access to the pool after detached except it has
1747 * other reference to the pool.
1748 */
1749static void worker_detach_from_pool(struct worker *worker,
1750 struct worker_pool *pool)
1751{
1752 struct completion *detach_completion = NULL;
1753
1754 mutex_lock(&pool->attach_mutex);
1755 list_del(&worker->node);
1756 if (list_empty(&pool->workers))
1757 detach_completion = pool->detach_completion;
1758 mutex_unlock(&pool->attach_mutex);
1759
1760 /* clear leftover flags without pool->lock after it is detached */
1761 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1762
1763 if (detach_completion)
1764 complete(detach_completion);
1765}
1766
1767/**
1768 * create_worker - create a new workqueue worker
1769 * @pool: pool the new worker will belong to
1770 *
1771 * Create and start a new worker which is attached to @pool.
1772 *
1773 * CONTEXT:
1774 * Might sleep. Does GFP_KERNEL allocations.
1775 *
1776 * Return:
1777 * Pointer to the newly created worker.
1778 */
1779static struct worker *create_worker(struct worker_pool *pool)
1780{
1781 struct worker *worker = NULL;
1782 int id = -1;
1783 char id_buf[16];
1784
1785 /* ID is needed to determine kthread name */
1786 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1787 if (id < 0)
1788 goto fail;
1789
1790 worker = alloc_worker(pool->node);
1791 if (!worker)
1792 goto fail;
1793
1794 worker->pool = pool;
1795 worker->id = id;
1796
1797 if (pool->cpu >= 0)
1798 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1799 pool->attrs->nice < 0 ? "H" : "");
1800 else
1801 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1802
1803 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1804 "kworker/%s", id_buf);
1805 if (IS_ERR(worker->task))
1806 goto fail;
1807
1808 set_user_nice(worker->task, pool->attrs->nice);
1809 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1810
1811 /* successful, attach the worker to the pool */
1812 worker_attach_to_pool(worker, pool);
1813
1814 /* start the newly created worker */
1815 spin_lock_irq(&pool->lock);
1816 worker->pool->nr_workers++;
1817 worker_enter_idle(worker);
1818 wake_up_process(worker->task);
1819 spin_unlock_irq(&pool->lock);
1820
1821 return worker;
1822
1823fail:
1824 if (id >= 0)
1825 ida_simple_remove(&pool->worker_ida, id);
1826 kfree(worker);
1827 return NULL;
1828}
1829
1830/**
1831 * destroy_worker - destroy a workqueue worker
1832 * @worker: worker to be destroyed
1833 *
1834 * Destroy @worker and adjust @pool stats accordingly. The worker should
1835 * be idle.
1836 *
1837 * CONTEXT:
1838 * spin_lock_irq(pool->lock).
1839 */
1840static void destroy_worker(struct worker *worker)
1841{
1842 struct worker_pool *pool = worker->pool;
1843
1844 lockdep_assert_held(&pool->lock);
1845
1846 /* sanity check frenzy */
1847 if (WARN_ON(worker->current_work) ||
1848 WARN_ON(!list_empty(&worker->scheduled)) ||
1849 WARN_ON(!(worker->flags & WORKER_IDLE)))
1850 return;
1851
1852 pool->nr_workers--;
1853 pool->nr_idle--;
1854
1855 list_del_init(&worker->entry);
1856 worker->flags |= WORKER_DIE;
1857 wake_up_process(worker->task);
1858}
1859
1860static void idle_worker_timeout(unsigned long __pool)
1861{
1862 struct worker_pool *pool = (void *)__pool;
1863
1864 spin_lock_irq(&pool->lock);
1865
1866 while (too_many_workers(pool)) {
1867 struct worker *worker;
1868 unsigned long expires;
1869
1870 /* idle_list is kept in LIFO order, check the last one */
1871 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1872 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873
1874 if (time_before(jiffies, expires)) {
1875 mod_timer(&pool->idle_timer, expires);
1876 break;
1877 }
1878
1879 destroy_worker(worker);
1880 }
1881
1882 spin_unlock_irq(&pool->lock);
1883}
1884
1885static void send_mayday(struct work_struct *work)
1886{
1887 struct pool_workqueue *pwq = get_work_pwq(work);
1888 struct workqueue_struct *wq = pwq->wq;
1889
1890 lockdep_assert_held(&wq_mayday_lock);
1891
1892 if (!wq->rescuer)
1893 return;
1894
1895 /* mayday mayday mayday */
1896 if (list_empty(&pwq->mayday_node)) {
1897 /*
1898 * If @pwq is for an unbound wq, its base ref may be put at
1899 * any time due to an attribute change. Pin @pwq until the
1900 * rescuer is done with it.
1901 */
1902 get_pwq(pwq);
1903 list_add_tail(&pwq->mayday_node, &wq->maydays);
1904 wake_up_process(wq->rescuer->task);
1905 }
1906}
1907
1908static void pool_mayday_timeout(unsigned long __pool)
1909{
1910 struct worker_pool *pool = (void *)__pool;
1911 struct work_struct *work;
1912
1913 spin_lock_irq(&pool->lock);
1914 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1915
1916 if (need_to_create_worker(pool)) {
1917 /*
1918 * We've been trying to create a new worker but
1919 * haven't been successful. We might be hitting an
1920 * allocation deadlock. Send distress signals to
1921 * rescuers.
1922 */
1923 list_for_each_entry(work, &pool->worklist, entry)
1924 send_mayday(work);
1925 }
1926
1927 spin_unlock(&wq_mayday_lock);
1928 spin_unlock_irq(&pool->lock);
1929
1930 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1931}
1932
1933/**
1934 * maybe_create_worker - create a new worker if necessary
1935 * @pool: pool to create a new worker for
1936 *
1937 * Create a new worker for @pool if necessary. @pool is guaranteed to
1938 * have at least one idle worker on return from this function. If
1939 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1940 * sent to all rescuers with works scheduled on @pool to resolve
1941 * possible allocation deadlock.
1942 *
1943 * On return, need_to_create_worker() is guaranteed to be %false and
1944 * may_start_working() %true.
1945 *
1946 * LOCKING:
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times. Does GFP_KERNEL allocations. Called only from
1949 * manager.
1950 */
1951static void maybe_create_worker(struct worker_pool *pool)
1952__releases(&pool->lock)
1953__acquires(&pool->lock)
1954{
1955restart:
1956 spin_unlock_irq(&pool->lock);
1957
1958 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1959 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1960
1961 while (true) {
1962 if (create_worker(pool) || !need_to_create_worker(pool))
1963 break;
1964
1965 schedule_timeout_interruptible(CREATE_COOLDOWN);
1966
1967 if (!need_to_create_worker(pool))
1968 break;
1969 }
1970
1971 del_timer_sync(&pool->mayday_timer);
1972 spin_lock_irq(&pool->lock);
1973 /*
1974 * This is necessary even after a new worker was just successfully
1975 * created as @pool->lock was dropped and the new worker might have
1976 * already become busy.
1977 */
1978 if (need_to_create_worker(pool))
1979 goto restart;
1980}
1981
1982/**
1983 * manage_workers - manage worker pool
1984 * @worker: self
1985 *
1986 * Assume the manager role and manage the worker pool @worker belongs
1987 * to. At any given time, there can be only zero or one manager per
1988 * pool. The exclusion is handled automatically by this function.
1989 *
1990 * The caller can safely start processing works on false return. On
1991 * true return, it's guaranteed that need_to_create_worker() is false
1992 * and may_start_working() is true.
1993 *
1994 * CONTEXT:
1995 * spin_lock_irq(pool->lock) which may be released and regrabbed
1996 * multiple times. Does GFP_KERNEL allocations.
1997 *
1998 * Return:
1999 * %false if the pool doesn't need management and the caller can safely
2000 * start processing works, %true if management function was performed and
2001 * the conditions that the caller verified before calling the function may
2002 * no longer be true.
2003 */
2004static bool manage_workers(struct worker *worker)
2005{
2006 struct worker_pool *pool = worker->pool;
2007
2008 /*
2009 * Anyone who successfully grabs manager_arb wins the arbitration
2010 * and becomes the manager. mutex_trylock() on pool->manager_arb
2011 * failure while holding pool->lock reliably indicates that someone
2012 * else is managing the pool and the worker which failed trylock
2013 * can proceed to executing work items. This means that anyone
2014 * grabbing manager_arb is responsible for actually performing
2015 * manager duties. If manager_arb is grabbed and released without
2016 * actual management, the pool may stall indefinitely.
2017 */
2018 if (!mutex_trylock(&pool->manager_arb))
2019 return false;
2020 pool->manager = worker;
2021
2022 maybe_create_worker(pool);
2023
2024 pool->manager = NULL;
2025 mutex_unlock(&pool->manager_arb);
2026 return true;
2027}
2028
2029/**
2030 * process_one_work - process single work
2031 * @worker: self
2032 * @work: work to process
2033 *
2034 * Process @work. This function contains all the logics necessary to
2035 * process a single work including synchronization against and
2036 * interaction with other workers on the same cpu, queueing and
2037 * flushing. As long as context requirement is met, any worker can
2038 * call this function to process a work.
2039 *
2040 * CONTEXT:
2041 * spin_lock_irq(pool->lock) which is released and regrabbed.
2042 */
2043static void process_one_work(struct worker *worker, struct work_struct *work)
2044__releases(&pool->lock)
2045__acquires(&pool->lock)
2046{
2047 struct pool_workqueue *pwq = get_work_pwq(work);
2048 struct worker_pool *pool = worker->pool;
2049 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2050 int work_color;
2051 struct worker *collision;
2052#ifdef CONFIG_LOCKDEP
2053 /*
2054 * It is permissible to free the struct work_struct from
2055 * inside the function that is called from it, this we need to
2056 * take into account for lockdep too. To avoid bogus "held
2057 * lock freed" warnings as well as problems when looking into
2058 * work->lockdep_map, make a copy and use that here.
2059 */
2060 struct lockdep_map lockdep_map;
2061
2062 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2063#endif
2064 /* ensure we're on the correct CPU */
2065 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2066 raw_smp_processor_id() != pool->cpu);
2067
2068 /*
2069 * A single work shouldn't be executed concurrently by
2070 * multiple workers on a single cpu. Check whether anyone is
2071 * already processing the work. If so, defer the work to the
2072 * currently executing one.
2073 */
2074 collision = find_worker_executing_work(pool, work);
2075 if (unlikely(collision)) {
2076 move_linked_works(work, &collision->scheduled, NULL);
2077 return;
2078 }
2079
2080 /* claim and dequeue */
2081 debug_work_deactivate(work);
2082 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2083 worker->current_work = work;
2084 worker->current_func = work->func;
2085 worker->current_pwq = pwq;
2086 work_color = get_work_color(work);
2087
2088 list_del_init(&work->entry);
2089
2090 /*
2091 * CPU intensive works don't participate in concurrency management.
2092 * They're the scheduler's responsibility. This takes @worker out
2093 * of concurrency management and the next code block will chain
2094 * execution of the pending work items.
2095 */
2096 if (unlikely(cpu_intensive))
2097 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2098
2099 /*
2100 * Wake up another worker if necessary. The condition is always
2101 * false for normal per-cpu workers since nr_running would always
2102 * be >= 1 at this point. This is used to chain execution of the
2103 * pending work items for WORKER_NOT_RUNNING workers such as the
2104 * UNBOUND and CPU_INTENSIVE ones.
2105 */
2106 if (need_more_worker(pool))
2107 wake_up_worker(pool);
2108
2109 /*
2110 * Record the last pool and clear PENDING which should be the last
2111 * update to @work. Also, do this inside @pool->lock so that
2112 * PENDING and queued state changes happen together while IRQ is
2113 * disabled.
2114 */
2115 set_work_pool_and_clear_pending(work, pool->id);
2116
2117 spin_unlock_irq(&pool->lock);
2118
2119 lock_map_acquire_read(&pwq->wq->lockdep_map);
2120 lock_map_acquire(&lockdep_map);
2121 trace_workqueue_execute_start(work);
2122 worker->current_func(work);
2123 /*
2124 * While we must be careful to not use "work" after this, the trace
2125 * point will only record its address.
2126 */
2127 trace_workqueue_execute_end(work);
2128 lock_map_release(&lockdep_map);
2129 lock_map_release(&pwq->wq->lockdep_map);
2130
2131 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2132 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2133 " last function: %pf\n",
2134 current->comm, preempt_count(), task_pid_nr(current),
2135 worker->current_func);
2136 debug_show_held_locks(current);
2137 dump_stack();
2138 }
2139
2140 /*
2141 * The following prevents a kworker from hogging CPU on !PREEMPT
2142 * kernels, where a requeueing work item waiting for something to
2143 * happen could deadlock with stop_machine as such work item could
2144 * indefinitely requeue itself while all other CPUs are trapped in
2145 * stop_machine. At the same time, report a quiescent RCU state so
2146 * the same condition doesn't freeze RCU.
2147 */
2148 cond_resched_rcu_qs();
2149
2150 spin_lock_irq(&pool->lock);
2151
2152 /* clear cpu intensive status */
2153 if (unlikely(cpu_intensive))
2154 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2155
2156 /* we're done with it, release */
2157 hash_del(&worker->hentry);
2158 worker->current_work = NULL;
2159 worker->current_func = NULL;
2160 worker->current_pwq = NULL;
2161 worker->desc_valid = false;
2162 pwq_dec_nr_in_flight(pwq, work_color);
2163}
2164
2165/**
2166 * process_scheduled_works - process scheduled works
2167 * @worker: self
2168 *
2169 * Process all scheduled works. Please note that the scheduled list
2170 * may change while processing a work, so this function repeatedly
2171 * fetches a work from the top and executes it.
2172 *
2173 * CONTEXT:
2174 * spin_lock_irq(pool->lock) which may be released and regrabbed
2175 * multiple times.
2176 */
2177static void process_scheduled_works(struct worker *worker)
2178{
2179 while (!list_empty(&worker->scheduled)) {
2180 struct work_struct *work = list_first_entry(&worker->scheduled,
2181 struct work_struct, entry);
2182 process_one_work(worker, work);
2183 }
2184}
2185
2186/**
2187 * worker_thread - the worker thread function
2188 * @__worker: self
2189 *
2190 * The worker thread function. All workers belong to a worker_pool -
2191 * either a per-cpu one or dynamic unbound one. These workers process all
2192 * work items regardless of their specific target workqueue. The only
2193 * exception is work items which belong to workqueues with a rescuer which
2194 * will be explained in rescuer_thread().
2195 *
2196 * Return: 0
2197 */
2198static int worker_thread(void *__worker)
2199{
2200 struct worker *worker = __worker;
2201 struct worker_pool *pool = worker->pool;
2202
2203 /* tell the scheduler that this is a workqueue worker */
2204 worker->task->flags |= PF_WQ_WORKER;
2205woke_up:
2206 spin_lock_irq(&pool->lock);
2207
2208 /* am I supposed to die? */
2209 if (unlikely(worker->flags & WORKER_DIE)) {
2210 spin_unlock_irq(&pool->lock);
2211 WARN_ON_ONCE(!list_empty(&worker->entry));
2212 worker->task->flags &= ~PF_WQ_WORKER;
2213
2214 set_task_comm(worker->task, "kworker/dying");
2215 ida_simple_remove(&pool->worker_ida, worker->id);
2216 worker_detach_from_pool(worker, pool);
2217 kfree(worker);
2218 return 0;
2219 }
2220
2221 worker_leave_idle(worker);
2222recheck:
2223 /* no more worker necessary? */
2224 if (!need_more_worker(pool))
2225 goto sleep;
2226
2227 /* do we need to manage? */
2228 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2229 goto recheck;
2230
2231 /*
2232 * ->scheduled list can only be filled while a worker is
2233 * preparing to process a work or actually processing it.
2234 * Make sure nobody diddled with it while I was sleeping.
2235 */
2236 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2237
2238 /*
2239 * Finish PREP stage. We're guaranteed to have at least one idle
2240 * worker or that someone else has already assumed the manager
2241 * role. This is where @worker starts participating in concurrency
2242 * management if applicable and concurrency management is restored
2243 * after being rebound. See rebind_workers() for details.
2244 */
2245 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2246
2247 do {
2248 struct work_struct *work =
2249 list_first_entry(&pool->worklist,
2250 struct work_struct, entry);
2251
2252 pool->watchdog_ts = jiffies;
2253
2254 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2255 /* optimization path, not strictly necessary */
2256 process_one_work(worker, work);
2257 if (unlikely(!list_empty(&worker->scheduled)))
2258 process_scheduled_works(worker);
2259 } else {
2260 move_linked_works(work, &worker->scheduled, NULL);
2261 process_scheduled_works(worker);
2262 }
2263 } while (keep_working(pool));
2264
2265 worker_set_flags(worker, WORKER_PREP);
2266sleep:
2267 /*
2268 * pool->lock is held and there's no work to process and no need to
2269 * manage, sleep. Workers are woken up only while holding
2270 * pool->lock or from local cpu, so setting the current state
2271 * before releasing pool->lock is enough to prevent losing any
2272 * event.
2273 */
2274 worker_enter_idle(worker);
2275 __set_current_state(TASK_INTERRUPTIBLE);
2276 spin_unlock_irq(&pool->lock);
2277 schedule();
2278 goto woke_up;
2279}
2280
2281/**
2282 * rescuer_thread - the rescuer thread function
2283 * @__rescuer: self
2284 *
2285 * Workqueue rescuer thread function. There's one rescuer for each
2286 * workqueue which has WQ_MEM_RECLAIM set.
2287 *
2288 * Regular work processing on a pool may block trying to create a new
2289 * worker which uses GFP_KERNEL allocation which has slight chance of
2290 * developing into deadlock if some works currently on the same queue
2291 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2292 * the problem rescuer solves.
2293 *
2294 * When such condition is possible, the pool summons rescuers of all
2295 * workqueues which have works queued on the pool and let them process
2296 * those works so that forward progress can be guaranteed.
2297 *
2298 * This should happen rarely.
2299 *
2300 * Return: 0
2301 */
2302static int rescuer_thread(void *__rescuer)
2303{
2304 struct worker *rescuer = __rescuer;
2305 struct workqueue_struct *wq = rescuer->rescue_wq;
2306 struct list_head *scheduled = &rescuer->scheduled;
2307 bool should_stop;
2308
2309 set_user_nice(current, RESCUER_NICE_LEVEL);
2310
2311 /*
2312 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2313 * doesn't participate in concurrency management.
2314 */
2315 rescuer->task->flags |= PF_WQ_WORKER;
2316repeat:
2317 set_current_state(TASK_INTERRUPTIBLE);
2318
2319 /*
2320 * By the time the rescuer is requested to stop, the workqueue
2321 * shouldn't have any work pending, but @wq->maydays may still have
2322 * pwq(s) queued. This can happen by non-rescuer workers consuming
2323 * all the work items before the rescuer got to them. Go through
2324 * @wq->maydays processing before acting on should_stop so that the
2325 * list is always empty on exit.
2326 */
2327 should_stop = kthread_should_stop();
2328
2329 /* see whether any pwq is asking for help */
2330 spin_lock_irq(&wq_mayday_lock);
2331
2332 while (!list_empty(&wq->maydays)) {
2333 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2334 struct pool_workqueue, mayday_node);
2335 struct worker_pool *pool = pwq->pool;
2336 struct work_struct *work, *n;
2337 bool first = true;
2338
2339 __set_current_state(TASK_RUNNING);
2340 list_del_init(&pwq->mayday_node);
2341
2342 spin_unlock_irq(&wq_mayday_lock);
2343
2344 worker_attach_to_pool(rescuer, pool);
2345
2346 spin_lock_irq(&pool->lock);
2347 rescuer->pool = pool;
2348
2349 /*
2350 * Slurp in all works issued via this workqueue and
2351 * process'em.
2352 */
2353 WARN_ON_ONCE(!list_empty(scheduled));
2354 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2355 if (get_work_pwq(work) == pwq) {
2356 if (first)
2357 pool->watchdog_ts = jiffies;
2358 move_linked_works(work, scheduled, &n);
2359 }
2360 first = false;
2361 }
2362
2363 if (!list_empty(scheduled)) {
2364 process_scheduled_works(rescuer);
2365
2366 /*
2367 * The above execution of rescued work items could
2368 * have created more to rescue through
2369 * pwq_activate_first_delayed() or chained
2370 * queueing. Let's put @pwq back on mayday list so
2371 * that such back-to-back work items, which may be
2372 * being used to relieve memory pressure, don't
2373 * incur MAYDAY_INTERVAL delay inbetween.
2374 */
2375 if (need_to_create_worker(pool)) {
2376 spin_lock(&wq_mayday_lock);
2377 get_pwq(pwq);
2378 list_move_tail(&pwq->mayday_node, &wq->maydays);
2379 spin_unlock(&wq_mayday_lock);
2380 }
2381 }
2382
2383 /*
2384 * Put the reference grabbed by send_mayday(). @pool won't
2385 * go away while we're still attached to it.
2386 */
2387 put_pwq(pwq);
2388
2389 /*
2390 * Leave this pool. If need_more_worker() is %true, notify a
2391 * regular worker; otherwise, we end up with 0 concurrency
2392 * and stalling the execution.
2393 */
2394 if (need_more_worker(pool))
2395 wake_up_worker(pool);
2396
2397 rescuer->pool = NULL;
2398 spin_unlock_irq(&pool->lock);
2399
2400 worker_detach_from_pool(rescuer, pool);
2401
2402 spin_lock_irq(&wq_mayday_lock);
2403 }
2404
2405 spin_unlock_irq(&wq_mayday_lock);
2406
2407 if (should_stop) {
2408 __set_current_state(TASK_RUNNING);
2409 rescuer->task->flags &= ~PF_WQ_WORKER;
2410 return 0;
2411 }
2412
2413 /* rescuers should never participate in concurrency management */
2414 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2415 schedule();
2416 goto repeat;
2417}
2418
2419/**
2420 * check_flush_dependency - check for flush dependency sanity
2421 * @target_wq: workqueue being flushed
2422 * @target_work: work item being flushed (NULL for workqueue flushes)
2423 *
2424 * %current is trying to flush the whole @target_wq or @target_work on it.
2425 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2426 * reclaiming memory or running on a workqueue which doesn't have
2427 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2428 * a deadlock.
2429 */
2430static void check_flush_dependency(struct workqueue_struct *target_wq,
2431 struct work_struct *target_work)
2432{
2433 work_func_t target_func = target_work ? target_work->func : NULL;
2434 struct worker *worker;
2435
2436 if (target_wq->flags & WQ_MEM_RECLAIM)
2437 return;
2438
2439 worker = current_wq_worker();
2440
2441 WARN_ONCE(current->flags & PF_MEMALLOC,
2442 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2443 current->pid, current->comm, target_wq->name, target_func);
2444 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2445 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2446 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2447 worker->current_pwq->wq->name, worker->current_func,
2448 target_wq->name, target_func);
2449}
2450
2451struct wq_barrier {
2452 struct work_struct work;
2453 struct completion done;
2454 struct task_struct *task; /* purely informational */
2455};
2456
2457static void wq_barrier_func(struct work_struct *work)
2458{
2459 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2460 complete(&barr->done);
2461}
2462
2463/**
2464 * insert_wq_barrier - insert a barrier work
2465 * @pwq: pwq to insert barrier into
2466 * @barr: wq_barrier to insert
2467 * @target: target work to attach @barr to
2468 * @worker: worker currently executing @target, NULL if @target is not executing
2469 *
2470 * @barr is linked to @target such that @barr is completed only after
2471 * @target finishes execution. Please note that the ordering
2472 * guarantee is observed only with respect to @target and on the local
2473 * cpu.
2474 *
2475 * Currently, a queued barrier can't be canceled. This is because
2476 * try_to_grab_pending() can't determine whether the work to be
2477 * grabbed is at the head of the queue and thus can't clear LINKED
2478 * flag of the previous work while there must be a valid next work
2479 * after a work with LINKED flag set.
2480 *
2481 * Note that when @worker is non-NULL, @target may be modified
2482 * underneath us, so we can't reliably determine pwq from @target.
2483 *
2484 * CONTEXT:
2485 * spin_lock_irq(pool->lock).
2486 */
2487static void insert_wq_barrier(struct pool_workqueue *pwq,
2488 struct wq_barrier *barr,
2489 struct work_struct *target, struct worker *worker)
2490{
2491 struct list_head *head;
2492 unsigned int linked = 0;
2493
2494 /*
2495 * debugobject calls are safe here even with pool->lock locked
2496 * as we know for sure that this will not trigger any of the
2497 * checks and call back into the fixup functions where we
2498 * might deadlock.
2499 */
2500 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2501 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2502 init_completion(&barr->done);
2503 barr->task = current;
2504
2505 /*
2506 * If @target is currently being executed, schedule the
2507 * barrier to the worker; otherwise, put it after @target.
2508 */
2509 if (worker)
2510 head = worker->scheduled.next;
2511 else {
2512 unsigned long *bits = work_data_bits(target);
2513
2514 head = target->entry.next;
2515 /* there can already be other linked works, inherit and set */
2516 linked = *bits & WORK_STRUCT_LINKED;
2517 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2518 }
2519
2520 debug_work_activate(&barr->work);
2521 insert_work(pwq, &barr->work, head,
2522 work_color_to_flags(WORK_NO_COLOR) | linked);
2523}
2524
2525/**
2526 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2527 * @wq: workqueue being flushed
2528 * @flush_color: new flush color, < 0 for no-op
2529 * @work_color: new work color, < 0 for no-op
2530 *
2531 * Prepare pwqs for workqueue flushing.
2532 *
2533 * If @flush_color is non-negative, flush_color on all pwqs should be
2534 * -1. If no pwq has in-flight commands at the specified color, all
2535 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2536 * has in flight commands, its pwq->flush_color is set to
2537 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2538 * wakeup logic is armed and %true is returned.
2539 *
2540 * The caller should have initialized @wq->first_flusher prior to
2541 * calling this function with non-negative @flush_color. If
2542 * @flush_color is negative, no flush color update is done and %false
2543 * is returned.
2544 *
2545 * If @work_color is non-negative, all pwqs should have the same
2546 * work_color which is previous to @work_color and all will be
2547 * advanced to @work_color.
2548 *
2549 * CONTEXT:
2550 * mutex_lock(wq->mutex).
2551 *
2552 * Return:
2553 * %true if @flush_color >= 0 and there's something to flush. %false
2554 * otherwise.
2555 */
2556static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2557 int flush_color, int work_color)
2558{
2559 bool wait = false;
2560 struct pool_workqueue *pwq;
2561
2562 if (flush_color >= 0) {
2563 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2564 atomic_set(&wq->nr_pwqs_to_flush, 1);
2565 }
2566
2567 for_each_pwq(pwq, wq) {
2568 struct worker_pool *pool = pwq->pool;
2569
2570 spin_lock_irq(&pool->lock);
2571
2572 if (flush_color >= 0) {
2573 WARN_ON_ONCE(pwq->flush_color != -1);
2574
2575 if (pwq->nr_in_flight[flush_color]) {
2576 pwq->flush_color = flush_color;
2577 atomic_inc(&wq->nr_pwqs_to_flush);
2578 wait = true;
2579 }
2580 }
2581
2582 if (work_color >= 0) {
2583 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2584 pwq->work_color = work_color;
2585 }
2586
2587 spin_unlock_irq(&pool->lock);
2588 }
2589
2590 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2591 complete(&wq->first_flusher->done);
2592
2593 return wait;
2594}
2595
2596/**
2597 * flush_workqueue - ensure that any scheduled work has run to completion.
2598 * @wq: workqueue to flush
2599 *
2600 * This function sleeps until all work items which were queued on entry
2601 * have finished execution, but it is not livelocked by new incoming ones.
2602 */
2603void flush_workqueue(struct workqueue_struct *wq)
2604{
2605 struct wq_flusher this_flusher = {
2606 .list = LIST_HEAD_INIT(this_flusher.list),
2607 .flush_color = -1,
2608 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2609 };
2610 int next_color;
2611
2612 lock_map_acquire(&wq->lockdep_map);
2613 lock_map_release(&wq->lockdep_map);
2614
2615 mutex_lock(&wq->mutex);
2616
2617 /*
2618 * Start-to-wait phase
2619 */
2620 next_color = work_next_color(wq->work_color);
2621
2622 if (next_color != wq->flush_color) {
2623 /*
2624 * Color space is not full. The current work_color
2625 * becomes our flush_color and work_color is advanced
2626 * by one.
2627 */
2628 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2629 this_flusher.flush_color = wq->work_color;
2630 wq->work_color = next_color;
2631
2632 if (!wq->first_flusher) {
2633 /* no flush in progress, become the first flusher */
2634 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2635
2636 wq->first_flusher = &this_flusher;
2637
2638 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2639 wq->work_color)) {
2640 /* nothing to flush, done */
2641 wq->flush_color = next_color;
2642 wq->first_flusher = NULL;
2643 goto out_unlock;
2644 }
2645 } else {
2646 /* wait in queue */
2647 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2648 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2649 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2650 }
2651 } else {
2652 /*
2653 * Oops, color space is full, wait on overflow queue.
2654 * The next flush completion will assign us
2655 * flush_color and transfer to flusher_queue.
2656 */
2657 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2658 }
2659
2660 check_flush_dependency(wq, NULL);
2661
2662 mutex_unlock(&wq->mutex);
2663
2664 wait_for_completion(&this_flusher.done);
2665
2666 /*
2667 * Wake-up-and-cascade phase
2668 *
2669 * First flushers are responsible for cascading flushes and
2670 * handling overflow. Non-first flushers can simply return.
2671 */
2672 if (wq->first_flusher != &this_flusher)
2673 return;
2674
2675 mutex_lock(&wq->mutex);
2676
2677 /* we might have raced, check again with mutex held */
2678 if (wq->first_flusher != &this_flusher)
2679 goto out_unlock;
2680
2681 wq->first_flusher = NULL;
2682
2683 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2684 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2685
2686 while (true) {
2687 struct wq_flusher *next, *tmp;
2688
2689 /* complete all the flushers sharing the current flush color */
2690 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2691 if (next->flush_color != wq->flush_color)
2692 break;
2693 list_del_init(&next->list);
2694 complete(&next->done);
2695 }
2696
2697 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2698 wq->flush_color != work_next_color(wq->work_color));
2699
2700 /* this flush_color is finished, advance by one */
2701 wq->flush_color = work_next_color(wq->flush_color);
2702
2703 /* one color has been freed, handle overflow queue */
2704 if (!list_empty(&wq->flusher_overflow)) {
2705 /*
2706 * Assign the same color to all overflowed
2707 * flushers, advance work_color and append to
2708 * flusher_queue. This is the start-to-wait
2709 * phase for these overflowed flushers.
2710 */
2711 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2712 tmp->flush_color = wq->work_color;
2713
2714 wq->work_color = work_next_color(wq->work_color);
2715
2716 list_splice_tail_init(&wq->flusher_overflow,
2717 &wq->flusher_queue);
2718 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2719 }
2720
2721 if (list_empty(&wq->flusher_queue)) {
2722 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2723 break;
2724 }
2725
2726 /*
2727 * Need to flush more colors. Make the next flusher
2728 * the new first flusher and arm pwqs.
2729 */
2730 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2731 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2732
2733 list_del_init(&next->list);
2734 wq->first_flusher = next;
2735
2736 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2737 break;
2738
2739 /*
2740 * Meh... this color is already done, clear first
2741 * flusher and repeat cascading.
2742 */
2743 wq->first_flusher = NULL;
2744 }
2745
2746out_unlock:
2747 mutex_unlock(&wq->mutex);
2748}
2749EXPORT_SYMBOL(flush_workqueue);
2750
2751/**
2752 * drain_workqueue - drain a workqueue
2753 * @wq: workqueue to drain
2754 *
2755 * Wait until the workqueue becomes empty. While draining is in progress,
2756 * only chain queueing is allowed. IOW, only currently pending or running
2757 * work items on @wq can queue further work items on it. @wq is flushed
2758 * repeatedly until it becomes empty. The number of flushing is determined
2759 * by the depth of chaining and should be relatively short. Whine if it
2760 * takes too long.
2761 */
2762void drain_workqueue(struct workqueue_struct *wq)
2763{
2764 unsigned int flush_cnt = 0;
2765 struct pool_workqueue *pwq;
2766
2767 /*
2768 * __queue_work() needs to test whether there are drainers, is much
2769 * hotter than drain_workqueue() and already looks at @wq->flags.
2770 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2771 */
2772 mutex_lock(&wq->mutex);
2773 if (!wq->nr_drainers++)
2774 wq->flags |= __WQ_DRAINING;
2775 mutex_unlock(&wq->mutex);
2776reflush:
2777 flush_workqueue(wq);
2778
2779 mutex_lock(&wq->mutex);
2780
2781 for_each_pwq(pwq, wq) {
2782 bool drained;
2783
2784 spin_lock_irq(&pwq->pool->lock);
2785 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2786 spin_unlock_irq(&pwq->pool->lock);
2787
2788 if (drained)
2789 continue;
2790
2791 if (++flush_cnt == 10 ||
2792 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2793 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2794 wq->name, flush_cnt);
2795
2796 mutex_unlock(&wq->mutex);
2797 goto reflush;
2798 }
2799
2800 if (!--wq->nr_drainers)
2801 wq->flags &= ~__WQ_DRAINING;
2802 mutex_unlock(&wq->mutex);
2803}
2804EXPORT_SYMBOL_GPL(drain_workqueue);
2805
2806static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2807{
2808 struct worker *worker = NULL;
2809 struct worker_pool *pool;
2810 struct pool_workqueue *pwq;
2811
2812 might_sleep();
2813
2814 local_irq_disable();
2815 pool = get_work_pool(work);
2816 if (!pool) {
2817 local_irq_enable();
2818 return false;
2819 }
2820
2821 spin_lock(&pool->lock);
2822 /* see the comment in try_to_grab_pending() with the same code */
2823 pwq = get_work_pwq(work);
2824 if (pwq) {
2825 if (unlikely(pwq->pool != pool))
2826 goto already_gone;
2827 } else {
2828 worker = find_worker_executing_work(pool, work);
2829 if (!worker)
2830 goto already_gone;
2831 pwq = worker->current_pwq;
2832 }
2833
2834 check_flush_dependency(pwq->wq, work);
2835
2836 insert_wq_barrier(pwq, barr, work, worker);
2837 spin_unlock_irq(&pool->lock);
2838
2839 /*
2840 * If @max_active is 1 or rescuer is in use, flushing another work
2841 * item on the same workqueue may lead to deadlock. Make sure the
2842 * flusher is not running on the same workqueue by verifying write
2843 * access.
2844 */
2845 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2846 lock_map_acquire(&pwq->wq->lockdep_map);
2847 else
2848 lock_map_acquire_read(&pwq->wq->lockdep_map);
2849 lock_map_release(&pwq->wq->lockdep_map);
2850
2851 return true;
2852already_gone:
2853 spin_unlock_irq(&pool->lock);
2854 return false;
2855}
2856
2857/**
2858 * flush_work - wait for a work to finish executing the last queueing instance
2859 * @work: the work to flush
2860 *
2861 * Wait until @work has finished execution. @work is guaranteed to be idle
2862 * on return if it hasn't been requeued since flush started.
2863 *
2864 * Return:
2865 * %true if flush_work() waited for the work to finish execution,
2866 * %false if it was already idle.
2867 */
2868bool flush_work(struct work_struct *work)
2869{
2870 struct wq_barrier barr;
2871
2872 lock_map_acquire(&work->lockdep_map);
2873 lock_map_release(&work->lockdep_map);
2874
2875 if (start_flush_work(work, &barr)) {
2876 wait_for_completion(&barr.done);
2877 destroy_work_on_stack(&barr.work);
2878 return true;
2879 } else {
2880 return false;
2881 }
2882}
2883EXPORT_SYMBOL_GPL(flush_work);
2884
2885struct cwt_wait {
2886 wait_queue_t wait;
2887 struct work_struct *work;
2888};
2889
2890static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2891{
2892 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2893
2894 if (cwait->work != key)
2895 return 0;
2896 return autoremove_wake_function(wait, mode, sync, key);
2897}
2898
2899static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2900{
2901 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2902 unsigned long flags;
2903 int ret;
2904
2905 do {
2906 ret = try_to_grab_pending(work, is_dwork, &flags);
2907 /*
2908 * If someone else is already canceling, wait for it to
2909 * finish. flush_work() doesn't work for PREEMPT_NONE
2910 * because we may get scheduled between @work's completion
2911 * and the other canceling task resuming and clearing
2912 * CANCELING - flush_work() will return false immediately
2913 * as @work is no longer busy, try_to_grab_pending() will
2914 * return -ENOENT as @work is still being canceled and the
2915 * other canceling task won't be able to clear CANCELING as
2916 * we're hogging the CPU.
2917 *
2918 * Let's wait for completion using a waitqueue. As this
2919 * may lead to the thundering herd problem, use a custom
2920 * wake function which matches @work along with exclusive
2921 * wait and wakeup.
2922 */
2923 if (unlikely(ret == -ENOENT)) {
2924 struct cwt_wait cwait;
2925
2926 init_wait(&cwait.wait);
2927 cwait.wait.func = cwt_wakefn;
2928 cwait.work = work;
2929
2930 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2931 TASK_UNINTERRUPTIBLE);
2932 if (work_is_canceling(work))
2933 schedule();
2934 finish_wait(&cancel_waitq, &cwait.wait);
2935 }
2936 } while (unlikely(ret < 0));
2937
2938 /* tell other tasks trying to grab @work to back off */
2939 mark_work_canceling(work);
2940 local_irq_restore(flags);
2941
2942 flush_work(work);
2943 clear_work_data(work);
2944
2945 /*
2946 * Paired with prepare_to_wait() above so that either
2947 * waitqueue_active() is visible here or !work_is_canceling() is
2948 * visible there.
2949 */
2950 smp_mb();
2951 if (waitqueue_active(&cancel_waitq))
2952 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2953
2954 return ret;
2955}
2956
2957/**
2958 * cancel_work_sync - cancel a work and wait for it to finish
2959 * @work: the work to cancel
2960 *
2961 * Cancel @work and wait for its execution to finish. This function
2962 * can be used even if the work re-queues itself or migrates to
2963 * another workqueue. On return from this function, @work is
2964 * guaranteed to be not pending or executing on any CPU.
2965 *
2966 * cancel_work_sync(&delayed_work->work) must not be used for
2967 * delayed_work's. Use cancel_delayed_work_sync() instead.
2968 *
2969 * The caller must ensure that the workqueue on which @work was last
2970 * queued can't be destroyed before this function returns.
2971 *
2972 * Return:
2973 * %true if @work was pending, %false otherwise.
2974 */
2975bool cancel_work_sync(struct work_struct *work)
2976{
2977 return __cancel_work_timer(work, false);
2978}
2979EXPORT_SYMBOL_GPL(cancel_work_sync);
2980
2981/**
2982 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2983 * @dwork: the delayed work to flush
2984 *
2985 * Delayed timer is cancelled and the pending work is queued for
2986 * immediate execution. Like flush_work(), this function only
2987 * considers the last queueing instance of @dwork.
2988 *
2989 * Return:
2990 * %true if flush_work() waited for the work to finish execution,
2991 * %false if it was already idle.
2992 */
2993bool flush_delayed_work(struct delayed_work *dwork)
2994{
2995 local_irq_disable();
2996 if (del_timer_sync(&dwork->timer))
2997 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2998 local_irq_enable();
2999 return flush_work(&dwork->work);
3000}
3001EXPORT_SYMBOL(flush_delayed_work);
3002
3003/**
3004 * cancel_delayed_work - cancel a delayed work
3005 * @dwork: delayed_work to cancel
3006 *
3007 * Kill off a pending delayed_work.
3008 *
3009 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3010 * pending.
3011 *
3012 * Note:
3013 * The work callback function may still be running on return, unless
3014 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3015 * use cancel_delayed_work_sync() to wait on it.
3016 *
3017 * This function is safe to call from any context including IRQ handler.
3018 */
3019bool cancel_delayed_work(struct delayed_work *dwork)
3020{
3021 unsigned long flags;
3022 int ret;
3023
3024 do {
3025 ret = try_to_grab_pending(&dwork->work, true, &flags);
3026 } while (unlikely(ret == -EAGAIN));
3027
3028 if (unlikely(ret < 0))
3029 return false;
3030
3031 set_work_pool_and_clear_pending(&dwork->work,
3032 get_work_pool_id(&dwork->work));
3033 local_irq_restore(flags);
3034 return ret;
3035}
3036EXPORT_SYMBOL(cancel_delayed_work);
3037
3038/**
3039 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3040 * @dwork: the delayed work cancel
3041 *
3042 * This is cancel_work_sync() for delayed works.
3043 *
3044 * Return:
3045 * %true if @dwork was pending, %false otherwise.
3046 */
3047bool cancel_delayed_work_sync(struct delayed_work *dwork)
3048{
3049 return __cancel_work_timer(&dwork->work, true);
3050}
3051EXPORT_SYMBOL(cancel_delayed_work_sync);
3052
3053/**
3054 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3055 * @func: the function to call
3056 *
3057 * schedule_on_each_cpu() executes @func on each online CPU using the
3058 * system workqueue and blocks until all CPUs have completed.
3059 * schedule_on_each_cpu() is very slow.
3060 *
3061 * Return:
3062 * 0 on success, -errno on failure.
3063 */
3064int schedule_on_each_cpu(work_func_t func)
3065{
3066 int cpu;
3067 struct work_struct __percpu *works;
3068
3069 works = alloc_percpu(struct work_struct);
3070 if (!works)
3071 return -ENOMEM;
3072
3073 get_online_cpus();
3074
3075 for_each_online_cpu(cpu) {
3076 struct work_struct *work = per_cpu_ptr(works, cpu);
3077
3078 INIT_WORK(work, func);
3079 schedule_work_on(cpu, work);
3080 }
3081
3082 for_each_online_cpu(cpu)
3083 flush_work(per_cpu_ptr(works, cpu));
3084
3085 put_online_cpus();
3086 free_percpu(works);
3087 return 0;
3088}
3089
3090/**
3091 * execute_in_process_context - reliably execute the routine with user context
3092 * @fn: the function to execute
3093 * @ew: guaranteed storage for the execute work structure (must
3094 * be available when the work executes)
3095 *
3096 * Executes the function immediately if process context is available,
3097 * otherwise schedules the function for delayed execution.
3098 *
3099 * Return: 0 - function was executed
3100 * 1 - function was scheduled for execution
3101 */
3102int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3103{
3104 if (!in_interrupt()) {
3105 fn(&ew->work);
3106 return 0;
3107 }
3108
3109 INIT_WORK(&ew->work, fn);
3110 schedule_work(&ew->work);
3111
3112 return 1;
3113}
3114EXPORT_SYMBOL_GPL(execute_in_process_context);
3115
3116/**
3117 * free_workqueue_attrs - free a workqueue_attrs
3118 * @attrs: workqueue_attrs to free
3119 *
3120 * Undo alloc_workqueue_attrs().
3121 */
3122void free_workqueue_attrs(struct workqueue_attrs *attrs)
3123{
3124 if (attrs) {
3125 free_cpumask_var(attrs->cpumask);
3126 kfree(attrs);
3127 }
3128}
3129
3130/**
3131 * alloc_workqueue_attrs - allocate a workqueue_attrs
3132 * @gfp_mask: allocation mask to use
3133 *
3134 * Allocate a new workqueue_attrs, initialize with default settings and
3135 * return it.
3136 *
3137 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3138 */
3139struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3140{
3141 struct workqueue_attrs *attrs;
3142
3143 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3144 if (!attrs)
3145 goto fail;
3146 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3147 goto fail;
3148
3149 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3150 return attrs;
3151fail:
3152 free_workqueue_attrs(attrs);
3153 return NULL;
3154}
3155
3156static void copy_workqueue_attrs(struct workqueue_attrs *to,
3157 const struct workqueue_attrs *from)
3158{
3159 to->nice = from->nice;
3160 cpumask_copy(to->cpumask, from->cpumask);
3161 /*
3162 * Unlike hash and equality test, this function doesn't ignore
3163 * ->no_numa as it is used for both pool and wq attrs. Instead,
3164 * get_unbound_pool() explicitly clears ->no_numa after copying.
3165 */
3166 to->no_numa = from->no_numa;
3167}
3168
3169/* hash value of the content of @attr */
3170static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3171{
3172 u32 hash = 0;
3173
3174 hash = jhash_1word(attrs->nice, hash);
3175 hash = jhash(cpumask_bits(attrs->cpumask),
3176 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3177 return hash;
3178}
3179
3180/* content equality test */
3181static bool wqattrs_equal(const struct workqueue_attrs *a,
3182 const struct workqueue_attrs *b)
3183{
3184 if (a->nice != b->nice)
3185 return false;
3186 if (!cpumask_equal(a->cpumask, b->cpumask))
3187 return false;
3188 return true;
3189}
3190
3191/**
3192 * init_worker_pool - initialize a newly zalloc'd worker_pool
3193 * @pool: worker_pool to initialize
3194 *
3195 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3196 *
3197 * Return: 0 on success, -errno on failure. Even on failure, all fields
3198 * inside @pool proper are initialized and put_unbound_pool() can be called
3199 * on @pool safely to release it.
3200 */
3201static int init_worker_pool(struct worker_pool *pool)
3202{
3203 spin_lock_init(&pool->lock);
3204 pool->id = -1;
3205 pool->cpu = -1;
3206 pool->node = NUMA_NO_NODE;
3207 pool->flags |= POOL_DISASSOCIATED;
3208 pool->watchdog_ts = jiffies;
3209 INIT_LIST_HEAD(&pool->worklist);
3210 INIT_LIST_HEAD(&pool->idle_list);
3211 hash_init(pool->busy_hash);
3212
3213 init_timer_deferrable(&pool->idle_timer);
3214 pool->idle_timer.function = idle_worker_timeout;
3215 pool->idle_timer.data = (unsigned long)pool;
3216
3217 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3218 (unsigned long)pool);
3219
3220 mutex_init(&pool->manager_arb);
3221 mutex_init(&pool->attach_mutex);
3222 INIT_LIST_HEAD(&pool->workers);
3223
3224 ida_init(&pool->worker_ida);
3225 INIT_HLIST_NODE(&pool->hash_node);
3226 pool->refcnt = 1;
3227
3228 /* shouldn't fail above this point */
3229 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3230 if (!pool->attrs)
3231 return -ENOMEM;
3232 return 0;
3233}
3234
3235static void rcu_free_wq(struct rcu_head *rcu)
3236{
3237 struct workqueue_struct *wq =
3238 container_of(rcu, struct workqueue_struct, rcu);
3239
3240 if (!(wq->flags & WQ_UNBOUND))
3241 free_percpu(wq->cpu_pwqs);
3242 else
3243 free_workqueue_attrs(wq->unbound_attrs);
3244
3245 kfree(wq->rescuer);
3246 kfree(wq);
3247}
3248
3249static void rcu_free_pool(struct rcu_head *rcu)
3250{
3251 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3252
3253 ida_destroy(&pool->worker_ida);
3254 free_workqueue_attrs(pool->attrs);
3255 kfree(pool);
3256}
3257
3258/**
3259 * put_unbound_pool - put a worker_pool
3260 * @pool: worker_pool to put
3261 *
3262 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3263 * safe manner. get_unbound_pool() calls this function on its failure path
3264 * and this function should be able to release pools which went through,
3265 * successfully or not, init_worker_pool().
3266 *
3267 * Should be called with wq_pool_mutex held.
3268 */
3269static void put_unbound_pool(struct worker_pool *pool)
3270{
3271 DECLARE_COMPLETION_ONSTACK(detach_completion);
3272 struct worker *worker;
3273
3274 lockdep_assert_held(&wq_pool_mutex);
3275
3276 if (--pool->refcnt)
3277 return;
3278
3279 /* sanity checks */
3280 if (WARN_ON(!(pool->cpu < 0)) ||
3281 WARN_ON(!list_empty(&pool->worklist)))
3282 return;
3283
3284 /* release id and unhash */
3285 if (pool->id >= 0)
3286 idr_remove(&worker_pool_idr, pool->id);
3287 hash_del(&pool->hash_node);
3288
3289 /*
3290 * Become the manager and destroy all workers. Grabbing
3291 * manager_arb prevents @pool's workers from blocking on
3292 * attach_mutex.
3293 */
3294 mutex_lock(&pool->manager_arb);
3295
3296 spin_lock_irq(&pool->lock);
3297 while ((worker = first_idle_worker(pool)))
3298 destroy_worker(worker);
3299 WARN_ON(pool->nr_workers || pool->nr_idle);
3300 spin_unlock_irq(&pool->lock);
3301
3302 mutex_lock(&pool->attach_mutex);
3303 if (!list_empty(&pool->workers))
3304 pool->detach_completion = &detach_completion;
3305 mutex_unlock(&pool->attach_mutex);
3306
3307 if (pool->detach_completion)
3308 wait_for_completion(pool->detach_completion);
3309
3310 mutex_unlock(&pool->manager_arb);
3311
3312 /* shut down the timers */
3313 del_timer_sync(&pool->idle_timer);
3314 del_timer_sync(&pool->mayday_timer);
3315
3316 /* sched-RCU protected to allow dereferences from get_work_pool() */
3317 call_rcu_sched(&pool->rcu, rcu_free_pool);
3318}
3319
3320/**
3321 * get_unbound_pool - get a worker_pool with the specified attributes
3322 * @attrs: the attributes of the worker_pool to get
3323 *
3324 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3325 * reference count and return it. If there already is a matching
3326 * worker_pool, it will be used; otherwise, this function attempts to
3327 * create a new one.
3328 *
3329 * Should be called with wq_pool_mutex held.
3330 *
3331 * Return: On success, a worker_pool with the same attributes as @attrs.
3332 * On failure, %NULL.
3333 */
3334static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3335{
3336 u32 hash = wqattrs_hash(attrs);
3337 struct worker_pool *pool;
3338 int node;
3339 int target_node = NUMA_NO_NODE;
3340
3341 lockdep_assert_held(&wq_pool_mutex);
3342
3343 /* do we already have a matching pool? */
3344 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3345 if (wqattrs_equal(pool->attrs, attrs)) {
3346 pool->refcnt++;
3347 return pool;
3348 }
3349 }
3350
3351 /* if cpumask is contained inside a NUMA node, we belong to that node */
3352 if (wq_numa_enabled) {
3353 for_each_node(node) {
3354 if (cpumask_subset(attrs->cpumask,
3355 wq_numa_possible_cpumask[node])) {
3356 target_node = node;
3357 break;
3358 }
3359 }
3360 }
3361
3362 /* nope, create a new one */
3363 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3364 if (!pool || init_worker_pool(pool) < 0)
3365 goto fail;
3366
3367 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3368 copy_workqueue_attrs(pool->attrs, attrs);
3369 pool->node = target_node;
3370
3371 /*
3372 * no_numa isn't a worker_pool attribute, always clear it. See
3373 * 'struct workqueue_attrs' comments for detail.
3374 */
3375 pool->attrs->no_numa = false;
3376
3377 if (worker_pool_assign_id(pool) < 0)
3378 goto fail;
3379
3380 /* create and start the initial worker */
3381 if (!create_worker(pool))
3382 goto fail;
3383
3384 /* install */
3385 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3386
3387 return pool;
3388fail:
3389 if (pool)
3390 put_unbound_pool(pool);
3391 return NULL;
3392}
3393
3394static void rcu_free_pwq(struct rcu_head *rcu)
3395{
3396 kmem_cache_free(pwq_cache,
3397 container_of(rcu, struct pool_workqueue, rcu));
3398}
3399
3400/*
3401 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3402 * and needs to be destroyed.
3403 */
3404static void pwq_unbound_release_workfn(struct work_struct *work)
3405{
3406 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3407 unbound_release_work);
3408 struct workqueue_struct *wq = pwq->wq;
3409 struct worker_pool *pool = pwq->pool;
3410 bool is_last;
3411
3412 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3413 return;
3414
3415 mutex_lock(&wq->mutex);
3416 list_del_rcu(&pwq->pwqs_node);
3417 is_last = list_empty(&wq->pwqs);
3418 mutex_unlock(&wq->mutex);
3419
3420 mutex_lock(&wq_pool_mutex);
3421 put_unbound_pool(pool);
3422 mutex_unlock(&wq_pool_mutex);
3423
3424 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3425
3426 /*
3427 * If we're the last pwq going away, @wq is already dead and no one
3428 * is gonna access it anymore. Schedule RCU free.
3429 */
3430 if (is_last)
3431 call_rcu_sched(&wq->rcu, rcu_free_wq);
3432}
3433
3434/**
3435 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3436 * @pwq: target pool_workqueue
3437 *
3438 * If @pwq isn't freezing, set @pwq->max_active to the associated
3439 * workqueue's saved_max_active and activate delayed work items
3440 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3441 */
3442static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3443{
3444 struct workqueue_struct *wq = pwq->wq;
3445 bool freezable = wq->flags & WQ_FREEZABLE;
3446
3447 /* for @wq->saved_max_active */
3448 lockdep_assert_held(&wq->mutex);
3449
3450 /* fast exit for non-freezable wqs */
3451 if (!freezable && pwq->max_active == wq->saved_max_active)
3452 return;
3453
3454 spin_lock_irq(&pwq->pool->lock);
3455
3456 /*
3457 * During [un]freezing, the caller is responsible for ensuring that
3458 * this function is called at least once after @workqueue_freezing
3459 * is updated and visible.
3460 */
3461 if (!freezable || !workqueue_freezing) {
3462 pwq->max_active = wq->saved_max_active;
3463
3464 while (!list_empty(&pwq->delayed_works) &&
3465 pwq->nr_active < pwq->max_active)
3466 pwq_activate_first_delayed(pwq);
3467
3468 /*
3469 * Need to kick a worker after thawed or an unbound wq's
3470 * max_active is bumped. It's a slow path. Do it always.
3471 */
3472 wake_up_worker(pwq->pool);
3473 } else {
3474 pwq->max_active = 0;
3475 }
3476
3477 spin_unlock_irq(&pwq->pool->lock);
3478}
3479
3480/* initialize newly alloced @pwq which is associated with @wq and @pool */
3481static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3482 struct worker_pool *pool)
3483{
3484 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
3486 memset(pwq, 0, sizeof(*pwq));
3487
3488 pwq->pool = pool;
3489 pwq->wq = wq;
3490 pwq->flush_color = -1;
3491 pwq->refcnt = 1;
3492 INIT_LIST_HEAD(&pwq->delayed_works);
3493 INIT_LIST_HEAD(&pwq->pwqs_node);
3494 INIT_LIST_HEAD(&pwq->mayday_node);
3495 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3496}
3497
3498/* sync @pwq with the current state of its associated wq and link it */
3499static void link_pwq(struct pool_workqueue *pwq)
3500{
3501 struct workqueue_struct *wq = pwq->wq;
3502
3503 lockdep_assert_held(&wq->mutex);
3504
3505 /* may be called multiple times, ignore if already linked */
3506 if (!list_empty(&pwq->pwqs_node))
3507 return;
3508
3509 /* set the matching work_color */
3510 pwq->work_color = wq->work_color;
3511
3512 /* sync max_active to the current setting */
3513 pwq_adjust_max_active(pwq);
3514
3515 /* link in @pwq */
3516 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3517}
3518
3519/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3521 const struct workqueue_attrs *attrs)
3522{
3523 struct worker_pool *pool;
3524 struct pool_workqueue *pwq;
3525
3526 lockdep_assert_held(&wq_pool_mutex);
3527
3528 pool = get_unbound_pool(attrs);
3529 if (!pool)
3530 return NULL;
3531
3532 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3533 if (!pwq) {
3534 put_unbound_pool(pool);
3535 return NULL;
3536 }
3537
3538 init_pwq(pwq, wq, pool);
3539 return pwq;
3540}
3541
3542/**
3543 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3544 * @attrs: the wq_attrs of the default pwq of the target workqueue
3545 * @node: the target NUMA node
3546 * @cpu_going_down: if >= 0, the CPU to consider as offline
3547 * @cpumask: outarg, the resulting cpumask
3548 *
3549 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3550 * @cpu_going_down is >= 0, that cpu is considered offline during
3551 * calculation. The result is stored in @cpumask.
3552 *
3553 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3554 * enabled and @node has online CPUs requested by @attrs, the returned
3555 * cpumask is the intersection of the possible CPUs of @node and
3556 * @attrs->cpumask.
3557 *
3558 * The caller is responsible for ensuring that the cpumask of @node stays
3559 * stable.
3560 *
3561 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562 * %false if equal.
3563 */
3564static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565 int cpu_going_down, cpumask_t *cpumask)
3566{
3567 if (!wq_numa_enabled || attrs->no_numa)
3568 goto use_dfl;
3569
3570 /* does @node have any online CPUs @attrs wants? */
3571 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572 if (cpu_going_down >= 0)
3573 cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575 if (cpumask_empty(cpumask))
3576 goto use_dfl;
3577
3578 /* yeap, return possible CPUs in @node that @attrs wants */
3579 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580 return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582use_dfl:
3583 cpumask_copy(cpumask, attrs->cpumask);
3584 return false;
3585}
3586
3587/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589 int node,
3590 struct pool_workqueue *pwq)
3591{
3592 struct pool_workqueue *old_pwq;
3593
3594 lockdep_assert_held(&wq_pool_mutex);
3595 lockdep_assert_held(&wq->mutex);
3596
3597 /* link_pwq() can handle duplicate calls */
3598 link_pwq(pwq);
3599
3600 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602 return old_pwq;
3603}
3604
3605/* context to store the prepared attrs & pwqs before applying */
3606struct apply_wqattrs_ctx {
3607 struct workqueue_struct *wq; /* target workqueue */
3608 struct workqueue_attrs *attrs; /* attrs to apply */
3609 struct list_head list; /* queued for batching commit */
3610 struct pool_workqueue *dfl_pwq;
3611 struct pool_workqueue *pwq_tbl[];
3612};
3613
3614/* free the resources after success or abort */
3615static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616{
3617 if (ctx) {
3618 int node;
3619
3620 for_each_node(node)
3621 put_pwq_unlocked(ctx->pwq_tbl[node]);
3622 put_pwq_unlocked(ctx->dfl_pwq);
3623
3624 free_workqueue_attrs(ctx->attrs);
3625
3626 kfree(ctx);
3627 }
3628}
3629
3630/* allocate the attrs and pwqs for later installation */
3631static struct apply_wqattrs_ctx *
3632apply_wqattrs_prepare(struct workqueue_struct *wq,
3633 const struct workqueue_attrs *attrs)
3634{
3635 struct apply_wqattrs_ctx *ctx;
3636 struct workqueue_attrs *new_attrs, *tmp_attrs;
3637 int node;
3638
3639 lockdep_assert_held(&wq_pool_mutex);
3640
3641 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642 GFP_KERNEL);
3643
3644 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646 if (!ctx || !new_attrs || !tmp_attrs)
3647 goto out_free;
3648
3649 /*
3650 * Calculate the attrs of the default pwq.
3651 * If the user configured cpumask doesn't overlap with the
3652 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653 */
3654 copy_workqueue_attrs(new_attrs, attrs);
3655 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658
3659 /*
3660 * We may create multiple pwqs with differing cpumasks. Make a
3661 * copy of @new_attrs which will be modified and used to obtain
3662 * pools.
3663 */
3664 copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666 /*
3667 * If something goes wrong during CPU up/down, we'll fall back to
3668 * the default pwq covering whole @attrs->cpumask. Always create
3669 * it even if we don't use it immediately.
3670 */
3671 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672 if (!ctx->dfl_pwq)
3673 goto out_free;
3674
3675 for_each_node(node) {
3676 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678 if (!ctx->pwq_tbl[node])
3679 goto out_free;
3680 } else {
3681 ctx->dfl_pwq->refcnt++;
3682 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683 }
3684 }
3685
3686 /* save the user configured attrs and sanitize it. */
3687 copy_workqueue_attrs(new_attrs, attrs);
3688 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689 ctx->attrs = new_attrs;
3690
3691 ctx->wq = wq;
3692 free_workqueue_attrs(tmp_attrs);
3693 return ctx;
3694
3695out_free:
3696 free_workqueue_attrs(tmp_attrs);
3697 free_workqueue_attrs(new_attrs);
3698 apply_wqattrs_cleanup(ctx);
3699 return NULL;
3700}
3701
3702/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704{
3705 int node;
3706
3707 /* all pwqs have been created successfully, let's install'em */
3708 mutex_lock(&ctx->wq->mutex);
3709
3710 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712 /* save the previous pwq and install the new one */
3713 for_each_node(node)
3714 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715 ctx->pwq_tbl[node]);
3716
3717 /* @dfl_pwq might not have been used, ensure it's linked */
3718 link_pwq(ctx->dfl_pwq);
3719 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721 mutex_unlock(&ctx->wq->mutex);
3722}
3723
3724static void apply_wqattrs_lock(void)
3725{
3726 /* CPUs should stay stable across pwq creations and installations */
3727 get_online_cpus();
3728 mutex_lock(&wq_pool_mutex);
3729}
3730
3731static void apply_wqattrs_unlock(void)
3732{
3733 mutex_unlock(&wq_pool_mutex);
3734 put_online_cpus();
3735}
3736
3737static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738 const struct workqueue_attrs *attrs)
3739{
3740 struct apply_wqattrs_ctx *ctx;
3741
3742 /* only unbound workqueues can change attributes */
3743 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744 return -EINVAL;
3745
3746 /* creating multiple pwqs breaks ordering guarantee */
3747 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748 return -EINVAL;
3749
3750 ctx = apply_wqattrs_prepare(wq, attrs);
3751 if (!ctx)
3752 return -ENOMEM;
3753
3754 /* the ctx has been prepared successfully, let's commit it */
3755 apply_wqattrs_commit(ctx);
3756 apply_wqattrs_cleanup(ctx);
3757
3758 return 0;
3759}
3760
3761/**
3762 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763 * @wq: the target workqueue
3764 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765 *
3766 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3767 * machines, this function maps a separate pwq to each NUMA node with
3768 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3769 * NUMA node it was issued on. Older pwqs are released as in-flight work
3770 * items finish. Note that a work item which repeatedly requeues itself
3771 * back-to-back will stay on its current pwq.
3772 *
3773 * Performs GFP_KERNEL allocations.
3774 *
3775 * Return: 0 on success and -errno on failure.
3776 */
3777int apply_workqueue_attrs(struct workqueue_struct *wq,
3778 const struct workqueue_attrs *attrs)
3779{
3780 int ret;
3781
3782 apply_wqattrs_lock();
3783 ret = apply_workqueue_attrs_locked(wq, attrs);
3784 apply_wqattrs_unlock();
3785
3786 return ret;
3787}
3788
3789/**
3790 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791 * @wq: the target workqueue
3792 * @cpu: the CPU coming up or going down
3793 * @online: whether @cpu is coming up or going down
3794 *
3795 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3797 * @wq accordingly.
3798 *
3799 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3800 * falls back to @wq->dfl_pwq which may not be optimal but is always
3801 * correct.
3802 *
3803 * Note that when the last allowed CPU of a NUMA node goes offline for a
3804 * workqueue with a cpumask spanning multiple nodes, the workers which were
3805 * already executing the work items for the workqueue will lose their CPU
3806 * affinity and may execute on any CPU. This is similar to how per-cpu
3807 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3808 * affinity, it's the user's responsibility to flush the work item from
3809 * CPU_DOWN_PREPARE.
3810 */
3811static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812 bool online)
3813{
3814 int node = cpu_to_node(cpu);
3815 int cpu_off = online ? -1 : cpu;
3816 struct pool_workqueue *old_pwq = NULL, *pwq;
3817 struct workqueue_attrs *target_attrs;
3818 cpumask_t *cpumask;
3819
3820 lockdep_assert_held(&wq_pool_mutex);
3821
3822 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3823 wq->unbound_attrs->no_numa)
3824 return;
3825
3826 /*
3827 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828 * Let's use a preallocated one. The following buf is protected by
3829 * CPU hotplug exclusion.
3830 */
3831 target_attrs = wq_update_unbound_numa_attrs_buf;
3832 cpumask = target_attrs->cpumask;
3833
3834 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835 pwq = unbound_pwq_by_node(wq, node);
3836
3837 /*
3838 * Let's determine what needs to be done. If the target cpumask is
3839 * different from the default pwq's, we need to compare it to @pwq's
3840 * and create a new one if they don't match. If the target cpumask
3841 * equals the default pwq's, the default pwq should be used.
3842 */
3843 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3844 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3845 return;
3846 } else {
3847 goto use_dfl_pwq;
3848 }
3849
3850 /* create a new pwq */
3851 pwq = alloc_unbound_pwq(wq, target_attrs);
3852 if (!pwq) {
3853 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854 wq->name);
3855 goto use_dfl_pwq;
3856 }
3857
3858 /* Install the new pwq. */
3859 mutex_lock(&wq->mutex);
3860 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861 goto out_unlock;
3862
3863use_dfl_pwq:
3864 mutex_lock(&wq->mutex);
3865 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866 get_pwq(wq->dfl_pwq);
3867 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869out_unlock:
3870 mutex_unlock(&wq->mutex);
3871 put_pwq_unlocked(old_pwq);
3872}
3873
3874static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3875{
3876 bool highpri = wq->flags & WQ_HIGHPRI;
3877 int cpu, ret;
3878
3879 if (!(wq->flags & WQ_UNBOUND)) {
3880 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881 if (!wq->cpu_pwqs)
3882 return -ENOMEM;
3883
3884 for_each_possible_cpu(cpu) {
3885 struct pool_workqueue *pwq =
3886 per_cpu_ptr(wq->cpu_pwqs, cpu);
3887 struct worker_pool *cpu_pools =
3888 per_cpu(cpu_worker_pools, cpu);
3889
3890 init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892 mutex_lock(&wq->mutex);
3893 link_pwq(pwq);
3894 mutex_unlock(&wq->mutex);
3895 }
3896 return 0;
3897 } else if (wq->flags & __WQ_ORDERED) {
3898 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899 /* there should only be single pwq for ordering guarantee */
3900 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3901 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902 "ordering guarantee broken for workqueue %s\n", wq->name);
3903 return ret;
3904 } else {
3905 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3906 }
3907}
3908
3909static int wq_clamp_max_active(int max_active, unsigned int flags,
3910 const char *name)
3911{
3912 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914 if (max_active < 1 || max_active > lim)
3915 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916 max_active, name, 1, lim);
3917
3918 return clamp_val(max_active, 1, lim);
3919}
3920
3921struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3922 unsigned int flags,
3923 int max_active,
3924 struct lock_class_key *key,
3925 const char *lock_name, ...)
3926{
3927 size_t tbl_size = 0;
3928 va_list args;
3929 struct workqueue_struct *wq;
3930 struct pool_workqueue *pwq;
3931
3932 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3933 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934 flags |= WQ_UNBOUND;
3935
3936 /* allocate wq and format name */
3937 if (flags & WQ_UNBOUND)
3938 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3939
3940 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3941 if (!wq)
3942 return NULL;
3943
3944 if (flags & WQ_UNBOUND) {
3945 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946 if (!wq->unbound_attrs)
3947 goto err_free_wq;
3948 }
3949
3950 va_start(args, lock_name);
3951 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3952 va_end(args);
3953
3954 max_active = max_active ?: WQ_DFL_ACTIVE;
3955 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3956
3957 /* init wq */
3958 wq->flags = flags;
3959 wq->saved_max_active = max_active;
3960 mutex_init(&wq->mutex);
3961 atomic_set(&wq->nr_pwqs_to_flush, 0);
3962 INIT_LIST_HEAD(&wq->pwqs);
3963 INIT_LIST_HEAD(&wq->flusher_queue);
3964 INIT_LIST_HEAD(&wq->flusher_overflow);
3965 INIT_LIST_HEAD(&wq->maydays);
3966
3967 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3968 INIT_LIST_HEAD(&wq->list);
3969
3970 if (alloc_and_link_pwqs(wq) < 0)
3971 goto err_free_wq;
3972
3973 /*
3974 * Workqueues which may be used during memory reclaim should
3975 * have a rescuer to guarantee forward progress.
3976 */
3977 if (flags & WQ_MEM_RECLAIM) {
3978 struct worker *rescuer;
3979
3980 rescuer = alloc_worker(NUMA_NO_NODE);
3981 if (!rescuer)
3982 goto err_destroy;
3983
3984 rescuer->rescue_wq = wq;
3985 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3986 wq->name);
3987 if (IS_ERR(rescuer->task)) {
3988 kfree(rescuer);
3989 goto err_destroy;
3990 }
3991
3992 wq->rescuer = rescuer;
3993 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3994 wake_up_process(rescuer->task);
3995 }
3996
3997 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998 goto err_destroy;
3999
4000 /*
4001 * wq_pool_mutex protects global freeze state and workqueues list.
4002 * Grab it, adjust max_active and add the new @wq to workqueues
4003 * list.
4004 */
4005 mutex_lock(&wq_pool_mutex);
4006
4007 mutex_lock(&wq->mutex);
4008 for_each_pwq(pwq, wq)
4009 pwq_adjust_max_active(pwq);
4010 mutex_unlock(&wq->mutex);
4011
4012 list_add_tail_rcu(&wq->list, &workqueues);
4013
4014 mutex_unlock(&wq_pool_mutex);
4015
4016 return wq;
4017
4018err_free_wq:
4019 free_workqueue_attrs(wq->unbound_attrs);
4020 kfree(wq);
4021 return NULL;
4022err_destroy:
4023 destroy_workqueue(wq);
4024 return NULL;
4025}
4026EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4027
4028/**
4029 * destroy_workqueue - safely terminate a workqueue
4030 * @wq: target workqueue
4031 *
4032 * Safely destroy a workqueue. All work currently pending will be done first.
4033 */
4034void destroy_workqueue(struct workqueue_struct *wq)
4035{
4036 struct pool_workqueue *pwq;
4037 int node;
4038
4039 /* drain it before proceeding with destruction */
4040 drain_workqueue(wq);
4041
4042 /* sanity checks */
4043 mutex_lock(&wq->mutex);
4044 for_each_pwq(pwq, wq) {
4045 int i;
4046
4047 for (i = 0; i < WORK_NR_COLORS; i++) {
4048 if (WARN_ON(pwq->nr_in_flight[i])) {
4049 mutex_unlock(&wq->mutex);
4050 return;
4051 }
4052 }
4053
4054 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4055 WARN_ON(pwq->nr_active) ||
4056 WARN_ON(!list_empty(&pwq->delayed_works))) {
4057 mutex_unlock(&wq->mutex);
4058 return;
4059 }
4060 }
4061 mutex_unlock(&wq->mutex);
4062
4063 /*
4064 * wq list is used to freeze wq, remove from list after
4065 * flushing is complete in case freeze races us.
4066 */
4067 mutex_lock(&wq_pool_mutex);
4068 list_del_rcu(&wq->list);
4069 mutex_unlock(&wq_pool_mutex);
4070
4071 workqueue_sysfs_unregister(wq);
4072
4073 if (wq->rescuer)
4074 kthread_stop(wq->rescuer->task);
4075
4076 if (!(wq->flags & WQ_UNBOUND)) {
4077 /*
4078 * The base ref is never dropped on per-cpu pwqs. Directly
4079 * schedule RCU free.
4080 */
4081 call_rcu_sched(&wq->rcu, rcu_free_wq);
4082 } else {
4083 /*
4084 * We're the sole accessor of @wq at this point. Directly
4085 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4086 * @wq will be freed when the last pwq is released.
4087 */
4088 for_each_node(node) {
4089 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4090 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4091 put_pwq_unlocked(pwq);
4092 }
4093
4094 /*
4095 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4096 * put. Don't access it afterwards.
4097 */
4098 pwq = wq->dfl_pwq;
4099 wq->dfl_pwq = NULL;
4100 put_pwq_unlocked(pwq);
4101 }
4102}
4103EXPORT_SYMBOL_GPL(destroy_workqueue);
4104
4105/**
4106 * workqueue_set_max_active - adjust max_active of a workqueue
4107 * @wq: target workqueue
4108 * @max_active: new max_active value.
4109 *
4110 * Set max_active of @wq to @max_active.
4111 *
4112 * CONTEXT:
4113 * Don't call from IRQ context.
4114 */
4115void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4116{
4117 struct pool_workqueue *pwq;
4118
4119 /* disallow meddling with max_active for ordered workqueues */
4120 if (WARN_ON(wq->flags & __WQ_ORDERED))
4121 return;
4122
4123 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4124
4125 mutex_lock(&wq->mutex);
4126
4127 wq->saved_max_active = max_active;
4128
4129 for_each_pwq(pwq, wq)
4130 pwq_adjust_max_active(pwq);
4131
4132 mutex_unlock(&wq->mutex);
4133}
4134EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4135
4136/**
4137 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4138 *
4139 * Determine whether %current is a workqueue rescuer. Can be used from
4140 * work functions to determine whether it's being run off the rescuer task.
4141 *
4142 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4143 */
4144bool current_is_workqueue_rescuer(void)
4145{
4146 struct worker *worker = current_wq_worker();
4147
4148 return worker && worker->rescue_wq;
4149}
4150
4151/**
4152 * workqueue_congested - test whether a workqueue is congested
4153 * @cpu: CPU in question
4154 * @wq: target workqueue
4155 *
4156 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4157 * no synchronization around this function and the test result is
4158 * unreliable and only useful as advisory hints or for debugging.
4159 *
4160 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4161 * Note that both per-cpu and unbound workqueues may be associated with
4162 * multiple pool_workqueues which have separate congested states. A
4163 * workqueue being congested on one CPU doesn't mean the workqueue is also
4164 * contested on other CPUs / NUMA nodes.
4165 *
4166 * Return:
4167 * %true if congested, %false otherwise.
4168 */
4169bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4170{
4171 struct pool_workqueue *pwq;
4172 bool ret;
4173
4174 rcu_read_lock_sched();
4175
4176 if (cpu == WORK_CPU_UNBOUND)
4177 cpu = smp_processor_id();
4178
4179 if (!(wq->flags & WQ_UNBOUND))
4180 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4181 else
4182 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4183
4184 ret = !list_empty(&pwq->delayed_works);
4185 rcu_read_unlock_sched();
4186
4187 return ret;
4188}
4189EXPORT_SYMBOL_GPL(workqueue_congested);
4190
4191/**
4192 * work_busy - test whether a work is currently pending or running
4193 * @work: the work to be tested
4194 *
4195 * Test whether @work is currently pending or running. There is no
4196 * synchronization around this function and the test result is
4197 * unreliable and only useful as advisory hints or for debugging.
4198 *
4199 * Return:
4200 * OR'd bitmask of WORK_BUSY_* bits.
4201 */
4202unsigned int work_busy(struct work_struct *work)
4203{
4204 struct worker_pool *pool;
4205 unsigned long flags;
4206 unsigned int ret = 0;
4207
4208 if (work_pending(work))
4209 ret |= WORK_BUSY_PENDING;
4210
4211 local_irq_save(flags);
4212 pool = get_work_pool(work);
4213 if (pool) {
4214 spin_lock(&pool->lock);
4215 if (find_worker_executing_work(pool, work))
4216 ret |= WORK_BUSY_RUNNING;
4217 spin_unlock(&pool->lock);
4218 }
4219 local_irq_restore(flags);
4220
4221 return ret;
4222}
4223EXPORT_SYMBOL_GPL(work_busy);
4224
4225/**
4226 * set_worker_desc - set description for the current work item
4227 * @fmt: printf-style format string
4228 * @...: arguments for the format string
4229 *
4230 * This function can be called by a running work function to describe what
4231 * the work item is about. If the worker task gets dumped, this
4232 * information will be printed out together to help debugging. The
4233 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4234 */
4235void set_worker_desc(const char *fmt, ...)
4236{
4237 struct worker *worker = current_wq_worker();
4238 va_list args;
4239
4240 if (worker) {
4241 va_start(args, fmt);
4242 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4243 va_end(args);
4244 worker->desc_valid = true;
4245 }
4246}
4247
4248/**
4249 * print_worker_info - print out worker information and description
4250 * @log_lvl: the log level to use when printing
4251 * @task: target task
4252 *
4253 * If @task is a worker and currently executing a work item, print out the
4254 * name of the workqueue being serviced and worker description set with
4255 * set_worker_desc() by the currently executing work item.
4256 *
4257 * This function can be safely called on any task as long as the
4258 * task_struct itself is accessible. While safe, this function isn't
4259 * synchronized and may print out mixups or garbages of limited length.
4260 */
4261void print_worker_info(const char *log_lvl, struct task_struct *task)
4262{
4263 work_func_t *fn = NULL;
4264 char name[WQ_NAME_LEN] = { };
4265 char desc[WORKER_DESC_LEN] = { };
4266 struct pool_workqueue *pwq = NULL;
4267 struct workqueue_struct *wq = NULL;
4268 bool desc_valid = false;
4269 struct worker *worker;
4270
4271 if (!(task->flags & PF_WQ_WORKER))
4272 return;
4273
4274 /*
4275 * This function is called without any synchronization and @task
4276 * could be in any state. Be careful with dereferences.
4277 */
4278 worker = probe_kthread_data(task);
4279
4280 /*
4281 * Carefully copy the associated workqueue's workfn and name. Keep
4282 * the original last '\0' in case the original contains garbage.
4283 */
4284 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4285 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4286 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4287 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4288
4289 /* copy worker description */
4290 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4291 if (desc_valid)
4292 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4293
4294 if (fn || name[0] || desc[0]) {
4295 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4296 if (desc[0])
4297 pr_cont(" (%s)", desc);
4298 pr_cont("\n");
4299 }
4300}
4301
4302static void pr_cont_pool_info(struct worker_pool *pool)
4303{
4304 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4305 if (pool->node != NUMA_NO_NODE)
4306 pr_cont(" node=%d", pool->node);
4307 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4308}
4309
4310static void pr_cont_work(bool comma, struct work_struct *work)
4311{
4312 if (work->func == wq_barrier_func) {
4313 struct wq_barrier *barr;
4314
4315 barr = container_of(work, struct wq_barrier, work);
4316
4317 pr_cont("%s BAR(%d)", comma ? "," : "",
4318 task_pid_nr(barr->task));
4319 } else {
4320 pr_cont("%s %pf", comma ? "," : "", work->func);
4321 }
4322}
4323
4324static void show_pwq(struct pool_workqueue *pwq)
4325{
4326 struct worker_pool *pool = pwq->pool;
4327 struct work_struct *work;
4328 struct worker *worker;
4329 bool has_in_flight = false, has_pending = false;
4330 int bkt;
4331
4332 pr_info(" pwq %d:", pool->id);
4333 pr_cont_pool_info(pool);
4334
4335 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4336 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4337
4338 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4339 if (worker->current_pwq == pwq) {
4340 has_in_flight = true;
4341 break;
4342 }
4343 }
4344 if (has_in_flight) {
4345 bool comma = false;
4346
4347 pr_info(" in-flight:");
4348 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4349 if (worker->current_pwq != pwq)
4350 continue;
4351
4352 pr_cont("%s %d%s:%pf", comma ? "," : "",
4353 task_pid_nr(worker->task),
4354 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4355 worker->current_func);
4356 list_for_each_entry(work, &worker->scheduled, entry)
4357 pr_cont_work(false, work);
4358 comma = true;
4359 }
4360 pr_cont("\n");
4361 }
4362
4363 list_for_each_entry(work, &pool->worklist, entry) {
4364 if (get_work_pwq(work) == pwq) {
4365 has_pending = true;
4366 break;
4367 }
4368 }
4369 if (has_pending) {
4370 bool comma = false;
4371
4372 pr_info(" pending:");
4373 list_for_each_entry(work, &pool->worklist, entry) {
4374 if (get_work_pwq(work) != pwq)
4375 continue;
4376
4377 pr_cont_work(comma, work);
4378 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4379 }
4380 pr_cont("\n");
4381 }
4382
4383 if (!list_empty(&pwq->delayed_works)) {
4384 bool comma = false;
4385
4386 pr_info(" delayed:");
4387 list_for_each_entry(work, &pwq->delayed_works, entry) {
4388 pr_cont_work(comma, work);
4389 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4390 }
4391 pr_cont("\n");
4392 }
4393}
4394
4395/**
4396 * show_workqueue_state - dump workqueue state
4397 *
4398 * Called from a sysrq handler and prints out all busy workqueues and
4399 * pools.
4400 */
4401void show_workqueue_state(void)
4402{
4403 struct workqueue_struct *wq;
4404 struct worker_pool *pool;
4405 unsigned long flags;
4406 int pi;
4407
4408 rcu_read_lock_sched();
4409
4410 pr_info("Showing busy workqueues and worker pools:\n");
4411
4412 list_for_each_entry_rcu(wq, &workqueues, list) {
4413 struct pool_workqueue *pwq;
4414 bool idle = true;
4415
4416 for_each_pwq(pwq, wq) {
4417 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4418 idle = false;
4419 break;
4420 }
4421 }
4422 if (idle)
4423 continue;
4424
4425 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4426
4427 for_each_pwq(pwq, wq) {
4428 spin_lock_irqsave(&pwq->pool->lock, flags);
4429 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4430 show_pwq(pwq);
4431 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4432 }
4433 }
4434
4435 for_each_pool(pool, pi) {
4436 struct worker *worker;
4437 bool first = true;
4438
4439 spin_lock_irqsave(&pool->lock, flags);
4440 if (pool->nr_workers == pool->nr_idle)
4441 goto next_pool;
4442
4443 pr_info("pool %d:", pool->id);
4444 pr_cont_pool_info(pool);
4445 pr_cont(" hung=%us workers=%d",
4446 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4447 pool->nr_workers);
4448 if (pool->manager)
4449 pr_cont(" manager: %d",
4450 task_pid_nr(pool->manager->task));
4451 list_for_each_entry(worker, &pool->idle_list, entry) {
4452 pr_cont(" %s%d", first ? "idle: " : "",
4453 task_pid_nr(worker->task));
4454 first = false;
4455 }
4456 pr_cont("\n");
4457 next_pool:
4458 spin_unlock_irqrestore(&pool->lock, flags);
4459 }
4460
4461 rcu_read_unlock_sched();
4462}
4463
4464/*
4465 * CPU hotplug.
4466 *
4467 * There are two challenges in supporting CPU hotplug. Firstly, there
4468 * are a lot of assumptions on strong associations among work, pwq and
4469 * pool which make migrating pending and scheduled works very
4470 * difficult to implement without impacting hot paths. Secondly,
4471 * worker pools serve mix of short, long and very long running works making
4472 * blocked draining impractical.
4473 *
4474 * This is solved by allowing the pools to be disassociated from the CPU
4475 * running as an unbound one and allowing it to be reattached later if the
4476 * cpu comes back online.
4477 */
4478
4479static void wq_unbind_fn(struct work_struct *work)
4480{
4481 int cpu = smp_processor_id();
4482 struct worker_pool *pool;
4483 struct worker *worker;
4484
4485 for_each_cpu_worker_pool(pool, cpu) {
4486 mutex_lock(&pool->attach_mutex);
4487 spin_lock_irq(&pool->lock);
4488
4489 /*
4490 * We've blocked all attach/detach operations. Make all workers
4491 * unbound and set DISASSOCIATED. Before this, all workers
4492 * except for the ones which are still executing works from
4493 * before the last CPU down must be on the cpu. After
4494 * this, they may become diasporas.
4495 */
4496 for_each_pool_worker(worker, pool)
4497 worker->flags |= WORKER_UNBOUND;
4498
4499 pool->flags |= POOL_DISASSOCIATED;
4500
4501 spin_unlock_irq(&pool->lock);
4502 mutex_unlock(&pool->attach_mutex);
4503
4504 /*
4505 * Call schedule() so that we cross rq->lock and thus can
4506 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4507 * This is necessary as scheduler callbacks may be invoked
4508 * from other cpus.
4509 */
4510 schedule();
4511
4512 /*
4513 * Sched callbacks are disabled now. Zap nr_running.
4514 * After this, nr_running stays zero and need_more_worker()
4515 * and keep_working() are always true as long as the
4516 * worklist is not empty. This pool now behaves as an
4517 * unbound (in terms of concurrency management) pool which
4518 * are served by workers tied to the pool.
4519 */
4520 atomic_set(&pool->nr_running, 0);
4521
4522 /*
4523 * With concurrency management just turned off, a busy
4524 * worker blocking could lead to lengthy stalls. Kick off
4525 * unbound chain execution of currently pending work items.
4526 */
4527 spin_lock_irq(&pool->lock);
4528 wake_up_worker(pool);
4529 spin_unlock_irq(&pool->lock);
4530 }
4531}
4532
4533/**
4534 * rebind_workers - rebind all workers of a pool to the associated CPU
4535 * @pool: pool of interest
4536 *
4537 * @pool->cpu is coming online. Rebind all workers to the CPU.
4538 */
4539static void rebind_workers(struct worker_pool *pool)
4540{
4541 struct worker *worker;
4542
4543 lockdep_assert_held(&pool->attach_mutex);
4544
4545 /*
4546 * Restore CPU affinity of all workers. As all idle workers should
4547 * be on the run-queue of the associated CPU before any local
4548 * wake-ups for concurrency management happen, restore CPU affinity
4549 * of all workers first and then clear UNBOUND. As we're called
4550 * from CPU_ONLINE, the following shouldn't fail.
4551 */
4552 for_each_pool_worker(worker, pool)
4553 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4554 pool->attrs->cpumask) < 0);
4555
4556 spin_lock_irq(&pool->lock);
4557
4558 /*
4559 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4560 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4561 * being reworked and this can go away in time.
4562 */
4563 if (!(pool->flags & POOL_DISASSOCIATED)) {
4564 spin_unlock_irq(&pool->lock);
4565 return;
4566 }
4567
4568 pool->flags &= ~POOL_DISASSOCIATED;
4569
4570 for_each_pool_worker(worker, pool) {
4571 unsigned int worker_flags = worker->flags;
4572
4573 /*
4574 * A bound idle worker should actually be on the runqueue
4575 * of the associated CPU for local wake-ups targeting it to
4576 * work. Kick all idle workers so that they migrate to the
4577 * associated CPU. Doing this in the same loop as
4578 * replacing UNBOUND with REBOUND is safe as no worker will
4579 * be bound before @pool->lock is released.
4580 */
4581 if (worker_flags & WORKER_IDLE)
4582 wake_up_process(worker->task);
4583
4584 /*
4585 * We want to clear UNBOUND but can't directly call
4586 * worker_clr_flags() or adjust nr_running. Atomically
4587 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4588 * @worker will clear REBOUND using worker_clr_flags() when
4589 * it initiates the next execution cycle thus restoring
4590 * concurrency management. Note that when or whether
4591 * @worker clears REBOUND doesn't affect correctness.
4592 *
4593 * ACCESS_ONCE() is necessary because @worker->flags may be
4594 * tested without holding any lock in
4595 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4596 * fail incorrectly leading to premature concurrency
4597 * management operations.
4598 */
4599 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4600 worker_flags |= WORKER_REBOUND;
4601 worker_flags &= ~WORKER_UNBOUND;
4602 ACCESS_ONCE(worker->flags) = worker_flags;
4603 }
4604
4605 spin_unlock_irq(&pool->lock);
4606}
4607
4608/**
4609 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4610 * @pool: unbound pool of interest
4611 * @cpu: the CPU which is coming up
4612 *
4613 * An unbound pool may end up with a cpumask which doesn't have any online
4614 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4615 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4616 * online CPU before, cpus_allowed of all its workers should be restored.
4617 */
4618static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4619{
4620 static cpumask_t cpumask;
4621 struct worker *worker;
4622
4623 lockdep_assert_held(&pool->attach_mutex);
4624
4625 /* is @cpu allowed for @pool? */
4626 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4627 return;
4628
4629 /* is @cpu the only online CPU? */
4630 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4631 if (cpumask_weight(&cpumask) != 1)
4632 return;
4633
4634 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4635 for_each_pool_worker(worker, pool)
4636 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4637 pool->attrs->cpumask) < 0);
4638}
4639
4640/*
4641 * Workqueues should be brought up before normal priority CPU notifiers.
4642 * This will be registered high priority CPU notifier.
4643 */
4644static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4645 unsigned long action,
4646 void *hcpu)
4647{
4648 int cpu = (unsigned long)hcpu;
4649 struct worker_pool *pool;
4650 struct workqueue_struct *wq;
4651 int pi;
4652
4653 switch (action & ~CPU_TASKS_FROZEN) {
4654 case CPU_UP_PREPARE:
4655 for_each_cpu_worker_pool(pool, cpu) {
4656 if (pool->nr_workers)
4657 continue;
4658 if (!create_worker(pool))
4659 return NOTIFY_BAD;
4660 }
4661 break;
4662
4663 case CPU_DOWN_FAILED:
4664 case CPU_ONLINE:
4665 mutex_lock(&wq_pool_mutex);
4666
4667 for_each_pool(pool, pi) {
4668 mutex_lock(&pool->attach_mutex);
4669
4670 if (pool->cpu == cpu)
4671 rebind_workers(pool);
4672 else if (pool->cpu < 0)
4673 restore_unbound_workers_cpumask(pool, cpu);
4674
4675 mutex_unlock(&pool->attach_mutex);
4676 }
4677
4678 /* update NUMA affinity of unbound workqueues */
4679 list_for_each_entry(wq, &workqueues, list)
4680 wq_update_unbound_numa(wq, cpu, true);
4681
4682 mutex_unlock(&wq_pool_mutex);
4683 break;
4684 }
4685 return NOTIFY_OK;
4686}
4687
4688/*
4689 * Workqueues should be brought down after normal priority CPU notifiers.
4690 * This will be registered as low priority CPU notifier.
4691 */
4692static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4693 unsigned long action,
4694 void *hcpu)
4695{
4696 int cpu = (unsigned long)hcpu;
4697 struct work_struct unbind_work;
4698 struct workqueue_struct *wq;
4699
4700 switch (action & ~CPU_TASKS_FROZEN) {
4701 case CPU_DOWN_PREPARE:
4702 /* unbinding per-cpu workers should happen on the local CPU */
4703 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4704 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4705
4706 /* update NUMA affinity of unbound workqueues */
4707 mutex_lock(&wq_pool_mutex);
4708 list_for_each_entry(wq, &workqueues, list)
4709 wq_update_unbound_numa(wq, cpu, false);
4710 mutex_unlock(&wq_pool_mutex);
4711
4712 /* wait for per-cpu unbinding to finish */
4713 flush_work(&unbind_work);
4714 destroy_work_on_stack(&unbind_work);
4715 break;
4716 }
4717 return NOTIFY_OK;
4718}
4719
4720#ifdef CONFIG_SMP
4721
4722struct work_for_cpu {
4723 struct work_struct work;
4724 long (*fn)(void *);
4725 void *arg;
4726 long ret;
4727};
4728
4729static void work_for_cpu_fn(struct work_struct *work)
4730{
4731 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4732
4733 wfc->ret = wfc->fn(wfc->arg);
4734}
4735
4736/**
4737 * work_on_cpu - run a function in thread context on a particular cpu
4738 * @cpu: the cpu to run on
4739 * @fn: the function to run
4740 * @arg: the function arg
4741 *
4742 * It is up to the caller to ensure that the cpu doesn't go offline.
4743 * The caller must not hold any locks which would prevent @fn from completing.
4744 *
4745 * Return: The value @fn returns.
4746 */
4747long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4748{
4749 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4750
4751 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4752 schedule_work_on(cpu, &wfc.work);
4753 flush_work(&wfc.work);
4754 destroy_work_on_stack(&wfc.work);
4755 return wfc.ret;
4756}
4757EXPORT_SYMBOL_GPL(work_on_cpu);
4758#endif /* CONFIG_SMP */
4759
4760#ifdef CONFIG_FREEZER
4761
4762/**
4763 * freeze_workqueues_begin - begin freezing workqueues
4764 *
4765 * Start freezing workqueues. After this function returns, all freezable
4766 * workqueues will queue new works to their delayed_works list instead of
4767 * pool->worklist.
4768 *
4769 * CONTEXT:
4770 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4771 */
4772void freeze_workqueues_begin(void)
4773{
4774 struct workqueue_struct *wq;
4775 struct pool_workqueue *pwq;
4776
4777 mutex_lock(&wq_pool_mutex);
4778
4779 WARN_ON_ONCE(workqueue_freezing);
4780 workqueue_freezing = true;
4781
4782 list_for_each_entry(wq, &workqueues, list) {
4783 mutex_lock(&wq->mutex);
4784 for_each_pwq(pwq, wq)
4785 pwq_adjust_max_active(pwq);
4786 mutex_unlock(&wq->mutex);
4787 }
4788
4789 mutex_unlock(&wq_pool_mutex);
4790}
4791
4792/**
4793 * freeze_workqueues_busy - are freezable workqueues still busy?
4794 *
4795 * Check whether freezing is complete. This function must be called
4796 * between freeze_workqueues_begin() and thaw_workqueues().
4797 *
4798 * CONTEXT:
4799 * Grabs and releases wq_pool_mutex.
4800 *
4801 * Return:
4802 * %true if some freezable workqueues are still busy. %false if freezing
4803 * is complete.
4804 */
4805bool freeze_workqueues_busy(void)
4806{
4807 bool busy = false;
4808 struct workqueue_struct *wq;
4809 struct pool_workqueue *pwq;
4810
4811 mutex_lock(&wq_pool_mutex);
4812
4813 WARN_ON_ONCE(!workqueue_freezing);
4814
4815 list_for_each_entry(wq, &workqueues, list) {
4816 if (!(wq->flags & WQ_FREEZABLE))
4817 continue;
4818 /*
4819 * nr_active is monotonically decreasing. It's safe
4820 * to peek without lock.
4821 */
4822 rcu_read_lock_sched();
4823 for_each_pwq(pwq, wq) {
4824 WARN_ON_ONCE(pwq->nr_active < 0);
4825 if (pwq->nr_active) {
4826 busy = true;
4827 rcu_read_unlock_sched();
4828 goto out_unlock;
4829 }
4830 }
4831 rcu_read_unlock_sched();
4832 }
4833out_unlock:
4834 mutex_unlock(&wq_pool_mutex);
4835 return busy;
4836}
4837
4838/**
4839 * thaw_workqueues - thaw workqueues
4840 *
4841 * Thaw workqueues. Normal queueing is restored and all collected
4842 * frozen works are transferred to their respective pool worklists.
4843 *
4844 * CONTEXT:
4845 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4846 */
4847void thaw_workqueues(void)
4848{
4849 struct workqueue_struct *wq;
4850 struct pool_workqueue *pwq;
4851
4852 mutex_lock(&wq_pool_mutex);
4853
4854 if (!workqueue_freezing)
4855 goto out_unlock;
4856
4857 workqueue_freezing = false;
4858
4859 /* restore max_active and repopulate worklist */
4860 list_for_each_entry(wq, &workqueues, list) {
4861 mutex_lock(&wq->mutex);
4862 for_each_pwq(pwq, wq)
4863 pwq_adjust_max_active(pwq);
4864 mutex_unlock(&wq->mutex);
4865 }
4866
4867out_unlock:
4868 mutex_unlock(&wq_pool_mutex);
4869}
4870#endif /* CONFIG_FREEZER */
4871
4872static int workqueue_apply_unbound_cpumask(void)
4873{
4874 LIST_HEAD(ctxs);
4875 int ret = 0;
4876 struct workqueue_struct *wq;
4877 struct apply_wqattrs_ctx *ctx, *n;
4878
4879 lockdep_assert_held(&wq_pool_mutex);
4880
4881 list_for_each_entry(wq, &workqueues, list) {
4882 if (!(wq->flags & WQ_UNBOUND))
4883 continue;
4884 /* creating multiple pwqs breaks ordering guarantee */
4885 if (wq->flags & __WQ_ORDERED)
4886 continue;
4887
4888 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4889 if (!ctx) {
4890 ret = -ENOMEM;
4891 break;
4892 }
4893
4894 list_add_tail(&ctx->list, &ctxs);
4895 }
4896
4897 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4898 if (!ret)
4899 apply_wqattrs_commit(ctx);
4900 apply_wqattrs_cleanup(ctx);
4901 }
4902
4903 return ret;
4904}
4905
4906/**
4907 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4908 * @cpumask: the cpumask to set
4909 *
4910 * The low-level workqueues cpumask is a global cpumask that limits
4911 * the affinity of all unbound workqueues. This function check the @cpumask
4912 * and apply it to all unbound workqueues and updates all pwqs of them.
4913 *
4914 * Retun: 0 - Success
4915 * -EINVAL - Invalid @cpumask
4916 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4917 */
4918int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4919{
4920 int ret = -EINVAL;
4921 cpumask_var_t saved_cpumask;
4922
4923 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4924 return -ENOMEM;
4925
4926 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4927 if (!cpumask_empty(cpumask)) {
4928 apply_wqattrs_lock();
4929
4930 /* save the old wq_unbound_cpumask. */
4931 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4932
4933 /* update wq_unbound_cpumask at first and apply it to wqs. */
4934 cpumask_copy(wq_unbound_cpumask, cpumask);
4935 ret = workqueue_apply_unbound_cpumask();
4936
4937 /* restore the wq_unbound_cpumask when failed. */
4938 if (ret < 0)
4939 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4940
4941 apply_wqattrs_unlock();
4942 }
4943
4944 free_cpumask_var(saved_cpumask);
4945 return ret;
4946}
4947
4948#ifdef CONFIG_SYSFS
4949/*
4950 * Workqueues with WQ_SYSFS flag set is visible to userland via
4951 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4952 * following attributes.
4953 *
4954 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4955 * max_active RW int : maximum number of in-flight work items
4956 *
4957 * Unbound workqueues have the following extra attributes.
4958 *
4959 * id RO int : the associated pool ID
4960 * nice RW int : nice value of the workers
4961 * cpumask RW mask : bitmask of allowed CPUs for the workers
4962 */
4963struct wq_device {
4964 struct workqueue_struct *wq;
4965 struct device dev;
4966};
4967
4968static struct workqueue_struct *dev_to_wq(struct device *dev)
4969{
4970 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4971
4972 return wq_dev->wq;
4973}
4974
4975static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4976 char *buf)
4977{
4978 struct workqueue_struct *wq = dev_to_wq(dev);
4979
4980 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4981}
4982static DEVICE_ATTR_RO(per_cpu);
4983
4984static ssize_t max_active_show(struct device *dev,
4985 struct device_attribute *attr, char *buf)
4986{
4987 struct workqueue_struct *wq = dev_to_wq(dev);
4988
4989 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4990}
4991
4992static ssize_t max_active_store(struct device *dev,
4993 struct device_attribute *attr, const char *buf,
4994 size_t count)
4995{
4996 struct workqueue_struct *wq = dev_to_wq(dev);
4997 int val;
4998
4999 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5000 return -EINVAL;
5001
5002 workqueue_set_max_active(wq, val);
5003 return count;
5004}
5005static DEVICE_ATTR_RW(max_active);
5006
5007static struct attribute *wq_sysfs_attrs[] = {
5008 &dev_attr_per_cpu.attr,
5009 &dev_attr_max_active.attr,
5010 NULL,
5011};
5012ATTRIBUTE_GROUPS(wq_sysfs);
5013
5014static ssize_t wq_pool_ids_show(struct device *dev,
5015 struct device_attribute *attr, char *buf)
5016{
5017 struct workqueue_struct *wq = dev_to_wq(dev);
5018 const char *delim = "";
5019 int node, written = 0;
5020
5021 rcu_read_lock_sched();
5022 for_each_node(node) {
5023 written += scnprintf(buf + written, PAGE_SIZE - written,
5024 "%s%d:%d", delim, node,
5025 unbound_pwq_by_node(wq, node)->pool->id);
5026 delim = " ";
5027 }
5028 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5029 rcu_read_unlock_sched();
5030
5031 return written;
5032}
5033
5034static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5035 char *buf)
5036{
5037 struct workqueue_struct *wq = dev_to_wq(dev);
5038 int written;
5039
5040 mutex_lock(&wq->mutex);
5041 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5042 mutex_unlock(&wq->mutex);
5043
5044 return written;
5045}
5046
5047/* prepare workqueue_attrs for sysfs store operations */
5048static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5049{
5050 struct workqueue_attrs *attrs;
5051
5052 lockdep_assert_held(&wq_pool_mutex);
5053
5054 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5055 if (!attrs)
5056 return NULL;
5057
5058 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5059 return attrs;
5060}
5061
5062static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5063 const char *buf, size_t count)
5064{
5065 struct workqueue_struct *wq = dev_to_wq(dev);
5066 struct workqueue_attrs *attrs;
5067 int ret = -ENOMEM;
5068
5069 apply_wqattrs_lock();
5070
5071 attrs = wq_sysfs_prep_attrs(wq);
5072 if (!attrs)
5073 goto out_unlock;
5074
5075 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5076 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5077 ret = apply_workqueue_attrs_locked(wq, attrs);
5078 else
5079 ret = -EINVAL;
5080
5081out_unlock:
5082 apply_wqattrs_unlock();
5083 free_workqueue_attrs(attrs);
5084 return ret ?: count;
5085}
5086
5087static ssize_t wq_cpumask_show(struct device *dev,
5088 struct device_attribute *attr, char *buf)
5089{
5090 struct workqueue_struct *wq = dev_to_wq(dev);
5091 int written;
5092
5093 mutex_lock(&wq->mutex);
5094 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5095 cpumask_pr_args(wq->unbound_attrs->cpumask));
5096 mutex_unlock(&wq->mutex);
5097 return written;
5098}
5099
5100static ssize_t wq_cpumask_store(struct device *dev,
5101 struct device_attribute *attr,
5102 const char *buf, size_t count)
5103{
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105 struct workqueue_attrs *attrs;
5106 int ret = -ENOMEM;
5107
5108 apply_wqattrs_lock();
5109
5110 attrs = wq_sysfs_prep_attrs(wq);
5111 if (!attrs)
5112 goto out_unlock;
5113
5114 ret = cpumask_parse(buf, attrs->cpumask);
5115 if (!ret)
5116 ret = apply_workqueue_attrs_locked(wq, attrs);
5117
5118out_unlock:
5119 apply_wqattrs_unlock();
5120 free_workqueue_attrs(attrs);
5121 return ret ?: count;
5122}
5123
5124static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5125 char *buf)
5126{
5127 struct workqueue_struct *wq = dev_to_wq(dev);
5128 int written;
5129
5130 mutex_lock(&wq->mutex);
5131 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5132 !wq->unbound_attrs->no_numa);
5133 mutex_unlock(&wq->mutex);
5134
5135 return written;
5136}
5137
5138static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5139 const char *buf, size_t count)
5140{
5141 struct workqueue_struct *wq = dev_to_wq(dev);
5142 struct workqueue_attrs *attrs;
5143 int v, ret = -ENOMEM;
5144
5145 apply_wqattrs_lock();
5146
5147 attrs = wq_sysfs_prep_attrs(wq);
5148 if (!attrs)
5149 goto out_unlock;
5150
5151 ret = -EINVAL;
5152 if (sscanf(buf, "%d", &v) == 1) {
5153 attrs->no_numa = !v;
5154 ret = apply_workqueue_attrs_locked(wq, attrs);
5155 }
5156
5157out_unlock:
5158 apply_wqattrs_unlock();
5159 free_workqueue_attrs(attrs);
5160 return ret ?: count;
5161}
5162
5163static struct device_attribute wq_sysfs_unbound_attrs[] = {
5164 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5165 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5166 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5167 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5168 __ATTR_NULL,
5169};
5170
5171static struct bus_type wq_subsys = {
5172 .name = "workqueue",
5173 .dev_groups = wq_sysfs_groups,
5174};
5175
5176static ssize_t wq_unbound_cpumask_show(struct device *dev,
5177 struct device_attribute *attr, char *buf)
5178{
5179 int written;
5180
5181 mutex_lock(&wq_pool_mutex);
5182 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5183 cpumask_pr_args(wq_unbound_cpumask));
5184 mutex_unlock(&wq_pool_mutex);
5185
5186 return written;
5187}
5188
5189static ssize_t wq_unbound_cpumask_store(struct device *dev,
5190 struct device_attribute *attr, const char *buf, size_t count)
5191{
5192 cpumask_var_t cpumask;
5193 int ret;
5194
5195 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5196 return -ENOMEM;
5197
5198 ret = cpumask_parse(buf, cpumask);
5199 if (!ret)
5200 ret = workqueue_set_unbound_cpumask(cpumask);
5201
5202 free_cpumask_var(cpumask);
5203 return ret ? ret : count;
5204}
5205
5206static struct device_attribute wq_sysfs_cpumask_attr =
5207 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5208 wq_unbound_cpumask_store);
5209
5210static int __init wq_sysfs_init(void)
5211{
5212 int err;
5213
5214 err = subsys_virtual_register(&wq_subsys, NULL);
5215 if (err)
5216 return err;
5217
5218 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5219}
5220core_initcall(wq_sysfs_init);
5221
5222static void wq_device_release(struct device *dev)
5223{
5224 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5225
5226 kfree(wq_dev);
5227}
5228
5229/**
5230 * workqueue_sysfs_register - make a workqueue visible in sysfs
5231 * @wq: the workqueue to register
5232 *
5233 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5234 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5235 * which is the preferred method.
5236 *
5237 * Workqueue user should use this function directly iff it wants to apply
5238 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5239 * apply_workqueue_attrs() may race against userland updating the
5240 * attributes.
5241 *
5242 * Return: 0 on success, -errno on failure.
5243 */
5244int workqueue_sysfs_register(struct workqueue_struct *wq)
5245{
5246 struct wq_device *wq_dev;
5247 int ret;
5248
5249 /*
5250 * Adjusting max_active or creating new pwqs by applying
5251 * attributes breaks ordering guarantee. Disallow exposing ordered
5252 * workqueues.
5253 */
5254 if (WARN_ON(wq->flags & __WQ_ORDERED))
5255 return -EINVAL;
5256
5257 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5258 if (!wq_dev)
5259 return -ENOMEM;
5260
5261 wq_dev->wq = wq;
5262 wq_dev->dev.bus = &wq_subsys;
5263 wq_dev->dev.release = wq_device_release;
5264 dev_set_name(&wq_dev->dev, "%s", wq->name);
5265
5266 /*
5267 * unbound_attrs are created separately. Suppress uevent until
5268 * everything is ready.
5269 */
5270 dev_set_uevent_suppress(&wq_dev->dev, true);
5271
5272 ret = device_register(&wq_dev->dev);
5273 if (ret) {
5274 kfree(wq_dev);
5275 wq->wq_dev = NULL;
5276 return ret;
5277 }
5278
5279 if (wq->flags & WQ_UNBOUND) {
5280 struct device_attribute *attr;
5281
5282 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5283 ret = device_create_file(&wq_dev->dev, attr);
5284 if (ret) {
5285 device_unregister(&wq_dev->dev);
5286 wq->wq_dev = NULL;
5287 return ret;
5288 }
5289 }
5290 }
5291
5292 dev_set_uevent_suppress(&wq_dev->dev, false);
5293 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5294 return 0;
5295}
5296
5297/**
5298 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5299 * @wq: the workqueue to unregister
5300 *
5301 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5302 */
5303static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5304{
5305 struct wq_device *wq_dev = wq->wq_dev;
5306
5307 if (!wq->wq_dev)
5308 return;
5309
5310 wq->wq_dev = NULL;
5311 device_unregister(&wq_dev->dev);
5312}
5313#else /* CONFIG_SYSFS */
5314static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5315#endif /* CONFIG_SYSFS */
5316
5317/*
5318 * Workqueue watchdog.
5319 *
5320 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5321 * flush dependency, a concurrency managed work item which stays RUNNING
5322 * indefinitely. Workqueue stalls can be very difficult to debug as the
5323 * usual warning mechanisms don't trigger and internal workqueue state is
5324 * largely opaque.
5325 *
5326 * Workqueue watchdog monitors all worker pools periodically and dumps
5327 * state if some pools failed to make forward progress for a while where
5328 * forward progress is defined as the first item on ->worklist changing.
5329 *
5330 * This mechanism is controlled through the kernel parameter
5331 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5332 * corresponding sysfs parameter file.
5333 */
5334#ifdef CONFIG_WQ_WATCHDOG
5335
5336static void wq_watchdog_timer_fn(unsigned long data);
5337
5338static unsigned long wq_watchdog_thresh = 30;
5339static struct timer_list wq_watchdog_timer =
5340 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5341
5342static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5343static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5344
5345static void wq_watchdog_reset_touched(void)
5346{
5347 int cpu;
5348
5349 wq_watchdog_touched = jiffies;
5350 for_each_possible_cpu(cpu)
5351 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5352}
5353
5354static void wq_watchdog_timer_fn(unsigned long data)
5355{
5356 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5357 bool lockup_detected = false;
5358 struct worker_pool *pool;
5359 int pi;
5360
5361 if (!thresh)
5362 return;
5363
5364 rcu_read_lock();
5365
5366 for_each_pool(pool, pi) {
5367 unsigned long pool_ts, touched, ts;
5368
5369 if (list_empty(&pool->worklist))
5370 continue;
5371
5372 /* get the latest of pool and touched timestamps */
5373 pool_ts = READ_ONCE(pool->watchdog_ts);
5374 touched = READ_ONCE(wq_watchdog_touched);
5375
5376 if (time_after(pool_ts, touched))
5377 ts = pool_ts;
5378 else
5379 ts = touched;
5380
5381 if (pool->cpu >= 0) {
5382 unsigned long cpu_touched =
5383 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5384 pool->cpu));
5385 if (time_after(cpu_touched, ts))
5386 ts = cpu_touched;
5387 }
5388
5389 /* did we stall? */
5390 if (time_after(jiffies, ts + thresh)) {
5391 lockup_detected = true;
5392 pr_emerg("BUG: workqueue lockup - pool");
5393 pr_cont_pool_info(pool);
5394 pr_cont(" stuck for %us!\n",
5395 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5396 }
5397 }
5398
5399 rcu_read_unlock();
5400
5401 if (lockup_detected)
5402 show_workqueue_state();
5403
5404 wq_watchdog_reset_touched();
5405 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5406}
5407
5408void wq_watchdog_touch(int cpu)
5409{
5410 if (cpu >= 0)
5411 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5412 else
5413 wq_watchdog_touched = jiffies;
5414}
5415
5416static void wq_watchdog_set_thresh(unsigned long thresh)
5417{
5418 wq_watchdog_thresh = 0;
5419 del_timer_sync(&wq_watchdog_timer);
5420
5421 if (thresh) {
5422 wq_watchdog_thresh = thresh;
5423 wq_watchdog_reset_touched();
5424 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5425 }
5426}
5427
5428static int wq_watchdog_param_set_thresh(const char *val,
5429 const struct kernel_param *kp)
5430{
5431 unsigned long thresh;
5432 int ret;
5433
5434 ret = kstrtoul(val, 0, &thresh);
5435 if (ret)
5436 return ret;
5437
5438 if (system_wq)
5439 wq_watchdog_set_thresh(thresh);
5440 else
5441 wq_watchdog_thresh = thresh;
5442
5443 return 0;
5444}
5445
5446static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5447 .set = wq_watchdog_param_set_thresh,
5448 .get = param_get_ulong,
5449};
5450
5451module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5452 0644);
5453
5454static void wq_watchdog_init(void)
5455{
5456 wq_watchdog_set_thresh(wq_watchdog_thresh);
5457}
5458
5459#else /* CONFIG_WQ_WATCHDOG */
5460
5461static inline void wq_watchdog_init(void) { }
5462
5463#endif /* CONFIG_WQ_WATCHDOG */
5464
5465static void __init wq_numa_init(void)
5466{
5467 cpumask_var_t *tbl;
5468 int node, cpu;
5469
5470 if (num_possible_nodes() <= 1)
5471 return;
5472
5473 if (wq_disable_numa) {
5474 pr_info("workqueue: NUMA affinity support disabled\n");
5475 return;
5476 }
5477
5478 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5479 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5480
5481 /*
5482 * We want masks of possible CPUs of each node which isn't readily
5483 * available. Build one from cpu_to_node() which should have been
5484 * fully initialized by now.
5485 */
5486 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5487 BUG_ON(!tbl);
5488
5489 for_each_node(node)
5490 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5491 node_online(node) ? node : NUMA_NO_NODE));
5492
5493 for_each_possible_cpu(cpu) {
5494 node = cpu_to_node(cpu);
5495 if (WARN_ON(node == NUMA_NO_NODE)) {
5496 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5497 /* happens iff arch is bonkers, let's just proceed */
5498 return;
5499 }
5500 cpumask_set_cpu(cpu, tbl[node]);
5501 }
5502
5503 wq_numa_possible_cpumask = tbl;
5504 wq_numa_enabled = true;
5505}
5506
5507static int __init init_workqueues(void)
5508{
5509 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5510 int i, cpu;
5511
5512 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5513
5514 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5515 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5516
5517 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5518
5519 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5520 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5521
5522 wq_numa_init();
5523
5524 /* initialize CPU pools */
5525 for_each_possible_cpu(cpu) {
5526 struct worker_pool *pool;
5527
5528 i = 0;
5529 for_each_cpu_worker_pool(pool, cpu) {
5530 BUG_ON(init_worker_pool(pool));
5531 pool->cpu = cpu;
5532 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5533 pool->attrs->nice = std_nice[i++];
5534 pool->node = cpu_to_node(cpu);
5535
5536 /* alloc pool ID */
5537 mutex_lock(&wq_pool_mutex);
5538 BUG_ON(worker_pool_assign_id(pool));
5539 mutex_unlock(&wq_pool_mutex);
5540 }
5541 }
5542
5543 /* create the initial worker */
5544 for_each_online_cpu(cpu) {
5545 struct worker_pool *pool;
5546
5547 for_each_cpu_worker_pool(pool, cpu) {
5548 pool->flags &= ~POOL_DISASSOCIATED;
5549 BUG_ON(!create_worker(pool));
5550 }
5551 }
5552
5553 /* create default unbound and ordered wq attrs */
5554 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5555 struct workqueue_attrs *attrs;
5556
5557 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5558 attrs->nice = std_nice[i];
5559 unbound_std_wq_attrs[i] = attrs;
5560
5561 /*
5562 * An ordered wq should have only one pwq as ordering is
5563 * guaranteed by max_active which is enforced by pwqs.
5564 * Turn off NUMA so that dfl_pwq is used for all nodes.
5565 */
5566 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5567 attrs->nice = std_nice[i];
5568 attrs->no_numa = true;
5569 ordered_wq_attrs[i] = attrs;
5570 }
5571
5572 system_wq = alloc_workqueue("events", 0, 0);
5573 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5574 system_long_wq = alloc_workqueue("events_long", 0, 0);
5575 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5576 WQ_UNBOUND_MAX_ACTIVE);
5577 system_freezable_wq = alloc_workqueue("events_freezable",
5578 WQ_FREEZABLE, 0);
5579 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5580 WQ_POWER_EFFICIENT, 0);
5581 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5582 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5583 0);
5584 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5585 !system_unbound_wq || !system_freezable_wq ||
5586 !system_power_efficient_wq ||
5587 !system_freezable_power_efficient_wq);
5588
5589 wq_watchdog_init();
5590
5591 return 0;
5592}
5593early_initcall(init_workqueues);
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 /*
158 * The counter is incremented in a process context on the associated CPU
159 * w/ preemption disabled, and decremented or reset in the same context
160 * but w/ pool->lock held. The readers grab pool->lock and are
161 * guaranteed to see if the counter reached zero.
162 */
163 int nr_running;
164
165 struct list_head worklist; /* L: list of pending works */
166
167 int nr_workers; /* L: total number of workers */
168 int nr_idle; /* L: currently idle workers */
169
170 struct list_head idle_list; /* L: list of idle workers */
171 struct timer_list idle_timer; /* L: worker idle timeout */
172 struct timer_list mayday_timer; /* L: SOS timer for workers */
173
174 /* a workers is either on busy_hash or idle_list, or the manager */
175 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
176 /* L: hash of busy workers */
177
178 struct worker *manager; /* L: purely informational */
179 struct list_head workers; /* A: attached workers */
180 struct completion *detach_completion; /* all workers detached */
181
182 struct ida worker_ida; /* worker IDs for task name */
183
184 struct workqueue_attrs *attrs; /* I: worker attributes */
185 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
186 int refcnt; /* PL: refcnt for unbound pools */
187
188 /*
189 * Destruction of pool is RCU protected to allow dereferences
190 * from get_work_pool().
191 */
192 struct rcu_head rcu;
193};
194
195/*
196 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
197 * of work_struct->data are used for flags and the remaining high bits
198 * point to the pwq; thus, pwqs need to be aligned at two's power of the
199 * number of flag bits.
200 */
201struct pool_workqueue {
202 struct worker_pool *pool; /* I: the associated pool */
203 struct workqueue_struct *wq; /* I: the owning workqueue */
204 int work_color; /* L: current color */
205 int flush_color; /* L: flushing color */
206 int refcnt; /* L: reference count */
207 int nr_in_flight[WORK_NR_COLORS];
208 /* L: nr of in_flight works */
209
210 /*
211 * nr_active management and WORK_STRUCT_INACTIVE:
212 *
213 * When pwq->nr_active >= max_active, new work item is queued to
214 * pwq->inactive_works instead of pool->worklist and marked with
215 * WORK_STRUCT_INACTIVE.
216 *
217 * All work items marked with WORK_STRUCT_INACTIVE do not participate
218 * in pwq->nr_active and all work items in pwq->inactive_works are
219 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
220 * work items are in pwq->inactive_works. Some of them are ready to
221 * run in pool->worklist or worker->scheduled. Those work itmes are
222 * only struct wq_barrier which is used for flush_work() and should
223 * not participate in pwq->nr_active. For non-barrier work item, it
224 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
225 */
226 int nr_active; /* L: nr of active works */
227 int max_active; /* L: max active works */
228 struct list_head inactive_works; /* L: inactive works */
229 struct list_head pwqs_node; /* WR: node on wq->pwqs */
230 struct list_head mayday_node; /* MD: node on wq->maydays */
231
232 /*
233 * Release of unbound pwq is punted to system_wq. See put_pwq()
234 * and pwq_unbound_release_workfn() for details. pool_workqueue
235 * itself is also RCU protected so that the first pwq can be
236 * determined without grabbing wq->mutex.
237 */
238 struct work_struct unbound_release_work;
239 struct rcu_head rcu;
240} __aligned(1 << WORK_STRUCT_FLAG_BITS);
241
242/*
243 * Structure used to wait for workqueue flush.
244 */
245struct wq_flusher {
246 struct list_head list; /* WQ: list of flushers */
247 int flush_color; /* WQ: flush color waiting for */
248 struct completion done; /* flush completion */
249};
250
251struct wq_device;
252
253/*
254 * The externally visible workqueue. It relays the issued work items to
255 * the appropriate worker_pool through its pool_workqueues.
256 */
257struct workqueue_struct {
258 struct list_head pwqs; /* WR: all pwqs of this wq */
259 struct list_head list; /* PR: list of all workqueues */
260
261 struct mutex mutex; /* protects this wq */
262 int work_color; /* WQ: current work color */
263 int flush_color; /* WQ: current flush color */
264 atomic_t nr_pwqs_to_flush; /* flush in progress */
265 struct wq_flusher *first_flusher; /* WQ: first flusher */
266 struct list_head flusher_queue; /* WQ: flush waiters */
267 struct list_head flusher_overflow; /* WQ: flush overflow list */
268
269 struct list_head maydays; /* MD: pwqs requesting rescue */
270 struct worker *rescuer; /* MD: rescue worker */
271
272 int nr_drainers; /* WQ: drain in progress */
273 int saved_max_active; /* WQ: saved pwq max_active */
274
275 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
276 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
277
278#ifdef CONFIG_SYSFS
279 struct wq_device *wq_dev; /* I: for sysfs interface */
280#endif
281#ifdef CONFIG_LOCKDEP
282 char *lock_name;
283 struct lock_class_key key;
284 struct lockdep_map lockdep_map;
285#endif
286 char name[WQ_NAME_LEN]; /* I: workqueue name */
287
288 /*
289 * Destruction of workqueue_struct is RCU protected to allow walking
290 * the workqueues list without grabbing wq_pool_mutex.
291 * This is used to dump all workqueues from sysrq.
292 */
293 struct rcu_head rcu;
294
295 /* hot fields used during command issue, aligned to cacheline */
296 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
297 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
298 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
299};
300
301static struct kmem_cache *pwq_cache;
302
303static cpumask_var_t *wq_numa_possible_cpumask;
304 /* possible CPUs of each node */
305
306static bool wq_disable_numa;
307module_param_named(disable_numa, wq_disable_numa, bool, 0444);
308
309/* see the comment above the definition of WQ_POWER_EFFICIENT */
310static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
311module_param_named(power_efficient, wq_power_efficient, bool, 0444);
312
313static bool wq_online; /* can kworkers be created yet? */
314
315static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
316
317/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
318static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
319
320static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
321static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
322static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
323/* wait for manager to go away */
324static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
325
326static LIST_HEAD(workqueues); /* PR: list of all workqueues */
327static bool workqueue_freezing; /* PL: have wqs started freezing? */
328
329/* PL: allowable cpus for unbound wqs and work items */
330static cpumask_var_t wq_unbound_cpumask;
331
332/* CPU where unbound work was last round robin scheduled from this CPU */
333static DEFINE_PER_CPU(int, wq_rr_cpu_last);
334
335/*
336 * Local execution of unbound work items is no longer guaranteed. The
337 * following always forces round-robin CPU selection on unbound work items
338 * to uncover usages which depend on it.
339 */
340#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
341static bool wq_debug_force_rr_cpu = true;
342#else
343static bool wq_debug_force_rr_cpu = false;
344#endif
345module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
346
347/* the per-cpu worker pools */
348static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
349
350static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
351
352/* PL: hash of all unbound pools keyed by pool->attrs */
353static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
354
355/* I: attributes used when instantiating standard unbound pools on demand */
356static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
357
358/* I: attributes used when instantiating ordered pools on demand */
359static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
360
361struct workqueue_struct *system_wq __read_mostly;
362EXPORT_SYMBOL(system_wq);
363struct workqueue_struct *system_highpri_wq __read_mostly;
364EXPORT_SYMBOL_GPL(system_highpri_wq);
365struct workqueue_struct *system_long_wq __read_mostly;
366EXPORT_SYMBOL_GPL(system_long_wq);
367struct workqueue_struct *system_unbound_wq __read_mostly;
368EXPORT_SYMBOL_GPL(system_unbound_wq);
369struct workqueue_struct *system_freezable_wq __read_mostly;
370EXPORT_SYMBOL_GPL(system_freezable_wq);
371struct workqueue_struct *system_power_efficient_wq __read_mostly;
372EXPORT_SYMBOL_GPL(system_power_efficient_wq);
373struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
374EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
375
376static int worker_thread(void *__worker);
377static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
378static void show_pwq(struct pool_workqueue *pwq);
379static void show_one_worker_pool(struct worker_pool *pool);
380
381#define CREATE_TRACE_POINTS
382#include <trace/events/workqueue.h>
383
384#define assert_rcu_or_pool_mutex() \
385 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
386 !lockdep_is_held(&wq_pool_mutex), \
387 "RCU or wq_pool_mutex should be held")
388
389#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
390 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
391 !lockdep_is_held(&wq->mutex) && \
392 !lockdep_is_held(&wq_pool_mutex), \
393 "RCU, wq->mutex or wq_pool_mutex should be held")
394
395#define for_each_cpu_worker_pool(pool, cpu) \
396 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
397 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
398 (pool)++)
399
400/**
401 * for_each_pool - iterate through all worker_pools in the system
402 * @pool: iteration cursor
403 * @pi: integer used for iteration
404 *
405 * This must be called either with wq_pool_mutex held or RCU read
406 * locked. If the pool needs to be used beyond the locking in effect, the
407 * caller is responsible for guaranteeing that the pool stays online.
408 *
409 * The if/else clause exists only for the lockdep assertion and can be
410 * ignored.
411 */
412#define for_each_pool(pool, pi) \
413 idr_for_each_entry(&worker_pool_idr, pool, pi) \
414 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
415 else
416
417/**
418 * for_each_pool_worker - iterate through all workers of a worker_pool
419 * @worker: iteration cursor
420 * @pool: worker_pool to iterate workers of
421 *
422 * This must be called with wq_pool_attach_mutex.
423 *
424 * The if/else clause exists only for the lockdep assertion and can be
425 * ignored.
426 */
427#define for_each_pool_worker(worker, pool) \
428 list_for_each_entry((worker), &(pool)->workers, node) \
429 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
430 else
431
432/**
433 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
434 * @pwq: iteration cursor
435 * @wq: the target workqueue
436 *
437 * This must be called either with wq->mutex held or RCU read locked.
438 * If the pwq needs to be used beyond the locking in effect, the caller is
439 * responsible for guaranteeing that the pwq stays online.
440 *
441 * The if/else clause exists only for the lockdep assertion and can be
442 * ignored.
443 */
444#define for_each_pwq(pwq, wq) \
445 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
446 lockdep_is_held(&(wq->mutex)))
447
448#ifdef CONFIG_DEBUG_OBJECTS_WORK
449
450static const struct debug_obj_descr work_debug_descr;
451
452static void *work_debug_hint(void *addr)
453{
454 return ((struct work_struct *) addr)->func;
455}
456
457static bool work_is_static_object(void *addr)
458{
459 struct work_struct *work = addr;
460
461 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
462}
463
464/*
465 * fixup_init is called when:
466 * - an active object is initialized
467 */
468static bool work_fixup_init(void *addr, enum debug_obj_state state)
469{
470 struct work_struct *work = addr;
471
472 switch (state) {
473 case ODEBUG_STATE_ACTIVE:
474 cancel_work_sync(work);
475 debug_object_init(work, &work_debug_descr);
476 return true;
477 default:
478 return false;
479 }
480}
481
482/*
483 * fixup_free is called when:
484 * - an active object is freed
485 */
486static bool work_fixup_free(void *addr, enum debug_obj_state state)
487{
488 struct work_struct *work = addr;
489
490 switch (state) {
491 case ODEBUG_STATE_ACTIVE:
492 cancel_work_sync(work);
493 debug_object_free(work, &work_debug_descr);
494 return true;
495 default:
496 return false;
497 }
498}
499
500static const struct debug_obj_descr work_debug_descr = {
501 .name = "work_struct",
502 .debug_hint = work_debug_hint,
503 .is_static_object = work_is_static_object,
504 .fixup_init = work_fixup_init,
505 .fixup_free = work_fixup_free,
506};
507
508static inline void debug_work_activate(struct work_struct *work)
509{
510 debug_object_activate(work, &work_debug_descr);
511}
512
513static inline void debug_work_deactivate(struct work_struct *work)
514{
515 debug_object_deactivate(work, &work_debug_descr);
516}
517
518void __init_work(struct work_struct *work, int onstack)
519{
520 if (onstack)
521 debug_object_init_on_stack(work, &work_debug_descr);
522 else
523 debug_object_init(work, &work_debug_descr);
524}
525EXPORT_SYMBOL_GPL(__init_work);
526
527void destroy_work_on_stack(struct work_struct *work)
528{
529 debug_object_free(work, &work_debug_descr);
530}
531EXPORT_SYMBOL_GPL(destroy_work_on_stack);
532
533void destroy_delayed_work_on_stack(struct delayed_work *work)
534{
535 destroy_timer_on_stack(&work->timer);
536 debug_object_free(&work->work, &work_debug_descr);
537}
538EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
539
540#else
541static inline void debug_work_activate(struct work_struct *work) { }
542static inline void debug_work_deactivate(struct work_struct *work) { }
543#endif
544
545/**
546 * worker_pool_assign_id - allocate ID and assign it to @pool
547 * @pool: the pool pointer of interest
548 *
549 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
550 * successfully, -errno on failure.
551 */
552static int worker_pool_assign_id(struct worker_pool *pool)
553{
554 int ret;
555
556 lockdep_assert_held(&wq_pool_mutex);
557
558 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
559 GFP_KERNEL);
560 if (ret >= 0) {
561 pool->id = ret;
562 return 0;
563 }
564 return ret;
565}
566
567/**
568 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
569 * @wq: the target workqueue
570 * @node: the node ID
571 *
572 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
573 * read locked.
574 * If the pwq needs to be used beyond the locking in effect, the caller is
575 * responsible for guaranteeing that the pwq stays online.
576 *
577 * Return: The unbound pool_workqueue for @node.
578 */
579static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
580 int node)
581{
582 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
583
584 /*
585 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
586 * delayed item is pending. The plan is to keep CPU -> NODE
587 * mapping valid and stable across CPU on/offlines. Once that
588 * happens, this workaround can be removed.
589 */
590 if (unlikely(node == NUMA_NO_NODE))
591 return wq->dfl_pwq;
592
593 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
594}
595
596static unsigned int work_color_to_flags(int color)
597{
598 return color << WORK_STRUCT_COLOR_SHIFT;
599}
600
601static int get_work_color(unsigned long work_data)
602{
603 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
604 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
605}
606
607static int work_next_color(int color)
608{
609 return (color + 1) % WORK_NR_COLORS;
610}
611
612/*
613 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
614 * contain the pointer to the queued pwq. Once execution starts, the flag
615 * is cleared and the high bits contain OFFQ flags and pool ID.
616 *
617 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
618 * and clear_work_data() can be used to set the pwq, pool or clear
619 * work->data. These functions should only be called while the work is
620 * owned - ie. while the PENDING bit is set.
621 *
622 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
623 * corresponding to a work. Pool is available once the work has been
624 * queued anywhere after initialization until it is sync canceled. pwq is
625 * available only while the work item is queued.
626 *
627 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
628 * canceled. While being canceled, a work item may have its PENDING set
629 * but stay off timer and worklist for arbitrarily long and nobody should
630 * try to steal the PENDING bit.
631 */
632static inline void set_work_data(struct work_struct *work, unsigned long data,
633 unsigned long flags)
634{
635 WARN_ON_ONCE(!work_pending(work));
636 atomic_long_set(&work->data, data | flags | work_static(work));
637}
638
639static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
640 unsigned long extra_flags)
641{
642 set_work_data(work, (unsigned long)pwq,
643 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
644}
645
646static void set_work_pool_and_keep_pending(struct work_struct *work,
647 int pool_id)
648{
649 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
650 WORK_STRUCT_PENDING);
651}
652
653static void set_work_pool_and_clear_pending(struct work_struct *work,
654 int pool_id)
655{
656 /*
657 * The following wmb is paired with the implied mb in
658 * test_and_set_bit(PENDING) and ensures all updates to @work made
659 * here are visible to and precede any updates by the next PENDING
660 * owner.
661 */
662 smp_wmb();
663 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
664 /*
665 * The following mb guarantees that previous clear of a PENDING bit
666 * will not be reordered with any speculative LOADS or STORES from
667 * work->current_func, which is executed afterwards. This possible
668 * reordering can lead to a missed execution on attempt to queue
669 * the same @work. E.g. consider this case:
670 *
671 * CPU#0 CPU#1
672 * ---------------------------- --------------------------------
673 *
674 * 1 STORE event_indicated
675 * 2 queue_work_on() {
676 * 3 test_and_set_bit(PENDING)
677 * 4 } set_..._and_clear_pending() {
678 * 5 set_work_data() # clear bit
679 * 6 smp_mb()
680 * 7 work->current_func() {
681 * 8 LOAD event_indicated
682 * }
683 *
684 * Without an explicit full barrier speculative LOAD on line 8 can
685 * be executed before CPU#0 does STORE on line 1. If that happens,
686 * CPU#0 observes the PENDING bit is still set and new execution of
687 * a @work is not queued in a hope, that CPU#1 will eventually
688 * finish the queued @work. Meanwhile CPU#1 does not see
689 * event_indicated is set, because speculative LOAD was executed
690 * before actual STORE.
691 */
692 smp_mb();
693}
694
695static void clear_work_data(struct work_struct *work)
696{
697 smp_wmb(); /* see set_work_pool_and_clear_pending() */
698 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
699}
700
701static struct pool_workqueue *get_work_pwq(struct work_struct *work)
702{
703 unsigned long data = atomic_long_read(&work->data);
704
705 if (data & WORK_STRUCT_PWQ)
706 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
707 else
708 return NULL;
709}
710
711/**
712 * get_work_pool - return the worker_pool a given work was associated with
713 * @work: the work item of interest
714 *
715 * Pools are created and destroyed under wq_pool_mutex, and allows read
716 * access under RCU read lock. As such, this function should be
717 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
718 *
719 * All fields of the returned pool are accessible as long as the above
720 * mentioned locking is in effect. If the returned pool needs to be used
721 * beyond the critical section, the caller is responsible for ensuring the
722 * returned pool is and stays online.
723 *
724 * Return: The worker_pool @work was last associated with. %NULL if none.
725 */
726static struct worker_pool *get_work_pool(struct work_struct *work)
727{
728 unsigned long data = atomic_long_read(&work->data);
729 int pool_id;
730
731 assert_rcu_or_pool_mutex();
732
733 if (data & WORK_STRUCT_PWQ)
734 return ((struct pool_workqueue *)
735 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
736
737 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
738 if (pool_id == WORK_OFFQ_POOL_NONE)
739 return NULL;
740
741 return idr_find(&worker_pool_idr, pool_id);
742}
743
744/**
745 * get_work_pool_id - return the worker pool ID a given work is associated with
746 * @work: the work item of interest
747 *
748 * Return: The worker_pool ID @work was last associated with.
749 * %WORK_OFFQ_POOL_NONE if none.
750 */
751static int get_work_pool_id(struct work_struct *work)
752{
753 unsigned long data = atomic_long_read(&work->data);
754
755 if (data & WORK_STRUCT_PWQ)
756 return ((struct pool_workqueue *)
757 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
758
759 return data >> WORK_OFFQ_POOL_SHIFT;
760}
761
762static void mark_work_canceling(struct work_struct *work)
763{
764 unsigned long pool_id = get_work_pool_id(work);
765
766 pool_id <<= WORK_OFFQ_POOL_SHIFT;
767 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
768}
769
770static bool work_is_canceling(struct work_struct *work)
771{
772 unsigned long data = atomic_long_read(&work->data);
773
774 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
775}
776
777/*
778 * Policy functions. These define the policies on how the global worker
779 * pools are managed. Unless noted otherwise, these functions assume that
780 * they're being called with pool->lock held.
781 */
782
783static bool __need_more_worker(struct worker_pool *pool)
784{
785 return !pool->nr_running;
786}
787
788/*
789 * Need to wake up a worker? Called from anything but currently
790 * running workers.
791 *
792 * Note that, because unbound workers never contribute to nr_running, this
793 * function will always return %true for unbound pools as long as the
794 * worklist isn't empty.
795 */
796static bool need_more_worker(struct worker_pool *pool)
797{
798 return !list_empty(&pool->worklist) && __need_more_worker(pool);
799}
800
801/* Can I start working? Called from busy but !running workers. */
802static bool may_start_working(struct worker_pool *pool)
803{
804 return pool->nr_idle;
805}
806
807/* Do I need to keep working? Called from currently running workers. */
808static bool keep_working(struct worker_pool *pool)
809{
810 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
811}
812
813/* Do we need a new worker? Called from manager. */
814static bool need_to_create_worker(struct worker_pool *pool)
815{
816 return need_more_worker(pool) && !may_start_working(pool);
817}
818
819/* Do we have too many workers and should some go away? */
820static bool too_many_workers(struct worker_pool *pool)
821{
822 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
823 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
824 int nr_busy = pool->nr_workers - nr_idle;
825
826 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
827}
828
829/*
830 * Wake up functions.
831 */
832
833/* Return the first idle worker. Called with pool->lock held. */
834static struct worker *first_idle_worker(struct worker_pool *pool)
835{
836 if (unlikely(list_empty(&pool->idle_list)))
837 return NULL;
838
839 return list_first_entry(&pool->idle_list, struct worker, entry);
840}
841
842/**
843 * wake_up_worker - wake up an idle worker
844 * @pool: worker pool to wake worker from
845 *
846 * Wake up the first idle worker of @pool.
847 *
848 * CONTEXT:
849 * raw_spin_lock_irq(pool->lock).
850 */
851static void wake_up_worker(struct worker_pool *pool)
852{
853 struct worker *worker = first_idle_worker(pool);
854
855 if (likely(worker))
856 wake_up_process(worker->task);
857}
858
859/**
860 * wq_worker_running - a worker is running again
861 * @task: task waking up
862 *
863 * This function is called when a worker returns from schedule()
864 */
865void wq_worker_running(struct task_struct *task)
866{
867 struct worker *worker = kthread_data(task);
868
869 if (!worker->sleeping)
870 return;
871
872 /*
873 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
874 * and the nr_running increment below, we may ruin the nr_running reset
875 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
876 * pool. Protect against such race.
877 */
878 preempt_disable();
879 if (!(worker->flags & WORKER_NOT_RUNNING))
880 worker->pool->nr_running++;
881 preempt_enable();
882 worker->sleeping = 0;
883}
884
885/**
886 * wq_worker_sleeping - a worker is going to sleep
887 * @task: task going to sleep
888 *
889 * This function is called from schedule() when a busy worker is
890 * going to sleep.
891 */
892void wq_worker_sleeping(struct task_struct *task)
893{
894 struct worker *worker = kthread_data(task);
895 struct worker_pool *pool;
896
897 /*
898 * Rescuers, which may not have all the fields set up like normal
899 * workers, also reach here, let's not access anything before
900 * checking NOT_RUNNING.
901 */
902 if (worker->flags & WORKER_NOT_RUNNING)
903 return;
904
905 pool = worker->pool;
906
907 /* Return if preempted before wq_worker_running() was reached */
908 if (worker->sleeping)
909 return;
910
911 worker->sleeping = 1;
912 raw_spin_lock_irq(&pool->lock);
913
914 /*
915 * Recheck in case unbind_workers() preempted us. We don't
916 * want to decrement nr_running after the worker is unbound
917 * and nr_running has been reset.
918 */
919 if (worker->flags & WORKER_NOT_RUNNING) {
920 raw_spin_unlock_irq(&pool->lock);
921 return;
922 }
923
924 pool->nr_running--;
925 if (need_more_worker(pool))
926 wake_up_worker(pool);
927 raw_spin_unlock_irq(&pool->lock);
928}
929
930/**
931 * wq_worker_last_func - retrieve worker's last work function
932 * @task: Task to retrieve last work function of.
933 *
934 * Determine the last function a worker executed. This is called from
935 * the scheduler to get a worker's last known identity.
936 *
937 * CONTEXT:
938 * raw_spin_lock_irq(rq->lock)
939 *
940 * This function is called during schedule() when a kworker is going
941 * to sleep. It's used by psi to identify aggregation workers during
942 * dequeuing, to allow periodic aggregation to shut-off when that
943 * worker is the last task in the system or cgroup to go to sleep.
944 *
945 * As this function doesn't involve any workqueue-related locking, it
946 * only returns stable values when called from inside the scheduler's
947 * queuing and dequeuing paths, when @task, which must be a kworker,
948 * is guaranteed to not be processing any works.
949 *
950 * Return:
951 * The last work function %current executed as a worker, NULL if it
952 * hasn't executed any work yet.
953 */
954work_func_t wq_worker_last_func(struct task_struct *task)
955{
956 struct worker *worker = kthread_data(task);
957
958 return worker->last_func;
959}
960
961/**
962 * worker_set_flags - set worker flags and adjust nr_running accordingly
963 * @worker: self
964 * @flags: flags to set
965 *
966 * Set @flags in @worker->flags and adjust nr_running accordingly.
967 *
968 * CONTEXT:
969 * raw_spin_lock_irq(pool->lock)
970 */
971static inline void worker_set_flags(struct worker *worker, unsigned int flags)
972{
973 struct worker_pool *pool = worker->pool;
974
975 WARN_ON_ONCE(worker->task != current);
976
977 /* If transitioning into NOT_RUNNING, adjust nr_running. */
978 if ((flags & WORKER_NOT_RUNNING) &&
979 !(worker->flags & WORKER_NOT_RUNNING)) {
980 pool->nr_running--;
981 }
982
983 worker->flags |= flags;
984}
985
986/**
987 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
988 * @worker: self
989 * @flags: flags to clear
990 *
991 * Clear @flags in @worker->flags and adjust nr_running accordingly.
992 *
993 * CONTEXT:
994 * raw_spin_lock_irq(pool->lock)
995 */
996static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
997{
998 struct worker_pool *pool = worker->pool;
999 unsigned int oflags = worker->flags;
1000
1001 WARN_ON_ONCE(worker->task != current);
1002
1003 worker->flags &= ~flags;
1004
1005 /*
1006 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1007 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1008 * of multiple flags, not a single flag.
1009 */
1010 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1011 if (!(worker->flags & WORKER_NOT_RUNNING))
1012 pool->nr_running++;
1013}
1014
1015/**
1016 * find_worker_executing_work - find worker which is executing a work
1017 * @pool: pool of interest
1018 * @work: work to find worker for
1019 *
1020 * Find a worker which is executing @work on @pool by searching
1021 * @pool->busy_hash which is keyed by the address of @work. For a worker
1022 * to match, its current execution should match the address of @work and
1023 * its work function. This is to avoid unwanted dependency between
1024 * unrelated work executions through a work item being recycled while still
1025 * being executed.
1026 *
1027 * This is a bit tricky. A work item may be freed once its execution
1028 * starts and nothing prevents the freed area from being recycled for
1029 * another work item. If the same work item address ends up being reused
1030 * before the original execution finishes, workqueue will identify the
1031 * recycled work item as currently executing and make it wait until the
1032 * current execution finishes, introducing an unwanted dependency.
1033 *
1034 * This function checks the work item address and work function to avoid
1035 * false positives. Note that this isn't complete as one may construct a
1036 * work function which can introduce dependency onto itself through a
1037 * recycled work item. Well, if somebody wants to shoot oneself in the
1038 * foot that badly, there's only so much we can do, and if such deadlock
1039 * actually occurs, it should be easy to locate the culprit work function.
1040 *
1041 * CONTEXT:
1042 * raw_spin_lock_irq(pool->lock).
1043 *
1044 * Return:
1045 * Pointer to worker which is executing @work if found, %NULL
1046 * otherwise.
1047 */
1048static struct worker *find_worker_executing_work(struct worker_pool *pool,
1049 struct work_struct *work)
1050{
1051 struct worker *worker;
1052
1053 hash_for_each_possible(pool->busy_hash, worker, hentry,
1054 (unsigned long)work)
1055 if (worker->current_work == work &&
1056 worker->current_func == work->func)
1057 return worker;
1058
1059 return NULL;
1060}
1061
1062/**
1063 * move_linked_works - move linked works to a list
1064 * @work: start of series of works to be scheduled
1065 * @head: target list to append @work to
1066 * @nextp: out parameter for nested worklist walking
1067 *
1068 * Schedule linked works starting from @work to @head. Work series to
1069 * be scheduled starts at @work and includes any consecutive work with
1070 * WORK_STRUCT_LINKED set in its predecessor.
1071 *
1072 * If @nextp is not NULL, it's updated to point to the next work of
1073 * the last scheduled work. This allows move_linked_works() to be
1074 * nested inside outer list_for_each_entry_safe().
1075 *
1076 * CONTEXT:
1077 * raw_spin_lock_irq(pool->lock).
1078 */
1079static void move_linked_works(struct work_struct *work, struct list_head *head,
1080 struct work_struct **nextp)
1081{
1082 struct work_struct *n;
1083
1084 /*
1085 * Linked worklist will always end before the end of the list,
1086 * use NULL for list head.
1087 */
1088 list_for_each_entry_safe_from(work, n, NULL, entry) {
1089 list_move_tail(&work->entry, head);
1090 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1091 break;
1092 }
1093
1094 /*
1095 * If we're already inside safe list traversal and have moved
1096 * multiple works to the scheduled queue, the next position
1097 * needs to be updated.
1098 */
1099 if (nextp)
1100 *nextp = n;
1101}
1102
1103/**
1104 * get_pwq - get an extra reference on the specified pool_workqueue
1105 * @pwq: pool_workqueue to get
1106 *
1107 * Obtain an extra reference on @pwq. The caller should guarantee that
1108 * @pwq has positive refcnt and be holding the matching pool->lock.
1109 */
1110static void get_pwq(struct pool_workqueue *pwq)
1111{
1112 lockdep_assert_held(&pwq->pool->lock);
1113 WARN_ON_ONCE(pwq->refcnt <= 0);
1114 pwq->refcnt++;
1115}
1116
1117/**
1118 * put_pwq - put a pool_workqueue reference
1119 * @pwq: pool_workqueue to put
1120 *
1121 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1122 * destruction. The caller should be holding the matching pool->lock.
1123 */
1124static void put_pwq(struct pool_workqueue *pwq)
1125{
1126 lockdep_assert_held(&pwq->pool->lock);
1127 if (likely(--pwq->refcnt))
1128 return;
1129 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1130 return;
1131 /*
1132 * @pwq can't be released under pool->lock, bounce to
1133 * pwq_unbound_release_workfn(). This never recurses on the same
1134 * pool->lock as this path is taken only for unbound workqueues and
1135 * the release work item is scheduled on a per-cpu workqueue. To
1136 * avoid lockdep warning, unbound pool->locks are given lockdep
1137 * subclass of 1 in get_unbound_pool().
1138 */
1139 schedule_work(&pwq->unbound_release_work);
1140}
1141
1142/**
1143 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1144 * @pwq: pool_workqueue to put (can be %NULL)
1145 *
1146 * put_pwq() with locking. This function also allows %NULL @pwq.
1147 */
1148static void put_pwq_unlocked(struct pool_workqueue *pwq)
1149{
1150 if (pwq) {
1151 /*
1152 * As both pwqs and pools are RCU protected, the
1153 * following lock operations are safe.
1154 */
1155 raw_spin_lock_irq(&pwq->pool->lock);
1156 put_pwq(pwq);
1157 raw_spin_unlock_irq(&pwq->pool->lock);
1158 }
1159}
1160
1161static void pwq_activate_inactive_work(struct work_struct *work)
1162{
1163 struct pool_workqueue *pwq = get_work_pwq(work);
1164
1165 trace_workqueue_activate_work(work);
1166 if (list_empty(&pwq->pool->worklist))
1167 pwq->pool->watchdog_ts = jiffies;
1168 move_linked_works(work, &pwq->pool->worklist, NULL);
1169 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1170 pwq->nr_active++;
1171}
1172
1173static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1174{
1175 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1176 struct work_struct, entry);
1177
1178 pwq_activate_inactive_work(work);
1179}
1180
1181/**
1182 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1183 * @pwq: pwq of interest
1184 * @work_data: work_data of work which left the queue
1185 *
1186 * A work either has completed or is removed from pending queue,
1187 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1188 *
1189 * CONTEXT:
1190 * raw_spin_lock_irq(pool->lock).
1191 */
1192static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1193{
1194 int color = get_work_color(work_data);
1195
1196 if (!(work_data & WORK_STRUCT_INACTIVE)) {
1197 pwq->nr_active--;
1198 if (!list_empty(&pwq->inactive_works)) {
1199 /* one down, submit an inactive one */
1200 if (pwq->nr_active < pwq->max_active)
1201 pwq_activate_first_inactive(pwq);
1202 }
1203 }
1204
1205 pwq->nr_in_flight[color]--;
1206
1207 /* is flush in progress and are we at the flushing tip? */
1208 if (likely(pwq->flush_color != color))
1209 goto out_put;
1210
1211 /* are there still in-flight works? */
1212 if (pwq->nr_in_flight[color])
1213 goto out_put;
1214
1215 /* this pwq is done, clear flush_color */
1216 pwq->flush_color = -1;
1217
1218 /*
1219 * If this was the last pwq, wake up the first flusher. It
1220 * will handle the rest.
1221 */
1222 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1223 complete(&pwq->wq->first_flusher->done);
1224out_put:
1225 put_pwq(pwq);
1226}
1227
1228/**
1229 * try_to_grab_pending - steal work item from worklist and disable irq
1230 * @work: work item to steal
1231 * @is_dwork: @work is a delayed_work
1232 * @flags: place to store irq state
1233 *
1234 * Try to grab PENDING bit of @work. This function can handle @work in any
1235 * stable state - idle, on timer or on worklist.
1236 *
1237 * Return:
1238 *
1239 * ======== ================================================================
1240 * 1 if @work was pending and we successfully stole PENDING
1241 * 0 if @work was idle and we claimed PENDING
1242 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1243 * -ENOENT if someone else is canceling @work, this state may persist
1244 * for arbitrarily long
1245 * ======== ================================================================
1246 *
1247 * Note:
1248 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1249 * interrupted while holding PENDING and @work off queue, irq must be
1250 * disabled on entry. This, combined with delayed_work->timer being
1251 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1252 *
1253 * On successful return, >= 0, irq is disabled and the caller is
1254 * responsible for releasing it using local_irq_restore(*@flags).
1255 *
1256 * This function is safe to call from any context including IRQ handler.
1257 */
1258static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1259 unsigned long *flags)
1260{
1261 struct worker_pool *pool;
1262 struct pool_workqueue *pwq;
1263
1264 local_irq_save(*flags);
1265
1266 /* try to steal the timer if it exists */
1267 if (is_dwork) {
1268 struct delayed_work *dwork = to_delayed_work(work);
1269
1270 /*
1271 * dwork->timer is irqsafe. If del_timer() fails, it's
1272 * guaranteed that the timer is not queued anywhere and not
1273 * running on the local CPU.
1274 */
1275 if (likely(del_timer(&dwork->timer)))
1276 return 1;
1277 }
1278
1279 /* try to claim PENDING the normal way */
1280 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1281 return 0;
1282
1283 rcu_read_lock();
1284 /*
1285 * The queueing is in progress, or it is already queued. Try to
1286 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1287 */
1288 pool = get_work_pool(work);
1289 if (!pool)
1290 goto fail;
1291
1292 raw_spin_lock(&pool->lock);
1293 /*
1294 * work->data is guaranteed to point to pwq only while the work
1295 * item is queued on pwq->wq, and both updating work->data to point
1296 * to pwq on queueing and to pool on dequeueing are done under
1297 * pwq->pool->lock. This in turn guarantees that, if work->data
1298 * points to pwq which is associated with a locked pool, the work
1299 * item is currently queued on that pool.
1300 */
1301 pwq = get_work_pwq(work);
1302 if (pwq && pwq->pool == pool) {
1303 debug_work_deactivate(work);
1304
1305 /*
1306 * A cancelable inactive work item must be in the
1307 * pwq->inactive_works since a queued barrier can't be
1308 * canceled (see the comments in insert_wq_barrier()).
1309 *
1310 * An inactive work item cannot be grabbed directly because
1311 * it might have linked barrier work items which, if left
1312 * on the inactive_works list, will confuse pwq->nr_active
1313 * management later on and cause stall. Make sure the work
1314 * item is activated before grabbing.
1315 */
1316 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1317 pwq_activate_inactive_work(work);
1318
1319 list_del_init(&work->entry);
1320 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1321
1322 /* work->data points to pwq iff queued, point to pool */
1323 set_work_pool_and_keep_pending(work, pool->id);
1324
1325 raw_spin_unlock(&pool->lock);
1326 rcu_read_unlock();
1327 return 1;
1328 }
1329 raw_spin_unlock(&pool->lock);
1330fail:
1331 rcu_read_unlock();
1332 local_irq_restore(*flags);
1333 if (work_is_canceling(work))
1334 return -ENOENT;
1335 cpu_relax();
1336 return -EAGAIN;
1337}
1338
1339/**
1340 * insert_work - insert a work into a pool
1341 * @pwq: pwq @work belongs to
1342 * @work: work to insert
1343 * @head: insertion point
1344 * @extra_flags: extra WORK_STRUCT_* flags to set
1345 *
1346 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1347 * work_struct flags.
1348 *
1349 * CONTEXT:
1350 * raw_spin_lock_irq(pool->lock).
1351 */
1352static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1353 struct list_head *head, unsigned int extra_flags)
1354{
1355 struct worker_pool *pool = pwq->pool;
1356
1357 /* record the work call stack in order to print it in KASAN reports */
1358 kasan_record_aux_stack_noalloc(work);
1359
1360 /* we own @work, set data and link */
1361 set_work_pwq(work, pwq, extra_flags);
1362 list_add_tail(&work->entry, head);
1363 get_pwq(pwq);
1364
1365 if (__need_more_worker(pool))
1366 wake_up_worker(pool);
1367}
1368
1369/*
1370 * Test whether @work is being queued from another work executing on the
1371 * same workqueue.
1372 */
1373static bool is_chained_work(struct workqueue_struct *wq)
1374{
1375 struct worker *worker;
1376
1377 worker = current_wq_worker();
1378 /*
1379 * Return %true iff I'm a worker executing a work item on @wq. If
1380 * I'm @worker, it's safe to dereference it without locking.
1381 */
1382 return worker && worker->current_pwq->wq == wq;
1383}
1384
1385/*
1386 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1387 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1388 * avoid perturbing sensitive tasks.
1389 */
1390static int wq_select_unbound_cpu(int cpu)
1391{
1392 static bool printed_dbg_warning;
1393 int new_cpu;
1394
1395 if (likely(!wq_debug_force_rr_cpu)) {
1396 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1397 return cpu;
1398 } else if (!printed_dbg_warning) {
1399 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1400 printed_dbg_warning = true;
1401 }
1402
1403 if (cpumask_empty(wq_unbound_cpumask))
1404 return cpu;
1405
1406 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1407 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1408 if (unlikely(new_cpu >= nr_cpu_ids)) {
1409 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1410 if (unlikely(new_cpu >= nr_cpu_ids))
1411 return cpu;
1412 }
1413 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1414
1415 return new_cpu;
1416}
1417
1418static void __queue_work(int cpu, struct workqueue_struct *wq,
1419 struct work_struct *work)
1420{
1421 struct pool_workqueue *pwq;
1422 struct worker_pool *last_pool;
1423 struct list_head *worklist;
1424 unsigned int work_flags;
1425 unsigned int req_cpu = cpu;
1426
1427 /*
1428 * While a work item is PENDING && off queue, a task trying to
1429 * steal the PENDING will busy-loop waiting for it to either get
1430 * queued or lose PENDING. Grabbing PENDING and queueing should
1431 * happen with IRQ disabled.
1432 */
1433 lockdep_assert_irqs_disabled();
1434
1435
1436 /* if draining, only works from the same workqueue are allowed */
1437 if (unlikely(wq->flags & __WQ_DRAINING) &&
1438 WARN_ON_ONCE(!is_chained_work(wq)))
1439 return;
1440 rcu_read_lock();
1441retry:
1442 /* pwq which will be used unless @work is executing elsewhere */
1443 if (wq->flags & WQ_UNBOUND) {
1444 if (req_cpu == WORK_CPU_UNBOUND)
1445 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1446 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1447 } else {
1448 if (req_cpu == WORK_CPU_UNBOUND)
1449 cpu = raw_smp_processor_id();
1450 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1451 }
1452
1453 /*
1454 * If @work was previously on a different pool, it might still be
1455 * running there, in which case the work needs to be queued on that
1456 * pool to guarantee non-reentrancy.
1457 */
1458 last_pool = get_work_pool(work);
1459 if (last_pool && last_pool != pwq->pool) {
1460 struct worker *worker;
1461
1462 raw_spin_lock(&last_pool->lock);
1463
1464 worker = find_worker_executing_work(last_pool, work);
1465
1466 if (worker && worker->current_pwq->wq == wq) {
1467 pwq = worker->current_pwq;
1468 } else {
1469 /* meh... not running there, queue here */
1470 raw_spin_unlock(&last_pool->lock);
1471 raw_spin_lock(&pwq->pool->lock);
1472 }
1473 } else {
1474 raw_spin_lock(&pwq->pool->lock);
1475 }
1476
1477 /*
1478 * pwq is determined and locked. For unbound pools, we could have
1479 * raced with pwq release and it could already be dead. If its
1480 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1481 * without another pwq replacing it in the numa_pwq_tbl or while
1482 * work items are executing on it, so the retrying is guaranteed to
1483 * make forward-progress.
1484 */
1485 if (unlikely(!pwq->refcnt)) {
1486 if (wq->flags & WQ_UNBOUND) {
1487 raw_spin_unlock(&pwq->pool->lock);
1488 cpu_relax();
1489 goto retry;
1490 }
1491 /* oops */
1492 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1493 wq->name, cpu);
1494 }
1495
1496 /* pwq determined, queue */
1497 trace_workqueue_queue_work(req_cpu, pwq, work);
1498
1499 if (WARN_ON(!list_empty(&work->entry)))
1500 goto out;
1501
1502 pwq->nr_in_flight[pwq->work_color]++;
1503 work_flags = work_color_to_flags(pwq->work_color);
1504
1505 if (likely(pwq->nr_active < pwq->max_active)) {
1506 trace_workqueue_activate_work(work);
1507 pwq->nr_active++;
1508 worklist = &pwq->pool->worklist;
1509 if (list_empty(worklist))
1510 pwq->pool->watchdog_ts = jiffies;
1511 } else {
1512 work_flags |= WORK_STRUCT_INACTIVE;
1513 worklist = &pwq->inactive_works;
1514 }
1515
1516 debug_work_activate(work);
1517 insert_work(pwq, work, worklist, work_flags);
1518
1519out:
1520 raw_spin_unlock(&pwq->pool->lock);
1521 rcu_read_unlock();
1522}
1523
1524/**
1525 * queue_work_on - queue work on specific cpu
1526 * @cpu: CPU number to execute work on
1527 * @wq: workqueue to use
1528 * @work: work to queue
1529 *
1530 * We queue the work to a specific CPU, the caller must ensure it
1531 * can't go away. Callers that fail to ensure that the specified
1532 * CPU cannot go away will execute on a randomly chosen CPU.
1533 *
1534 * Return: %false if @work was already on a queue, %true otherwise.
1535 */
1536bool queue_work_on(int cpu, struct workqueue_struct *wq,
1537 struct work_struct *work)
1538{
1539 bool ret = false;
1540 unsigned long flags;
1541
1542 local_irq_save(flags);
1543
1544 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1545 __queue_work(cpu, wq, work);
1546 ret = true;
1547 }
1548
1549 local_irq_restore(flags);
1550 return ret;
1551}
1552EXPORT_SYMBOL(queue_work_on);
1553
1554/**
1555 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1556 * @node: NUMA node ID that we want to select a CPU from
1557 *
1558 * This function will attempt to find a "random" cpu available on a given
1559 * node. If there are no CPUs available on the given node it will return
1560 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1561 * available CPU if we need to schedule this work.
1562 */
1563static int workqueue_select_cpu_near(int node)
1564{
1565 int cpu;
1566
1567 /* No point in doing this if NUMA isn't enabled for workqueues */
1568 if (!wq_numa_enabled)
1569 return WORK_CPU_UNBOUND;
1570
1571 /* Delay binding to CPU if node is not valid or online */
1572 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1573 return WORK_CPU_UNBOUND;
1574
1575 /* Use local node/cpu if we are already there */
1576 cpu = raw_smp_processor_id();
1577 if (node == cpu_to_node(cpu))
1578 return cpu;
1579
1580 /* Use "random" otherwise know as "first" online CPU of node */
1581 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1582
1583 /* If CPU is valid return that, otherwise just defer */
1584 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1585}
1586
1587/**
1588 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1589 * @node: NUMA node that we are targeting the work for
1590 * @wq: workqueue to use
1591 * @work: work to queue
1592 *
1593 * We queue the work to a "random" CPU within a given NUMA node. The basic
1594 * idea here is to provide a way to somehow associate work with a given
1595 * NUMA node.
1596 *
1597 * This function will only make a best effort attempt at getting this onto
1598 * the right NUMA node. If no node is requested or the requested node is
1599 * offline then we just fall back to standard queue_work behavior.
1600 *
1601 * Currently the "random" CPU ends up being the first available CPU in the
1602 * intersection of cpu_online_mask and the cpumask of the node, unless we
1603 * are running on the node. In that case we just use the current CPU.
1604 *
1605 * Return: %false if @work was already on a queue, %true otherwise.
1606 */
1607bool queue_work_node(int node, struct workqueue_struct *wq,
1608 struct work_struct *work)
1609{
1610 unsigned long flags;
1611 bool ret = false;
1612
1613 /*
1614 * This current implementation is specific to unbound workqueues.
1615 * Specifically we only return the first available CPU for a given
1616 * node instead of cycling through individual CPUs within the node.
1617 *
1618 * If this is used with a per-cpu workqueue then the logic in
1619 * workqueue_select_cpu_near would need to be updated to allow for
1620 * some round robin type logic.
1621 */
1622 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1623
1624 local_irq_save(flags);
1625
1626 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1627 int cpu = workqueue_select_cpu_near(node);
1628
1629 __queue_work(cpu, wq, work);
1630 ret = true;
1631 }
1632
1633 local_irq_restore(flags);
1634 return ret;
1635}
1636EXPORT_SYMBOL_GPL(queue_work_node);
1637
1638void delayed_work_timer_fn(struct timer_list *t)
1639{
1640 struct delayed_work *dwork = from_timer(dwork, t, timer);
1641
1642 /* should have been called from irqsafe timer with irq already off */
1643 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1644}
1645EXPORT_SYMBOL(delayed_work_timer_fn);
1646
1647static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1648 struct delayed_work *dwork, unsigned long delay)
1649{
1650 struct timer_list *timer = &dwork->timer;
1651 struct work_struct *work = &dwork->work;
1652
1653 WARN_ON_ONCE(!wq);
1654 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1655 WARN_ON_ONCE(timer_pending(timer));
1656 WARN_ON_ONCE(!list_empty(&work->entry));
1657
1658 /*
1659 * If @delay is 0, queue @dwork->work immediately. This is for
1660 * both optimization and correctness. The earliest @timer can
1661 * expire is on the closest next tick and delayed_work users depend
1662 * on that there's no such delay when @delay is 0.
1663 */
1664 if (!delay) {
1665 __queue_work(cpu, wq, &dwork->work);
1666 return;
1667 }
1668
1669 dwork->wq = wq;
1670 dwork->cpu = cpu;
1671 timer->expires = jiffies + delay;
1672
1673 if (unlikely(cpu != WORK_CPU_UNBOUND))
1674 add_timer_on(timer, cpu);
1675 else
1676 add_timer(timer);
1677}
1678
1679/**
1680 * queue_delayed_work_on - queue work on specific CPU after delay
1681 * @cpu: CPU number to execute work on
1682 * @wq: workqueue to use
1683 * @dwork: work to queue
1684 * @delay: number of jiffies to wait before queueing
1685 *
1686 * Return: %false if @work was already on a queue, %true otherwise. If
1687 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1688 * execution.
1689 */
1690bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1691 struct delayed_work *dwork, unsigned long delay)
1692{
1693 struct work_struct *work = &dwork->work;
1694 bool ret = false;
1695 unsigned long flags;
1696
1697 /* read the comment in __queue_work() */
1698 local_irq_save(flags);
1699
1700 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1701 __queue_delayed_work(cpu, wq, dwork, delay);
1702 ret = true;
1703 }
1704
1705 local_irq_restore(flags);
1706 return ret;
1707}
1708EXPORT_SYMBOL(queue_delayed_work_on);
1709
1710/**
1711 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1712 * @cpu: CPU number to execute work on
1713 * @wq: workqueue to use
1714 * @dwork: work to queue
1715 * @delay: number of jiffies to wait before queueing
1716 *
1717 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1718 * modify @dwork's timer so that it expires after @delay. If @delay is
1719 * zero, @work is guaranteed to be scheduled immediately regardless of its
1720 * current state.
1721 *
1722 * Return: %false if @dwork was idle and queued, %true if @dwork was
1723 * pending and its timer was modified.
1724 *
1725 * This function is safe to call from any context including IRQ handler.
1726 * See try_to_grab_pending() for details.
1727 */
1728bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1729 struct delayed_work *dwork, unsigned long delay)
1730{
1731 unsigned long flags;
1732 int ret;
1733
1734 do {
1735 ret = try_to_grab_pending(&dwork->work, true, &flags);
1736 } while (unlikely(ret == -EAGAIN));
1737
1738 if (likely(ret >= 0)) {
1739 __queue_delayed_work(cpu, wq, dwork, delay);
1740 local_irq_restore(flags);
1741 }
1742
1743 /* -ENOENT from try_to_grab_pending() becomes %true */
1744 return ret;
1745}
1746EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1747
1748static void rcu_work_rcufn(struct rcu_head *rcu)
1749{
1750 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1751
1752 /* read the comment in __queue_work() */
1753 local_irq_disable();
1754 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1755 local_irq_enable();
1756}
1757
1758/**
1759 * queue_rcu_work - queue work after a RCU grace period
1760 * @wq: workqueue to use
1761 * @rwork: work to queue
1762 *
1763 * Return: %false if @rwork was already pending, %true otherwise. Note
1764 * that a full RCU grace period is guaranteed only after a %true return.
1765 * While @rwork is guaranteed to be executed after a %false return, the
1766 * execution may happen before a full RCU grace period has passed.
1767 */
1768bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1769{
1770 struct work_struct *work = &rwork->work;
1771
1772 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1773 rwork->wq = wq;
1774 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
1775 return true;
1776 }
1777
1778 return false;
1779}
1780EXPORT_SYMBOL(queue_rcu_work);
1781
1782/**
1783 * worker_enter_idle - enter idle state
1784 * @worker: worker which is entering idle state
1785 *
1786 * @worker is entering idle state. Update stats and idle timer if
1787 * necessary.
1788 *
1789 * LOCKING:
1790 * raw_spin_lock_irq(pool->lock).
1791 */
1792static void worker_enter_idle(struct worker *worker)
1793{
1794 struct worker_pool *pool = worker->pool;
1795
1796 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1797 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1798 (worker->hentry.next || worker->hentry.pprev)))
1799 return;
1800
1801 /* can't use worker_set_flags(), also called from create_worker() */
1802 worker->flags |= WORKER_IDLE;
1803 pool->nr_idle++;
1804 worker->last_active = jiffies;
1805
1806 /* idle_list is LIFO */
1807 list_add(&worker->entry, &pool->idle_list);
1808
1809 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1810 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1811
1812 /* Sanity check nr_running. */
1813 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1814}
1815
1816/**
1817 * worker_leave_idle - leave idle state
1818 * @worker: worker which is leaving idle state
1819 *
1820 * @worker is leaving idle state. Update stats.
1821 *
1822 * LOCKING:
1823 * raw_spin_lock_irq(pool->lock).
1824 */
1825static void worker_leave_idle(struct worker *worker)
1826{
1827 struct worker_pool *pool = worker->pool;
1828
1829 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1830 return;
1831 worker_clr_flags(worker, WORKER_IDLE);
1832 pool->nr_idle--;
1833 list_del_init(&worker->entry);
1834}
1835
1836static struct worker *alloc_worker(int node)
1837{
1838 struct worker *worker;
1839
1840 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1841 if (worker) {
1842 INIT_LIST_HEAD(&worker->entry);
1843 INIT_LIST_HEAD(&worker->scheduled);
1844 INIT_LIST_HEAD(&worker->node);
1845 /* on creation a worker is in !idle && prep state */
1846 worker->flags = WORKER_PREP;
1847 }
1848 return worker;
1849}
1850
1851/**
1852 * worker_attach_to_pool() - attach a worker to a pool
1853 * @worker: worker to be attached
1854 * @pool: the target pool
1855 *
1856 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1857 * cpu-binding of @worker are kept coordinated with the pool across
1858 * cpu-[un]hotplugs.
1859 */
1860static void worker_attach_to_pool(struct worker *worker,
1861 struct worker_pool *pool)
1862{
1863 mutex_lock(&wq_pool_attach_mutex);
1864
1865 /*
1866 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1867 * stable across this function. See the comments above the flag
1868 * definition for details.
1869 */
1870 if (pool->flags & POOL_DISASSOCIATED)
1871 worker->flags |= WORKER_UNBOUND;
1872 else
1873 kthread_set_per_cpu(worker->task, pool->cpu);
1874
1875 if (worker->rescue_wq)
1876 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1877
1878 list_add_tail(&worker->node, &pool->workers);
1879 worker->pool = pool;
1880
1881 mutex_unlock(&wq_pool_attach_mutex);
1882}
1883
1884/**
1885 * worker_detach_from_pool() - detach a worker from its pool
1886 * @worker: worker which is attached to its pool
1887 *
1888 * Undo the attaching which had been done in worker_attach_to_pool(). The
1889 * caller worker shouldn't access to the pool after detached except it has
1890 * other reference to the pool.
1891 */
1892static void worker_detach_from_pool(struct worker *worker)
1893{
1894 struct worker_pool *pool = worker->pool;
1895 struct completion *detach_completion = NULL;
1896
1897 mutex_lock(&wq_pool_attach_mutex);
1898
1899 kthread_set_per_cpu(worker->task, -1);
1900 list_del(&worker->node);
1901 worker->pool = NULL;
1902
1903 if (list_empty(&pool->workers))
1904 detach_completion = pool->detach_completion;
1905 mutex_unlock(&wq_pool_attach_mutex);
1906
1907 /* clear leftover flags without pool->lock after it is detached */
1908 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1909
1910 if (detach_completion)
1911 complete(detach_completion);
1912}
1913
1914/**
1915 * create_worker - create a new workqueue worker
1916 * @pool: pool the new worker will belong to
1917 *
1918 * Create and start a new worker which is attached to @pool.
1919 *
1920 * CONTEXT:
1921 * Might sleep. Does GFP_KERNEL allocations.
1922 *
1923 * Return:
1924 * Pointer to the newly created worker.
1925 */
1926static struct worker *create_worker(struct worker_pool *pool)
1927{
1928 struct worker *worker;
1929 int id;
1930 char id_buf[16];
1931
1932 /* ID is needed to determine kthread name */
1933 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
1934 if (id < 0)
1935 return NULL;
1936
1937 worker = alloc_worker(pool->node);
1938 if (!worker)
1939 goto fail;
1940
1941 worker->id = id;
1942
1943 if (pool->cpu >= 0)
1944 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1945 pool->attrs->nice < 0 ? "H" : "");
1946 else
1947 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1948
1949 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1950 "kworker/%s", id_buf);
1951 if (IS_ERR(worker->task))
1952 goto fail;
1953
1954 set_user_nice(worker->task, pool->attrs->nice);
1955 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1956
1957 /* successful, attach the worker to the pool */
1958 worker_attach_to_pool(worker, pool);
1959
1960 /* start the newly created worker */
1961 raw_spin_lock_irq(&pool->lock);
1962 worker->pool->nr_workers++;
1963 worker_enter_idle(worker);
1964 wake_up_process(worker->task);
1965 raw_spin_unlock_irq(&pool->lock);
1966
1967 return worker;
1968
1969fail:
1970 ida_free(&pool->worker_ida, id);
1971 kfree(worker);
1972 return NULL;
1973}
1974
1975/**
1976 * destroy_worker - destroy a workqueue worker
1977 * @worker: worker to be destroyed
1978 *
1979 * Destroy @worker and adjust @pool stats accordingly. The worker should
1980 * be idle.
1981 *
1982 * CONTEXT:
1983 * raw_spin_lock_irq(pool->lock).
1984 */
1985static void destroy_worker(struct worker *worker)
1986{
1987 struct worker_pool *pool = worker->pool;
1988
1989 lockdep_assert_held(&pool->lock);
1990
1991 /* sanity check frenzy */
1992 if (WARN_ON(worker->current_work) ||
1993 WARN_ON(!list_empty(&worker->scheduled)) ||
1994 WARN_ON(!(worker->flags & WORKER_IDLE)))
1995 return;
1996
1997 pool->nr_workers--;
1998 pool->nr_idle--;
1999
2000 list_del_init(&worker->entry);
2001 worker->flags |= WORKER_DIE;
2002 wake_up_process(worker->task);
2003}
2004
2005static void idle_worker_timeout(struct timer_list *t)
2006{
2007 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2008
2009 raw_spin_lock_irq(&pool->lock);
2010
2011 while (too_many_workers(pool)) {
2012 struct worker *worker;
2013 unsigned long expires;
2014
2015 /* idle_list is kept in LIFO order, check the last one */
2016 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2017 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2018
2019 if (time_before(jiffies, expires)) {
2020 mod_timer(&pool->idle_timer, expires);
2021 break;
2022 }
2023
2024 destroy_worker(worker);
2025 }
2026
2027 raw_spin_unlock_irq(&pool->lock);
2028}
2029
2030static void send_mayday(struct work_struct *work)
2031{
2032 struct pool_workqueue *pwq = get_work_pwq(work);
2033 struct workqueue_struct *wq = pwq->wq;
2034
2035 lockdep_assert_held(&wq_mayday_lock);
2036
2037 if (!wq->rescuer)
2038 return;
2039
2040 /* mayday mayday mayday */
2041 if (list_empty(&pwq->mayday_node)) {
2042 /*
2043 * If @pwq is for an unbound wq, its base ref may be put at
2044 * any time due to an attribute change. Pin @pwq until the
2045 * rescuer is done with it.
2046 */
2047 get_pwq(pwq);
2048 list_add_tail(&pwq->mayday_node, &wq->maydays);
2049 wake_up_process(wq->rescuer->task);
2050 }
2051}
2052
2053static void pool_mayday_timeout(struct timer_list *t)
2054{
2055 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2056 struct work_struct *work;
2057
2058 raw_spin_lock_irq(&pool->lock);
2059 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2060
2061 if (need_to_create_worker(pool)) {
2062 /*
2063 * We've been trying to create a new worker but
2064 * haven't been successful. We might be hitting an
2065 * allocation deadlock. Send distress signals to
2066 * rescuers.
2067 */
2068 list_for_each_entry(work, &pool->worklist, entry)
2069 send_mayday(work);
2070 }
2071
2072 raw_spin_unlock(&wq_mayday_lock);
2073 raw_spin_unlock_irq(&pool->lock);
2074
2075 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2076}
2077
2078/**
2079 * maybe_create_worker - create a new worker if necessary
2080 * @pool: pool to create a new worker for
2081 *
2082 * Create a new worker for @pool if necessary. @pool is guaranteed to
2083 * have at least one idle worker on return from this function. If
2084 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2085 * sent to all rescuers with works scheduled on @pool to resolve
2086 * possible allocation deadlock.
2087 *
2088 * On return, need_to_create_worker() is guaranteed to be %false and
2089 * may_start_working() %true.
2090 *
2091 * LOCKING:
2092 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2093 * multiple times. Does GFP_KERNEL allocations. Called only from
2094 * manager.
2095 */
2096static void maybe_create_worker(struct worker_pool *pool)
2097__releases(&pool->lock)
2098__acquires(&pool->lock)
2099{
2100restart:
2101 raw_spin_unlock_irq(&pool->lock);
2102
2103 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2104 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2105
2106 while (true) {
2107 if (create_worker(pool) || !need_to_create_worker(pool))
2108 break;
2109
2110 schedule_timeout_interruptible(CREATE_COOLDOWN);
2111
2112 if (!need_to_create_worker(pool))
2113 break;
2114 }
2115
2116 del_timer_sync(&pool->mayday_timer);
2117 raw_spin_lock_irq(&pool->lock);
2118 /*
2119 * This is necessary even after a new worker was just successfully
2120 * created as @pool->lock was dropped and the new worker might have
2121 * already become busy.
2122 */
2123 if (need_to_create_worker(pool))
2124 goto restart;
2125}
2126
2127/**
2128 * manage_workers - manage worker pool
2129 * @worker: self
2130 *
2131 * Assume the manager role and manage the worker pool @worker belongs
2132 * to. At any given time, there can be only zero or one manager per
2133 * pool. The exclusion is handled automatically by this function.
2134 *
2135 * The caller can safely start processing works on false return. On
2136 * true return, it's guaranteed that need_to_create_worker() is false
2137 * and may_start_working() is true.
2138 *
2139 * CONTEXT:
2140 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2141 * multiple times. Does GFP_KERNEL allocations.
2142 *
2143 * Return:
2144 * %false if the pool doesn't need management and the caller can safely
2145 * start processing works, %true if management function was performed and
2146 * the conditions that the caller verified before calling the function may
2147 * no longer be true.
2148 */
2149static bool manage_workers(struct worker *worker)
2150{
2151 struct worker_pool *pool = worker->pool;
2152
2153 if (pool->flags & POOL_MANAGER_ACTIVE)
2154 return false;
2155
2156 pool->flags |= POOL_MANAGER_ACTIVE;
2157 pool->manager = worker;
2158
2159 maybe_create_worker(pool);
2160
2161 pool->manager = NULL;
2162 pool->flags &= ~POOL_MANAGER_ACTIVE;
2163 rcuwait_wake_up(&manager_wait);
2164 return true;
2165}
2166
2167/**
2168 * process_one_work - process single work
2169 * @worker: self
2170 * @work: work to process
2171 *
2172 * Process @work. This function contains all the logics necessary to
2173 * process a single work including synchronization against and
2174 * interaction with other workers on the same cpu, queueing and
2175 * flushing. As long as context requirement is met, any worker can
2176 * call this function to process a work.
2177 *
2178 * CONTEXT:
2179 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2180 */
2181static void process_one_work(struct worker *worker, struct work_struct *work)
2182__releases(&pool->lock)
2183__acquires(&pool->lock)
2184{
2185 struct pool_workqueue *pwq = get_work_pwq(work);
2186 struct worker_pool *pool = worker->pool;
2187 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2188 unsigned long work_data;
2189 struct worker *collision;
2190#ifdef CONFIG_LOCKDEP
2191 /*
2192 * It is permissible to free the struct work_struct from
2193 * inside the function that is called from it, this we need to
2194 * take into account for lockdep too. To avoid bogus "held
2195 * lock freed" warnings as well as problems when looking into
2196 * work->lockdep_map, make a copy and use that here.
2197 */
2198 struct lockdep_map lockdep_map;
2199
2200 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2201#endif
2202 /* ensure we're on the correct CPU */
2203 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2204 raw_smp_processor_id() != pool->cpu);
2205
2206 /*
2207 * A single work shouldn't be executed concurrently by
2208 * multiple workers on a single cpu. Check whether anyone is
2209 * already processing the work. If so, defer the work to the
2210 * currently executing one.
2211 */
2212 collision = find_worker_executing_work(pool, work);
2213 if (unlikely(collision)) {
2214 move_linked_works(work, &collision->scheduled, NULL);
2215 return;
2216 }
2217
2218 /* claim and dequeue */
2219 debug_work_deactivate(work);
2220 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2221 worker->current_work = work;
2222 worker->current_func = work->func;
2223 worker->current_pwq = pwq;
2224 work_data = *work_data_bits(work);
2225 worker->current_color = get_work_color(work_data);
2226
2227 /*
2228 * Record wq name for cmdline and debug reporting, may get
2229 * overridden through set_worker_desc().
2230 */
2231 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2232
2233 list_del_init(&work->entry);
2234
2235 /*
2236 * CPU intensive works don't participate in concurrency management.
2237 * They're the scheduler's responsibility. This takes @worker out
2238 * of concurrency management and the next code block will chain
2239 * execution of the pending work items.
2240 */
2241 if (unlikely(cpu_intensive))
2242 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2243
2244 /*
2245 * Wake up another worker if necessary. The condition is always
2246 * false for normal per-cpu workers since nr_running would always
2247 * be >= 1 at this point. This is used to chain execution of the
2248 * pending work items for WORKER_NOT_RUNNING workers such as the
2249 * UNBOUND and CPU_INTENSIVE ones.
2250 */
2251 if (need_more_worker(pool))
2252 wake_up_worker(pool);
2253
2254 /*
2255 * Record the last pool and clear PENDING which should be the last
2256 * update to @work. Also, do this inside @pool->lock so that
2257 * PENDING and queued state changes happen together while IRQ is
2258 * disabled.
2259 */
2260 set_work_pool_and_clear_pending(work, pool->id);
2261
2262 raw_spin_unlock_irq(&pool->lock);
2263
2264 lock_map_acquire(&pwq->wq->lockdep_map);
2265 lock_map_acquire(&lockdep_map);
2266 /*
2267 * Strictly speaking we should mark the invariant state without holding
2268 * any locks, that is, before these two lock_map_acquire()'s.
2269 *
2270 * However, that would result in:
2271 *
2272 * A(W1)
2273 * WFC(C)
2274 * A(W1)
2275 * C(C)
2276 *
2277 * Which would create W1->C->W1 dependencies, even though there is no
2278 * actual deadlock possible. There are two solutions, using a
2279 * read-recursive acquire on the work(queue) 'locks', but this will then
2280 * hit the lockdep limitation on recursive locks, or simply discard
2281 * these locks.
2282 *
2283 * AFAICT there is no possible deadlock scenario between the
2284 * flush_work() and complete() primitives (except for single-threaded
2285 * workqueues), so hiding them isn't a problem.
2286 */
2287 lockdep_invariant_state(true);
2288 trace_workqueue_execute_start(work);
2289 worker->current_func(work);
2290 /*
2291 * While we must be careful to not use "work" after this, the trace
2292 * point will only record its address.
2293 */
2294 trace_workqueue_execute_end(work, worker->current_func);
2295 lock_map_release(&lockdep_map);
2296 lock_map_release(&pwq->wq->lockdep_map);
2297
2298 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2299 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2300 " last function: %ps\n",
2301 current->comm, preempt_count(), task_pid_nr(current),
2302 worker->current_func);
2303 debug_show_held_locks(current);
2304 dump_stack();
2305 }
2306
2307 /*
2308 * The following prevents a kworker from hogging CPU on !PREEMPTION
2309 * kernels, where a requeueing work item waiting for something to
2310 * happen could deadlock with stop_machine as such work item could
2311 * indefinitely requeue itself while all other CPUs are trapped in
2312 * stop_machine. At the same time, report a quiescent RCU state so
2313 * the same condition doesn't freeze RCU.
2314 */
2315 cond_resched();
2316
2317 raw_spin_lock_irq(&pool->lock);
2318
2319 /* clear cpu intensive status */
2320 if (unlikely(cpu_intensive))
2321 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2322
2323 /* tag the worker for identification in schedule() */
2324 worker->last_func = worker->current_func;
2325
2326 /* we're done with it, release */
2327 hash_del(&worker->hentry);
2328 worker->current_work = NULL;
2329 worker->current_func = NULL;
2330 worker->current_pwq = NULL;
2331 worker->current_color = INT_MAX;
2332 pwq_dec_nr_in_flight(pwq, work_data);
2333}
2334
2335/**
2336 * process_scheduled_works - process scheduled works
2337 * @worker: self
2338 *
2339 * Process all scheduled works. Please note that the scheduled list
2340 * may change while processing a work, so this function repeatedly
2341 * fetches a work from the top and executes it.
2342 *
2343 * CONTEXT:
2344 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2345 * multiple times.
2346 */
2347static void process_scheduled_works(struct worker *worker)
2348{
2349 while (!list_empty(&worker->scheduled)) {
2350 struct work_struct *work = list_first_entry(&worker->scheduled,
2351 struct work_struct, entry);
2352 process_one_work(worker, work);
2353 }
2354}
2355
2356static void set_pf_worker(bool val)
2357{
2358 mutex_lock(&wq_pool_attach_mutex);
2359 if (val)
2360 current->flags |= PF_WQ_WORKER;
2361 else
2362 current->flags &= ~PF_WQ_WORKER;
2363 mutex_unlock(&wq_pool_attach_mutex);
2364}
2365
2366/**
2367 * worker_thread - the worker thread function
2368 * @__worker: self
2369 *
2370 * The worker thread function. All workers belong to a worker_pool -
2371 * either a per-cpu one or dynamic unbound one. These workers process all
2372 * work items regardless of their specific target workqueue. The only
2373 * exception is work items which belong to workqueues with a rescuer which
2374 * will be explained in rescuer_thread().
2375 *
2376 * Return: 0
2377 */
2378static int worker_thread(void *__worker)
2379{
2380 struct worker *worker = __worker;
2381 struct worker_pool *pool = worker->pool;
2382
2383 /* tell the scheduler that this is a workqueue worker */
2384 set_pf_worker(true);
2385woke_up:
2386 raw_spin_lock_irq(&pool->lock);
2387
2388 /* am I supposed to die? */
2389 if (unlikely(worker->flags & WORKER_DIE)) {
2390 raw_spin_unlock_irq(&pool->lock);
2391 WARN_ON_ONCE(!list_empty(&worker->entry));
2392 set_pf_worker(false);
2393
2394 set_task_comm(worker->task, "kworker/dying");
2395 ida_free(&pool->worker_ida, worker->id);
2396 worker_detach_from_pool(worker);
2397 kfree(worker);
2398 return 0;
2399 }
2400
2401 worker_leave_idle(worker);
2402recheck:
2403 /* no more worker necessary? */
2404 if (!need_more_worker(pool))
2405 goto sleep;
2406
2407 /* do we need to manage? */
2408 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2409 goto recheck;
2410
2411 /*
2412 * ->scheduled list can only be filled while a worker is
2413 * preparing to process a work or actually processing it.
2414 * Make sure nobody diddled with it while I was sleeping.
2415 */
2416 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2417
2418 /*
2419 * Finish PREP stage. We're guaranteed to have at least one idle
2420 * worker or that someone else has already assumed the manager
2421 * role. This is where @worker starts participating in concurrency
2422 * management if applicable and concurrency management is restored
2423 * after being rebound. See rebind_workers() for details.
2424 */
2425 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2426
2427 do {
2428 struct work_struct *work =
2429 list_first_entry(&pool->worklist,
2430 struct work_struct, entry);
2431
2432 pool->watchdog_ts = jiffies;
2433
2434 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2435 /* optimization path, not strictly necessary */
2436 process_one_work(worker, work);
2437 if (unlikely(!list_empty(&worker->scheduled)))
2438 process_scheduled_works(worker);
2439 } else {
2440 move_linked_works(work, &worker->scheduled, NULL);
2441 process_scheduled_works(worker);
2442 }
2443 } while (keep_working(pool));
2444
2445 worker_set_flags(worker, WORKER_PREP);
2446sleep:
2447 /*
2448 * pool->lock is held and there's no work to process and no need to
2449 * manage, sleep. Workers are woken up only while holding
2450 * pool->lock or from local cpu, so setting the current state
2451 * before releasing pool->lock is enough to prevent losing any
2452 * event.
2453 */
2454 worker_enter_idle(worker);
2455 __set_current_state(TASK_IDLE);
2456 raw_spin_unlock_irq(&pool->lock);
2457 schedule();
2458 goto woke_up;
2459}
2460
2461/**
2462 * rescuer_thread - the rescuer thread function
2463 * @__rescuer: self
2464 *
2465 * Workqueue rescuer thread function. There's one rescuer for each
2466 * workqueue which has WQ_MEM_RECLAIM set.
2467 *
2468 * Regular work processing on a pool may block trying to create a new
2469 * worker which uses GFP_KERNEL allocation which has slight chance of
2470 * developing into deadlock if some works currently on the same queue
2471 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2472 * the problem rescuer solves.
2473 *
2474 * When such condition is possible, the pool summons rescuers of all
2475 * workqueues which have works queued on the pool and let them process
2476 * those works so that forward progress can be guaranteed.
2477 *
2478 * This should happen rarely.
2479 *
2480 * Return: 0
2481 */
2482static int rescuer_thread(void *__rescuer)
2483{
2484 struct worker *rescuer = __rescuer;
2485 struct workqueue_struct *wq = rescuer->rescue_wq;
2486 struct list_head *scheduled = &rescuer->scheduled;
2487 bool should_stop;
2488
2489 set_user_nice(current, RESCUER_NICE_LEVEL);
2490
2491 /*
2492 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2493 * doesn't participate in concurrency management.
2494 */
2495 set_pf_worker(true);
2496repeat:
2497 set_current_state(TASK_IDLE);
2498
2499 /*
2500 * By the time the rescuer is requested to stop, the workqueue
2501 * shouldn't have any work pending, but @wq->maydays may still have
2502 * pwq(s) queued. This can happen by non-rescuer workers consuming
2503 * all the work items before the rescuer got to them. Go through
2504 * @wq->maydays processing before acting on should_stop so that the
2505 * list is always empty on exit.
2506 */
2507 should_stop = kthread_should_stop();
2508
2509 /* see whether any pwq is asking for help */
2510 raw_spin_lock_irq(&wq_mayday_lock);
2511
2512 while (!list_empty(&wq->maydays)) {
2513 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2514 struct pool_workqueue, mayday_node);
2515 struct worker_pool *pool = pwq->pool;
2516 struct work_struct *work, *n;
2517 bool first = true;
2518
2519 __set_current_state(TASK_RUNNING);
2520 list_del_init(&pwq->mayday_node);
2521
2522 raw_spin_unlock_irq(&wq_mayday_lock);
2523
2524 worker_attach_to_pool(rescuer, pool);
2525
2526 raw_spin_lock_irq(&pool->lock);
2527
2528 /*
2529 * Slurp in all works issued via this workqueue and
2530 * process'em.
2531 */
2532 WARN_ON_ONCE(!list_empty(scheduled));
2533 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2534 if (get_work_pwq(work) == pwq) {
2535 if (first)
2536 pool->watchdog_ts = jiffies;
2537 move_linked_works(work, scheduled, &n);
2538 }
2539 first = false;
2540 }
2541
2542 if (!list_empty(scheduled)) {
2543 process_scheduled_works(rescuer);
2544
2545 /*
2546 * The above execution of rescued work items could
2547 * have created more to rescue through
2548 * pwq_activate_first_inactive() or chained
2549 * queueing. Let's put @pwq back on mayday list so
2550 * that such back-to-back work items, which may be
2551 * being used to relieve memory pressure, don't
2552 * incur MAYDAY_INTERVAL delay inbetween.
2553 */
2554 if (pwq->nr_active && need_to_create_worker(pool)) {
2555 raw_spin_lock(&wq_mayday_lock);
2556 /*
2557 * Queue iff we aren't racing destruction
2558 * and somebody else hasn't queued it already.
2559 */
2560 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2561 get_pwq(pwq);
2562 list_add_tail(&pwq->mayday_node, &wq->maydays);
2563 }
2564 raw_spin_unlock(&wq_mayday_lock);
2565 }
2566 }
2567
2568 /*
2569 * Put the reference grabbed by send_mayday(). @pool won't
2570 * go away while we're still attached to it.
2571 */
2572 put_pwq(pwq);
2573
2574 /*
2575 * Leave this pool. If need_more_worker() is %true, notify a
2576 * regular worker; otherwise, we end up with 0 concurrency
2577 * and stalling the execution.
2578 */
2579 if (need_more_worker(pool))
2580 wake_up_worker(pool);
2581
2582 raw_spin_unlock_irq(&pool->lock);
2583
2584 worker_detach_from_pool(rescuer);
2585
2586 raw_spin_lock_irq(&wq_mayday_lock);
2587 }
2588
2589 raw_spin_unlock_irq(&wq_mayday_lock);
2590
2591 if (should_stop) {
2592 __set_current_state(TASK_RUNNING);
2593 set_pf_worker(false);
2594 return 0;
2595 }
2596
2597 /* rescuers should never participate in concurrency management */
2598 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2599 schedule();
2600 goto repeat;
2601}
2602
2603/**
2604 * check_flush_dependency - check for flush dependency sanity
2605 * @target_wq: workqueue being flushed
2606 * @target_work: work item being flushed (NULL for workqueue flushes)
2607 *
2608 * %current is trying to flush the whole @target_wq or @target_work on it.
2609 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2610 * reclaiming memory or running on a workqueue which doesn't have
2611 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2612 * a deadlock.
2613 */
2614static void check_flush_dependency(struct workqueue_struct *target_wq,
2615 struct work_struct *target_work)
2616{
2617 work_func_t target_func = target_work ? target_work->func : NULL;
2618 struct worker *worker;
2619
2620 if (target_wq->flags & WQ_MEM_RECLAIM)
2621 return;
2622
2623 worker = current_wq_worker();
2624
2625 WARN_ONCE(current->flags & PF_MEMALLOC,
2626 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2627 current->pid, current->comm, target_wq->name, target_func);
2628 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2629 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2630 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2631 worker->current_pwq->wq->name, worker->current_func,
2632 target_wq->name, target_func);
2633}
2634
2635struct wq_barrier {
2636 struct work_struct work;
2637 struct completion done;
2638 struct task_struct *task; /* purely informational */
2639};
2640
2641static void wq_barrier_func(struct work_struct *work)
2642{
2643 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2644 complete(&barr->done);
2645}
2646
2647/**
2648 * insert_wq_barrier - insert a barrier work
2649 * @pwq: pwq to insert barrier into
2650 * @barr: wq_barrier to insert
2651 * @target: target work to attach @barr to
2652 * @worker: worker currently executing @target, NULL if @target is not executing
2653 *
2654 * @barr is linked to @target such that @barr is completed only after
2655 * @target finishes execution. Please note that the ordering
2656 * guarantee is observed only with respect to @target and on the local
2657 * cpu.
2658 *
2659 * Currently, a queued barrier can't be canceled. This is because
2660 * try_to_grab_pending() can't determine whether the work to be
2661 * grabbed is at the head of the queue and thus can't clear LINKED
2662 * flag of the previous work while there must be a valid next work
2663 * after a work with LINKED flag set.
2664 *
2665 * Note that when @worker is non-NULL, @target may be modified
2666 * underneath us, so we can't reliably determine pwq from @target.
2667 *
2668 * CONTEXT:
2669 * raw_spin_lock_irq(pool->lock).
2670 */
2671static void insert_wq_barrier(struct pool_workqueue *pwq,
2672 struct wq_barrier *barr,
2673 struct work_struct *target, struct worker *worker)
2674{
2675 unsigned int work_flags = 0;
2676 unsigned int work_color;
2677 struct list_head *head;
2678
2679 /*
2680 * debugobject calls are safe here even with pool->lock locked
2681 * as we know for sure that this will not trigger any of the
2682 * checks and call back into the fixup functions where we
2683 * might deadlock.
2684 */
2685 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2686 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2687
2688 init_completion_map(&barr->done, &target->lockdep_map);
2689
2690 barr->task = current;
2691
2692 /* The barrier work item does not participate in pwq->nr_active. */
2693 work_flags |= WORK_STRUCT_INACTIVE;
2694
2695 /*
2696 * If @target is currently being executed, schedule the
2697 * barrier to the worker; otherwise, put it after @target.
2698 */
2699 if (worker) {
2700 head = worker->scheduled.next;
2701 work_color = worker->current_color;
2702 } else {
2703 unsigned long *bits = work_data_bits(target);
2704
2705 head = target->entry.next;
2706 /* there can already be other linked works, inherit and set */
2707 work_flags |= *bits & WORK_STRUCT_LINKED;
2708 work_color = get_work_color(*bits);
2709 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2710 }
2711
2712 pwq->nr_in_flight[work_color]++;
2713 work_flags |= work_color_to_flags(work_color);
2714
2715 debug_work_activate(&barr->work);
2716 insert_work(pwq, &barr->work, head, work_flags);
2717}
2718
2719/**
2720 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2721 * @wq: workqueue being flushed
2722 * @flush_color: new flush color, < 0 for no-op
2723 * @work_color: new work color, < 0 for no-op
2724 *
2725 * Prepare pwqs for workqueue flushing.
2726 *
2727 * If @flush_color is non-negative, flush_color on all pwqs should be
2728 * -1. If no pwq has in-flight commands at the specified color, all
2729 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2730 * has in flight commands, its pwq->flush_color is set to
2731 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2732 * wakeup logic is armed and %true is returned.
2733 *
2734 * The caller should have initialized @wq->first_flusher prior to
2735 * calling this function with non-negative @flush_color. If
2736 * @flush_color is negative, no flush color update is done and %false
2737 * is returned.
2738 *
2739 * If @work_color is non-negative, all pwqs should have the same
2740 * work_color which is previous to @work_color and all will be
2741 * advanced to @work_color.
2742 *
2743 * CONTEXT:
2744 * mutex_lock(wq->mutex).
2745 *
2746 * Return:
2747 * %true if @flush_color >= 0 and there's something to flush. %false
2748 * otherwise.
2749 */
2750static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2751 int flush_color, int work_color)
2752{
2753 bool wait = false;
2754 struct pool_workqueue *pwq;
2755
2756 if (flush_color >= 0) {
2757 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2758 atomic_set(&wq->nr_pwqs_to_flush, 1);
2759 }
2760
2761 for_each_pwq(pwq, wq) {
2762 struct worker_pool *pool = pwq->pool;
2763
2764 raw_spin_lock_irq(&pool->lock);
2765
2766 if (flush_color >= 0) {
2767 WARN_ON_ONCE(pwq->flush_color != -1);
2768
2769 if (pwq->nr_in_flight[flush_color]) {
2770 pwq->flush_color = flush_color;
2771 atomic_inc(&wq->nr_pwqs_to_flush);
2772 wait = true;
2773 }
2774 }
2775
2776 if (work_color >= 0) {
2777 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2778 pwq->work_color = work_color;
2779 }
2780
2781 raw_spin_unlock_irq(&pool->lock);
2782 }
2783
2784 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2785 complete(&wq->first_flusher->done);
2786
2787 return wait;
2788}
2789
2790/**
2791 * __flush_workqueue - ensure that any scheduled work has run to completion.
2792 * @wq: workqueue to flush
2793 *
2794 * This function sleeps until all work items which were queued on entry
2795 * have finished execution, but it is not livelocked by new incoming ones.
2796 */
2797void __flush_workqueue(struct workqueue_struct *wq)
2798{
2799 struct wq_flusher this_flusher = {
2800 .list = LIST_HEAD_INIT(this_flusher.list),
2801 .flush_color = -1,
2802 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2803 };
2804 int next_color;
2805
2806 if (WARN_ON(!wq_online))
2807 return;
2808
2809 lock_map_acquire(&wq->lockdep_map);
2810 lock_map_release(&wq->lockdep_map);
2811
2812 mutex_lock(&wq->mutex);
2813
2814 /*
2815 * Start-to-wait phase
2816 */
2817 next_color = work_next_color(wq->work_color);
2818
2819 if (next_color != wq->flush_color) {
2820 /*
2821 * Color space is not full. The current work_color
2822 * becomes our flush_color and work_color is advanced
2823 * by one.
2824 */
2825 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2826 this_flusher.flush_color = wq->work_color;
2827 wq->work_color = next_color;
2828
2829 if (!wq->first_flusher) {
2830 /* no flush in progress, become the first flusher */
2831 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2832
2833 wq->first_flusher = &this_flusher;
2834
2835 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2836 wq->work_color)) {
2837 /* nothing to flush, done */
2838 wq->flush_color = next_color;
2839 wq->first_flusher = NULL;
2840 goto out_unlock;
2841 }
2842 } else {
2843 /* wait in queue */
2844 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2845 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2846 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2847 }
2848 } else {
2849 /*
2850 * Oops, color space is full, wait on overflow queue.
2851 * The next flush completion will assign us
2852 * flush_color and transfer to flusher_queue.
2853 */
2854 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2855 }
2856
2857 check_flush_dependency(wq, NULL);
2858
2859 mutex_unlock(&wq->mutex);
2860
2861 wait_for_completion(&this_flusher.done);
2862
2863 /*
2864 * Wake-up-and-cascade phase
2865 *
2866 * First flushers are responsible for cascading flushes and
2867 * handling overflow. Non-first flushers can simply return.
2868 */
2869 if (READ_ONCE(wq->first_flusher) != &this_flusher)
2870 return;
2871
2872 mutex_lock(&wq->mutex);
2873
2874 /* we might have raced, check again with mutex held */
2875 if (wq->first_flusher != &this_flusher)
2876 goto out_unlock;
2877
2878 WRITE_ONCE(wq->first_flusher, NULL);
2879
2880 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2881 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2882
2883 while (true) {
2884 struct wq_flusher *next, *tmp;
2885
2886 /* complete all the flushers sharing the current flush color */
2887 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2888 if (next->flush_color != wq->flush_color)
2889 break;
2890 list_del_init(&next->list);
2891 complete(&next->done);
2892 }
2893
2894 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2895 wq->flush_color != work_next_color(wq->work_color));
2896
2897 /* this flush_color is finished, advance by one */
2898 wq->flush_color = work_next_color(wq->flush_color);
2899
2900 /* one color has been freed, handle overflow queue */
2901 if (!list_empty(&wq->flusher_overflow)) {
2902 /*
2903 * Assign the same color to all overflowed
2904 * flushers, advance work_color and append to
2905 * flusher_queue. This is the start-to-wait
2906 * phase for these overflowed flushers.
2907 */
2908 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2909 tmp->flush_color = wq->work_color;
2910
2911 wq->work_color = work_next_color(wq->work_color);
2912
2913 list_splice_tail_init(&wq->flusher_overflow,
2914 &wq->flusher_queue);
2915 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2916 }
2917
2918 if (list_empty(&wq->flusher_queue)) {
2919 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2920 break;
2921 }
2922
2923 /*
2924 * Need to flush more colors. Make the next flusher
2925 * the new first flusher and arm pwqs.
2926 */
2927 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2928 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2929
2930 list_del_init(&next->list);
2931 wq->first_flusher = next;
2932
2933 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2934 break;
2935
2936 /*
2937 * Meh... this color is already done, clear first
2938 * flusher and repeat cascading.
2939 */
2940 wq->first_flusher = NULL;
2941 }
2942
2943out_unlock:
2944 mutex_unlock(&wq->mutex);
2945}
2946EXPORT_SYMBOL(__flush_workqueue);
2947
2948/**
2949 * drain_workqueue - drain a workqueue
2950 * @wq: workqueue to drain
2951 *
2952 * Wait until the workqueue becomes empty. While draining is in progress,
2953 * only chain queueing is allowed. IOW, only currently pending or running
2954 * work items on @wq can queue further work items on it. @wq is flushed
2955 * repeatedly until it becomes empty. The number of flushing is determined
2956 * by the depth of chaining and should be relatively short. Whine if it
2957 * takes too long.
2958 */
2959void drain_workqueue(struct workqueue_struct *wq)
2960{
2961 unsigned int flush_cnt = 0;
2962 struct pool_workqueue *pwq;
2963
2964 /*
2965 * __queue_work() needs to test whether there are drainers, is much
2966 * hotter than drain_workqueue() and already looks at @wq->flags.
2967 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2968 */
2969 mutex_lock(&wq->mutex);
2970 if (!wq->nr_drainers++)
2971 wq->flags |= __WQ_DRAINING;
2972 mutex_unlock(&wq->mutex);
2973reflush:
2974 __flush_workqueue(wq);
2975
2976 mutex_lock(&wq->mutex);
2977
2978 for_each_pwq(pwq, wq) {
2979 bool drained;
2980
2981 raw_spin_lock_irq(&pwq->pool->lock);
2982 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
2983 raw_spin_unlock_irq(&pwq->pool->lock);
2984
2985 if (drained)
2986 continue;
2987
2988 if (++flush_cnt == 10 ||
2989 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2990 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
2991 wq->name, __func__, flush_cnt);
2992
2993 mutex_unlock(&wq->mutex);
2994 goto reflush;
2995 }
2996
2997 if (!--wq->nr_drainers)
2998 wq->flags &= ~__WQ_DRAINING;
2999 mutex_unlock(&wq->mutex);
3000}
3001EXPORT_SYMBOL_GPL(drain_workqueue);
3002
3003static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3004 bool from_cancel)
3005{
3006 struct worker *worker = NULL;
3007 struct worker_pool *pool;
3008 struct pool_workqueue *pwq;
3009
3010 might_sleep();
3011
3012 rcu_read_lock();
3013 pool = get_work_pool(work);
3014 if (!pool) {
3015 rcu_read_unlock();
3016 return false;
3017 }
3018
3019 raw_spin_lock_irq(&pool->lock);
3020 /* see the comment in try_to_grab_pending() with the same code */
3021 pwq = get_work_pwq(work);
3022 if (pwq) {
3023 if (unlikely(pwq->pool != pool))
3024 goto already_gone;
3025 } else {
3026 worker = find_worker_executing_work(pool, work);
3027 if (!worker)
3028 goto already_gone;
3029 pwq = worker->current_pwq;
3030 }
3031
3032 check_flush_dependency(pwq->wq, work);
3033
3034 insert_wq_barrier(pwq, barr, work, worker);
3035 raw_spin_unlock_irq(&pool->lock);
3036
3037 /*
3038 * Force a lock recursion deadlock when using flush_work() inside a
3039 * single-threaded or rescuer equipped workqueue.
3040 *
3041 * For single threaded workqueues the deadlock happens when the work
3042 * is after the work issuing the flush_work(). For rescuer equipped
3043 * workqueues the deadlock happens when the rescuer stalls, blocking
3044 * forward progress.
3045 */
3046 if (!from_cancel &&
3047 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3048 lock_map_acquire(&pwq->wq->lockdep_map);
3049 lock_map_release(&pwq->wq->lockdep_map);
3050 }
3051 rcu_read_unlock();
3052 return true;
3053already_gone:
3054 raw_spin_unlock_irq(&pool->lock);
3055 rcu_read_unlock();
3056 return false;
3057}
3058
3059static bool __flush_work(struct work_struct *work, bool from_cancel)
3060{
3061 struct wq_barrier barr;
3062
3063 if (WARN_ON(!wq_online))
3064 return false;
3065
3066 if (WARN_ON(!work->func))
3067 return false;
3068
3069 lock_map_acquire(&work->lockdep_map);
3070 lock_map_release(&work->lockdep_map);
3071
3072 if (start_flush_work(work, &barr, from_cancel)) {
3073 wait_for_completion(&barr.done);
3074 destroy_work_on_stack(&barr.work);
3075 return true;
3076 } else {
3077 return false;
3078 }
3079}
3080
3081/**
3082 * flush_work - wait for a work to finish executing the last queueing instance
3083 * @work: the work to flush
3084 *
3085 * Wait until @work has finished execution. @work is guaranteed to be idle
3086 * on return if it hasn't been requeued since flush started.
3087 *
3088 * Return:
3089 * %true if flush_work() waited for the work to finish execution,
3090 * %false if it was already idle.
3091 */
3092bool flush_work(struct work_struct *work)
3093{
3094 return __flush_work(work, false);
3095}
3096EXPORT_SYMBOL_GPL(flush_work);
3097
3098struct cwt_wait {
3099 wait_queue_entry_t wait;
3100 struct work_struct *work;
3101};
3102
3103static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3104{
3105 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3106
3107 if (cwait->work != key)
3108 return 0;
3109 return autoremove_wake_function(wait, mode, sync, key);
3110}
3111
3112static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3113{
3114 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3115 unsigned long flags;
3116 int ret;
3117
3118 do {
3119 ret = try_to_grab_pending(work, is_dwork, &flags);
3120 /*
3121 * If someone else is already canceling, wait for it to
3122 * finish. flush_work() doesn't work for PREEMPT_NONE
3123 * because we may get scheduled between @work's completion
3124 * and the other canceling task resuming and clearing
3125 * CANCELING - flush_work() will return false immediately
3126 * as @work is no longer busy, try_to_grab_pending() will
3127 * return -ENOENT as @work is still being canceled and the
3128 * other canceling task won't be able to clear CANCELING as
3129 * we're hogging the CPU.
3130 *
3131 * Let's wait for completion using a waitqueue. As this
3132 * may lead to the thundering herd problem, use a custom
3133 * wake function which matches @work along with exclusive
3134 * wait and wakeup.
3135 */
3136 if (unlikely(ret == -ENOENT)) {
3137 struct cwt_wait cwait;
3138
3139 init_wait(&cwait.wait);
3140 cwait.wait.func = cwt_wakefn;
3141 cwait.work = work;
3142
3143 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3144 TASK_UNINTERRUPTIBLE);
3145 if (work_is_canceling(work))
3146 schedule();
3147 finish_wait(&cancel_waitq, &cwait.wait);
3148 }
3149 } while (unlikely(ret < 0));
3150
3151 /* tell other tasks trying to grab @work to back off */
3152 mark_work_canceling(work);
3153 local_irq_restore(flags);
3154
3155 /*
3156 * This allows canceling during early boot. We know that @work
3157 * isn't executing.
3158 */
3159 if (wq_online)
3160 __flush_work(work, true);
3161
3162 clear_work_data(work);
3163
3164 /*
3165 * Paired with prepare_to_wait() above so that either
3166 * waitqueue_active() is visible here or !work_is_canceling() is
3167 * visible there.
3168 */
3169 smp_mb();
3170 if (waitqueue_active(&cancel_waitq))
3171 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3172
3173 return ret;
3174}
3175
3176/**
3177 * cancel_work_sync - cancel a work and wait for it to finish
3178 * @work: the work to cancel
3179 *
3180 * Cancel @work and wait for its execution to finish. This function
3181 * can be used even if the work re-queues itself or migrates to
3182 * another workqueue. On return from this function, @work is
3183 * guaranteed to be not pending or executing on any CPU.
3184 *
3185 * cancel_work_sync(&delayed_work->work) must not be used for
3186 * delayed_work's. Use cancel_delayed_work_sync() instead.
3187 *
3188 * The caller must ensure that the workqueue on which @work was last
3189 * queued can't be destroyed before this function returns.
3190 *
3191 * Return:
3192 * %true if @work was pending, %false otherwise.
3193 */
3194bool cancel_work_sync(struct work_struct *work)
3195{
3196 return __cancel_work_timer(work, false);
3197}
3198EXPORT_SYMBOL_GPL(cancel_work_sync);
3199
3200/**
3201 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3202 * @dwork: the delayed work to flush
3203 *
3204 * Delayed timer is cancelled and the pending work is queued for
3205 * immediate execution. Like flush_work(), this function only
3206 * considers the last queueing instance of @dwork.
3207 *
3208 * Return:
3209 * %true if flush_work() waited for the work to finish execution,
3210 * %false if it was already idle.
3211 */
3212bool flush_delayed_work(struct delayed_work *dwork)
3213{
3214 local_irq_disable();
3215 if (del_timer_sync(&dwork->timer))
3216 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3217 local_irq_enable();
3218 return flush_work(&dwork->work);
3219}
3220EXPORT_SYMBOL(flush_delayed_work);
3221
3222/**
3223 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3224 * @rwork: the rcu work to flush
3225 *
3226 * Return:
3227 * %true if flush_rcu_work() waited for the work to finish execution,
3228 * %false if it was already idle.
3229 */
3230bool flush_rcu_work(struct rcu_work *rwork)
3231{
3232 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3233 rcu_barrier();
3234 flush_work(&rwork->work);
3235 return true;
3236 } else {
3237 return flush_work(&rwork->work);
3238 }
3239}
3240EXPORT_SYMBOL(flush_rcu_work);
3241
3242static bool __cancel_work(struct work_struct *work, bool is_dwork)
3243{
3244 unsigned long flags;
3245 int ret;
3246
3247 do {
3248 ret = try_to_grab_pending(work, is_dwork, &flags);
3249 } while (unlikely(ret == -EAGAIN));
3250
3251 if (unlikely(ret < 0))
3252 return false;
3253
3254 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3255 local_irq_restore(flags);
3256 return ret;
3257}
3258
3259/*
3260 * See cancel_delayed_work()
3261 */
3262bool cancel_work(struct work_struct *work)
3263{
3264 return __cancel_work(work, false);
3265}
3266EXPORT_SYMBOL(cancel_work);
3267
3268/**
3269 * cancel_delayed_work - cancel a delayed work
3270 * @dwork: delayed_work to cancel
3271 *
3272 * Kill off a pending delayed_work.
3273 *
3274 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3275 * pending.
3276 *
3277 * Note:
3278 * The work callback function may still be running on return, unless
3279 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3280 * use cancel_delayed_work_sync() to wait on it.
3281 *
3282 * This function is safe to call from any context including IRQ handler.
3283 */
3284bool cancel_delayed_work(struct delayed_work *dwork)
3285{
3286 return __cancel_work(&dwork->work, true);
3287}
3288EXPORT_SYMBOL(cancel_delayed_work);
3289
3290/**
3291 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3292 * @dwork: the delayed work cancel
3293 *
3294 * This is cancel_work_sync() for delayed works.
3295 *
3296 * Return:
3297 * %true if @dwork was pending, %false otherwise.
3298 */
3299bool cancel_delayed_work_sync(struct delayed_work *dwork)
3300{
3301 return __cancel_work_timer(&dwork->work, true);
3302}
3303EXPORT_SYMBOL(cancel_delayed_work_sync);
3304
3305/**
3306 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3307 * @func: the function to call
3308 *
3309 * schedule_on_each_cpu() executes @func on each online CPU using the
3310 * system workqueue and blocks until all CPUs have completed.
3311 * schedule_on_each_cpu() is very slow.
3312 *
3313 * Return:
3314 * 0 on success, -errno on failure.
3315 */
3316int schedule_on_each_cpu(work_func_t func)
3317{
3318 int cpu;
3319 struct work_struct __percpu *works;
3320
3321 works = alloc_percpu(struct work_struct);
3322 if (!works)
3323 return -ENOMEM;
3324
3325 cpus_read_lock();
3326
3327 for_each_online_cpu(cpu) {
3328 struct work_struct *work = per_cpu_ptr(works, cpu);
3329
3330 INIT_WORK(work, func);
3331 schedule_work_on(cpu, work);
3332 }
3333
3334 for_each_online_cpu(cpu)
3335 flush_work(per_cpu_ptr(works, cpu));
3336
3337 cpus_read_unlock();
3338 free_percpu(works);
3339 return 0;
3340}
3341
3342/**
3343 * execute_in_process_context - reliably execute the routine with user context
3344 * @fn: the function to execute
3345 * @ew: guaranteed storage for the execute work structure (must
3346 * be available when the work executes)
3347 *
3348 * Executes the function immediately if process context is available,
3349 * otherwise schedules the function for delayed execution.
3350 *
3351 * Return: 0 - function was executed
3352 * 1 - function was scheduled for execution
3353 */
3354int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3355{
3356 if (!in_interrupt()) {
3357 fn(&ew->work);
3358 return 0;
3359 }
3360
3361 INIT_WORK(&ew->work, fn);
3362 schedule_work(&ew->work);
3363
3364 return 1;
3365}
3366EXPORT_SYMBOL_GPL(execute_in_process_context);
3367
3368/**
3369 * free_workqueue_attrs - free a workqueue_attrs
3370 * @attrs: workqueue_attrs to free
3371 *
3372 * Undo alloc_workqueue_attrs().
3373 */
3374void free_workqueue_attrs(struct workqueue_attrs *attrs)
3375{
3376 if (attrs) {
3377 free_cpumask_var(attrs->cpumask);
3378 kfree(attrs);
3379 }
3380}
3381
3382/**
3383 * alloc_workqueue_attrs - allocate a workqueue_attrs
3384 *
3385 * Allocate a new workqueue_attrs, initialize with default settings and
3386 * return it.
3387 *
3388 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3389 */
3390struct workqueue_attrs *alloc_workqueue_attrs(void)
3391{
3392 struct workqueue_attrs *attrs;
3393
3394 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3395 if (!attrs)
3396 goto fail;
3397 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3398 goto fail;
3399
3400 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3401 return attrs;
3402fail:
3403 free_workqueue_attrs(attrs);
3404 return NULL;
3405}
3406
3407static void copy_workqueue_attrs(struct workqueue_attrs *to,
3408 const struct workqueue_attrs *from)
3409{
3410 to->nice = from->nice;
3411 cpumask_copy(to->cpumask, from->cpumask);
3412 /*
3413 * Unlike hash and equality test, this function doesn't ignore
3414 * ->no_numa as it is used for both pool and wq attrs. Instead,
3415 * get_unbound_pool() explicitly clears ->no_numa after copying.
3416 */
3417 to->no_numa = from->no_numa;
3418}
3419
3420/* hash value of the content of @attr */
3421static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3422{
3423 u32 hash = 0;
3424
3425 hash = jhash_1word(attrs->nice, hash);
3426 hash = jhash(cpumask_bits(attrs->cpumask),
3427 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3428 return hash;
3429}
3430
3431/* content equality test */
3432static bool wqattrs_equal(const struct workqueue_attrs *a,
3433 const struct workqueue_attrs *b)
3434{
3435 if (a->nice != b->nice)
3436 return false;
3437 if (!cpumask_equal(a->cpumask, b->cpumask))
3438 return false;
3439 return true;
3440}
3441
3442/**
3443 * init_worker_pool - initialize a newly zalloc'd worker_pool
3444 * @pool: worker_pool to initialize
3445 *
3446 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3447 *
3448 * Return: 0 on success, -errno on failure. Even on failure, all fields
3449 * inside @pool proper are initialized and put_unbound_pool() can be called
3450 * on @pool safely to release it.
3451 */
3452static int init_worker_pool(struct worker_pool *pool)
3453{
3454 raw_spin_lock_init(&pool->lock);
3455 pool->id = -1;
3456 pool->cpu = -1;
3457 pool->node = NUMA_NO_NODE;
3458 pool->flags |= POOL_DISASSOCIATED;
3459 pool->watchdog_ts = jiffies;
3460 INIT_LIST_HEAD(&pool->worklist);
3461 INIT_LIST_HEAD(&pool->idle_list);
3462 hash_init(pool->busy_hash);
3463
3464 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3465
3466 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3467
3468 INIT_LIST_HEAD(&pool->workers);
3469
3470 ida_init(&pool->worker_ida);
3471 INIT_HLIST_NODE(&pool->hash_node);
3472 pool->refcnt = 1;
3473
3474 /* shouldn't fail above this point */
3475 pool->attrs = alloc_workqueue_attrs();
3476 if (!pool->attrs)
3477 return -ENOMEM;
3478 return 0;
3479}
3480
3481#ifdef CONFIG_LOCKDEP
3482static void wq_init_lockdep(struct workqueue_struct *wq)
3483{
3484 char *lock_name;
3485
3486 lockdep_register_key(&wq->key);
3487 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3488 if (!lock_name)
3489 lock_name = wq->name;
3490
3491 wq->lock_name = lock_name;
3492 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3493}
3494
3495static void wq_unregister_lockdep(struct workqueue_struct *wq)
3496{
3497 lockdep_unregister_key(&wq->key);
3498}
3499
3500static void wq_free_lockdep(struct workqueue_struct *wq)
3501{
3502 if (wq->lock_name != wq->name)
3503 kfree(wq->lock_name);
3504}
3505#else
3506static void wq_init_lockdep(struct workqueue_struct *wq)
3507{
3508}
3509
3510static void wq_unregister_lockdep(struct workqueue_struct *wq)
3511{
3512}
3513
3514static void wq_free_lockdep(struct workqueue_struct *wq)
3515{
3516}
3517#endif
3518
3519static void rcu_free_wq(struct rcu_head *rcu)
3520{
3521 struct workqueue_struct *wq =
3522 container_of(rcu, struct workqueue_struct, rcu);
3523
3524 wq_free_lockdep(wq);
3525
3526 if (!(wq->flags & WQ_UNBOUND))
3527 free_percpu(wq->cpu_pwqs);
3528 else
3529 free_workqueue_attrs(wq->unbound_attrs);
3530
3531 kfree(wq);
3532}
3533
3534static void rcu_free_pool(struct rcu_head *rcu)
3535{
3536 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3537
3538 ida_destroy(&pool->worker_ida);
3539 free_workqueue_attrs(pool->attrs);
3540 kfree(pool);
3541}
3542
3543/* This returns with the lock held on success (pool manager is inactive). */
3544static bool wq_manager_inactive(struct worker_pool *pool)
3545{
3546 raw_spin_lock_irq(&pool->lock);
3547
3548 if (pool->flags & POOL_MANAGER_ACTIVE) {
3549 raw_spin_unlock_irq(&pool->lock);
3550 return false;
3551 }
3552 return true;
3553}
3554
3555/**
3556 * put_unbound_pool - put a worker_pool
3557 * @pool: worker_pool to put
3558 *
3559 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3560 * safe manner. get_unbound_pool() calls this function on its failure path
3561 * and this function should be able to release pools which went through,
3562 * successfully or not, init_worker_pool().
3563 *
3564 * Should be called with wq_pool_mutex held.
3565 */
3566static void put_unbound_pool(struct worker_pool *pool)
3567{
3568 DECLARE_COMPLETION_ONSTACK(detach_completion);
3569 struct worker *worker;
3570
3571 lockdep_assert_held(&wq_pool_mutex);
3572
3573 if (--pool->refcnt)
3574 return;
3575
3576 /* sanity checks */
3577 if (WARN_ON(!(pool->cpu < 0)) ||
3578 WARN_ON(!list_empty(&pool->worklist)))
3579 return;
3580
3581 /* release id and unhash */
3582 if (pool->id >= 0)
3583 idr_remove(&worker_pool_idr, pool->id);
3584 hash_del(&pool->hash_node);
3585
3586 /*
3587 * Become the manager and destroy all workers. This prevents
3588 * @pool's workers from blocking on attach_mutex. We're the last
3589 * manager and @pool gets freed with the flag set.
3590 * Because of how wq_manager_inactive() works, we will hold the
3591 * spinlock after a successful wait.
3592 */
3593 rcuwait_wait_event(&manager_wait, wq_manager_inactive(pool),
3594 TASK_UNINTERRUPTIBLE);
3595 pool->flags |= POOL_MANAGER_ACTIVE;
3596
3597 while ((worker = first_idle_worker(pool)))
3598 destroy_worker(worker);
3599 WARN_ON(pool->nr_workers || pool->nr_idle);
3600 raw_spin_unlock_irq(&pool->lock);
3601
3602 mutex_lock(&wq_pool_attach_mutex);
3603 if (!list_empty(&pool->workers))
3604 pool->detach_completion = &detach_completion;
3605 mutex_unlock(&wq_pool_attach_mutex);
3606
3607 if (pool->detach_completion)
3608 wait_for_completion(pool->detach_completion);
3609
3610 /* shut down the timers */
3611 del_timer_sync(&pool->idle_timer);
3612 del_timer_sync(&pool->mayday_timer);
3613
3614 /* RCU protected to allow dereferences from get_work_pool() */
3615 call_rcu(&pool->rcu, rcu_free_pool);
3616}
3617
3618/**
3619 * get_unbound_pool - get a worker_pool with the specified attributes
3620 * @attrs: the attributes of the worker_pool to get
3621 *
3622 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3623 * reference count and return it. If there already is a matching
3624 * worker_pool, it will be used; otherwise, this function attempts to
3625 * create a new one.
3626 *
3627 * Should be called with wq_pool_mutex held.
3628 *
3629 * Return: On success, a worker_pool with the same attributes as @attrs.
3630 * On failure, %NULL.
3631 */
3632static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3633{
3634 u32 hash = wqattrs_hash(attrs);
3635 struct worker_pool *pool;
3636 int node;
3637 int target_node = NUMA_NO_NODE;
3638
3639 lockdep_assert_held(&wq_pool_mutex);
3640
3641 /* do we already have a matching pool? */
3642 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3643 if (wqattrs_equal(pool->attrs, attrs)) {
3644 pool->refcnt++;
3645 return pool;
3646 }
3647 }
3648
3649 /* if cpumask is contained inside a NUMA node, we belong to that node */
3650 if (wq_numa_enabled) {
3651 for_each_node(node) {
3652 if (cpumask_subset(attrs->cpumask,
3653 wq_numa_possible_cpumask[node])) {
3654 target_node = node;
3655 break;
3656 }
3657 }
3658 }
3659
3660 /* nope, create a new one */
3661 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3662 if (!pool || init_worker_pool(pool) < 0)
3663 goto fail;
3664
3665 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3666 copy_workqueue_attrs(pool->attrs, attrs);
3667 pool->node = target_node;
3668
3669 /*
3670 * no_numa isn't a worker_pool attribute, always clear it. See
3671 * 'struct workqueue_attrs' comments for detail.
3672 */
3673 pool->attrs->no_numa = false;
3674
3675 if (worker_pool_assign_id(pool) < 0)
3676 goto fail;
3677
3678 /* create and start the initial worker */
3679 if (wq_online && !create_worker(pool))
3680 goto fail;
3681
3682 /* install */
3683 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3684
3685 return pool;
3686fail:
3687 if (pool)
3688 put_unbound_pool(pool);
3689 return NULL;
3690}
3691
3692static void rcu_free_pwq(struct rcu_head *rcu)
3693{
3694 kmem_cache_free(pwq_cache,
3695 container_of(rcu, struct pool_workqueue, rcu));
3696}
3697
3698/*
3699 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3700 * and needs to be destroyed.
3701 */
3702static void pwq_unbound_release_workfn(struct work_struct *work)
3703{
3704 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3705 unbound_release_work);
3706 struct workqueue_struct *wq = pwq->wq;
3707 struct worker_pool *pool = pwq->pool;
3708 bool is_last = false;
3709
3710 /*
3711 * when @pwq is not linked, it doesn't hold any reference to the
3712 * @wq, and @wq is invalid to access.
3713 */
3714 if (!list_empty(&pwq->pwqs_node)) {
3715 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3716 return;
3717
3718 mutex_lock(&wq->mutex);
3719 list_del_rcu(&pwq->pwqs_node);
3720 is_last = list_empty(&wq->pwqs);
3721 mutex_unlock(&wq->mutex);
3722 }
3723
3724 mutex_lock(&wq_pool_mutex);
3725 put_unbound_pool(pool);
3726 mutex_unlock(&wq_pool_mutex);
3727
3728 call_rcu(&pwq->rcu, rcu_free_pwq);
3729
3730 /*
3731 * If we're the last pwq going away, @wq is already dead and no one
3732 * is gonna access it anymore. Schedule RCU free.
3733 */
3734 if (is_last) {
3735 wq_unregister_lockdep(wq);
3736 call_rcu(&wq->rcu, rcu_free_wq);
3737 }
3738}
3739
3740/**
3741 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3742 * @pwq: target pool_workqueue
3743 *
3744 * If @pwq isn't freezing, set @pwq->max_active to the associated
3745 * workqueue's saved_max_active and activate inactive work items
3746 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3747 */
3748static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3749{
3750 struct workqueue_struct *wq = pwq->wq;
3751 bool freezable = wq->flags & WQ_FREEZABLE;
3752 unsigned long flags;
3753
3754 /* for @wq->saved_max_active */
3755 lockdep_assert_held(&wq->mutex);
3756
3757 /* fast exit for non-freezable wqs */
3758 if (!freezable && pwq->max_active == wq->saved_max_active)
3759 return;
3760
3761 /* this function can be called during early boot w/ irq disabled */
3762 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
3763
3764 /*
3765 * During [un]freezing, the caller is responsible for ensuring that
3766 * this function is called at least once after @workqueue_freezing
3767 * is updated and visible.
3768 */
3769 if (!freezable || !workqueue_freezing) {
3770 bool kick = false;
3771
3772 pwq->max_active = wq->saved_max_active;
3773
3774 while (!list_empty(&pwq->inactive_works) &&
3775 pwq->nr_active < pwq->max_active) {
3776 pwq_activate_first_inactive(pwq);
3777 kick = true;
3778 }
3779
3780 /*
3781 * Need to kick a worker after thawed or an unbound wq's
3782 * max_active is bumped. In realtime scenarios, always kicking a
3783 * worker will cause interference on the isolated cpu cores, so
3784 * let's kick iff work items were activated.
3785 */
3786 if (kick)
3787 wake_up_worker(pwq->pool);
3788 } else {
3789 pwq->max_active = 0;
3790 }
3791
3792 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
3793}
3794
3795/* initialize newly allocated @pwq which is associated with @wq and @pool */
3796static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3797 struct worker_pool *pool)
3798{
3799 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3800
3801 memset(pwq, 0, sizeof(*pwq));
3802
3803 pwq->pool = pool;
3804 pwq->wq = wq;
3805 pwq->flush_color = -1;
3806 pwq->refcnt = 1;
3807 INIT_LIST_HEAD(&pwq->inactive_works);
3808 INIT_LIST_HEAD(&pwq->pwqs_node);
3809 INIT_LIST_HEAD(&pwq->mayday_node);
3810 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3811}
3812
3813/* sync @pwq with the current state of its associated wq and link it */
3814static void link_pwq(struct pool_workqueue *pwq)
3815{
3816 struct workqueue_struct *wq = pwq->wq;
3817
3818 lockdep_assert_held(&wq->mutex);
3819
3820 /* may be called multiple times, ignore if already linked */
3821 if (!list_empty(&pwq->pwqs_node))
3822 return;
3823
3824 /* set the matching work_color */
3825 pwq->work_color = wq->work_color;
3826
3827 /* sync max_active to the current setting */
3828 pwq_adjust_max_active(pwq);
3829
3830 /* link in @pwq */
3831 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3832}
3833
3834/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3835static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3836 const struct workqueue_attrs *attrs)
3837{
3838 struct worker_pool *pool;
3839 struct pool_workqueue *pwq;
3840
3841 lockdep_assert_held(&wq_pool_mutex);
3842
3843 pool = get_unbound_pool(attrs);
3844 if (!pool)
3845 return NULL;
3846
3847 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3848 if (!pwq) {
3849 put_unbound_pool(pool);
3850 return NULL;
3851 }
3852
3853 init_pwq(pwq, wq, pool);
3854 return pwq;
3855}
3856
3857/**
3858 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3859 * @attrs: the wq_attrs of the default pwq of the target workqueue
3860 * @node: the target NUMA node
3861 * @cpu_going_down: if >= 0, the CPU to consider as offline
3862 * @cpumask: outarg, the resulting cpumask
3863 *
3864 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3865 * @cpu_going_down is >= 0, that cpu is considered offline during
3866 * calculation. The result is stored in @cpumask.
3867 *
3868 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3869 * enabled and @node has online CPUs requested by @attrs, the returned
3870 * cpumask is the intersection of the possible CPUs of @node and
3871 * @attrs->cpumask.
3872 *
3873 * The caller is responsible for ensuring that the cpumask of @node stays
3874 * stable.
3875 *
3876 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3877 * %false if equal.
3878 */
3879static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3880 int cpu_going_down, cpumask_t *cpumask)
3881{
3882 if (!wq_numa_enabled || attrs->no_numa)
3883 goto use_dfl;
3884
3885 /* does @node have any online CPUs @attrs wants? */
3886 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3887 if (cpu_going_down >= 0)
3888 cpumask_clear_cpu(cpu_going_down, cpumask);
3889
3890 if (cpumask_empty(cpumask))
3891 goto use_dfl;
3892
3893 /* yeap, return possible CPUs in @node that @attrs wants */
3894 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3895
3896 if (cpumask_empty(cpumask)) {
3897 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3898 "possible intersect\n");
3899 return false;
3900 }
3901
3902 return !cpumask_equal(cpumask, attrs->cpumask);
3903
3904use_dfl:
3905 cpumask_copy(cpumask, attrs->cpumask);
3906 return false;
3907}
3908
3909/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3910static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3911 int node,
3912 struct pool_workqueue *pwq)
3913{
3914 struct pool_workqueue *old_pwq;
3915
3916 lockdep_assert_held(&wq_pool_mutex);
3917 lockdep_assert_held(&wq->mutex);
3918
3919 /* link_pwq() can handle duplicate calls */
3920 link_pwq(pwq);
3921
3922 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3923 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3924 return old_pwq;
3925}
3926
3927/* context to store the prepared attrs & pwqs before applying */
3928struct apply_wqattrs_ctx {
3929 struct workqueue_struct *wq; /* target workqueue */
3930 struct workqueue_attrs *attrs; /* attrs to apply */
3931 struct list_head list; /* queued for batching commit */
3932 struct pool_workqueue *dfl_pwq;
3933 struct pool_workqueue *pwq_tbl[];
3934};
3935
3936/* free the resources after success or abort */
3937static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3938{
3939 if (ctx) {
3940 int node;
3941
3942 for_each_node(node)
3943 put_pwq_unlocked(ctx->pwq_tbl[node]);
3944 put_pwq_unlocked(ctx->dfl_pwq);
3945
3946 free_workqueue_attrs(ctx->attrs);
3947
3948 kfree(ctx);
3949 }
3950}
3951
3952/* allocate the attrs and pwqs for later installation */
3953static struct apply_wqattrs_ctx *
3954apply_wqattrs_prepare(struct workqueue_struct *wq,
3955 const struct workqueue_attrs *attrs)
3956{
3957 struct apply_wqattrs_ctx *ctx;
3958 struct workqueue_attrs *new_attrs, *tmp_attrs;
3959 int node;
3960
3961 lockdep_assert_held(&wq_pool_mutex);
3962
3963 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
3964
3965 new_attrs = alloc_workqueue_attrs();
3966 tmp_attrs = alloc_workqueue_attrs();
3967 if (!ctx || !new_attrs || !tmp_attrs)
3968 goto out_free;
3969
3970 /*
3971 * Calculate the attrs of the default pwq.
3972 * If the user configured cpumask doesn't overlap with the
3973 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3974 */
3975 copy_workqueue_attrs(new_attrs, attrs);
3976 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3977 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3978 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3979
3980 /*
3981 * We may create multiple pwqs with differing cpumasks. Make a
3982 * copy of @new_attrs which will be modified and used to obtain
3983 * pools.
3984 */
3985 copy_workqueue_attrs(tmp_attrs, new_attrs);
3986
3987 /*
3988 * If something goes wrong during CPU up/down, we'll fall back to
3989 * the default pwq covering whole @attrs->cpumask. Always create
3990 * it even if we don't use it immediately.
3991 */
3992 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3993 if (!ctx->dfl_pwq)
3994 goto out_free;
3995
3996 for_each_node(node) {
3997 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3998 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3999 if (!ctx->pwq_tbl[node])
4000 goto out_free;
4001 } else {
4002 ctx->dfl_pwq->refcnt++;
4003 ctx->pwq_tbl[node] = ctx->dfl_pwq;
4004 }
4005 }
4006
4007 /* save the user configured attrs and sanitize it. */
4008 copy_workqueue_attrs(new_attrs, attrs);
4009 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4010 ctx->attrs = new_attrs;
4011
4012 ctx->wq = wq;
4013 free_workqueue_attrs(tmp_attrs);
4014 return ctx;
4015
4016out_free:
4017 free_workqueue_attrs(tmp_attrs);
4018 free_workqueue_attrs(new_attrs);
4019 apply_wqattrs_cleanup(ctx);
4020 return NULL;
4021}
4022
4023/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4024static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4025{
4026 int node;
4027
4028 /* all pwqs have been created successfully, let's install'em */
4029 mutex_lock(&ctx->wq->mutex);
4030
4031 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4032
4033 /* save the previous pwq and install the new one */
4034 for_each_node(node)
4035 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
4036 ctx->pwq_tbl[node]);
4037
4038 /* @dfl_pwq might not have been used, ensure it's linked */
4039 link_pwq(ctx->dfl_pwq);
4040 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4041
4042 mutex_unlock(&ctx->wq->mutex);
4043}
4044
4045static void apply_wqattrs_lock(void)
4046{
4047 /* CPUs should stay stable across pwq creations and installations */
4048 cpus_read_lock();
4049 mutex_lock(&wq_pool_mutex);
4050}
4051
4052static void apply_wqattrs_unlock(void)
4053{
4054 mutex_unlock(&wq_pool_mutex);
4055 cpus_read_unlock();
4056}
4057
4058static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4059 const struct workqueue_attrs *attrs)
4060{
4061 struct apply_wqattrs_ctx *ctx;
4062
4063 /* only unbound workqueues can change attributes */
4064 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4065 return -EINVAL;
4066
4067 /* creating multiple pwqs breaks ordering guarantee */
4068 if (!list_empty(&wq->pwqs)) {
4069 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4070 return -EINVAL;
4071
4072 wq->flags &= ~__WQ_ORDERED;
4073 }
4074
4075 ctx = apply_wqattrs_prepare(wq, attrs);
4076 if (!ctx)
4077 return -ENOMEM;
4078
4079 /* the ctx has been prepared successfully, let's commit it */
4080 apply_wqattrs_commit(ctx);
4081 apply_wqattrs_cleanup(ctx);
4082
4083 return 0;
4084}
4085
4086/**
4087 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4088 * @wq: the target workqueue
4089 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4090 *
4091 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
4092 * machines, this function maps a separate pwq to each NUMA node with
4093 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4094 * NUMA node it was issued on. Older pwqs are released as in-flight work
4095 * items finish. Note that a work item which repeatedly requeues itself
4096 * back-to-back will stay on its current pwq.
4097 *
4098 * Performs GFP_KERNEL allocations.
4099 *
4100 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4101 *
4102 * Return: 0 on success and -errno on failure.
4103 */
4104int apply_workqueue_attrs(struct workqueue_struct *wq,
4105 const struct workqueue_attrs *attrs)
4106{
4107 int ret;
4108
4109 lockdep_assert_cpus_held();
4110
4111 mutex_lock(&wq_pool_mutex);
4112 ret = apply_workqueue_attrs_locked(wq, attrs);
4113 mutex_unlock(&wq_pool_mutex);
4114
4115 return ret;
4116}
4117
4118/**
4119 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4120 * @wq: the target workqueue
4121 * @cpu: the CPU coming up or going down
4122 * @online: whether @cpu is coming up or going down
4123 *
4124 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4125 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4126 * @wq accordingly.
4127 *
4128 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4129 * falls back to @wq->dfl_pwq which may not be optimal but is always
4130 * correct.
4131 *
4132 * Note that when the last allowed CPU of a NUMA node goes offline for a
4133 * workqueue with a cpumask spanning multiple nodes, the workers which were
4134 * already executing the work items for the workqueue will lose their CPU
4135 * affinity and may execute on any CPU. This is similar to how per-cpu
4136 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4137 * affinity, it's the user's responsibility to flush the work item from
4138 * CPU_DOWN_PREPARE.
4139 */
4140static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4141 bool online)
4142{
4143 int node = cpu_to_node(cpu);
4144 int cpu_off = online ? -1 : cpu;
4145 struct pool_workqueue *old_pwq = NULL, *pwq;
4146 struct workqueue_attrs *target_attrs;
4147 cpumask_t *cpumask;
4148
4149 lockdep_assert_held(&wq_pool_mutex);
4150
4151 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4152 wq->unbound_attrs->no_numa)
4153 return;
4154
4155 /*
4156 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4157 * Let's use a preallocated one. The following buf is protected by
4158 * CPU hotplug exclusion.
4159 */
4160 target_attrs = wq_update_unbound_numa_attrs_buf;
4161 cpumask = target_attrs->cpumask;
4162
4163 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4164 pwq = unbound_pwq_by_node(wq, node);
4165
4166 /*
4167 * Let's determine what needs to be done. If the target cpumask is
4168 * different from the default pwq's, we need to compare it to @pwq's
4169 * and create a new one if they don't match. If the target cpumask
4170 * equals the default pwq's, the default pwq should be used.
4171 */
4172 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4173 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4174 return;
4175 } else {
4176 goto use_dfl_pwq;
4177 }
4178
4179 /* create a new pwq */
4180 pwq = alloc_unbound_pwq(wq, target_attrs);
4181 if (!pwq) {
4182 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4183 wq->name);
4184 goto use_dfl_pwq;
4185 }
4186
4187 /* Install the new pwq. */
4188 mutex_lock(&wq->mutex);
4189 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4190 goto out_unlock;
4191
4192use_dfl_pwq:
4193 mutex_lock(&wq->mutex);
4194 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4195 get_pwq(wq->dfl_pwq);
4196 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4197 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4198out_unlock:
4199 mutex_unlock(&wq->mutex);
4200 put_pwq_unlocked(old_pwq);
4201}
4202
4203static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4204{
4205 bool highpri = wq->flags & WQ_HIGHPRI;
4206 int cpu, ret;
4207
4208 if (!(wq->flags & WQ_UNBOUND)) {
4209 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4210 if (!wq->cpu_pwqs)
4211 return -ENOMEM;
4212
4213 for_each_possible_cpu(cpu) {
4214 struct pool_workqueue *pwq =
4215 per_cpu_ptr(wq->cpu_pwqs, cpu);
4216 struct worker_pool *cpu_pools =
4217 per_cpu(cpu_worker_pools, cpu);
4218
4219 init_pwq(pwq, wq, &cpu_pools[highpri]);
4220
4221 mutex_lock(&wq->mutex);
4222 link_pwq(pwq);
4223 mutex_unlock(&wq->mutex);
4224 }
4225 return 0;
4226 }
4227
4228 cpus_read_lock();
4229 if (wq->flags & __WQ_ORDERED) {
4230 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4231 /* there should only be single pwq for ordering guarantee */
4232 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4233 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4234 "ordering guarantee broken for workqueue %s\n", wq->name);
4235 } else {
4236 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4237 }
4238 cpus_read_unlock();
4239
4240 return ret;
4241}
4242
4243static int wq_clamp_max_active(int max_active, unsigned int flags,
4244 const char *name)
4245{
4246 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4247
4248 if (max_active < 1 || max_active > lim)
4249 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4250 max_active, name, 1, lim);
4251
4252 return clamp_val(max_active, 1, lim);
4253}
4254
4255/*
4256 * Workqueues which may be used during memory reclaim should have a rescuer
4257 * to guarantee forward progress.
4258 */
4259static int init_rescuer(struct workqueue_struct *wq)
4260{
4261 struct worker *rescuer;
4262 int ret;
4263
4264 if (!(wq->flags & WQ_MEM_RECLAIM))
4265 return 0;
4266
4267 rescuer = alloc_worker(NUMA_NO_NODE);
4268 if (!rescuer)
4269 return -ENOMEM;
4270
4271 rescuer->rescue_wq = wq;
4272 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4273 if (IS_ERR(rescuer->task)) {
4274 ret = PTR_ERR(rescuer->task);
4275 kfree(rescuer);
4276 return ret;
4277 }
4278
4279 wq->rescuer = rescuer;
4280 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4281 wake_up_process(rescuer->task);
4282
4283 return 0;
4284}
4285
4286__printf(1, 4)
4287struct workqueue_struct *alloc_workqueue(const char *fmt,
4288 unsigned int flags,
4289 int max_active, ...)
4290{
4291 size_t tbl_size = 0;
4292 va_list args;
4293 struct workqueue_struct *wq;
4294 struct pool_workqueue *pwq;
4295
4296 /*
4297 * Unbound && max_active == 1 used to imply ordered, which is no
4298 * longer the case on NUMA machines due to per-node pools. While
4299 * alloc_ordered_workqueue() is the right way to create an ordered
4300 * workqueue, keep the previous behavior to avoid subtle breakages
4301 * on NUMA.
4302 */
4303 if ((flags & WQ_UNBOUND) && max_active == 1)
4304 flags |= __WQ_ORDERED;
4305
4306 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4307 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4308 flags |= WQ_UNBOUND;
4309
4310 /* allocate wq and format name */
4311 if (flags & WQ_UNBOUND)
4312 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4313
4314 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4315 if (!wq)
4316 return NULL;
4317
4318 if (flags & WQ_UNBOUND) {
4319 wq->unbound_attrs = alloc_workqueue_attrs();
4320 if (!wq->unbound_attrs)
4321 goto err_free_wq;
4322 }
4323
4324 va_start(args, max_active);
4325 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4326 va_end(args);
4327
4328 max_active = max_active ?: WQ_DFL_ACTIVE;
4329 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4330
4331 /* init wq */
4332 wq->flags = flags;
4333 wq->saved_max_active = max_active;
4334 mutex_init(&wq->mutex);
4335 atomic_set(&wq->nr_pwqs_to_flush, 0);
4336 INIT_LIST_HEAD(&wq->pwqs);
4337 INIT_LIST_HEAD(&wq->flusher_queue);
4338 INIT_LIST_HEAD(&wq->flusher_overflow);
4339 INIT_LIST_HEAD(&wq->maydays);
4340
4341 wq_init_lockdep(wq);
4342 INIT_LIST_HEAD(&wq->list);
4343
4344 if (alloc_and_link_pwqs(wq) < 0)
4345 goto err_unreg_lockdep;
4346
4347 if (wq_online && init_rescuer(wq) < 0)
4348 goto err_destroy;
4349
4350 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4351 goto err_destroy;
4352
4353 /*
4354 * wq_pool_mutex protects global freeze state and workqueues list.
4355 * Grab it, adjust max_active and add the new @wq to workqueues
4356 * list.
4357 */
4358 mutex_lock(&wq_pool_mutex);
4359
4360 mutex_lock(&wq->mutex);
4361 for_each_pwq(pwq, wq)
4362 pwq_adjust_max_active(pwq);
4363 mutex_unlock(&wq->mutex);
4364
4365 list_add_tail_rcu(&wq->list, &workqueues);
4366
4367 mutex_unlock(&wq_pool_mutex);
4368
4369 return wq;
4370
4371err_unreg_lockdep:
4372 wq_unregister_lockdep(wq);
4373 wq_free_lockdep(wq);
4374err_free_wq:
4375 free_workqueue_attrs(wq->unbound_attrs);
4376 kfree(wq);
4377 return NULL;
4378err_destroy:
4379 destroy_workqueue(wq);
4380 return NULL;
4381}
4382EXPORT_SYMBOL_GPL(alloc_workqueue);
4383
4384static bool pwq_busy(struct pool_workqueue *pwq)
4385{
4386 int i;
4387
4388 for (i = 0; i < WORK_NR_COLORS; i++)
4389 if (pwq->nr_in_flight[i])
4390 return true;
4391
4392 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4393 return true;
4394 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4395 return true;
4396
4397 return false;
4398}
4399
4400/**
4401 * destroy_workqueue - safely terminate a workqueue
4402 * @wq: target workqueue
4403 *
4404 * Safely destroy a workqueue. All work currently pending will be done first.
4405 */
4406void destroy_workqueue(struct workqueue_struct *wq)
4407{
4408 struct pool_workqueue *pwq;
4409 int node;
4410
4411 /*
4412 * Remove it from sysfs first so that sanity check failure doesn't
4413 * lead to sysfs name conflicts.
4414 */
4415 workqueue_sysfs_unregister(wq);
4416
4417 /* drain it before proceeding with destruction */
4418 drain_workqueue(wq);
4419
4420 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4421 if (wq->rescuer) {
4422 struct worker *rescuer = wq->rescuer;
4423
4424 /* this prevents new queueing */
4425 raw_spin_lock_irq(&wq_mayday_lock);
4426 wq->rescuer = NULL;
4427 raw_spin_unlock_irq(&wq_mayday_lock);
4428
4429 /* rescuer will empty maydays list before exiting */
4430 kthread_stop(rescuer->task);
4431 kfree(rescuer);
4432 }
4433
4434 /*
4435 * Sanity checks - grab all the locks so that we wait for all
4436 * in-flight operations which may do put_pwq().
4437 */
4438 mutex_lock(&wq_pool_mutex);
4439 mutex_lock(&wq->mutex);
4440 for_each_pwq(pwq, wq) {
4441 raw_spin_lock_irq(&pwq->pool->lock);
4442 if (WARN_ON(pwq_busy(pwq))) {
4443 pr_warn("%s: %s has the following busy pwq\n",
4444 __func__, wq->name);
4445 show_pwq(pwq);
4446 raw_spin_unlock_irq(&pwq->pool->lock);
4447 mutex_unlock(&wq->mutex);
4448 mutex_unlock(&wq_pool_mutex);
4449 show_one_workqueue(wq);
4450 return;
4451 }
4452 raw_spin_unlock_irq(&pwq->pool->lock);
4453 }
4454 mutex_unlock(&wq->mutex);
4455
4456 /*
4457 * wq list is used to freeze wq, remove from list after
4458 * flushing is complete in case freeze races us.
4459 */
4460 list_del_rcu(&wq->list);
4461 mutex_unlock(&wq_pool_mutex);
4462
4463 if (!(wq->flags & WQ_UNBOUND)) {
4464 wq_unregister_lockdep(wq);
4465 /*
4466 * The base ref is never dropped on per-cpu pwqs. Directly
4467 * schedule RCU free.
4468 */
4469 call_rcu(&wq->rcu, rcu_free_wq);
4470 } else {
4471 /*
4472 * We're the sole accessor of @wq at this point. Directly
4473 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4474 * @wq will be freed when the last pwq is released.
4475 */
4476 for_each_node(node) {
4477 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4478 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4479 put_pwq_unlocked(pwq);
4480 }
4481
4482 /*
4483 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4484 * put. Don't access it afterwards.
4485 */
4486 pwq = wq->dfl_pwq;
4487 wq->dfl_pwq = NULL;
4488 put_pwq_unlocked(pwq);
4489 }
4490}
4491EXPORT_SYMBOL_GPL(destroy_workqueue);
4492
4493/**
4494 * workqueue_set_max_active - adjust max_active of a workqueue
4495 * @wq: target workqueue
4496 * @max_active: new max_active value.
4497 *
4498 * Set max_active of @wq to @max_active.
4499 *
4500 * CONTEXT:
4501 * Don't call from IRQ context.
4502 */
4503void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4504{
4505 struct pool_workqueue *pwq;
4506
4507 /* disallow meddling with max_active for ordered workqueues */
4508 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4509 return;
4510
4511 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4512
4513 mutex_lock(&wq->mutex);
4514
4515 wq->flags &= ~__WQ_ORDERED;
4516 wq->saved_max_active = max_active;
4517
4518 for_each_pwq(pwq, wq)
4519 pwq_adjust_max_active(pwq);
4520
4521 mutex_unlock(&wq->mutex);
4522}
4523EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4524
4525/**
4526 * current_work - retrieve %current task's work struct
4527 *
4528 * Determine if %current task is a workqueue worker and what it's working on.
4529 * Useful to find out the context that the %current task is running in.
4530 *
4531 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4532 */
4533struct work_struct *current_work(void)
4534{
4535 struct worker *worker = current_wq_worker();
4536
4537 return worker ? worker->current_work : NULL;
4538}
4539EXPORT_SYMBOL(current_work);
4540
4541/**
4542 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4543 *
4544 * Determine whether %current is a workqueue rescuer. Can be used from
4545 * work functions to determine whether it's being run off the rescuer task.
4546 *
4547 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4548 */
4549bool current_is_workqueue_rescuer(void)
4550{
4551 struct worker *worker = current_wq_worker();
4552
4553 return worker && worker->rescue_wq;
4554}
4555
4556/**
4557 * workqueue_congested - test whether a workqueue is congested
4558 * @cpu: CPU in question
4559 * @wq: target workqueue
4560 *
4561 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4562 * no synchronization around this function and the test result is
4563 * unreliable and only useful as advisory hints or for debugging.
4564 *
4565 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4566 * Note that both per-cpu and unbound workqueues may be associated with
4567 * multiple pool_workqueues which have separate congested states. A
4568 * workqueue being congested on one CPU doesn't mean the workqueue is also
4569 * contested on other CPUs / NUMA nodes.
4570 *
4571 * Return:
4572 * %true if congested, %false otherwise.
4573 */
4574bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4575{
4576 struct pool_workqueue *pwq;
4577 bool ret;
4578
4579 rcu_read_lock();
4580 preempt_disable();
4581
4582 if (cpu == WORK_CPU_UNBOUND)
4583 cpu = smp_processor_id();
4584
4585 if (!(wq->flags & WQ_UNBOUND))
4586 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4587 else
4588 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4589
4590 ret = !list_empty(&pwq->inactive_works);
4591 preempt_enable();
4592 rcu_read_unlock();
4593
4594 return ret;
4595}
4596EXPORT_SYMBOL_GPL(workqueue_congested);
4597
4598/**
4599 * work_busy - test whether a work is currently pending or running
4600 * @work: the work to be tested
4601 *
4602 * Test whether @work is currently pending or running. There is no
4603 * synchronization around this function and the test result is
4604 * unreliable and only useful as advisory hints or for debugging.
4605 *
4606 * Return:
4607 * OR'd bitmask of WORK_BUSY_* bits.
4608 */
4609unsigned int work_busy(struct work_struct *work)
4610{
4611 struct worker_pool *pool;
4612 unsigned long flags;
4613 unsigned int ret = 0;
4614
4615 if (work_pending(work))
4616 ret |= WORK_BUSY_PENDING;
4617
4618 rcu_read_lock();
4619 pool = get_work_pool(work);
4620 if (pool) {
4621 raw_spin_lock_irqsave(&pool->lock, flags);
4622 if (find_worker_executing_work(pool, work))
4623 ret |= WORK_BUSY_RUNNING;
4624 raw_spin_unlock_irqrestore(&pool->lock, flags);
4625 }
4626 rcu_read_unlock();
4627
4628 return ret;
4629}
4630EXPORT_SYMBOL_GPL(work_busy);
4631
4632/**
4633 * set_worker_desc - set description for the current work item
4634 * @fmt: printf-style format string
4635 * @...: arguments for the format string
4636 *
4637 * This function can be called by a running work function to describe what
4638 * the work item is about. If the worker task gets dumped, this
4639 * information will be printed out together to help debugging. The
4640 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4641 */
4642void set_worker_desc(const char *fmt, ...)
4643{
4644 struct worker *worker = current_wq_worker();
4645 va_list args;
4646
4647 if (worker) {
4648 va_start(args, fmt);
4649 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4650 va_end(args);
4651 }
4652}
4653EXPORT_SYMBOL_GPL(set_worker_desc);
4654
4655/**
4656 * print_worker_info - print out worker information and description
4657 * @log_lvl: the log level to use when printing
4658 * @task: target task
4659 *
4660 * If @task is a worker and currently executing a work item, print out the
4661 * name of the workqueue being serviced and worker description set with
4662 * set_worker_desc() by the currently executing work item.
4663 *
4664 * This function can be safely called on any task as long as the
4665 * task_struct itself is accessible. While safe, this function isn't
4666 * synchronized and may print out mixups or garbages of limited length.
4667 */
4668void print_worker_info(const char *log_lvl, struct task_struct *task)
4669{
4670 work_func_t *fn = NULL;
4671 char name[WQ_NAME_LEN] = { };
4672 char desc[WORKER_DESC_LEN] = { };
4673 struct pool_workqueue *pwq = NULL;
4674 struct workqueue_struct *wq = NULL;
4675 struct worker *worker;
4676
4677 if (!(task->flags & PF_WQ_WORKER))
4678 return;
4679
4680 /*
4681 * This function is called without any synchronization and @task
4682 * could be in any state. Be careful with dereferences.
4683 */
4684 worker = kthread_probe_data(task);
4685
4686 /*
4687 * Carefully copy the associated workqueue's workfn, name and desc.
4688 * Keep the original last '\0' in case the original is garbage.
4689 */
4690 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
4691 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
4692 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
4693 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
4694 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
4695
4696 if (fn || name[0] || desc[0]) {
4697 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4698 if (strcmp(name, desc))
4699 pr_cont(" (%s)", desc);
4700 pr_cont("\n");
4701 }
4702}
4703
4704static void pr_cont_pool_info(struct worker_pool *pool)
4705{
4706 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4707 if (pool->node != NUMA_NO_NODE)
4708 pr_cont(" node=%d", pool->node);
4709 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4710}
4711
4712static void pr_cont_work(bool comma, struct work_struct *work)
4713{
4714 if (work->func == wq_barrier_func) {
4715 struct wq_barrier *barr;
4716
4717 barr = container_of(work, struct wq_barrier, work);
4718
4719 pr_cont("%s BAR(%d)", comma ? "," : "",
4720 task_pid_nr(barr->task));
4721 } else {
4722 pr_cont("%s %ps", comma ? "," : "", work->func);
4723 }
4724}
4725
4726static void show_pwq(struct pool_workqueue *pwq)
4727{
4728 struct worker_pool *pool = pwq->pool;
4729 struct work_struct *work;
4730 struct worker *worker;
4731 bool has_in_flight = false, has_pending = false;
4732 int bkt;
4733
4734 pr_info(" pwq %d:", pool->id);
4735 pr_cont_pool_info(pool);
4736
4737 pr_cont(" active=%d/%d refcnt=%d%s\n",
4738 pwq->nr_active, pwq->max_active, pwq->refcnt,
4739 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4740
4741 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4742 if (worker->current_pwq == pwq) {
4743 has_in_flight = true;
4744 break;
4745 }
4746 }
4747 if (has_in_flight) {
4748 bool comma = false;
4749
4750 pr_info(" in-flight:");
4751 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4752 if (worker->current_pwq != pwq)
4753 continue;
4754
4755 pr_cont("%s %d%s:%ps", comma ? "," : "",
4756 task_pid_nr(worker->task),
4757 worker->rescue_wq ? "(RESCUER)" : "",
4758 worker->current_func);
4759 list_for_each_entry(work, &worker->scheduled, entry)
4760 pr_cont_work(false, work);
4761 comma = true;
4762 }
4763 pr_cont("\n");
4764 }
4765
4766 list_for_each_entry(work, &pool->worklist, entry) {
4767 if (get_work_pwq(work) == pwq) {
4768 has_pending = true;
4769 break;
4770 }
4771 }
4772 if (has_pending) {
4773 bool comma = false;
4774
4775 pr_info(" pending:");
4776 list_for_each_entry(work, &pool->worklist, entry) {
4777 if (get_work_pwq(work) != pwq)
4778 continue;
4779
4780 pr_cont_work(comma, work);
4781 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4782 }
4783 pr_cont("\n");
4784 }
4785
4786 if (!list_empty(&pwq->inactive_works)) {
4787 bool comma = false;
4788
4789 pr_info(" inactive:");
4790 list_for_each_entry(work, &pwq->inactive_works, entry) {
4791 pr_cont_work(comma, work);
4792 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4793 }
4794 pr_cont("\n");
4795 }
4796}
4797
4798/**
4799 * show_one_workqueue - dump state of specified workqueue
4800 * @wq: workqueue whose state will be printed
4801 */
4802void show_one_workqueue(struct workqueue_struct *wq)
4803{
4804 struct pool_workqueue *pwq;
4805 bool idle = true;
4806 unsigned long flags;
4807
4808 for_each_pwq(pwq, wq) {
4809 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4810 idle = false;
4811 break;
4812 }
4813 }
4814 if (idle) /* Nothing to print for idle workqueue */
4815 return;
4816
4817 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4818
4819 for_each_pwq(pwq, wq) {
4820 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4821 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
4822 /*
4823 * Defer printing to avoid deadlocks in console
4824 * drivers that queue work while holding locks
4825 * also taken in their write paths.
4826 */
4827 printk_deferred_enter();
4828 show_pwq(pwq);
4829 printk_deferred_exit();
4830 }
4831 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4832 /*
4833 * We could be printing a lot from atomic context, e.g.
4834 * sysrq-t -> show_all_workqueues(). Avoid triggering
4835 * hard lockup.
4836 */
4837 touch_nmi_watchdog();
4838 }
4839
4840}
4841
4842/**
4843 * show_one_worker_pool - dump state of specified worker pool
4844 * @pool: worker pool whose state will be printed
4845 */
4846static void show_one_worker_pool(struct worker_pool *pool)
4847{
4848 struct worker *worker;
4849 bool first = true;
4850 unsigned long flags;
4851
4852 raw_spin_lock_irqsave(&pool->lock, flags);
4853 if (pool->nr_workers == pool->nr_idle)
4854 goto next_pool;
4855 /*
4856 * Defer printing to avoid deadlocks in console drivers that
4857 * queue work while holding locks also taken in their write
4858 * paths.
4859 */
4860 printk_deferred_enter();
4861 pr_info("pool %d:", pool->id);
4862 pr_cont_pool_info(pool);
4863 pr_cont(" hung=%us workers=%d",
4864 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4865 pool->nr_workers);
4866 if (pool->manager)
4867 pr_cont(" manager: %d",
4868 task_pid_nr(pool->manager->task));
4869 list_for_each_entry(worker, &pool->idle_list, entry) {
4870 pr_cont(" %s%d", first ? "idle: " : "",
4871 task_pid_nr(worker->task));
4872 first = false;
4873 }
4874 pr_cont("\n");
4875 printk_deferred_exit();
4876next_pool:
4877 raw_spin_unlock_irqrestore(&pool->lock, flags);
4878 /*
4879 * We could be printing a lot from atomic context, e.g.
4880 * sysrq-t -> show_all_workqueues(). Avoid triggering
4881 * hard lockup.
4882 */
4883 touch_nmi_watchdog();
4884
4885}
4886
4887/**
4888 * show_all_workqueues - dump workqueue state
4889 *
4890 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4891 * all busy workqueues and pools.
4892 */
4893void show_all_workqueues(void)
4894{
4895 struct workqueue_struct *wq;
4896 struct worker_pool *pool;
4897 int pi;
4898
4899 rcu_read_lock();
4900
4901 pr_info("Showing busy workqueues and worker pools:\n");
4902
4903 list_for_each_entry_rcu(wq, &workqueues, list)
4904 show_one_workqueue(wq);
4905
4906 for_each_pool(pool, pi)
4907 show_one_worker_pool(pool);
4908
4909 rcu_read_unlock();
4910}
4911
4912/* used to show worker information through /proc/PID/{comm,stat,status} */
4913void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4914{
4915 int off;
4916
4917 /* always show the actual comm */
4918 off = strscpy(buf, task->comm, size);
4919 if (off < 0)
4920 return;
4921
4922 /* stabilize PF_WQ_WORKER and worker pool association */
4923 mutex_lock(&wq_pool_attach_mutex);
4924
4925 if (task->flags & PF_WQ_WORKER) {
4926 struct worker *worker = kthread_data(task);
4927 struct worker_pool *pool = worker->pool;
4928
4929 if (pool) {
4930 raw_spin_lock_irq(&pool->lock);
4931 /*
4932 * ->desc tracks information (wq name or
4933 * set_worker_desc()) for the latest execution. If
4934 * current, prepend '+', otherwise '-'.
4935 */
4936 if (worker->desc[0] != '\0') {
4937 if (worker->current_work)
4938 scnprintf(buf + off, size - off, "+%s",
4939 worker->desc);
4940 else
4941 scnprintf(buf + off, size - off, "-%s",
4942 worker->desc);
4943 }
4944 raw_spin_unlock_irq(&pool->lock);
4945 }
4946 }
4947
4948 mutex_unlock(&wq_pool_attach_mutex);
4949}
4950
4951#ifdef CONFIG_SMP
4952
4953/*
4954 * CPU hotplug.
4955 *
4956 * There are two challenges in supporting CPU hotplug. Firstly, there
4957 * are a lot of assumptions on strong associations among work, pwq and
4958 * pool which make migrating pending and scheduled works very
4959 * difficult to implement without impacting hot paths. Secondly,
4960 * worker pools serve mix of short, long and very long running works making
4961 * blocked draining impractical.
4962 *
4963 * This is solved by allowing the pools to be disassociated from the CPU
4964 * running as an unbound one and allowing it to be reattached later if the
4965 * cpu comes back online.
4966 */
4967
4968static void unbind_workers(int cpu)
4969{
4970 struct worker_pool *pool;
4971 struct worker *worker;
4972
4973 for_each_cpu_worker_pool(pool, cpu) {
4974 mutex_lock(&wq_pool_attach_mutex);
4975 raw_spin_lock_irq(&pool->lock);
4976
4977 /*
4978 * We've blocked all attach/detach operations. Make all workers
4979 * unbound and set DISASSOCIATED. Before this, all workers
4980 * must be on the cpu. After this, they may become diasporas.
4981 * And the preemption disabled section in their sched callbacks
4982 * are guaranteed to see WORKER_UNBOUND since the code here
4983 * is on the same cpu.
4984 */
4985 for_each_pool_worker(worker, pool)
4986 worker->flags |= WORKER_UNBOUND;
4987
4988 pool->flags |= POOL_DISASSOCIATED;
4989
4990 /*
4991 * The handling of nr_running in sched callbacks are disabled
4992 * now. Zap nr_running. After this, nr_running stays zero and
4993 * need_more_worker() and keep_working() are always true as
4994 * long as the worklist is not empty. This pool now behaves as
4995 * an unbound (in terms of concurrency management) pool which
4996 * are served by workers tied to the pool.
4997 */
4998 pool->nr_running = 0;
4999
5000 /*
5001 * With concurrency management just turned off, a busy
5002 * worker blocking could lead to lengthy stalls. Kick off
5003 * unbound chain execution of currently pending work items.
5004 */
5005 wake_up_worker(pool);
5006
5007 raw_spin_unlock_irq(&pool->lock);
5008
5009 for_each_pool_worker(worker, pool) {
5010 kthread_set_per_cpu(worker->task, -1);
5011 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
5012 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
5013 else
5014 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
5015 }
5016
5017 mutex_unlock(&wq_pool_attach_mutex);
5018 }
5019}
5020
5021/**
5022 * rebind_workers - rebind all workers of a pool to the associated CPU
5023 * @pool: pool of interest
5024 *
5025 * @pool->cpu is coming online. Rebind all workers to the CPU.
5026 */
5027static void rebind_workers(struct worker_pool *pool)
5028{
5029 struct worker *worker;
5030
5031 lockdep_assert_held(&wq_pool_attach_mutex);
5032
5033 /*
5034 * Restore CPU affinity of all workers. As all idle workers should
5035 * be on the run-queue of the associated CPU before any local
5036 * wake-ups for concurrency management happen, restore CPU affinity
5037 * of all workers first and then clear UNBOUND. As we're called
5038 * from CPU_ONLINE, the following shouldn't fail.
5039 */
5040 for_each_pool_worker(worker, pool) {
5041 kthread_set_per_cpu(worker->task, pool->cpu);
5042 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5043 pool->attrs->cpumask) < 0);
5044 }
5045
5046 raw_spin_lock_irq(&pool->lock);
5047
5048 pool->flags &= ~POOL_DISASSOCIATED;
5049
5050 for_each_pool_worker(worker, pool) {
5051 unsigned int worker_flags = worker->flags;
5052
5053 /*
5054 * We want to clear UNBOUND but can't directly call
5055 * worker_clr_flags() or adjust nr_running. Atomically
5056 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5057 * @worker will clear REBOUND using worker_clr_flags() when
5058 * it initiates the next execution cycle thus restoring
5059 * concurrency management. Note that when or whether
5060 * @worker clears REBOUND doesn't affect correctness.
5061 *
5062 * WRITE_ONCE() is necessary because @worker->flags may be
5063 * tested without holding any lock in
5064 * wq_worker_running(). Without it, NOT_RUNNING test may
5065 * fail incorrectly leading to premature concurrency
5066 * management operations.
5067 */
5068 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5069 worker_flags |= WORKER_REBOUND;
5070 worker_flags &= ~WORKER_UNBOUND;
5071 WRITE_ONCE(worker->flags, worker_flags);
5072 }
5073
5074 raw_spin_unlock_irq(&pool->lock);
5075}
5076
5077/**
5078 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5079 * @pool: unbound pool of interest
5080 * @cpu: the CPU which is coming up
5081 *
5082 * An unbound pool may end up with a cpumask which doesn't have any online
5083 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5084 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5085 * online CPU before, cpus_allowed of all its workers should be restored.
5086 */
5087static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5088{
5089 static cpumask_t cpumask;
5090 struct worker *worker;
5091
5092 lockdep_assert_held(&wq_pool_attach_mutex);
5093
5094 /* is @cpu allowed for @pool? */
5095 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5096 return;
5097
5098 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5099
5100 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5101 for_each_pool_worker(worker, pool)
5102 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5103}
5104
5105int workqueue_prepare_cpu(unsigned int cpu)
5106{
5107 struct worker_pool *pool;
5108
5109 for_each_cpu_worker_pool(pool, cpu) {
5110 if (pool->nr_workers)
5111 continue;
5112 if (!create_worker(pool))
5113 return -ENOMEM;
5114 }
5115 return 0;
5116}
5117
5118int workqueue_online_cpu(unsigned int cpu)
5119{
5120 struct worker_pool *pool;
5121 struct workqueue_struct *wq;
5122 int pi;
5123
5124 mutex_lock(&wq_pool_mutex);
5125
5126 for_each_pool(pool, pi) {
5127 mutex_lock(&wq_pool_attach_mutex);
5128
5129 if (pool->cpu == cpu)
5130 rebind_workers(pool);
5131 else if (pool->cpu < 0)
5132 restore_unbound_workers_cpumask(pool, cpu);
5133
5134 mutex_unlock(&wq_pool_attach_mutex);
5135 }
5136
5137 /* update NUMA affinity of unbound workqueues */
5138 list_for_each_entry(wq, &workqueues, list)
5139 wq_update_unbound_numa(wq, cpu, true);
5140
5141 mutex_unlock(&wq_pool_mutex);
5142 return 0;
5143}
5144
5145int workqueue_offline_cpu(unsigned int cpu)
5146{
5147 struct workqueue_struct *wq;
5148
5149 /* unbinding per-cpu workers should happen on the local CPU */
5150 if (WARN_ON(cpu != smp_processor_id()))
5151 return -1;
5152
5153 unbind_workers(cpu);
5154
5155 /* update NUMA affinity of unbound workqueues */
5156 mutex_lock(&wq_pool_mutex);
5157 list_for_each_entry(wq, &workqueues, list)
5158 wq_update_unbound_numa(wq, cpu, false);
5159 mutex_unlock(&wq_pool_mutex);
5160
5161 return 0;
5162}
5163
5164struct work_for_cpu {
5165 struct work_struct work;
5166 long (*fn)(void *);
5167 void *arg;
5168 long ret;
5169};
5170
5171static void work_for_cpu_fn(struct work_struct *work)
5172{
5173 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5174
5175 wfc->ret = wfc->fn(wfc->arg);
5176}
5177
5178/**
5179 * work_on_cpu - run a function in thread context on a particular cpu
5180 * @cpu: the cpu to run on
5181 * @fn: the function to run
5182 * @arg: the function arg
5183 *
5184 * It is up to the caller to ensure that the cpu doesn't go offline.
5185 * The caller must not hold any locks which would prevent @fn from completing.
5186 *
5187 * Return: The value @fn returns.
5188 */
5189long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5190{
5191 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5192
5193 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5194 schedule_work_on(cpu, &wfc.work);
5195 flush_work(&wfc.work);
5196 destroy_work_on_stack(&wfc.work);
5197 return wfc.ret;
5198}
5199EXPORT_SYMBOL_GPL(work_on_cpu);
5200
5201/**
5202 * work_on_cpu_safe - run a function in thread context on a particular cpu
5203 * @cpu: the cpu to run on
5204 * @fn: the function to run
5205 * @arg: the function argument
5206 *
5207 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5208 * any locks which would prevent @fn from completing.
5209 *
5210 * Return: The value @fn returns.
5211 */
5212long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5213{
5214 long ret = -ENODEV;
5215
5216 cpus_read_lock();
5217 if (cpu_online(cpu))
5218 ret = work_on_cpu(cpu, fn, arg);
5219 cpus_read_unlock();
5220 return ret;
5221}
5222EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5223#endif /* CONFIG_SMP */
5224
5225#ifdef CONFIG_FREEZER
5226
5227/**
5228 * freeze_workqueues_begin - begin freezing workqueues
5229 *
5230 * Start freezing workqueues. After this function returns, all freezable
5231 * workqueues will queue new works to their inactive_works list instead of
5232 * pool->worklist.
5233 *
5234 * CONTEXT:
5235 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5236 */
5237void freeze_workqueues_begin(void)
5238{
5239 struct workqueue_struct *wq;
5240 struct pool_workqueue *pwq;
5241
5242 mutex_lock(&wq_pool_mutex);
5243
5244 WARN_ON_ONCE(workqueue_freezing);
5245 workqueue_freezing = true;
5246
5247 list_for_each_entry(wq, &workqueues, list) {
5248 mutex_lock(&wq->mutex);
5249 for_each_pwq(pwq, wq)
5250 pwq_adjust_max_active(pwq);
5251 mutex_unlock(&wq->mutex);
5252 }
5253
5254 mutex_unlock(&wq_pool_mutex);
5255}
5256
5257/**
5258 * freeze_workqueues_busy - are freezable workqueues still busy?
5259 *
5260 * Check whether freezing is complete. This function must be called
5261 * between freeze_workqueues_begin() and thaw_workqueues().
5262 *
5263 * CONTEXT:
5264 * Grabs and releases wq_pool_mutex.
5265 *
5266 * Return:
5267 * %true if some freezable workqueues are still busy. %false if freezing
5268 * is complete.
5269 */
5270bool freeze_workqueues_busy(void)
5271{
5272 bool busy = false;
5273 struct workqueue_struct *wq;
5274 struct pool_workqueue *pwq;
5275
5276 mutex_lock(&wq_pool_mutex);
5277
5278 WARN_ON_ONCE(!workqueue_freezing);
5279
5280 list_for_each_entry(wq, &workqueues, list) {
5281 if (!(wq->flags & WQ_FREEZABLE))
5282 continue;
5283 /*
5284 * nr_active is monotonically decreasing. It's safe
5285 * to peek without lock.
5286 */
5287 rcu_read_lock();
5288 for_each_pwq(pwq, wq) {
5289 WARN_ON_ONCE(pwq->nr_active < 0);
5290 if (pwq->nr_active) {
5291 busy = true;
5292 rcu_read_unlock();
5293 goto out_unlock;
5294 }
5295 }
5296 rcu_read_unlock();
5297 }
5298out_unlock:
5299 mutex_unlock(&wq_pool_mutex);
5300 return busy;
5301}
5302
5303/**
5304 * thaw_workqueues - thaw workqueues
5305 *
5306 * Thaw workqueues. Normal queueing is restored and all collected
5307 * frozen works are transferred to their respective pool worklists.
5308 *
5309 * CONTEXT:
5310 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5311 */
5312void thaw_workqueues(void)
5313{
5314 struct workqueue_struct *wq;
5315 struct pool_workqueue *pwq;
5316
5317 mutex_lock(&wq_pool_mutex);
5318
5319 if (!workqueue_freezing)
5320 goto out_unlock;
5321
5322 workqueue_freezing = false;
5323
5324 /* restore max_active and repopulate worklist */
5325 list_for_each_entry(wq, &workqueues, list) {
5326 mutex_lock(&wq->mutex);
5327 for_each_pwq(pwq, wq)
5328 pwq_adjust_max_active(pwq);
5329 mutex_unlock(&wq->mutex);
5330 }
5331
5332out_unlock:
5333 mutex_unlock(&wq_pool_mutex);
5334}
5335#endif /* CONFIG_FREEZER */
5336
5337static int workqueue_apply_unbound_cpumask(void)
5338{
5339 LIST_HEAD(ctxs);
5340 int ret = 0;
5341 struct workqueue_struct *wq;
5342 struct apply_wqattrs_ctx *ctx, *n;
5343
5344 lockdep_assert_held(&wq_pool_mutex);
5345
5346 list_for_each_entry(wq, &workqueues, list) {
5347 if (!(wq->flags & WQ_UNBOUND))
5348 continue;
5349 /* creating multiple pwqs breaks ordering guarantee */
5350 if (wq->flags & __WQ_ORDERED)
5351 continue;
5352
5353 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5354 if (!ctx) {
5355 ret = -ENOMEM;
5356 break;
5357 }
5358
5359 list_add_tail(&ctx->list, &ctxs);
5360 }
5361
5362 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5363 if (!ret)
5364 apply_wqattrs_commit(ctx);
5365 apply_wqattrs_cleanup(ctx);
5366 }
5367
5368 return ret;
5369}
5370
5371/**
5372 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5373 * @cpumask: the cpumask to set
5374 *
5375 * The low-level workqueues cpumask is a global cpumask that limits
5376 * the affinity of all unbound workqueues. This function check the @cpumask
5377 * and apply it to all unbound workqueues and updates all pwqs of them.
5378 *
5379 * Return: 0 - Success
5380 * -EINVAL - Invalid @cpumask
5381 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
5382 */
5383int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5384{
5385 int ret = -EINVAL;
5386 cpumask_var_t saved_cpumask;
5387
5388 /*
5389 * Not excluding isolated cpus on purpose.
5390 * If the user wishes to include them, we allow that.
5391 */
5392 cpumask_and(cpumask, cpumask, cpu_possible_mask);
5393 if (!cpumask_empty(cpumask)) {
5394 apply_wqattrs_lock();
5395 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
5396 ret = 0;
5397 goto out_unlock;
5398 }
5399
5400 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL)) {
5401 ret = -ENOMEM;
5402 goto out_unlock;
5403 }
5404
5405 /* save the old wq_unbound_cpumask. */
5406 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5407
5408 /* update wq_unbound_cpumask at first and apply it to wqs. */
5409 cpumask_copy(wq_unbound_cpumask, cpumask);
5410 ret = workqueue_apply_unbound_cpumask();
5411
5412 /* restore the wq_unbound_cpumask when failed. */
5413 if (ret < 0)
5414 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5415
5416 free_cpumask_var(saved_cpumask);
5417out_unlock:
5418 apply_wqattrs_unlock();
5419 }
5420
5421 return ret;
5422}
5423
5424#ifdef CONFIG_SYSFS
5425/*
5426 * Workqueues with WQ_SYSFS flag set is visible to userland via
5427 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5428 * following attributes.
5429 *
5430 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5431 * max_active RW int : maximum number of in-flight work items
5432 *
5433 * Unbound workqueues have the following extra attributes.
5434 *
5435 * pool_ids RO int : the associated pool IDs for each node
5436 * nice RW int : nice value of the workers
5437 * cpumask RW mask : bitmask of allowed CPUs for the workers
5438 * numa RW bool : whether enable NUMA affinity
5439 */
5440struct wq_device {
5441 struct workqueue_struct *wq;
5442 struct device dev;
5443};
5444
5445static struct workqueue_struct *dev_to_wq(struct device *dev)
5446{
5447 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5448
5449 return wq_dev->wq;
5450}
5451
5452static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5453 char *buf)
5454{
5455 struct workqueue_struct *wq = dev_to_wq(dev);
5456
5457 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5458}
5459static DEVICE_ATTR_RO(per_cpu);
5460
5461static ssize_t max_active_show(struct device *dev,
5462 struct device_attribute *attr, char *buf)
5463{
5464 struct workqueue_struct *wq = dev_to_wq(dev);
5465
5466 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5467}
5468
5469static ssize_t max_active_store(struct device *dev,
5470 struct device_attribute *attr, const char *buf,
5471 size_t count)
5472{
5473 struct workqueue_struct *wq = dev_to_wq(dev);
5474 int val;
5475
5476 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5477 return -EINVAL;
5478
5479 workqueue_set_max_active(wq, val);
5480 return count;
5481}
5482static DEVICE_ATTR_RW(max_active);
5483
5484static struct attribute *wq_sysfs_attrs[] = {
5485 &dev_attr_per_cpu.attr,
5486 &dev_attr_max_active.attr,
5487 NULL,
5488};
5489ATTRIBUTE_GROUPS(wq_sysfs);
5490
5491static ssize_t wq_pool_ids_show(struct device *dev,
5492 struct device_attribute *attr, char *buf)
5493{
5494 struct workqueue_struct *wq = dev_to_wq(dev);
5495 const char *delim = "";
5496 int node, written = 0;
5497
5498 cpus_read_lock();
5499 rcu_read_lock();
5500 for_each_node(node) {
5501 written += scnprintf(buf + written, PAGE_SIZE - written,
5502 "%s%d:%d", delim, node,
5503 unbound_pwq_by_node(wq, node)->pool->id);
5504 delim = " ";
5505 }
5506 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5507 rcu_read_unlock();
5508 cpus_read_unlock();
5509
5510 return written;
5511}
5512
5513static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5514 char *buf)
5515{
5516 struct workqueue_struct *wq = dev_to_wq(dev);
5517 int written;
5518
5519 mutex_lock(&wq->mutex);
5520 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5521 mutex_unlock(&wq->mutex);
5522
5523 return written;
5524}
5525
5526/* prepare workqueue_attrs for sysfs store operations */
5527static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5528{
5529 struct workqueue_attrs *attrs;
5530
5531 lockdep_assert_held(&wq_pool_mutex);
5532
5533 attrs = alloc_workqueue_attrs();
5534 if (!attrs)
5535 return NULL;
5536
5537 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5538 return attrs;
5539}
5540
5541static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5542 const char *buf, size_t count)
5543{
5544 struct workqueue_struct *wq = dev_to_wq(dev);
5545 struct workqueue_attrs *attrs;
5546 int ret = -ENOMEM;
5547
5548 apply_wqattrs_lock();
5549
5550 attrs = wq_sysfs_prep_attrs(wq);
5551 if (!attrs)
5552 goto out_unlock;
5553
5554 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5555 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5556 ret = apply_workqueue_attrs_locked(wq, attrs);
5557 else
5558 ret = -EINVAL;
5559
5560out_unlock:
5561 apply_wqattrs_unlock();
5562 free_workqueue_attrs(attrs);
5563 return ret ?: count;
5564}
5565
5566static ssize_t wq_cpumask_show(struct device *dev,
5567 struct device_attribute *attr, char *buf)
5568{
5569 struct workqueue_struct *wq = dev_to_wq(dev);
5570 int written;
5571
5572 mutex_lock(&wq->mutex);
5573 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5574 cpumask_pr_args(wq->unbound_attrs->cpumask));
5575 mutex_unlock(&wq->mutex);
5576 return written;
5577}
5578
5579static ssize_t wq_cpumask_store(struct device *dev,
5580 struct device_attribute *attr,
5581 const char *buf, size_t count)
5582{
5583 struct workqueue_struct *wq = dev_to_wq(dev);
5584 struct workqueue_attrs *attrs;
5585 int ret = -ENOMEM;
5586
5587 apply_wqattrs_lock();
5588
5589 attrs = wq_sysfs_prep_attrs(wq);
5590 if (!attrs)
5591 goto out_unlock;
5592
5593 ret = cpumask_parse(buf, attrs->cpumask);
5594 if (!ret)
5595 ret = apply_workqueue_attrs_locked(wq, attrs);
5596
5597out_unlock:
5598 apply_wqattrs_unlock();
5599 free_workqueue_attrs(attrs);
5600 return ret ?: count;
5601}
5602
5603static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5604 char *buf)
5605{
5606 struct workqueue_struct *wq = dev_to_wq(dev);
5607 int written;
5608
5609 mutex_lock(&wq->mutex);
5610 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5611 !wq->unbound_attrs->no_numa);
5612 mutex_unlock(&wq->mutex);
5613
5614 return written;
5615}
5616
5617static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5618 const char *buf, size_t count)
5619{
5620 struct workqueue_struct *wq = dev_to_wq(dev);
5621 struct workqueue_attrs *attrs;
5622 int v, ret = -ENOMEM;
5623
5624 apply_wqattrs_lock();
5625
5626 attrs = wq_sysfs_prep_attrs(wq);
5627 if (!attrs)
5628 goto out_unlock;
5629
5630 ret = -EINVAL;
5631 if (sscanf(buf, "%d", &v) == 1) {
5632 attrs->no_numa = !v;
5633 ret = apply_workqueue_attrs_locked(wq, attrs);
5634 }
5635
5636out_unlock:
5637 apply_wqattrs_unlock();
5638 free_workqueue_attrs(attrs);
5639 return ret ?: count;
5640}
5641
5642static struct device_attribute wq_sysfs_unbound_attrs[] = {
5643 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5644 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5645 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5646 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5647 __ATTR_NULL,
5648};
5649
5650static struct bus_type wq_subsys = {
5651 .name = "workqueue",
5652 .dev_groups = wq_sysfs_groups,
5653};
5654
5655static ssize_t wq_unbound_cpumask_show(struct device *dev,
5656 struct device_attribute *attr, char *buf)
5657{
5658 int written;
5659
5660 mutex_lock(&wq_pool_mutex);
5661 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5662 cpumask_pr_args(wq_unbound_cpumask));
5663 mutex_unlock(&wq_pool_mutex);
5664
5665 return written;
5666}
5667
5668static ssize_t wq_unbound_cpumask_store(struct device *dev,
5669 struct device_attribute *attr, const char *buf, size_t count)
5670{
5671 cpumask_var_t cpumask;
5672 int ret;
5673
5674 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5675 return -ENOMEM;
5676
5677 ret = cpumask_parse(buf, cpumask);
5678 if (!ret)
5679 ret = workqueue_set_unbound_cpumask(cpumask);
5680
5681 free_cpumask_var(cpumask);
5682 return ret ? ret : count;
5683}
5684
5685static struct device_attribute wq_sysfs_cpumask_attr =
5686 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5687 wq_unbound_cpumask_store);
5688
5689static int __init wq_sysfs_init(void)
5690{
5691 int err;
5692
5693 err = subsys_virtual_register(&wq_subsys, NULL);
5694 if (err)
5695 return err;
5696
5697 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5698}
5699core_initcall(wq_sysfs_init);
5700
5701static void wq_device_release(struct device *dev)
5702{
5703 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5704
5705 kfree(wq_dev);
5706}
5707
5708/**
5709 * workqueue_sysfs_register - make a workqueue visible in sysfs
5710 * @wq: the workqueue to register
5711 *
5712 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5713 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5714 * which is the preferred method.
5715 *
5716 * Workqueue user should use this function directly iff it wants to apply
5717 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5718 * apply_workqueue_attrs() may race against userland updating the
5719 * attributes.
5720 *
5721 * Return: 0 on success, -errno on failure.
5722 */
5723int workqueue_sysfs_register(struct workqueue_struct *wq)
5724{
5725 struct wq_device *wq_dev;
5726 int ret;
5727
5728 /*
5729 * Adjusting max_active or creating new pwqs by applying
5730 * attributes breaks ordering guarantee. Disallow exposing ordered
5731 * workqueues.
5732 */
5733 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5734 return -EINVAL;
5735
5736 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5737 if (!wq_dev)
5738 return -ENOMEM;
5739
5740 wq_dev->wq = wq;
5741 wq_dev->dev.bus = &wq_subsys;
5742 wq_dev->dev.release = wq_device_release;
5743 dev_set_name(&wq_dev->dev, "%s", wq->name);
5744
5745 /*
5746 * unbound_attrs are created separately. Suppress uevent until
5747 * everything is ready.
5748 */
5749 dev_set_uevent_suppress(&wq_dev->dev, true);
5750
5751 ret = device_register(&wq_dev->dev);
5752 if (ret) {
5753 put_device(&wq_dev->dev);
5754 wq->wq_dev = NULL;
5755 return ret;
5756 }
5757
5758 if (wq->flags & WQ_UNBOUND) {
5759 struct device_attribute *attr;
5760
5761 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5762 ret = device_create_file(&wq_dev->dev, attr);
5763 if (ret) {
5764 device_unregister(&wq_dev->dev);
5765 wq->wq_dev = NULL;
5766 return ret;
5767 }
5768 }
5769 }
5770
5771 dev_set_uevent_suppress(&wq_dev->dev, false);
5772 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5773 return 0;
5774}
5775
5776/**
5777 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5778 * @wq: the workqueue to unregister
5779 *
5780 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5781 */
5782static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5783{
5784 struct wq_device *wq_dev = wq->wq_dev;
5785
5786 if (!wq->wq_dev)
5787 return;
5788
5789 wq->wq_dev = NULL;
5790 device_unregister(&wq_dev->dev);
5791}
5792#else /* CONFIG_SYSFS */
5793static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5794#endif /* CONFIG_SYSFS */
5795
5796/*
5797 * Workqueue watchdog.
5798 *
5799 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5800 * flush dependency, a concurrency managed work item which stays RUNNING
5801 * indefinitely. Workqueue stalls can be very difficult to debug as the
5802 * usual warning mechanisms don't trigger and internal workqueue state is
5803 * largely opaque.
5804 *
5805 * Workqueue watchdog monitors all worker pools periodically and dumps
5806 * state if some pools failed to make forward progress for a while where
5807 * forward progress is defined as the first item on ->worklist changing.
5808 *
5809 * This mechanism is controlled through the kernel parameter
5810 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5811 * corresponding sysfs parameter file.
5812 */
5813#ifdef CONFIG_WQ_WATCHDOG
5814
5815static unsigned long wq_watchdog_thresh = 30;
5816static struct timer_list wq_watchdog_timer;
5817
5818static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5819static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5820
5821static void wq_watchdog_reset_touched(void)
5822{
5823 int cpu;
5824
5825 wq_watchdog_touched = jiffies;
5826 for_each_possible_cpu(cpu)
5827 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5828}
5829
5830static void wq_watchdog_timer_fn(struct timer_list *unused)
5831{
5832 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5833 bool lockup_detected = false;
5834 unsigned long now = jiffies;
5835 struct worker_pool *pool;
5836 int pi;
5837
5838 if (!thresh)
5839 return;
5840
5841 rcu_read_lock();
5842
5843 for_each_pool(pool, pi) {
5844 unsigned long pool_ts, touched, ts;
5845
5846 if (list_empty(&pool->worklist))
5847 continue;
5848
5849 /*
5850 * If a virtual machine is stopped by the host it can look to
5851 * the watchdog like a stall.
5852 */
5853 kvm_check_and_clear_guest_paused();
5854
5855 /* get the latest of pool and touched timestamps */
5856 if (pool->cpu >= 0)
5857 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
5858 else
5859 touched = READ_ONCE(wq_watchdog_touched);
5860 pool_ts = READ_ONCE(pool->watchdog_ts);
5861
5862 if (time_after(pool_ts, touched))
5863 ts = pool_ts;
5864 else
5865 ts = touched;
5866
5867 /* did we stall? */
5868 if (time_after(now, ts + thresh)) {
5869 lockup_detected = true;
5870 pr_emerg("BUG: workqueue lockup - pool");
5871 pr_cont_pool_info(pool);
5872 pr_cont(" stuck for %us!\n",
5873 jiffies_to_msecs(now - pool_ts) / 1000);
5874 }
5875 }
5876
5877 rcu_read_unlock();
5878
5879 if (lockup_detected)
5880 show_all_workqueues();
5881
5882 wq_watchdog_reset_touched();
5883 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5884}
5885
5886notrace void wq_watchdog_touch(int cpu)
5887{
5888 if (cpu >= 0)
5889 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5890
5891 wq_watchdog_touched = jiffies;
5892}
5893
5894static void wq_watchdog_set_thresh(unsigned long thresh)
5895{
5896 wq_watchdog_thresh = 0;
5897 del_timer_sync(&wq_watchdog_timer);
5898
5899 if (thresh) {
5900 wq_watchdog_thresh = thresh;
5901 wq_watchdog_reset_touched();
5902 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5903 }
5904}
5905
5906static int wq_watchdog_param_set_thresh(const char *val,
5907 const struct kernel_param *kp)
5908{
5909 unsigned long thresh;
5910 int ret;
5911
5912 ret = kstrtoul(val, 0, &thresh);
5913 if (ret)
5914 return ret;
5915
5916 if (system_wq)
5917 wq_watchdog_set_thresh(thresh);
5918 else
5919 wq_watchdog_thresh = thresh;
5920
5921 return 0;
5922}
5923
5924static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5925 .set = wq_watchdog_param_set_thresh,
5926 .get = param_get_ulong,
5927};
5928
5929module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5930 0644);
5931
5932static void wq_watchdog_init(void)
5933{
5934 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5935 wq_watchdog_set_thresh(wq_watchdog_thresh);
5936}
5937
5938#else /* CONFIG_WQ_WATCHDOG */
5939
5940static inline void wq_watchdog_init(void) { }
5941
5942#endif /* CONFIG_WQ_WATCHDOG */
5943
5944static void __init wq_numa_init(void)
5945{
5946 cpumask_var_t *tbl;
5947 int node, cpu;
5948
5949 if (num_possible_nodes() <= 1)
5950 return;
5951
5952 if (wq_disable_numa) {
5953 pr_info("workqueue: NUMA affinity support disabled\n");
5954 return;
5955 }
5956
5957 for_each_possible_cpu(cpu) {
5958 if (WARN_ON(cpu_to_node(cpu) == NUMA_NO_NODE)) {
5959 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5960 return;
5961 }
5962 }
5963
5964 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5965 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5966
5967 /*
5968 * We want masks of possible CPUs of each node which isn't readily
5969 * available. Build one from cpu_to_node() which should have been
5970 * fully initialized by now.
5971 */
5972 tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5973 BUG_ON(!tbl);
5974
5975 for_each_node(node)
5976 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5977 node_online(node) ? node : NUMA_NO_NODE));
5978
5979 for_each_possible_cpu(cpu) {
5980 node = cpu_to_node(cpu);
5981 cpumask_set_cpu(cpu, tbl[node]);
5982 }
5983
5984 wq_numa_possible_cpumask = tbl;
5985 wq_numa_enabled = true;
5986}
5987
5988/**
5989 * workqueue_init_early - early init for workqueue subsystem
5990 *
5991 * This is the first half of two-staged workqueue subsystem initialization
5992 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5993 * idr are up. It sets up all the data structures and system workqueues
5994 * and allows early boot code to create workqueues and queue/cancel work
5995 * items. Actual work item execution starts only after kthreads can be
5996 * created and scheduled right before early initcalls.
5997 */
5998void __init workqueue_init_early(void)
5999{
6000 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6001 int i, cpu;
6002
6003 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6004
6005 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6006 cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ));
6007 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN));
6008
6009 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6010
6011 /* initialize CPU pools */
6012 for_each_possible_cpu(cpu) {
6013 struct worker_pool *pool;
6014
6015 i = 0;
6016 for_each_cpu_worker_pool(pool, cpu) {
6017 BUG_ON(init_worker_pool(pool));
6018 pool->cpu = cpu;
6019 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6020 pool->attrs->nice = std_nice[i++];
6021 pool->node = cpu_to_node(cpu);
6022
6023 /* alloc pool ID */
6024 mutex_lock(&wq_pool_mutex);
6025 BUG_ON(worker_pool_assign_id(pool));
6026 mutex_unlock(&wq_pool_mutex);
6027 }
6028 }
6029
6030 /* create default unbound and ordered wq attrs */
6031 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6032 struct workqueue_attrs *attrs;
6033
6034 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6035 attrs->nice = std_nice[i];
6036 unbound_std_wq_attrs[i] = attrs;
6037
6038 /*
6039 * An ordered wq should have only one pwq as ordering is
6040 * guaranteed by max_active which is enforced by pwqs.
6041 * Turn off NUMA so that dfl_pwq is used for all nodes.
6042 */
6043 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6044 attrs->nice = std_nice[i];
6045 attrs->no_numa = true;
6046 ordered_wq_attrs[i] = attrs;
6047 }
6048
6049 system_wq = alloc_workqueue("events", 0, 0);
6050 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6051 system_long_wq = alloc_workqueue("events_long", 0, 0);
6052 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6053 WQ_UNBOUND_MAX_ACTIVE);
6054 system_freezable_wq = alloc_workqueue("events_freezable",
6055 WQ_FREEZABLE, 0);
6056 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6057 WQ_POWER_EFFICIENT, 0);
6058 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6059 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6060 0);
6061 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6062 !system_unbound_wq || !system_freezable_wq ||
6063 !system_power_efficient_wq ||
6064 !system_freezable_power_efficient_wq);
6065}
6066
6067/**
6068 * workqueue_init - bring workqueue subsystem fully online
6069 *
6070 * This is the latter half of two-staged workqueue subsystem initialization
6071 * and invoked as soon as kthreads can be created and scheduled.
6072 * Workqueues have been created and work items queued on them, but there
6073 * are no kworkers executing the work items yet. Populate the worker pools
6074 * with the initial workers and enable future kworker creations.
6075 */
6076void __init workqueue_init(void)
6077{
6078 struct workqueue_struct *wq;
6079 struct worker_pool *pool;
6080 int cpu, bkt;
6081
6082 /*
6083 * It'd be simpler to initialize NUMA in workqueue_init_early() but
6084 * CPU to node mapping may not be available that early on some
6085 * archs such as power and arm64. As per-cpu pools created
6086 * previously could be missing node hint and unbound pools NUMA
6087 * affinity, fix them up.
6088 *
6089 * Also, while iterating workqueues, create rescuers if requested.
6090 */
6091 wq_numa_init();
6092
6093 mutex_lock(&wq_pool_mutex);
6094
6095 for_each_possible_cpu(cpu) {
6096 for_each_cpu_worker_pool(pool, cpu) {
6097 pool->node = cpu_to_node(cpu);
6098 }
6099 }
6100
6101 list_for_each_entry(wq, &workqueues, list) {
6102 wq_update_unbound_numa(wq, smp_processor_id(), true);
6103 WARN(init_rescuer(wq),
6104 "workqueue: failed to create early rescuer for %s",
6105 wq->name);
6106 }
6107
6108 mutex_unlock(&wq_pool_mutex);
6109
6110 /* create the initial workers */
6111 for_each_online_cpu(cpu) {
6112 for_each_cpu_worker_pool(pool, cpu) {
6113 pool->flags &= ~POOL_DISASSOCIATED;
6114 BUG_ON(!create_worker(pool));
6115 }
6116 }
6117
6118 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6119 BUG_ON(!create_worker(pool));
6120
6121 wq_online = true;
6122 wq_watchdog_init();
6123}
6124
6125/*
6126 * Despite the naming, this is a no-op function which is here only for avoiding
6127 * link error. Since compile-time warning may fail to catch, we will need to
6128 * emit run-time warning from __flush_workqueue().
6129 */
6130void __warn_flushing_systemwide_wq(void) { }
6131EXPORT_SYMBOL(__warn_flushing_systemwide_wq);