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/interrupt.h>
33#include <linux/signal.h>
34#include <linux/completion.h>
35#include <linux/workqueue.h>
36#include <linux/slab.h>
37#include <linux/cpu.h>
38#include <linux/notifier.h>
39#include <linux/kthread.h>
40#include <linux/hardirq.h>
41#include <linux/mempolicy.h>
42#include <linux/freezer.h>
43#include <linux/debug_locks.h>
44#include <linux/lockdep.h>
45#include <linux/idr.h>
46#include <linux/jhash.h>
47#include <linux/hashtable.h>
48#include <linux/rculist.h>
49#include <linux/nodemask.h>
50#include <linux/moduleparam.h>
51#include <linux/uaccess.h>
52#include <linux/sched/isolation.h>
53#include <linux/sched/debug.h>
54#include <linux/nmi.h>
55#include <linux/kvm_para.h>
56#include <linux/delay.h>
57#include <linux/irq_work.h>
58
59#include "workqueue_internal.h"
60
61enum worker_pool_flags {
62 /*
63 * worker_pool flags
64 *
65 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect.
69 *
70 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool behaves as an unbound one.
73 *
74 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progress.
77 *
78 * As there can only be one concurrent BH execution context per CPU, a
79 * BH pool is per-CPU and always DISASSOCIATED.
80 */
81 POOL_BH = 1 << 0, /* is a BH pool */
82 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
83 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
85};
86
87enum worker_flags {
88 /* worker flags */
89 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95
96 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOUND | WORKER_REBOUND,
98};
99
100enum work_cancel_flags {
101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102};
103
104enum wq_internal_consts {
105 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
106
107 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
108 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
109
110 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
111 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
112
113 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
114 /* call for help after 10ms
115 (min two ticks) */
116 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
117 CREATE_COOLDOWN = HZ, /* time to breath after fail */
118
119 /*
120 * Rescue workers are used only on emergencies and shared by
121 * all cpus. Give MIN_NICE.
122 */
123 RESCUER_NICE_LEVEL = MIN_NICE,
124 HIGHPRI_NICE_LEVEL = MIN_NICE,
125
126 WQ_NAME_LEN = 32,
127};
128
129/*
130 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
131 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
132 * msecs_to_jiffies() can't be an initializer.
133 */
134#define BH_WORKER_JIFFIES msecs_to_jiffies(2)
135#define BH_WORKER_RESTARTS 10
136
137/*
138 * Structure fields follow one of the following exclusion rules.
139 *
140 * I: Modifiable by initialization/destruction paths and read-only for
141 * everyone else.
142 *
143 * P: Preemption protected. Disabling preemption is enough and should
144 * only be modified and accessed from the local cpu.
145 *
146 * L: pool->lock protected. Access with pool->lock held.
147 *
148 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
149 * reads.
150 *
151 * K: Only modified by worker while holding pool->lock. Can be safely read by
152 * self, while holding pool->lock or from IRQ context if %current is the
153 * kworker.
154 *
155 * S: Only modified by worker self.
156 *
157 * A: wq_pool_attach_mutex protected.
158 *
159 * PL: wq_pool_mutex protected.
160 *
161 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
162 *
163 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
164 *
165 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
166 * RCU for reads.
167 *
168 * WQ: wq->mutex protected.
169 *
170 * WR: wq->mutex protected for writes. RCU protected for reads.
171 *
172 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
173 * with READ_ONCE() without locking.
174 *
175 * MD: wq_mayday_lock protected.
176 *
177 * WD: Used internally by the watchdog.
178 */
179
180/* struct worker is defined in workqueue_internal.h */
181
182struct worker_pool {
183 raw_spinlock_t lock; /* the pool lock */
184 int cpu; /* I: the associated cpu */
185 int node; /* I: the associated node ID */
186 int id; /* I: pool ID */
187 unsigned int flags; /* L: flags */
188
189 unsigned long watchdog_ts; /* L: watchdog timestamp */
190 bool cpu_stall; /* WD: stalled cpu bound pool */
191
192 /*
193 * The counter is incremented in a process context on the associated CPU
194 * w/ preemption disabled, and decremented or reset in the same context
195 * but w/ pool->lock held. The readers grab pool->lock and are
196 * guaranteed to see if the counter reached zero.
197 */
198 int nr_running;
199
200 struct list_head worklist; /* L: list of pending works */
201
202 int nr_workers; /* L: total number of workers */
203 int nr_idle; /* L: currently idle workers */
204
205 struct list_head idle_list; /* L: list of idle workers */
206 struct timer_list idle_timer; /* L: worker idle timeout */
207 struct work_struct idle_cull_work; /* L: worker idle cleanup */
208
209 struct timer_list mayday_timer; /* L: SOS timer for workers */
210
211 /* a workers is either on busy_hash or idle_list, or the manager */
212 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
213 /* L: hash of busy workers */
214
215 struct worker *manager; /* L: purely informational */
216 struct list_head workers; /* A: attached workers */
217 struct list_head dying_workers; /* A: workers about to die */
218 struct completion *detach_completion; /* all workers detached */
219
220 struct ida worker_ida; /* worker IDs for task name */
221
222 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; /* PL: refcnt for unbound pools */
225
226 /*
227 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool().
229 */
230 struct rcu_head rcu;
231};
232
233/*
234 * Per-pool_workqueue statistics. These can be monitored using
235 * tools/workqueue/wq_monitor.py.
236 */
237enum pool_workqueue_stats {
238 PWQ_STAT_STARTED, /* work items started execution */
239 PWQ_STAT_COMPLETED, /* work items completed execution */
240 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244 PWQ_STAT_MAYDAY, /* maydays to rescuer */
245 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
246
247 PWQ_NR_STATS,
248};
249
250/*
251 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
252 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits.
255 */
256struct pool_workqueue {
257 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; /* L: current color */
260 int flush_color; /* L: flushing color */
261 int refcnt; /* L: reference count */
262 int nr_in_flight[WORK_NR_COLORS];
263 /* L: nr of in_flight works */
264 bool plugged; /* L: execution suspended */
265
266 /*
267 * nr_active management and WORK_STRUCT_INACTIVE:
268 *
269 * When pwq->nr_active >= max_active, new work item is queued to
270 * pwq->inactive_works instead of pool->worklist and marked with
271 * WORK_STRUCT_INACTIVE.
272 *
273 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
274 * nr_active and all work items in pwq->inactive_works are marked with
275 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276 * in pwq->inactive_works. Some of them are ready to run in
277 * pool->worklist or worker->scheduled. Those work itmes are only struct
278 * wq_barrier which is used for flush_work() and should not participate
279 * in nr_active. For non-barrier work item, it is marked with
280 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281 */
282 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works; /* L: inactive works */
284 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
285 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; /* MD: node on wq->maydays */
287
288 u64 stats[PWQ_NR_STATS];
289
290 /*
291 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 * RCU protected so that the first pwq can be determined without
294 * grabbing wq->mutex.
295 */
296 struct kthread_work release_work;
297 struct rcu_head rcu;
298} __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
299
300/*
301 * Structure used to wait for workqueue flush.
302 */
303struct wq_flusher {
304 struct list_head list; /* WQ: list of flushers */
305 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; /* flush completion */
307};
308
309struct wq_device;
310
311/*
312 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
313 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
314 * As sharing a single nr_active across multiple sockets can be very expensive,
315 * the counting and enforcement is per NUMA node.
316 *
317 * The following struct is used to enforce per-node max_active. When a pwq wants
318 * to start executing a work item, it should increment ->nr using
319 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
322 * round-robin order.
323 */
324struct wq_node_nr_active {
325 int max; /* per-node max_active */
326 atomic_t nr; /* per-node nr_active */
327 raw_spinlock_t lock; /* nests inside pool locks */
328 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
329};
330
331/*
332 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its pool_workqueues.
334 */
335struct workqueue_struct {
336 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; /* PR: list of all workqueues */
338
339 struct mutex mutex; /* protects this wq */
340 int work_color; /* WQ: current work color */
341 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overflow; /* WQ: flush overflow list */
346
347 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; /* MD: rescue worker */
349
350 int nr_drainers; /* WQ: drain in progress */
351
352 /* See alloc_workqueue() function comment for info on min/max_active */
353 int max_active; /* WO: max active works */
354 int min_active; /* WO: min active works */
355 int saved_max_active; /* WQ: saved max_active */
356 int saved_min_active; /* WQ: saved min_active */
357
358 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
360
361#ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; /* I: for sysfs interface */
363#endif
364#ifdef CONFIG_LOCKDEP
365 char *lock_name;
366 struct lock_class_key key;
367 struct lockdep_map lockdep_map;
368#endif
369 char name[WQ_NAME_LEN]; /* I: workqueue name */
370
371 /*
372 * Destruction of workqueue_struct is RCU protected to allow walking
373 * the workqueues list without grabbing wq_pool_mutex.
374 * This is used to dump all workqueues from sysrq.
375 */
376 struct rcu_head rcu;
377
378 /* hot fields used during command issue, aligned to cacheline */
379 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
380 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
381 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
382};
383
384/*
385 * Each pod type describes how CPUs should be grouped for unbound workqueues.
386 * See the comment above workqueue_attrs->affn_scope.
387 */
388struct wq_pod_type {
389 int nr_pods; /* number of pods */
390 cpumask_var_t *pod_cpus; /* pod -> cpus */
391 int *pod_node; /* pod -> node */
392 int *cpu_pod; /* cpu -> pod */
393};
394
395static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
396 [WQ_AFFN_DFL] = "default",
397 [WQ_AFFN_CPU] = "cpu",
398 [WQ_AFFN_SMT] = "smt",
399 [WQ_AFFN_CACHE] = "cache",
400 [WQ_AFFN_NUMA] = "numa",
401 [WQ_AFFN_SYSTEM] = "system",
402};
403
404/*
405 * Per-cpu work items which run for longer than the following threshold are
406 * automatically considered CPU intensive and excluded from concurrency
407 * management to prevent them from noticeably delaying other per-cpu work items.
408 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
409 * The actual value is initialized in wq_cpu_intensive_thresh_init().
410 */
411static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
412module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
413#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
414static unsigned int wq_cpu_intensive_warning_thresh = 4;
415module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
416#endif
417
418/* see the comment above the definition of WQ_POWER_EFFICIENT */
419static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
420module_param_named(power_efficient, wq_power_efficient, bool, 0444);
421
422static bool wq_online; /* can kworkers be created yet? */
423static bool wq_topo_initialized __read_mostly = false;
424
425static struct kmem_cache *pwq_cache;
426
427static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
428static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
429
430/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
431static struct workqueue_attrs *wq_update_pod_attrs_buf;
432
433static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
434static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
435static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
436/* wait for manager to go away */
437static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
438
439static LIST_HEAD(workqueues); /* PR: list of all workqueues */
440static bool workqueue_freezing; /* PL: have wqs started freezing? */
441
442/* PL&A: allowable cpus for unbound wqs and work items */
443static cpumask_var_t wq_unbound_cpumask;
444
445/* PL: user requested unbound cpumask via sysfs */
446static cpumask_var_t wq_requested_unbound_cpumask;
447
448/* PL: isolated cpumask to be excluded from unbound cpumask */
449static cpumask_var_t wq_isolated_cpumask;
450
451/* for further constrain wq_unbound_cpumask by cmdline parameter*/
452static struct cpumask wq_cmdline_cpumask __initdata;
453
454/* CPU where unbound work was last round robin scheduled from this CPU */
455static DEFINE_PER_CPU(int, wq_rr_cpu_last);
456
457/*
458 * Local execution of unbound work items is no longer guaranteed. The
459 * following always forces round-robin CPU selection on unbound work items
460 * to uncover usages which depend on it.
461 */
462#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
463static bool wq_debug_force_rr_cpu = true;
464#else
465static bool wq_debug_force_rr_cpu = false;
466#endif
467module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
468
469/* to raise softirq for the BH worker pools on other CPUs */
470static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS],
471 bh_pool_irq_works);
472
473/* the BH worker pools */
474static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
475 bh_worker_pools);
476
477/* the per-cpu worker pools */
478static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
479 cpu_worker_pools);
480
481static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
482
483/* PL: hash of all unbound pools keyed by pool->attrs */
484static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
485
486/* I: attributes used when instantiating standard unbound pools on demand */
487static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
488
489/* I: attributes used when instantiating ordered pools on demand */
490static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
491
492/*
493 * Used to synchronize multiple cancel_sync attempts on the same work item. See
494 * work_grab_pending() and __cancel_work_sync().
495 */
496static DECLARE_WAIT_QUEUE_HEAD(wq_cancel_waitq);
497
498/*
499 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
500 * process context while holding a pool lock. Bounce to a dedicated kthread
501 * worker to avoid A-A deadlocks.
502 */
503static struct kthread_worker *pwq_release_worker __ro_after_init;
504
505struct workqueue_struct *system_wq __ro_after_init;
506EXPORT_SYMBOL(system_wq);
507struct workqueue_struct *system_highpri_wq __ro_after_init;
508EXPORT_SYMBOL_GPL(system_highpri_wq);
509struct workqueue_struct *system_long_wq __ro_after_init;
510EXPORT_SYMBOL_GPL(system_long_wq);
511struct workqueue_struct *system_unbound_wq __ro_after_init;
512EXPORT_SYMBOL_GPL(system_unbound_wq);
513struct workqueue_struct *system_freezable_wq __ro_after_init;
514EXPORT_SYMBOL_GPL(system_freezable_wq);
515struct workqueue_struct *system_power_efficient_wq __ro_after_init;
516EXPORT_SYMBOL_GPL(system_power_efficient_wq);
517struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
518EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
519struct workqueue_struct *system_bh_wq;
520EXPORT_SYMBOL_GPL(system_bh_wq);
521struct workqueue_struct *system_bh_highpri_wq;
522EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
523
524static int worker_thread(void *__worker);
525static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
526static void show_pwq(struct pool_workqueue *pwq);
527static void show_one_worker_pool(struct worker_pool *pool);
528
529#define CREATE_TRACE_POINTS
530#include <trace/events/workqueue.h>
531
532#define assert_rcu_or_pool_mutex() \
533 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
534 !lockdep_is_held(&wq_pool_mutex), \
535 "RCU or wq_pool_mutex should be held")
536
537#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
538 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
539 !lockdep_is_held(&wq->mutex) && \
540 !lockdep_is_held(&wq_pool_mutex), \
541 "RCU, wq->mutex or wq_pool_mutex should be held")
542
543#define for_each_bh_worker_pool(pool, cpu) \
544 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
545 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
546 (pool)++)
547
548#define for_each_cpu_worker_pool(pool, cpu) \
549 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
550 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
551 (pool)++)
552
553/**
554 * for_each_pool - iterate through all worker_pools in the system
555 * @pool: iteration cursor
556 * @pi: integer used for iteration
557 *
558 * This must be called either with wq_pool_mutex held or RCU read
559 * locked. If the pool needs to be used beyond the locking in effect, the
560 * caller is responsible for guaranteeing that the pool stays online.
561 *
562 * The if/else clause exists only for the lockdep assertion and can be
563 * ignored.
564 */
565#define for_each_pool(pool, pi) \
566 idr_for_each_entry(&worker_pool_idr, pool, pi) \
567 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
568 else
569
570/**
571 * for_each_pool_worker - iterate through all workers of a worker_pool
572 * @worker: iteration cursor
573 * @pool: worker_pool to iterate workers of
574 *
575 * This must be called with wq_pool_attach_mutex.
576 *
577 * The if/else clause exists only for the lockdep assertion and can be
578 * ignored.
579 */
580#define for_each_pool_worker(worker, pool) \
581 list_for_each_entry((worker), &(pool)->workers, node) \
582 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
583 else
584
585/**
586 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
587 * @pwq: iteration cursor
588 * @wq: the target workqueue
589 *
590 * This must be called either with wq->mutex held or RCU read locked.
591 * If the pwq needs to be used beyond the locking in effect, the caller is
592 * responsible for guaranteeing that the pwq stays online.
593 *
594 * The if/else clause exists only for the lockdep assertion and can be
595 * ignored.
596 */
597#define for_each_pwq(pwq, wq) \
598 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
599 lockdep_is_held(&(wq->mutex)))
600
601#ifdef CONFIG_DEBUG_OBJECTS_WORK
602
603static const struct debug_obj_descr work_debug_descr;
604
605static void *work_debug_hint(void *addr)
606{
607 return ((struct work_struct *) addr)->func;
608}
609
610static bool work_is_static_object(void *addr)
611{
612 struct work_struct *work = addr;
613
614 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
615}
616
617/*
618 * fixup_init is called when:
619 * - an active object is initialized
620 */
621static bool work_fixup_init(void *addr, enum debug_obj_state state)
622{
623 struct work_struct *work = addr;
624
625 switch (state) {
626 case ODEBUG_STATE_ACTIVE:
627 cancel_work_sync(work);
628 debug_object_init(work, &work_debug_descr);
629 return true;
630 default:
631 return false;
632 }
633}
634
635/*
636 * fixup_free is called when:
637 * - an active object is freed
638 */
639static bool work_fixup_free(void *addr, enum debug_obj_state state)
640{
641 struct work_struct *work = addr;
642
643 switch (state) {
644 case ODEBUG_STATE_ACTIVE:
645 cancel_work_sync(work);
646 debug_object_free(work, &work_debug_descr);
647 return true;
648 default:
649 return false;
650 }
651}
652
653static const struct debug_obj_descr work_debug_descr = {
654 .name = "work_struct",
655 .debug_hint = work_debug_hint,
656 .is_static_object = work_is_static_object,
657 .fixup_init = work_fixup_init,
658 .fixup_free = work_fixup_free,
659};
660
661static inline void debug_work_activate(struct work_struct *work)
662{
663 debug_object_activate(work, &work_debug_descr);
664}
665
666static inline void debug_work_deactivate(struct work_struct *work)
667{
668 debug_object_deactivate(work, &work_debug_descr);
669}
670
671void __init_work(struct work_struct *work, int onstack)
672{
673 if (onstack)
674 debug_object_init_on_stack(work, &work_debug_descr);
675 else
676 debug_object_init(work, &work_debug_descr);
677}
678EXPORT_SYMBOL_GPL(__init_work);
679
680void destroy_work_on_stack(struct work_struct *work)
681{
682 debug_object_free(work, &work_debug_descr);
683}
684EXPORT_SYMBOL_GPL(destroy_work_on_stack);
685
686void destroy_delayed_work_on_stack(struct delayed_work *work)
687{
688 destroy_timer_on_stack(&work->timer);
689 debug_object_free(&work->work, &work_debug_descr);
690}
691EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
692
693#else
694static inline void debug_work_activate(struct work_struct *work) { }
695static inline void debug_work_deactivate(struct work_struct *work) { }
696#endif
697
698/**
699 * worker_pool_assign_id - allocate ID and assign it to @pool
700 * @pool: the pool pointer of interest
701 *
702 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
703 * successfully, -errno on failure.
704 */
705static int worker_pool_assign_id(struct worker_pool *pool)
706{
707 int ret;
708
709 lockdep_assert_held(&wq_pool_mutex);
710
711 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
712 GFP_KERNEL);
713 if (ret >= 0) {
714 pool->id = ret;
715 return 0;
716 }
717 return ret;
718}
719
720static struct pool_workqueue __rcu **
721unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
722{
723 if (cpu >= 0)
724 return per_cpu_ptr(wq->cpu_pwq, cpu);
725 else
726 return &wq->dfl_pwq;
727}
728
729/* @cpu < 0 for dfl_pwq */
730static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
731{
732 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
733 lockdep_is_held(&wq_pool_mutex) ||
734 lockdep_is_held(&wq->mutex));
735}
736
737/**
738 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
739 * @wq: workqueue of interest
740 *
741 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
742 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
743 * default pwq is always mapped to the pool with the current effective cpumask.
744 */
745static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
746{
747 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
748}
749
750static unsigned int work_color_to_flags(int color)
751{
752 return color << WORK_STRUCT_COLOR_SHIFT;
753}
754
755static int get_work_color(unsigned long work_data)
756{
757 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
758 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
759}
760
761static int work_next_color(int color)
762{
763 return (color + 1) % WORK_NR_COLORS;
764}
765
766/*
767 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
768 * contain the pointer to the queued pwq. Once execution starts, the flag
769 * is cleared and the high bits contain OFFQ flags and pool ID.
770 *
771 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
772 * can be used to set the pwq, pool or clear work->data. These functions should
773 * only be called while the work is owned - ie. while the PENDING bit is set.
774 *
775 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
776 * corresponding to a work. Pool is available once the work has been
777 * queued anywhere after initialization until it is sync canceled. pwq is
778 * available only while the work item is queued.
779 *
780 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
781 * canceled. While being canceled, a work item may have its PENDING set
782 * but stay off timer and worklist for arbitrarily long and nobody should
783 * try to steal the PENDING bit.
784 */
785static inline void set_work_data(struct work_struct *work, unsigned long data)
786{
787 WARN_ON_ONCE(!work_pending(work));
788 atomic_long_set(&work->data, data | work_static(work));
789}
790
791static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
792 unsigned long flags)
793{
794 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
795 WORK_STRUCT_PWQ | flags);
796}
797
798static void set_work_pool_and_keep_pending(struct work_struct *work,
799 int pool_id, unsigned long flags)
800{
801 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
802 WORK_STRUCT_PENDING | flags);
803}
804
805static void set_work_pool_and_clear_pending(struct work_struct *work,
806 int pool_id, unsigned long flags)
807{
808 /*
809 * The following wmb is paired with the implied mb in
810 * test_and_set_bit(PENDING) and ensures all updates to @work made
811 * here are visible to and precede any updates by the next PENDING
812 * owner.
813 */
814 smp_wmb();
815 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
816 flags);
817 /*
818 * The following mb guarantees that previous clear of a PENDING bit
819 * will not be reordered with any speculative LOADS or STORES from
820 * work->current_func, which is executed afterwards. This possible
821 * reordering can lead to a missed execution on attempt to queue
822 * the same @work. E.g. consider this case:
823 *
824 * CPU#0 CPU#1
825 * ---------------------------- --------------------------------
826 *
827 * 1 STORE event_indicated
828 * 2 queue_work_on() {
829 * 3 test_and_set_bit(PENDING)
830 * 4 } set_..._and_clear_pending() {
831 * 5 set_work_data() # clear bit
832 * 6 smp_mb()
833 * 7 work->current_func() {
834 * 8 LOAD event_indicated
835 * }
836 *
837 * Without an explicit full barrier speculative LOAD on line 8 can
838 * be executed before CPU#0 does STORE on line 1. If that happens,
839 * CPU#0 observes the PENDING bit is still set and new execution of
840 * a @work is not queued in a hope, that CPU#1 will eventually
841 * finish the queued @work. Meanwhile CPU#1 does not see
842 * event_indicated is set, because speculative LOAD was executed
843 * before actual STORE.
844 */
845 smp_mb();
846}
847
848static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
849{
850 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
851}
852
853static struct pool_workqueue *get_work_pwq(struct work_struct *work)
854{
855 unsigned long data = atomic_long_read(&work->data);
856
857 if (data & WORK_STRUCT_PWQ)
858 return work_struct_pwq(data);
859 else
860 return NULL;
861}
862
863/**
864 * get_work_pool - return the worker_pool a given work was associated with
865 * @work: the work item of interest
866 *
867 * Pools are created and destroyed under wq_pool_mutex, and allows read
868 * access under RCU read lock. As such, this function should be
869 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
870 *
871 * All fields of the returned pool are accessible as long as the above
872 * mentioned locking is in effect. If the returned pool needs to be used
873 * beyond the critical section, the caller is responsible for ensuring the
874 * returned pool is and stays online.
875 *
876 * Return: The worker_pool @work was last associated with. %NULL if none.
877 */
878static struct worker_pool *get_work_pool(struct work_struct *work)
879{
880 unsigned long data = atomic_long_read(&work->data);
881 int pool_id;
882
883 assert_rcu_or_pool_mutex();
884
885 if (data & WORK_STRUCT_PWQ)
886 return work_struct_pwq(data)->pool;
887
888 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
889 if (pool_id == WORK_OFFQ_POOL_NONE)
890 return NULL;
891
892 return idr_find(&worker_pool_idr, pool_id);
893}
894
895/**
896 * get_work_pool_id - return the worker pool ID a given work is associated with
897 * @work: the work item of interest
898 *
899 * Return: The worker_pool ID @work was last associated with.
900 * %WORK_OFFQ_POOL_NONE if none.
901 */
902static int get_work_pool_id(struct work_struct *work)
903{
904 unsigned long data = atomic_long_read(&work->data);
905
906 if (data & WORK_STRUCT_PWQ)
907 return work_struct_pwq(data)->pool->id;
908
909 return data >> WORK_OFFQ_POOL_SHIFT;
910}
911
912static void mark_work_canceling(struct work_struct *work)
913{
914 unsigned long pool_id = get_work_pool_id(work);
915
916 pool_id <<= WORK_OFFQ_POOL_SHIFT;
917 set_work_data(work, pool_id | WORK_STRUCT_PENDING | WORK_OFFQ_CANCELING);
918}
919
920static bool work_is_canceling(struct work_struct *work)
921{
922 unsigned long data = atomic_long_read(&work->data);
923
924 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
925}
926
927/*
928 * Policy functions. These define the policies on how the global worker
929 * pools are managed. Unless noted otherwise, these functions assume that
930 * they're being called with pool->lock held.
931 */
932
933/*
934 * Need to wake up a worker? Called from anything but currently
935 * running workers.
936 *
937 * Note that, because unbound workers never contribute to nr_running, this
938 * function will always return %true for unbound pools as long as the
939 * worklist isn't empty.
940 */
941static bool need_more_worker(struct worker_pool *pool)
942{
943 return !list_empty(&pool->worklist) && !pool->nr_running;
944}
945
946/* Can I start working? Called from busy but !running workers. */
947static bool may_start_working(struct worker_pool *pool)
948{
949 return pool->nr_idle;
950}
951
952/* Do I need to keep working? Called from currently running workers. */
953static bool keep_working(struct worker_pool *pool)
954{
955 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
956}
957
958/* Do we need a new worker? Called from manager. */
959static bool need_to_create_worker(struct worker_pool *pool)
960{
961 return need_more_worker(pool) && !may_start_working(pool);
962}
963
964/* Do we have too many workers and should some go away? */
965static bool too_many_workers(struct worker_pool *pool)
966{
967 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
968 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
969 int nr_busy = pool->nr_workers - nr_idle;
970
971 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
972}
973
974/**
975 * worker_set_flags - set worker flags and adjust nr_running accordingly
976 * @worker: self
977 * @flags: flags to set
978 *
979 * Set @flags in @worker->flags and adjust nr_running accordingly.
980 */
981static inline void worker_set_flags(struct worker *worker, unsigned int flags)
982{
983 struct worker_pool *pool = worker->pool;
984
985 lockdep_assert_held(&pool->lock);
986
987 /* If transitioning into NOT_RUNNING, adjust nr_running. */
988 if ((flags & WORKER_NOT_RUNNING) &&
989 !(worker->flags & WORKER_NOT_RUNNING)) {
990 pool->nr_running--;
991 }
992
993 worker->flags |= flags;
994}
995
996/**
997 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
998 * @worker: self
999 * @flags: flags to clear
1000 *
1001 * Clear @flags in @worker->flags and adjust nr_running accordingly.
1002 */
1003static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
1004{
1005 struct worker_pool *pool = worker->pool;
1006 unsigned int oflags = worker->flags;
1007
1008 lockdep_assert_held(&pool->lock);
1009
1010 worker->flags &= ~flags;
1011
1012 /*
1013 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1014 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1015 * of multiple flags, not a single flag.
1016 */
1017 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1018 if (!(worker->flags & WORKER_NOT_RUNNING))
1019 pool->nr_running++;
1020}
1021
1022/* Return the first idle worker. Called with pool->lock held. */
1023static struct worker *first_idle_worker(struct worker_pool *pool)
1024{
1025 if (unlikely(list_empty(&pool->idle_list)))
1026 return NULL;
1027
1028 return list_first_entry(&pool->idle_list, struct worker, entry);
1029}
1030
1031/**
1032 * worker_enter_idle - enter idle state
1033 * @worker: worker which is entering idle state
1034 *
1035 * @worker is entering idle state. Update stats and idle timer if
1036 * necessary.
1037 *
1038 * LOCKING:
1039 * raw_spin_lock_irq(pool->lock).
1040 */
1041static void worker_enter_idle(struct worker *worker)
1042{
1043 struct worker_pool *pool = worker->pool;
1044
1045 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1046 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1047 (worker->hentry.next || worker->hentry.pprev)))
1048 return;
1049
1050 /* can't use worker_set_flags(), also called from create_worker() */
1051 worker->flags |= WORKER_IDLE;
1052 pool->nr_idle++;
1053 worker->last_active = jiffies;
1054
1055 /* idle_list is LIFO */
1056 list_add(&worker->entry, &pool->idle_list);
1057
1058 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1059 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1060
1061 /* Sanity check nr_running. */
1062 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1063}
1064
1065/**
1066 * worker_leave_idle - leave idle state
1067 * @worker: worker which is leaving idle state
1068 *
1069 * @worker is leaving idle state. Update stats.
1070 *
1071 * LOCKING:
1072 * raw_spin_lock_irq(pool->lock).
1073 */
1074static void worker_leave_idle(struct worker *worker)
1075{
1076 struct worker_pool *pool = worker->pool;
1077
1078 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1079 return;
1080 worker_clr_flags(worker, WORKER_IDLE);
1081 pool->nr_idle--;
1082 list_del_init(&worker->entry);
1083}
1084
1085/**
1086 * find_worker_executing_work - find worker which is executing a work
1087 * @pool: pool of interest
1088 * @work: work to find worker for
1089 *
1090 * Find a worker which is executing @work on @pool by searching
1091 * @pool->busy_hash which is keyed by the address of @work. For a worker
1092 * to match, its current execution should match the address of @work and
1093 * its work function. This is to avoid unwanted dependency between
1094 * unrelated work executions through a work item being recycled while still
1095 * being executed.
1096 *
1097 * This is a bit tricky. A work item may be freed once its execution
1098 * starts and nothing prevents the freed area from being recycled for
1099 * another work item. If the same work item address ends up being reused
1100 * before the original execution finishes, workqueue will identify the
1101 * recycled work item as currently executing and make it wait until the
1102 * current execution finishes, introducing an unwanted dependency.
1103 *
1104 * This function checks the work item address and work function to avoid
1105 * false positives. Note that this isn't complete as one may construct a
1106 * work function which can introduce dependency onto itself through a
1107 * recycled work item. Well, if somebody wants to shoot oneself in the
1108 * foot that badly, there's only so much we can do, and if such deadlock
1109 * actually occurs, it should be easy to locate the culprit work function.
1110 *
1111 * CONTEXT:
1112 * raw_spin_lock_irq(pool->lock).
1113 *
1114 * Return:
1115 * Pointer to worker which is executing @work if found, %NULL
1116 * otherwise.
1117 */
1118static struct worker *find_worker_executing_work(struct worker_pool *pool,
1119 struct work_struct *work)
1120{
1121 struct worker *worker;
1122
1123 hash_for_each_possible(pool->busy_hash, worker, hentry,
1124 (unsigned long)work)
1125 if (worker->current_work == work &&
1126 worker->current_func == work->func)
1127 return worker;
1128
1129 return NULL;
1130}
1131
1132/**
1133 * move_linked_works - move linked works to a list
1134 * @work: start of series of works to be scheduled
1135 * @head: target list to append @work to
1136 * @nextp: out parameter for nested worklist walking
1137 *
1138 * Schedule linked works starting from @work to @head. Work series to be
1139 * scheduled starts at @work and includes any consecutive work with
1140 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1141 * @nextp.
1142 *
1143 * CONTEXT:
1144 * raw_spin_lock_irq(pool->lock).
1145 */
1146static void move_linked_works(struct work_struct *work, struct list_head *head,
1147 struct work_struct **nextp)
1148{
1149 struct work_struct *n;
1150
1151 /*
1152 * Linked worklist will always end before the end of the list,
1153 * use NULL for list head.
1154 */
1155 list_for_each_entry_safe_from(work, n, NULL, entry) {
1156 list_move_tail(&work->entry, head);
1157 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1158 break;
1159 }
1160
1161 /*
1162 * If we're already inside safe list traversal and have moved
1163 * multiple works to the scheduled queue, the next position
1164 * needs to be updated.
1165 */
1166 if (nextp)
1167 *nextp = n;
1168}
1169
1170/**
1171 * assign_work - assign a work item and its linked work items to a worker
1172 * @work: work to assign
1173 * @worker: worker to assign to
1174 * @nextp: out parameter for nested worklist walking
1175 *
1176 * Assign @work and its linked work items to @worker. If @work is already being
1177 * executed by another worker in the same pool, it'll be punted there.
1178 *
1179 * If @nextp is not NULL, it's updated to point to the next work of the last
1180 * scheduled work. This allows assign_work() to be nested inside
1181 * list_for_each_entry_safe().
1182 *
1183 * Returns %true if @work was successfully assigned to @worker. %false if @work
1184 * was punted to another worker already executing it.
1185 */
1186static bool assign_work(struct work_struct *work, struct worker *worker,
1187 struct work_struct **nextp)
1188{
1189 struct worker_pool *pool = worker->pool;
1190 struct worker *collision;
1191
1192 lockdep_assert_held(&pool->lock);
1193
1194 /*
1195 * A single work shouldn't be executed concurrently by multiple workers.
1196 * __queue_work() ensures that @work doesn't jump to a different pool
1197 * while still running in the previous pool. Here, we should ensure that
1198 * @work is not executed concurrently by multiple workers from the same
1199 * pool. Check whether anyone is already processing the work. If so,
1200 * defer the work to the currently executing one.
1201 */
1202 collision = find_worker_executing_work(pool, work);
1203 if (unlikely(collision)) {
1204 move_linked_works(work, &collision->scheduled, nextp);
1205 return false;
1206 }
1207
1208 move_linked_works(work, &worker->scheduled, nextp);
1209 return true;
1210}
1211
1212static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1213{
1214 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1215
1216 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1217}
1218
1219static void kick_bh_pool(struct worker_pool *pool)
1220{
1221#ifdef CONFIG_SMP
1222 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1223 if (unlikely(pool->cpu != smp_processor_id() &&
1224 !(pool->flags & POOL_BH_DRAINING))) {
1225 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1226 return;
1227 }
1228#endif
1229 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1230 raise_softirq_irqoff(HI_SOFTIRQ);
1231 else
1232 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1233}
1234
1235/**
1236 * kick_pool - wake up an idle worker if necessary
1237 * @pool: pool to kick
1238 *
1239 * @pool may have pending work items. Wake up worker if necessary. Returns
1240 * whether a worker was woken up.
1241 */
1242static bool kick_pool(struct worker_pool *pool)
1243{
1244 struct worker *worker = first_idle_worker(pool);
1245 struct task_struct *p;
1246
1247 lockdep_assert_held(&pool->lock);
1248
1249 if (!need_more_worker(pool) || !worker)
1250 return false;
1251
1252 if (pool->flags & POOL_BH) {
1253 kick_bh_pool(pool);
1254 return true;
1255 }
1256
1257 p = worker->task;
1258
1259#ifdef CONFIG_SMP
1260 /*
1261 * Idle @worker is about to execute @work and waking up provides an
1262 * opportunity to migrate @worker at a lower cost by setting the task's
1263 * wake_cpu field. Let's see if we want to move @worker to improve
1264 * execution locality.
1265 *
1266 * We're waking the worker that went idle the latest and there's some
1267 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1268 * so, setting the wake_cpu won't do anything. As this is a best-effort
1269 * optimization and the race window is narrow, let's leave as-is for
1270 * now. If this becomes pronounced, we can skip over workers which are
1271 * still on cpu when picking an idle worker.
1272 *
1273 * If @pool has non-strict affinity, @worker might have ended up outside
1274 * its affinity scope. Repatriate.
1275 */
1276 if (!pool->attrs->affn_strict &&
1277 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1278 struct work_struct *work = list_first_entry(&pool->worklist,
1279 struct work_struct, entry);
1280 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1281 cpu_online_mask);
1282 if (wake_cpu < nr_cpu_ids) {
1283 p->wake_cpu = wake_cpu;
1284 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1285 }
1286 }
1287#endif
1288 wake_up_process(p);
1289 return true;
1290}
1291
1292#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1293
1294/*
1295 * Concurrency-managed per-cpu work items that hog CPU for longer than
1296 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1297 * which prevents them from stalling other concurrency-managed work items. If a
1298 * work function keeps triggering this mechanism, it's likely that the work item
1299 * should be using an unbound workqueue instead.
1300 *
1301 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1302 * and report them so that they can be examined and converted to use unbound
1303 * workqueues as appropriate. To avoid flooding the console, each violating work
1304 * function is tracked and reported with exponential backoff.
1305 */
1306#define WCI_MAX_ENTS 128
1307
1308struct wci_ent {
1309 work_func_t func;
1310 atomic64_t cnt;
1311 struct hlist_node hash_node;
1312};
1313
1314static struct wci_ent wci_ents[WCI_MAX_ENTS];
1315static int wci_nr_ents;
1316static DEFINE_RAW_SPINLOCK(wci_lock);
1317static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1318
1319static struct wci_ent *wci_find_ent(work_func_t func)
1320{
1321 struct wci_ent *ent;
1322
1323 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1324 (unsigned long)func) {
1325 if (ent->func == func)
1326 return ent;
1327 }
1328 return NULL;
1329}
1330
1331static void wq_cpu_intensive_report(work_func_t func)
1332{
1333 struct wci_ent *ent;
1334
1335restart:
1336 ent = wci_find_ent(func);
1337 if (ent) {
1338 u64 cnt;
1339
1340 /*
1341 * Start reporting from the warning_thresh and back off
1342 * exponentially.
1343 */
1344 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1345 if (wq_cpu_intensive_warning_thresh &&
1346 cnt >= wq_cpu_intensive_warning_thresh &&
1347 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1348 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1349 ent->func, wq_cpu_intensive_thresh_us,
1350 atomic64_read(&ent->cnt));
1351 return;
1352 }
1353
1354 /*
1355 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1356 * is exhausted, something went really wrong and we probably made enough
1357 * noise already.
1358 */
1359 if (wci_nr_ents >= WCI_MAX_ENTS)
1360 return;
1361
1362 raw_spin_lock(&wci_lock);
1363
1364 if (wci_nr_ents >= WCI_MAX_ENTS) {
1365 raw_spin_unlock(&wci_lock);
1366 return;
1367 }
1368
1369 if (wci_find_ent(func)) {
1370 raw_spin_unlock(&wci_lock);
1371 goto restart;
1372 }
1373
1374 ent = &wci_ents[wci_nr_ents++];
1375 ent->func = func;
1376 atomic64_set(&ent->cnt, 0);
1377 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1378
1379 raw_spin_unlock(&wci_lock);
1380
1381 goto restart;
1382}
1383
1384#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1385static void wq_cpu_intensive_report(work_func_t func) {}
1386#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1387
1388/**
1389 * wq_worker_running - a worker is running again
1390 * @task: task waking up
1391 *
1392 * This function is called when a worker returns from schedule()
1393 */
1394void wq_worker_running(struct task_struct *task)
1395{
1396 struct worker *worker = kthread_data(task);
1397
1398 if (!READ_ONCE(worker->sleeping))
1399 return;
1400
1401 /*
1402 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1403 * and the nr_running increment below, we may ruin the nr_running reset
1404 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1405 * pool. Protect against such race.
1406 */
1407 preempt_disable();
1408 if (!(worker->flags & WORKER_NOT_RUNNING))
1409 worker->pool->nr_running++;
1410 preempt_enable();
1411
1412 /*
1413 * CPU intensive auto-detection cares about how long a work item hogged
1414 * CPU without sleeping. Reset the starting timestamp on wakeup.
1415 */
1416 worker->current_at = worker->task->se.sum_exec_runtime;
1417
1418 WRITE_ONCE(worker->sleeping, 0);
1419}
1420
1421/**
1422 * wq_worker_sleeping - a worker is going to sleep
1423 * @task: task going to sleep
1424 *
1425 * This function is called from schedule() when a busy worker is
1426 * going to sleep.
1427 */
1428void wq_worker_sleeping(struct task_struct *task)
1429{
1430 struct worker *worker = kthread_data(task);
1431 struct worker_pool *pool;
1432
1433 /*
1434 * Rescuers, which may not have all the fields set up like normal
1435 * workers, also reach here, let's not access anything before
1436 * checking NOT_RUNNING.
1437 */
1438 if (worker->flags & WORKER_NOT_RUNNING)
1439 return;
1440
1441 pool = worker->pool;
1442
1443 /* Return if preempted before wq_worker_running() was reached */
1444 if (READ_ONCE(worker->sleeping))
1445 return;
1446
1447 WRITE_ONCE(worker->sleeping, 1);
1448 raw_spin_lock_irq(&pool->lock);
1449
1450 /*
1451 * Recheck in case unbind_workers() preempted us. We don't
1452 * want to decrement nr_running after the worker is unbound
1453 * and nr_running has been reset.
1454 */
1455 if (worker->flags & WORKER_NOT_RUNNING) {
1456 raw_spin_unlock_irq(&pool->lock);
1457 return;
1458 }
1459
1460 pool->nr_running--;
1461 if (kick_pool(pool))
1462 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1463
1464 raw_spin_unlock_irq(&pool->lock);
1465}
1466
1467/**
1468 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1469 * @task: task currently running
1470 *
1471 * Called from scheduler_tick(). We're in the IRQ context and the current
1472 * worker's fields which follow the 'K' locking rule can be accessed safely.
1473 */
1474void wq_worker_tick(struct task_struct *task)
1475{
1476 struct worker *worker = kthread_data(task);
1477 struct pool_workqueue *pwq = worker->current_pwq;
1478 struct worker_pool *pool = worker->pool;
1479
1480 if (!pwq)
1481 return;
1482
1483 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1484
1485 if (!wq_cpu_intensive_thresh_us)
1486 return;
1487
1488 /*
1489 * If the current worker is concurrency managed and hogged the CPU for
1490 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1491 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1492 *
1493 * Set @worker->sleeping means that @worker is in the process of
1494 * switching out voluntarily and won't be contributing to
1495 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1496 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1497 * double decrements. The task is releasing the CPU anyway. Let's skip.
1498 * We probably want to make this prettier in the future.
1499 */
1500 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1501 worker->task->se.sum_exec_runtime - worker->current_at <
1502 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1503 return;
1504
1505 raw_spin_lock(&pool->lock);
1506
1507 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1508 wq_cpu_intensive_report(worker->current_func);
1509 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1510
1511 if (kick_pool(pool))
1512 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1513
1514 raw_spin_unlock(&pool->lock);
1515}
1516
1517/**
1518 * wq_worker_last_func - retrieve worker's last work function
1519 * @task: Task to retrieve last work function of.
1520 *
1521 * Determine the last function a worker executed. This is called from
1522 * the scheduler to get a worker's last known identity.
1523 *
1524 * CONTEXT:
1525 * raw_spin_lock_irq(rq->lock)
1526 *
1527 * This function is called during schedule() when a kworker is going
1528 * to sleep. It's used by psi to identify aggregation workers during
1529 * dequeuing, to allow periodic aggregation to shut-off when that
1530 * worker is the last task in the system or cgroup to go to sleep.
1531 *
1532 * As this function doesn't involve any workqueue-related locking, it
1533 * only returns stable values when called from inside the scheduler's
1534 * queuing and dequeuing paths, when @task, which must be a kworker,
1535 * is guaranteed to not be processing any works.
1536 *
1537 * Return:
1538 * The last work function %current executed as a worker, NULL if it
1539 * hasn't executed any work yet.
1540 */
1541work_func_t wq_worker_last_func(struct task_struct *task)
1542{
1543 struct worker *worker = kthread_data(task);
1544
1545 return worker->last_func;
1546}
1547
1548/**
1549 * wq_node_nr_active - Determine wq_node_nr_active to use
1550 * @wq: workqueue of interest
1551 * @node: NUMA node, can be %NUMA_NO_NODE
1552 *
1553 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1554 *
1555 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1556 *
1557 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1558 *
1559 * - Otherwise, node_nr_active[@node].
1560 */
1561static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1562 int node)
1563{
1564 if (!(wq->flags & WQ_UNBOUND))
1565 return NULL;
1566
1567 if (node == NUMA_NO_NODE)
1568 node = nr_node_ids;
1569
1570 return wq->node_nr_active[node];
1571}
1572
1573/**
1574 * wq_update_node_max_active - Update per-node max_actives to use
1575 * @wq: workqueue to update
1576 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1577 *
1578 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1579 * distributed among nodes according to the proportions of numbers of online
1580 * cpus. The result is always between @wq->min_active and max_active.
1581 */
1582static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1583{
1584 struct cpumask *effective = unbound_effective_cpumask(wq);
1585 int min_active = READ_ONCE(wq->min_active);
1586 int max_active = READ_ONCE(wq->max_active);
1587 int total_cpus, node;
1588
1589 lockdep_assert_held(&wq->mutex);
1590
1591 if (!wq_topo_initialized)
1592 return;
1593
1594 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1595 off_cpu = -1;
1596
1597 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1598 if (off_cpu >= 0)
1599 total_cpus--;
1600
1601 /* If all CPUs of the wq get offline, use the default values */
1602 if (unlikely(!total_cpus)) {
1603 for_each_node(node)
1604 wq_node_nr_active(wq, node)->max = min_active;
1605
1606 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1607 return;
1608 }
1609
1610 for_each_node(node) {
1611 int node_cpus;
1612
1613 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1614 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1615 node_cpus--;
1616
1617 wq_node_nr_active(wq, node)->max =
1618 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1619 min_active, max_active);
1620 }
1621
1622 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1623}
1624
1625/**
1626 * get_pwq - get an extra reference on the specified pool_workqueue
1627 * @pwq: pool_workqueue to get
1628 *
1629 * Obtain an extra reference on @pwq. The caller should guarantee that
1630 * @pwq has positive refcnt and be holding the matching pool->lock.
1631 */
1632static void get_pwq(struct pool_workqueue *pwq)
1633{
1634 lockdep_assert_held(&pwq->pool->lock);
1635 WARN_ON_ONCE(pwq->refcnt <= 0);
1636 pwq->refcnt++;
1637}
1638
1639/**
1640 * put_pwq - put a pool_workqueue reference
1641 * @pwq: pool_workqueue to put
1642 *
1643 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1644 * destruction. The caller should be holding the matching pool->lock.
1645 */
1646static void put_pwq(struct pool_workqueue *pwq)
1647{
1648 lockdep_assert_held(&pwq->pool->lock);
1649 if (likely(--pwq->refcnt))
1650 return;
1651 /*
1652 * @pwq can't be released under pool->lock, bounce to a dedicated
1653 * kthread_worker to avoid A-A deadlocks.
1654 */
1655 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1656}
1657
1658/**
1659 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1660 * @pwq: pool_workqueue to put (can be %NULL)
1661 *
1662 * put_pwq() with locking. This function also allows %NULL @pwq.
1663 */
1664static void put_pwq_unlocked(struct pool_workqueue *pwq)
1665{
1666 if (pwq) {
1667 /*
1668 * As both pwqs and pools are RCU protected, the
1669 * following lock operations are safe.
1670 */
1671 raw_spin_lock_irq(&pwq->pool->lock);
1672 put_pwq(pwq);
1673 raw_spin_unlock_irq(&pwq->pool->lock);
1674 }
1675}
1676
1677static bool pwq_is_empty(struct pool_workqueue *pwq)
1678{
1679 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1680}
1681
1682static void __pwq_activate_work(struct pool_workqueue *pwq,
1683 struct work_struct *work)
1684{
1685 unsigned long *wdb = work_data_bits(work);
1686
1687 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1688 trace_workqueue_activate_work(work);
1689 if (list_empty(&pwq->pool->worklist))
1690 pwq->pool->watchdog_ts = jiffies;
1691 move_linked_works(work, &pwq->pool->worklist, NULL);
1692 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1693}
1694
1695/**
1696 * pwq_activate_work - Activate a work item if inactive
1697 * @pwq: pool_workqueue @work belongs to
1698 * @work: work item to activate
1699 *
1700 * Returns %true if activated. %false if already active.
1701 */
1702static bool pwq_activate_work(struct pool_workqueue *pwq,
1703 struct work_struct *work)
1704{
1705 struct worker_pool *pool = pwq->pool;
1706 struct wq_node_nr_active *nna;
1707
1708 lockdep_assert_held(&pool->lock);
1709
1710 if (!(*work_data_bits(work) & WORK_STRUCT_INACTIVE))
1711 return false;
1712
1713 nna = wq_node_nr_active(pwq->wq, pool->node);
1714 if (nna)
1715 atomic_inc(&nna->nr);
1716
1717 pwq->nr_active++;
1718 __pwq_activate_work(pwq, work);
1719 return true;
1720}
1721
1722static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1723{
1724 int max = READ_ONCE(nna->max);
1725
1726 while (true) {
1727 int old, tmp;
1728
1729 old = atomic_read(&nna->nr);
1730 if (old >= max)
1731 return false;
1732 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1733 if (tmp == old)
1734 return true;
1735 }
1736}
1737
1738/**
1739 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1740 * @pwq: pool_workqueue of interest
1741 * @fill: max_active may have increased, try to increase concurrency level
1742 *
1743 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1744 * successfully obtained. %false otherwise.
1745 */
1746static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1747{
1748 struct workqueue_struct *wq = pwq->wq;
1749 struct worker_pool *pool = pwq->pool;
1750 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1751 bool obtained = false;
1752
1753 lockdep_assert_held(&pool->lock);
1754
1755 if (!nna) {
1756 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1757 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1758 goto out;
1759 }
1760
1761 if (unlikely(pwq->plugged))
1762 return false;
1763
1764 /*
1765 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1766 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1767 * concurrency level. Don't jump the line.
1768 *
1769 * We need to ignore the pending test after max_active has increased as
1770 * pwq_dec_nr_active() can only maintain the concurrency level but not
1771 * increase it. This is indicated by @fill.
1772 */
1773 if (!list_empty(&pwq->pending_node) && likely(!fill))
1774 goto out;
1775
1776 obtained = tryinc_node_nr_active(nna);
1777 if (obtained)
1778 goto out;
1779
1780 /*
1781 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1782 * and try again. The smp_mb() is paired with the implied memory barrier
1783 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1784 * we see the decremented $nna->nr or they see non-empty
1785 * $nna->pending_pwqs.
1786 */
1787 raw_spin_lock(&nna->lock);
1788
1789 if (list_empty(&pwq->pending_node))
1790 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1791 else if (likely(!fill))
1792 goto out_unlock;
1793
1794 smp_mb();
1795
1796 obtained = tryinc_node_nr_active(nna);
1797
1798 /*
1799 * If @fill, @pwq might have already been pending. Being spuriously
1800 * pending in cold paths doesn't affect anything. Let's leave it be.
1801 */
1802 if (obtained && likely(!fill))
1803 list_del_init(&pwq->pending_node);
1804
1805out_unlock:
1806 raw_spin_unlock(&nna->lock);
1807out:
1808 if (obtained)
1809 pwq->nr_active++;
1810 return obtained;
1811}
1812
1813/**
1814 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1815 * @pwq: pool_workqueue of interest
1816 * @fill: max_active may have increased, try to increase concurrency level
1817 *
1818 * Activate the first inactive work item of @pwq if available and allowed by
1819 * max_active limit.
1820 *
1821 * Returns %true if an inactive work item has been activated. %false if no
1822 * inactive work item is found or max_active limit is reached.
1823 */
1824static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1825{
1826 struct work_struct *work =
1827 list_first_entry_or_null(&pwq->inactive_works,
1828 struct work_struct, entry);
1829
1830 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1831 __pwq_activate_work(pwq, work);
1832 return true;
1833 } else {
1834 return false;
1835 }
1836}
1837
1838/**
1839 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1840 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1841 *
1842 * This function should only be called for ordered workqueues where only the
1843 * oldest pwq is unplugged, the others are plugged to suspend execution to
1844 * ensure proper work item ordering::
1845 *
1846 * dfl_pwq --------------+ [P] - plugged
1847 * |
1848 * v
1849 * pwqs -> A -> B [P] -> C [P] (newest)
1850 * | | |
1851 * 1 3 5
1852 * | | |
1853 * 2 4 6
1854 *
1855 * When the oldest pwq is drained and removed, this function should be called
1856 * to unplug the next oldest one to start its work item execution. Note that
1857 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1858 * the list is the oldest.
1859 */
1860static void unplug_oldest_pwq(struct workqueue_struct *wq)
1861{
1862 struct pool_workqueue *pwq;
1863
1864 lockdep_assert_held(&wq->mutex);
1865
1866 /* Caller should make sure that pwqs isn't empty before calling */
1867 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1868 pwqs_node);
1869 raw_spin_lock_irq(&pwq->pool->lock);
1870 if (pwq->plugged) {
1871 pwq->plugged = false;
1872 if (pwq_activate_first_inactive(pwq, true))
1873 kick_pool(pwq->pool);
1874 }
1875 raw_spin_unlock_irq(&pwq->pool->lock);
1876}
1877
1878/**
1879 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1880 * @nna: wq_node_nr_active to activate a pending pwq for
1881 * @caller_pool: worker_pool the caller is locking
1882 *
1883 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1884 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1885 */
1886static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1887 struct worker_pool *caller_pool)
1888{
1889 struct worker_pool *locked_pool = caller_pool;
1890 struct pool_workqueue *pwq;
1891 struct work_struct *work;
1892
1893 lockdep_assert_held(&caller_pool->lock);
1894
1895 raw_spin_lock(&nna->lock);
1896retry:
1897 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1898 struct pool_workqueue, pending_node);
1899 if (!pwq)
1900 goto out_unlock;
1901
1902 /*
1903 * If @pwq is for a different pool than @locked_pool, we need to lock
1904 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1905 * / lock dance. For that, we also need to release @nna->lock as it's
1906 * nested inside pool locks.
1907 */
1908 if (pwq->pool != locked_pool) {
1909 raw_spin_unlock(&locked_pool->lock);
1910 locked_pool = pwq->pool;
1911 if (!raw_spin_trylock(&locked_pool->lock)) {
1912 raw_spin_unlock(&nna->lock);
1913 raw_spin_lock(&locked_pool->lock);
1914 raw_spin_lock(&nna->lock);
1915 goto retry;
1916 }
1917 }
1918
1919 /*
1920 * $pwq may not have any inactive work items due to e.g. cancellations.
1921 * Drop it from pending_pwqs and see if there's another one.
1922 */
1923 work = list_first_entry_or_null(&pwq->inactive_works,
1924 struct work_struct, entry);
1925 if (!work) {
1926 list_del_init(&pwq->pending_node);
1927 goto retry;
1928 }
1929
1930 /*
1931 * Acquire an nr_active count and activate the inactive work item. If
1932 * $pwq still has inactive work items, rotate it to the end of the
1933 * pending_pwqs so that we round-robin through them. This means that
1934 * inactive work items are not activated in queueing order which is fine
1935 * given that there has never been any ordering across different pwqs.
1936 */
1937 if (likely(tryinc_node_nr_active(nna))) {
1938 pwq->nr_active++;
1939 __pwq_activate_work(pwq, work);
1940
1941 if (list_empty(&pwq->inactive_works))
1942 list_del_init(&pwq->pending_node);
1943 else
1944 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1945
1946 /* if activating a foreign pool, make sure it's running */
1947 if (pwq->pool != caller_pool)
1948 kick_pool(pwq->pool);
1949 }
1950
1951out_unlock:
1952 raw_spin_unlock(&nna->lock);
1953 if (locked_pool != caller_pool) {
1954 raw_spin_unlock(&locked_pool->lock);
1955 raw_spin_lock(&caller_pool->lock);
1956 }
1957}
1958
1959/**
1960 * pwq_dec_nr_active - Retire an active count
1961 * @pwq: pool_workqueue of interest
1962 *
1963 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1964 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1965 */
1966static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1967{
1968 struct worker_pool *pool = pwq->pool;
1969 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1970
1971 lockdep_assert_held(&pool->lock);
1972
1973 /*
1974 * @pwq->nr_active should be decremented for both percpu and unbound
1975 * workqueues.
1976 */
1977 pwq->nr_active--;
1978
1979 /*
1980 * For a percpu workqueue, it's simple. Just need to kick the first
1981 * inactive work item on @pwq itself.
1982 */
1983 if (!nna) {
1984 pwq_activate_first_inactive(pwq, false);
1985 return;
1986 }
1987
1988 /*
1989 * If @pwq is for an unbound workqueue, it's more complicated because
1990 * multiple pwqs and pools may be sharing the nr_active count. When a
1991 * pwq needs to wait for an nr_active count, it puts itself on
1992 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1993 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1994 * guarantee that either we see non-empty pending_pwqs or they see
1995 * decremented $nna->nr.
1996 *
1997 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1998 * max_active gets updated. However, it is guaranteed to be equal to or
1999 * larger than @pwq->wq->min_active which is above zero unless freezing.
2000 * This maintains the forward progress guarantee.
2001 */
2002 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
2003 return;
2004
2005 if (!list_empty(&nna->pending_pwqs))
2006 node_activate_pending_pwq(nna, pool);
2007}
2008
2009/**
2010 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
2011 * @pwq: pwq of interest
2012 * @work_data: work_data of work which left the queue
2013 *
2014 * A work either has completed or is removed from pending queue,
2015 * decrement nr_in_flight of its pwq and handle workqueue flushing.
2016 *
2017 * NOTE:
2018 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
2019 * and thus should be called after all other state updates for the in-flight
2020 * work item is complete.
2021 *
2022 * CONTEXT:
2023 * raw_spin_lock_irq(pool->lock).
2024 */
2025static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
2026{
2027 int color = get_work_color(work_data);
2028
2029 if (!(work_data & WORK_STRUCT_INACTIVE))
2030 pwq_dec_nr_active(pwq);
2031
2032 pwq->nr_in_flight[color]--;
2033
2034 /* is flush in progress and are we at the flushing tip? */
2035 if (likely(pwq->flush_color != color))
2036 goto out_put;
2037
2038 /* are there still in-flight works? */
2039 if (pwq->nr_in_flight[color])
2040 goto out_put;
2041
2042 /* this pwq is done, clear flush_color */
2043 pwq->flush_color = -1;
2044
2045 /*
2046 * If this was the last pwq, wake up the first flusher. It
2047 * will handle the rest.
2048 */
2049 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2050 complete(&pwq->wq->first_flusher->done);
2051out_put:
2052 put_pwq(pwq);
2053}
2054
2055/**
2056 * try_to_grab_pending - steal work item from worklist and disable irq
2057 * @work: work item to steal
2058 * @cflags: %WORK_CANCEL_ flags
2059 * @irq_flags: place to store irq state
2060 *
2061 * Try to grab PENDING bit of @work. This function can handle @work in any
2062 * stable state - idle, on timer or on worklist.
2063 *
2064 * Return:
2065 *
2066 * ======== ================================================================
2067 * 1 if @work was pending and we successfully stole PENDING
2068 * 0 if @work was idle and we claimed PENDING
2069 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2070 * -ENOENT if someone else is canceling @work, this state may persist
2071 * for arbitrarily long
2072 * ======== ================================================================
2073 *
2074 * Note:
2075 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2076 * interrupted while holding PENDING and @work off queue, irq must be
2077 * disabled on entry. This, combined with delayed_work->timer being
2078 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2079 *
2080 * On successful return, >= 0, irq is disabled and the caller is
2081 * responsible for releasing it using local_irq_restore(*@irq_flags).
2082 *
2083 * This function is safe to call from any context including IRQ handler.
2084 */
2085static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2086 unsigned long *irq_flags)
2087{
2088 struct worker_pool *pool;
2089 struct pool_workqueue *pwq;
2090
2091 local_irq_save(*irq_flags);
2092
2093 /* try to steal the timer if it exists */
2094 if (cflags & WORK_CANCEL_DELAYED) {
2095 struct delayed_work *dwork = to_delayed_work(work);
2096
2097 /*
2098 * dwork->timer is irqsafe. If del_timer() fails, it's
2099 * guaranteed that the timer is not queued anywhere and not
2100 * running on the local CPU.
2101 */
2102 if (likely(del_timer(&dwork->timer)))
2103 return 1;
2104 }
2105
2106 /* try to claim PENDING the normal way */
2107 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2108 return 0;
2109
2110 rcu_read_lock();
2111 /*
2112 * The queueing is in progress, or it is already queued. Try to
2113 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2114 */
2115 pool = get_work_pool(work);
2116 if (!pool)
2117 goto fail;
2118
2119 raw_spin_lock(&pool->lock);
2120 /*
2121 * work->data is guaranteed to point to pwq only while the work
2122 * item is queued on pwq->wq, and both updating work->data to point
2123 * to pwq on queueing and to pool on dequeueing are done under
2124 * pwq->pool->lock. This in turn guarantees that, if work->data
2125 * points to pwq which is associated with a locked pool, the work
2126 * item is currently queued on that pool.
2127 */
2128 pwq = get_work_pwq(work);
2129 if (pwq && pwq->pool == pool) {
2130 unsigned long work_data;
2131
2132 debug_work_deactivate(work);
2133
2134 /*
2135 * A cancelable inactive work item must be in the
2136 * pwq->inactive_works since a queued barrier can't be
2137 * canceled (see the comments in insert_wq_barrier()).
2138 *
2139 * An inactive work item cannot be grabbed directly because
2140 * it might have linked barrier work items which, if left
2141 * on the inactive_works list, will confuse pwq->nr_active
2142 * management later on and cause stall. Make sure the work
2143 * item is activated before grabbing.
2144 */
2145 pwq_activate_work(pwq, work);
2146
2147 list_del_init(&work->entry);
2148
2149 /*
2150 * work->data points to pwq iff queued. Let's point to pool. As
2151 * this destroys work->data needed by the next step, stash it.
2152 */
2153 work_data = *work_data_bits(work);
2154 set_work_pool_and_keep_pending(work, pool->id, 0);
2155
2156 /* must be the last step, see the function comment */
2157 pwq_dec_nr_in_flight(pwq, work_data);
2158
2159 raw_spin_unlock(&pool->lock);
2160 rcu_read_unlock();
2161 return 1;
2162 }
2163 raw_spin_unlock(&pool->lock);
2164fail:
2165 rcu_read_unlock();
2166 local_irq_restore(*irq_flags);
2167 if (work_is_canceling(work))
2168 return -ENOENT;
2169 cpu_relax();
2170 return -EAGAIN;
2171}
2172
2173struct cwt_wait {
2174 wait_queue_entry_t wait;
2175 struct work_struct *work;
2176};
2177
2178static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
2179{
2180 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2181
2182 if (cwait->work != key)
2183 return 0;
2184 return autoremove_wake_function(wait, mode, sync, key);
2185}
2186
2187/**
2188 * work_grab_pending - steal work item from worklist and disable irq
2189 * @work: work item to steal
2190 * @cflags: %WORK_CANCEL_ flags
2191 * @irq_flags: place to store IRQ state
2192 *
2193 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2194 * or on worklist.
2195 *
2196 * Must be called in process context. IRQ is disabled on return with IRQ state
2197 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2198 * local_irq_restore().
2199 *
2200 * Returns %true if @work was pending. %false if idle.
2201 */
2202static bool work_grab_pending(struct work_struct *work, u32 cflags,
2203 unsigned long *irq_flags)
2204{
2205 struct cwt_wait cwait;
2206 int ret;
2207
2208 might_sleep();
2209repeat:
2210 ret = try_to_grab_pending(work, cflags, irq_flags);
2211 if (likely(ret >= 0))
2212 return ret;
2213 if (ret != -ENOENT)
2214 goto repeat;
2215
2216 /*
2217 * Someone is already canceling. Wait for it to finish. flush_work()
2218 * doesn't work for PREEMPT_NONE because we may get woken up between
2219 * @work's completion and the other canceling task resuming and clearing
2220 * CANCELING - flush_work() will return false immediately as @work is no
2221 * longer busy, try_to_grab_pending() will return -ENOENT as @work is
2222 * still being canceled and the other canceling task won't be able to
2223 * clear CANCELING as we're hogging the CPU.
2224 *
2225 * Let's wait for completion using a waitqueue. As this may lead to the
2226 * thundering herd problem, use a custom wake function which matches
2227 * @work along with exclusive wait and wakeup.
2228 */
2229 init_wait(&cwait.wait);
2230 cwait.wait.func = cwt_wakefn;
2231 cwait.work = work;
2232
2233 prepare_to_wait_exclusive(&wq_cancel_waitq, &cwait.wait,
2234 TASK_UNINTERRUPTIBLE);
2235 if (work_is_canceling(work))
2236 schedule();
2237 finish_wait(&wq_cancel_waitq, &cwait.wait);
2238
2239 goto repeat;
2240}
2241
2242/**
2243 * insert_work - insert a work into a pool
2244 * @pwq: pwq @work belongs to
2245 * @work: work to insert
2246 * @head: insertion point
2247 * @extra_flags: extra WORK_STRUCT_* flags to set
2248 *
2249 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2250 * work_struct flags.
2251 *
2252 * CONTEXT:
2253 * raw_spin_lock_irq(pool->lock).
2254 */
2255static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2256 struct list_head *head, unsigned int extra_flags)
2257{
2258 debug_work_activate(work);
2259
2260 /* record the work call stack in order to print it in KASAN reports */
2261 kasan_record_aux_stack_noalloc(work);
2262
2263 /* we own @work, set data and link */
2264 set_work_pwq(work, pwq, extra_flags);
2265 list_add_tail(&work->entry, head);
2266 get_pwq(pwq);
2267}
2268
2269/*
2270 * Test whether @work is being queued from another work executing on the
2271 * same workqueue.
2272 */
2273static bool is_chained_work(struct workqueue_struct *wq)
2274{
2275 struct worker *worker;
2276
2277 worker = current_wq_worker();
2278 /*
2279 * Return %true iff I'm a worker executing a work item on @wq. If
2280 * I'm @worker, it's safe to dereference it without locking.
2281 */
2282 return worker && worker->current_pwq->wq == wq;
2283}
2284
2285/*
2286 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2287 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2288 * avoid perturbing sensitive tasks.
2289 */
2290static int wq_select_unbound_cpu(int cpu)
2291{
2292 int new_cpu;
2293
2294 if (likely(!wq_debug_force_rr_cpu)) {
2295 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2296 return cpu;
2297 } else {
2298 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2299 }
2300
2301 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2302 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2303 if (unlikely(new_cpu >= nr_cpu_ids)) {
2304 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2305 if (unlikely(new_cpu >= nr_cpu_ids))
2306 return cpu;
2307 }
2308 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2309
2310 return new_cpu;
2311}
2312
2313static void __queue_work(int cpu, struct workqueue_struct *wq,
2314 struct work_struct *work)
2315{
2316 struct pool_workqueue *pwq;
2317 struct worker_pool *last_pool, *pool;
2318 unsigned int work_flags;
2319 unsigned int req_cpu = cpu;
2320
2321 /*
2322 * While a work item is PENDING && off queue, a task trying to
2323 * steal the PENDING will busy-loop waiting for it to either get
2324 * queued or lose PENDING. Grabbing PENDING and queueing should
2325 * happen with IRQ disabled.
2326 */
2327 lockdep_assert_irqs_disabled();
2328
2329 /*
2330 * For a draining wq, only works from the same workqueue are
2331 * allowed. The __WQ_DESTROYING helps to spot the issue that
2332 * queues a new work item to a wq after destroy_workqueue(wq).
2333 */
2334 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2335 WARN_ON_ONCE(!is_chained_work(wq))))
2336 return;
2337 rcu_read_lock();
2338retry:
2339 /* pwq which will be used unless @work is executing elsewhere */
2340 if (req_cpu == WORK_CPU_UNBOUND) {
2341 if (wq->flags & WQ_UNBOUND)
2342 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2343 else
2344 cpu = raw_smp_processor_id();
2345 }
2346
2347 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2348 pool = pwq->pool;
2349
2350 /*
2351 * If @work was previously on a different pool, it might still be
2352 * running there, in which case the work needs to be queued on that
2353 * pool to guarantee non-reentrancy.
2354 */
2355 last_pool = get_work_pool(work);
2356 if (last_pool && last_pool != pool) {
2357 struct worker *worker;
2358
2359 raw_spin_lock(&last_pool->lock);
2360
2361 worker = find_worker_executing_work(last_pool, work);
2362
2363 if (worker && worker->current_pwq->wq == wq) {
2364 pwq = worker->current_pwq;
2365 pool = pwq->pool;
2366 WARN_ON_ONCE(pool != last_pool);
2367 } else {
2368 /* meh... not running there, queue here */
2369 raw_spin_unlock(&last_pool->lock);
2370 raw_spin_lock(&pool->lock);
2371 }
2372 } else {
2373 raw_spin_lock(&pool->lock);
2374 }
2375
2376 /*
2377 * pwq is determined and locked. For unbound pools, we could have raced
2378 * with pwq release and it could already be dead. If its refcnt is zero,
2379 * repeat pwq selection. Note that unbound pwqs never die without
2380 * another pwq replacing it in cpu_pwq or while work items are executing
2381 * on it, so the retrying is guaranteed to make forward-progress.
2382 */
2383 if (unlikely(!pwq->refcnt)) {
2384 if (wq->flags & WQ_UNBOUND) {
2385 raw_spin_unlock(&pool->lock);
2386 cpu_relax();
2387 goto retry;
2388 }
2389 /* oops */
2390 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2391 wq->name, cpu);
2392 }
2393
2394 /* pwq determined, queue */
2395 trace_workqueue_queue_work(req_cpu, pwq, work);
2396
2397 if (WARN_ON(!list_empty(&work->entry)))
2398 goto out;
2399
2400 pwq->nr_in_flight[pwq->work_color]++;
2401 work_flags = work_color_to_flags(pwq->work_color);
2402
2403 /*
2404 * Limit the number of concurrently active work items to max_active.
2405 * @work must also queue behind existing inactive work items to maintain
2406 * ordering when max_active changes. See wq_adjust_max_active().
2407 */
2408 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2409 if (list_empty(&pool->worklist))
2410 pool->watchdog_ts = jiffies;
2411
2412 trace_workqueue_activate_work(work);
2413 insert_work(pwq, work, &pool->worklist, work_flags);
2414 kick_pool(pool);
2415 } else {
2416 work_flags |= WORK_STRUCT_INACTIVE;
2417 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2418 }
2419
2420out:
2421 raw_spin_unlock(&pool->lock);
2422 rcu_read_unlock();
2423}
2424
2425/**
2426 * queue_work_on - queue work on specific cpu
2427 * @cpu: CPU number to execute work on
2428 * @wq: workqueue to use
2429 * @work: work to queue
2430 *
2431 * We queue the work to a specific CPU, the caller must ensure it
2432 * can't go away. Callers that fail to ensure that the specified
2433 * CPU cannot go away will execute on a randomly chosen CPU.
2434 * But note well that callers specifying a CPU that never has been
2435 * online will get a splat.
2436 *
2437 * Return: %false if @work was already on a queue, %true otherwise.
2438 */
2439bool queue_work_on(int cpu, struct workqueue_struct *wq,
2440 struct work_struct *work)
2441{
2442 bool ret = false;
2443 unsigned long irq_flags;
2444
2445 local_irq_save(irq_flags);
2446
2447 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2448 __queue_work(cpu, wq, work);
2449 ret = true;
2450 }
2451
2452 local_irq_restore(irq_flags);
2453 return ret;
2454}
2455EXPORT_SYMBOL(queue_work_on);
2456
2457/**
2458 * select_numa_node_cpu - Select a CPU based on NUMA node
2459 * @node: NUMA node ID that we want to select a CPU from
2460 *
2461 * This function will attempt to find a "random" cpu available on a given
2462 * node. If there are no CPUs available on the given node it will return
2463 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2464 * available CPU if we need to schedule this work.
2465 */
2466static int select_numa_node_cpu(int node)
2467{
2468 int cpu;
2469
2470 /* Delay binding to CPU if node is not valid or online */
2471 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2472 return WORK_CPU_UNBOUND;
2473
2474 /* Use local node/cpu if we are already there */
2475 cpu = raw_smp_processor_id();
2476 if (node == cpu_to_node(cpu))
2477 return cpu;
2478
2479 /* Use "random" otherwise know as "first" online CPU of node */
2480 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2481
2482 /* If CPU is valid return that, otherwise just defer */
2483 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2484}
2485
2486/**
2487 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2488 * @node: NUMA node that we are targeting the work for
2489 * @wq: workqueue to use
2490 * @work: work to queue
2491 *
2492 * We queue the work to a "random" CPU within a given NUMA node. The basic
2493 * idea here is to provide a way to somehow associate work with a given
2494 * NUMA node.
2495 *
2496 * This function will only make a best effort attempt at getting this onto
2497 * the right NUMA node. If no node is requested or the requested node is
2498 * offline then we just fall back to standard queue_work behavior.
2499 *
2500 * Currently the "random" CPU ends up being the first available CPU in the
2501 * intersection of cpu_online_mask and the cpumask of the node, unless we
2502 * are running on the node. In that case we just use the current CPU.
2503 *
2504 * Return: %false if @work was already on a queue, %true otherwise.
2505 */
2506bool queue_work_node(int node, struct workqueue_struct *wq,
2507 struct work_struct *work)
2508{
2509 unsigned long irq_flags;
2510 bool ret = false;
2511
2512 /*
2513 * This current implementation is specific to unbound workqueues.
2514 * Specifically we only return the first available CPU for a given
2515 * node instead of cycling through individual CPUs within the node.
2516 *
2517 * If this is used with a per-cpu workqueue then the logic in
2518 * workqueue_select_cpu_near would need to be updated to allow for
2519 * some round robin type logic.
2520 */
2521 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2522
2523 local_irq_save(irq_flags);
2524
2525 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2526 int cpu = select_numa_node_cpu(node);
2527
2528 __queue_work(cpu, wq, work);
2529 ret = true;
2530 }
2531
2532 local_irq_restore(irq_flags);
2533 return ret;
2534}
2535EXPORT_SYMBOL_GPL(queue_work_node);
2536
2537void delayed_work_timer_fn(struct timer_list *t)
2538{
2539 struct delayed_work *dwork = from_timer(dwork, t, timer);
2540
2541 /* should have been called from irqsafe timer with irq already off */
2542 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2543}
2544EXPORT_SYMBOL(delayed_work_timer_fn);
2545
2546static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2547 struct delayed_work *dwork, unsigned long delay)
2548{
2549 struct timer_list *timer = &dwork->timer;
2550 struct work_struct *work = &dwork->work;
2551
2552 WARN_ON_ONCE(!wq);
2553 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2554 WARN_ON_ONCE(timer_pending(timer));
2555 WARN_ON_ONCE(!list_empty(&work->entry));
2556
2557 /*
2558 * If @delay is 0, queue @dwork->work immediately. This is for
2559 * both optimization and correctness. The earliest @timer can
2560 * expire is on the closest next tick and delayed_work users depend
2561 * on that there's no such delay when @delay is 0.
2562 */
2563 if (!delay) {
2564 __queue_work(cpu, wq, &dwork->work);
2565 return;
2566 }
2567
2568 dwork->wq = wq;
2569 dwork->cpu = cpu;
2570 timer->expires = jiffies + delay;
2571
2572 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2573 /* If the current cpu is a housekeeping cpu, use it. */
2574 cpu = smp_processor_id();
2575 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2576 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2577 add_timer_on(timer, cpu);
2578 } else {
2579 if (likely(cpu == WORK_CPU_UNBOUND))
2580 add_timer_global(timer);
2581 else
2582 add_timer_on(timer, cpu);
2583 }
2584}
2585
2586/**
2587 * queue_delayed_work_on - queue work on specific CPU after delay
2588 * @cpu: CPU number to execute work on
2589 * @wq: workqueue to use
2590 * @dwork: work to queue
2591 * @delay: number of jiffies to wait before queueing
2592 *
2593 * Return: %false if @work was already on a queue, %true otherwise. If
2594 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2595 * execution.
2596 */
2597bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2598 struct delayed_work *dwork, unsigned long delay)
2599{
2600 struct work_struct *work = &dwork->work;
2601 bool ret = false;
2602 unsigned long irq_flags;
2603
2604 /* read the comment in __queue_work() */
2605 local_irq_save(irq_flags);
2606
2607 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2608 __queue_delayed_work(cpu, wq, dwork, delay);
2609 ret = true;
2610 }
2611
2612 local_irq_restore(irq_flags);
2613 return ret;
2614}
2615EXPORT_SYMBOL(queue_delayed_work_on);
2616
2617/**
2618 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2619 * @cpu: CPU number to execute work on
2620 * @wq: workqueue to use
2621 * @dwork: work to queue
2622 * @delay: number of jiffies to wait before queueing
2623 *
2624 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2625 * modify @dwork's timer so that it expires after @delay. If @delay is
2626 * zero, @work is guaranteed to be scheduled immediately regardless of its
2627 * current state.
2628 *
2629 * Return: %false if @dwork was idle and queued, %true if @dwork was
2630 * pending and its timer was modified.
2631 *
2632 * This function is safe to call from any context including IRQ handler.
2633 * See try_to_grab_pending() for details.
2634 */
2635bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2636 struct delayed_work *dwork, unsigned long delay)
2637{
2638 unsigned long irq_flags;
2639 int ret;
2640
2641 do {
2642 ret = try_to_grab_pending(&dwork->work, WORK_CANCEL_DELAYED,
2643 &irq_flags);
2644 } while (unlikely(ret == -EAGAIN));
2645
2646 if (likely(ret >= 0)) {
2647 __queue_delayed_work(cpu, wq, dwork, delay);
2648 local_irq_restore(irq_flags);
2649 }
2650
2651 /* -ENOENT from try_to_grab_pending() becomes %true */
2652 return ret;
2653}
2654EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2655
2656static void rcu_work_rcufn(struct rcu_head *rcu)
2657{
2658 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2659
2660 /* read the comment in __queue_work() */
2661 local_irq_disable();
2662 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2663 local_irq_enable();
2664}
2665
2666/**
2667 * queue_rcu_work - queue work after a RCU grace period
2668 * @wq: workqueue to use
2669 * @rwork: work to queue
2670 *
2671 * Return: %false if @rwork was already pending, %true otherwise. Note
2672 * that a full RCU grace period is guaranteed only after a %true return.
2673 * While @rwork is guaranteed to be executed after a %false return, the
2674 * execution may happen before a full RCU grace period has passed.
2675 */
2676bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2677{
2678 struct work_struct *work = &rwork->work;
2679
2680 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2681 rwork->wq = wq;
2682 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2683 return true;
2684 }
2685
2686 return false;
2687}
2688EXPORT_SYMBOL(queue_rcu_work);
2689
2690static struct worker *alloc_worker(int node)
2691{
2692 struct worker *worker;
2693
2694 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2695 if (worker) {
2696 INIT_LIST_HEAD(&worker->entry);
2697 INIT_LIST_HEAD(&worker->scheduled);
2698 INIT_LIST_HEAD(&worker->node);
2699 /* on creation a worker is in !idle && prep state */
2700 worker->flags = WORKER_PREP;
2701 }
2702 return worker;
2703}
2704
2705static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2706{
2707 if (pool->cpu < 0 && pool->attrs->affn_strict)
2708 return pool->attrs->__pod_cpumask;
2709 else
2710 return pool->attrs->cpumask;
2711}
2712
2713/**
2714 * worker_attach_to_pool() - attach a worker to a pool
2715 * @worker: worker to be attached
2716 * @pool: the target pool
2717 *
2718 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2719 * cpu-binding of @worker are kept coordinated with the pool across
2720 * cpu-[un]hotplugs.
2721 */
2722static void worker_attach_to_pool(struct worker *worker,
2723 struct worker_pool *pool)
2724{
2725 mutex_lock(&wq_pool_attach_mutex);
2726
2727 /*
2728 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2729 * across this function. See the comments above the flag definition for
2730 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2731 */
2732 if (pool->flags & POOL_DISASSOCIATED) {
2733 worker->flags |= WORKER_UNBOUND;
2734 } else {
2735 WARN_ON_ONCE(pool->flags & POOL_BH);
2736 kthread_set_per_cpu(worker->task, pool->cpu);
2737 }
2738
2739 if (worker->rescue_wq)
2740 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2741
2742 list_add_tail(&worker->node, &pool->workers);
2743 worker->pool = pool;
2744
2745 mutex_unlock(&wq_pool_attach_mutex);
2746}
2747
2748/**
2749 * worker_detach_from_pool() - detach a worker from its pool
2750 * @worker: worker which is attached to its pool
2751 *
2752 * Undo the attaching which had been done in worker_attach_to_pool(). The
2753 * caller worker shouldn't access to the pool after detached except it has
2754 * other reference to the pool.
2755 */
2756static void worker_detach_from_pool(struct worker *worker)
2757{
2758 struct worker_pool *pool = worker->pool;
2759 struct completion *detach_completion = NULL;
2760
2761 /* there is one permanent BH worker per CPU which should never detach */
2762 WARN_ON_ONCE(pool->flags & POOL_BH);
2763
2764 mutex_lock(&wq_pool_attach_mutex);
2765
2766 kthread_set_per_cpu(worker->task, -1);
2767 list_del(&worker->node);
2768 worker->pool = NULL;
2769
2770 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2771 detach_completion = pool->detach_completion;
2772 mutex_unlock(&wq_pool_attach_mutex);
2773
2774 /* clear leftover flags without pool->lock after it is detached */
2775 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2776
2777 if (detach_completion)
2778 complete(detach_completion);
2779}
2780
2781/**
2782 * create_worker - create a new workqueue worker
2783 * @pool: pool the new worker will belong to
2784 *
2785 * Create and start a new worker which is attached to @pool.
2786 *
2787 * CONTEXT:
2788 * Might sleep. Does GFP_KERNEL allocations.
2789 *
2790 * Return:
2791 * Pointer to the newly created worker.
2792 */
2793static struct worker *create_worker(struct worker_pool *pool)
2794{
2795 struct worker *worker;
2796 int id;
2797 char id_buf[23];
2798
2799 /* ID is needed to determine kthread name */
2800 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2801 if (id < 0) {
2802 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2803 ERR_PTR(id));
2804 return NULL;
2805 }
2806
2807 worker = alloc_worker(pool->node);
2808 if (!worker) {
2809 pr_err_once("workqueue: Failed to allocate a worker\n");
2810 goto fail;
2811 }
2812
2813 worker->id = id;
2814
2815 if (!(pool->flags & POOL_BH)) {
2816 if (pool->cpu >= 0)
2817 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2818 pool->attrs->nice < 0 ? "H" : "");
2819 else
2820 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2821
2822 worker->task = kthread_create_on_node(worker_thread, worker,
2823 pool->node, "kworker/%s", id_buf);
2824 if (IS_ERR(worker->task)) {
2825 if (PTR_ERR(worker->task) == -EINTR) {
2826 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2827 id_buf);
2828 } else {
2829 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2830 worker->task);
2831 }
2832 goto fail;
2833 }
2834
2835 set_user_nice(worker->task, pool->attrs->nice);
2836 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2837 }
2838
2839 /* successful, attach the worker to the pool */
2840 worker_attach_to_pool(worker, pool);
2841
2842 /* start the newly created worker */
2843 raw_spin_lock_irq(&pool->lock);
2844
2845 worker->pool->nr_workers++;
2846 worker_enter_idle(worker);
2847
2848 /*
2849 * @worker is waiting on a completion in kthread() and will trigger hung
2850 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2851 * wake it up explicitly.
2852 */
2853 if (worker->task)
2854 wake_up_process(worker->task);
2855
2856 raw_spin_unlock_irq(&pool->lock);
2857
2858 return worker;
2859
2860fail:
2861 ida_free(&pool->worker_ida, id);
2862 kfree(worker);
2863 return NULL;
2864}
2865
2866static void unbind_worker(struct worker *worker)
2867{
2868 lockdep_assert_held(&wq_pool_attach_mutex);
2869
2870 kthread_set_per_cpu(worker->task, -1);
2871 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2872 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2873 else
2874 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2875}
2876
2877static void wake_dying_workers(struct list_head *cull_list)
2878{
2879 struct worker *worker, *tmp;
2880
2881 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2882 list_del_init(&worker->entry);
2883 unbind_worker(worker);
2884 /*
2885 * If the worker was somehow already running, then it had to be
2886 * in pool->idle_list when set_worker_dying() happened or we
2887 * wouldn't have gotten here.
2888 *
2889 * Thus, the worker must either have observed the WORKER_DIE
2890 * flag, or have set its state to TASK_IDLE. Either way, the
2891 * below will be observed by the worker and is safe to do
2892 * outside of pool->lock.
2893 */
2894 wake_up_process(worker->task);
2895 }
2896}
2897
2898/**
2899 * set_worker_dying - Tag a worker for destruction
2900 * @worker: worker to be destroyed
2901 * @list: transfer worker away from its pool->idle_list and into list
2902 *
2903 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2904 * should be idle.
2905 *
2906 * CONTEXT:
2907 * raw_spin_lock_irq(pool->lock).
2908 */
2909static void set_worker_dying(struct worker *worker, struct list_head *list)
2910{
2911 struct worker_pool *pool = worker->pool;
2912
2913 lockdep_assert_held(&pool->lock);
2914 lockdep_assert_held(&wq_pool_attach_mutex);
2915
2916 /* sanity check frenzy */
2917 if (WARN_ON(worker->current_work) ||
2918 WARN_ON(!list_empty(&worker->scheduled)) ||
2919 WARN_ON(!(worker->flags & WORKER_IDLE)))
2920 return;
2921
2922 pool->nr_workers--;
2923 pool->nr_idle--;
2924
2925 worker->flags |= WORKER_DIE;
2926
2927 list_move(&worker->entry, list);
2928 list_move(&worker->node, &pool->dying_workers);
2929}
2930
2931/**
2932 * idle_worker_timeout - check if some idle workers can now be deleted.
2933 * @t: The pool's idle_timer that just expired
2934 *
2935 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2936 * worker_leave_idle(), as a worker flicking between idle and active while its
2937 * pool is at the too_many_workers() tipping point would cause too much timer
2938 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2939 * it expire and re-evaluate things from there.
2940 */
2941static void idle_worker_timeout(struct timer_list *t)
2942{
2943 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2944 bool do_cull = false;
2945
2946 if (work_pending(&pool->idle_cull_work))
2947 return;
2948
2949 raw_spin_lock_irq(&pool->lock);
2950
2951 if (too_many_workers(pool)) {
2952 struct worker *worker;
2953 unsigned long expires;
2954
2955 /* idle_list is kept in LIFO order, check the last one */
2956 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2957 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2958 do_cull = !time_before(jiffies, expires);
2959
2960 if (!do_cull)
2961 mod_timer(&pool->idle_timer, expires);
2962 }
2963 raw_spin_unlock_irq(&pool->lock);
2964
2965 if (do_cull)
2966 queue_work(system_unbound_wq, &pool->idle_cull_work);
2967}
2968
2969/**
2970 * idle_cull_fn - cull workers that have been idle for too long.
2971 * @work: the pool's work for handling these idle workers
2972 *
2973 * This goes through a pool's idle workers and gets rid of those that have been
2974 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2975 *
2976 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2977 * culled, so this also resets worker affinity. This requires a sleepable
2978 * context, hence the split between timer callback and work item.
2979 */
2980static void idle_cull_fn(struct work_struct *work)
2981{
2982 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2983 LIST_HEAD(cull_list);
2984
2985 /*
2986 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2987 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2988 * path. This is required as a previously-preempted worker could run after
2989 * set_worker_dying() has happened but before wake_dying_workers() did.
2990 */
2991 mutex_lock(&wq_pool_attach_mutex);
2992 raw_spin_lock_irq(&pool->lock);
2993
2994 while (too_many_workers(pool)) {
2995 struct worker *worker;
2996 unsigned long expires;
2997
2998 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2999 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
3000
3001 if (time_before(jiffies, expires)) {
3002 mod_timer(&pool->idle_timer, expires);
3003 break;
3004 }
3005
3006 set_worker_dying(worker, &cull_list);
3007 }
3008
3009 raw_spin_unlock_irq(&pool->lock);
3010 wake_dying_workers(&cull_list);
3011 mutex_unlock(&wq_pool_attach_mutex);
3012}
3013
3014static void send_mayday(struct work_struct *work)
3015{
3016 struct pool_workqueue *pwq = get_work_pwq(work);
3017 struct workqueue_struct *wq = pwq->wq;
3018
3019 lockdep_assert_held(&wq_mayday_lock);
3020
3021 if (!wq->rescuer)
3022 return;
3023
3024 /* mayday mayday mayday */
3025 if (list_empty(&pwq->mayday_node)) {
3026 /*
3027 * If @pwq is for an unbound wq, its base ref may be put at
3028 * any time due to an attribute change. Pin @pwq until the
3029 * rescuer is done with it.
3030 */
3031 get_pwq(pwq);
3032 list_add_tail(&pwq->mayday_node, &wq->maydays);
3033 wake_up_process(wq->rescuer->task);
3034 pwq->stats[PWQ_STAT_MAYDAY]++;
3035 }
3036}
3037
3038static void pool_mayday_timeout(struct timer_list *t)
3039{
3040 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3041 struct work_struct *work;
3042
3043 raw_spin_lock_irq(&pool->lock);
3044 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3045
3046 if (need_to_create_worker(pool)) {
3047 /*
3048 * We've been trying to create a new worker but
3049 * haven't been successful. We might be hitting an
3050 * allocation deadlock. Send distress signals to
3051 * rescuers.
3052 */
3053 list_for_each_entry(work, &pool->worklist, entry)
3054 send_mayday(work);
3055 }
3056
3057 raw_spin_unlock(&wq_mayday_lock);
3058 raw_spin_unlock_irq(&pool->lock);
3059
3060 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3061}
3062
3063/**
3064 * maybe_create_worker - create a new worker if necessary
3065 * @pool: pool to create a new worker for
3066 *
3067 * Create a new worker for @pool if necessary. @pool is guaranteed to
3068 * have at least one idle worker on return from this function. If
3069 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3070 * sent to all rescuers with works scheduled on @pool to resolve
3071 * possible allocation deadlock.
3072 *
3073 * On return, need_to_create_worker() is guaranteed to be %false and
3074 * may_start_working() %true.
3075 *
3076 * LOCKING:
3077 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3078 * multiple times. Does GFP_KERNEL allocations. Called only from
3079 * manager.
3080 */
3081static void maybe_create_worker(struct worker_pool *pool)
3082__releases(&pool->lock)
3083__acquires(&pool->lock)
3084{
3085restart:
3086 raw_spin_unlock_irq(&pool->lock);
3087
3088 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3089 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3090
3091 while (true) {
3092 if (create_worker(pool) || !need_to_create_worker(pool))
3093 break;
3094
3095 schedule_timeout_interruptible(CREATE_COOLDOWN);
3096
3097 if (!need_to_create_worker(pool))
3098 break;
3099 }
3100
3101 del_timer_sync(&pool->mayday_timer);
3102 raw_spin_lock_irq(&pool->lock);
3103 /*
3104 * This is necessary even after a new worker was just successfully
3105 * created as @pool->lock was dropped and the new worker might have
3106 * already become busy.
3107 */
3108 if (need_to_create_worker(pool))
3109 goto restart;
3110}
3111
3112/**
3113 * manage_workers - manage worker pool
3114 * @worker: self
3115 *
3116 * Assume the manager role and manage the worker pool @worker belongs
3117 * to. At any given time, there can be only zero or one manager per
3118 * pool. The exclusion is handled automatically by this function.
3119 *
3120 * The caller can safely start processing works on false return. On
3121 * true return, it's guaranteed that need_to_create_worker() is false
3122 * and may_start_working() is true.
3123 *
3124 * CONTEXT:
3125 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3126 * multiple times. Does GFP_KERNEL allocations.
3127 *
3128 * Return:
3129 * %false if the pool doesn't need management and the caller can safely
3130 * start processing works, %true if management function was performed and
3131 * the conditions that the caller verified before calling the function may
3132 * no longer be true.
3133 */
3134static bool manage_workers(struct worker *worker)
3135{
3136 struct worker_pool *pool = worker->pool;
3137
3138 if (pool->flags & POOL_MANAGER_ACTIVE)
3139 return false;
3140
3141 pool->flags |= POOL_MANAGER_ACTIVE;
3142 pool->manager = worker;
3143
3144 maybe_create_worker(pool);
3145
3146 pool->manager = NULL;
3147 pool->flags &= ~POOL_MANAGER_ACTIVE;
3148 rcuwait_wake_up(&manager_wait);
3149 return true;
3150}
3151
3152/**
3153 * process_one_work - process single work
3154 * @worker: self
3155 * @work: work to process
3156 *
3157 * Process @work. This function contains all the logics necessary to
3158 * process a single work including synchronization against and
3159 * interaction with other workers on the same cpu, queueing and
3160 * flushing. As long as context requirement is met, any worker can
3161 * call this function to process a work.
3162 *
3163 * CONTEXT:
3164 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3165 */
3166static void process_one_work(struct worker *worker, struct work_struct *work)
3167__releases(&pool->lock)
3168__acquires(&pool->lock)
3169{
3170 struct pool_workqueue *pwq = get_work_pwq(work);
3171 struct worker_pool *pool = worker->pool;
3172 unsigned long work_data;
3173 int lockdep_start_depth, rcu_start_depth;
3174 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3175#ifdef CONFIG_LOCKDEP
3176 /*
3177 * It is permissible to free the struct work_struct from
3178 * inside the function that is called from it, this we need to
3179 * take into account for lockdep too. To avoid bogus "held
3180 * lock freed" warnings as well as problems when looking into
3181 * work->lockdep_map, make a copy and use that here.
3182 */
3183 struct lockdep_map lockdep_map;
3184
3185 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3186#endif
3187 /* ensure we're on the correct CPU */
3188 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3189 raw_smp_processor_id() != pool->cpu);
3190
3191 /* claim and dequeue */
3192 debug_work_deactivate(work);
3193 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3194 worker->current_work = work;
3195 worker->current_func = work->func;
3196 worker->current_pwq = pwq;
3197 if (worker->task)
3198 worker->current_at = worker->task->se.sum_exec_runtime;
3199 work_data = *work_data_bits(work);
3200 worker->current_color = get_work_color(work_data);
3201
3202 /*
3203 * Record wq name for cmdline and debug reporting, may get
3204 * overridden through set_worker_desc().
3205 */
3206 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3207
3208 list_del_init(&work->entry);
3209
3210 /*
3211 * CPU intensive works don't participate in concurrency management.
3212 * They're the scheduler's responsibility. This takes @worker out
3213 * of concurrency management and the next code block will chain
3214 * execution of the pending work items.
3215 */
3216 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3217 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3218
3219 /*
3220 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3221 * since nr_running would always be >= 1 at this point. This is used to
3222 * chain execution of the pending work items for WORKER_NOT_RUNNING
3223 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3224 */
3225 kick_pool(pool);
3226
3227 /*
3228 * Record the last pool and clear PENDING which should be the last
3229 * update to @work. Also, do this inside @pool->lock so that
3230 * PENDING and queued state changes happen together while IRQ is
3231 * disabled.
3232 */
3233 set_work_pool_and_clear_pending(work, pool->id, 0);
3234
3235 pwq->stats[PWQ_STAT_STARTED]++;
3236 raw_spin_unlock_irq(&pool->lock);
3237
3238 rcu_start_depth = rcu_preempt_depth();
3239 lockdep_start_depth = lockdep_depth(current);
3240 /* see drain_dead_softirq_workfn() */
3241 if (!bh_draining)
3242 lock_map_acquire(&pwq->wq->lockdep_map);
3243 lock_map_acquire(&lockdep_map);
3244 /*
3245 * Strictly speaking we should mark the invariant state without holding
3246 * any locks, that is, before these two lock_map_acquire()'s.
3247 *
3248 * However, that would result in:
3249 *
3250 * A(W1)
3251 * WFC(C)
3252 * A(W1)
3253 * C(C)
3254 *
3255 * Which would create W1->C->W1 dependencies, even though there is no
3256 * actual deadlock possible. There are two solutions, using a
3257 * read-recursive acquire on the work(queue) 'locks', but this will then
3258 * hit the lockdep limitation on recursive locks, or simply discard
3259 * these locks.
3260 *
3261 * AFAICT there is no possible deadlock scenario between the
3262 * flush_work() and complete() primitives (except for single-threaded
3263 * workqueues), so hiding them isn't a problem.
3264 */
3265 lockdep_invariant_state(true);
3266 trace_workqueue_execute_start(work);
3267 worker->current_func(work);
3268 /*
3269 * While we must be careful to not use "work" after this, the trace
3270 * point will only record its address.
3271 */
3272 trace_workqueue_execute_end(work, worker->current_func);
3273 pwq->stats[PWQ_STAT_COMPLETED]++;
3274 lock_map_release(&lockdep_map);
3275 if (!bh_draining)
3276 lock_map_release(&pwq->wq->lockdep_map);
3277
3278 if (unlikely((worker->task && in_atomic()) ||
3279 lockdep_depth(current) != lockdep_start_depth ||
3280 rcu_preempt_depth() != rcu_start_depth)) {
3281 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3282 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3283 current->comm, task_pid_nr(current), preempt_count(),
3284 lockdep_start_depth, lockdep_depth(current),
3285 rcu_start_depth, rcu_preempt_depth(),
3286 worker->current_func);
3287 debug_show_held_locks(current);
3288 dump_stack();
3289 }
3290
3291 /*
3292 * The following prevents a kworker from hogging CPU on !PREEMPTION
3293 * kernels, where a requeueing work item waiting for something to
3294 * happen could deadlock with stop_machine as such work item could
3295 * indefinitely requeue itself while all other CPUs are trapped in
3296 * stop_machine. At the same time, report a quiescent RCU state so
3297 * the same condition doesn't freeze RCU.
3298 */
3299 if (worker->task)
3300 cond_resched();
3301
3302 raw_spin_lock_irq(&pool->lock);
3303
3304 /*
3305 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3306 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3307 * wq_cpu_intensive_thresh_us. Clear it.
3308 */
3309 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3310
3311 /* tag the worker for identification in schedule() */
3312 worker->last_func = worker->current_func;
3313
3314 /* we're done with it, release */
3315 hash_del(&worker->hentry);
3316 worker->current_work = NULL;
3317 worker->current_func = NULL;
3318 worker->current_pwq = NULL;
3319 worker->current_color = INT_MAX;
3320
3321 /* must be the last step, see the function comment */
3322 pwq_dec_nr_in_flight(pwq, work_data);
3323}
3324
3325/**
3326 * process_scheduled_works - process scheduled works
3327 * @worker: self
3328 *
3329 * Process all scheduled works. Please note that the scheduled list
3330 * may change while processing a work, so this function repeatedly
3331 * fetches a work from the top and executes it.
3332 *
3333 * CONTEXT:
3334 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3335 * multiple times.
3336 */
3337static void process_scheduled_works(struct worker *worker)
3338{
3339 struct work_struct *work;
3340 bool first = true;
3341
3342 while ((work = list_first_entry_or_null(&worker->scheduled,
3343 struct work_struct, entry))) {
3344 if (first) {
3345 worker->pool->watchdog_ts = jiffies;
3346 first = false;
3347 }
3348 process_one_work(worker, work);
3349 }
3350}
3351
3352static void set_pf_worker(bool val)
3353{
3354 mutex_lock(&wq_pool_attach_mutex);
3355 if (val)
3356 current->flags |= PF_WQ_WORKER;
3357 else
3358 current->flags &= ~PF_WQ_WORKER;
3359 mutex_unlock(&wq_pool_attach_mutex);
3360}
3361
3362/**
3363 * worker_thread - the worker thread function
3364 * @__worker: self
3365 *
3366 * The worker thread function. All workers belong to a worker_pool -
3367 * either a per-cpu one or dynamic unbound one. These workers process all
3368 * work items regardless of their specific target workqueue. The only
3369 * exception is work items which belong to workqueues with a rescuer which
3370 * will be explained in rescuer_thread().
3371 *
3372 * Return: 0
3373 */
3374static int worker_thread(void *__worker)
3375{
3376 struct worker *worker = __worker;
3377 struct worker_pool *pool = worker->pool;
3378
3379 /* tell the scheduler that this is a workqueue worker */
3380 set_pf_worker(true);
3381woke_up:
3382 raw_spin_lock_irq(&pool->lock);
3383
3384 /* am I supposed to die? */
3385 if (unlikely(worker->flags & WORKER_DIE)) {
3386 raw_spin_unlock_irq(&pool->lock);
3387 set_pf_worker(false);
3388
3389 set_task_comm(worker->task, "kworker/dying");
3390 ida_free(&pool->worker_ida, worker->id);
3391 worker_detach_from_pool(worker);
3392 WARN_ON_ONCE(!list_empty(&worker->entry));
3393 kfree(worker);
3394 return 0;
3395 }
3396
3397 worker_leave_idle(worker);
3398recheck:
3399 /* no more worker necessary? */
3400 if (!need_more_worker(pool))
3401 goto sleep;
3402
3403 /* do we need to manage? */
3404 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3405 goto recheck;
3406
3407 /*
3408 * ->scheduled list can only be filled while a worker is
3409 * preparing to process a work or actually processing it.
3410 * Make sure nobody diddled with it while I was sleeping.
3411 */
3412 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3413
3414 /*
3415 * Finish PREP stage. We're guaranteed to have at least one idle
3416 * worker or that someone else has already assumed the manager
3417 * role. This is where @worker starts participating in concurrency
3418 * management if applicable and concurrency management is restored
3419 * after being rebound. See rebind_workers() for details.
3420 */
3421 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3422
3423 do {
3424 struct work_struct *work =
3425 list_first_entry(&pool->worklist,
3426 struct work_struct, entry);
3427
3428 if (assign_work(work, worker, NULL))
3429 process_scheduled_works(worker);
3430 } while (keep_working(pool));
3431
3432 worker_set_flags(worker, WORKER_PREP);
3433sleep:
3434 /*
3435 * pool->lock is held and there's no work to process and no need to
3436 * manage, sleep. Workers are woken up only while holding
3437 * pool->lock or from local cpu, so setting the current state
3438 * before releasing pool->lock is enough to prevent losing any
3439 * event.
3440 */
3441 worker_enter_idle(worker);
3442 __set_current_state(TASK_IDLE);
3443 raw_spin_unlock_irq(&pool->lock);
3444 schedule();
3445 goto woke_up;
3446}
3447
3448/**
3449 * rescuer_thread - the rescuer thread function
3450 * @__rescuer: self
3451 *
3452 * Workqueue rescuer thread function. There's one rescuer for each
3453 * workqueue which has WQ_MEM_RECLAIM set.
3454 *
3455 * Regular work processing on a pool may block trying to create a new
3456 * worker which uses GFP_KERNEL allocation which has slight chance of
3457 * developing into deadlock if some works currently on the same queue
3458 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3459 * the problem rescuer solves.
3460 *
3461 * When such condition is possible, the pool summons rescuers of all
3462 * workqueues which have works queued on the pool and let them process
3463 * those works so that forward progress can be guaranteed.
3464 *
3465 * This should happen rarely.
3466 *
3467 * Return: 0
3468 */
3469static int rescuer_thread(void *__rescuer)
3470{
3471 struct worker *rescuer = __rescuer;
3472 struct workqueue_struct *wq = rescuer->rescue_wq;
3473 bool should_stop;
3474
3475 set_user_nice(current, RESCUER_NICE_LEVEL);
3476
3477 /*
3478 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3479 * doesn't participate in concurrency management.
3480 */
3481 set_pf_worker(true);
3482repeat:
3483 set_current_state(TASK_IDLE);
3484
3485 /*
3486 * By the time the rescuer is requested to stop, the workqueue
3487 * shouldn't have any work pending, but @wq->maydays may still have
3488 * pwq(s) queued. This can happen by non-rescuer workers consuming
3489 * all the work items before the rescuer got to them. Go through
3490 * @wq->maydays processing before acting on should_stop so that the
3491 * list is always empty on exit.
3492 */
3493 should_stop = kthread_should_stop();
3494
3495 /* see whether any pwq is asking for help */
3496 raw_spin_lock_irq(&wq_mayday_lock);
3497
3498 while (!list_empty(&wq->maydays)) {
3499 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3500 struct pool_workqueue, mayday_node);
3501 struct worker_pool *pool = pwq->pool;
3502 struct work_struct *work, *n;
3503
3504 __set_current_state(TASK_RUNNING);
3505 list_del_init(&pwq->mayday_node);
3506
3507 raw_spin_unlock_irq(&wq_mayday_lock);
3508
3509 worker_attach_to_pool(rescuer, pool);
3510
3511 raw_spin_lock_irq(&pool->lock);
3512
3513 /*
3514 * Slurp in all works issued via this workqueue and
3515 * process'em.
3516 */
3517 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3518 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3519 if (get_work_pwq(work) == pwq &&
3520 assign_work(work, rescuer, &n))
3521 pwq->stats[PWQ_STAT_RESCUED]++;
3522 }
3523
3524 if (!list_empty(&rescuer->scheduled)) {
3525 process_scheduled_works(rescuer);
3526
3527 /*
3528 * The above execution of rescued work items could
3529 * have created more to rescue through
3530 * pwq_activate_first_inactive() or chained
3531 * queueing. Let's put @pwq back on mayday list so
3532 * that such back-to-back work items, which may be
3533 * being used to relieve memory pressure, don't
3534 * incur MAYDAY_INTERVAL delay inbetween.
3535 */
3536 if (pwq->nr_active && need_to_create_worker(pool)) {
3537 raw_spin_lock(&wq_mayday_lock);
3538 /*
3539 * Queue iff we aren't racing destruction
3540 * and somebody else hasn't queued it already.
3541 */
3542 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3543 get_pwq(pwq);
3544 list_add_tail(&pwq->mayday_node, &wq->maydays);
3545 }
3546 raw_spin_unlock(&wq_mayday_lock);
3547 }
3548 }
3549
3550 /*
3551 * Put the reference grabbed by send_mayday(). @pool won't
3552 * go away while we're still attached to it.
3553 */
3554 put_pwq(pwq);
3555
3556 /*
3557 * Leave this pool. Notify regular workers; otherwise, we end up
3558 * with 0 concurrency and stalling the execution.
3559 */
3560 kick_pool(pool);
3561
3562 raw_spin_unlock_irq(&pool->lock);
3563
3564 worker_detach_from_pool(rescuer);
3565
3566 raw_spin_lock_irq(&wq_mayday_lock);
3567 }
3568
3569 raw_spin_unlock_irq(&wq_mayday_lock);
3570
3571 if (should_stop) {
3572 __set_current_state(TASK_RUNNING);
3573 set_pf_worker(false);
3574 return 0;
3575 }
3576
3577 /* rescuers should never participate in concurrency management */
3578 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3579 schedule();
3580 goto repeat;
3581}
3582
3583static void bh_worker(struct worker *worker)
3584{
3585 struct worker_pool *pool = worker->pool;
3586 int nr_restarts = BH_WORKER_RESTARTS;
3587 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3588
3589 raw_spin_lock_irq(&pool->lock);
3590 worker_leave_idle(worker);
3591
3592 /*
3593 * This function follows the structure of worker_thread(). See there for
3594 * explanations on each step.
3595 */
3596 if (!need_more_worker(pool))
3597 goto done;
3598
3599 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3600 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3601
3602 do {
3603 struct work_struct *work =
3604 list_first_entry(&pool->worklist,
3605 struct work_struct, entry);
3606
3607 if (assign_work(work, worker, NULL))
3608 process_scheduled_works(worker);
3609 } while (keep_working(pool) &&
3610 --nr_restarts && time_before(jiffies, end));
3611
3612 worker_set_flags(worker, WORKER_PREP);
3613done:
3614 worker_enter_idle(worker);
3615 kick_pool(pool);
3616 raw_spin_unlock_irq(&pool->lock);
3617}
3618
3619/*
3620 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3621 *
3622 * This is currently called from tasklet[_hi]action() and thus is also called
3623 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3624 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3625 * can be dropped.
3626 *
3627 * After full conversion, we'll add worker->softirq_action, directly use the
3628 * softirq action and obtain the worker pointer from the softirq_action pointer.
3629 */
3630void workqueue_softirq_action(bool highpri)
3631{
3632 struct worker_pool *pool =
3633 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3634 if (need_more_worker(pool))
3635 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3636}
3637
3638struct wq_drain_dead_softirq_work {
3639 struct work_struct work;
3640 struct worker_pool *pool;
3641 struct completion done;
3642};
3643
3644static void drain_dead_softirq_workfn(struct work_struct *work)
3645{
3646 struct wq_drain_dead_softirq_work *dead_work =
3647 container_of(work, struct wq_drain_dead_softirq_work, work);
3648 struct worker_pool *pool = dead_work->pool;
3649 bool repeat;
3650
3651 /*
3652 * @pool's CPU is dead and we want to execute its still pending work
3653 * items from this BH work item which is running on a different CPU. As
3654 * its CPU is dead, @pool can't be kicked and, as work execution path
3655 * will be nested, a lockdep annotation needs to be suppressed. Mark
3656 * @pool with %POOL_BH_DRAINING for the special treatments.
3657 */
3658 raw_spin_lock_irq(&pool->lock);
3659 pool->flags |= POOL_BH_DRAINING;
3660 raw_spin_unlock_irq(&pool->lock);
3661
3662 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3663
3664 raw_spin_lock_irq(&pool->lock);
3665 pool->flags &= ~POOL_BH_DRAINING;
3666 repeat = need_more_worker(pool);
3667 raw_spin_unlock_irq(&pool->lock);
3668
3669 /*
3670 * bh_worker() might hit consecutive execution limit and bail. If there
3671 * still are pending work items, reschedule self and return so that we
3672 * don't hog this CPU's BH.
3673 */
3674 if (repeat) {
3675 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3676 queue_work(system_bh_highpri_wq, work);
3677 else
3678 queue_work(system_bh_wq, work);
3679 } else {
3680 complete(&dead_work->done);
3681 }
3682}
3683
3684/*
3685 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3686 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3687 * have to worry about draining overlapping with CPU coming back online or
3688 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3689 * on). Let's keep it simple and drain them synchronously. These are BH work
3690 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3691 */
3692void workqueue_softirq_dead(unsigned int cpu)
3693{
3694 int i;
3695
3696 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3697 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3698 struct wq_drain_dead_softirq_work dead_work;
3699
3700 if (!need_more_worker(pool))
3701 continue;
3702
3703 INIT_WORK(&dead_work.work, drain_dead_softirq_workfn);
3704 dead_work.pool = pool;
3705 init_completion(&dead_work.done);
3706
3707 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3708 queue_work(system_bh_highpri_wq, &dead_work.work);
3709 else
3710 queue_work(system_bh_wq, &dead_work.work);
3711
3712 wait_for_completion(&dead_work.done);
3713 }
3714}
3715
3716/**
3717 * check_flush_dependency - check for flush dependency sanity
3718 * @target_wq: workqueue being flushed
3719 * @target_work: work item being flushed (NULL for workqueue flushes)
3720 *
3721 * %current is trying to flush the whole @target_wq or @target_work on it.
3722 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
3723 * reclaiming memory or running on a workqueue which doesn't have
3724 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
3725 * a deadlock.
3726 */
3727static void check_flush_dependency(struct workqueue_struct *target_wq,
3728 struct work_struct *target_work)
3729{
3730 work_func_t target_func = target_work ? target_work->func : NULL;
3731 struct worker *worker;
3732
3733 if (target_wq->flags & WQ_MEM_RECLAIM)
3734 return;
3735
3736 worker = current_wq_worker();
3737
3738 WARN_ONCE(current->flags & PF_MEMALLOC,
3739 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3740 current->pid, current->comm, target_wq->name, target_func);
3741 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3742 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3743 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3744 worker->current_pwq->wq->name, worker->current_func,
3745 target_wq->name, target_func);
3746}
3747
3748struct wq_barrier {
3749 struct work_struct work;
3750 struct completion done;
3751 struct task_struct *task; /* purely informational */
3752};
3753
3754static void wq_barrier_func(struct work_struct *work)
3755{
3756 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3757 complete(&barr->done);
3758}
3759
3760/**
3761 * insert_wq_barrier - insert a barrier work
3762 * @pwq: pwq to insert barrier into
3763 * @barr: wq_barrier to insert
3764 * @target: target work to attach @barr to
3765 * @worker: worker currently executing @target, NULL if @target is not executing
3766 *
3767 * @barr is linked to @target such that @barr is completed only after
3768 * @target finishes execution. Please note that the ordering
3769 * guarantee is observed only with respect to @target and on the local
3770 * cpu.
3771 *
3772 * Currently, a queued barrier can't be canceled. This is because
3773 * try_to_grab_pending() can't determine whether the work to be
3774 * grabbed is at the head of the queue and thus can't clear LINKED
3775 * flag of the previous work while there must be a valid next work
3776 * after a work with LINKED flag set.
3777 *
3778 * Note that when @worker is non-NULL, @target may be modified
3779 * underneath us, so we can't reliably determine pwq from @target.
3780 *
3781 * CONTEXT:
3782 * raw_spin_lock_irq(pool->lock).
3783 */
3784static void insert_wq_barrier(struct pool_workqueue *pwq,
3785 struct wq_barrier *barr,
3786 struct work_struct *target, struct worker *worker)
3787{
3788 static __maybe_unused struct lock_class_key bh_key, thr_key;
3789 unsigned int work_flags = 0;
3790 unsigned int work_color;
3791 struct list_head *head;
3792
3793 /*
3794 * debugobject calls are safe here even with pool->lock locked
3795 * as we know for sure that this will not trigger any of the
3796 * checks and call back into the fixup functions where we
3797 * might deadlock.
3798 *
3799 * BH and threaded workqueues need separate lockdep keys to avoid
3800 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3801 * usage".
3802 */
3803 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3804 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3805 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3806
3807 init_completion_map(&barr->done, &target->lockdep_map);
3808
3809 barr->task = current;
3810
3811 /* The barrier work item does not participate in nr_active. */
3812 work_flags |= WORK_STRUCT_INACTIVE;
3813
3814 /*
3815 * If @target is currently being executed, schedule the
3816 * barrier to the worker; otherwise, put it after @target.
3817 */
3818 if (worker) {
3819 head = worker->scheduled.next;
3820 work_color = worker->current_color;
3821 } else {
3822 unsigned long *bits = work_data_bits(target);
3823
3824 head = target->entry.next;
3825 /* there can already be other linked works, inherit and set */
3826 work_flags |= *bits & WORK_STRUCT_LINKED;
3827 work_color = get_work_color(*bits);
3828 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3829 }
3830
3831 pwq->nr_in_flight[work_color]++;
3832 work_flags |= work_color_to_flags(work_color);
3833
3834 insert_work(pwq, &barr->work, head, work_flags);
3835}
3836
3837/**
3838 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3839 * @wq: workqueue being flushed
3840 * @flush_color: new flush color, < 0 for no-op
3841 * @work_color: new work color, < 0 for no-op
3842 *
3843 * Prepare pwqs for workqueue flushing.
3844 *
3845 * If @flush_color is non-negative, flush_color on all pwqs should be
3846 * -1. If no pwq has in-flight commands at the specified color, all
3847 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3848 * has in flight commands, its pwq->flush_color is set to
3849 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3850 * wakeup logic is armed and %true is returned.
3851 *
3852 * The caller should have initialized @wq->first_flusher prior to
3853 * calling this function with non-negative @flush_color. If
3854 * @flush_color is negative, no flush color update is done and %false
3855 * is returned.
3856 *
3857 * If @work_color is non-negative, all pwqs should have the same
3858 * work_color which is previous to @work_color and all will be
3859 * advanced to @work_color.
3860 *
3861 * CONTEXT:
3862 * mutex_lock(wq->mutex).
3863 *
3864 * Return:
3865 * %true if @flush_color >= 0 and there's something to flush. %false
3866 * otherwise.
3867 */
3868static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3869 int flush_color, int work_color)
3870{
3871 bool wait = false;
3872 struct pool_workqueue *pwq;
3873
3874 if (flush_color >= 0) {
3875 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3876 atomic_set(&wq->nr_pwqs_to_flush, 1);
3877 }
3878
3879 for_each_pwq(pwq, wq) {
3880 struct worker_pool *pool = pwq->pool;
3881
3882 raw_spin_lock_irq(&pool->lock);
3883
3884 if (flush_color >= 0) {
3885 WARN_ON_ONCE(pwq->flush_color != -1);
3886
3887 if (pwq->nr_in_flight[flush_color]) {
3888 pwq->flush_color = flush_color;
3889 atomic_inc(&wq->nr_pwqs_to_flush);
3890 wait = true;
3891 }
3892 }
3893
3894 if (work_color >= 0) {
3895 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3896 pwq->work_color = work_color;
3897 }
3898
3899 raw_spin_unlock_irq(&pool->lock);
3900 }
3901
3902 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3903 complete(&wq->first_flusher->done);
3904
3905 return wait;
3906}
3907
3908static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3909{
3910#ifdef CONFIG_LOCKDEP
3911 if (wq->flags & WQ_BH)
3912 local_bh_disable();
3913
3914 lock_map_acquire(&wq->lockdep_map);
3915 lock_map_release(&wq->lockdep_map);
3916
3917 if (wq->flags & WQ_BH)
3918 local_bh_enable();
3919#endif
3920}
3921
3922static void touch_work_lockdep_map(struct work_struct *work,
3923 struct workqueue_struct *wq)
3924{
3925#ifdef CONFIG_LOCKDEP
3926 if (wq->flags & WQ_BH)
3927 local_bh_disable();
3928
3929 lock_map_acquire(&work->lockdep_map);
3930 lock_map_release(&work->lockdep_map);
3931
3932 if (wq->flags & WQ_BH)
3933 local_bh_enable();
3934#endif
3935}
3936
3937/**
3938 * __flush_workqueue - ensure that any scheduled work has run to completion.
3939 * @wq: workqueue to flush
3940 *
3941 * This function sleeps until all work items which were queued on entry
3942 * have finished execution, but it is not livelocked by new incoming ones.
3943 */
3944void __flush_workqueue(struct workqueue_struct *wq)
3945{
3946 struct wq_flusher this_flusher = {
3947 .list = LIST_HEAD_INIT(this_flusher.list),
3948 .flush_color = -1,
3949 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3950 };
3951 int next_color;
3952
3953 if (WARN_ON(!wq_online))
3954 return;
3955
3956 touch_wq_lockdep_map(wq);
3957
3958 mutex_lock(&wq->mutex);
3959
3960 /*
3961 * Start-to-wait phase
3962 */
3963 next_color = work_next_color(wq->work_color);
3964
3965 if (next_color != wq->flush_color) {
3966 /*
3967 * Color space is not full. The current work_color
3968 * becomes our flush_color and work_color is advanced
3969 * by one.
3970 */
3971 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3972 this_flusher.flush_color = wq->work_color;
3973 wq->work_color = next_color;
3974
3975 if (!wq->first_flusher) {
3976 /* no flush in progress, become the first flusher */
3977 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3978
3979 wq->first_flusher = &this_flusher;
3980
3981 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3982 wq->work_color)) {
3983 /* nothing to flush, done */
3984 wq->flush_color = next_color;
3985 wq->first_flusher = NULL;
3986 goto out_unlock;
3987 }
3988 } else {
3989 /* wait in queue */
3990 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3991 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3992 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3993 }
3994 } else {
3995 /*
3996 * Oops, color space is full, wait on overflow queue.
3997 * The next flush completion will assign us
3998 * flush_color and transfer to flusher_queue.
3999 */
4000 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
4001 }
4002
4003 check_flush_dependency(wq, NULL);
4004
4005 mutex_unlock(&wq->mutex);
4006
4007 wait_for_completion(&this_flusher.done);
4008
4009 /*
4010 * Wake-up-and-cascade phase
4011 *
4012 * First flushers are responsible for cascading flushes and
4013 * handling overflow. Non-first flushers can simply return.
4014 */
4015 if (READ_ONCE(wq->first_flusher) != &this_flusher)
4016 return;
4017
4018 mutex_lock(&wq->mutex);
4019
4020 /* we might have raced, check again with mutex held */
4021 if (wq->first_flusher != &this_flusher)
4022 goto out_unlock;
4023
4024 WRITE_ONCE(wq->first_flusher, NULL);
4025
4026 WARN_ON_ONCE(!list_empty(&this_flusher.list));
4027 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4028
4029 while (true) {
4030 struct wq_flusher *next, *tmp;
4031
4032 /* complete all the flushers sharing the current flush color */
4033 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4034 if (next->flush_color != wq->flush_color)
4035 break;
4036 list_del_init(&next->list);
4037 complete(&next->done);
4038 }
4039
4040 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4041 wq->flush_color != work_next_color(wq->work_color));
4042
4043 /* this flush_color is finished, advance by one */
4044 wq->flush_color = work_next_color(wq->flush_color);
4045
4046 /* one color has been freed, handle overflow queue */
4047 if (!list_empty(&wq->flusher_overflow)) {
4048 /*
4049 * Assign the same color to all overflowed
4050 * flushers, advance work_color and append to
4051 * flusher_queue. This is the start-to-wait
4052 * phase for these overflowed flushers.
4053 */
4054 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4055 tmp->flush_color = wq->work_color;
4056
4057 wq->work_color = work_next_color(wq->work_color);
4058
4059 list_splice_tail_init(&wq->flusher_overflow,
4060 &wq->flusher_queue);
4061 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4062 }
4063
4064 if (list_empty(&wq->flusher_queue)) {
4065 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4066 break;
4067 }
4068
4069 /*
4070 * Need to flush more colors. Make the next flusher
4071 * the new first flusher and arm pwqs.
4072 */
4073 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4074 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4075
4076 list_del_init(&next->list);
4077 wq->first_flusher = next;
4078
4079 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4080 break;
4081
4082 /*
4083 * Meh... this color is already done, clear first
4084 * flusher and repeat cascading.
4085 */
4086 wq->first_flusher = NULL;
4087 }
4088
4089out_unlock:
4090 mutex_unlock(&wq->mutex);
4091}
4092EXPORT_SYMBOL(__flush_workqueue);
4093
4094/**
4095 * drain_workqueue - drain a workqueue
4096 * @wq: workqueue to drain
4097 *
4098 * Wait until the workqueue becomes empty. While draining is in progress,
4099 * only chain queueing is allowed. IOW, only currently pending or running
4100 * work items on @wq can queue further work items on it. @wq is flushed
4101 * repeatedly until it becomes empty. The number of flushing is determined
4102 * by the depth of chaining and should be relatively short. Whine if it
4103 * takes too long.
4104 */
4105void drain_workqueue(struct workqueue_struct *wq)
4106{
4107 unsigned int flush_cnt = 0;
4108 struct pool_workqueue *pwq;
4109
4110 /*
4111 * __queue_work() needs to test whether there are drainers, is much
4112 * hotter than drain_workqueue() and already looks at @wq->flags.
4113 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4114 */
4115 mutex_lock(&wq->mutex);
4116 if (!wq->nr_drainers++)
4117 wq->flags |= __WQ_DRAINING;
4118 mutex_unlock(&wq->mutex);
4119reflush:
4120 __flush_workqueue(wq);
4121
4122 mutex_lock(&wq->mutex);
4123
4124 for_each_pwq(pwq, wq) {
4125 bool drained;
4126
4127 raw_spin_lock_irq(&pwq->pool->lock);
4128 drained = pwq_is_empty(pwq);
4129 raw_spin_unlock_irq(&pwq->pool->lock);
4130
4131 if (drained)
4132 continue;
4133
4134 if (++flush_cnt == 10 ||
4135 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4136 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4137 wq->name, __func__, flush_cnt);
4138
4139 mutex_unlock(&wq->mutex);
4140 goto reflush;
4141 }
4142
4143 if (!--wq->nr_drainers)
4144 wq->flags &= ~__WQ_DRAINING;
4145 mutex_unlock(&wq->mutex);
4146}
4147EXPORT_SYMBOL_GPL(drain_workqueue);
4148
4149static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4150 bool from_cancel)
4151{
4152 struct worker *worker = NULL;
4153 struct worker_pool *pool;
4154 struct pool_workqueue *pwq;
4155 struct workqueue_struct *wq;
4156
4157 might_sleep();
4158
4159 rcu_read_lock();
4160 pool = get_work_pool(work);
4161 if (!pool) {
4162 rcu_read_unlock();
4163 return false;
4164 }
4165
4166 raw_spin_lock_irq(&pool->lock);
4167 /* see the comment in try_to_grab_pending() with the same code */
4168 pwq = get_work_pwq(work);
4169 if (pwq) {
4170 if (unlikely(pwq->pool != pool))
4171 goto already_gone;
4172 } else {
4173 worker = find_worker_executing_work(pool, work);
4174 if (!worker)
4175 goto already_gone;
4176 pwq = worker->current_pwq;
4177 }
4178
4179 wq = pwq->wq;
4180 check_flush_dependency(wq, work);
4181
4182 insert_wq_barrier(pwq, barr, work, worker);
4183 raw_spin_unlock_irq(&pool->lock);
4184
4185 touch_work_lockdep_map(work, wq);
4186
4187 /*
4188 * Force a lock recursion deadlock when using flush_work() inside a
4189 * single-threaded or rescuer equipped workqueue.
4190 *
4191 * For single threaded workqueues the deadlock happens when the work
4192 * is after the work issuing the flush_work(). For rescuer equipped
4193 * workqueues the deadlock happens when the rescuer stalls, blocking
4194 * forward progress.
4195 */
4196 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4197 touch_wq_lockdep_map(wq);
4198
4199 rcu_read_unlock();
4200 return true;
4201already_gone:
4202 raw_spin_unlock_irq(&pool->lock);
4203 rcu_read_unlock();
4204 return false;
4205}
4206
4207static bool __flush_work(struct work_struct *work, bool from_cancel)
4208{
4209 struct wq_barrier barr;
4210
4211 if (WARN_ON(!wq_online))
4212 return false;
4213
4214 if (WARN_ON(!work->func))
4215 return false;
4216
4217 if (start_flush_work(work, &barr, from_cancel)) {
4218 wait_for_completion(&barr.done);
4219 destroy_work_on_stack(&barr.work);
4220 return true;
4221 } else {
4222 return false;
4223 }
4224}
4225
4226/**
4227 * flush_work - wait for a work to finish executing the last queueing instance
4228 * @work: the work to flush
4229 *
4230 * Wait until @work has finished execution. @work is guaranteed to be idle
4231 * on return if it hasn't been requeued since flush started.
4232 *
4233 * Return:
4234 * %true if flush_work() waited for the work to finish execution,
4235 * %false if it was already idle.
4236 */
4237bool flush_work(struct work_struct *work)
4238{
4239 return __flush_work(work, false);
4240}
4241EXPORT_SYMBOL_GPL(flush_work);
4242
4243/**
4244 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4245 * @dwork: the delayed work to flush
4246 *
4247 * Delayed timer is cancelled and the pending work is queued for
4248 * immediate execution. Like flush_work(), this function only
4249 * considers the last queueing instance of @dwork.
4250 *
4251 * Return:
4252 * %true if flush_work() waited for the work to finish execution,
4253 * %false if it was already idle.
4254 */
4255bool flush_delayed_work(struct delayed_work *dwork)
4256{
4257 local_irq_disable();
4258 if (del_timer_sync(&dwork->timer))
4259 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4260 local_irq_enable();
4261 return flush_work(&dwork->work);
4262}
4263EXPORT_SYMBOL(flush_delayed_work);
4264
4265/**
4266 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4267 * @rwork: the rcu work to flush
4268 *
4269 * Return:
4270 * %true if flush_rcu_work() waited for the work to finish execution,
4271 * %false if it was already idle.
4272 */
4273bool flush_rcu_work(struct rcu_work *rwork)
4274{
4275 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4276 rcu_barrier();
4277 flush_work(&rwork->work);
4278 return true;
4279 } else {
4280 return flush_work(&rwork->work);
4281 }
4282}
4283EXPORT_SYMBOL(flush_rcu_work);
4284
4285static bool __cancel_work(struct work_struct *work, u32 cflags)
4286{
4287 unsigned long irq_flags;
4288 int ret;
4289
4290 do {
4291 ret = try_to_grab_pending(work, cflags, &irq_flags);
4292 } while (unlikely(ret == -EAGAIN));
4293
4294 if (unlikely(ret < 0))
4295 return false;
4296
4297 set_work_pool_and_clear_pending(work, get_work_pool_id(work), 0);
4298 local_irq_restore(irq_flags);
4299 return ret;
4300}
4301
4302static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4303{
4304 unsigned long irq_flags;
4305 bool ret;
4306
4307 /* claim @work and tell other tasks trying to grab @work to back off */
4308 ret = work_grab_pending(work, cflags, &irq_flags);
4309 mark_work_canceling(work);
4310 local_irq_restore(irq_flags);
4311
4312 /*
4313 * Skip __flush_work() during early boot when we know that @work isn't
4314 * executing. This allows canceling during early boot.
4315 */
4316 if (wq_online)
4317 __flush_work(work, true);
4318
4319 /*
4320 * smp_mb() at the end of set_work_pool_and_clear_pending() is paired
4321 * with prepare_to_wait() above so that either waitqueue_active() is
4322 * visible here or !work_is_canceling() is visible there.
4323 */
4324 set_work_pool_and_clear_pending(work, WORK_OFFQ_POOL_NONE, 0);
4325
4326 if (waitqueue_active(&wq_cancel_waitq))
4327 __wake_up(&wq_cancel_waitq, TASK_NORMAL, 1, work);
4328
4329 return ret;
4330}
4331
4332/*
4333 * See cancel_delayed_work()
4334 */
4335bool cancel_work(struct work_struct *work)
4336{
4337 return __cancel_work(work, 0);
4338}
4339EXPORT_SYMBOL(cancel_work);
4340
4341/**
4342 * cancel_work_sync - cancel a work and wait for it to finish
4343 * @work: the work to cancel
4344 *
4345 * Cancel @work and wait for its execution to finish. This function
4346 * can be used even if the work re-queues itself or migrates to
4347 * another workqueue. On return from this function, @work is
4348 * guaranteed to be not pending or executing on any CPU.
4349 *
4350 * cancel_work_sync(&delayed_work->work) must not be used for
4351 * delayed_work's. Use cancel_delayed_work_sync() instead.
4352 *
4353 * The caller must ensure that the workqueue on which @work was last
4354 * queued can't be destroyed before this function returns.
4355 *
4356 * Return:
4357 * %true if @work was pending, %false otherwise.
4358 */
4359bool cancel_work_sync(struct work_struct *work)
4360{
4361 return __cancel_work_sync(work, 0);
4362}
4363EXPORT_SYMBOL_GPL(cancel_work_sync);
4364
4365/**
4366 * cancel_delayed_work - cancel a delayed work
4367 * @dwork: delayed_work to cancel
4368 *
4369 * Kill off a pending delayed_work.
4370 *
4371 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4372 * pending.
4373 *
4374 * Note:
4375 * The work callback function may still be running on return, unless
4376 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4377 * use cancel_delayed_work_sync() to wait on it.
4378 *
4379 * This function is safe to call from any context including IRQ handler.
4380 */
4381bool cancel_delayed_work(struct delayed_work *dwork)
4382{
4383 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4384}
4385EXPORT_SYMBOL(cancel_delayed_work);
4386
4387/**
4388 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4389 * @dwork: the delayed work cancel
4390 *
4391 * This is cancel_work_sync() for delayed works.
4392 *
4393 * Return:
4394 * %true if @dwork was pending, %false otherwise.
4395 */
4396bool cancel_delayed_work_sync(struct delayed_work *dwork)
4397{
4398 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4399}
4400EXPORT_SYMBOL(cancel_delayed_work_sync);
4401
4402/**
4403 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4404 * @func: the function to call
4405 *
4406 * schedule_on_each_cpu() executes @func on each online CPU using the
4407 * system workqueue and blocks until all CPUs have completed.
4408 * schedule_on_each_cpu() is very slow.
4409 *
4410 * Return:
4411 * 0 on success, -errno on failure.
4412 */
4413int schedule_on_each_cpu(work_func_t func)
4414{
4415 int cpu;
4416 struct work_struct __percpu *works;
4417
4418 works = alloc_percpu(struct work_struct);
4419 if (!works)
4420 return -ENOMEM;
4421
4422 cpus_read_lock();
4423
4424 for_each_online_cpu(cpu) {
4425 struct work_struct *work = per_cpu_ptr(works, cpu);
4426
4427 INIT_WORK(work, func);
4428 schedule_work_on(cpu, work);
4429 }
4430
4431 for_each_online_cpu(cpu)
4432 flush_work(per_cpu_ptr(works, cpu));
4433
4434 cpus_read_unlock();
4435 free_percpu(works);
4436 return 0;
4437}
4438
4439/**
4440 * execute_in_process_context - reliably execute the routine with user context
4441 * @fn: the function to execute
4442 * @ew: guaranteed storage for the execute work structure (must
4443 * be available when the work executes)
4444 *
4445 * Executes the function immediately if process context is available,
4446 * otherwise schedules the function for delayed execution.
4447 *
4448 * Return: 0 - function was executed
4449 * 1 - function was scheduled for execution
4450 */
4451int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4452{
4453 if (!in_interrupt()) {
4454 fn(&ew->work);
4455 return 0;
4456 }
4457
4458 INIT_WORK(&ew->work, fn);
4459 schedule_work(&ew->work);
4460
4461 return 1;
4462}
4463EXPORT_SYMBOL_GPL(execute_in_process_context);
4464
4465/**
4466 * free_workqueue_attrs - free a workqueue_attrs
4467 * @attrs: workqueue_attrs to free
4468 *
4469 * Undo alloc_workqueue_attrs().
4470 */
4471void free_workqueue_attrs(struct workqueue_attrs *attrs)
4472{
4473 if (attrs) {
4474 free_cpumask_var(attrs->cpumask);
4475 free_cpumask_var(attrs->__pod_cpumask);
4476 kfree(attrs);
4477 }
4478}
4479
4480/**
4481 * alloc_workqueue_attrs - allocate a workqueue_attrs
4482 *
4483 * Allocate a new workqueue_attrs, initialize with default settings and
4484 * return it.
4485 *
4486 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4487 */
4488struct workqueue_attrs *alloc_workqueue_attrs(void)
4489{
4490 struct workqueue_attrs *attrs;
4491
4492 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4493 if (!attrs)
4494 goto fail;
4495 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4496 goto fail;
4497 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4498 goto fail;
4499
4500 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4501 attrs->affn_scope = WQ_AFFN_DFL;
4502 return attrs;
4503fail:
4504 free_workqueue_attrs(attrs);
4505 return NULL;
4506}
4507
4508static void copy_workqueue_attrs(struct workqueue_attrs *to,
4509 const struct workqueue_attrs *from)
4510{
4511 to->nice = from->nice;
4512 cpumask_copy(to->cpumask, from->cpumask);
4513 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4514 to->affn_strict = from->affn_strict;
4515
4516 /*
4517 * Unlike hash and equality test, copying shouldn't ignore wq-only
4518 * fields as copying is used for both pool and wq attrs. Instead,
4519 * get_unbound_pool() explicitly clears the fields.
4520 */
4521 to->affn_scope = from->affn_scope;
4522 to->ordered = from->ordered;
4523}
4524
4525/*
4526 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4527 * comments in 'struct workqueue_attrs' definition.
4528 */
4529static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4530{
4531 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4532 attrs->ordered = false;
4533}
4534
4535/* hash value of the content of @attr */
4536static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4537{
4538 u32 hash = 0;
4539
4540 hash = jhash_1word(attrs->nice, hash);
4541 hash = jhash(cpumask_bits(attrs->cpumask),
4542 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4543 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4544 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4545 hash = jhash_1word(attrs->affn_strict, hash);
4546 return hash;
4547}
4548
4549/* content equality test */
4550static bool wqattrs_equal(const struct workqueue_attrs *a,
4551 const struct workqueue_attrs *b)
4552{
4553 if (a->nice != b->nice)
4554 return false;
4555 if (!cpumask_equal(a->cpumask, b->cpumask))
4556 return false;
4557 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4558 return false;
4559 if (a->affn_strict != b->affn_strict)
4560 return false;
4561 return true;
4562}
4563
4564/* Update @attrs with actually available CPUs */
4565static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4566 const cpumask_t *unbound_cpumask)
4567{
4568 /*
4569 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4570 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4571 * @unbound_cpumask.
4572 */
4573 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4574 if (unlikely(cpumask_empty(attrs->cpumask)))
4575 cpumask_copy(attrs->cpumask, unbound_cpumask);
4576}
4577
4578/* find wq_pod_type to use for @attrs */
4579static const struct wq_pod_type *
4580wqattrs_pod_type(const struct workqueue_attrs *attrs)
4581{
4582 enum wq_affn_scope scope;
4583 struct wq_pod_type *pt;
4584
4585 /* to synchronize access to wq_affn_dfl */
4586 lockdep_assert_held(&wq_pool_mutex);
4587
4588 if (attrs->affn_scope == WQ_AFFN_DFL)
4589 scope = wq_affn_dfl;
4590 else
4591 scope = attrs->affn_scope;
4592
4593 pt = &wq_pod_types[scope];
4594
4595 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4596 likely(pt->nr_pods))
4597 return pt;
4598
4599 /*
4600 * Before workqueue_init_topology(), only SYSTEM is available which is
4601 * initialized in workqueue_init_early().
4602 */
4603 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4604 BUG_ON(!pt->nr_pods);
4605 return pt;
4606}
4607
4608/**
4609 * init_worker_pool - initialize a newly zalloc'd worker_pool
4610 * @pool: worker_pool to initialize
4611 *
4612 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4613 *
4614 * Return: 0 on success, -errno on failure. Even on failure, all fields
4615 * inside @pool proper are initialized and put_unbound_pool() can be called
4616 * on @pool safely to release it.
4617 */
4618static int init_worker_pool(struct worker_pool *pool)
4619{
4620 raw_spin_lock_init(&pool->lock);
4621 pool->id = -1;
4622 pool->cpu = -1;
4623 pool->node = NUMA_NO_NODE;
4624 pool->flags |= POOL_DISASSOCIATED;
4625 pool->watchdog_ts = jiffies;
4626 INIT_LIST_HEAD(&pool->worklist);
4627 INIT_LIST_HEAD(&pool->idle_list);
4628 hash_init(pool->busy_hash);
4629
4630 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4631 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4632
4633 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4634
4635 INIT_LIST_HEAD(&pool->workers);
4636 INIT_LIST_HEAD(&pool->dying_workers);
4637
4638 ida_init(&pool->worker_ida);
4639 INIT_HLIST_NODE(&pool->hash_node);
4640 pool->refcnt = 1;
4641
4642 /* shouldn't fail above this point */
4643 pool->attrs = alloc_workqueue_attrs();
4644 if (!pool->attrs)
4645 return -ENOMEM;
4646
4647 wqattrs_clear_for_pool(pool->attrs);
4648
4649 return 0;
4650}
4651
4652#ifdef CONFIG_LOCKDEP
4653static void wq_init_lockdep(struct workqueue_struct *wq)
4654{
4655 char *lock_name;
4656
4657 lockdep_register_key(&wq->key);
4658 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4659 if (!lock_name)
4660 lock_name = wq->name;
4661
4662 wq->lock_name = lock_name;
4663 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
4664}
4665
4666static void wq_unregister_lockdep(struct workqueue_struct *wq)
4667{
4668 lockdep_unregister_key(&wq->key);
4669}
4670
4671static void wq_free_lockdep(struct workqueue_struct *wq)
4672{
4673 if (wq->lock_name != wq->name)
4674 kfree(wq->lock_name);
4675}
4676#else
4677static void wq_init_lockdep(struct workqueue_struct *wq)
4678{
4679}
4680
4681static void wq_unregister_lockdep(struct workqueue_struct *wq)
4682{
4683}
4684
4685static void wq_free_lockdep(struct workqueue_struct *wq)
4686{
4687}
4688#endif
4689
4690static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4691{
4692 int node;
4693
4694 for_each_node(node) {
4695 kfree(nna_ar[node]);
4696 nna_ar[node] = NULL;
4697 }
4698
4699 kfree(nna_ar[nr_node_ids]);
4700 nna_ar[nr_node_ids] = NULL;
4701}
4702
4703static void init_node_nr_active(struct wq_node_nr_active *nna)
4704{
4705 nna->max = WQ_DFL_MIN_ACTIVE;
4706 atomic_set(&nna->nr, 0);
4707 raw_spin_lock_init(&nna->lock);
4708 INIT_LIST_HEAD(&nna->pending_pwqs);
4709}
4710
4711/*
4712 * Each node's nr_active counter will be accessed mostly from its own node and
4713 * should be allocated in the node.
4714 */
4715static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4716{
4717 struct wq_node_nr_active *nna;
4718 int node;
4719
4720 for_each_node(node) {
4721 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4722 if (!nna)
4723 goto err_free;
4724 init_node_nr_active(nna);
4725 nna_ar[node] = nna;
4726 }
4727
4728 /* [nr_node_ids] is used as the fallback */
4729 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4730 if (!nna)
4731 goto err_free;
4732 init_node_nr_active(nna);
4733 nna_ar[nr_node_ids] = nna;
4734
4735 return 0;
4736
4737err_free:
4738 free_node_nr_active(nna_ar);
4739 return -ENOMEM;
4740}
4741
4742static void rcu_free_wq(struct rcu_head *rcu)
4743{
4744 struct workqueue_struct *wq =
4745 container_of(rcu, struct workqueue_struct, rcu);
4746
4747 if (wq->flags & WQ_UNBOUND)
4748 free_node_nr_active(wq->node_nr_active);
4749
4750 wq_free_lockdep(wq);
4751 free_percpu(wq->cpu_pwq);
4752 free_workqueue_attrs(wq->unbound_attrs);
4753 kfree(wq);
4754}
4755
4756static void rcu_free_pool(struct rcu_head *rcu)
4757{
4758 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4759
4760 ida_destroy(&pool->worker_ida);
4761 free_workqueue_attrs(pool->attrs);
4762 kfree(pool);
4763}
4764
4765/**
4766 * put_unbound_pool - put a worker_pool
4767 * @pool: worker_pool to put
4768 *
4769 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4770 * safe manner. get_unbound_pool() calls this function on its failure path
4771 * and this function should be able to release pools which went through,
4772 * successfully or not, init_worker_pool().
4773 *
4774 * Should be called with wq_pool_mutex held.
4775 */
4776static void put_unbound_pool(struct worker_pool *pool)
4777{
4778 DECLARE_COMPLETION_ONSTACK(detach_completion);
4779 struct worker *worker;
4780 LIST_HEAD(cull_list);
4781
4782 lockdep_assert_held(&wq_pool_mutex);
4783
4784 if (--pool->refcnt)
4785 return;
4786
4787 /* sanity checks */
4788 if (WARN_ON(!(pool->cpu < 0)) ||
4789 WARN_ON(!list_empty(&pool->worklist)))
4790 return;
4791
4792 /* release id and unhash */
4793 if (pool->id >= 0)
4794 idr_remove(&worker_pool_idr, pool->id);
4795 hash_del(&pool->hash_node);
4796
4797 /*
4798 * Become the manager and destroy all workers. This prevents
4799 * @pool's workers from blocking on attach_mutex. We're the last
4800 * manager and @pool gets freed with the flag set.
4801 *
4802 * Having a concurrent manager is quite unlikely to happen as we can
4803 * only get here with
4804 * pwq->refcnt == pool->refcnt == 0
4805 * which implies no work queued to the pool, which implies no worker can
4806 * become the manager. However a worker could have taken the role of
4807 * manager before the refcnts dropped to 0, since maybe_create_worker()
4808 * drops pool->lock
4809 */
4810 while (true) {
4811 rcuwait_wait_event(&manager_wait,
4812 !(pool->flags & POOL_MANAGER_ACTIVE),
4813 TASK_UNINTERRUPTIBLE);
4814
4815 mutex_lock(&wq_pool_attach_mutex);
4816 raw_spin_lock_irq(&pool->lock);
4817 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4818 pool->flags |= POOL_MANAGER_ACTIVE;
4819 break;
4820 }
4821 raw_spin_unlock_irq(&pool->lock);
4822 mutex_unlock(&wq_pool_attach_mutex);
4823 }
4824
4825 while ((worker = first_idle_worker(pool)))
4826 set_worker_dying(worker, &cull_list);
4827 WARN_ON(pool->nr_workers || pool->nr_idle);
4828 raw_spin_unlock_irq(&pool->lock);
4829
4830 wake_dying_workers(&cull_list);
4831
4832 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4833 pool->detach_completion = &detach_completion;
4834 mutex_unlock(&wq_pool_attach_mutex);
4835
4836 if (pool->detach_completion)
4837 wait_for_completion(pool->detach_completion);
4838
4839 /* shut down the timers */
4840 del_timer_sync(&pool->idle_timer);
4841 cancel_work_sync(&pool->idle_cull_work);
4842 del_timer_sync(&pool->mayday_timer);
4843
4844 /* RCU protected to allow dereferences from get_work_pool() */
4845 call_rcu(&pool->rcu, rcu_free_pool);
4846}
4847
4848/**
4849 * get_unbound_pool - get a worker_pool with the specified attributes
4850 * @attrs: the attributes of the worker_pool to get
4851 *
4852 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4853 * reference count and return it. If there already is a matching
4854 * worker_pool, it will be used; otherwise, this function attempts to
4855 * create a new one.
4856 *
4857 * Should be called with wq_pool_mutex held.
4858 *
4859 * Return: On success, a worker_pool with the same attributes as @attrs.
4860 * On failure, %NULL.
4861 */
4862static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4863{
4864 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4865 u32 hash = wqattrs_hash(attrs);
4866 struct worker_pool *pool;
4867 int pod, node = NUMA_NO_NODE;
4868
4869 lockdep_assert_held(&wq_pool_mutex);
4870
4871 /* do we already have a matching pool? */
4872 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4873 if (wqattrs_equal(pool->attrs, attrs)) {
4874 pool->refcnt++;
4875 return pool;
4876 }
4877 }
4878
4879 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4880 for (pod = 0; pod < pt->nr_pods; pod++) {
4881 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4882 node = pt->pod_node[pod];
4883 break;
4884 }
4885 }
4886
4887 /* nope, create a new one */
4888 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4889 if (!pool || init_worker_pool(pool) < 0)
4890 goto fail;
4891
4892 pool->node = node;
4893 copy_workqueue_attrs(pool->attrs, attrs);
4894 wqattrs_clear_for_pool(pool->attrs);
4895
4896 if (worker_pool_assign_id(pool) < 0)
4897 goto fail;
4898
4899 /* create and start the initial worker */
4900 if (wq_online && !create_worker(pool))
4901 goto fail;
4902
4903 /* install */
4904 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4905
4906 return pool;
4907fail:
4908 if (pool)
4909 put_unbound_pool(pool);
4910 return NULL;
4911}
4912
4913static void rcu_free_pwq(struct rcu_head *rcu)
4914{
4915 kmem_cache_free(pwq_cache,
4916 container_of(rcu, struct pool_workqueue, rcu));
4917}
4918
4919/*
4920 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4921 * refcnt and needs to be destroyed.
4922 */
4923static void pwq_release_workfn(struct kthread_work *work)
4924{
4925 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4926 release_work);
4927 struct workqueue_struct *wq = pwq->wq;
4928 struct worker_pool *pool = pwq->pool;
4929 bool is_last = false;
4930
4931 /*
4932 * When @pwq is not linked, it doesn't hold any reference to the
4933 * @wq, and @wq is invalid to access.
4934 */
4935 if (!list_empty(&pwq->pwqs_node)) {
4936 mutex_lock(&wq->mutex);
4937 list_del_rcu(&pwq->pwqs_node);
4938 is_last = list_empty(&wq->pwqs);
4939
4940 /*
4941 * For ordered workqueue with a plugged dfl_pwq, restart it now.
4942 */
4943 if (!is_last && (wq->flags & __WQ_ORDERED))
4944 unplug_oldest_pwq(wq);
4945
4946 mutex_unlock(&wq->mutex);
4947 }
4948
4949 if (wq->flags & WQ_UNBOUND) {
4950 mutex_lock(&wq_pool_mutex);
4951 put_unbound_pool(pool);
4952 mutex_unlock(&wq_pool_mutex);
4953 }
4954
4955 if (!list_empty(&pwq->pending_node)) {
4956 struct wq_node_nr_active *nna =
4957 wq_node_nr_active(pwq->wq, pwq->pool->node);
4958
4959 raw_spin_lock_irq(&nna->lock);
4960 list_del_init(&pwq->pending_node);
4961 raw_spin_unlock_irq(&nna->lock);
4962 }
4963
4964 call_rcu(&pwq->rcu, rcu_free_pwq);
4965
4966 /*
4967 * If we're the last pwq going away, @wq is already dead and no one
4968 * is gonna access it anymore. Schedule RCU free.
4969 */
4970 if (is_last) {
4971 wq_unregister_lockdep(wq);
4972 call_rcu(&wq->rcu, rcu_free_wq);
4973 }
4974}
4975
4976/* initialize newly allocated @pwq which is associated with @wq and @pool */
4977static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4978 struct worker_pool *pool)
4979{
4980 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
4981
4982 memset(pwq, 0, sizeof(*pwq));
4983
4984 pwq->pool = pool;
4985 pwq->wq = wq;
4986 pwq->flush_color = -1;
4987 pwq->refcnt = 1;
4988 INIT_LIST_HEAD(&pwq->inactive_works);
4989 INIT_LIST_HEAD(&pwq->pending_node);
4990 INIT_LIST_HEAD(&pwq->pwqs_node);
4991 INIT_LIST_HEAD(&pwq->mayday_node);
4992 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4993}
4994
4995/* sync @pwq with the current state of its associated wq and link it */
4996static void link_pwq(struct pool_workqueue *pwq)
4997{
4998 struct workqueue_struct *wq = pwq->wq;
4999
5000 lockdep_assert_held(&wq->mutex);
5001
5002 /* may be called multiple times, ignore if already linked */
5003 if (!list_empty(&pwq->pwqs_node))
5004 return;
5005
5006 /* set the matching work_color */
5007 pwq->work_color = wq->work_color;
5008
5009 /* link in @pwq */
5010 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5011}
5012
5013/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5014static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5015 const struct workqueue_attrs *attrs)
5016{
5017 struct worker_pool *pool;
5018 struct pool_workqueue *pwq;
5019
5020 lockdep_assert_held(&wq_pool_mutex);
5021
5022 pool = get_unbound_pool(attrs);
5023 if (!pool)
5024 return NULL;
5025
5026 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5027 if (!pwq) {
5028 put_unbound_pool(pool);
5029 return NULL;
5030 }
5031
5032 init_pwq(pwq, wq, pool);
5033 return pwq;
5034}
5035
5036/**
5037 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5038 * @attrs: the wq_attrs of the default pwq of the target workqueue
5039 * @cpu: the target CPU
5040 * @cpu_going_down: if >= 0, the CPU to consider as offline
5041 *
5042 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
5043 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
5044 * The result is stored in @attrs->__pod_cpumask.
5045 *
5046 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5047 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5048 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5049 *
5050 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5051 */
5052static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
5053 int cpu_going_down)
5054{
5055 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5056 int pod = pt->cpu_pod[cpu];
5057
5058 /* does @pod have any online CPUs @attrs wants? */
5059 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5060 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
5061 if (cpu_going_down >= 0)
5062 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
5063
5064 if (cpumask_empty(attrs->__pod_cpumask)) {
5065 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5066 return;
5067 }
5068
5069 /* yeap, return possible CPUs in @pod that @attrs wants */
5070 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
5071
5072 if (cpumask_empty(attrs->__pod_cpumask))
5073 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
5074 "possible intersect\n");
5075}
5076
5077/* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5078static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5079 int cpu, struct pool_workqueue *pwq)
5080{
5081 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5082 struct pool_workqueue *old_pwq;
5083
5084 lockdep_assert_held(&wq_pool_mutex);
5085 lockdep_assert_held(&wq->mutex);
5086
5087 /* link_pwq() can handle duplicate calls */
5088 link_pwq(pwq);
5089
5090 old_pwq = rcu_access_pointer(*slot);
5091 rcu_assign_pointer(*slot, pwq);
5092 return old_pwq;
5093}
5094
5095/* context to store the prepared attrs & pwqs before applying */
5096struct apply_wqattrs_ctx {
5097 struct workqueue_struct *wq; /* target workqueue */
5098 struct workqueue_attrs *attrs; /* attrs to apply */
5099 struct list_head list; /* queued for batching commit */
5100 struct pool_workqueue *dfl_pwq;
5101 struct pool_workqueue *pwq_tbl[];
5102};
5103
5104/* free the resources after success or abort */
5105static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5106{
5107 if (ctx) {
5108 int cpu;
5109
5110 for_each_possible_cpu(cpu)
5111 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5112 put_pwq_unlocked(ctx->dfl_pwq);
5113
5114 free_workqueue_attrs(ctx->attrs);
5115
5116 kfree(ctx);
5117 }
5118}
5119
5120/* allocate the attrs and pwqs for later installation */
5121static struct apply_wqattrs_ctx *
5122apply_wqattrs_prepare(struct workqueue_struct *wq,
5123 const struct workqueue_attrs *attrs,
5124 const cpumask_var_t unbound_cpumask)
5125{
5126 struct apply_wqattrs_ctx *ctx;
5127 struct workqueue_attrs *new_attrs;
5128 int cpu;
5129
5130 lockdep_assert_held(&wq_pool_mutex);
5131
5132 if (WARN_ON(attrs->affn_scope < 0 ||
5133 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5134 return ERR_PTR(-EINVAL);
5135
5136 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5137
5138 new_attrs = alloc_workqueue_attrs();
5139 if (!ctx || !new_attrs)
5140 goto out_free;
5141
5142 /*
5143 * If something goes wrong during CPU up/down, we'll fall back to
5144 * the default pwq covering whole @attrs->cpumask. Always create
5145 * it even if we don't use it immediately.
5146 */
5147 copy_workqueue_attrs(new_attrs, attrs);
5148 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5149 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5150 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5151 if (!ctx->dfl_pwq)
5152 goto out_free;
5153
5154 for_each_possible_cpu(cpu) {
5155 if (new_attrs->ordered) {
5156 ctx->dfl_pwq->refcnt++;
5157 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5158 } else {
5159 wq_calc_pod_cpumask(new_attrs, cpu, -1);
5160 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5161 if (!ctx->pwq_tbl[cpu])
5162 goto out_free;
5163 }
5164 }
5165
5166 /* save the user configured attrs and sanitize it. */
5167 copy_workqueue_attrs(new_attrs, attrs);
5168 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5169 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5170 ctx->attrs = new_attrs;
5171
5172 /*
5173 * For initialized ordered workqueues, there should only be one pwq
5174 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5175 * of newly queued work items until execution of older work items in
5176 * the old pwq's have completed.
5177 */
5178 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5179 ctx->dfl_pwq->plugged = true;
5180
5181 ctx->wq = wq;
5182 return ctx;
5183
5184out_free:
5185 free_workqueue_attrs(new_attrs);
5186 apply_wqattrs_cleanup(ctx);
5187 return ERR_PTR(-ENOMEM);
5188}
5189
5190/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5191static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5192{
5193 int cpu;
5194
5195 /* all pwqs have been created successfully, let's install'em */
5196 mutex_lock(&ctx->wq->mutex);
5197
5198 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5199
5200 /* save the previous pwqs and install the new ones */
5201 for_each_possible_cpu(cpu)
5202 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5203 ctx->pwq_tbl[cpu]);
5204 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5205
5206 /* update node_nr_active->max */
5207 wq_update_node_max_active(ctx->wq, -1);
5208
5209 /* rescuer needs to respect wq cpumask changes */
5210 if (ctx->wq->rescuer)
5211 set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5212 unbound_effective_cpumask(ctx->wq));
5213
5214 mutex_unlock(&ctx->wq->mutex);
5215}
5216
5217static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5218 const struct workqueue_attrs *attrs)
5219{
5220 struct apply_wqattrs_ctx *ctx;
5221
5222 /* only unbound workqueues can change attributes */
5223 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5224 return -EINVAL;
5225
5226 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5227 if (IS_ERR(ctx))
5228 return PTR_ERR(ctx);
5229
5230 /* the ctx has been prepared successfully, let's commit it */
5231 apply_wqattrs_commit(ctx);
5232 apply_wqattrs_cleanup(ctx);
5233
5234 return 0;
5235}
5236
5237/**
5238 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5239 * @wq: the target workqueue
5240 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5241 *
5242 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5243 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5244 * work items are affine to the pod it was issued on. Older pwqs are released as
5245 * in-flight work items finish. Note that a work item which repeatedly requeues
5246 * itself back-to-back will stay on its current pwq.
5247 *
5248 * Performs GFP_KERNEL allocations.
5249 *
5250 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
5251 *
5252 * Return: 0 on success and -errno on failure.
5253 */
5254int apply_workqueue_attrs(struct workqueue_struct *wq,
5255 const struct workqueue_attrs *attrs)
5256{
5257 int ret;
5258
5259 lockdep_assert_cpus_held();
5260
5261 mutex_lock(&wq_pool_mutex);
5262 ret = apply_workqueue_attrs_locked(wq, attrs);
5263 mutex_unlock(&wq_pool_mutex);
5264
5265 return ret;
5266}
5267
5268/**
5269 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
5270 * @wq: the target workqueue
5271 * @cpu: the CPU to update pool association for
5272 * @hotplug_cpu: the CPU coming up or going down
5273 * @online: whether @cpu is coming up or going down
5274 *
5275 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5276 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
5277 * @wq accordingly.
5278 *
5279 *
5280 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5281 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5282 *
5283 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5284 * with a cpumask spanning multiple pods, the workers which were already
5285 * executing the work items for the workqueue will lose their CPU affinity and
5286 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5287 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5288 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5289 */
5290static void wq_update_pod(struct workqueue_struct *wq, int cpu,
5291 int hotplug_cpu, bool online)
5292{
5293 int off_cpu = online ? -1 : hotplug_cpu;
5294 struct pool_workqueue *old_pwq = NULL, *pwq;
5295 struct workqueue_attrs *target_attrs;
5296
5297 lockdep_assert_held(&wq_pool_mutex);
5298
5299 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5300 return;
5301
5302 /*
5303 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5304 * Let's use a preallocated one. The following buf is protected by
5305 * CPU hotplug exclusion.
5306 */
5307 target_attrs = wq_update_pod_attrs_buf;
5308
5309 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5310 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5311
5312 /* nothing to do if the target cpumask matches the current pwq */
5313 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
5314 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5315 return;
5316
5317 /* create a new pwq */
5318 pwq = alloc_unbound_pwq(wq, target_attrs);
5319 if (!pwq) {
5320 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5321 wq->name);
5322 goto use_dfl_pwq;
5323 }
5324
5325 /* Install the new pwq. */
5326 mutex_lock(&wq->mutex);
5327 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5328 goto out_unlock;
5329
5330use_dfl_pwq:
5331 mutex_lock(&wq->mutex);
5332 pwq = unbound_pwq(wq, -1);
5333 raw_spin_lock_irq(&pwq->pool->lock);
5334 get_pwq(pwq);
5335 raw_spin_unlock_irq(&pwq->pool->lock);
5336 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5337out_unlock:
5338 mutex_unlock(&wq->mutex);
5339 put_pwq_unlocked(old_pwq);
5340}
5341
5342static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5343{
5344 bool highpri = wq->flags & WQ_HIGHPRI;
5345 int cpu, ret;
5346
5347 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5348 if (!wq->cpu_pwq)
5349 goto enomem;
5350
5351 if (!(wq->flags & WQ_UNBOUND)) {
5352 for_each_possible_cpu(cpu) {
5353 struct pool_workqueue **pwq_p;
5354 struct worker_pool __percpu *pools;
5355 struct worker_pool *pool;
5356
5357 if (wq->flags & WQ_BH)
5358 pools = bh_worker_pools;
5359 else
5360 pools = cpu_worker_pools;
5361
5362 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5363 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5364
5365 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5366 pool->node);
5367 if (!*pwq_p)
5368 goto enomem;
5369
5370 init_pwq(*pwq_p, wq, pool);
5371
5372 mutex_lock(&wq->mutex);
5373 link_pwq(*pwq_p);
5374 mutex_unlock(&wq->mutex);
5375 }
5376 return 0;
5377 }
5378
5379 cpus_read_lock();
5380 if (wq->flags & __WQ_ORDERED) {
5381 struct pool_workqueue *dfl_pwq;
5382
5383 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
5384 /* there should only be single pwq for ordering guarantee */
5385 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5386 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5387 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5388 "ordering guarantee broken for workqueue %s\n", wq->name);
5389 } else {
5390 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
5391 }
5392 cpus_read_unlock();
5393
5394 /* for unbound pwq, flush the pwq_release_worker ensures that the
5395 * pwq_release_workfn() completes before calling kfree(wq).
5396 */
5397 if (ret)
5398 kthread_flush_worker(pwq_release_worker);
5399
5400 return ret;
5401
5402enomem:
5403 if (wq->cpu_pwq) {
5404 for_each_possible_cpu(cpu) {
5405 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5406
5407 if (pwq)
5408 kmem_cache_free(pwq_cache, pwq);
5409 }
5410 free_percpu(wq->cpu_pwq);
5411 wq->cpu_pwq = NULL;
5412 }
5413 return -ENOMEM;
5414}
5415
5416static int wq_clamp_max_active(int max_active, unsigned int flags,
5417 const char *name)
5418{
5419 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5420 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5421 max_active, name, 1, WQ_MAX_ACTIVE);
5422
5423 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5424}
5425
5426/*
5427 * Workqueues which may be used during memory reclaim should have a rescuer
5428 * to guarantee forward progress.
5429 */
5430static int init_rescuer(struct workqueue_struct *wq)
5431{
5432 struct worker *rescuer;
5433 int ret;
5434
5435 if (!(wq->flags & WQ_MEM_RECLAIM))
5436 return 0;
5437
5438 rescuer = alloc_worker(NUMA_NO_NODE);
5439 if (!rescuer) {
5440 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5441 wq->name);
5442 return -ENOMEM;
5443 }
5444
5445 rescuer->rescue_wq = wq;
5446 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
5447 if (IS_ERR(rescuer->task)) {
5448 ret = PTR_ERR(rescuer->task);
5449 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5450 wq->name, ERR_PTR(ret));
5451 kfree(rescuer);
5452 return ret;
5453 }
5454
5455 wq->rescuer = rescuer;
5456 if (wq->flags & WQ_UNBOUND)
5457 kthread_bind_mask(rescuer->task, wq_unbound_cpumask);
5458 else
5459 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5460 wake_up_process(rescuer->task);
5461
5462 return 0;
5463}
5464
5465/**
5466 * wq_adjust_max_active - update a wq's max_active to the current setting
5467 * @wq: target workqueue
5468 *
5469 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5470 * activate inactive work items accordingly. If @wq is freezing, clear
5471 * @wq->max_active to zero.
5472 */
5473static void wq_adjust_max_active(struct workqueue_struct *wq)
5474{
5475 bool activated;
5476 int new_max, new_min;
5477
5478 lockdep_assert_held(&wq->mutex);
5479
5480 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5481 new_max = 0;
5482 new_min = 0;
5483 } else {
5484 new_max = wq->saved_max_active;
5485 new_min = wq->saved_min_active;
5486 }
5487
5488 if (wq->max_active == new_max && wq->min_active == new_min)
5489 return;
5490
5491 /*
5492 * Update @wq->max/min_active and then kick inactive work items if more
5493 * active work items are allowed. This doesn't break work item ordering
5494 * because new work items are always queued behind existing inactive
5495 * work items if there are any.
5496 */
5497 WRITE_ONCE(wq->max_active, new_max);
5498 WRITE_ONCE(wq->min_active, new_min);
5499
5500 if (wq->flags & WQ_UNBOUND)
5501 wq_update_node_max_active(wq, -1);
5502
5503 if (new_max == 0)
5504 return;
5505
5506 /*
5507 * Round-robin through pwq's activating the first inactive work item
5508 * until max_active is filled.
5509 */
5510 do {
5511 struct pool_workqueue *pwq;
5512
5513 activated = false;
5514 for_each_pwq(pwq, wq) {
5515 unsigned long irq_flags;
5516
5517 /* can be called during early boot w/ irq disabled */
5518 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5519 if (pwq_activate_first_inactive(pwq, true)) {
5520 activated = true;
5521 kick_pool(pwq->pool);
5522 }
5523 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5524 }
5525 } while (activated);
5526}
5527
5528__printf(1, 4)
5529struct workqueue_struct *alloc_workqueue(const char *fmt,
5530 unsigned int flags,
5531 int max_active, ...)
5532{
5533 va_list args;
5534 struct workqueue_struct *wq;
5535 size_t wq_size;
5536 int name_len;
5537
5538 if (flags & WQ_BH) {
5539 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5540 return NULL;
5541 if (WARN_ON_ONCE(max_active))
5542 return NULL;
5543 }
5544
5545 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5546 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5547 flags |= WQ_UNBOUND;
5548
5549 /* allocate wq and format name */
5550 if (flags & WQ_UNBOUND)
5551 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5552 else
5553 wq_size = sizeof(*wq);
5554
5555 wq = kzalloc(wq_size, GFP_KERNEL);
5556 if (!wq)
5557 return NULL;
5558
5559 if (flags & WQ_UNBOUND) {
5560 wq->unbound_attrs = alloc_workqueue_attrs();
5561 if (!wq->unbound_attrs)
5562 goto err_free_wq;
5563 }
5564
5565 va_start(args, max_active);
5566 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5567 va_end(args);
5568
5569 if (name_len >= WQ_NAME_LEN)
5570 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5571 wq->name);
5572
5573 if (flags & WQ_BH) {
5574 /*
5575 * BH workqueues always share a single execution context per CPU
5576 * and don't impose any max_active limit.
5577 */
5578 max_active = INT_MAX;
5579 } else {
5580 max_active = max_active ?: WQ_DFL_ACTIVE;
5581 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5582 }
5583
5584 /* init wq */
5585 wq->flags = flags;
5586 wq->max_active = max_active;
5587 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5588 wq->saved_max_active = wq->max_active;
5589 wq->saved_min_active = wq->min_active;
5590 mutex_init(&wq->mutex);
5591 atomic_set(&wq->nr_pwqs_to_flush, 0);
5592 INIT_LIST_HEAD(&wq->pwqs);
5593 INIT_LIST_HEAD(&wq->flusher_queue);
5594 INIT_LIST_HEAD(&wq->flusher_overflow);
5595 INIT_LIST_HEAD(&wq->maydays);
5596
5597 wq_init_lockdep(wq);
5598 INIT_LIST_HEAD(&wq->list);
5599
5600 if (flags & WQ_UNBOUND) {
5601 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5602 goto err_unreg_lockdep;
5603 }
5604
5605 if (alloc_and_link_pwqs(wq) < 0)
5606 goto err_free_node_nr_active;
5607
5608 if (wq_online && init_rescuer(wq) < 0)
5609 goto err_destroy;
5610
5611 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5612 goto err_destroy;
5613
5614 /*
5615 * wq_pool_mutex protects global freeze state and workqueues list.
5616 * Grab it, adjust max_active and add the new @wq to workqueues
5617 * list.
5618 */
5619 mutex_lock(&wq_pool_mutex);
5620
5621 mutex_lock(&wq->mutex);
5622 wq_adjust_max_active(wq);
5623 mutex_unlock(&wq->mutex);
5624
5625 list_add_tail_rcu(&wq->list, &workqueues);
5626
5627 mutex_unlock(&wq_pool_mutex);
5628
5629 return wq;
5630
5631err_free_node_nr_active:
5632 if (wq->flags & WQ_UNBOUND)
5633 free_node_nr_active(wq->node_nr_active);
5634err_unreg_lockdep:
5635 wq_unregister_lockdep(wq);
5636 wq_free_lockdep(wq);
5637err_free_wq:
5638 free_workqueue_attrs(wq->unbound_attrs);
5639 kfree(wq);
5640 return NULL;
5641err_destroy:
5642 destroy_workqueue(wq);
5643 return NULL;
5644}
5645EXPORT_SYMBOL_GPL(alloc_workqueue);
5646
5647static bool pwq_busy(struct pool_workqueue *pwq)
5648{
5649 int i;
5650
5651 for (i = 0; i < WORK_NR_COLORS; i++)
5652 if (pwq->nr_in_flight[i])
5653 return true;
5654
5655 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5656 return true;
5657 if (!pwq_is_empty(pwq))
5658 return true;
5659
5660 return false;
5661}
5662
5663/**
5664 * destroy_workqueue - safely terminate a workqueue
5665 * @wq: target workqueue
5666 *
5667 * Safely destroy a workqueue. All work currently pending will be done first.
5668 */
5669void destroy_workqueue(struct workqueue_struct *wq)
5670{
5671 struct pool_workqueue *pwq;
5672 int cpu;
5673
5674 /*
5675 * Remove it from sysfs first so that sanity check failure doesn't
5676 * lead to sysfs name conflicts.
5677 */
5678 workqueue_sysfs_unregister(wq);
5679
5680 /* mark the workqueue destruction is in progress */
5681 mutex_lock(&wq->mutex);
5682 wq->flags |= __WQ_DESTROYING;
5683 mutex_unlock(&wq->mutex);
5684
5685 /* drain it before proceeding with destruction */
5686 drain_workqueue(wq);
5687
5688 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5689 if (wq->rescuer) {
5690 struct worker *rescuer = wq->rescuer;
5691
5692 /* this prevents new queueing */
5693 raw_spin_lock_irq(&wq_mayday_lock);
5694 wq->rescuer = NULL;
5695 raw_spin_unlock_irq(&wq_mayday_lock);
5696
5697 /* rescuer will empty maydays list before exiting */
5698 kthread_stop(rescuer->task);
5699 kfree(rescuer);
5700 }
5701
5702 /*
5703 * Sanity checks - grab all the locks so that we wait for all
5704 * in-flight operations which may do put_pwq().
5705 */
5706 mutex_lock(&wq_pool_mutex);
5707 mutex_lock(&wq->mutex);
5708 for_each_pwq(pwq, wq) {
5709 raw_spin_lock_irq(&pwq->pool->lock);
5710 if (WARN_ON(pwq_busy(pwq))) {
5711 pr_warn("%s: %s has the following busy pwq\n",
5712 __func__, wq->name);
5713 show_pwq(pwq);
5714 raw_spin_unlock_irq(&pwq->pool->lock);
5715 mutex_unlock(&wq->mutex);
5716 mutex_unlock(&wq_pool_mutex);
5717 show_one_workqueue(wq);
5718 return;
5719 }
5720 raw_spin_unlock_irq(&pwq->pool->lock);
5721 }
5722 mutex_unlock(&wq->mutex);
5723
5724 /*
5725 * wq list is used to freeze wq, remove from list after
5726 * flushing is complete in case freeze races us.
5727 */
5728 list_del_rcu(&wq->list);
5729 mutex_unlock(&wq_pool_mutex);
5730
5731 /*
5732 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5733 * to put the base refs. @wq will be auto-destroyed from the last
5734 * pwq_put. RCU read lock prevents @wq from going away from under us.
5735 */
5736 rcu_read_lock();
5737
5738 for_each_possible_cpu(cpu) {
5739 put_pwq_unlocked(unbound_pwq(wq, cpu));
5740 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5741 }
5742
5743 put_pwq_unlocked(unbound_pwq(wq, -1));
5744 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5745
5746 rcu_read_unlock();
5747}
5748EXPORT_SYMBOL_GPL(destroy_workqueue);
5749
5750/**
5751 * workqueue_set_max_active - adjust max_active of a workqueue
5752 * @wq: target workqueue
5753 * @max_active: new max_active value.
5754 *
5755 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5756 * comment.
5757 *
5758 * CONTEXT:
5759 * Don't call from IRQ context.
5760 */
5761void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5762{
5763 /* max_active doesn't mean anything for BH workqueues */
5764 if (WARN_ON(wq->flags & WQ_BH))
5765 return;
5766 /* disallow meddling with max_active for ordered workqueues */
5767 if (WARN_ON(wq->flags & __WQ_ORDERED))
5768 return;
5769
5770 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5771
5772 mutex_lock(&wq->mutex);
5773
5774 wq->saved_max_active = max_active;
5775 if (wq->flags & WQ_UNBOUND)
5776 wq->saved_min_active = min(wq->saved_min_active, max_active);
5777
5778 wq_adjust_max_active(wq);
5779
5780 mutex_unlock(&wq->mutex);
5781}
5782EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5783
5784/**
5785 * workqueue_set_min_active - adjust min_active of an unbound workqueue
5786 * @wq: target unbound workqueue
5787 * @min_active: new min_active value
5788 *
5789 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5790 * unbound workqueue is not guaranteed to be able to process max_active
5791 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5792 * able to process min_active number of interdependent work items which is
5793 * %WQ_DFL_MIN_ACTIVE by default.
5794 *
5795 * Use this function to adjust the min_active value between 0 and the current
5796 * max_active.
5797 */
5798void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5799{
5800 /* min_active is only meaningful for non-ordered unbound workqueues */
5801 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5802 WQ_UNBOUND))
5803 return;
5804
5805 mutex_lock(&wq->mutex);
5806 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5807 wq_adjust_max_active(wq);
5808 mutex_unlock(&wq->mutex);
5809}
5810
5811/**
5812 * current_work - retrieve %current task's work struct
5813 *
5814 * Determine if %current task is a workqueue worker and what it's working on.
5815 * Useful to find out the context that the %current task is running in.
5816 *
5817 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5818 */
5819struct work_struct *current_work(void)
5820{
5821 struct worker *worker = current_wq_worker();
5822
5823 return worker ? worker->current_work : NULL;
5824}
5825EXPORT_SYMBOL(current_work);
5826
5827/**
5828 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5829 *
5830 * Determine whether %current is a workqueue rescuer. Can be used from
5831 * work functions to determine whether it's being run off the rescuer task.
5832 *
5833 * Return: %true if %current is a workqueue rescuer. %false otherwise.
5834 */
5835bool current_is_workqueue_rescuer(void)
5836{
5837 struct worker *worker = current_wq_worker();
5838
5839 return worker && worker->rescue_wq;
5840}
5841
5842/**
5843 * workqueue_congested - test whether a workqueue is congested
5844 * @cpu: CPU in question
5845 * @wq: target workqueue
5846 *
5847 * Test whether @wq's cpu workqueue for @cpu is congested. There is
5848 * no synchronization around this function and the test result is
5849 * unreliable and only useful as advisory hints or for debugging.
5850 *
5851 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
5852 *
5853 * With the exception of ordered workqueues, all workqueues have per-cpu
5854 * pool_workqueues, each with its own congested state. A workqueue being
5855 * congested on one CPU doesn't mean that the workqueue is contested on any
5856 * other CPUs.
5857 *
5858 * Return:
5859 * %true if congested, %false otherwise.
5860 */
5861bool workqueue_congested(int cpu, struct workqueue_struct *wq)
5862{
5863 struct pool_workqueue *pwq;
5864 bool ret;
5865
5866 rcu_read_lock();
5867 preempt_disable();
5868
5869 if (cpu == WORK_CPU_UNBOUND)
5870 cpu = smp_processor_id();
5871
5872 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5873 ret = !list_empty(&pwq->inactive_works);
5874
5875 preempt_enable();
5876 rcu_read_unlock();
5877
5878 return ret;
5879}
5880EXPORT_SYMBOL_GPL(workqueue_congested);
5881
5882/**
5883 * work_busy - test whether a work is currently pending or running
5884 * @work: the work to be tested
5885 *
5886 * Test whether @work is currently pending or running. There is no
5887 * synchronization around this function and the test result is
5888 * unreliable and only useful as advisory hints or for debugging.
5889 *
5890 * Return:
5891 * OR'd bitmask of WORK_BUSY_* bits.
5892 */
5893unsigned int work_busy(struct work_struct *work)
5894{
5895 struct worker_pool *pool;
5896 unsigned long irq_flags;
5897 unsigned int ret = 0;
5898
5899 if (work_pending(work))
5900 ret |= WORK_BUSY_PENDING;
5901
5902 rcu_read_lock();
5903 pool = get_work_pool(work);
5904 if (pool) {
5905 raw_spin_lock_irqsave(&pool->lock, irq_flags);
5906 if (find_worker_executing_work(pool, work))
5907 ret |= WORK_BUSY_RUNNING;
5908 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
5909 }
5910 rcu_read_unlock();
5911
5912 return ret;
5913}
5914EXPORT_SYMBOL_GPL(work_busy);
5915
5916/**
5917 * set_worker_desc - set description for the current work item
5918 * @fmt: printf-style format string
5919 * @...: arguments for the format string
5920 *
5921 * This function can be called by a running work function to describe what
5922 * the work item is about. If the worker task gets dumped, this
5923 * information will be printed out together to help debugging. The
5924 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5925 */
5926void set_worker_desc(const char *fmt, ...)
5927{
5928 struct worker *worker = current_wq_worker();
5929 va_list args;
5930
5931 if (worker) {
5932 va_start(args, fmt);
5933 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5934 va_end(args);
5935 }
5936}
5937EXPORT_SYMBOL_GPL(set_worker_desc);
5938
5939/**
5940 * print_worker_info - print out worker information and description
5941 * @log_lvl: the log level to use when printing
5942 * @task: target task
5943 *
5944 * If @task is a worker and currently executing a work item, print out the
5945 * name of the workqueue being serviced and worker description set with
5946 * set_worker_desc() by the currently executing work item.
5947 *
5948 * This function can be safely called on any task as long as the
5949 * task_struct itself is accessible. While safe, this function isn't
5950 * synchronized and may print out mixups or garbages of limited length.
5951 */
5952void print_worker_info(const char *log_lvl, struct task_struct *task)
5953{
5954 work_func_t *fn = NULL;
5955 char name[WQ_NAME_LEN] = { };
5956 char desc[WORKER_DESC_LEN] = { };
5957 struct pool_workqueue *pwq = NULL;
5958 struct workqueue_struct *wq = NULL;
5959 struct worker *worker;
5960
5961 if (!(task->flags & PF_WQ_WORKER))
5962 return;
5963
5964 /*
5965 * This function is called without any synchronization and @task
5966 * could be in any state. Be careful with dereferences.
5967 */
5968 worker = kthread_probe_data(task);
5969
5970 /*
5971 * Carefully copy the associated workqueue's workfn, name and desc.
5972 * Keep the original last '\0' in case the original is garbage.
5973 */
5974 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5975 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5976 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5977 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5978 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5979
5980 if (fn || name[0] || desc[0]) {
5981 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5982 if (strcmp(name, desc))
5983 pr_cont(" (%s)", desc);
5984 pr_cont("\n");
5985 }
5986}
5987
5988static void pr_cont_pool_info(struct worker_pool *pool)
5989{
5990 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5991 if (pool->node != NUMA_NO_NODE)
5992 pr_cont(" node=%d", pool->node);
5993 pr_cont(" flags=0x%x", pool->flags);
5994 if (pool->flags & POOL_BH)
5995 pr_cont(" bh%s",
5996 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
5997 else
5998 pr_cont(" nice=%d", pool->attrs->nice);
5999}
6000
6001static void pr_cont_worker_id(struct worker *worker)
6002{
6003 struct worker_pool *pool = worker->pool;
6004
6005 if (pool->flags & WQ_BH)
6006 pr_cont("bh%s",
6007 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6008 else
6009 pr_cont("%d%s", task_pid_nr(worker->task),
6010 worker->rescue_wq ? "(RESCUER)" : "");
6011}
6012
6013struct pr_cont_work_struct {
6014 bool comma;
6015 work_func_t func;
6016 long ctr;
6017};
6018
6019static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6020{
6021 if (!pcwsp->ctr)
6022 goto out_record;
6023 if (func == pcwsp->func) {
6024 pcwsp->ctr++;
6025 return;
6026 }
6027 if (pcwsp->ctr == 1)
6028 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6029 else
6030 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6031 pcwsp->ctr = 0;
6032out_record:
6033 if ((long)func == -1L)
6034 return;
6035 pcwsp->comma = comma;
6036 pcwsp->func = func;
6037 pcwsp->ctr = 1;
6038}
6039
6040static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6041{
6042 if (work->func == wq_barrier_func) {
6043 struct wq_barrier *barr;
6044
6045 barr = container_of(work, struct wq_barrier, work);
6046
6047 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6048 pr_cont("%s BAR(%d)", comma ? "," : "",
6049 task_pid_nr(barr->task));
6050 } else {
6051 if (!comma)
6052 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6053 pr_cont_work_flush(comma, work->func, pcwsp);
6054 }
6055}
6056
6057static void show_pwq(struct pool_workqueue *pwq)
6058{
6059 struct pr_cont_work_struct pcws = { .ctr = 0, };
6060 struct worker_pool *pool = pwq->pool;
6061 struct work_struct *work;
6062 struct worker *worker;
6063 bool has_in_flight = false, has_pending = false;
6064 int bkt;
6065
6066 pr_info(" pwq %d:", pool->id);
6067 pr_cont_pool_info(pool);
6068
6069 pr_cont(" active=%d refcnt=%d%s\n",
6070 pwq->nr_active, pwq->refcnt,
6071 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6072
6073 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6074 if (worker->current_pwq == pwq) {
6075 has_in_flight = true;
6076 break;
6077 }
6078 }
6079 if (has_in_flight) {
6080 bool comma = false;
6081
6082 pr_info(" in-flight:");
6083 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6084 if (worker->current_pwq != pwq)
6085 continue;
6086
6087 pr_cont(" %s", comma ? "," : "");
6088 pr_cont_worker_id(worker);
6089 pr_cont(":%ps", worker->current_func);
6090 list_for_each_entry(work, &worker->scheduled, entry)
6091 pr_cont_work(false, work, &pcws);
6092 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6093 comma = true;
6094 }
6095 pr_cont("\n");
6096 }
6097
6098 list_for_each_entry(work, &pool->worklist, entry) {
6099 if (get_work_pwq(work) == pwq) {
6100 has_pending = true;
6101 break;
6102 }
6103 }
6104 if (has_pending) {
6105 bool comma = false;
6106
6107 pr_info(" pending:");
6108 list_for_each_entry(work, &pool->worklist, entry) {
6109 if (get_work_pwq(work) != pwq)
6110 continue;
6111
6112 pr_cont_work(comma, work, &pcws);
6113 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6114 }
6115 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6116 pr_cont("\n");
6117 }
6118
6119 if (!list_empty(&pwq->inactive_works)) {
6120 bool comma = false;
6121
6122 pr_info(" inactive:");
6123 list_for_each_entry(work, &pwq->inactive_works, entry) {
6124 pr_cont_work(comma, work, &pcws);
6125 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6126 }
6127 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6128 pr_cont("\n");
6129 }
6130}
6131
6132/**
6133 * show_one_workqueue - dump state of specified workqueue
6134 * @wq: workqueue whose state will be printed
6135 */
6136void show_one_workqueue(struct workqueue_struct *wq)
6137{
6138 struct pool_workqueue *pwq;
6139 bool idle = true;
6140 unsigned long irq_flags;
6141
6142 for_each_pwq(pwq, wq) {
6143 if (!pwq_is_empty(pwq)) {
6144 idle = false;
6145 break;
6146 }
6147 }
6148 if (idle) /* Nothing to print for idle workqueue */
6149 return;
6150
6151 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6152
6153 for_each_pwq(pwq, wq) {
6154 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6155 if (!pwq_is_empty(pwq)) {
6156 /*
6157 * Defer printing to avoid deadlocks in console
6158 * drivers that queue work while holding locks
6159 * also taken in their write paths.
6160 */
6161 printk_deferred_enter();
6162 show_pwq(pwq);
6163 printk_deferred_exit();
6164 }
6165 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6166 /*
6167 * We could be printing a lot from atomic context, e.g.
6168 * sysrq-t -> show_all_workqueues(). Avoid triggering
6169 * hard lockup.
6170 */
6171 touch_nmi_watchdog();
6172 }
6173
6174}
6175
6176/**
6177 * show_one_worker_pool - dump state of specified worker pool
6178 * @pool: worker pool whose state will be printed
6179 */
6180static void show_one_worker_pool(struct worker_pool *pool)
6181{
6182 struct worker *worker;
6183 bool first = true;
6184 unsigned long irq_flags;
6185 unsigned long hung = 0;
6186
6187 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6188 if (pool->nr_workers == pool->nr_idle)
6189 goto next_pool;
6190
6191 /* How long the first pending work is waiting for a worker. */
6192 if (!list_empty(&pool->worklist))
6193 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6194
6195 /*
6196 * Defer printing to avoid deadlocks in console drivers that
6197 * queue work while holding locks also taken in their write
6198 * paths.
6199 */
6200 printk_deferred_enter();
6201 pr_info("pool %d:", pool->id);
6202 pr_cont_pool_info(pool);
6203 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6204 if (pool->manager)
6205 pr_cont(" manager: %d",
6206 task_pid_nr(pool->manager->task));
6207 list_for_each_entry(worker, &pool->idle_list, entry) {
6208 pr_cont(" %s", first ? "idle: " : "");
6209 pr_cont_worker_id(worker);
6210 first = false;
6211 }
6212 pr_cont("\n");
6213 printk_deferred_exit();
6214next_pool:
6215 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6216 /*
6217 * We could be printing a lot from atomic context, e.g.
6218 * sysrq-t -> show_all_workqueues(). Avoid triggering
6219 * hard lockup.
6220 */
6221 touch_nmi_watchdog();
6222
6223}
6224
6225/**
6226 * show_all_workqueues - dump workqueue state
6227 *
6228 * Called from a sysrq handler and prints out all busy workqueues and pools.
6229 */
6230void show_all_workqueues(void)
6231{
6232 struct workqueue_struct *wq;
6233 struct worker_pool *pool;
6234 int pi;
6235
6236 rcu_read_lock();
6237
6238 pr_info("Showing busy workqueues and worker pools:\n");
6239
6240 list_for_each_entry_rcu(wq, &workqueues, list)
6241 show_one_workqueue(wq);
6242
6243 for_each_pool(pool, pi)
6244 show_one_worker_pool(pool);
6245
6246 rcu_read_unlock();
6247}
6248
6249/**
6250 * show_freezable_workqueues - dump freezable workqueue state
6251 *
6252 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6253 * still busy.
6254 */
6255void show_freezable_workqueues(void)
6256{
6257 struct workqueue_struct *wq;
6258
6259 rcu_read_lock();
6260
6261 pr_info("Showing freezable workqueues that are still busy:\n");
6262
6263 list_for_each_entry_rcu(wq, &workqueues, list) {
6264 if (!(wq->flags & WQ_FREEZABLE))
6265 continue;
6266 show_one_workqueue(wq);
6267 }
6268
6269 rcu_read_unlock();
6270}
6271
6272/* used to show worker information through /proc/PID/{comm,stat,status} */
6273void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6274{
6275 int off;
6276
6277 /* always show the actual comm */
6278 off = strscpy(buf, task->comm, size);
6279 if (off < 0)
6280 return;
6281
6282 /* stabilize PF_WQ_WORKER and worker pool association */
6283 mutex_lock(&wq_pool_attach_mutex);
6284
6285 if (task->flags & PF_WQ_WORKER) {
6286 struct worker *worker = kthread_data(task);
6287 struct worker_pool *pool = worker->pool;
6288
6289 if (pool) {
6290 raw_spin_lock_irq(&pool->lock);
6291 /*
6292 * ->desc tracks information (wq name or
6293 * set_worker_desc()) for the latest execution. If
6294 * current, prepend '+', otherwise '-'.
6295 */
6296 if (worker->desc[0] != '\0') {
6297 if (worker->current_work)
6298 scnprintf(buf + off, size - off, "+%s",
6299 worker->desc);
6300 else
6301 scnprintf(buf + off, size - off, "-%s",
6302 worker->desc);
6303 }
6304 raw_spin_unlock_irq(&pool->lock);
6305 }
6306 }
6307
6308 mutex_unlock(&wq_pool_attach_mutex);
6309}
6310
6311#ifdef CONFIG_SMP
6312
6313/*
6314 * CPU hotplug.
6315 *
6316 * There are two challenges in supporting CPU hotplug. Firstly, there
6317 * are a lot of assumptions on strong associations among work, pwq and
6318 * pool which make migrating pending and scheduled works very
6319 * difficult to implement without impacting hot paths. Secondly,
6320 * worker pools serve mix of short, long and very long running works making
6321 * blocked draining impractical.
6322 *
6323 * This is solved by allowing the pools to be disassociated from the CPU
6324 * running as an unbound one and allowing it to be reattached later if the
6325 * cpu comes back online.
6326 */
6327
6328static void unbind_workers(int cpu)
6329{
6330 struct worker_pool *pool;
6331 struct worker *worker;
6332
6333 for_each_cpu_worker_pool(pool, cpu) {
6334 mutex_lock(&wq_pool_attach_mutex);
6335 raw_spin_lock_irq(&pool->lock);
6336
6337 /*
6338 * We've blocked all attach/detach operations. Make all workers
6339 * unbound and set DISASSOCIATED. Before this, all workers
6340 * must be on the cpu. After this, they may become diasporas.
6341 * And the preemption disabled section in their sched callbacks
6342 * are guaranteed to see WORKER_UNBOUND since the code here
6343 * is on the same cpu.
6344 */
6345 for_each_pool_worker(worker, pool)
6346 worker->flags |= WORKER_UNBOUND;
6347
6348 pool->flags |= POOL_DISASSOCIATED;
6349
6350 /*
6351 * The handling of nr_running in sched callbacks are disabled
6352 * now. Zap nr_running. After this, nr_running stays zero and
6353 * need_more_worker() and keep_working() are always true as
6354 * long as the worklist is not empty. This pool now behaves as
6355 * an unbound (in terms of concurrency management) pool which
6356 * are served by workers tied to the pool.
6357 */
6358 pool->nr_running = 0;
6359
6360 /*
6361 * With concurrency management just turned off, a busy
6362 * worker blocking could lead to lengthy stalls. Kick off
6363 * unbound chain execution of currently pending work items.
6364 */
6365 kick_pool(pool);
6366
6367 raw_spin_unlock_irq(&pool->lock);
6368
6369 for_each_pool_worker(worker, pool)
6370 unbind_worker(worker);
6371
6372 mutex_unlock(&wq_pool_attach_mutex);
6373 }
6374}
6375
6376/**
6377 * rebind_workers - rebind all workers of a pool to the associated CPU
6378 * @pool: pool of interest
6379 *
6380 * @pool->cpu is coming online. Rebind all workers to the CPU.
6381 */
6382static void rebind_workers(struct worker_pool *pool)
6383{
6384 struct worker *worker;
6385
6386 lockdep_assert_held(&wq_pool_attach_mutex);
6387
6388 /*
6389 * Restore CPU affinity of all workers. As all idle workers should
6390 * be on the run-queue of the associated CPU before any local
6391 * wake-ups for concurrency management happen, restore CPU affinity
6392 * of all workers first and then clear UNBOUND. As we're called
6393 * from CPU_ONLINE, the following shouldn't fail.
6394 */
6395 for_each_pool_worker(worker, pool) {
6396 kthread_set_per_cpu(worker->task, pool->cpu);
6397 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6398 pool_allowed_cpus(pool)) < 0);
6399 }
6400
6401 raw_spin_lock_irq(&pool->lock);
6402
6403 pool->flags &= ~POOL_DISASSOCIATED;
6404
6405 for_each_pool_worker(worker, pool) {
6406 unsigned int worker_flags = worker->flags;
6407
6408 /*
6409 * We want to clear UNBOUND but can't directly call
6410 * worker_clr_flags() or adjust nr_running. Atomically
6411 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6412 * @worker will clear REBOUND using worker_clr_flags() when
6413 * it initiates the next execution cycle thus restoring
6414 * concurrency management. Note that when or whether
6415 * @worker clears REBOUND doesn't affect correctness.
6416 *
6417 * WRITE_ONCE() is necessary because @worker->flags may be
6418 * tested without holding any lock in
6419 * wq_worker_running(). Without it, NOT_RUNNING test may
6420 * fail incorrectly leading to premature concurrency
6421 * management operations.
6422 */
6423 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6424 worker_flags |= WORKER_REBOUND;
6425 worker_flags &= ~WORKER_UNBOUND;
6426 WRITE_ONCE(worker->flags, worker_flags);
6427 }
6428
6429 raw_spin_unlock_irq(&pool->lock);
6430}
6431
6432/**
6433 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6434 * @pool: unbound pool of interest
6435 * @cpu: the CPU which is coming up
6436 *
6437 * An unbound pool may end up with a cpumask which doesn't have any online
6438 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6439 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6440 * online CPU before, cpus_allowed of all its workers should be restored.
6441 */
6442static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6443{
6444 static cpumask_t cpumask;
6445 struct worker *worker;
6446
6447 lockdep_assert_held(&wq_pool_attach_mutex);
6448
6449 /* is @cpu allowed for @pool? */
6450 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6451 return;
6452
6453 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6454
6455 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6456 for_each_pool_worker(worker, pool)
6457 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6458}
6459
6460int workqueue_prepare_cpu(unsigned int cpu)
6461{
6462 struct worker_pool *pool;
6463
6464 for_each_cpu_worker_pool(pool, cpu) {
6465 if (pool->nr_workers)
6466 continue;
6467 if (!create_worker(pool))
6468 return -ENOMEM;
6469 }
6470 return 0;
6471}
6472
6473int workqueue_online_cpu(unsigned int cpu)
6474{
6475 struct worker_pool *pool;
6476 struct workqueue_struct *wq;
6477 int pi;
6478
6479 mutex_lock(&wq_pool_mutex);
6480
6481 for_each_pool(pool, pi) {
6482 /* BH pools aren't affected by hotplug */
6483 if (pool->flags & POOL_BH)
6484 continue;
6485
6486 mutex_lock(&wq_pool_attach_mutex);
6487 if (pool->cpu == cpu)
6488 rebind_workers(pool);
6489 else if (pool->cpu < 0)
6490 restore_unbound_workers_cpumask(pool, cpu);
6491 mutex_unlock(&wq_pool_attach_mutex);
6492 }
6493
6494 /* update pod affinity of unbound workqueues */
6495 list_for_each_entry(wq, &workqueues, list) {
6496 struct workqueue_attrs *attrs = wq->unbound_attrs;
6497
6498 if (attrs) {
6499 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6500 int tcpu;
6501
6502 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6503 wq_update_pod(wq, tcpu, cpu, true);
6504
6505 mutex_lock(&wq->mutex);
6506 wq_update_node_max_active(wq, -1);
6507 mutex_unlock(&wq->mutex);
6508 }
6509 }
6510
6511 mutex_unlock(&wq_pool_mutex);
6512 return 0;
6513}
6514
6515int workqueue_offline_cpu(unsigned int cpu)
6516{
6517 struct workqueue_struct *wq;
6518
6519 /* unbinding per-cpu workers should happen on the local CPU */
6520 if (WARN_ON(cpu != smp_processor_id()))
6521 return -1;
6522
6523 unbind_workers(cpu);
6524
6525 /* update pod affinity of unbound workqueues */
6526 mutex_lock(&wq_pool_mutex);
6527 list_for_each_entry(wq, &workqueues, list) {
6528 struct workqueue_attrs *attrs = wq->unbound_attrs;
6529
6530 if (attrs) {
6531 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6532 int tcpu;
6533
6534 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6535 wq_update_pod(wq, tcpu, cpu, false);
6536
6537 mutex_lock(&wq->mutex);
6538 wq_update_node_max_active(wq, cpu);
6539 mutex_unlock(&wq->mutex);
6540 }
6541 }
6542 mutex_unlock(&wq_pool_mutex);
6543
6544 return 0;
6545}
6546
6547struct work_for_cpu {
6548 struct work_struct work;
6549 long (*fn)(void *);
6550 void *arg;
6551 long ret;
6552};
6553
6554static void work_for_cpu_fn(struct work_struct *work)
6555{
6556 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6557
6558 wfc->ret = wfc->fn(wfc->arg);
6559}
6560
6561/**
6562 * work_on_cpu_key - run a function in thread context on a particular cpu
6563 * @cpu: the cpu to run on
6564 * @fn: the function to run
6565 * @arg: the function arg
6566 * @key: The lock class key for lock debugging purposes
6567 *
6568 * It is up to the caller to ensure that the cpu doesn't go offline.
6569 * The caller must not hold any locks which would prevent @fn from completing.
6570 *
6571 * Return: The value @fn returns.
6572 */
6573long work_on_cpu_key(int cpu, long (*fn)(void *),
6574 void *arg, struct lock_class_key *key)
6575{
6576 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6577
6578 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6579 schedule_work_on(cpu, &wfc.work);
6580 flush_work(&wfc.work);
6581 destroy_work_on_stack(&wfc.work);
6582 return wfc.ret;
6583}
6584EXPORT_SYMBOL_GPL(work_on_cpu_key);
6585
6586/**
6587 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6588 * @cpu: the cpu to run on
6589 * @fn: the function to run
6590 * @arg: the function argument
6591 * @key: The lock class key for lock debugging purposes
6592 *
6593 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6594 * any locks which would prevent @fn from completing.
6595 *
6596 * Return: The value @fn returns.
6597 */
6598long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6599 void *arg, struct lock_class_key *key)
6600{
6601 long ret = -ENODEV;
6602
6603 cpus_read_lock();
6604 if (cpu_online(cpu))
6605 ret = work_on_cpu_key(cpu, fn, arg, key);
6606 cpus_read_unlock();
6607 return ret;
6608}
6609EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6610#endif /* CONFIG_SMP */
6611
6612#ifdef CONFIG_FREEZER
6613
6614/**
6615 * freeze_workqueues_begin - begin freezing workqueues
6616 *
6617 * Start freezing workqueues. After this function returns, all freezable
6618 * workqueues will queue new works to their inactive_works list instead of
6619 * pool->worklist.
6620 *
6621 * CONTEXT:
6622 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6623 */
6624void freeze_workqueues_begin(void)
6625{
6626 struct workqueue_struct *wq;
6627
6628 mutex_lock(&wq_pool_mutex);
6629
6630 WARN_ON_ONCE(workqueue_freezing);
6631 workqueue_freezing = true;
6632
6633 list_for_each_entry(wq, &workqueues, list) {
6634 mutex_lock(&wq->mutex);
6635 wq_adjust_max_active(wq);
6636 mutex_unlock(&wq->mutex);
6637 }
6638
6639 mutex_unlock(&wq_pool_mutex);
6640}
6641
6642/**
6643 * freeze_workqueues_busy - are freezable workqueues still busy?
6644 *
6645 * Check whether freezing is complete. This function must be called
6646 * between freeze_workqueues_begin() and thaw_workqueues().
6647 *
6648 * CONTEXT:
6649 * Grabs and releases wq_pool_mutex.
6650 *
6651 * Return:
6652 * %true if some freezable workqueues are still busy. %false if freezing
6653 * is complete.
6654 */
6655bool freeze_workqueues_busy(void)
6656{
6657 bool busy = false;
6658 struct workqueue_struct *wq;
6659 struct pool_workqueue *pwq;
6660
6661 mutex_lock(&wq_pool_mutex);
6662
6663 WARN_ON_ONCE(!workqueue_freezing);
6664
6665 list_for_each_entry(wq, &workqueues, list) {
6666 if (!(wq->flags & WQ_FREEZABLE))
6667 continue;
6668 /*
6669 * nr_active is monotonically decreasing. It's safe
6670 * to peek without lock.
6671 */
6672 rcu_read_lock();
6673 for_each_pwq(pwq, wq) {
6674 WARN_ON_ONCE(pwq->nr_active < 0);
6675 if (pwq->nr_active) {
6676 busy = true;
6677 rcu_read_unlock();
6678 goto out_unlock;
6679 }
6680 }
6681 rcu_read_unlock();
6682 }
6683out_unlock:
6684 mutex_unlock(&wq_pool_mutex);
6685 return busy;
6686}
6687
6688/**
6689 * thaw_workqueues - thaw workqueues
6690 *
6691 * Thaw workqueues. Normal queueing is restored and all collected
6692 * frozen works are transferred to their respective pool worklists.
6693 *
6694 * CONTEXT:
6695 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6696 */
6697void thaw_workqueues(void)
6698{
6699 struct workqueue_struct *wq;
6700
6701 mutex_lock(&wq_pool_mutex);
6702
6703 if (!workqueue_freezing)
6704 goto out_unlock;
6705
6706 workqueue_freezing = false;
6707
6708 /* restore max_active and repopulate worklist */
6709 list_for_each_entry(wq, &workqueues, list) {
6710 mutex_lock(&wq->mutex);
6711 wq_adjust_max_active(wq);
6712 mutex_unlock(&wq->mutex);
6713 }
6714
6715out_unlock:
6716 mutex_unlock(&wq_pool_mutex);
6717}
6718#endif /* CONFIG_FREEZER */
6719
6720static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6721{
6722 LIST_HEAD(ctxs);
6723 int ret = 0;
6724 struct workqueue_struct *wq;
6725 struct apply_wqattrs_ctx *ctx, *n;
6726
6727 lockdep_assert_held(&wq_pool_mutex);
6728
6729 list_for_each_entry(wq, &workqueues, list) {
6730 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6731 continue;
6732
6733 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6734 if (IS_ERR(ctx)) {
6735 ret = PTR_ERR(ctx);
6736 break;
6737 }
6738
6739 list_add_tail(&ctx->list, &ctxs);
6740 }
6741
6742 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6743 if (!ret)
6744 apply_wqattrs_commit(ctx);
6745 apply_wqattrs_cleanup(ctx);
6746 }
6747
6748 if (!ret) {
6749 mutex_lock(&wq_pool_attach_mutex);
6750 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6751 mutex_unlock(&wq_pool_attach_mutex);
6752 }
6753 return ret;
6754}
6755
6756/**
6757 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6758 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6759 *
6760 * This function can be called from cpuset code to provide a set of isolated
6761 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
6762 * either cpus_read_lock or cpus_write_lock.
6763 */
6764int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6765{
6766 cpumask_var_t cpumask;
6767 int ret = 0;
6768
6769 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6770 return -ENOMEM;
6771
6772 lockdep_assert_cpus_held();
6773 mutex_lock(&wq_pool_mutex);
6774
6775 /* Save the current isolated cpumask & export it via sysfs */
6776 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6777
6778 /*
6779 * If the operation fails, it will fall back to
6780 * wq_requested_unbound_cpumask which is initially set to
6781 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6782 * by any subsequent write to workqueue/cpumask sysfs file.
6783 */
6784 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6785 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6786 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6787 ret = workqueue_apply_unbound_cpumask(cpumask);
6788
6789 mutex_unlock(&wq_pool_mutex);
6790 free_cpumask_var(cpumask);
6791 return ret;
6792}
6793
6794static int parse_affn_scope(const char *val)
6795{
6796 int i;
6797
6798 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6799 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6800 return i;
6801 }
6802 return -EINVAL;
6803}
6804
6805static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6806{
6807 struct workqueue_struct *wq;
6808 int affn, cpu;
6809
6810 affn = parse_affn_scope(val);
6811 if (affn < 0)
6812 return affn;
6813 if (affn == WQ_AFFN_DFL)
6814 return -EINVAL;
6815
6816 cpus_read_lock();
6817 mutex_lock(&wq_pool_mutex);
6818
6819 wq_affn_dfl = affn;
6820
6821 list_for_each_entry(wq, &workqueues, list) {
6822 for_each_online_cpu(cpu) {
6823 wq_update_pod(wq, cpu, cpu, true);
6824 }
6825 }
6826
6827 mutex_unlock(&wq_pool_mutex);
6828 cpus_read_unlock();
6829
6830 return 0;
6831}
6832
6833static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
6834{
6835 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
6836}
6837
6838static const struct kernel_param_ops wq_affn_dfl_ops = {
6839 .set = wq_affn_dfl_set,
6840 .get = wq_affn_dfl_get,
6841};
6842
6843module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
6844
6845#ifdef CONFIG_SYSFS
6846/*
6847 * Workqueues with WQ_SYSFS flag set is visible to userland via
6848 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
6849 * following attributes.
6850 *
6851 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
6852 * max_active RW int : maximum number of in-flight work items
6853 *
6854 * Unbound workqueues have the following extra attributes.
6855 *
6856 * nice RW int : nice value of the workers
6857 * cpumask RW mask : bitmask of allowed CPUs for the workers
6858 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
6859 * affinity_strict RW bool : worker CPU affinity is strict
6860 */
6861struct wq_device {
6862 struct workqueue_struct *wq;
6863 struct device dev;
6864};
6865
6866static struct workqueue_struct *dev_to_wq(struct device *dev)
6867{
6868 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6869
6870 return wq_dev->wq;
6871}
6872
6873static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
6874 char *buf)
6875{
6876 struct workqueue_struct *wq = dev_to_wq(dev);
6877
6878 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
6879}
6880static DEVICE_ATTR_RO(per_cpu);
6881
6882static ssize_t max_active_show(struct device *dev,
6883 struct device_attribute *attr, char *buf)
6884{
6885 struct workqueue_struct *wq = dev_to_wq(dev);
6886
6887 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
6888}
6889
6890static ssize_t max_active_store(struct device *dev,
6891 struct device_attribute *attr, const char *buf,
6892 size_t count)
6893{
6894 struct workqueue_struct *wq = dev_to_wq(dev);
6895 int val;
6896
6897 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
6898 return -EINVAL;
6899
6900 workqueue_set_max_active(wq, val);
6901 return count;
6902}
6903static DEVICE_ATTR_RW(max_active);
6904
6905static struct attribute *wq_sysfs_attrs[] = {
6906 &dev_attr_per_cpu.attr,
6907 &dev_attr_max_active.attr,
6908 NULL,
6909};
6910ATTRIBUTE_GROUPS(wq_sysfs);
6911
6912static void apply_wqattrs_lock(void)
6913{
6914 /* CPUs should stay stable across pwq creations and installations */
6915 cpus_read_lock();
6916 mutex_lock(&wq_pool_mutex);
6917}
6918
6919static void apply_wqattrs_unlock(void)
6920{
6921 mutex_unlock(&wq_pool_mutex);
6922 cpus_read_unlock();
6923}
6924
6925static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
6926 char *buf)
6927{
6928 struct workqueue_struct *wq = dev_to_wq(dev);
6929 int written;
6930
6931 mutex_lock(&wq->mutex);
6932 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
6933 mutex_unlock(&wq->mutex);
6934
6935 return written;
6936}
6937
6938/* prepare workqueue_attrs for sysfs store operations */
6939static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6940{
6941 struct workqueue_attrs *attrs;
6942
6943 lockdep_assert_held(&wq_pool_mutex);
6944
6945 attrs = alloc_workqueue_attrs();
6946 if (!attrs)
6947 return NULL;
6948
6949 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6950 return attrs;
6951}
6952
6953static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6954 const char *buf, size_t count)
6955{
6956 struct workqueue_struct *wq = dev_to_wq(dev);
6957 struct workqueue_attrs *attrs;
6958 int ret = -ENOMEM;
6959
6960 apply_wqattrs_lock();
6961
6962 attrs = wq_sysfs_prep_attrs(wq);
6963 if (!attrs)
6964 goto out_unlock;
6965
6966 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6967 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6968 ret = apply_workqueue_attrs_locked(wq, attrs);
6969 else
6970 ret = -EINVAL;
6971
6972out_unlock:
6973 apply_wqattrs_unlock();
6974 free_workqueue_attrs(attrs);
6975 return ret ?: count;
6976}
6977
6978static ssize_t wq_cpumask_show(struct device *dev,
6979 struct device_attribute *attr, char *buf)
6980{
6981 struct workqueue_struct *wq = dev_to_wq(dev);
6982 int written;
6983
6984 mutex_lock(&wq->mutex);
6985 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6986 cpumask_pr_args(wq->unbound_attrs->cpumask));
6987 mutex_unlock(&wq->mutex);
6988 return written;
6989}
6990
6991static ssize_t wq_cpumask_store(struct device *dev,
6992 struct device_attribute *attr,
6993 const char *buf, size_t count)
6994{
6995 struct workqueue_struct *wq = dev_to_wq(dev);
6996 struct workqueue_attrs *attrs;
6997 int ret = -ENOMEM;
6998
6999 apply_wqattrs_lock();
7000
7001 attrs = wq_sysfs_prep_attrs(wq);
7002 if (!attrs)
7003 goto out_unlock;
7004
7005 ret = cpumask_parse(buf, attrs->cpumask);
7006 if (!ret)
7007 ret = apply_workqueue_attrs_locked(wq, attrs);
7008
7009out_unlock:
7010 apply_wqattrs_unlock();
7011 free_workqueue_attrs(attrs);
7012 return ret ?: count;
7013}
7014
7015static ssize_t wq_affn_scope_show(struct device *dev,
7016 struct device_attribute *attr, char *buf)
7017{
7018 struct workqueue_struct *wq = dev_to_wq(dev);
7019 int written;
7020
7021 mutex_lock(&wq->mutex);
7022 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7023 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7024 wq_affn_names[WQ_AFFN_DFL],
7025 wq_affn_names[wq_affn_dfl]);
7026 else
7027 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7028 wq_affn_names[wq->unbound_attrs->affn_scope]);
7029 mutex_unlock(&wq->mutex);
7030
7031 return written;
7032}
7033
7034static ssize_t wq_affn_scope_store(struct device *dev,
7035 struct device_attribute *attr,
7036 const char *buf, size_t count)
7037{
7038 struct workqueue_struct *wq = dev_to_wq(dev);
7039 struct workqueue_attrs *attrs;
7040 int affn, ret = -ENOMEM;
7041
7042 affn = parse_affn_scope(buf);
7043 if (affn < 0)
7044 return affn;
7045
7046 apply_wqattrs_lock();
7047 attrs = wq_sysfs_prep_attrs(wq);
7048 if (attrs) {
7049 attrs->affn_scope = affn;
7050 ret = apply_workqueue_attrs_locked(wq, attrs);
7051 }
7052 apply_wqattrs_unlock();
7053 free_workqueue_attrs(attrs);
7054 return ret ?: count;
7055}
7056
7057static ssize_t wq_affinity_strict_show(struct device *dev,
7058 struct device_attribute *attr, char *buf)
7059{
7060 struct workqueue_struct *wq = dev_to_wq(dev);
7061
7062 return scnprintf(buf, PAGE_SIZE, "%d\n",
7063 wq->unbound_attrs->affn_strict);
7064}
7065
7066static ssize_t wq_affinity_strict_store(struct device *dev,
7067 struct device_attribute *attr,
7068 const char *buf, size_t count)
7069{
7070 struct workqueue_struct *wq = dev_to_wq(dev);
7071 struct workqueue_attrs *attrs;
7072 int v, ret = -ENOMEM;
7073
7074 if (sscanf(buf, "%d", &v) != 1)
7075 return -EINVAL;
7076
7077 apply_wqattrs_lock();
7078 attrs = wq_sysfs_prep_attrs(wq);
7079 if (attrs) {
7080 attrs->affn_strict = (bool)v;
7081 ret = apply_workqueue_attrs_locked(wq, attrs);
7082 }
7083 apply_wqattrs_unlock();
7084 free_workqueue_attrs(attrs);
7085 return ret ?: count;
7086}
7087
7088static struct device_attribute wq_sysfs_unbound_attrs[] = {
7089 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7090 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7091 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7092 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7093 __ATTR_NULL,
7094};
7095
7096static const struct bus_type wq_subsys = {
7097 .name = "workqueue",
7098 .dev_groups = wq_sysfs_groups,
7099};
7100
7101/**
7102 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7103 * @cpumask: the cpumask to set
7104 *
7105 * The low-level workqueues cpumask is a global cpumask that limits
7106 * the affinity of all unbound workqueues. This function check the @cpumask
7107 * and apply it to all unbound workqueues and updates all pwqs of them.
7108 *
7109 * Return: 0 - Success
7110 * -EINVAL - Invalid @cpumask
7111 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7112 */
7113static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7114{
7115 int ret = -EINVAL;
7116
7117 /*
7118 * Not excluding isolated cpus on purpose.
7119 * If the user wishes to include them, we allow that.
7120 */
7121 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7122 if (!cpumask_empty(cpumask)) {
7123 apply_wqattrs_lock();
7124 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7125 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
7126 ret = 0;
7127 goto out_unlock;
7128 }
7129
7130 ret = workqueue_apply_unbound_cpumask(cpumask);
7131
7132out_unlock:
7133 apply_wqattrs_unlock();
7134 }
7135
7136 return ret;
7137}
7138
7139static ssize_t __wq_cpumask_show(struct device *dev,
7140 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7141{
7142 int written;
7143
7144 mutex_lock(&wq_pool_mutex);
7145 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7146 mutex_unlock(&wq_pool_mutex);
7147
7148 return written;
7149}
7150
7151static ssize_t wq_unbound_cpumask_show(struct device *dev,
7152 struct device_attribute *attr, char *buf)
7153{
7154 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7155}
7156
7157static ssize_t wq_requested_cpumask_show(struct device *dev,
7158 struct device_attribute *attr, char *buf)
7159{
7160 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7161}
7162
7163static ssize_t wq_isolated_cpumask_show(struct device *dev,
7164 struct device_attribute *attr, char *buf)
7165{
7166 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7167}
7168
7169static ssize_t wq_unbound_cpumask_store(struct device *dev,
7170 struct device_attribute *attr, const char *buf, size_t count)
7171{
7172 cpumask_var_t cpumask;
7173 int ret;
7174
7175 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7176 return -ENOMEM;
7177
7178 ret = cpumask_parse(buf, cpumask);
7179 if (!ret)
7180 ret = workqueue_set_unbound_cpumask(cpumask);
7181
7182 free_cpumask_var(cpumask);
7183 return ret ? ret : count;
7184}
7185
7186static struct device_attribute wq_sysfs_cpumask_attrs[] = {
7187 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
7188 wq_unbound_cpumask_store),
7189 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
7190 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
7191 __ATTR_NULL,
7192};
7193
7194static int __init wq_sysfs_init(void)
7195{
7196 struct device *dev_root;
7197 int err;
7198
7199 err = subsys_virtual_register(&wq_subsys, NULL);
7200 if (err)
7201 return err;
7202
7203 dev_root = bus_get_dev_root(&wq_subsys);
7204 if (dev_root) {
7205 struct device_attribute *attr;
7206
7207 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
7208 err = device_create_file(dev_root, attr);
7209 if (err)
7210 break;
7211 }
7212 put_device(dev_root);
7213 }
7214 return err;
7215}
7216core_initcall(wq_sysfs_init);
7217
7218static void wq_device_release(struct device *dev)
7219{
7220 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7221
7222 kfree(wq_dev);
7223}
7224
7225/**
7226 * workqueue_sysfs_register - make a workqueue visible in sysfs
7227 * @wq: the workqueue to register
7228 *
7229 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7230 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7231 * which is the preferred method.
7232 *
7233 * Workqueue user should use this function directly iff it wants to apply
7234 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7235 * apply_workqueue_attrs() may race against userland updating the
7236 * attributes.
7237 *
7238 * Return: 0 on success, -errno on failure.
7239 */
7240int workqueue_sysfs_register(struct workqueue_struct *wq)
7241{
7242 struct wq_device *wq_dev;
7243 int ret;
7244
7245 /*
7246 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7247 * ordered workqueues.
7248 */
7249 if (WARN_ON(wq->flags & __WQ_ORDERED))
7250 return -EINVAL;
7251
7252 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7253 if (!wq_dev)
7254 return -ENOMEM;
7255
7256 wq_dev->wq = wq;
7257 wq_dev->dev.bus = &wq_subsys;
7258 wq_dev->dev.release = wq_device_release;
7259 dev_set_name(&wq_dev->dev, "%s", wq->name);
7260
7261 /*
7262 * unbound_attrs are created separately. Suppress uevent until
7263 * everything is ready.
7264 */
7265 dev_set_uevent_suppress(&wq_dev->dev, true);
7266
7267 ret = device_register(&wq_dev->dev);
7268 if (ret) {
7269 put_device(&wq_dev->dev);
7270 wq->wq_dev = NULL;
7271 return ret;
7272 }
7273
7274 if (wq->flags & WQ_UNBOUND) {
7275 struct device_attribute *attr;
7276
7277 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7278 ret = device_create_file(&wq_dev->dev, attr);
7279 if (ret) {
7280 device_unregister(&wq_dev->dev);
7281 wq->wq_dev = NULL;
7282 return ret;
7283 }
7284 }
7285 }
7286
7287 dev_set_uevent_suppress(&wq_dev->dev, false);
7288 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7289 return 0;
7290}
7291
7292/**
7293 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7294 * @wq: the workqueue to unregister
7295 *
7296 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7297 */
7298static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7299{
7300 struct wq_device *wq_dev = wq->wq_dev;
7301
7302 if (!wq->wq_dev)
7303 return;
7304
7305 wq->wq_dev = NULL;
7306 device_unregister(&wq_dev->dev);
7307}
7308#else /* CONFIG_SYSFS */
7309static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7310#endif /* CONFIG_SYSFS */
7311
7312/*
7313 * Workqueue watchdog.
7314 *
7315 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7316 * flush dependency, a concurrency managed work item which stays RUNNING
7317 * indefinitely. Workqueue stalls can be very difficult to debug as the
7318 * usual warning mechanisms don't trigger and internal workqueue state is
7319 * largely opaque.
7320 *
7321 * Workqueue watchdog monitors all worker pools periodically and dumps
7322 * state if some pools failed to make forward progress for a while where
7323 * forward progress is defined as the first item on ->worklist changing.
7324 *
7325 * This mechanism is controlled through the kernel parameter
7326 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7327 * corresponding sysfs parameter file.
7328 */
7329#ifdef CONFIG_WQ_WATCHDOG
7330
7331static unsigned long wq_watchdog_thresh = 30;
7332static struct timer_list wq_watchdog_timer;
7333
7334static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7335static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7336
7337/*
7338 * Show workers that might prevent the processing of pending work items.
7339 * The only candidates are CPU-bound workers in the running state.
7340 * Pending work items should be handled by another idle worker
7341 * in all other situations.
7342 */
7343static void show_cpu_pool_hog(struct worker_pool *pool)
7344{
7345 struct worker *worker;
7346 unsigned long irq_flags;
7347 int bkt;
7348
7349 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7350
7351 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7352 if (task_is_running(worker->task)) {
7353 /*
7354 * Defer printing to avoid deadlocks in console
7355 * drivers that queue work while holding locks
7356 * also taken in their write paths.
7357 */
7358 printk_deferred_enter();
7359
7360 pr_info("pool %d:\n", pool->id);
7361 sched_show_task(worker->task);
7362
7363 printk_deferred_exit();
7364 }
7365 }
7366
7367 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7368}
7369
7370static void show_cpu_pools_hogs(void)
7371{
7372 struct worker_pool *pool;
7373 int pi;
7374
7375 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7376
7377 rcu_read_lock();
7378
7379 for_each_pool(pool, pi) {
7380 if (pool->cpu_stall)
7381 show_cpu_pool_hog(pool);
7382
7383 }
7384
7385 rcu_read_unlock();
7386}
7387
7388static void wq_watchdog_reset_touched(void)
7389{
7390 int cpu;
7391
7392 wq_watchdog_touched = jiffies;
7393 for_each_possible_cpu(cpu)
7394 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7395}
7396
7397static void wq_watchdog_timer_fn(struct timer_list *unused)
7398{
7399 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7400 bool lockup_detected = false;
7401 bool cpu_pool_stall = false;
7402 unsigned long now = jiffies;
7403 struct worker_pool *pool;
7404 int pi;
7405
7406 if (!thresh)
7407 return;
7408
7409 rcu_read_lock();
7410
7411 for_each_pool(pool, pi) {
7412 unsigned long pool_ts, touched, ts;
7413
7414 pool->cpu_stall = false;
7415 if (list_empty(&pool->worklist))
7416 continue;
7417
7418 /*
7419 * If a virtual machine is stopped by the host it can look to
7420 * the watchdog like a stall.
7421 */
7422 kvm_check_and_clear_guest_paused();
7423
7424 /* get the latest of pool and touched timestamps */
7425 if (pool->cpu >= 0)
7426 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7427 else
7428 touched = READ_ONCE(wq_watchdog_touched);
7429 pool_ts = READ_ONCE(pool->watchdog_ts);
7430
7431 if (time_after(pool_ts, touched))
7432 ts = pool_ts;
7433 else
7434 ts = touched;
7435
7436 /* did we stall? */
7437 if (time_after(now, ts + thresh)) {
7438 lockup_detected = true;
7439 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7440 pool->cpu_stall = true;
7441 cpu_pool_stall = true;
7442 }
7443 pr_emerg("BUG: workqueue lockup - pool");
7444 pr_cont_pool_info(pool);
7445 pr_cont(" stuck for %us!\n",
7446 jiffies_to_msecs(now - pool_ts) / 1000);
7447 }
7448
7449
7450 }
7451
7452 rcu_read_unlock();
7453
7454 if (lockup_detected)
7455 show_all_workqueues();
7456
7457 if (cpu_pool_stall)
7458 show_cpu_pools_hogs();
7459
7460 wq_watchdog_reset_touched();
7461 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7462}
7463
7464notrace void wq_watchdog_touch(int cpu)
7465{
7466 if (cpu >= 0)
7467 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7468
7469 wq_watchdog_touched = jiffies;
7470}
7471
7472static void wq_watchdog_set_thresh(unsigned long thresh)
7473{
7474 wq_watchdog_thresh = 0;
7475 del_timer_sync(&wq_watchdog_timer);
7476
7477 if (thresh) {
7478 wq_watchdog_thresh = thresh;
7479 wq_watchdog_reset_touched();
7480 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7481 }
7482}
7483
7484static int wq_watchdog_param_set_thresh(const char *val,
7485 const struct kernel_param *kp)
7486{
7487 unsigned long thresh;
7488 int ret;
7489
7490 ret = kstrtoul(val, 0, &thresh);
7491 if (ret)
7492 return ret;
7493
7494 if (system_wq)
7495 wq_watchdog_set_thresh(thresh);
7496 else
7497 wq_watchdog_thresh = thresh;
7498
7499 return 0;
7500}
7501
7502static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7503 .set = wq_watchdog_param_set_thresh,
7504 .get = param_get_ulong,
7505};
7506
7507module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7508 0644);
7509
7510static void wq_watchdog_init(void)
7511{
7512 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7513 wq_watchdog_set_thresh(wq_watchdog_thresh);
7514}
7515
7516#else /* CONFIG_WQ_WATCHDOG */
7517
7518static inline void wq_watchdog_init(void) { }
7519
7520#endif /* CONFIG_WQ_WATCHDOG */
7521
7522static void bh_pool_kick_normal(struct irq_work *irq_work)
7523{
7524 raise_softirq_irqoff(TASKLET_SOFTIRQ);
7525}
7526
7527static void bh_pool_kick_highpri(struct irq_work *irq_work)
7528{
7529 raise_softirq_irqoff(HI_SOFTIRQ);
7530}
7531
7532static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7533{
7534 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7535 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7536 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7537 return;
7538 }
7539
7540 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7541}
7542
7543static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7544{
7545 BUG_ON(init_worker_pool(pool));
7546 pool->cpu = cpu;
7547 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7548 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7549 pool->attrs->nice = nice;
7550 pool->attrs->affn_strict = true;
7551 pool->node = cpu_to_node(cpu);
7552
7553 /* alloc pool ID */
7554 mutex_lock(&wq_pool_mutex);
7555 BUG_ON(worker_pool_assign_id(pool));
7556 mutex_unlock(&wq_pool_mutex);
7557}
7558
7559/**
7560 * workqueue_init_early - early init for workqueue subsystem
7561 *
7562 * This is the first step of three-staged workqueue subsystem initialization and
7563 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7564 * up. It sets up all the data structures and system workqueues and allows early
7565 * boot code to create workqueues and queue/cancel work items. Actual work item
7566 * execution starts only after kthreads can be created and scheduled right
7567 * before early initcalls.
7568 */
7569void __init workqueue_init_early(void)
7570{
7571 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7572 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7573 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7574 bh_pool_kick_highpri };
7575 int i, cpu;
7576
7577 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7578
7579 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7580 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7581 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7582
7583 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7584 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7585 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7586 if (!cpumask_empty(&wq_cmdline_cpumask))
7587 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7588
7589 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7590
7591 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7592
7593 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
7594 BUG_ON(!wq_update_pod_attrs_buf);
7595
7596 /*
7597 * If nohz_full is enabled, set power efficient workqueue as unbound.
7598 * This allows workqueue items to be moved to HK CPUs.
7599 */
7600 if (housekeeping_enabled(HK_TYPE_TICK))
7601 wq_power_efficient = true;
7602
7603 /* initialize WQ_AFFN_SYSTEM pods */
7604 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7605 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7606 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7607 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7608
7609 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7610
7611 pt->nr_pods = 1;
7612 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7613 pt->pod_node[0] = NUMA_NO_NODE;
7614 pt->cpu_pod[0] = 0;
7615
7616 /* initialize BH and CPU pools */
7617 for_each_possible_cpu(cpu) {
7618 struct worker_pool *pool;
7619
7620 i = 0;
7621 for_each_bh_worker_pool(pool, cpu) {
7622 init_cpu_worker_pool(pool, cpu, std_nice[i]);
7623 pool->flags |= POOL_BH;
7624 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7625 i++;
7626 }
7627
7628 i = 0;
7629 for_each_cpu_worker_pool(pool, cpu)
7630 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7631 }
7632
7633 /* create default unbound and ordered wq attrs */
7634 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7635 struct workqueue_attrs *attrs;
7636
7637 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7638 attrs->nice = std_nice[i];
7639 unbound_std_wq_attrs[i] = attrs;
7640
7641 /*
7642 * An ordered wq should have only one pwq as ordering is
7643 * guaranteed by max_active which is enforced by pwqs.
7644 */
7645 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7646 attrs->nice = std_nice[i];
7647 attrs->ordered = true;
7648 ordered_wq_attrs[i] = attrs;
7649 }
7650
7651 system_wq = alloc_workqueue("events", 0, 0);
7652 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7653 system_long_wq = alloc_workqueue("events_long", 0, 0);
7654 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7655 WQ_MAX_ACTIVE);
7656 system_freezable_wq = alloc_workqueue("events_freezable",
7657 WQ_FREEZABLE, 0);
7658 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7659 WQ_POWER_EFFICIENT, 0);
7660 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7661 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7662 0);
7663 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7664 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7665 WQ_BH | WQ_HIGHPRI, 0);
7666 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7667 !system_unbound_wq || !system_freezable_wq ||
7668 !system_power_efficient_wq ||
7669 !system_freezable_power_efficient_wq ||
7670 !system_bh_wq || !system_bh_highpri_wq);
7671}
7672
7673static void __init wq_cpu_intensive_thresh_init(void)
7674{
7675 unsigned long thresh;
7676 unsigned long bogo;
7677
7678 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7679 BUG_ON(IS_ERR(pwq_release_worker));
7680
7681 /* if the user set it to a specific value, keep it */
7682 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7683 return;
7684
7685 /*
7686 * The default of 10ms is derived from the fact that most modern (as of
7687 * 2023) processors can do a lot in 10ms and that it's just below what
7688 * most consider human-perceivable. However, the kernel also runs on a
7689 * lot slower CPUs including microcontrollers where the threshold is way
7690 * too low.
7691 *
7692 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7693 * This is by no means accurate but it doesn't have to be. The mechanism
7694 * is still useful even when the threshold is fully scaled up. Also, as
7695 * the reports would usually be applicable to everyone, some machines
7696 * operating on longer thresholds won't significantly diminish their
7697 * usefulness.
7698 */
7699 thresh = 10 * USEC_PER_MSEC;
7700
7701 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7702 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7703 if (bogo < 4000)
7704 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7705
7706 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7707 loops_per_jiffy, bogo, thresh);
7708
7709 wq_cpu_intensive_thresh_us = thresh;
7710}
7711
7712/**
7713 * workqueue_init - bring workqueue subsystem fully online
7714 *
7715 * This is the second step of three-staged workqueue subsystem initialization
7716 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7717 * been created and work items queued on them, but there are no kworkers
7718 * executing the work items yet. Populate the worker pools with the initial
7719 * workers and enable future kworker creations.
7720 */
7721void __init workqueue_init(void)
7722{
7723 struct workqueue_struct *wq;
7724 struct worker_pool *pool;
7725 int cpu, bkt;
7726
7727 wq_cpu_intensive_thresh_init();
7728
7729 mutex_lock(&wq_pool_mutex);
7730
7731 /*
7732 * Per-cpu pools created earlier could be missing node hint. Fix them
7733 * up. Also, create a rescuer for workqueues that requested it.
7734 */
7735 for_each_possible_cpu(cpu) {
7736 for_each_bh_worker_pool(pool, cpu)
7737 pool->node = cpu_to_node(cpu);
7738 for_each_cpu_worker_pool(pool, cpu)
7739 pool->node = cpu_to_node(cpu);
7740 }
7741
7742 list_for_each_entry(wq, &workqueues, list) {
7743 WARN(init_rescuer(wq),
7744 "workqueue: failed to create early rescuer for %s",
7745 wq->name);
7746 }
7747
7748 mutex_unlock(&wq_pool_mutex);
7749
7750 /*
7751 * Create the initial workers. A BH pool has one pseudo worker that
7752 * represents the shared BH execution context and thus doesn't get
7753 * affected by hotplug events. Create the BH pseudo workers for all
7754 * possible CPUs here.
7755 */
7756 for_each_possible_cpu(cpu)
7757 for_each_bh_worker_pool(pool, cpu)
7758 BUG_ON(!create_worker(pool));
7759
7760 for_each_online_cpu(cpu) {
7761 for_each_cpu_worker_pool(pool, cpu) {
7762 pool->flags &= ~POOL_DISASSOCIATED;
7763 BUG_ON(!create_worker(pool));
7764 }
7765 }
7766
7767 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7768 BUG_ON(!create_worker(pool));
7769
7770 wq_online = true;
7771 wq_watchdog_init();
7772}
7773
7774/*
7775 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7776 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7777 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7778 */
7779static void __init init_pod_type(struct wq_pod_type *pt,
7780 bool (*cpus_share_pod)(int, int))
7781{
7782 int cur, pre, cpu, pod;
7783
7784 pt->nr_pods = 0;
7785
7786 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7787 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7788 BUG_ON(!pt->cpu_pod);
7789
7790 for_each_possible_cpu(cur) {
7791 for_each_possible_cpu(pre) {
7792 if (pre >= cur) {
7793 pt->cpu_pod[cur] = pt->nr_pods++;
7794 break;
7795 }
7796 if (cpus_share_pod(cur, pre)) {
7797 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7798 break;
7799 }
7800 }
7801 }
7802
7803 /* init the rest to match @pt->cpu_pod[] */
7804 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7805 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7806 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7807
7808 for (pod = 0; pod < pt->nr_pods; pod++)
7809 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7810
7811 for_each_possible_cpu(cpu) {
7812 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7813 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7814 }
7815}
7816
7817static bool __init cpus_dont_share(int cpu0, int cpu1)
7818{
7819 return false;
7820}
7821
7822static bool __init cpus_share_smt(int cpu0, int cpu1)
7823{
7824#ifdef CONFIG_SCHED_SMT
7825 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7826#else
7827 return false;
7828#endif
7829}
7830
7831static bool __init cpus_share_numa(int cpu0, int cpu1)
7832{
7833 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7834}
7835
7836/**
7837 * workqueue_init_topology - initialize CPU pods for unbound workqueues
7838 *
7839 * This is the third step of three-staged workqueue subsystem initialization and
7840 * invoked after SMP and topology information are fully initialized. It
7841 * initializes the unbound CPU pods accordingly.
7842 */
7843void __init workqueue_init_topology(void)
7844{
7845 struct workqueue_struct *wq;
7846 int cpu;
7847
7848 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
7849 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
7850 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
7851 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
7852
7853 wq_topo_initialized = true;
7854
7855 mutex_lock(&wq_pool_mutex);
7856
7857 /*
7858 * Workqueues allocated earlier would have all CPUs sharing the default
7859 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
7860 * combinations to apply per-pod sharing.
7861 */
7862 list_for_each_entry(wq, &workqueues, list) {
7863 for_each_online_cpu(cpu)
7864 wq_update_pod(wq, cpu, cpu, true);
7865 if (wq->flags & WQ_UNBOUND) {
7866 mutex_lock(&wq->mutex);
7867 wq_update_node_max_active(wq, -1);
7868 mutex_unlock(&wq->mutex);
7869 }
7870 }
7871
7872 mutex_unlock(&wq_pool_mutex);
7873}
7874
7875void __warn_flushing_systemwide_wq(void)
7876{
7877 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
7878 dump_stack();
7879}
7880EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
7881
7882static int __init workqueue_unbound_cpus_setup(char *str)
7883{
7884 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
7885 cpumask_clear(&wq_cmdline_cpumask);
7886 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
7887 }
7888
7889 return 1;
7890}
7891__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);