Loading...
1/*
2 * kernel/workqueue.c - generic async execution with shared worker pool
3 *
4 * Copyright (C) 2002 Ingo Molnar
5 *
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
11 *
12 * Made to use alloc_percpu by Christoph Lameter.
13 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
23 *
24 * Please read Documentation/workqueue.txt for details.
25 */
26
27#include <linux/export.h>
28#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
38#include <linux/hardirq.h>
39#include <linux/mempolicy.h>
40#include <linux/freezer.h>
41#include <linux/kallsyms.h>
42#include <linux/debug_locks.h>
43#include <linux/lockdep.h>
44#include <linux/idr.h>
45#include <linux/jhash.h>
46#include <linux/hashtable.h>
47#include <linux/rculist.h>
48#include <linux/nodemask.h>
49#include <linux/moduleparam.h>
50#include <linux/uaccess.h>
51
52#include "workqueue_internal.h"
53
54enum {
55 /*
56 * worker_pool flags
57 *
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * is in effect.
62 *
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
66 *
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
70 */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72
73 /* worker flags */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
80
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
83
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
94 (min two ticks) */
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97
98 /*
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
101 */
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
104
105 WQ_NAME_LEN = 24,
106};
107
108/*
109 * Structure fields follow one of the following exclusion rules.
110 *
111 * I: Modifiable by initialization/destruction paths and read-only for
112 * everyone else.
113 *
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
116 *
117 * L: pool->lock protected. Access with pool->lock held.
118 *
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
123 *
124 * A: pool->attach_mutex protected.
125 *
126 * PL: wq_pool_mutex protected.
127 *
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129 *
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131 *
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
134 *
135 * WQ: wq->mutex protected.
136 *
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 *
139 * MD: wq_mayday_lock protected.
140 */
141
142/* struct worker is defined in workqueue_internal.h */
143
144struct worker_pool {
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
150
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
152
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
155
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
158
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
162
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
166
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
173
174 struct ida worker_ida; /* worker IDs for task name */
175
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
179
180 /*
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
183 * cacheline.
184 */
185 atomic_t nr_running ____cacheline_aligned_in_smp;
186
187 /*
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
190 */
191 struct rcu_head rcu;
192} ____cacheline_aligned_in_smp;
193
194/*
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
199 */
200struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
213
214 /*
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
219 */
220 struct work_struct unbound_release_work;
221 struct rcu_head rcu;
222} __aligned(1 << WORK_STRUCT_FLAG_BITS);
223
224/*
225 * Structure used to wait for workqueue flush.
226 */
227struct wq_flusher {
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
231};
232
233struct wq_device;
234
235/*
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
238 */
239struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
242
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
250
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
253
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
256
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259
260#ifdef CONFIG_SYSFS
261 struct wq_device *wq_dev; /* I: for sysfs interface */
262#endif
263#ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
265#endif
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268 /*
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
272 */
273 struct rcu_head rcu;
274
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279};
280
281static struct kmem_cache *pwq_cache;
282
283static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
285
286static bool wq_disable_numa;
287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289/* see the comment above the definition of WQ_POWER_EFFICIENT */
290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
294
295/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
297
298static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
300
301static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302static bool workqueue_freezing; /* PL: have wqs started freezing? */
303
304/* PL: allowable cpus for unbound wqs and work items */
305static cpumask_var_t wq_unbound_cpumask;
306
307/* CPU where unbound work was last round robin scheduled from this CPU */
308static DEFINE_PER_CPU(int, wq_rr_cpu_last);
309
310/*
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
314 */
315#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316static bool wq_debug_force_rr_cpu = true;
317#else
318static bool wq_debug_force_rr_cpu = false;
319#endif
320module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
321
322/* the per-cpu worker pools */
323static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
324
325static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
326
327/* PL: hash of all unbound pools keyed by pool->attrs */
328static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
329
330/* I: attributes used when instantiating standard unbound pools on demand */
331static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
332
333/* I: attributes used when instantiating ordered pools on demand */
334static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
335
336struct workqueue_struct *system_wq __read_mostly;
337EXPORT_SYMBOL(system_wq);
338struct workqueue_struct *system_highpri_wq __read_mostly;
339EXPORT_SYMBOL_GPL(system_highpri_wq);
340struct workqueue_struct *system_long_wq __read_mostly;
341EXPORT_SYMBOL_GPL(system_long_wq);
342struct workqueue_struct *system_unbound_wq __read_mostly;
343EXPORT_SYMBOL_GPL(system_unbound_wq);
344struct workqueue_struct *system_freezable_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_freezable_wq);
346struct workqueue_struct *system_power_efficient_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_power_efficient_wq);
348struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
350
351static int worker_thread(void *__worker);
352static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
353
354#define CREATE_TRACE_POINTS
355#include <trace/events/workqueue.h>
356
357#define assert_rcu_or_pool_mutex() \
358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU or wq_pool_mutex should be held")
361
362#define assert_rcu_or_wq_mutex(wq) \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq->mutex), \
365 "sched RCU or wq->mutex should be held")
366
367#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex) && \
370 !lockdep_is_held(&wq_pool_mutex), \
371 "sched RCU, wq->mutex or wq_pool_mutex should be held")
372
373#define for_each_cpu_worker_pool(pool, cpu) \
374 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
376 (pool)++)
377
378/**
379 * for_each_pool - iterate through all worker_pools in the system
380 * @pool: iteration cursor
381 * @pi: integer used for iteration
382 *
383 * This must be called either with wq_pool_mutex held or sched RCU read
384 * locked. If the pool needs to be used beyond the locking in effect, the
385 * caller is responsible for guaranteeing that the pool stays online.
386 *
387 * The if/else clause exists only for the lockdep assertion and can be
388 * ignored.
389 */
390#define for_each_pool(pool, pi) \
391 idr_for_each_entry(&worker_pool_idr, pool, pi) \
392 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
393 else
394
395/**
396 * for_each_pool_worker - iterate through all workers of a worker_pool
397 * @worker: iteration cursor
398 * @pool: worker_pool to iterate workers of
399 *
400 * This must be called with @pool->attach_mutex.
401 *
402 * The if/else clause exists only for the lockdep assertion and can be
403 * ignored.
404 */
405#define for_each_pool_worker(worker, pool) \
406 list_for_each_entry((worker), &(pool)->workers, node) \
407 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
408 else
409
410/**
411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412 * @pwq: iteration cursor
413 * @wq: the target workqueue
414 *
415 * This must be called either with wq->mutex held or sched RCU read locked.
416 * If the pwq needs to be used beyond the locking in effect, the caller is
417 * responsible for guaranteeing that the pwq stays online.
418 *
419 * The if/else clause exists only for the lockdep assertion and can be
420 * ignored.
421 */
422#define for_each_pwq(pwq, wq) \
423 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
425 else
426
427#ifdef CONFIG_DEBUG_OBJECTS_WORK
428
429static struct debug_obj_descr work_debug_descr;
430
431static void *work_debug_hint(void *addr)
432{
433 return ((struct work_struct *) addr)->func;
434}
435
436/*
437 * fixup_init is called when:
438 * - an active object is initialized
439 */
440static int work_fixup_init(void *addr, enum debug_obj_state state)
441{
442 struct work_struct *work = addr;
443
444 switch (state) {
445 case ODEBUG_STATE_ACTIVE:
446 cancel_work_sync(work);
447 debug_object_init(work, &work_debug_descr);
448 return 1;
449 default:
450 return 0;
451 }
452}
453
454/*
455 * fixup_activate is called when:
456 * - an active object is activated
457 * - an unknown object is activated (might be a statically initialized object)
458 */
459static int work_fixup_activate(void *addr, enum debug_obj_state state)
460{
461 struct work_struct *work = addr;
462
463 switch (state) {
464
465 case ODEBUG_STATE_NOTAVAILABLE:
466 /*
467 * This is not really a fixup. The work struct was
468 * statically initialized. We just make sure that it
469 * is tracked in the object tracker.
470 */
471 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
472 debug_object_init(work, &work_debug_descr);
473 debug_object_activate(work, &work_debug_descr);
474 return 0;
475 }
476 WARN_ON_ONCE(1);
477 return 0;
478
479 case ODEBUG_STATE_ACTIVE:
480 WARN_ON(1);
481
482 default:
483 return 0;
484 }
485}
486
487/*
488 * fixup_free is called when:
489 * - an active object is freed
490 */
491static int work_fixup_free(void *addr, enum debug_obj_state state)
492{
493 struct work_struct *work = addr;
494
495 switch (state) {
496 case ODEBUG_STATE_ACTIVE:
497 cancel_work_sync(work);
498 debug_object_free(work, &work_debug_descr);
499 return 1;
500 default:
501 return 0;
502 }
503}
504
505static struct debug_obj_descr work_debug_descr = {
506 .name = "work_struct",
507 .debug_hint = work_debug_hint,
508 .fixup_init = work_fixup_init,
509 .fixup_activate = work_fixup_activate,
510 .fixup_free = work_fixup_free,
511};
512
513static inline void debug_work_activate(struct work_struct *work)
514{
515 debug_object_activate(work, &work_debug_descr);
516}
517
518static inline void debug_work_deactivate(struct work_struct *work)
519{
520 debug_object_deactivate(work, &work_debug_descr);
521}
522
523void __init_work(struct work_struct *work, int onstack)
524{
525 if (onstack)
526 debug_object_init_on_stack(work, &work_debug_descr);
527 else
528 debug_object_init(work, &work_debug_descr);
529}
530EXPORT_SYMBOL_GPL(__init_work);
531
532void destroy_work_on_stack(struct work_struct *work)
533{
534 debug_object_free(work, &work_debug_descr);
535}
536EXPORT_SYMBOL_GPL(destroy_work_on_stack);
537
538void destroy_delayed_work_on_stack(struct delayed_work *work)
539{
540 destroy_timer_on_stack(&work->timer);
541 debug_object_free(&work->work, &work_debug_descr);
542}
543EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
544
545#else
546static inline void debug_work_activate(struct work_struct *work) { }
547static inline void debug_work_deactivate(struct work_struct *work) { }
548#endif
549
550/**
551 * worker_pool_assign_id - allocate ID and assing it to @pool
552 * @pool: the pool pointer of interest
553 *
554 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
555 * successfully, -errno on failure.
556 */
557static int worker_pool_assign_id(struct worker_pool *pool)
558{
559 int ret;
560
561 lockdep_assert_held(&wq_pool_mutex);
562
563 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
564 GFP_KERNEL);
565 if (ret >= 0) {
566 pool->id = ret;
567 return 0;
568 }
569 return ret;
570}
571
572/**
573 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
574 * @wq: the target workqueue
575 * @node: the node ID
576 *
577 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
578 * read locked.
579 * If the pwq needs to be used beyond the locking in effect, the caller is
580 * responsible for guaranteeing that the pwq stays online.
581 *
582 * Return: The unbound pool_workqueue for @node.
583 */
584static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
585 int node)
586{
587 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
588
589 /*
590 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
591 * delayed item is pending. The plan is to keep CPU -> NODE
592 * mapping valid and stable across CPU on/offlines. Once that
593 * happens, this workaround can be removed.
594 */
595 if (unlikely(node == NUMA_NO_NODE))
596 return wq->dfl_pwq;
597
598 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
599}
600
601static unsigned int work_color_to_flags(int color)
602{
603 return color << WORK_STRUCT_COLOR_SHIFT;
604}
605
606static int get_work_color(struct work_struct *work)
607{
608 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
609 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
610}
611
612static int work_next_color(int color)
613{
614 return (color + 1) % WORK_NR_COLORS;
615}
616
617/*
618 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
619 * contain the pointer to the queued pwq. Once execution starts, the flag
620 * is cleared and the high bits contain OFFQ flags and pool ID.
621 *
622 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
623 * and clear_work_data() can be used to set the pwq, pool or clear
624 * work->data. These functions should only be called while the work is
625 * owned - ie. while the PENDING bit is set.
626 *
627 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
628 * corresponding to a work. Pool is available once the work has been
629 * queued anywhere after initialization until it is sync canceled. pwq is
630 * available only while the work item is queued.
631 *
632 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
633 * canceled. While being canceled, a work item may have its PENDING set
634 * but stay off timer and worklist for arbitrarily long and nobody should
635 * try to steal the PENDING bit.
636 */
637static inline void set_work_data(struct work_struct *work, unsigned long data,
638 unsigned long flags)
639{
640 WARN_ON_ONCE(!work_pending(work));
641 atomic_long_set(&work->data, data | flags | work_static(work));
642}
643
644static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
645 unsigned long extra_flags)
646{
647 set_work_data(work, (unsigned long)pwq,
648 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
649}
650
651static void set_work_pool_and_keep_pending(struct work_struct *work,
652 int pool_id)
653{
654 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
655 WORK_STRUCT_PENDING);
656}
657
658static void set_work_pool_and_clear_pending(struct work_struct *work,
659 int pool_id)
660{
661 /*
662 * The following wmb is paired with the implied mb in
663 * test_and_set_bit(PENDING) and ensures all updates to @work made
664 * here are visible to and precede any updates by the next PENDING
665 * owner.
666 */
667 smp_wmb();
668 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
669 /*
670 * The following mb guarantees that previous clear of a PENDING bit
671 * will not be reordered with any speculative LOADS or STORES from
672 * work->current_func, which is executed afterwards. This possible
673 * reordering can lead to a missed execution on attempt to qeueue
674 * the same @work. E.g. consider this case:
675 *
676 * CPU#0 CPU#1
677 * ---------------------------- --------------------------------
678 *
679 * 1 STORE event_indicated
680 * 2 queue_work_on() {
681 * 3 test_and_set_bit(PENDING)
682 * 4 } set_..._and_clear_pending() {
683 * 5 set_work_data() # clear bit
684 * 6 smp_mb()
685 * 7 work->current_func() {
686 * 8 LOAD event_indicated
687 * }
688 *
689 * Without an explicit full barrier speculative LOAD on line 8 can
690 * be executed before CPU#0 does STORE on line 1. If that happens,
691 * CPU#0 observes the PENDING bit is still set and new execution of
692 * a @work is not queued in a hope, that CPU#1 will eventually
693 * finish the queued @work. Meanwhile CPU#1 does not see
694 * event_indicated is set, because speculative LOAD was executed
695 * before actual STORE.
696 */
697 smp_mb();
698}
699
700static void clear_work_data(struct work_struct *work)
701{
702 smp_wmb(); /* see set_work_pool_and_clear_pending() */
703 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
704}
705
706static struct pool_workqueue *get_work_pwq(struct work_struct *work)
707{
708 unsigned long data = atomic_long_read(&work->data);
709
710 if (data & WORK_STRUCT_PWQ)
711 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
712 else
713 return NULL;
714}
715
716/**
717 * get_work_pool - return the worker_pool a given work was associated with
718 * @work: the work item of interest
719 *
720 * Pools are created and destroyed under wq_pool_mutex, and allows read
721 * access under sched-RCU read lock. As such, this function should be
722 * called under wq_pool_mutex or with preemption disabled.
723 *
724 * All fields of the returned pool are accessible as long as the above
725 * mentioned locking is in effect. If the returned pool needs to be used
726 * beyond the critical section, the caller is responsible for ensuring the
727 * returned pool is and stays online.
728 *
729 * Return: The worker_pool @work was last associated with. %NULL if none.
730 */
731static struct worker_pool *get_work_pool(struct work_struct *work)
732{
733 unsigned long data = atomic_long_read(&work->data);
734 int pool_id;
735
736 assert_rcu_or_pool_mutex();
737
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
741
742 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
743 if (pool_id == WORK_OFFQ_POOL_NONE)
744 return NULL;
745
746 return idr_find(&worker_pool_idr, pool_id);
747}
748
749/**
750 * get_work_pool_id - return the worker pool ID a given work is associated with
751 * @work: the work item of interest
752 *
753 * Return: The worker_pool ID @work was last associated with.
754 * %WORK_OFFQ_POOL_NONE if none.
755 */
756static int get_work_pool_id(struct work_struct *work)
757{
758 unsigned long data = atomic_long_read(&work->data);
759
760 if (data & WORK_STRUCT_PWQ)
761 return ((struct pool_workqueue *)
762 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
763
764 return data >> WORK_OFFQ_POOL_SHIFT;
765}
766
767static void mark_work_canceling(struct work_struct *work)
768{
769 unsigned long pool_id = get_work_pool_id(work);
770
771 pool_id <<= WORK_OFFQ_POOL_SHIFT;
772 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
773}
774
775static bool work_is_canceling(struct work_struct *work)
776{
777 unsigned long data = atomic_long_read(&work->data);
778
779 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
780}
781
782/*
783 * Policy functions. These define the policies on how the global worker
784 * pools are managed. Unless noted otherwise, these functions assume that
785 * they're being called with pool->lock held.
786 */
787
788static bool __need_more_worker(struct worker_pool *pool)
789{
790 return !atomic_read(&pool->nr_running);
791}
792
793/*
794 * Need to wake up a worker? Called from anything but currently
795 * running workers.
796 *
797 * Note that, because unbound workers never contribute to nr_running, this
798 * function will always return %true for unbound pools as long as the
799 * worklist isn't empty.
800 */
801static bool need_more_worker(struct worker_pool *pool)
802{
803 return !list_empty(&pool->worklist) && __need_more_worker(pool);
804}
805
806/* Can I start working? Called from busy but !running workers. */
807static bool may_start_working(struct worker_pool *pool)
808{
809 return pool->nr_idle;
810}
811
812/* Do I need to keep working? Called from currently running workers. */
813static bool keep_working(struct worker_pool *pool)
814{
815 return !list_empty(&pool->worklist) &&
816 atomic_read(&pool->nr_running) <= 1;
817}
818
819/* Do we need a new worker? Called from manager. */
820static bool need_to_create_worker(struct worker_pool *pool)
821{
822 return need_more_worker(pool) && !may_start_working(pool);
823}
824
825/* Do we have too many workers and should some go away? */
826static bool too_many_workers(struct worker_pool *pool)
827{
828 bool managing = mutex_is_locked(&pool->manager_arb);
829 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
830 int nr_busy = pool->nr_workers - nr_idle;
831
832 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
833}
834
835/*
836 * Wake up functions.
837 */
838
839/* Return the first idle worker. Safe with preemption disabled */
840static struct worker *first_idle_worker(struct worker_pool *pool)
841{
842 if (unlikely(list_empty(&pool->idle_list)))
843 return NULL;
844
845 return list_first_entry(&pool->idle_list, struct worker, entry);
846}
847
848/**
849 * wake_up_worker - wake up an idle worker
850 * @pool: worker pool to wake worker from
851 *
852 * Wake up the first idle worker of @pool.
853 *
854 * CONTEXT:
855 * spin_lock_irq(pool->lock).
856 */
857static void wake_up_worker(struct worker_pool *pool)
858{
859 struct worker *worker = first_idle_worker(pool);
860
861 if (likely(worker))
862 wake_up_process(worker->task);
863}
864
865/**
866 * wq_worker_waking_up - a worker is waking up
867 * @task: task waking up
868 * @cpu: CPU @task is waking up to
869 *
870 * This function is called during try_to_wake_up() when a worker is
871 * being awoken.
872 *
873 * CONTEXT:
874 * spin_lock_irq(rq->lock)
875 */
876void wq_worker_waking_up(struct task_struct *task, int cpu)
877{
878 struct worker *worker = kthread_data(task);
879
880 if (!(worker->flags & WORKER_NOT_RUNNING)) {
881 WARN_ON_ONCE(worker->pool->cpu != cpu);
882 atomic_inc(&worker->pool->nr_running);
883 }
884}
885
886/**
887 * wq_worker_sleeping - a worker is going to sleep
888 * @task: task going to sleep
889 *
890 * This function is called during schedule() when a busy worker is
891 * going to sleep. Worker on the same cpu can be woken up by
892 * returning pointer to its task.
893 *
894 * CONTEXT:
895 * spin_lock_irq(rq->lock)
896 *
897 * Return:
898 * Worker task on @cpu to wake up, %NULL if none.
899 */
900struct task_struct *wq_worker_sleeping(struct task_struct *task)
901{
902 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
903 struct worker_pool *pool;
904
905 /*
906 * Rescuers, which may not have all the fields set up like normal
907 * workers, also reach here, let's not access anything before
908 * checking NOT_RUNNING.
909 */
910 if (worker->flags & WORKER_NOT_RUNNING)
911 return NULL;
912
913 pool = worker->pool;
914
915 /* this can only happen on the local cpu */
916 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
917 return NULL;
918
919 /*
920 * The counterpart of the following dec_and_test, implied mb,
921 * worklist not empty test sequence is in insert_work().
922 * Please read comment there.
923 *
924 * NOT_RUNNING is clear. This means that we're bound to and
925 * running on the local cpu w/ rq lock held and preemption
926 * disabled, which in turn means that none else could be
927 * manipulating idle_list, so dereferencing idle_list without pool
928 * lock is safe.
929 */
930 if (atomic_dec_and_test(&pool->nr_running) &&
931 !list_empty(&pool->worklist))
932 to_wakeup = first_idle_worker(pool);
933 return to_wakeup ? to_wakeup->task : NULL;
934}
935
936/**
937 * worker_set_flags - set worker flags and adjust nr_running accordingly
938 * @worker: self
939 * @flags: flags to set
940 *
941 * Set @flags in @worker->flags and adjust nr_running accordingly.
942 *
943 * CONTEXT:
944 * spin_lock_irq(pool->lock)
945 */
946static inline void worker_set_flags(struct worker *worker, unsigned int flags)
947{
948 struct worker_pool *pool = worker->pool;
949
950 WARN_ON_ONCE(worker->task != current);
951
952 /* If transitioning into NOT_RUNNING, adjust nr_running. */
953 if ((flags & WORKER_NOT_RUNNING) &&
954 !(worker->flags & WORKER_NOT_RUNNING)) {
955 atomic_dec(&pool->nr_running);
956 }
957
958 worker->flags |= flags;
959}
960
961/**
962 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
963 * @worker: self
964 * @flags: flags to clear
965 *
966 * Clear @flags in @worker->flags and adjust nr_running accordingly.
967 *
968 * CONTEXT:
969 * spin_lock_irq(pool->lock)
970 */
971static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
972{
973 struct worker_pool *pool = worker->pool;
974 unsigned int oflags = worker->flags;
975
976 WARN_ON_ONCE(worker->task != current);
977
978 worker->flags &= ~flags;
979
980 /*
981 * If transitioning out of NOT_RUNNING, increment nr_running. Note
982 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
983 * of multiple flags, not a single flag.
984 */
985 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
986 if (!(worker->flags & WORKER_NOT_RUNNING))
987 atomic_inc(&pool->nr_running);
988}
989
990/**
991 * find_worker_executing_work - find worker which is executing a work
992 * @pool: pool of interest
993 * @work: work to find worker for
994 *
995 * Find a worker which is executing @work on @pool by searching
996 * @pool->busy_hash which is keyed by the address of @work. For a worker
997 * to match, its current execution should match the address of @work and
998 * its work function. This is to avoid unwanted dependency between
999 * unrelated work executions through a work item being recycled while still
1000 * being executed.
1001 *
1002 * This is a bit tricky. A work item may be freed once its execution
1003 * starts and nothing prevents the freed area from being recycled for
1004 * another work item. If the same work item address ends up being reused
1005 * before the original execution finishes, workqueue will identify the
1006 * recycled work item as currently executing and make it wait until the
1007 * current execution finishes, introducing an unwanted dependency.
1008 *
1009 * This function checks the work item address and work function to avoid
1010 * false positives. Note that this isn't complete as one may construct a
1011 * work function which can introduce dependency onto itself through a
1012 * recycled work item. Well, if somebody wants to shoot oneself in the
1013 * foot that badly, there's only so much we can do, and if such deadlock
1014 * actually occurs, it should be easy to locate the culprit work function.
1015 *
1016 * CONTEXT:
1017 * spin_lock_irq(pool->lock).
1018 *
1019 * Return:
1020 * Pointer to worker which is executing @work if found, %NULL
1021 * otherwise.
1022 */
1023static struct worker *find_worker_executing_work(struct worker_pool *pool,
1024 struct work_struct *work)
1025{
1026 struct worker *worker;
1027
1028 hash_for_each_possible(pool->busy_hash, worker, hentry,
1029 (unsigned long)work)
1030 if (worker->current_work == work &&
1031 worker->current_func == work->func)
1032 return worker;
1033
1034 return NULL;
1035}
1036
1037/**
1038 * move_linked_works - move linked works to a list
1039 * @work: start of series of works to be scheduled
1040 * @head: target list to append @work to
1041 * @nextp: out parameter for nested worklist walking
1042 *
1043 * Schedule linked works starting from @work to @head. Work series to
1044 * be scheduled starts at @work and includes any consecutive work with
1045 * WORK_STRUCT_LINKED set in its predecessor.
1046 *
1047 * If @nextp is not NULL, it's updated to point to the next work of
1048 * the last scheduled work. This allows move_linked_works() to be
1049 * nested inside outer list_for_each_entry_safe().
1050 *
1051 * CONTEXT:
1052 * spin_lock_irq(pool->lock).
1053 */
1054static void move_linked_works(struct work_struct *work, struct list_head *head,
1055 struct work_struct **nextp)
1056{
1057 struct work_struct *n;
1058
1059 /*
1060 * Linked worklist will always end before the end of the list,
1061 * use NULL for list head.
1062 */
1063 list_for_each_entry_safe_from(work, n, NULL, entry) {
1064 list_move_tail(&work->entry, head);
1065 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1066 break;
1067 }
1068
1069 /*
1070 * If we're already inside safe list traversal and have moved
1071 * multiple works to the scheduled queue, the next position
1072 * needs to be updated.
1073 */
1074 if (nextp)
1075 *nextp = n;
1076}
1077
1078/**
1079 * get_pwq - get an extra reference on the specified pool_workqueue
1080 * @pwq: pool_workqueue to get
1081 *
1082 * Obtain an extra reference on @pwq. The caller should guarantee that
1083 * @pwq has positive refcnt and be holding the matching pool->lock.
1084 */
1085static void get_pwq(struct pool_workqueue *pwq)
1086{
1087 lockdep_assert_held(&pwq->pool->lock);
1088 WARN_ON_ONCE(pwq->refcnt <= 0);
1089 pwq->refcnt++;
1090}
1091
1092/**
1093 * put_pwq - put a pool_workqueue reference
1094 * @pwq: pool_workqueue to put
1095 *
1096 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1097 * destruction. The caller should be holding the matching pool->lock.
1098 */
1099static void put_pwq(struct pool_workqueue *pwq)
1100{
1101 lockdep_assert_held(&pwq->pool->lock);
1102 if (likely(--pwq->refcnt))
1103 return;
1104 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105 return;
1106 /*
1107 * @pwq can't be released under pool->lock, bounce to
1108 * pwq_unbound_release_workfn(). This never recurses on the same
1109 * pool->lock as this path is taken only for unbound workqueues and
1110 * the release work item is scheduled on a per-cpu workqueue. To
1111 * avoid lockdep warning, unbound pool->locks are given lockdep
1112 * subclass of 1 in get_unbound_pool().
1113 */
1114 schedule_work(&pwq->unbound_release_work);
1115}
1116
1117/**
1118 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1119 * @pwq: pool_workqueue to put (can be %NULL)
1120 *
1121 * put_pwq() with locking. This function also allows %NULL @pwq.
1122 */
1123static void put_pwq_unlocked(struct pool_workqueue *pwq)
1124{
1125 if (pwq) {
1126 /*
1127 * As both pwqs and pools are sched-RCU protected, the
1128 * following lock operations are safe.
1129 */
1130 spin_lock_irq(&pwq->pool->lock);
1131 put_pwq(pwq);
1132 spin_unlock_irq(&pwq->pool->lock);
1133 }
1134}
1135
1136static void pwq_activate_delayed_work(struct work_struct *work)
1137{
1138 struct pool_workqueue *pwq = get_work_pwq(work);
1139
1140 trace_workqueue_activate_work(work);
1141 if (list_empty(&pwq->pool->worklist))
1142 pwq->pool->watchdog_ts = jiffies;
1143 move_linked_works(work, &pwq->pool->worklist, NULL);
1144 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1145 pwq->nr_active++;
1146}
1147
1148static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1149{
1150 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1151 struct work_struct, entry);
1152
1153 pwq_activate_delayed_work(work);
1154}
1155
1156/**
1157 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1158 * @pwq: pwq of interest
1159 * @color: color of work which left the queue
1160 *
1161 * A work either has completed or is removed from pending queue,
1162 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1163 *
1164 * CONTEXT:
1165 * spin_lock_irq(pool->lock).
1166 */
1167static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1168{
1169 /* uncolored work items don't participate in flushing or nr_active */
1170 if (color == WORK_NO_COLOR)
1171 goto out_put;
1172
1173 pwq->nr_in_flight[color]--;
1174
1175 pwq->nr_active--;
1176 if (!list_empty(&pwq->delayed_works)) {
1177 /* one down, submit a delayed one */
1178 if (pwq->nr_active < pwq->max_active)
1179 pwq_activate_first_delayed(pwq);
1180 }
1181
1182 /* is flush in progress and are we at the flushing tip? */
1183 if (likely(pwq->flush_color != color))
1184 goto out_put;
1185
1186 /* are there still in-flight works? */
1187 if (pwq->nr_in_flight[color])
1188 goto out_put;
1189
1190 /* this pwq is done, clear flush_color */
1191 pwq->flush_color = -1;
1192
1193 /*
1194 * If this was the last pwq, wake up the first flusher. It
1195 * will handle the rest.
1196 */
1197 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1198 complete(&pwq->wq->first_flusher->done);
1199out_put:
1200 put_pwq(pwq);
1201}
1202
1203/**
1204 * try_to_grab_pending - steal work item from worklist and disable irq
1205 * @work: work item to steal
1206 * @is_dwork: @work is a delayed_work
1207 * @flags: place to store irq state
1208 *
1209 * Try to grab PENDING bit of @work. This function can handle @work in any
1210 * stable state - idle, on timer or on worklist.
1211 *
1212 * Return:
1213 * 1 if @work was pending and we successfully stole PENDING
1214 * 0 if @work was idle and we claimed PENDING
1215 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1216 * -ENOENT if someone else is canceling @work, this state may persist
1217 * for arbitrarily long
1218 *
1219 * Note:
1220 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1221 * interrupted while holding PENDING and @work off queue, irq must be
1222 * disabled on entry. This, combined with delayed_work->timer being
1223 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1224 *
1225 * On successful return, >= 0, irq is disabled and the caller is
1226 * responsible for releasing it using local_irq_restore(*@flags).
1227 *
1228 * This function is safe to call from any context including IRQ handler.
1229 */
1230static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1231 unsigned long *flags)
1232{
1233 struct worker_pool *pool;
1234 struct pool_workqueue *pwq;
1235
1236 local_irq_save(*flags);
1237
1238 /* try to steal the timer if it exists */
1239 if (is_dwork) {
1240 struct delayed_work *dwork = to_delayed_work(work);
1241
1242 /*
1243 * dwork->timer is irqsafe. If del_timer() fails, it's
1244 * guaranteed that the timer is not queued anywhere and not
1245 * running on the local CPU.
1246 */
1247 if (likely(del_timer(&dwork->timer)))
1248 return 1;
1249 }
1250
1251 /* try to claim PENDING the normal way */
1252 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1253 return 0;
1254
1255 /*
1256 * The queueing is in progress, or it is already queued. Try to
1257 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258 */
1259 pool = get_work_pool(work);
1260 if (!pool)
1261 goto fail;
1262
1263 spin_lock(&pool->lock);
1264 /*
1265 * work->data is guaranteed to point to pwq only while the work
1266 * item is queued on pwq->wq, and both updating work->data to point
1267 * to pwq on queueing and to pool on dequeueing are done under
1268 * pwq->pool->lock. This in turn guarantees that, if work->data
1269 * points to pwq which is associated with a locked pool, the work
1270 * item is currently queued on that pool.
1271 */
1272 pwq = get_work_pwq(work);
1273 if (pwq && pwq->pool == pool) {
1274 debug_work_deactivate(work);
1275
1276 /*
1277 * A delayed work item cannot be grabbed directly because
1278 * it might have linked NO_COLOR work items which, if left
1279 * on the delayed_list, will confuse pwq->nr_active
1280 * management later on and cause stall. Make sure the work
1281 * item is activated before grabbing.
1282 */
1283 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1284 pwq_activate_delayed_work(work);
1285
1286 list_del_init(&work->entry);
1287 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1288
1289 /* work->data points to pwq iff queued, point to pool */
1290 set_work_pool_and_keep_pending(work, pool->id);
1291
1292 spin_unlock(&pool->lock);
1293 return 1;
1294 }
1295 spin_unlock(&pool->lock);
1296fail:
1297 local_irq_restore(*flags);
1298 if (work_is_canceling(work))
1299 return -ENOENT;
1300 cpu_relax();
1301 return -EAGAIN;
1302}
1303
1304/**
1305 * insert_work - insert a work into a pool
1306 * @pwq: pwq @work belongs to
1307 * @work: work to insert
1308 * @head: insertion point
1309 * @extra_flags: extra WORK_STRUCT_* flags to set
1310 *
1311 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1312 * work_struct flags.
1313 *
1314 * CONTEXT:
1315 * spin_lock_irq(pool->lock).
1316 */
1317static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1318 struct list_head *head, unsigned int extra_flags)
1319{
1320 struct worker_pool *pool = pwq->pool;
1321
1322 /* we own @work, set data and link */
1323 set_work_pwq(work, pwq, extra_flags);
1324 list_add_tail(&work->entry, head);
1325 get_pwq(pwq);
1326
1327 /*
1328 * Ensure either wq_worker_sleeping() sees the above
1329 * list_add_tail() or we see zero nr_running to avoid workers lying
1330 * around lazily while there are works to be processed.
1331 */
1332 smp_mb();
1333
1334 if (__need_more_worker(pool))
1335 wake_up_worker(pool);
1336}
1337
1338/*
1339 * Test whether @work is being queued from another work executing on the
1340 * same workqueue.
1341 */
1342static bool is_chained_work(struct workqueue_struct *wq)
1343{
1344 struct worker *worker;
1345
1346 worker = current_wq_worker();
1347 /*
1348 * Return %true iff I'm a worker execuing a work item on @wq. If
1349 * I'm @worker, it's safe to dereference it without locking.
1350 */
1351 return worker && worker->current_pwq->wq == wq;
1352}
1353
1354/*
1355 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1356 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1357 * avoid perturbing sensitive tasks.
1358 */
1359static int wq_select_unbound_cpu(int cpu)
1360{
1361 static bool printed_dbg_warning;
1362 int new_cpu;
1363
1364 if (likely(!wq_debug_force_rr_cpu)) {
1365 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1366 return cpu;
1367 } else if (!printed_dbg_warning) {
1368 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1369 printed_dbg_warning = true;
1370 }
1371
1372 if (cpumask_empty(wq_unbound_cpumask))
1373 return cpu;
1374
1375 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1376 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1377 if (unlikely(new_cpu >= nr_cpu_ids)) {
1378 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1379 if (unlikely(new_cpu >= nr_cpu_ids))
1380 return cpu;
1381 }
1382 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1383
1384 return new_cpu;
1385}
1386
1387static void __queue_work(int cpu, struct workqueue_struct *wq,
1388 struct work_struct *work)
1389{
1390 struct pool_workqueue *pwq;
1391 struct worker_pool *last_pool;
1392 struct list_head *worklist;
1393 unsigned int work_flags;
1394 unsigned int req_cpu = cpu;
1395
1396 /*
1397 * While a work item is PENDING && off queue, a task trying to
1398 * steal the PENDING will busy-loop waiting for it to either get
1399 * queued or lose PENDING. Grabbing PENDING and queueing should
1400 * happen with IRQ disabled.
1401 */
1402 WARN_ON_ONCE(!irqs_disabled());
1403
1404 debug_work_activate(work);
1405
1406 /* if draining, only works from the same workqueue are allowed */
1407 if (unlikely(wq->flags & __WQ_DRAINING) &&
1408 WARN_ON_ONCE(!is_chained_work(wq)))
1409 return;
1410retry:
1411 if (req_cpu == WORK_CPU_UNBOUND)
1412 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1413
1414 /* pwq which will be used unless @work is executing elsewhere */
1415 if (!(wq->flags & WQ_UNBOUND))
1416 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1417 else
1418 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419
1420 /*
1421 * If @work was previously on a different pool, it might still be
1422 * running there, in which case the work needs to be queued on that
1423 * pool to guarantee non-reentrancy.
1424 */
1425 last_pool = get_work_pool(work);
1426 if (last_pool && last_pool != pwq->pool) {
1427 struct worker *worker;
1428
1429 spin_lock(&last_pool->lock);
1430
1431 worker = find_worker_executing_work(last_pool, work);
1432
1433 if (worker && worker->current_pwq->wq == wq) {
1434 pwq = worker->current_pwq;
1435 } else {
1436 /* meh... not running there, queue here */
1437 spin_unlock(&last_pool->lock);
1438 spin_lock(&pwq->pool->lock);
1439 }
1440 } else {
1441 spin_lock(&pwq->pool->lock);
1442 }
1443
1444 /*
1445 * pwq is determined and locked. For unbound pools, we could have
1446 * raced with pwq release and it could already be dead. If its
1447 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1448 * without another pwq replacing it in the numa_pwq_tbl or while
1449 * work items are executing on it, so the retrying is guaranteed to
1450 * make forward-progress.
1451 */
1452 if (unlikely(!pwq->refcnt)) {
1453 if (wq->flags & WQ_UNBOUND) {
1454 spin_unlock(&pwq->pool->lock);
1455 cpu_relax();
1456 goto retry;
1457 }
1458 /* oops */
1459 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1460 wq->name, cpu);
1461 }
1462
1463 /* pwq determined, queue */
1464 trace_workqueue_queue_work(req_cpu, pwq, work);
1465
1466 if (WARN_ON(!list_empty(&work->entry))) {
1467 spin_unlock(&pwq->pool->lock);
1468 return;
1469 }
1470
1471 pwq->nr_in_flight[pwq->work_color]++;
1472 work_flags = work_color_to_flags(pwq->work_color);
1473
1474 if (likely(pwq->nr_active < pwq->max_active)) {
1475 trace_workqueue_activate_work(work);
1476 pwq->nr_active++;
1477 worklist = &pwq->pool->worklist;
1478 if (list_empty(worklist))
1479 pwq->pool->watchdog_ts = jiffies;
1480 } else {
1481 work_flags |= WORK_STRUCT_DELAYED;
1482 worklist = &pwq->delayed_works;
1483 }
1484
1485 insert_work(pwq, work, worklist, work_flags);
1486
1487 spin_unlock(&pwq->pool->lock);
1488}
1489
1490/**
1491 * queue_work_on - queue work on specific cpu
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @work: work to queue
1495 *
1496 * We queue the work to a specific CPU, the caller must ensure it
1497 * can't go away.
1498 *
1499 * Return: %false if @work was already on a queue, %true otherwise.
1500 */
1501bool queue_work_on(int cpu, struct workqueue_struct *wq,
1502 struct work_struct *work)
1503{
1504 bool ret = false;
1505 unsigned long flags;
1506
1507 local_irq_save(flags);
1508
1509 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1510 __queue_work(cpu, wq, work);
1511 ret = true;
1512 }
1513
1514 local_irq_restore(flags);
1515 return ret;
1516}
1517EXPORT_SYMBOL(queue_work_on);
1518
1519void delayed_work_timer_fn(unsigned long __data)
1520{
1521 struct delayed_work *dwork = (struct delayed_work *)__data;
1522
1523 /* should have been called from irqsafe timer with irq already off */
1524 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1525}
1526EXPORT_SYMBOL(delayed_work_timer_fn);
1527
1528static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1529 struct delayed_work *dwork, unsigned long delay)
1530{
1531 struct timer_list *timer = &dwork->timer;
1532 struct work_struct *work = &dwork->work;
1533
1534 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1535 timer->data != (unsigned long)dwork);
1536 WARN_ON_ONCE(timer_pending(timer));
1537 WARN_ON_ONCE(!list_empty(&work->entry));
1538
1539 /*
1540 * If @delay is 0, queue @dwork->work immediately. This is for
1541 * both optimization and correctness. The earliest @timer can
1542 * expire is on the closest next tick and delayed_work users depend
1543 * on that there's no such delay when @delay is 0.
1544 */
1545 if (!delay) {
1546 __queue_work(cpu, wq, &dwork->work);
1547 return;
1548 }
1549
1550 timer_stats_timer_set_start_info(&dwork->timer);
1551
1552 dwork->wq = wq;
1553 dwork->cpu = cpu;
1554 timer->expires = jiffies + delay;
1555
1556 if (unlikely(cpu != WORK_CPU_UNBOUND))
1557 add_timer_on(timer, cpu);
1558 else
1559 add_timer(timer);
1560}
1561
1562/**
1563 * queue_delayed_work_on - queue work on specific CPU after delay
1564 * @cpu: CPU number to execute work on
1565 * @wq: workqueue to use
1566 * @dwork: work to queue
1567 * @delay: number of jiffies to wait before queueing
1568 *
1569 * Return: %false if @work was already on a queue, %true otherwise. If
1570 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1571 * execution.
1572 */
1573bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1574 struct delayed_work *dwork, unsigned long delay)
1575{
1576 struct work_struct *work = &dwork->work;
1577 bool ret = false;
1578 unsigned long flags;
1579
1580 /* read the comment in __queue_work() */
1581 local_irq_save(flags);
1582
1583 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1584 __queue_delayed_work(cpu, wq, dwork, delay);
1585 ret = true;
1586 }
1587
1588 local_irq_restore(flags);
1589 return ret;
1590}
1591EXPORT_SYMBOL(queue_delayed_work_on);
1592
1593/**
1594 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1595 * @cpu: CPU number to execute work on
1596 * @wq: workqueue to use
1597 * @dwork: work to queue
1598 * @delay: number of jiffies to wait before queueing
1599 *
1600 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1601 * modify @dwork's timer so that it expires after @delay. If @delay is
1602 * zero, @work is guaranteed to be scheduled immediately regardless of its
1603 * current state.
1604 *
1605 * Return: %false if @dwork was idle and queued, %true if @dwork was
1606 * pending and its timer was modified.
1607 *
1608 * This function is safe to call from any context including IRQ handler.
1609 * See try_to_grab_pending() for details.
1610 */
1611bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1612 struct delayed_work *dwork, unsigned long delay)
1613{
1614 unsigned long flags;
1615 int ret;
1616
1617 do {
1618 ret = try_to_grab_pending(&dwork->work, true, &flags);
1619 } while (unlikely(ret == -EAGAIN));
1620
1621 if (likely(ret >= 0)) {
1622 __queue_delayed_work(cpu, wq, dwork, delay);
1623 local_irq_restore(flags);
1624 }
1625
1626 /* -ENOENT from try_to_grab_pending() becomes %true */
1627 return ret;
1628}
1629EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1630
1631/**
1632 * worker_enter_idle - enter idle state
1633 * @worker: worker which is entering idle state
1634 *
1635 * @worker is entering idle state. Update stats and idle timer if
1636 * necessary.
1637 *
1638 * LOCKING:
1639 * spin_lock_irq(pool->lock).
1640 */
1641static void worker_enter_idle(struct worker *worker)
1642{
1643 struct worker_pool *pool = worker->pool;
1644
1645 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1646 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1647 (worker->hentry.next || worker->hentry.pprev)))
1648 return;
1649
1650 /* can't use worker_set_flags(), also called from create_worker() */
1651 worker->flags |= WORKER_IDLE;
1652 pool->nr_idle++;
1653 worker->last_active = jiffies;
1654
1655 /* idle_list is LIFO */
1656 list_add(&worker->entry, &pool->idle_list);
1657
1658 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1659 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1660
1661 /*
1662 * Sanity check nr_running. Because wq_unbind_fn() releases
1663 * pool->lock between setting %WORKER_UNBOUND and zapping
1664 * nr_running, the warning may trigger spuriously. Check iff
1665 * unbind is not in progress.
1666 */
1667 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1668 pool->nr_workers == pool->nr_idle &&
1669 atomic_read(&pool->nr_running));
1670}
1671
1672/**
1673 * worker_leave_idle - leave idle state
1674 * @worker: worker which is leaving idle state
1675 *
1676 * @worker is leaving idle state. Update stats.
1677 *
1678 * LOCKING:
1679 * spin_lock_irq(pool->lock).
1680 */
1681static void worker_leave_idle(struct worker *worker)
1682{
1683 struct worker_pool *pool = worker->pool;
1684
1685 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1686 return;
1687 worker_clr_flags(worker, WORKER_IDLE);
1688 pool->nr_idle--;
1689 list_del_init(&worker->entry);
1690}
1691
1692static struct worker *alloc_worker(int node)
1693{
1694 struct worker *worker;
1695
1696 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1697 if (worker) {
1698 INIT_LIST_HEAD(&worker->entry);
1699 INIT_LIST_HEAD(&worker->scheduled);
1700 INIT_LIST_HEAD(&worker->node);
1701 /* on creation a worker is in !idle && prep state */
1702 worker->flags = WORKER_PREP;
1703 }
1704 return worker;
1705}
1706
1707/**
1708 * worker_attach_to_pool() - attach a worker to a pool
1709 * @worker: worker to be attached
1710 * @pool: the target pool
1711 *
1712 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1713 * cpu-binding of @worker are kept coordinated with the pool across
1714 * cpu-[un]hotplugs.
1715 */
1716static void worker_attach_to_pool(struct worker *worker,
1717 struct worker_pool *pool)
1718{
1719 mutex_lock(&pool->attach_mutex);
1720
1721 /*
1722 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723 * online CPUs. It'll be re-applied when any of the CPUs come up.
1724 */
1725 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1726
1727 /*
1728 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1729 * stable across this function. See the comments above the
1730 * flag definition for details.
1731 */
1732 if (pool->flags & POOL_DISASSOCIATED)
1733 worker->flags |= WORKER_UNBOUND;
1734
1735 list_add_tail(&worker->node, &pool->workers);
1736
1737 mutex_unlock(&pool->attach_mutex);
1738}
1739
1740/**
1741 * worker_detach_from_pool() - detach a worker from its pool
1742 * @worker: worker which is attached to its pool
1743 * @pool: the pool @worker is attached to
1744 *
1745 * Undo the attaching which had been done in worker_attach_to_pool(). The
1746 * caller worker shouldn't access to the pool after detached except it has
1747 * other reference to the pool.
1748 */
1749static void worker_detach_from_pool(struct worker *worker,
1750 struct worker_pool *pool)
1751{
1752 struct completion *detach_completion = NULL;
1753
1754 mutex_lock(&pool->attach_mutex);
1755 list_del(&worker->node);
1756 if (list_empty(&pool->workers))
1757 detach_completion = pool->detach_completion;
1758 mutex_unlock(&pool->attach_mutex);
1759
1760 /* clear leftover flags without pool->lock after it is detached */
1761 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1762
1763 if (detach_completion)
1764 complete(detach_completion);
1765}
1766
1767/**
1768 * create_worker - create a new workqueue worker
1769 * @pool: pool the new worker will belong to
1770 *
1771 * Create and start a new worker which is attached to @pool.
1772 *
1773 * CONTEXT:
1774 * Might sleep. Does GFP_KERNEL allocations.
1775 *
1776 * Return:
1777 * Pointer to the newly created worker.
1778 */
1779static struct worker *create_worker(struct worker_pool *pool)
1780{
1781 struct worker *worker = NULL;
1782 int id = -1;
1783 char id_buf[16];
1784
1785 /* ID is needed to determine kthread name */
1786 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1787 if (id < 0)
1788 goto fail;
1789
1790 worker = alloc_worker(pool->node);
1791 if (!worker)
1792 goto fail;
1793
1794 worker->pool = pool;
1795 worker->id = id;
1796
1797 if (pool->cpu >= 0)
1798 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1799 pool->attrs->nice < 0 ? "H" : "");
1800 else
1801 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1802
1803 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1804 "kworker/%s", id_buf);
1805 if (IS_ERR(worker->task))
1806 goto fail;
1807
1808 set_user_nice(worker->task, pool->attrs->nice);
1809 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1810
1811 /* successful, attach the worker to the pool */
1812 worker_attach_to_pool(worker, pool);
1813
1814 /* start the newly created worker */
1815 spin_lock_irq(&pool->lock);
1816 worker->pool->nr_workers++;
1817 worker_enter_idle(worker);
1818 wake_up_process(worker->task);
1819 spin_unlock_irq(&pool->lock);
1820
1821 return worker;
1822
1823fail:
1824 if (id >= 0)
1825 ida_simple_remove(&pool->worker_ida, id);
1826 kfree(worker);
1827 return NULL;
1828}
1829
1830/**
1831 * destroy_worker - destroy a workqueue worker
1832 * @worker: worker to be destroyed
1833 *
1834 * Destroy @worker and adjust @pool stats accordingly. The worker should
1835 * be idle.
1836 *
1837 * CONTEXT:
1838 * spin_lock_irq(pool->lock).
1839 */
1840static void destroy_worker(struct worker *worker)
1841{
1842 struct worker_pool *pool = worker->pool;
1843
1844 lockdep_assert_held(&pool->lock);
1845
1846 /* sanity check frenzy */
1847 if (WARN_ON(worker->current_work) ||
1848 WARN_ON(!list_empty(&worker->scheduled)) ||
1849 WARN_ON(!(worker->flags & WORKER_IDLE)))
1850 return;
1851
1852 pool->nr_workers--;
1853 pool->nr_idle--;
1854
1855 list_del_init(&worker->entry);
1856 worker->flags |= WORKER_DIE;
1857 wake_up_process(worker->task);
1858}
1859
1860static void idle_worker_timeout(unsigned long __pool)
1861{
1862 struct worker_pool *pool = (void *)__pool;
1863
1864 spin_lock_irq(&pool->lock);
1865
1866 while (too_many_workers(pool)) {
1867 struct worker *worker;
1868 unsigned long expires;
1869
1870 /* idle_list is kept in LIFO order, check the last one */
1871 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1872 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873
1874 if (time_before(jiffies, expires)) {
1875 mod_timer(&pool->idle_timer, expires);
1876 break;
1877 }
1878
1879 destroy_worker(worker);
1880 }
1881
1882 spin_unlock_irq(&pool->lock);
1883}
1884
1885static void send_mayday(struct work_struct *work)
1886{
1887 struct pool_workqueue *pwq = get_work_pwq(work);
1888 struct workqueue_struct *wq = pwq->wq;
1889
1890 lockdep_assert_held(&wq_mayday_lock);
1891
1892 if (!wq->rescuer)
1893 return;
1894
1895 /* mayday mayday mayday */
1896 if (list_empty(&pwq->mayday_node)) {
1897 /*
1898 * If @pwq is for an unbound wq, its base ref may be put at
1899 * any time due to an attribute change. Pin @pwq until the
1900 * rescuer is done with it.
1901 */
1902 get_pwq(pwq);
1903 list_add_tail(&pwq->mayday_node, &wq->maydays);
1904 wake_up_process(wq->rescuer->task);
1905 }
1906}
1907
1908static void pool_mayday_timeout(unsigned long __pool)
1909{
1910 struct worker_pool *pool = (void *)__pool;
1911 struct work_struct *work;
1912
1913 spin_lock_irq(&pool->lock);
1914 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1915
1916 if (need_to_create_worker(pool)) {
1917 /*
1918 * We've been trying to create a new worker but
1919 * haven't been successful. We might be hitting an
1920 * allocation deadlock. Send distress signals to
1921 * rescuers.
1922 */
1923 list_for_each_entry(work, &pool->worklist, entry)
1924 send_mayday(work);
1925 }
1926
1927 spin_unlock(&wq_mayday_lock);
1928 spin_unlock_irq(&pool->lock);
1929
1930 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1931}
1932
1933/**
1934 * maybe_create_worker - create a new worker if necessary
1935 * @pool: pool to create a new worker for
1936 *
1937 * Create a new worker for @pool if necessary. @pool is guaranteed to
1938 * have at least one idle worker on return from this function. If
1939 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1940 * sent to all rescuers with works scheduled on @pool to resolve
1941 * possible allocation deadlock.
1942 *
1943 * On return, need_to_create_worker() is guaranteed to be %false and
1944 * may_start_working() %true.
1945 *
1946 * LOCKING:
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times. Does GFP_KERNEL allocations. Called only from
1949 * manager.
1950 */
1951static void maybe_create_worker(struct worker_pool *pool)
1952__releases(&pool->lock)
1953__acquires(&pool->lock)
1954{
1955restart:
1956 spin_unlock_irq(&pool->lock);
1957
1958 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1959 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1960
1961 while (true) {
1962 if (create_worker(pool) || !need_to_create_worker(pool))
1963 break;
1964
1965 schedule_timeout_interruptible(CREATE_COOLDOWN);
1966
1967 if (!need_to_create_worker(pool))
1968 break;
1969 }
1970
1971 del_timer_sync(&pool->mayday_timer);
1972 spin_lock_irq(&pool->lock);
1973 /*
1974 * This is necessary even after a new worker was just successfully
1975 * created as @pool->lock was dropped and the new worker might have
1976 * already become busy.
1977 */
1978 if (need_to_create_worker(pool))
1979 goto restart;
1980}
1981
1982/**
1983 * manage_workers - manage worker pool
1984 * @worker: self
1985 *
1986 * Assume the manager role and manage the worker pool @worker belongs
1987 * to. At any given time, there can be only zero or one manager per
1988 * pool. The exclusion is handled automatically by this function.
1989 *
1990 * The caller can safely start processing works on false return. On
1991 * true return, it's guaranteed that need_to_create_worker() is false
1992 * and may_start_working() is true.
1993 *
1994 * CONTEXT:
1995 * spin_lock_irq(pool->lock) which may be released and regrabbed
1996 * multiple times. Does GFP_KERNEL allocations.
1997 *
1998 * Return:
1999 * %false if the pool doesn't need management and the caller can safely
2000 * start processing works, %true if management function was performed and
2001 * the conditions that the caller verified before calling the function may
2002 * no longer be true.
2003 */
2004static bool manage_workers(struct worker *worker)
2005{
2006 struct worker_pool *pool = worker->pool;
2007
2008 /*
2009 * Anyone who successfully grabs manager_arb wins the arbitration
2010 * and becomes the manager. mutex_trylock() on pool->manager_arb
2011 * failure while holding pool->lock reliably indicates that someone
2012 * else is managing the pool and the worker which failed trylock
2013 * can proceed to executing work items. This means that anyone
2014 * grabbing manager_arb is responsible for actually performing
2015 * manager duties. If manager_arb is grabbed and released without
2016 * actual management, the pool may stall indefinitely.
2017 */
2018 if (!mutex_trylock(&pool->manager_arb))
2019 return false;
2020 pool->manager = worker;
2021
2022 maybe_create_worker(pool);
2023
2024 pool->manager = NULL;
2025 mutex_unlock(&pool->manager_arb);
2026 return true;
2027}
2028
2029/**
2030 * process_one_work - process single work
2031 * @worker: self
2032 * @work: work to process
2033 *
2034 * Process @work. This function contains all the logics necessary to
2035 * process a single work including synchronization against and
2036 * interaction with other workers on the same cpu, queueing and
2037 * flushing. As long as context requirement is met, any worker can
2038 * call this function to process a work.
2039 *
2040 * CONTEXT:
2041 * spin_lock_irq(pool->lock) which is released and regrabbed.
2042 */
2043static void process_one_work(struct worker *worker, struct work_struct *work)
2044__releases(&pool->lock)
2045__acquires(&pool->lock)
2046{
2047 struct pool_workqueue *pwq = get_work_pwq(work);
2048 struct worker_pool *pool = worker->pool;
2049 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2050 int work_color;
2051 struct worker *collision;
2052#ifdef CONFIG_LOCKDEP
2053 /*
2054 * It is permissible to free the struct work_struct from
2055 * inside the function that is called from it, this we need to
2056 * take into account for lockdep too. To avoid bogus "held
2057 * lock freed" warnings as well as problems when looking into
2058 * work->lockdep_map, make a copy and use that here.
2059 */
2060 struct lockdep_map lockdep_map;
2061
2062 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2063#endif
2064 /* ensure we're on the correct CPU */
2065 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2066 raw_smp_processor_id() != pool->cpu);
2067
2068 /*
2069 * A single work shouldn't be executed concurrently by
2070 * multiple workers on a single cpu. Check whether anyone is
2071 * already processing the work. If so, defer the work to the
2072 * currently executing one.
2073 */
2074 collision = find_worker_executing_work(pool, work);
2075 if (unlikely(collision)) {
2076 move_linked_works(work, &collision->scheduled, NULL);
2077 return;
2078 }
2079
2080 /* claim and dequeue */
2081 debug_work_deactivate(work);
2082 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2083 worker->current_work = work;
2084 worker->current_func = work->func;
2085 worker->current_pwq = pwq;
2086 work_color = get_work_color(work);
2087
2088 list_del_init(&work->entry);
2089
2090 /*
2091 * CPU intensive works don't participate in concurrency management.
2092 * They're the scheduler's responsibility. This takes @worker out
2093 * of concurrency management and the next code block will chain
2094 * execution of the pending work items.
2095 */
2096 if (unlikely(cpu_intensive))
2097 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2098
2099 /*
2100 * Wake up another worker if necessary. The condition is always
2101 * false for normal per-cpu workers since nr_running would always
2102 * be >= 1 at this point. This is used to chain execution of the
2103 * pending work items for WORKER_NOT_RUNNING workers such as the
2104 * UNBOUND and CPU_INTENSIVE ones.
2105 */
2106 if (need_more_worker(pool))
2107 wake_up_worker(pool);
2108
2109 /*
2110 * Record the last pool and clear PENDING which should be the last
2111 * update to @work. Also, do this inside @pool->lock so that
2112 * PENDING and queued state changes happen together while IRQ is
2113 * disabled.
2114 */
2115 set_work_pool_and_clear_pending(work, pool->id);
2116
2117 spin_unlock_irq(&pool->lock);
2118
2119 lock_map_acquire_read(&pwq->wq->lockdep_map);
2120 lock_map_acquire(&lockdep_map);
2121 trace_workqueue_execute_start(work);
2122 worker->current_func(work);
2123 /*
2124 * While we must be careful to not use "work" after this, the trace
2125 * point will only record its address.
2126 */
2127 trace_workqueue_execute_end(work);
2128 lock_map_release(&lockdep_map);
2129 lock_map_release(&pwq->wq->lockdep_map);
2130
2131 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2132 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2133 " last function: %pf\n",
2134 current->comm, preempt_count(), task_pid_nr(current),
2135 worker->current_func);
2136 debug_show_held_locks(current);
2137 dump_stack();
2138 }
2139
2140 /*
2141 * The following prevents a kworker from hogging CPU on !PREEMPT
2142 * kernels, where a requeueing work item waiting for something to
2143 * happen could deadlock with stop_machine as such work item could
2144 * indefinitely requeue itself while all other CPUs are trapped in
2145 * stop_machine. At the same time, report a quiescent RCU state so
2146 * the same condition doesn't freeze RCU.
2147 */
2148 cond_resched_rcu_qs();
2149
2150 spin_lock_irq(&pool->lock);
2151
2152 /* clear cpu intensive status */
2153 if (unlikely(cpu_intensive))
2154 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2155
2156 /* we're done with it, release */
2157 hash_del(&worker->hentry);
2158 worker->current_work = NULL;
2159 worker->current_func = NULL;
2160 worker->current_pwq = NULL;
2161 worker->desc_valid = false;
2162 pwq_dec_nr_in_flight(pwq, work_color);
2163}
2164
2165/**
2166 * process_scheduled_works - process scheduled works
2167 * @worker: self
2168 *
2169 * Process all scheduled works. Please note that the scheduled list
2170 * may change while processing a work, so this function repeatedly
2171 * fetches a work from the top and executes it.
2172 *
2173 * CONTEXT:
2174 * spin_lock_irq(pool->lock) which may be released and regrabbed
2175 * multiple times.
2176 */
2177static void process_scheduled_works(struct worker *worker)
2178{
2179 while (!list_empty(&worker->scheduled)) {
2180 struct work_struct *work = list_first_entry(&worker->scheduled,
2181 struct work_struct, entry);
2182 process_one_work(worker, work);
2183 }
2184}
2185
2186/**
2187 * worker_thread - the worker thread function
2188 * @__worker: self
2189 *
2190 * The worker thread function. All workers belong to a worker_pool -
2191 * either a per-cpu one or dynamic unbound one. These workers process all
2192 * work items regardless of their specific target workqueue. The only
2193 * exception is work items which belong to workqueues with a rescuer which
2194 * will be explained in rescuer_thread().
2195 *
2196 * Return: 0
2197 */
2198static int worker_thread(void *__worker)
2199{
2200 struct worker *worker = __worker;
2201 struct worker_pool *pool = worker->pool;
2202
2203 /* tell the scheduler that this is a workqueue worker */
2204 worker->task->flags |= PF_WQ_WORKER;
2205woke_up:
2206 spin_lock_irq(&pool->lock);
2207
2208 /* am I supposed to die? */
2209 if (unlikely(worker->flags & WORKER_DIE)) {
2210 spin_unlock_irq(&pool->lock);
2211 WARN_ON_ONCE(!list_empty(&worker->entry));
2212 worker->task->flags &= ~PF_WQ_WORKER;
2213
2214 set_task_comm(worker->task, "kworker/dying");
2215 ida_simple_remove(&pool->worker_ida, worker->id);
2216 worker_detach_from_pool(worker, pool);
2217 kfree(worker);
2218 return 0;
2219 }
2220
2221 worker_leave_idle(worker);
2222recheck:
2223 /* no more worker necessary? */
2224 if (!need_more_worker(pool))
2225 goto sleep;
2226
2227 /* do we need to manage? */
2228 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2229 goto recheck;
2230
2231 /*
2232 * ->scheduled list can only be filled while a worker is
2233 * preparing to process a work or actually processing it.
2234 * Make sure nobody diddled with it while I was sleeping.
2235 */
2236 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2237
2238 /*
2239 * Finish PREP stage. We're guaranteed to have at least one idle
2240 * worker or that someone else has already assumed the manager
2241 * role. This is where @worker starts participating in concurrency
2242 * management if applicable and concurrency management is restored
2243 * after being rebound. See rebind_workers() for details.
2244 */
2245 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2246
2247 do {
2248 struct work_struct *work =
2249 list_first_entry(&pool->worklist,
2250 struct work_struct, entry);
2251
2252 pool->watchdog_ts = jiffies;
2253
2254 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2255 /* optimization path, not strictly necessary */
2256 process_one_work(worker, work);
2257 if (unlikely(!list_empty(&worker->scheduled)))
2258 process_scheduled_works(worker);
2259 } else {
2260 move_linked_works(work, &worker->scheduled, NULL);
2261 process_scheduled_works(worker);
2262 }
2263 } while (keep_working(pool));
2264
2265 worker_set_flags(worker, WORKER_PREP);
2266sleep:
2267 /*
2268 * pool->lock is held and there's no work to process and no need to
2269 * manage, sleep. Workers are woken up only while holding
2270 * pool->lock or from local cpu, so setting the current state
2271 * before releasing pool->lock is enough to prevent losing any
2272 * event.
2273 */
2274 worker_enter_idle(worker);
2275 __set_current_state(TASK_INTERRUPTIBLE);
2276 spin_unlock_irq(&pool->lock);
2277 schedule();
2278 goto woke_up;
2279}
2280
2281/**
2282 * rescuer_thread - the rescuer thread function
2283 * @__rescuer: self
2284 *
2285 * Workqueue rescuer thread function. There's one rescuer for each
2286 * workqueue which has WQ_MEM_RECLAIM set.
2287 *
2288 * Regular work processing on a pool may block trying to create a new
2289 * worker which uses GFP_KERNEL allocation which has slight chance of
2290 * developing into deadlock if some works currently on the same queue
2291 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2292 * the problem rescuer solves.
2293 *
2294 * When such condition is possible, the pool summons rescuers of all
2295 * workqueues which have works queued on the pool and let them process
2296 * those works so that forward progress can be guaranteed.
2297 *
2298 * This should happen rarely.
2299 *
2300 * Return: 0
2301 */
2302static int rescuer_thread(void *__rescuer)
2303{
2304 struct worker *rescuer = __rescuer;
2305 struct workqueue_struct *wq = rescuer->rescue_wq;
2306 struct list_head *scheduled = &rescuer->scheduled;
2307 bool should_stop;
2308
2309 set_user_nice(current, RESCUER_NICE_LEVEL);
2310
2311 /*
2312 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2313 * doesn't participate in concurrency management.
2314 */
2315 rescuer->task->flags |= PF_WQ_WORKER;
2316repeat:
2317 set_current_state(TASK_INTERRUPTIBLE);
2318
2319 /*
2320 * By the time the rescuer is requested to stop, the workqueue
2321 * shouldn't have any work pending, but @wq->maydays may still have
2322 * pwq(s) queued. This can happen by non-rescuer workers consuming
2323 * all the work items before the rescuer got to them. Go through
2324 * @wq->maydays processing before acting on should_stop so that the
2325 * list is always empty on exit.
2326 */
2327 should_stop = kthread_should_stop();
2328
2329 /* see whether any pwq is asking for help */
2330 spin_lock_irq(&wq_mayday_lock);
2331
2332 while (!list_empty(&wq->maydays)) {
2333 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2334 struct pool_workqueue, mayday_node);
2335 struct worker_pool *pool = pwq->pool;
2336 struct work_struct *work, *n;
2337 bool first = true;
2338
2339 __set_current_state(TASK_RUNNING);
2340 list_del_init(&pwq->mayday_node);
2341
2342 spin_unlock_irq(&wq_mayday_lock);
2343
2344 worker_attach_to_pool(rescuer, pool);
2345
2346 spin_lock_irq(&pool->lock);
2347 rescuer->pool = pool;
2348
2349 /*
2350 * Slurp in all works issued via this workqueue and
2351 * process'em.
2352 */
2353 WARN_ON_ONCE(!list_empty(scheduled));
2354 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2355 if (get_work_pwq(work) == pwq) {
2356 if (first)
2357 pool->watchdog_ts = jiffies;
2358 move_linked_works(work, scheduled, &n);
2359 }
2360 first = false;
2361 }
2362
2363 if (!list_empty(scheduled)) {
2364 process_scheduled_works(rescuer);
2365
2366 /*
2367 * The above execution of rescued work items could
2368 * have created more to rescue through
2369 * pwq_activate_first_delayed() or chained
2370 * queueing. Let's put @pwq back on mayday list so
2371 * that such back-to-back work items, which may be
2372 * being used to relieve memory pressure, don't
2373 * incur MAYDAY_INTERVAL delay inbetween.
2374 */
2375 if (need_to_create_worker(pool)) {
2376 spin_lock(&wq_mayday_lock);
2377 get_pwq(pwq);
2378 list_move_tail(&pwq->mayday_node, &wq->maydays);
2379 spin_unlock(&wq_mayday_lock);
2380 }
2381 }
2382
2383 /*
2384 * Put the reference grabbed by send_mayday(). @pool won't
2385 * go away while we're still attached to it.
2386 */
2387 put_pwq(pwq);
2388
2389 /*
2390 * Leave this pool. If need_more_worker() is %true, notify a
2391 * regular worker; otherwise, we end up with 0 concurrency
2392 * and stalling the execution.
2393 */
2394 if (need_more_worker(pool))
2395 wake_up_worker(pool);
2396
2397 rescuer->pool = NULL;
2398 spin_unlock_irq(&pool->lock);
2399
2400 worker_detach_from_pool(rescuer, pool);
2401
2402 spin_lock_irq(&wq_mayday_lock);
2403 }
2404
2405 spin_unlock_irq(&wq_mayday_lock);
2406
2407 if (should_stop) {
2408 __set_current_state(TASK_RUNNING);
2409 rescuer->task->flags &= ~PF_WQ_WORKER;
2410 return 0;
2411 }
2412
2413 /* rescuers should never participate in concurrency management */
2414 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2415 schedule();
2416 goto repeat;
2417}
2418
2419/**
2420 * check_flush_dependency - check for flush dependency sanity
2421 * @target_wq: workqueue being flushed
2422 * @target_work: work item being flushed (NULL for workqueue flushes)
2423 *
2424 * %current is trying to flush the whole @target_wq or @target_work on it.
2425 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2426 * reclaiming memory or running on a workqueue which doesn't have
2427 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2428 * a deadlock.
2429 */
2430static void check_flush_dependency(struct workqueue_struct *target_wq,
2431 struct work_struct *target_work)
2432{
2433 work_func_t target_func = target_work ? target_work->func : NULL;
2434 struct worker *worker;
2435
2436 if (target_wq->flags & WQ_MEM_RECLAIM)
2437 return;
2438
2439 worker = current_wq_worker();
2440
2441 WARN_ONCE(current->flags & PF_MEMALLOC,
2442 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2443 current->pid, current->comm, target_wq->name, target_func);
2444 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2445 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2446 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2447 worker->current_pwq->wq->name, worker->current_func,
2448 target_wq->name, target_func);
2449}
2450
2451struct wq_barrier {
2452 struct work_struct work;
2453 struct completion done;
2454 struct task_struct *task; /* purely informational */
2455};
2456
2457static void wq_barrier_func(struct work_struct *work)
2458{
2459 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2460 complete(&barr->done);
2461}
2462
2463/**
2464 * insert_wq_barrier - insert a barrier work
2465 * @pwq: pwq to insert barrier into
2466 * @barr: wq_barrier to insert
2467 * @target: target work to attach @barr to
2468 * @worker: worker currently executing @target, NULL if @target is not executing
2469 *
2470 * @barr is linked to @target such that @barr is completed only after
2471 * @target finishes execution. Please note that the ordering
2472 * guarantee is observed only with respect to @target and on the local
2473 * cpu.
2474 *
2475 * Currently, a queued barrier can't be canceled. This is because
2476 * try_to_grab_pending() can't determine whether the work to be
2477 * grabbed is at the head of the queue and thus can't clear LINKED
2478 * flag of the previous work while there must be a valid next work
2479 * after a work with LINKED flag set.
2480 *
2481 * Note that when @worker is non-NULL, @target may be modified
2482 * underneath us, so we can't reliably determine pwq from @target.
2483 *
2484 * CONTEXT:
2485 * spin_lock_irq(pool->lock).
2486 */
2487static void insert_wq_barrier(struct pool_workqueue *pwq,
2488 struct wq_barrier *barr,
2489 struct work_struct *target, struct worker *worker)
2490{
2491 struct list_head *head;
2492 unsigned int linked = 0;
2493
2494 /*
2495 * debugobject calls are safe here even with pool->lock locked
2496 * as we know for sure that this will not trigger any of the
2497 * checks and call back into the fixup functions where we
2498 * might deadlock.
2499 */
2500 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2501 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2502 init_completion(&barr->done);
2503 barr->task = current;
2504
2505 /*
2506 * If @target is currently being executed, schedule the
2507 * barrier to the worker; otherwise, put it after @target.
2508 */
2509 if (worker)
2510 head = worker->scheduled.next;
2511 else {
2512 unsigned long *bits = work_data_bits(target);
2513
2514 head = target->entry.next;
2515 /* there can already be other linked works, inherit and set */
2516 linked = *bits & WORK_STRUCT_LINKED;
2517 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2518 }
2519
2520 debug_work_activate(&barr->work);
2521 insert_work(pwq, &barr->work, head,
2522 work_color_to_flags(WORK_NO_COLOR) | linked);
2523}
2524
2525/**
2526 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2527 * @wq: workqueue being flushed
2528 * @flush_color: new flush color, < 0 for no-op
2529 * @work_color: new work color, < 0 for no-op
2530 *
2531 * Prepare pwqs for workqueue flushing.
2532 *
2533 * If @flush_color is non-negative, flush_color on all pwqs should be
2534 * -1. If no pwq has in-flight commands at the specified color, all
2535 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2536 * has in flight commands, its pwq->flush_color is set to
2537 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2538 * wakeup logic is armed and %true is returned.
2539 *
2540 * The caller should have initialized @wq->first_flusher prior to
2541 * calling this function with non-negative @flush_color. If
2542 * @flush_color is negative, no flush color update is done and %false
2543 * is returned.
2544 *
2545 * If @work_color is non-negative, all pwqs should have the same
2546 * work_color which is previous to @work_color and all will be
2547 * advanced to @work_color.
2548 *
2549 * CONTEXT:
2550 * mutex_lock(wq->mutex).
2551 *
2552 * Return:
2553 * %true if @flush_color >= 0 and there's something to flush. %false
2554 * otherwise.
2555 */
2556static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2557 int flush_color, int work_color)
2558{
2559 bool wait = false;
2560 struct pool_workqueue *pwq;
2561
2562 if (flush_color >= 0) {
2563 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2564 atomic_set(&wq->nr_pwqs_to_flush, 1);
2565 }
2566
2567 for_each_pwq(pwq, wq) {
2568 struct worker_pool *pool = pwq->pool;
2569
2570 spin_lock_irq(&pool->lock);
2571
2572 if (flush_color >= 0) {
2573 WARN_ON_ONCE(pwq->flush_color != -1);
2574
2575 if (pwq->nr_in_flight[flush_color]) {
2576 pwq->flush_color = flush_color;
2577 atomic_inc(&wq->nr_pwqs_to_flush);
2578 wait = true;
2579 }
2580 }
2581
2582 if (work_color >= 0) {
2583 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2584 pwq->work_color = work_color;
2585 }
2586
2587 spin_unlock_irq(&pool->lock);
2588 }
2589
2590 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2591 complete(&wq->first_flusher->done);
2592
2593 return wait;
2594}
2595
2596/**
2597 * flush_workqueue - ensure that any scheduled work has run to completion.
2598 * @wq: workqueue to flush
2599 *
2600 * This function sleeps until all work items which were queued on entry
2601 * have finished execution, but it is not livelocked by new incoming ones.
2602 */
2603void flush_workqueue(struct workqueue_struct *wq)
2604{
2605 struct wq_flusher this_flusher = {
2606 .list = LIST_HEAD_INIT(this_flusher.list),
2607 .flush_color = -1,
2608 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2609 };
2610 int next_color;
2611
2612 lock_map_acquire(&wq->lockdep_map);
2613 lock_map_release(&wq->lockdep_map);
2614
2615 mutex_lock(&wq->mutex);
2616
2617 /*
2618 * Start-to-wait phase
2619 */
2620 next_color = work_next_color(wq->work_color);
2621
2622 if (next_color != wq->flush_color) {
2623 /*
2624 * Color space is not full. The current work_color
2625 * becomes our flush_color and work_color is advanced
2626 * by one.
2627 */
2628 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2629 this_flusher.flush_color = wq->work_color;
2630 wq->work_color = next_color;
2631
2632 if (!wq->first_flusher) {
2633 /* no flush in progress, become the first flusher */
2634 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2635
2636 wq->first_flusher = &this_flusher;
2637
2638 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2639 wq->work_color)) {
2640 /* nothing to flush, done */
2641 wq->flush_color = next_color;
2642 wq->first_flusher = NULL;
2643 goto out_unlock;
2644 }
2645 } else {
2646 /* wait in queue */
2647 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2648 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2649 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2650 }
2651 } else {
2652 /*
2653 * Oops, color space is full, wait on overflow queue.
2654 * The next flush completion will assign us
2655 * flush_color and transfer to flusher_queue.
2656 */
2657 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2658 }
2659
2660 check_flush_dependency(wq, NULL);
2661
2662 mutex_unlock(&wq->mutex);
2663
2664 wait_for_completion(&this_flusher.done);
2665
2666 /*
2667 * Wake-up-and-cascade phase
2668 *
2669 * First flushers are responsible for cascading flushes and
2670 * handling overflow. Non-first flushers can simply return.
2671 */
2672 if (wq->first_flusher != &this_flusher)
2673 return;
2674
2675 mutex_lock(&wq->mutex);
2676
2677 /* we might have raced, check again with mutex held */
2678 if (wq->first_flusher != &this_flusher)
2679 goto out_unlock;
2680
2681 wq->first_flusher = NULL;
2682
2683 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2684 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2685
2686 while (true) {
2687 struct wq_flusher *next, *tmp;
2688
2689 /* complete all the flushers sharing the current flush color */
2690 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2691 if (next->flush_color != wq->flush_color)
2692 break;
2693 list_del_init(&next->list);
2694 complete(&next->done);
2695 }
2696
2697 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2698 wq->flush_color != work_next_color(wq->work_color));
2699
2700 /* this flush_color is finished, advance by one */
2701 wq->flush_color = work_next_color(wq->flush_color);
2702
2703 /* one color has been freed, handle overflow queue */
2704 if (!list_empty(&wq->flusher_overflow)) {
2705 /*
2706 * Assign the same color to all overflowed
2707 * flushers, advance work_color and append to
2708 * flusher_queue. This is the start-to-wait
2709 * phase for these overflowed flushers.
2710 */
2711 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2712 tmp->flush_color = wq->work_color;
2713
2714 wq->work_color = work_next_color(wq->work_color);
2715
2716 list_splice_tail_init(&wq->flusher_overflow,
2717 &wq->flusher_queue);
2718 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2719 }
2720
2721 if (list_empty(&wq->flusher_queue)) {
2722 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2723 break;
2724 }
2725
2726 /*
2727 * Need to flush more colors. Make the next flusher
2728 * the new first flusher and arm pwqs.
2729 */
2730 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2731 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2732
2733 list_del_init(&next->list);
2734 wq->first_flusher = next;
2735
2736 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2737 break;
2738
2739 /*
2740 * Meh... this color is already done, clear first
2741 * flusher and repeat cascading.
2742 */
2743 wq->first_flusher = NULL;
2744 }
2745
2746out_unlock:
2747 mutex_unlock(&wq->mutex);
2748}
2749EXPORT_SYMBOL(flush_workqueue);
2750
2751/**
2752 * drain_workqueue - drain a workqueue
2753 * @wq: workqueue to drain
2754 *
2755 * Wait until the workqueue becomes empty. While draining is in progress,
2756 * only chain queueing is allowed. IOW, only currently pending or running
2757 * work items on @wq can queue further work items on it. @wq is flushed
2758 * repeatedly until it becomes empty. The number of flushing is determined
2759 * by the depth of chaining and should be relatively short. Whine if it
2760 * takes too long.
2761 */
2762void drain_workqueue(struct workqueue_struct *wq)
2763{
2764 unsigned int flush_cnt = 0;
2765 struct pool_workqueue *pwq;
2766
2767 /*
2768 * __queue_work() needs to test whether there are drainers, is much
2769 * hotter than drain_workqueue() and already looks at @wq->flags.
2770 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2771 */
2772 mutex_lock(&wq->mutex);
2773 if (!wq->nr_drainers++)
2774 wq->flags |= __WQ_DRAINING;
2775 mutex_unlock(&wq->mutex);
2776reflush:
2777 flush_workqueue(wq);
2778
2779 mutex_lock(&wq->mutex);
2780
2781 for_each_pwq(pwq, wq) {
2782 bool drained;
2783
2784 spin_lock_irq(&pwq->pool->lock);
2785 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2786 spin_unlock_irq(&pwq->pool->lock);
2787
2788 if (drained)
2789 continue;
2790
2791 if (++flush_cnt == 10 ||
2792 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2793 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2794 wq->name, flush_cnt);
2795
2796 mutex_unlock(&wq->mutex);
2797 goto reflush;
2798 }
2799
2800 if (!--wq->nr_drainers)
2801 wq->flags &= ~__WQ_DRAINING;
2802 mutex_unlock(&wq->mutex);
2803}
2804EXPORT_SYMBOL_GPL(drain_workqueue);
2805
2806static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2807{
2808 struct worker *worker = NULL;
2809 struct worker_pool *pool;
2810 struct pool_workqueue *pwq;
2811
2812 might_sleep();
2813
2814 local_irq_disable();
2815 pool = get_work_pool(work);
2816 if (!pool) {
2817 local_irq_enable();
2818 return false;
2819 }
2820
2821 spin_lock(&pool->lock);
2822 /* see the comment in try_to_grab_pending() with the same code */
2823 pwq = get_work_pwq(work);
2824 if (pwq) {
2825 if (unlikely(pwq->pool != pool))
2826 goto already_gone;
2827 } else {
2828 worker = find_worker_executing_work(pool, work);
2829 if (!worker)
2830 goto already_gone;
2831 pwq = worker->current_pwq;
2832 }
2833
2834 check_flush_dependency(pwq->wq, work);
2835
2836 insert_wq_barrier(pwq, barr, work, worker);
2837 spin_unlock_irq(&pool->lock);
2838
2839 /*
2840 * If @max_active is 1 or rescuer is in use, flushing another work
2841 * item on the same workqueue may lead to deadlock. Make sure the
2842 * flusher is not running on the same workqueue by verifying write
2843 * access.
2844 */
2845 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2846 lock_map_acquire(&pwq->wq->lockdep_map);
2847 else
2848 lock_map_acquire_read(&pwq->wq->lockdep_map);
2849 lock_map_release(&pwq->wq->lockdep_map);
2850
2851 return true;
2852already_gone:
2853 spin_unlock_irq(&pool->lock);
2854 return false;
2855}
2856
2857/**
2858 * flush_work - wait for a work to finish executing the last queueing instance
2859 * @work: the work to flush
2860 *
2861 * Wait until @work has finished execution. @work is guaranteed to be idle
2862 * on return if it hasn't been requeued since flush started.
2863 *
2864 * Return:
2865 * %true if flush_work() waited for the work to finish execution,
2866 * %false if it was already idle.
2867 */
2868bool flush_work(struct work_struct *work)
2869{
2870 struct wq_barrier barr;
2871
2872 lock_map_acquire(&work->lockdep_map);
2873 lock_map_release(&work->lockdep_map);
2874
2875 if (start_flush_work(work, &barr)) {
2876 wait_for_completion(&barr.done);
2877 destroy_work_on_stack(&barr.work);
2878 return true;
2879 } else {
2880 return false;
2881 }
2882}
2883EXPORT_SYMBOL_GPL(flush_work);
2884
2885struct cwt_wait {
2886 wait_queue_t wait;
2887 struct work_struct *work;
2888};
2889
2890static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2891{
2892 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2893
2894 if (cwait->work != key)
2895 return 0;
2896 return autoremove_wake_function(wait, mode, sync, key);
2897}
2898
2899static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2900{
2901 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2902 unsigned long flags;
2903 int ret;
2904
2905 do {
2906 ret = try_to_grab_pending(work, is_dwork, &flags);
2907 /*
2908 * If someone else is already canceling, wait for it to
2909 * finish. flush_work() doesn't work for PREEMPT_NONE
2910 * because we may get scheduled between @work's completion
2911 * and the other canceling task resuming and clearing
2912 * CANCELING - flush_work() will return false immediately
2913 * as @work is no longer busy, try_to_grab_pending() will
2914 * return -ENOENT as @work is still being canceled and the
2915 * other canceling task won't be able to clear CANCELING as
2916 * we're hogging the CPU.
2917 *
2918 * Let's wait for completion using a waitqueue. As this
2919 * may lead to the thundering herd problem, use a custom
2920 * wake function which matches @work along with exclusive
2921 * wait and wakeup.
2922 */
2923 if (unlikely(ret == -ENOENT)) {
2924 struct cwt_wait cwait;
2925
2926 init_wait(&cwait.wait);
2927 cwait.wait.func = cwt_wakefn;
2928 cwait.work = work;
2929
2930 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2931 TASK_UNINTERRUPTIBLE);
2932 if (work_is_canceling(work))
2933 schedule();
2934 finish_wait(&cancel_waitq, &cwait.wait);
2935 }
2936 } while (unlikely(ret < 0));
2937
2938 /* tell other tasks trying to grab @work to back off */
2939 mark_work_canceling(work);
2940 local_irq_restore(flags);
2941
2942 flush_work(work);
2943 clear_work_data(work);
2944
2945 /*
2946 * Paired with prepare_to_wait() above so that either
2947 * waitqueue_active() is visible here or !work_is_canceling() is
2948 * visible there.
2949 */
2950 smp_mb();
2951 if (waitqueue_active(&cancel_waitq))
2952 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2953
2954 return ret;
2955}
2956
2957/**
2958 * cancel_work_sync - cancel a work and wait for it to finish
2959 * @work: the work to cancel
2960 *
2961 * Cancel @work and wait for its execution to finish. This function
2962 * can be used even if the work re-queues itself or migrates to
2963 * another workqueue. On return from this function, @work is
2964 * guaranteed to be not pending or executing on any CPU.
2965 *
2966 * cancel_work_sync(&delayed_work->work) must not be used for
2967 * delayed_work's. Use cancel_delayed_work_sync() instead.
2968 *
2969 * The caller must ensure that the workqueue on which @work was last
2970 * queued can't be destroyed before this function returns.
2971 *
2972 * Return:
2973 * %true if @work was pending, %false otherwise.
2974 */
2975bool cancel_work_sync(struct work_struct *work)
2976{
2977 return __cancel_work_timer(work, false);
2978}
2979EXPORT_SYMBOL_GPL(cancel_work_sync);
2980
2981/**
2982 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2983 * @dwork: the delayed work to flush
2984 *
2985 * Delayed timer is cancelled and the pending work is queued for
2986 * immediate execution. Like flush_work(), this function only
2987 * considers the last queueing instance of @dwork.
2988 *
2989 * Return:
2990 * %true if flush_work() waited for the work to finish execution,
2991 * %false if it was already idle.
2992 */
2993bool flush_delayed_work(struct delayed_work *dwork)
2994{
2995 local_irq_disable();
2996 if (del_timer_sync(&dwork->timer))
2997 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2998 local_irq_enable();
2999 return flush_work(&dwork->work);
3000}
3001EXPORT_SYMBOL(flush_delayed_work);
3002
3003/**
3004 * cancel_delayed_work - cancel a delayed work
3005 * @dwork: delayed_work to cancel
3006 *
3007 * Kill off a pending delayed_work.
3008 *
3009 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3010 * pending.
3011 *
3012 * Note:
3013 * The work callback function may still be running on return, unless
3014 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3015 * use cancel_delayed_work_sync() to wait on it.
3016 *
3017 * This function is safe to call from any context including IRQ handler.
3018 */
3019bool cancel_delayed_work(struct delayed_work *dwork)
3020{
3021 unsigned long flags;
3022 int ret;
3023
3024 do {
3025 ret = try_to_grab_pending(&dwork->work, true, &flags);
3026 } while (unlikely(ret == -EAGAIN));
3027
3028 if (unlikely(ret < 0))
3029 return false;
3030
3031 set_work_pool_and_clear_pending(&dwork->work,
3032 get_work_pool_id(&dwork->work));
3033 local_irq_restore(flags);
3034 return ret;
3035}
3036EXPORT_SYMBOL(cancel_delayed_work);
3037
3038/**
3039 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3040 * @dwork: the delayed work cancel
3041 *
3042 * This is cancel_work_sync() for delayed works.
3043 *
3044 * Return:
3045 * %true if @dwork was pending, %false otherwise.
3046 */
3047bool cancel_delayed_work_sync(struct delayed_work *dwork)
3048{
3049 return __cancel_work_timer(&dwork->work, true);
3050}
3051EXPORT_SYMBOL(cancel_delayed_work_sync);
3052
3053/**
3054 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3055 * @func: the function to call
3056 *
3057 * schedule_on_each_cpu() executes @func on each online CPU using the
3058 * system workqueue and blocks until all CPUs have completed.
3059 * schedule_on_each_cpu() is very slow.
3060 *
3061 * Return:
3062 * 0 on success, -errno on failure.
3063 */
3064int schedule_on_each_cpu(work_func_t func)
3065{
3066 int cpu;
3067 struct work_struct __percpu *works;
3068
3069 works = alloc_percpu(struct work_struct);
3070 if (!works)
3071 return -ENOMEM;
3072
3073 get_online_cpus();
3074
3075 for_each_online_cpu(cpu) {
3076 struct work_struct *work = per_cpu_ptr(works, cpu);
3077
3078 INIT_WORK(work, func);
3079 schedule_work_on(cpu, work);
3080 }
3081
3082 for_each_online_cpu(cpu)
3083 flush_work(per_cpu_ptr(works, cpu));
3084
3085 put_online_cpus();
3086 free_percpu(works);
3087 return 0;
3088}
3089
3090/**
3091 * execute_in_process_context - reliably execute the routine with user context
3092 * @fn: the function to execute
3093 * @ew: guaranteed storage for the execute work structure (must
3094 * be available when the work executes)
3095 *
3096 * Executes the function immediately if process context is available,
3097 * otherwise schedules the function for delayed execution.
3098 *
3099 * Return: 0 - function was executed
3100 * 1 - function was scheduled for execution
3101 */
3102int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3103{
3104 if (!in_interrupt()) {
3105 fn(&ew->work);
3106 return 0;
3107 }
3108
3109 INIT_WORK(&ew->work, fn);
3110 schedule_work(&ew->work);
3111
3112 return 1;
3113}
3114EXPORT_SYMBOL_GPL(execute_in_process_context);
3115
3116/**
3117 * free_workqueue_attrs - free a workqueue_attrs
3118 * @attrs: workqueue_attrs to free
3119 *
3120 * Undo alloc_workqueue_attrs().
3121 */
3122void free_workqueue_attrs(struct workqueue_attrs *attrs)
3123{
3124 if (attrs) {
3125 free_cpumask_var(attrs->cpumask);
3126 kfree(attrs);
3127 }
3128}
3129
3130/**
3131 * alloc_workqueue_attrs - allocate a workqueue_attrs
3132 * @gfp_mask: allocation mask to use
3133 *
3134 * Allocate a new workqueue_attrs, initialize with default settings and
3135 * return it.
3136 *
3137 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3138 */
3139struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3140{
3141 struct workqueue_attrs *attrs;
3142
3143 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3144 if (!attrs)
3145 goto fail;
3146 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3147 goto fail;
3148
3149 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3150 return attrs;
3151fail:
3152 free_workqueue_attrs(attrs);
3153 return NULL;
3154}
3155
3156static void copy_workqueue_attrs(struct workqueue_attrs *to,
3157 const struct workqueue_attrs *from)
3158{
3159 to->nice = from->nice;
3160 cpumask_copy(to->cpumask, from->cpumask);
3161 /*
3162 * Unlike hash and equality test, this function doesn't ignore
3163 * ->no_numa as it is used for both pool and wq attrs. Instead,
3164 * get_unbound_pool() explicitly clears ->no_numa after copying.
3165 */
3166 to->no_numa = from->no_numa;
3167}
3168
3169/* hash value of the content of @attr */
3170static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3171{
3172 u32 hash = 0;
3173
3174 hash = jhash_1word(attrs->nice, hash);
3175 hash = jhash(cpumask_bits(attrs->cpumask),
3176 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3177 return hash;
3178}
3179
3180/* content equality test */
3181static bool wqattrs_equal(const struct workqueue_attrs *a,
3182 const struct workqueue_attrs *b)
3183{
3184 if (a->nice != b->nice)
3185 return false;
3186 if (!cpumask_equal(a->cpumask, b->cpumask))
3187 return false;
3188 return true;
3189}
3190
3191/**
3192 * init_worker_pool - initialize a newly zalloc'd worker_pool
3193 * @pool: worker_pool to initialize
3194 *
3195 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3196 *
3197 * Return: 0 on success, -errno on failure. Even on failure, all fields
3198 * inside @pool proper are initialized and put_unbound_pool() can be called
3199 * on @pool safely to release it.
3200 */
3201static int init_worker_pool(struct worker_pool *pool)
3202{
3203 spin_lock_init(&pool->lock);
3204 pool->id = -1;
3205 pool->cpu = -1;
3206 pool->node = NUMA_NO_NODE;
3207 pool->flags |= POOL_DISASSOCIATED;
3208 pool->watchdog_ts = jiffies;
3209 INIT_LIST_HEAD(&pool->worklist);
3210 INIT_LIST_HEAD(&pool->idle_list);
3211 hash_init(pool->busy_hash);
3212
3213 init_timer_deferrable(&pool->idle_timer);
3214 pool->idle_timer.function = idle_worker_timeout;
3215 pool->idle_timer.data = (unsigned long)pool;
3216
3217 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3218 (unsigned long)pool);
3219
3220 mutex_init(&pool->manager_arb);
3221 mutex_init(&pool->attach_mutex);
3222 INIT_LIST_HEAD(&pool->workers);
3223
3224 ida_init(&pool->worker_ida);
3225 INIT_HLIST_NODE(&pool->hash_node);
3226 pool->refcnt = 1;
3227
3228 /* shouldn't fail above this point */
3229 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3230 if (!pool->attrs)
3231 return -ENOMEM;
3232 return 0;
3233}
3234
3235static void rcu_free_wq(struct rcu_head *rcu)
3236{
3237 struct workqueue_struct *wq =
3238 container_of(rcu, struct workqueue_struct, rcu);
3239
3240 if (!(wq->flags & WQ_UNBOUND))
3241 free_percpu(wq->cpu_pwqs);
3242 else
3243 free_workqueue_attrs(wq->unbound_attrs);
3244
3245 kfree(wq->rescuer);
3246 kfree(wq);
3247}
3248
3249static void rcu_free_pool(struct rcu_head *rcu)
3250{
3251 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3252
3253 ida_destroy(&pool->worker_ida);
3254 free_workqueue_attrs(pool->attrs);
3255 kfree(pool);
3256}
3257
3258/**
3259 * put_unbound_pool - put a worker_pool
3260 * @pool: worker_pool to put
3261 *
3262 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3263 * safe manner. get_unbound_pool() calls this function on its failure path
3264 * and this function should be able to release pools which went through,
3265 * successfully or not, init_worker_pool().
3266 *
3267 * Should be called with wq_pool_mutex held.
3268 */
3269static void put_unbound_pool(struct worker_pool *pool)
3270{
3271 DECLARE_COMPLETION_ONSTACK(detach_completion);
3272 struct worker *worker;
3273
3274 lockdep_assert_held(&wq_pool_mutex);
3275
3276 if (--pool->refcnt)
3277 return;
3278
3279 /* sanity checks */
3280 if (WARN_ON(!(pool->cpu < 0)) ||
3281 WARN_ON(!list_empty(&pool->worklist)))
3282 return;
3283
3284 /* release id and unhash */
3285 if (pool->id >= 0)
3286 idr_remove(&worker_pool_idr, pool->id);
3287 hash_del(&pool->hash_node);
3288
3289 /*
3290 * Become the manager and destroy all workers. Grabbing
3291 * manager_arb prevents @pool's workers from blocking on
3292 * attach_mutex.
3293 */
3294 mutex_lock(&pool->manager_arb);
3295
3296 spin_lock_irq(&pool->lock);
3297 while ((worker = first_idle_worker(pool)))
3298 destroy_worker(worker);
3299 WARN_ON(pool->nr_workers || pool->nr_idle);
3300 spin_unlock_irq(&pool->lock);
3301
3302 mutex_lock(&pool->attach_mutex);
3303 if (!list_empty(&pool->workers))
3304 pool->detach_completion = &detach_completion;
3305 mutex_unlock(&pool->attach_mutex);
3306
3307 if (pool->detach_completion)
3308 wait_for_completion(pool->detach_completion);
3309
3310 mutex_unlock(&pool->manager_arb);
3311
3312 /* shut down the timers */
3313 del_timer_sync(&pool->idle_timer);
3314 del_timer_sync(&pool->mayday_timer);
3315
3316 /* sched-RCU protected to allow dereferences from get_work_pool() */
3317 call_rcu_sched(&pool->rcu, rcu_free_pool);
3318}
3319
3320/**
3321 * get_unbound_pool - get a worker_pool with the specified attributes
3322 * @attrs: the attributes of the worker_pool to get
3323 *
3324 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3325 * reference count and return it. If there already is a matching
3326 * worker_pool, it will be used; otherwise, this function attempts to
3327 * create a new one.
3328 *
3329 * Should be called with wq_pool_mutex held.
3330 *
3331 * Return: On success, a worker_pool with the same attributes as @attrs.
3332 * On failure, %NULL.
3333 */
3334static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3335{
3336 u32 hash = wqattrs_hash(attrs);
3337 struct worker_pool *pool;
3338 int node;
3339 int target_node = NUMA_NO_NODE;
3340
3341 lockdep_assert_held(&wq_pool_mutex);
3342
3343 /* do we already have a matching pool? */
3344 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3345 if (wqattrs_equal(pool->attrs, attrs)) {
3346 pool->refcnt++;
3347 return pool;
3348 }
3349 }
3350
3351 /* if cpumask is contained inside a NUMA node, we belong to that node */
3352 if (wq_numa_enabled) {
3353 for_each_node(node) {
3354 if (cpumask_subset(attrs->cpumask,
3355 wq_numa_possible_cpumask[node])) {
3356 target_node = node;
3357 break;
3358 }
3359 }
3360 }
3361
3362 /* nope, create a new one */
3363 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3364 if (!pool || init_worker_pool(pool) < 0)
3365 goto fail;
3366
3367 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3368 copy_workqueue_attrs(pool->attrs, attrs);
3369 pool->node = target_node;
3370
3371 /*
3372 * no_numa isn't a worker_pool attribute, always clear it. See
3373 * 'struct workqueue_attrs' comments for detail.
3374 */
3375 pool->attrs->no_numa = false;
3376
3377 if (worker_pool_assign_id(pool) < 0)
3378 goto fail;
3379
3380 /* create and start the initial worker */
3381 if (!create_worker(pool))
3382 goto fail;
3383
3384 /* install */
3385 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3386
3387 return pool;
3388fail:
3389 if (pool)
3390 put_unbound_pool(pool);
3391 return NULL;
3392}
3393
3394static void rcu_free_pwq(struct rcu_head *rcu)
3395{
3396 kmem_cache_free(pwq_cache,
3397 container_of(rcu, struct pool_workqueue, rcu));
3398}
3399
3400/*
3401 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3402 * and needs to be destroyed.
3403 */
3404static void pwq_unbound_release_workfn(struct work_struct *work)
3405{
3406 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3407 unbound_release_work);
3408 struct workqueue_struct *wq = pwq->wq;
3409 struct worker_pool *pool = pwq->pool;
3410 bool is_last;
3411
3412 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3413 return;
3414
3415 mutex_lock(&wq->mutex);
3416 list_del_rcu(&pwq->pwqs_node);
3417 is_last = list_empty(&wq->pwqs);
3418 mutex_unlock(&wq->mutex);
3419
3420 mutex_lock(&wq_pool_mutex);
3421 put_unbound_pool(pool);
3422 mutex_unlock(&wq_pool_mutex);
3423
3424 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3425
3426 /*
3427 * If we're the last pwq going away, @wq is already dead and no one
3428 * is gonna access it anymore. Schedule RCU free.
3429 */
3430 if (is_last)
3431 call_rcu_sched(&wq->rcu, rcu_free_wq);
3432}
3433
3434/**
3435 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3436 * @pwq: target pool_workqueue
3437 *
3438 * If @pwq isn't freezing, set @pwq->max_active to the associated
3439 * workqueue's saved_max_active and activate delayed work items
3440 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3441 */
3442static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3443{
3444 struct workqueue_struct *wq = pwq->wq;
3445 bool freezable = wq->flags & WQ_FREEZABLE;
3446
3447 /* for @wq->saved_max_active */
3448 lockdep_assert_held(&wq->mutex);
3449
3450 /* fast exit for non-freezable wqs */
3451 if (!freezable && pwq->max_active == wq->saved_max_active)
3452 return;
3453
3454 spin_lock_irq(&pwq->pool->lock);
3455
3456 /*
3457 * During [un]freezing, the caller is responsible for ensuring that
3458 * this function is called at least once after @workqueue_freezing
3459 * is updated and visible.
3460 */
3461 if (!freezable || !workqueue_freezing) {
3462 pwq->max_active = wq->saved_max_active;
3463
3464 while (!list_empty(&pwq->delayed_works) &&
3465 pwq->nr_active < pwq->max_active)
3466 pwq_activate_first_delayed(pwq);
3467
3468 /*
3469 * Need to kick a worker after thawed or an unbound wq's
3470 * max_active is bumped. It's a slow path. Do it always.
3471 */
3472 wake_up_worker(pwq->pool);
3473 } else {
3474 pwq->max_active = 0;
3475 }
3476
3477 spin_unlock_irq(&pwq->pool->lock);
3478}
3479
3480/* initialize newly alloced @pwq which is associated with @wq and @pool */
3481static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3482 struct worker_pool *pool)
3483{
3484 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
3486 memset(pwq, 0, sizeof(*pwq));
3487
3488 pwq->pool = pool;
3489 pwq->wq = wq;
3490 pwq->flush_color = -1;
3491 pwq->refcnt = 1;
3492 INIT_LIST_HEAD(&pwq->delayed_works);
3493 INIT_LIST_HEAD(&pwq->pwqs_node);
3494 INIT_LIST_HEAD(&pwq->mayday_node);
3495 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3496}
3497
3498/* sync @pwq with the current state of its associated wq and link it */
3499static void link_pwq(struct pool_workqueue *pwq)
3500{
3501 struct workqueue_struct *wq = pwq->wq;
3502
3503 lockdep_assert_held(&wq->mutex);
3504
3505 /* may be called multiple times, ignore if already linked */
3506 if (!list_empty(&pwq->pwqs_node))
3507 return;
3508
3509 /* set the matching work_color */
3510 pwq->work_color = wq->work_color;
3511
3512 /* sync max_active to the current setting */
3513 pwq_adjust_max_active(pwq);
3514
3515 /* link in @pwq */
3516 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3517}
3518
3519/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3521 const struct workqueue_attrs *attrs)
3522{
3523 struct worker_pool *pool;
3524 struct pool_workqueue *pwq;
3525
3526 lockdep_assert_held(&wq_pool_mutex);
3527
3528 pool = get_unbound_pool(attrs);
3529 if (!pool)
3530 return NULL;
3531
3532 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3533 if (!pwq) {
3534 put_unbound_pool(pool);
3535 return NULL;
3536 }
3537
3538 init_pwq(pwq, wq, pool);
3539 return pwq;
3540}
3541
3542/**
3543 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3544 * @attrs: the wq_attrs of the default pwq of the target workqueue
3545 * @node: the target NUMA node
3546 * @cpu_going_down: if >= 0, the CPU to consider as offline
3547 * @cpumask: outarg, the resulting cpumask
3548 *
3549 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3550 * @cpu_going_down is >= 0, that cpu is considered offline during
3551 * calculation. The result is stored in @cpumask.
3552 *
3553 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3554 * enabled and @node has online CPUs requested by @attrs, the returned
3555 * cpumask is the intersection of the possible CPUs of @node and
3556 * @attrs->cpumask.
3557 *
3558 * The caller is responsible for ensuring that the cpumask of @node stays
3559 * stable.
3560 *
3561 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562 * %false if equal.
3563 */
3564static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565 int cpu_going_down, cpumask_t *cpumask)
3566{
3567 if (!wq_numa_enabled || attrs->no_numa)
3568 goto use_dfl;
3569
3570 /* does @node have any online CPUs @attrs wants? */
3571 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572 if (cpu_going_down >= 0)
3573 cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575 if (cpumask_empty(cpumask))
3576 goto use_dfl;
3577
3578 /* yeap, return possible CPUs in @node that @attrs wants */
3579 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580 return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582use_dfl:
3583 cpumask_copy(cpumask, attrs->cpumask);
3584 return false;
3585}
3586
3587/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589 int node,
3590 struct pool_workqueue *pwq)
3591{
3592 struct pool_workqueue *old_pwq;
3593
3594 lockdep_assert_held(&wq_pool_mutex);
3595 lockdep_assert_held(&wq->mutex);
3596
3597 /* link_pwq() can handle duplicate calls */
3598 link_pwq(pwq);
3599
3600 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602 return old_pwq;
3603}
3604
3605/* context to store the prepared attrs & pwqs before applying */
3606struct apply_wqattrs_ctx {
3607 struct workqueue_struct *wq; /* target workqueue */
3608 struct workqueue_attrs *attrs; /* attrs to apply */
3609 struct list_head list; /* queued for batching commit */
3610 struct pool_workqueue *dfl_pwq;
3611 struct pool_workqueue *pwq_tbl[];
3612};
3613
3614/* free the resources after success or abort */
3615static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616{
3617 if (ctx) {
3618 int node;
3619
3620 for_each_node(node)
3621 put_pwq_unlocked(ctx->pwq_tbl[node]);
3622 put_pwq_unlocked(ctx->dfl_pwq);
3623
3624 free_workqueue_attrs(ctx->attrs);
3625
3626 kfree(ctx);
3627 }
3628}
3629
3630/* allocate the attrs and pwqs for later installation */
3631static struct apply_wqattrs_ctx *
3632apply_wqattrs_prepare(struct workqueue_struct *wq,
3633 const struct workqueue_attrs *attrs)
3634{
3635 struct apply_wqattrs_ctx *ctx;
3636 struct workqueue_attrs *new_attrs, *tmp_attrs;
3637 int node;
3638
3639 lockdep_assert_held(&wq_pool_mutex);
3640
3641 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642 GFP_KERNEL);
3643
3644 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646 if (!ctx || !new_attrs || !tmp_attrs)
3647 goto out_free;
3648
3649 /*
3650 * Calculate the attrs of the default pwq.
3651 * If the user configured cpumask doesn't overlap with the
3652 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653 */
3654 copy_workqueue_attrs(new_attrs, attrs);
3655 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658
3659 /*
3660 * We may create multiple pwqs with differing cpumasks. Make a
3661 * copy of @new_attrs which will be modified and used to obtain
3662 * pools.
3663 */
3664 copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666 /*
3667 * If something goes wrong during CPU up/down, we'll fall back to
3668 * the default pwq covering whole @attrs->cpumask. Always create
3669 * it even if we don't use it immediately.
3670 */
3671 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672 if (!ctx->dfl_pwq)
3673 goto out_free;
3674
3675 for_each_node(node) {
3676 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678 if (!ctx->pwq_tbl[node])
3679 goto out_free;
3680 } else {
3681 ctx->dfl_pwq->refcnt++;
3682 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683 }
3684 }
3685
3686 /* save the user configured attrs and sanitize it. */
3687 copy_workqueue_attrs(new_attrs, attrs);
3688 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689 ctx->attrs = new_attrs;
3690
3691 ctx->wq = wq;
3692 free_workqueue_attrs(tmp_attrs);
3693 return ctx;
3694
3695out_free:
3696 free_workqueue_attrs(tmp_attrs);
3697 free_workqueue_attrs(new_attrs);
3698 apply_wqattrs_cleanup(ctx);
3699 return NULL;
3700}
3701
3702/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704{
3705 int node;
3706
3707 /* all pwqs have been created successfully, let's install'em */
3708 mutex_lock(&ctx->wq->mutex);
3709
3710 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712 /* save the previous pwq and install the new one */
3713 for_each_node(node)
3714 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715 ctx->pwq_tbl[node]);
3716
3717 /* @dfl_pwq might not have been used, ensure it's linked */
3718 link_pwq(ctx->dfl_pwq);
3719 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721 mutex_unlock(&ctx->wq->mutex);
3722}
3723
3724static void apply_wqattrs_lock(void)
3725{
3726 /* CPUs should stay stable across pwq creations and installations */
3727 get_online_cpus();
3728 mutex_lock(&wq_pool_mutex);
3729}
3730
3731static void apply_wqattrs_unlock(void)
3732{
3733 mutex_unlock(&wq_pool_mutex);
3734 put_online_cpus();
3735}
3736
3737static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738 const struct workqueue_attrs *attrs)
3739{
3740 struct apply_wqattrs_ctx *ctx;
3741
3742 /* only unbound workqueues can change attributes */
3743 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744 return -EINVAL;
3745
3746 /* creating multiple pwqs breaks ordering guarantee */
3747 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748 return -EINVAL;
3749
3750 ctx = apply_wqattrs_prepare(wq, attrs);
3751 if (!ctx)
3752 return -ENOMEM;
3753
3754 /* the ctx has been prepared successfully, let's commit it */
3755 apply_wqattrs_commit(ctx);
3756 apply_wqattrs_cleanup(ctx);
3757
3758 return 0;
3759}
3760
3761/**
3762 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763 * @wq: the target workqueue
3764 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765 *
3766 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3767 * machines, this function maps a separate pwq to each NUMA node with
3768 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3769 * NUMA node it was issued on. Older pwqs are released as in-flight work
3770 * items finish. Note that a work item which repeatedly requeues itself
3771 * back-to-back will stay on its current pwq.
3772 *
3773 * Performs GFP_KERNEL allocations.
3774 *
3775 * Return: 0 on success and -errno on failure.
3776 */
3777int apply_workqueue_attrs(struct workqueue_struct *wq,
3778 const struct workqueue_attrs *attrs)
3779{
3780 int ret;
3781
3782 apply_wqattrs_lock();
3783 ret = apply_workqueue_attrs_locked(wq, attrs);
3784 apply_wqattrs_unlock();
3785
3786 return ret;
3787}
3788
3789/**
3790 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791 * @wq: the target workqueue
3792 * @cpu: the CPU coming up or going down
3793 * @online: whether @cpu is coming up or going down
3794 *
3795 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3797 * @wq accordingly.
3798 *
3799 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3800 * falls back to @wq->dfl_pwq which may not be optimal but is always
3801 * correct.
3802 *
3803 * Note that when the last allowed CPU of a NUMA node goes offline for a
3804 * workqueue with a cpumask spanning multiple nodes, the workers which were
3805 * already executing the work items for the workqueue will lose their CPU
3806 * affinity and may execute on any CPU. This is similar to how per-cpu
3807 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3808 * affinity, it's the user's responsibility to flush the work item from
3809 * CPU_DOWN_PREPARE.
3810 */
3811static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812 bool online)
3813{
3814 int node = cpu_to_node(cpu);
3815 int cpu_off = online ? -1 : cpu;
3816 struct pool_workqueue *old_pwq = NULL, *pwq;
3817 struct workqueue_attrs *target_attrs;
3818 cpumask_t *cpumask;
3819
3820 lockdep_assert_held(&wq_pool_mutex);
3821
3822 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3823 wq->unbound_attrs->no_numa)
3824 return;
3825
3826 /*
3827 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828 * Let's use a preallocated one. The following buf is protected by
3829 * CPU hotplug exclusion.
3830 */
3831 target_attrs = wq_update_unbound_numa_attrs_buf;
3832 cpumask = target_attrs->cpumask;
3833
3834 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835 pwq = unbound_pwq_by_node(wq, node);
3836
3837 /*
3838 * Let's determine what needs to be done. If the target cpumask is
3839 * different from the default pwq's, we need to compare it to @pwq's
3840 * and create a new one if they don't match. If the target cpumask
3841 * equals the default pwq's, the default pwq should be used.
3842 */
3843 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3844 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3845 return;
3846 } else {
3847 goto use_dfl_pwq;
3848 }
3849
3850 /* create a new pwq */
3851 pwq = alloc_unbound_pwq(wq, target_attrs);
3852 if (!pwq) {
3853 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854 wq->name);
3855 goto use_dfl_pwq;
3856 }
3857
3858 /* Install the new pwq. */
3859 mutex_lock(&wq->mutex);
3860 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861 goto out_unlock;
3862
3863use_dfl_pwq:
3864 mutex_lock(&wq->mutex);
3865 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866 get_pwq(wq->dfl_pwq);
3867 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869out_unlock:
3870 mutex_unlock(&wq->mutex);
3871 put_pwq_unlocked(old_pwq);
3872}
3873
3874static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3875{
3876 bool highpri = wq->flags & WQ_HIGHPRI;
3877 int cpu, ret;
3878
3879 if (!(wq->flags & WQ_UNBOUND)) {
3880 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881 if (!wq->cpu_pwqs)
3882 return -ENOMEM;
3883
3884 for_each_possible_cpu(cpu) {
3885 struct pool_workqueue *pwq =
3886 per_cpu_ptr(wq->cpu_pwqs, cpu);
3887 struct worker_pool *cpu_pools =
3888 per_cpu(cpu_worker_pools, cpu);
3889
3890 init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892 mutex_lock(&wq->mutex);
3893 link_pwq(pwq);
3894 mutex_unlock(&wq->mutex);
3895 }
3896 return 0;
3897 } else if (wq->flags & __WQ_ORDERED) {
3898 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899 /* there should only be single pwq for ordering guarantee */
3900 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3901 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902 "ordering guarantee broken for workqueue %s\n", wq->name);
3903 return ret;
3904 } else {
3905 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3906 }
3907}
3908
3909static int wq_clamp_max_active(int max_active, unsigned int flags,
3910 const char *name)
3911{
3912 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914 if (max_active < 1 || max_active > lim)
3915 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916 max_active, name, 1, lim);
3917
3918 return clamp_val(max_active, 1, lim);
3919}
3920
3921struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3922 unsigned int flags,
3923 int max_active,
3924 struct lock_class_key *key,
3925 const char *lock_name, ...)
3926{
3927 size_t tbl_size = 0;
3928 va_list args;
3929 struct workqueue_struct *wq;
3930 struct pool_workqueue *pwq;
3931
3932 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3933 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934 flags |= WQ_UNBOUND;
3935
3936 /* allocate wq and format name */
3937 if (flags & WQ_UNBOUND)
3938 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3939
3940 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3941 if (!wq)
3942 return NULL;
3943
3944 if (flags & WQ_UNBOUND) {
3945 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946 if (!wq->unbound_attrs)
3947 goto err_free_wq;
3948 }
3949
3950 va_start(args, lock_name);
3951 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3952 va_end(args);
3953
3954 max_active = max_active ?: WQ_DFL_ACTIVE;
3955 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3956
3957 /* init wq */
3958 wq->flags = flags;
3959 wq->saved_max_active = max_active;
3960 mutex_init(&wq->mutex);
3961 atomic_set(&wq->nr_pwqs_to_flush, 0);
3962 INIT_LIST_HEAD(&wq->pwqs);
3963 INIT_LIST_HEAD(&wq->flusher_queue);
3964 INIT_LIST_HEAD(&wq->flusher_overflow);
3965 INIT_LIST_HEAD(&wq->maydays);
3966
3967 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3968 INIT_LIST_HEAD(&wq->list);
3969
3970 if (alloc_and_link_pwqs(wq) < 0)
3971 goto err_free_wq;
3972
3973 /*
3974 * Workqueues which may be used during memory reclaim should
3975 * have a rescuer to guarantee forward progress.
3976 */
3977 if (flags & WQ_MEM_RECLAIM) {
3978 struct worker *rescuer;
3979
3980 rescuer = alloc_worker(NUMA_NO_NODE);
3981 if (!rescuer)
3982 goto err_destroy;
3983
3984 rescuer->rescue_wq = wq;
3985 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3986 wq->name);
3987 if (IS_ERR(rescuer->task)) {
3988 kfree(rescuer);
3989 goto err_destroy;
3990 }
3991
3992 wq->rescuer = rescuer;
3993 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3994 wake_up_process(rescuer->task);
3995 }
3996
3997 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998 goto err_destroy;
3999
4000 /*
4001 * wq_pool_mutex protects global freeze state and workqueues list.
4002 * Grab it, adjust max_active and add the new @wq to workqueues
4003 * list.
4004 */
4005 mutex_lock(&wq_pool_mutex);
4006
4007 mutex_lock(&wq->mutex);
4008 for_each_pwq(pwq, wq)
4009 pwq_adjust_max_active(pwq);
4010 mutex_unlock(&wq->mutex);
4011
4012 list_add_tail_rcu(&wq->list, &workqueues);
4013
4014 mutex_unlock(&wq_pool_mutex);
4015
4016 return wq;
4017
4018err_free_wq:
4019 free_workqueue_attrs(wq->unbound_attrs);
4020 kfree(wq);
4021 return NULL;
4022err_destroy:
4023 destroy_workqueue(wq);
4024 return NULL;
4025}
4026EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4027
4028/**
4029 * destroy_workqueue - safely terminate a workqueue
4030 * @wq: target workqueue
4031 *
4032 * Safely destroy a workqueue. All work currently pending will be done first.
4033 */
4034void destroy_workqueue(struct workqueue_struct *wq)
4035{
4036 struct pool_workqueue *pwq;
4037 int node;
4038
4039 /* drain it before proceeding with destruction */
4040 drain_workqueue(wq);
4041
4042 /* sanity checks */
4043 mutex_lock(&wq->mutex);
4044 for_each_pwq(pwq, wq) {
4045 int i;
4046
4047 for (i = 0; i < WORK_NR_COLORS; i++) {
4048 if (WARN_ON(pwq->nr_in_flight[i])) {
4049 mutex_unlock(&wq->mutex);
4050 return;
4051 }
4052 }
4053
4054 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4055 WARN_ON(pwq->nr_active) ||
4056 WARN_ON(!list_empty(&pwq->delayed_works))) {
4057 mutex_unlock(&wq->mutex);
4058 return;
4059 }
4060 }
4061 mutex_unlock(&wq->mutex);
4062
4063 /*
4064 * wq list is used to freeze wq, remove from list after
4065 * flushing is complete in case freeze races us.
4066 */
4067 mutex_lock(&wq_pool_mutex);
4068 list_del_rcu(&wq->list);
4069 mutex_unlock(&wq_pool_mutex);
4070
4071 workqueue_sysfs_unregister(wq);
4072
4073 if (wq->rescuer)
4074 kthread_stop(wq->rescuer->task);
4075
4076 if (!(wq->flags & WQ_UNBOUND)) {
4077 /*
4078 * The base ref is never dropped on per-cpu pwqs. Directly
4079 * schedule RCU free.
4080 */
4081 call_rcu_sched(&wq->rcu, rcu_free_wq);
4082 } else {
4083 /*
4084 * We're the sole accessor of @wq at this point. Directly
4085 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4086 * @wq will be freed when the last pwq is released.
4087 */
4088 for_each_node(node) {
4089 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4090 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4091 put_pwq_unlocked(pwq);
4092 }
4093
4094 /*
4095 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4096 * put. Don't access it afterwards.
4097 */
4098 pwq = wq->dfl_pwq;
4099 wq->dfl_pwq = NULL;
4100 put_pwq_unlocked(pwq);
4101 }
4102}
4103EXPORT_SYMBOL_GPL(destroy_workqueue);
4104
4105/**
4106 * workqueue_set_max_active - adjust max_active of a workqueue
4107 * @wq: target workqueue
4108 * @max_active: new max_active value.
4109 *
4110 * Set max_active of @wq to @max_active.
4111 *
4112 * CONTEXT:
4113 * Don't call from IRQ context.
4114 */
4115void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4116{
4117 struct pool_workqueue *pwq;
4118
4119 /* disallow meddling with max_active for ordered workqueues */
4120 if (WARN_ON(wq->flags & __WQ_ORDERED))
4121 return;
4122
4123 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4124
4125 mutex_lock(&wq->mutex);
4126
4127 wq->saved_max_active = max_active;
4128
4129 for_each_pwq(pwq, wq)
4130 pwq_adjust_max_active(pwq);
4131
4132 mutex_unlock(&wq->mutex);
4133}
4134EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4135
4136/**
4137 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4138 *
4139 * Determine whether %current is a workqueue rescuer. Can be used from
4140 * work functions to determine whether it's being run off the rescuer task.
4141 *
4142 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4143 */
4144bool current_is_workqueue_rescuer(void)
4145{
4146 struct worker *worker = current_wq_worker();
4147
4148 return worker && worker->rescue_wq;
4149}
4150
4151/**
4152 * workqueue_congested - test whether a workqueue is congested
4153 * @cpu: CPU in question
4154 * @wq: target workqueue
4155 *
4156 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4157 * no synchronization around this function and the test result is
4158 * unreliable and only useful as advisory hints or for debugging.
4159 *
4160 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4161 * Note that both per-cpu and unbound workqueues may be associated with
4162 * multiple pool_workqueues which have separate congested states. A
4163 * workqueue being congested on one CPU doesn't mean the workqueue is also
4164 * contested on other CPUs / NUMA nodes.
4165 *
4166 * Return:
4167 * %true if congested, %false otherwise.
4168 */
4169bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4170{
4171 struct pool_workqueue *pwq;
4172 bool ret;
4173
4174 rcu_read_lock_sched();
4175
4176 if (cpu == WORK_CPU_UNBOUND)
4177 cpu = smp_processor_id();
4178
4179 if (!(wq->flags & WQ_UNBOUND))
4180 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4181 else
4182 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4183
4184 ret = !list_empty(&pwq->delayed_works);
4185 rcu_read_unlock_sched();
4186
4187 return ret;
4188}
4189EXPORT_SYMBOL_GPL(workqueue_congested);
4190
4191/**
4192 * work_busy - test whether a work is currently pending or running
4193 * @work: the work to be tested
4194 *
4195 * Test whether @work is currently pending or running. There is no
4196 * synchronization around this function and the test result is
4197 * unreliable and only useful as advisory hints or for debugging.
4198 *
4199 * Return:
4200 * OR'd bitmask of WORK_BUSY_* bits.
4201 */
4202unsigned int work_busy(struct work_struct *work)
4203{
4204 struct worker_pool *pool;
4205 unsigned long flags;
4206 unsigned int ret = 0;
4207
4208 if (work_pending(work))
4209 ret |= WORK_BUSY_PENDING;
4210
4211 local_irq_save(flags);
4212 pool = get_work_pool(work);
4213 if (pool) {
4214 spin_lock(&pool->lock);
4215 if (find_worker_executing_work(pool, work))
4216 ret |= WORK_BUSY_RUNNING;
4217 spin_unlock(&pool->lock);
4218 }
4219 local_irq_restore(flags);
4220
4221 return ret;
4222}
4223EXPORT_SYMBOL_GPL(work_busy);
4224
4225/**
4226 * set_worker_desc - set description for the current work item
4227 * @fmt: printf-style format string
4228 * @...: arguments for the format string
4229 *
4230 * This function can be called by a running work function to describe what
4231 * the work item is about. If the worker task gets dumped, this
4232 * information will be printed out together to help debugging. The
4233 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4234 */
4235void set_worker_desc(const char *fmt, ...)
4236{
4237 struct worker *worker = current_wq_worker();
4238 va_list args;
4239
4240 if (worker) {
4241 va_start(args, fmt);
4242 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4243 va_end(args);
4244 worker->desc_valid = true;
4245 }
4246}
4247
4248/**
4249 * print_worker_info - print out worker information and description
4250 * @log_lvl: the log level to use when printing
4251 * @task: target task
4252 *
4253 * If @task is a worker and currently executing a work item, print out the
4254 * name of the workqueue being serviced and worker description set with
4255 * set_worker_desc() by the currently executing work item.
4256 *
4257 * This function can be safely called on any task as long as the
4258 * task_struct itself is accessible. While safe, this function isn't
4259 * synchronized and may print out mixups or garbages of limited length.
4260 */
4261void print_worker_info(const char *log_lvl, struct task_struct *task)
4262{
4263 work_func_t *fn = NULL;
4264 char name[WQ_NAME_LEN] = { };
4265 char desc[WORKER_DESC_LEN] = { };
4266 struct pool_workqueue *pwq = NULL;
4267 struct workqueue_struct *wq = NULL;
4268 bool desc_valid = false;
4269 struct worker *worker;
4270
4271 if (!(task->flags & PF_WQ_WORKER))
4272 return;
4273
4274 /*
4275 * This function is called without any synchronization and @task
4276 * could be in any state. Be careful with dereferences.
4277 */
4278 worker = probe_kthread_data(task);
4279
4280 /*
4281 * Carefully copy the associated workqueue's workfn and name. Keep
4282 * the original last '\0' in case the original contains garbage.
4283 */
4284 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4285 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4286 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4287 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4288
4289 /* copy worker description */
4290 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4291 if (desc_valid)
4292 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4293
4294 if (fn || name[0] || desc[0]) {
4295 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4296 if (desc[0])
4297 pr_cont(" (%s)", desc);
4298 pr_cont("\n");
4299 }
4300}
4301
4302static void pr_cont_pool_info(struct worker_pool *pool)
4303{
4304 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4305 if (pool->node != NUMA_NO_NODE)
4306 pr_cont(" node=%d", pool->node);
4307 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4308}
4309
4310static void pr_cont_work(bool comma, struct work_struct *work)
4311{
4312 if (work->func == wq_barrier_func) {
4313 struct wq_barrier *barr;
4314
4315 barr = container_of(work, struct wq_barrier, work);
4316
4317 pr_cont("%s BAR(%d)", comma ? "," : "",
4318 task_pid_nr(barr->task));
4319 } else {
4320 pr_cont("%s %pf", comma ? "," : "", work->func);
4321 }
4322}
4323
4324static void show_pwq(struct pool_workqueue *pwq)
4325{
4326 struct worker_pool *pool = pwq->pool;
4327 struct work_struct *work;
4328 struct worker *worker;
4329 bool has_in_flight = false, has_pending = false;
4330 int bkt;
4331
4332 pr_info(" pwq %d:", pool->id);
4333 pr_cont_pool_info(pool);
4334
4335 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4336 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4337
4338 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4339 if (worker->current_pwq == pwq) {
4340 has_in_flight = true;
4341 break;
4342 }
4343 }
4344 if (has_in_flight) {
4345 bool comma = false;
4346
4347 pr_info(" in-flight:");
4348 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4349 if (worker->current_pwq != pwq)
4350 continue;
4351
4352 pr_cont("%s %d%s:%pf", comma ? "," : "",
4353 task_pid_nr(worker->task),
4354 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4355 worker->current_func);
4356 list_for_each_entry(work, &worker->scheduled, entry)
4357 pr_cont_work(false, work);
4358 comma = true;
4359 }
4360 pr_cont("\n");
4361 }
4362
4363 list_for_each_entry(work, &pool->worklist, entry) {
4364 if (get_work_pwq(work) == pwq) {
4365 has_pending = true;
4366 break;
4367 }
4368 }
4369 if (has_pending) {
4370 bool comma = false;
4371
4372 pr_info(" pending:");
4373 list_for_each_entry(work, &pool->worklist, entry) {
4374 if (get_work_pwq(work) != pwq)
4375 continue;
4376
4377 pr_cont_work(comma, work);
4378 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4379 }
4380 pr_cont("\n");
4381 }
4382
4383 if (!list_empty(&pwq->delayed_works)) {
4384 bool comma = false;
4385
4386 pr_info(" delayed:");
4387 list_for_each_entry(work, &pwq->delayed_works, entry) {
4388 pr_cont_work(comma, work);
4389 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4390 }
4391 pr_cont("\n");
4392 }
4393}
4394
4395/**
4396 * show_workqueue_state - dump workqueue state
4397 *
4398 * Called from a sysrq handler and prints out all busy workqueues and
4399 * pools.
4400 */
4401void show_workqueue_state(void)
4402{
4403 struct workqueue_struct *wq;
4404 struct worker_pool *pool;
4405 unsigned long flags;
4406 int pi;
4407
4408 rcu_read_lock_sched();
4409
4410 pr_info("Showing busy workqueues and worker pools:\n");
4411
4412 list_for_each_entry_rcu(wq, &workqueues, list) {
4413 struct pool_workqueue *pwq;
4414 bool idle = true;
4415
4416 for_each_pwq(pwq, wq) {
4417 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4418 idle = false;
4419 break;
4420 }
4421 }
4422 if (idle)
4423 continue;
4424
4425 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4426
4427 for_each_pwq(pwq, wq) {
4428 spin_lock_irqsave(&pwq->pool->lock, flags);
4429 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4430 show_pwq(pwq);
4431 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4432 }
4433 }
4434
4435 for_each_pool(pool, pi) {
4436 struct worker *worker;
4437 bool first = true;
4438
4439 spin_lock_irqsave(&pool->lock, flags);
4440 if (pool->nr_workers == pool->nr_idle)
4441 goto next_pool;
4442
4443 pr_info("pool %d:", pool->id);
4444 pr_cont_pool_info(pool);
4445 pr_cont(" hung=%us workers=%d",
4446 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4447 pool->nr_workers);
4448 if (pool->manager)
4449 pr_cont(" manager: %d",
4450 task_pid_nr(pool->manager->task));
4451 list_for_each_entry(worker, &pool->idle_list, entry) {
4452 pr_cont(" %s%d", first ? "idle: " : "",
4453 task_pid_nr(worker->task));
4454 first = false;
4455 }
4456 pr_cont("\n");
4457 next_pool:
4458 spin_unlock_irqrestore(&pool->lock, flags);
4459 }
4460
4461 rcu_read_unlock_sched();
4462}
4463
4464/*
4465 * CPU hotplug.
4466 *
4467 * There are two challenges in supporting CPU hotplug. Firstly, there
4468 * are a lot of assumptions on strong associations among work, pwq and
4469 * pool which make migrating pending and scheduled works very
4470 * difficult to implement without impacting hot paths. Secondly,
4471 * worker pools serve mix of short, long and very long running works making
4472 * blocked draining impractical.
4473 *
4474 * This is solved by allowing the pools to be disassociated from the CPU
4475 * running as an unbound one and allowing it to be reattached later if the
4476 * cpu comes back online.
4477 */
4478
4479static void wq_unbind_fn(struct work_struct *work)
4480{
4481 int cpu = smp_processor_id();
4482 struct worker_pool *pool;
4483 struct worker *worker;
4484
4485 for_each_cpu_worker_pool(pool, cpu) {
4486 mutex_lock(&pool->attach_mutex);
4487 spin_lock_irq(&pool->lock);
4488
4489 /*
4490 * We've blocked all attach/detach operations. Make all workers
4491 * unbound and set DISASSOCIATED. Before this, all workers
4492 * except for the ones which are still executing works from
4493 * before the last CPU down must be on the cpu. After
4494 * this, they may become diasporas.
4495 */
4496 for_each_pool_worker(worker, pool)
4497 worker->flags |= WORKER_UNBOUND;
4498
4499 pool->flags |= POOL_DISASSOCIATED;
4500
4501 spin_unlock_irq(&pool->lock);
4502 mutex_unlock(&pool->attach_mutex);
4503
4504 /*
4505 * Call schedule() so that we cross rq->lock and thus can
4506 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4507 * This is necessary as scheduler callbacks may be invoked
4508 * from other cpus.
4509 */
4510 schedule();
4511
4512 /*
4513 * Sched callbacks are disabled now. Zap nr_running.
4514 * After this, nr_running stays zero and need_more_worker()
4515 * and keep_working() are always true as long as the
4516 * worklist is not empty. This pool now behaves as an
4517 * unbound (in terms of concurrency management) pool which
4518 * are served by workers tied to the pool.
4519 */
4520 atomic_set(&pool->nr_running, 0);
4521
4522 /*
4523 * With concurrency management just turned off, a busy
4524 * worker blocking could lead to lengthy stalls. Kick off
4525 * unbound chain execution of currently pending work items.
4526 */
4527 spin_lock_irq(&pool->lock);
4528 wake_up_worker(pool);
4529 spin_unlock_irq(&pool->lock);
4530 }
4531}
4532
4533/**
4534 * rebind_workers - rebind all workers of a pool to the associated CPU
4535 * @pool: pool of interest
4536 *
4537 * @pool->cpu is coming online. Rebind all workers to the CPU.
4538 */
4539static void rebind_workers(struct worker_pool *pool)
4540{
4541 struct worker *worker;
4542
4543 lockdep_assert_held(&pool->attach_mutex);
4544
4545 /*
4546 * Restore CPU affinity of all workers. As all idle workers should
4547 * be on the run-queue of the associated CPU before any local
4548 * wake-ups for concurrency management happen, restore CPU affinity
4549 * of all workers first and then clear UNBOUND. As we're called
4550 * from CPU_ONLINE, the following shouldn't fail.
4551 */
4552 for_each_pool_worker(worker, pool)
4553 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4554 pool->attrs->cpumask) < 0);
4555
4556 spin_lock_irq(&pool->lock);
4557
4558 /*
4559 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4560 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4561 * being reworked and this can go away in time.
4562 */
4563 if (!(pool->flags & POOL_DISASSOCIATED)) {
4564 spin_unlock_irq(&pool->lock);
4565 return;
4566 }
4567
4568 pool->flags &= ~POOL_DISASSOCIATED;
4569
4570 for_each_pool_worker(worker, pool) {
4571 unsigned int worker_flags = worker->flags;
4572
4573 /*
4574 * A bound idle worker should actually be on the runqueue
4575 * of the associated CPU for local wake-ups targeting it to
4576 * work. Kick all idle workers so that they migrate to the
4577 * associated CPU. Doing this in the same loop as
4578 * replacing UNBOUND with REBOUND is safe as no worker will
4579 * be bound before @pool->lock is released.
4580 */
4581 if (worker_flags & WORKER_IDLE)
4582 wake_up_process(worker->task);
4583
4584 /*
4585 * We want to clear UNBOUND but can't directly call
4586 * worker_clr_flags() or adjust nr_running. Atomically
4587 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4588 * @worker will clear REBOUND using worker_clr_flags() when
4589 * it initiates the next execution cycle thus restoring
4590 * concurrency management. Note that when or whether
4591 * @worker clears REBOUND doesn't affect correctness.
4592 *
4593 * ACCESS_ONCE() is necessary because @worker->flags may be
4594 * tested without holding any lock in
4595 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4596 * fail incorrectly leading to premature concurrency
4597 * management operations.
4598 */
4599 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4600 worker_flags |= WORKER_REBOUND;
4601 worker_flags &= ~WORKER_UNBOUND;
4602 ACCESS_ONCE(worker->flags) = worker_flags;
4603 }
4604
4605 spin_unlock_irq(&pool->lock);
4606}
4607
4608/**
4609 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4610 * @pool: unbound pool of interest
4611 * @cpu: the CPU which is coming up
4612 *
4613 * An unbound pool may end up with a cpumask which doesn't have any online
4614 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4615 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4616 * online CPU before, cpus_allowed of all its workers should be restored.
4617 */
4618static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4619{
4620 static cpumask_t cpumask;
4621 struct worker *worker;
4622
4623 lockdep_assert_held(&pool->attach_mutex);
4624
4625 /* is @cpu allowed for @pool? */
4626 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4627 return;
4628
4629 /* is @cpu the only online CPU? */
4630 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4631 if (cpumask_weight(&cpumask) != 1)
4632 return;
4633
4634 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4635 for_each_pool_worker(worker, pool)
4636 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4637 pool->attrs->cpumask) < 0);
4638}
4639
4640/*
4641 * Workqueues should be brought up before normal priority CPU notifiers.
4642 * This will be registered high priority CPU notifier.
4643 */
4644static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4645 unsigned long action,
4646 void *hcpu)
4647{
4648 int cpu = (unsigned long)hcpu;
4649 struct worker_pool *pool;
4650 struct workqueue_struct *wq;
4651 int pi;
4652
4653 switch (action & ~CPU_TASKS_FROZEN) {
4654 case CPU_UP_PREPARE:
4655 for_each_cpu_worker_pool(pool, cpu) {
4656 if (pool->nr_workers)
4657 continue;
4658 if (!create_worker(pool))
4659 return NOTIFY_BAD;
4660 }
4661 break;
4662
4663 case CPU_DOWN_FAILED:
4664 case CPU_ONLINE:
4665 mutex_lock(&wq_pool_mutex);
4666
4667 for_each_pool(pool, pi) {
4668 mutex_lock(&pool->attach_mutex);
4669
4670 if (pool->cpu == cpu)
4671 rebind_workers(pool);
4672 else if (pool->cpu < 0)
4673 restore_unbound_workers_cpumask(pool, cpu);
4674
4675 mutex_unlock(&pool->attach_mutex);
4676 }
4677
4678 /* update NUMA affinity of unbound workqueues */
4679 list_for_each_entry(wq, &workqueues, list)
4680 wq_update_unbound_numa(wq, cpu, true);
4681
4682 mutex_unlock(&wq_pool_mutex);
4683 break;
4684 }
4685 return NOTIFY_OK;
4686}
4687
4688/*
4689 * Workqueues should be brought down after normal priority CPU notifiers.
4690 * This will be registered as low priority CPU notifier.
4691 */
4692static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4693 unsigned long action,
4694 void *hcpu)
4695{
4696 int cpu = (unsigned long)hcpu;
4697 struct work_struct unbind_work;
4698 struct workqueue_struct *wq;
4699
4700 switch (action & ~CPU_TASKS_FROZEN) {
4701 case CPU_DOWN_PREPARE:
4702 /* unbinding per-cpu workers should happen on the local CPU */
4703 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4704 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4705
4706 /* update NUMA affinity of unbound workqueues */
4707 mutex_lock(&wq_pool_mutex);
4708 list_for_each_entry(wq, &workqueues, list)
4709 wq_update_unbound_numa(wq, cpu, false);
4710 mutex_unlock(&wq_pool_mutex);
4711
4712 /* wait for per-cpu unbinding to finish */
4713 flush_work(&unbind_work);
4714 destroy_work_on_stack(&unbind_work);
4715 break;
4716 }
4717 return NOTIFY_OK;
4718}
4719
4720#ifdef CONFIG_SMP
4721
4722struct work_for_cpu {
4723 struct work_struct work;
4724 long (*fn)(void *);
4725 void *arg;
4726 long ret;
4727};
4728
4729static void work_for_cpu_fn(struct work_struct *work)
4730{
4731 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4732
4733 wfc->ret = wfc->fn(wfc->arg);
4734}
4735
4736/**
4737 * work_on_cpu - run a function in thread context on a particular cpu
4738 * @cpu: the cpu to run on
4739 * @fn: the function to run
4740 * @arg: the function arg
4741 *
4742 * It is up to the caller to ensure that the cpu doesn't go offline.
4743 * The caller must not hold any locks which would prevent @fn from completing.
4744 *
4745 * Return: The value @fn returns.
4746 */
4747long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4748{
4749 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4750
4751 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4752 schedule_work_on(cpu, &wfc.work);
4753 flush_work(&wfc.work);
4754 destroy_work_on_stack(&wfc.work);
4755 return wfc.ret;
4756}
4757EXPORT_SYMBOL_GPL(work_on_cpu);
4758#endif /* CONFIG_SMP */
4759
4760#ifdef CONFIG_FREEZER
4761
4762/**
4763 * freeze_workqueues_begin - begin freezing workqueues
4764 *
4765 * Start freezing workqueues. After this function returns, all freezable
4766 * workqueues will queue new works to their delayed_works list instead of
4767 * pool->worklist.
4768 *
4769 * CONTEXT:
4770 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4771 */
4772void freeze_workqueues_begin(void)
4773{
4774 struct workqueue_struct *wq;
4775 struct pool_workqueue *pwq;
4776
4777 mutex_lock(&wq_pool_mutex);
4778
4779 WARN_ON_ONCE(workqueue_freezing);
4780 workqueue_freezing = true;
4781
4782 list_for_each_entry(wq, &workqueues, list) {
4783 mutex_lock(&wq->mutex);
4784 for_each_pwq(pwq, wq)
4785 pwq_adjust_max_active(pwq);
4786 mutex_unlock(&wq->mutex);
4787 }
4788
4789 mutex_unlock(&wq_pool_mutex);
4790}
4791
4792/**
4793 * freeze_workqueues_busy - are freezable workqueues still busy?
4794 *
4795 * Check whether freezing is complete. This function must be called
4796 * between freeze_workqueues_begin() and thaw_workqueues().
4797 *
4798 * CONTEXT:
4799 * Grabs and releases wq_pool_mutex.
4800 *
4801 * Return:
4802 * %true if some freezable workqueues are still busy. %false if freezing
4803 * is complete.
4804 */
4805bool freeze_workqueues_busy(void)
4806{
4807 bool busy = false;
4808 struct workqueue_struct *wq;
4809 struct pool_workqueue *pwq;
4810
4811 mutex_lock(&wq_pool_mutex);
4812
4813 WARN_ON_ONCE(!workqueue_freezing);
4814
4815 list_for_each_entry(wq, &workqueues, list) {
4816 if (!(wq->flags & WQ_FREEZABLE))
4817 continue;
4818 /*
4819 * nr_active is monotonically decreasing. It's safe
4820 * to peek without lock.
4821 */
4822 rcu_read_lock_sched();
4823 for_each_pwq(pwq, wq) {
4824 WARN_ON_ONCE(pwq->nr_active < 0);
4825 if (pwq->nr_active) {
4826 busy = true;
4827 rcu_read_unlock_sched();
4828 goto out_unlock;
4829 }
4830 }
4831 rcu_read_unlock_sched();
4832 }
4833out_unlock:
4834 mutex_unlock(&wq_pool_mutex);
4835 return busy;
4836}
4837
4838/**
4839 * thaw_workqueues - thaw workqueues
4840 *
4841 * Thaw workqueues. Normal queueing is restored and all collected
4842 * frozen works are transferred to their respective pool worklists.
4843 *
4844 * CONTEXT:
4845 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4846 */
4847void thaw_workqueues(void)
4848{
4849 struct workqueue_struct *wq;
4850 struct pool_workqueue *pwq;
4851
4852 mutex_lock(&wq_pool_mutex);
4853
4854 if (!workqueue_freezing)
4855 goto out_unlock;
4856
4857 workqueue_freezing = false;
4858
4859 /* restore max_active and repopulate worklist */
4860 list_for_each_entry(wq, &workqueues, list) {
4861 mutex_lock(&wq->mutex);
4862 for_each_pwq(pwq, wq)
4863 pwq_adjust_max_active(pwq);
4864 mutex_unlock(&wq->mutex);
4865 }
4866
4867out_unlock:
4868 mutex_unlock(&wq_pool_mutex);
4869}
4870#endif /* CONFIG_FREEZER */
4871
4872static int workqueue_apply_unbound_cpumask(void)
4873{
4874 LIST_HEAD(ctxs);
4875 int ret = 0;
4876 struct workqueue_struct *wq;
4877 struct apply_wqattrs_ctx *ctx, *n;
4878
4879 lockdep_assert_held(&wq_pool_mutex);
4880
4881 list_for_each_entry(wq, &workqueues, list) {
4882 if (!(wq->flags & WQ_UNBOUND))
4883 continue;
4884 /* creating multiple pwqs breaks ordering guarantee */
4885 if (wq->flags & __WQ_ORDERED)
4886 continue;
4887
4888 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4889 if (!ctx) {
4890 ret = -ENOMEM;
4891 break;
4892 }
4893
4894 list_add_tail(&ctx->list, &ctxs);
4895 }
4896
4897 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4898 if (!ret)
4899 apply_wqattrs_commit(ctx);
4900 apply_wqattrs_cleanup(ctx);
4901 }
4902
4903 return ret;
4904}
4905
4906/**
4907 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4908 * @cpumask: the cpumask to set
4909 *
4910 * The low-level workqueues cpumask is a global cpumask that limits
4911 * the affinity of all unbound workqueues. This function check the @cpumask
4912 * and apply it to all unbound workqueues and updates all pwqs of them.
4913 *
4914 * Retun: 0 - Success
4915 * -EINVAL - Invalid @cpumask
4916 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4917 */
4918int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4919{
4920 int ret = -EINVAL;
4921 cpumask_var_t saved_cpumask;
4922
4923 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4924 return -ENOMEM;
4925
4926 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4927 if (!cpumask_empty(cpumask)) {
4928 apply_wqattrs_lock();
4929
4930 /* save the old wq_unbound_cpumask. */
4931 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4932
4933 /* update wq_unbound_cpumask at first and apply it to wqs. */
4934 cpumask_copy(wq_unbound_cpumask, cpumask);
4935 ret = workqueue_apply_unbound_cpumask();
4936
4937 /* restore the wq_unbound_cpumask when failed. */
4938 if (ret < 0)
4939 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4940
4941 apply_wqattrs_unlock();
4942 }
4943
4944 free_cpumask_var(saved_cpumask);
4945 return ret;
4946}
4947
4948#ifdef CONFIG_SYSFS
4949/*
4950 * Workqueues with WQ_SYSFS flag set is visible to userland via
4951 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4952 * following attributes.
4953 *
4954 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4955 * max_active RW int : maximum number of in-flight work items
4956 *
4957 * Unbound workqueues have the following extra attributes.
4958 *
4959 * id RO int : the associated pool ID
4960 * nice RW int : nice value of the workers
4961 * cpumask RW mask : bitmask of allowed CPUs for the workers
4962 */
4963struct wq_device {
4964 struct workqueue_struct *wq;
4965 struct device dev;
4966};
4967
4968static struct workqueue_struct *dev_to_wq(struct device *dev)
4969{
4970 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4971
4972 return wq_dev->wq;
4973}
4974
4975static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4976 char *buf)
4977{
4978 struct workqueue_struct *wq = dev_to_wq(dev);
4979
4980 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4981}
4982static DEVICE_ATTR_RO(per_cpu);
4983
4984static ssize_t max_active_show(struct device *dev,
4985 struct device_attribute *attr, char *buf)
4986{
4987 struct workqueue_struct *wq = dev_to_wq(dev);
4988
4989 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4990}
4991
4992static ssize_t max_active_store(struct device *dev,
4993 struct device_attribute *attr, const char *buf,
4994 size_t count)
4995{
4996 struct workqueue_struct *wq = dev_to_wq(dev);
4997 int val;
4998
4999 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5000 return -EINVAL;
5001
5002 workqueue_set_max_active(wq, val);
5003 return count;
5004}
5005static DEVICE_ATTR_RW(max_active);
5006
5007static struct attribute *wq_sysfs_attrs[] = {
5008 &dev_attr_per_cpu.attr,
5009 &dev_attr_max_active.attr,
5010 NULL,
5011};
5012ATTRIBUTE_GROUPS(wq_sysfs);
5013
5014static ssize_t wq_pool_ids_show(struct device *dev,
5015 struct device_attribute *attr, char *buf)
5016{
5017 struct workqueue_struct *wq = dev_to_wq(dev);
5018 const char *delim = "";
5019 int node, written = 0;
5020
5021 rcu_read_lock_sched();
5022 for_each_node(node) {
5023 written += scnprintf(buf + written, PAGE_SIZE - written,
5024 "%s%d:%d", delim, node,
5025 unbound_pwq_by_node(wq, node)->pool->id);
5026 delim = " ";
5027 }
5028 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5029 rcu_read_unlock_sched();
5030
5031 return written;
5032}
5033
5034static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5035 char *buf)
5036{
5037 struct workqueue_struct *wq = dev_to_wq(dev);
5038 int written;
5039
5040 mutex_lock(&wq->mutex);
5041 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5042 mutex_unlock(&wq->mutex);
5043
5044 return written;
5045}
5046
5047/* prepare workqueue_attrs for sysfs store operations */
5048static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5049{
5050 struct workqueue_attrs *attrs;
5051
5052 lockdep_assert_held(&wq_pool_mutex);
5053
5054 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5055 if (!attrs)
5056 return NULL;
5057
5058 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5059 return attrs;
5060}
5061
5062static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5063 const char *buf, size_t count)
5064{
5065 struct workqueue_struct *wq = dev_to_wq(dev);
5066 struct workqueue_attrs *attrs;
5067 int ret = -ENOMEM;
5068
5069 apply_wqattrs_lock();
5070
5071 attrs = wq_sysfs_prep_attrs(wq);
5072 if (!attrs)
5073 goto out_unlock;
5074
5075 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5076 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5077 ret = apply_workqueue_attrs_locked(wq, attrs);
5078 else
5079 ret = -EINVAL;
5080
5081out_unlock:
5082 apply_wqattrs_unlock();
5083 free_workqueue_attrs(attrs);
5084 return ret ?: count;
5085}
5086
5087static ssize_t wq_cpumask_show(struct device *dev,
5088 struct device_attribute *attr, char *buf)
5089{
5090 struct workqueue_struct *wq = dev_to_wq(dev);
5091 int written;
5092
5093 mutex_lock(&wq->mutex);
5094 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5095 cpumask_pr_args(wq->unbound_attrs->cpumask));
5096 mutex_unlock(&wq->mutex);
5097 return written;
5098}
5099
5100static ssize_t wq_cpumask_store(struct device *dev,
5101 struct device_attribute *attr,
5102 const char *buf, size_t count)
5103{
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105 struct workqueue_attrs *attrs;
5106 int ret = -ENOMEM;
5107
5108 apply_wqattrs_lock();
5109
5110 attrs = wq_sysfs_prep_attrs(wq);
5111 if (!attrs)
5112 goto out_unlock;
5113
5114 ret = cpumask_parse(buf, attrs->cpumask);
5115 if (!ret)
5116 ret = apply_workqueue_attrs_locked(wq, attrs);
5117
5118out_unlock:
5119 apply_wqattrs_unlock();
5120 free_workqueue_attrs(attrs);
5121 return ret ?: count;
5122}
5123
5124static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5125 char *buf)
5126{
5127 struct workqueue_struct *wq = dev_to_wq(dev);
5128 int written;
5129
5130 mutex_lock(&wq->mutex);
5131 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5132 !wq->unbound_attrs->no_numa);
5133 mutex_unlock(&wq->mutex);
5134
5135 return written;
5136}
5137
5138static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5139 const char *buf, size_t count)
5140{
5141 struct workqueue_struct *wq = dev_to_wq(dev);
5142 struct workqueue_attrs *attrs;
5143 int v, ret = -ENOMEM;
5144
5145 apply_wqattrs_lock();
5146
5147 attrs = wq_sysfs_prep_attrs(wq);
5148 if (!attrs)
5149 goto out_unlock;
5150
5151 ret = -EINVAL;
5152 if (sscanf(buf, "%d", &v) == 1) {
5153 attrs->no_numa = !v;
5154 ret = apply_workqueue_attrs_locked(wq, attrs);
5155 }
5156
5157out_unlock:
5158 apply_wqattrs_unlock();
5159 free_workqueue_attrs(attrs);
5160 return ret ?: count;
5161}
5162
5163static struct device_attribute wq_sysfs_unbound_attrs[] = {
5164 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5165 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5166 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5167 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5168 __ATTR_NULL,
5169};
5170
5171static struct bus_type wq_subsys = {
5172 .name = "workqueue",
5173 .dev_groups = wq_sysfs_groups,
5174};
5175
5176static ssize_t wq_unbound_cpumask_show(struct device *dev,
5177 struct device_attribute *attr, char *buf)
5178{
5179 int written;
5180
5181 mutex_lock(&wq_pool_mutex);
5182 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5183 cpumask_pr_args(wq_unbound_cpumask));
5184 mutex_unlock(&wq_pool_mutex);
5185
5186 return written;
5187}
5188
5189static ssize_t wq_unbound_cpumask_store(struct device *dev,
5190 struct device_attribute *attr, const char *buf, size_t count)
5191{
5192 cpumask_var_t cpumask;
5193 int ret;
5194
5195 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5196 return -ENOMEM;
5197
5198 ret = cpumask_parse(buf, cpumask);
5199 if (!ret)
5200 ret = workqueue_set_unbound_cpumask(cpumask);
5201
5202 free_cpumask_var(cpumask);
5203 return ret ? ret : count;
5204}
5205
5206static struct device_attribute wq_sysfs_cpumask_attr =
5207 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5208 wq_unbound_cpumask_store);
5209
5210static int __init wq_sysfs_init(void)
5211{
5212 int err;
5213
5214 err = subsys_virtual_register(&wq_subsys, NULL);
5215 if (err)
5216 return err;
5217
5218 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5219}
5220core_initcall(wq_sysfs_init);
5221
5222static void wq_device_release(struct device *dev)
5223{
5224 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5225
5226 kfree(wq_dev);
5227}
5228
5229/**
5230 * workqueue_sysfs_register - make a workqueue visible in sysfs
5231 * @wq: the workqueue to register
5232 *
5233 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5234 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5235 * which is the preferred method.
5236 *
5237 * Workqueue user should use this function directly iff it wants to apply
5238 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5239 * apply_workqueue_attrs() may race against userland updating the
5240 * attributes.
5241 *
5242 * Return: 0 on success, -errno on failure.
5243 */
5244int workqueue_sysfs_register(struct workqueue_struct *wq)
5245{
5246 struct wq_device *wq_dev;
5247 int ret;
5248
5249 /*
5250 * Adjusting max_active or creating new pwqs by applying
5251 * attributes breaks ordering guarantee. Disallow exposing ordered
5252 * workqueues.
5253 */
5254 if (WARN_ON(wq->flags & __WQ_ORDERED))
5255 return -EINVAL;
5256
5257 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5258 if (!wq_dev)
5259 return -ENOMEM;
5260
5261 wq_dev->wq = wq;
5262 wq_dev->dev.bus = &wq_subsys;
5263 wq_dev->dev.release = wq_device_release;
5264 dev_set_name(&wq_dev->dev, "%s", wq->name);
5265
5266 /*
5267 * unbound_attrs are created separately. Suppress uevent until
5268 * everything is ready.
5269 */
5270 dev_set_uevent_suppress(&wq_dev->dev, true);
5271
5272 ret = device_register(&wq_dev->dev);
5273 if (ret) {
5274 kfree(wq_dev);
5275 wq->wq_dev = NULL;
5276 return ret;
5277 }
5278
5279 if (wq->flags & WQ_UNBOUND) {
5280 struct device_attribute *attr;
5281
5282 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5283 ret = device_create_file(&wq_dev->dev, attr);
5284 if (ret) {
5285 device_unregister(&wq_dev->dev);
5286 wq->wq_dev = NULL;
5287 return ret;
5288 }
5289 }
5290 }
5291
5292 dev_set_uevent_suppress(&wq_dev->dev, false);
5293 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5294 return 0;
5295}
5296
5297/**
5298 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5299 * @wq: the workqueue to unregister
5300 *
5301 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5302 */
5303static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5304{
5305 struct wq_device *wq_dev = wq->wq_dev;
5306
5307 if (!wq->wq_dev)
5308 return;
5309
5310 wq->wq_dev = NULL;
5311 device_unregister(&wq_dev->dev);
5312}
5313#else /* CONFIG_SYSFS */
5314static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5315#endif /* CONFIG_SYSFS */
5316
5317/*
5318 * Workqueue watchdog.
5319 *
5320 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5321 * flush dependency, a concurrency managed work item which stays RUNNING
5322 * indefinitely. Workqueue stalls can be very difficult to debug as the
5323 * usual warning mechanisms don't trigger and internal workqueue state is
5324 * largely opaque.
5325 *
5326 * Workqueue watchdog monitors all worker pools periodically and dumps
5327 * state if some pools failed to make forward progress for a while where
5328 * forward progress is defined as the first item on ->worklist changing.
5329 *
5330 * This mechanism is controlled through the kernel parameter
5331 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5332 * corresponding sysfs parameter file.
5333 */
5334#ifdef CONFIG_WQ_WATCHDOG
5335
5336static void wq_watchdog_timer_fn(unsigned long data);
5337
5338static unsigned long wq_watchdog_thresh = 30;
5339static struct timer_list wq_watchdog_timer =
5340 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5341
5342static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5343static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5344
5345static void wq_watchdog_reset_touched(void)
5346{
5347 int cpu;
5348
5349 wq_watchdog_touched = jiffies;
5350 for_each_possible_cpu(cpu)
5351 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5352}
5353
5354static void wq_watchdog_timer_fn(unsigned long data)
5355{
5356 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5357 bool lockup_detected = false;
5358 struct worker_pool *pool;
5359 int pi;
5360
5361 if (!thresh)
5362 return;
5363
5364 rcu_read_lock();
5365
5366 for_each_pool(pool, pi) {
5367 unsigned long pool_ts, touched, ts;
5368
5369 if (list_empty(&pool->worklist))
5370 continue;
5371
5372 /* get the latest of pool and touched timestamps */
5373 pool_ts = READ_ONCE(pool->watchdog_ts);
5374 touched = READ_ONCE(wq_watchdog_touched);
5375
5376 if (time_after(pool_ts, touched))
5377 ts = pool_ts;
5378 else
5379 ts = touched;
5380
5381 if (pool->cpu >= 0) {
5382 unsigned long cpu_touched =
5383 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5384 pool->cpu));
5385 if (time_after(cpu_touched, ts))
5386 ts = cpu_touched;
5387 }
5388
5389 /* did we stall? */
5390 if (time_after(jiffies, ts + thresh)) {
5391 lockup_detected = true;
5392 pr_emerg("BUG: workqueue lockup - pool");
5393 pr_cont_pool_info(pool);
5394 pr_cont(" stuck for %us!\n",
5395 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5396 }
5397 }
5398
5399 rcu_read_unlock();
5400
5401 if (lockup_detected)
5402 show_workqueue_state();
5403
5404 wq_watchdog_reset_touched();
5405 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5406}
5407
5408void wq_watchdog_touch(int cpu)
5409{
5410 if (cpu >= 0)
5411 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5412 else
5413 wq_watchdog_touched = jiffies;
5414}
5415
5416static void wq_watchdog_set_thresh(unsigned long thresh)
5417{
5418 wq_watchdog_thresh = 0;
5419 del_timer_sync(&wq_watchdog_timer);
5420
5421 if (thresh) {
5422 wq_watchdog_thresh = thresh;
5423 wq_watchdog_reset_touched();
5424 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5425 }
5426}
5427
5428static int wq_watchdog_param_set_thresh(const char *val,
5429 const struct kernel_param *kp)
5430{
5431 unsigned long thresh;
5432 int ret;
5433
5434 ret = kstrtoul(val, 0, &thresh);
5435 if (ret)
5436 return ret;
5437
5438 if (system_wq)
5439 wq_watchdog_set_thresh(thresh);
5440 else
5441 wq_watchdog_thresh = thresh;
5442
5443 return 0;
5444}
5445
5446static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5447 .set = wq_watchdog_param_set_thresh,
5448 .get = param_get_ulong,
5449};
5450
5451module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5452 0644);
5453
5454static void wq_watchdog_init(void)
5455{
5456 wq_watchdog_set_thresh(wq_watchdog_thresh);
5457}
5458
5459#else /* CONFIG_WQ_WATCHDOG */
5460
5461static inline void wq_watchdog_init(void) { }
5462
5463#endif /* CONFIG_WQ_WATCHDOG */
5464
5465static void __init wq_numa_init(void)
5466{
5467 cpumask_var_t *tbl;
5468 int node, cpu;
5469
5470 if (num_possible_nodes() <= 1)
5471 return;
5472
5473 if (wq_disable_numa) {
5474 pr_info("workqueue: NUMA affinity support disabled\n");
5475 return;
5476 }
5477
5478 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5479 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5480
5481 /*
5482 * We want masks of possible CPUs of each node which isn't readily
5483 * available. Build one from cpu_to_node() which should have been
5484 * fully initialized by now.
5485 */
5486 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5487 BUG_ON(!tbl);
5488
5489 for_each_node(node)
5490 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5491 node_online(node) ? node : NUMA_NO_NODE));
5492
5493 for_each_possible_cpu(cpu) {
5494 node = cpu_to_node(cpu);
5495 if (WARN_ON(node == NUMA_NO_NODE)) {
5496 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5497 /* happens iff arch is bonkers, let's just proceed */
5498 return;
5499 }
5500 cpumask_set_cpu(cpu, tbl[node]);
5501 }
5502
5503 wq_numa_possible_cpumask = tbl;
5504 wq_numa_enabled = true;
5505}
5506
5507static int __init init_workqueues(void)
5508{
5509 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5510 int i, cpu;
5511
5512 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5513
5514 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5515 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5516
5517 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5518
5519 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5520 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5521
5522 wq_numa_init();
5523
5524 /* initialize CPU pools */
5525 for_each_possible_cpu(cpu) {
5526 struct worker_pool *pool;
5527
5528 i = 0;
5529 for_each_cpu_worker_pool(pool, cpu) {
5530 BUG_ON(init_worker_pool(pool));
5531 pool->cpu = cpu;
5532 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5533 pool->attrs->nice = std_nice[i++];
5534 pool->node = cpu_to_node(cpu);
5535
5536 /* alloc pool ID */
5537 mutex_lock(&wq_pool_mutex);
5538 BUG_ON(worker_pool_assign_id(pool));
5539 mutex_unlock(&wq_pool_mutex);
5540 }
5541 }
5542
5543 /* create the initial worker */
5544 for_each_online_cpu(cpu) {
5545 struct worker_pool *pool;
5546
5547 for_each_cpu_worker_pool(pool, cpu) {
5548 pool->flags &= ~POOL_DISASSOCIATED;
5549 BUG_ON(!create_worker(pool));
5550 }
5551 }
5552
5553 /* create default unbound and ordered wq attrs */
5554 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5555 struct workqueue_attrs *attrs;
5556
5557 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5558 attrs->nice = std_nice[i];
5559 unbound_std_wq_attrs[i] = attrs;
5560
5561 /*
5562 * An ordered wq should have only one pwq as ordering is
5563 * guaranteed by max_active which is enforced by pwqs.
5564 * Turn off NUMA so that dfl_pwq is used for all nodes.
5565 */
5566 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5567 attrs->nice = std_nice[i];
5568 attrs->no_numa = true;
5569 ordered_wq_attrs[i] = attrs;
5570 }
5571
5572 system_wq = alloc_workqueue("events", 0, 0);
5573 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5574 system_long_wq = alloc_workqueue("events_long", 0, 0);
5575 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5576 WQ_UNBOUND_MAX_ACTIVE);
5577 system_freezable_wq = alloc_workqueue("events_freezable",
5578 WQ_FREEZABLE, 0);
5579 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5580 WQ_POWER_EFFICIENT, 0);
5581 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5582 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5583 0);
5584 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5585 !system_unbound_wq || !system_freezable_wq ||
5586 !system_power_efficient_wq ||
5587 !system_freezable_power_efficient_wq);
5588
5589 wq_watchdog_init();
5590
5591 return 0;
5592}
5593early_initcall(init_workqueues);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/workqueue.c - generic async execution with shared worker pool
4 *
5 * Copyright (C) 2002 Ingo Molnar
6 *
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
12 *
13 * Made to use alloc_percpu by Christoph Lameter.
14 *
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 *
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
24 *
25 * Please read Documentation/core-api/workqueue.rst for details.
26 */
27
28#include <linux/export.h>
29#include <linux/kernel.h>
30#include <linux/sched.h>
31#include <linux/init.h>
32#include <linux/signal.h>
33#include <linux/completion.h>
34#include <linux/workqueue.h>
35#include <linux/slab.h>
36#include <linux/cpu.h>
37#include <linux/notifier.h>
38#include <linux/kthread.h>
39#include <linux/hardirq.h>
40#include <linux/mempolicy.h>
41#include <linux/freezer.h>
42#include <linux/debug_locks.h>
43#include <linux/lockdep.h>
44#include <linux/idr.h>
45#include <linux/jhash.h>
46#include <linux/hashtable.h>
47#include <linux/rculist.h>
48#include <linux/nodemask.h>
49#include <linux/moduleparam.h>
50#include <linux/uaccess.h>
51#include <linux/sched/isolation.h>
52#include <linux/sched/debug.h>
53#include <linux/nmi.h>
54#include <linux/kvm_para.h>
55#include <linux/delay.h>
56
57#include "workqueue_internal.h"
58
59enum {
60 /*
61 * worker_pool flags
62 *
63 * A bound pool is either associated or disassociated with its CPU.
64 * While associated (!DISASSOCIATED), all workers are bound to the
65 * CPU and none has %WORKER_UNBOUND set and concurrency management
66 * is in effect.
67 *
68 * While DISASSOCIATED, the cpu may be offline and all workers have
69 * %WORKER_UNBOUND set and concurrency management disabled, and may
70 * be executing on any CPU. The pool behaves as an unbound one.
71 *
72 * Note that DISASSOCIATED should be flipped only while holding
73 * wq_pool_attach_mutex to avoid changing binding state while
74 * worker_attach_to_pool() is in progress.
75 */
76 POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */
77 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
78
79 /* worker flags */
80 WORKER_DIE = 1 << 1, /* die die die */
81 WORKER_IDLE = 1 << 2, /* is idle */
82 WORKER_PREP = 1 << 3, /* preparing to run works */
83 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
84 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
85 WORKER_REBOUND = 1 << 8, /* worker was rebound */
86
87 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
88 WORKER_UNBOUND | WORKER_REBOUND,
89
90 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
91
92 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
93 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
94
95 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
96 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
97
98 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
99 /* call for help after 10ms
100 (min two ticks) */
101 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
102 CREATE_COOLDOWN = HZ, /* time to breath after fail */
103
104 /*
105 * Rescue workers are used only on emergencies and shared by
106 * all cpus. Give MIN_NICE.
107 */
108 RESCUER_NICE_LEVEL = MIN_NICE,
109 HIGHPRI_NICE_LEVEL = MIN_NICE,
110
111 WQ_NAME_LEN = 24,
112};
113
114/*
115 * Structure fields follow one of the following exclusion rules.
116 *
117 * I: Modifiable by initialization/destruction paths and read-only for
118 * everyone else.
119 *
120 * P: Preemption protected. Disabling preemption is enough and should
121 * only be modified and accessed from the local cpu.
122 *
123 * L: pool->lock protected. Access with pool->lock held.
124 *
125 * K: Only modified by worker while holding pool->lock. Can be safely read by
126 * self, while holding pool->lock or from IRQ context if %current is the
127 * kworker.
128 *
129 * S: Only modified by worker self.
130 *
131 * A: wq_pool_attach_mutex protected.
132 *
133 * PL: wq_pool_mutex protected.
134 *
135 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
136 *
137 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
138 *
139 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
140 * RCU for reads.
141 *
142 * WQ: wq->mutex protected.
143 *
144 * WR: wq->mutex protected for writes. RCU protected for reads.
145 *
146 * MD: wq_mayday_lock protected.
147 *
148 * WD: Used internally by the watchdog.
149 */
150
151/* struct worker is defined in workqueue_internal.h */
152
153struct worker_pool {
154 raw_spinlock_t lock; /* the pool lock */
155 int cpu; /* I: the associated cpu */
156 int node; /* I: the associated node ID */
157 int id; /* I: pool ID */
158 unsigned int flags; /* L: flags */
159
160 unsigned long watchdog_ts; /* L: watchdog timestamp */
161 bool cpu_stall; /* WD: stalled cpu bound pool */
162
163 /*
164 * The counter is incremented in a process context on the associated CPU
165 * w/ preemption disabled, and decremented or reset in the same context
166 * but w/ pool->lock held. The readers grab pool->lock and are
167 * guaranteed to see if the counter reached zero.
168 */
169 int nr_running;
170
171 struct list_head worklist; /* L: list of pending works */
172
173 int nr_workers; /* L: total number of workers */
174 int nr_idle; /* L: currently idle workers */
175
176 struct list_head idle_list; /* L: list of idle workers */
177 struct timer_list idle_timer; /* L: worker idle timeout */
178 struct work_struct idle_cull_work; /* L: worker idle cleanup */
179
180 struct timer_list mayday_timer; /* L: SOS timer for workers */
181
182 /* a workers is either on busy_hash or idle_list, or the manager */
183 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
184 /* L: hash of busy workers */
185
186 struct worker *manager; /* L: purely informational */
187 struct list_head workers; /* A: attached workers */
188 struct list_head dying_workers; /* A: workers about to die */
189 struct completion *detach_completion; /* all workers detached */
190
191 struct ida worker_ida; /* worker IDs for task name */
192
193 struct workqueue_attrs *attrs; /* I: worker attributes */
194 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
195 int refcnt; /* PL: refcnt for unbound pools */
196
197 /*
198 * Destruction of pool is RCU protected to allow dereferences
199 * from get_work_pool().
200 */
201 struct rcu_head rcu;
202};
203
204/*
205 * Per-pool_workqueue statistics. These can be monitored using
206 * tools/workqueue/wq_monitor.py.
207 */
208enum pool_workqueue_stats {
209 PWQ_STAT_STARTED, /* work items started execution */
210 PWQ_STAT_COMPLETED, /* work items completed execution */
211 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
212 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
213 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
214 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
215 PWQ_STAT_MAYDAY, /* maydays to rescuer */
216 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
217
218 PWQ_NR_STATS,
219};
220
221/*
222 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
223 * of work_struct->data are used for flags and the remaining high bits
224 * point to the pwq; thus, pwqs need to be aligned at two's power of the
225 * number of flag bits.
226 */
227struct pool_workqueue {
228 struct worker_pool *pool; /* I: the associated pool */
229 struct workqueue_struct *wq; /* I: the owning workqueue */
230 int work_color; /* L: current color */
231 int flush_color; /* L: flushing color */
232 int refcnt; /* L: reference count */
233 int nr_in_flight[WORK_NR_COLORS];
234 /* L: nr of in_flight works */
235
236 /*
237 * nr_active management and WORK_STRUCT_INACTIVE:
238 *
239 * When pwq->nr_active >= max_active, new work item is queued to
240 * pwq->inactive_works instead of pool->worklist and marked with
241 * WORK_STRUCT_INACTIVE.
242 *
243 * All work items marked with WORK_STRUCT_INACTIVE do not participate
244 * in pwq->nr_active and all work items in pwq->inactive_works are
245 * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE
246 * work items are in pwq->inactive_works. Some of them are ready to
247 * run in pool->worklist or worker->scheduled. Those work itmes are
248 * only struct wq_barrier which is used for flush_work() and should
249 * not participate in pwq->nr_active. For non-barrier work item, it
250 * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
251 */
252 int nr_active; /* L: nr of active works */
253 int max_active; /* L: max active works */
254 struct list_head inactive_works; /* L: inactive works */
255 struct list_head pwqs_node; /* WR: node on wq->pwqs */
256 struct list_head mayday_node; /* MD: node on wq->maydays */
257
258 u64 stats[PWQ_NR_STATS];
259
260 /*
261 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
262 * and pwq_release_workfn() for details. pool_workqueue itself is also
263 * RCU protected so that the first pwq can be determined without
264 * grabbing wq->mutex.
265 */
266 struct kthread_work release_work;
267 struct rcu_head rcu;
268} __aligned(1 << WORK_STRUCT_FLAG_BITS);
269
270/*
271 * Structure used to wait for workqueue flush.
272 */
273struct wq_flusher {
274 struct list_head list; /* WQ: list of flushers */
275 int flush_color; /* WQ: flush color waiting for */
276 struct completion done; /* flush completion */
277};
278
279struct wq_device;
280
281/*
282 * The externally visible workqueue. It relays the issued work items to
283 * the appropriate worker_pool through its pool_workqueues.
284 */
285struct workqueue_struct {
286 struct list_head pwqs; /* WR: all pwqs of this wq */
287 struct list_head list; /* PR: list of all workqueues */
288
289 struct mutex mutex; /* protects this wq */
290 int work_color; /* WQ: current work color */
291 int flush_color; /* WQ: current flush color */
292 atomic_t nr_pwqs_to_flush; /* flush in progress */
293 struct wq_flusher *first_flusher; /* WQ: first flusher */
294 struct list_head flusher_queue; /* WQ: flush waiters */
295 struct list_head flusher_overflow; /* WQ: flush overflow list */
296
297 struct list_head maydays; /* MD: pwqs requesting rescue */
298 struct worker *rescuer; /* MD: rescue worker */
299
300 int nr_drainers; /* WQ: drain in progress */
301 int saved_max_active; /* WQ: saved pwq max_active */
302
303 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
304 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
305
306#ifdef CONFIG_SYSFS
307 struct wq_device *wq_dev; /* I: for sysfs interface */
308#endif
309#ifdef CONFIG_LOCKDEP
310 char *lock_name;
311 struct lock_class_key key;
312 struct lockdep_map lockdep_map;
313#endif
314 char name[WQ_NAME_LEN]; /* I: workqueue name */
315
316 /*
317 * Destruction of workqueue_struct is RCU protected to allow walking
318 * the workqueues list without grabbing wq_pool_mutex.
319 * This is used to dump all workqueues from sysrq.
320 */
321 struct rcu_head rcu;
322
323 /* hot fields used during command issue, aligned to cacheline */
324 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
325 struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */
326};
327
328static struct kmem_cache *pwq_cache;
329
330/*
331 * Each pod type describes how CPUs should be grouped for unbound workqueues.
332 * See the comment above workqueue_attrs->affn_scope.
333 */
334struct wq_pod_type {
335 int nr_pods; /* number of pods */
336 cpumask_var_t *pod_cpus; /* pod -> cpus */
337 int *pod_node; /* pod -> node */
338 int *cpu_pod; /* cpu -> pod */
339};
340
341static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
342static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
343
344static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
345 [WQ_AFFN_DFL] = "default",
346 [WQ_AFFN_CPU] = "cpu",
347 [WQ_AFFN_SMT] = "smt",
348 [WQ_AFFN_CACHE] = "cache",
349 [WQ_AFFN_NUMA] = "numa",
350 [WQ_AFFN_SYSTEM] = "system",
351};
352
353/*
354 * Per-cpu work items which run for longer than the following threshold are
355 * automatically considered CPU intensive and excluded from concurrency
356 * management to prevent them from noticeably delaying other per-cpu work items.
357 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
358 * The actual value is initialized in wq_cpu_intensive_thresh_init().
359 */
360static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
361module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
362
363/* see the comment above the definition of WQ_POWER_EFFICIENT */
364static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
365module_param_named(power_efficient, wq_power_efficient, bool, 0444);
366
367static bool wq_online; /* can kworkers be created yet? */
368
369/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
370static struct workqueue_attrs *wq_update_pod_attrs_buf;
371
372static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
373static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
374static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
375/* wait for manager to go away */
376static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
377
378static LIST_HEAD(workqueues); /* PR: list of all workqueues */
379static bool workqueue_freezing; /* PL: have wqs started freezing? */
380
381/* PL&A: allowable cpus for unbound wqs and work items */
382static cpumask_var_t wq_unbound_cpumask;
383
384/* PL: user requested unbound cpumask via sysfs */
385static cpumask_var_t wq_requested_unbound_cpumask;
386
387/* PL: isolated cpumask to be excluded from unbound cpumask */
388static cpumask_var_t wq_isolated_cpumask;
389
390/* for further constrain wq_unbound_cpumask by cmdline parameter*/
391static struct cpumask wq_cmdline_cpumask __initdata;
392
393/* CPU where unbound work was last round robin scheduled from this CPU */
394static DEFINE_PER_CPU(int, wq_rr_cpu_last);
395
396/*
397 * Local execution of unbound work items is no longer guaranteed. The
398 * following always forces round-robin CPU selection on unbound work items
399 * to uncover usages which depend on it.
400 */
401#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
402static bool wq_debug_force_rr_cpu = true;
403#else
404static bool wq_debug_force_rr_cpu = false;
405#endif
406module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
407
408/* the per-cpu worker pools */
409static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
410
411static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
412
413/* PL: hash of all unbound pools keyed by pool->attrs */
414static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
415
416/* I: attributes used when instantiating standard unbound pools on demand */
417static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
418
419/* I: attributes used when instantiating ordered pools on demand */
420static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
421
422/*
423 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
424 * process context while holding a pool lock. Bounce to a dedicated kthread
425 * worker to avoid A-A deadlocks.
426 */
427static struct kthread_worker *pwq_release_worker __ro_after_init;
428
429struct workqueue_struct *system_wq __ro_after_init;
430EXPORT_SYMBOL(system_wq);
431struct workqueue_struct *system_highpri_wq __ro_after_init;
432EXPORT_SYMBOL_GPL(system_highpri_wq);
433struct workqueue_struct *system_long_wq __ro_after_init;
434EXPORT_SYMBOL_GPL(system_long_wq);
435struct workqueue_struct *system_unbound_wq __ro_after_init;
436EXPORT_SYMBOL_GPL(system_unbound_wq);
437struct workqueue_struct *system_freezable_wq __ro_after_init;
438EXPORT_SYMBOL_GPL(system_freezable_wq);
439struct workqueue_struct *system_power_efficient_wq __ro_after_init;
440EXPORT_SYMBOL_GPL(system_power_efficient_wq);
441struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
442EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
443
444static int worker_thread(void *__worker);
445static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
446static void show_pwq(struct pool_workqueue *pwq);
447static void show_one_worker_pool(struct worker_pool *pool);
448
449#define CREATE_TRACE_POINTS
450#include <trace/events/workqueue.h>
451
452#define assert_rcu_or_pool_mutex() \
453 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
454 !lockdep_is_held(&wq_pool_mutex), \
455 "RCU or wq_pool_mutex should be held")
456
457#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
458 RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
459 !lockdep_is_held(&wq->mutex) && \
460 !lockdep_is_held(&wq_pool_mutex), \
461 "RCU, wq->mutex or wq_pool_mutex should be held")
462
463#define for_each_cpu_worker_pool(pool, cpu) \
464 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
465 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
466 (pool)++)
467
468/**
469 * for_each_pool - iterate through all worker_pools in the system
470 * @pool: iteration cursor
471 * @pi: integer used for iteration
472 *
473 * This must be called either with wq_pool_mutex held or RCU read
474 * locked. If the pool needs to be used beyond the locking in effect, the
475 * caller is responsible for guaranteeing that the pool stays online.
476 *
477 * The if/else clause exists only for the lockdep assertion and can be
478 * ignored.
479 */
480#define for_each_pool(pool, pi) \
481 idr_for_each_entry(&worker_pool_idr, pool, pi) \
482 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
483 else
484
485/**
486 * for_each_pool_worker - iterate through all workers of a worker_pool
487 * @worker: iteration cursor
488 * @pool: worker_pool to iterate workers of
489 *
490 * This must be called with wq_pool_attach_mutex.
491 *
492 * The if/else clause exists only for the lockdep assertion and can be
493 * ignored.
494 */
495#define for_each_pool_worker(worker, pool) \
496 list_for_each_entry((worker), &(pool)->workers, node) \
497 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
498 else
499
500/**
501 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
502 * @pwq: iteration cursor
503 * @wq: the target workqueue
504 *
505 * This must be called either with wq->mutex held or RCU read locked.
506 * If the pwq needs to be used beyond the locking in effect, the caller is
507 * responsible for guaranteeing that the pwq stays online.
508 *
509 * The if/else clause exists only for the lockdep assertion and can be
510 * ignored.
511 */
512#define for_each_pwq(pwq, wq) \
513 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
514 lockdep_is_held(&(wq->mutex)))
515
516#ifdef CONFIG_DEBUG_OBJECTS_WORK
517
518static const struct debug_obj_descr work_debug_descr;
519
520static void *work_debug_hint(void *addr)
521{
522 return ((struct work_struct *) addr)->func;
523}
524
525static bool work_is_static_object(void *addr)
526{
527 struct work_struct *work = addr;
528
529 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
530}
531
532/*
533 * fixup_init is called when:
534 * - an active object is initialized
535 */
536static bool work_fixup_init(void *addr, enum debug_obj_state state)
537{
538 struct work_struct *work = addr;
539
540 switch (state) {
541 case ODEBUG_STATE_ACTIVE:
542 cancel_work_sync(work);
543 debug_object_init(work, &work_debug_descr);
544 return true;
545 default:
546 return false;
547 }
548}
549
550/*
551 * fixup_free is called when:
552 * - an active object is freed
553 */
554static bool work_fixup_free(void *addr, enum debug_obj_state state)
555{
556 struct work_struct *work = addr;
557
558 switch (state) {
559 case ODEBUG_STATE_ACTIVE:
560 cancel_work_sync(work);
561 debug_object_free(work, &work_debug_descr);
562 return true;
563 default:
564 return false;
565 }
566}
567
568static const struct debug_obj_descr work_debug_descr = {
569 .name = "work_struct",
570 .debug_hint = work_debug_hint,
571 .is_static_object = work_is_static_object,
572 .fixup_init = work_fixup_init,
573 .fixup_free = work_fixup_free,
574};
575
576static inline void debug_work_activate(struct work_struct *work)
577{
578 debug_object_activate(work, &work_debug_descr);
579}
580
581static inline void debug_work_deactivate(struct work_struct *work)
582{
583 debug_object_deactivate(work, &work_debug_descr);
584}
585
586void __init_work(struct work_struct *work, int onstack)
587{
588 if (onstack)
589 debug_object_init_on_stack(work, &work_debug_descr);
590 else
591 debug_object_init(work, &work_debug_descr);
592}
593EXPORT_SYMBOL_GPL(__init_work);
594
595void destroy_work_on_stack(struct work_struct *work)
596{
597 debug_object_free(work, &work_debug_descr);
598}
599EXPORT_SYMBOL_GPL(destroy_work_on_stack);
600
601void destroy_delayed_work_on_stack(struct delayed_work *work)
602{
603 destroy_timer_on_stack(&work->timer);
604 debug_object_free(&work->work, &work_debug_descr);
605}
606EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
607
608#else
609static inline void debug_work_activate(struct work_struct *work) { }
610static inline void debug_work_deactivate(struct work_struct *work) { }
611#endif
612
613/**
614 * worker_pool_assign_id - allocate ID and assign it to @pool
615 * @pool: the pool pointer of interest
616 *
617 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
618 * successfully, -errno on failure.
619 */
620static int worker_pool_assign_id(struct worker_pool *pool)
621{
622 int ret;
623
624 lockdep_assert_held(&wq_pool_mutex);
625
626 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
627 GFP_KERNEL);
628 if (ret >= 0) {
629 pool->id = ret;
630 return 0;
631 }
632 return ret;
633}
634
635static unsigned int work_color_to_flags(int color)
636{
637 return color << WORK_STRUCT_COLOR_SHIFT;
638}
639
640static int get_work_color(unsigned long work_data)
641{
642 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
643 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
644}
645
646static int work_next_color(int color)
647{
648 return (color + 1) % WORK_NR_COLORS;
649}
650
651/*
652 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
653 * contain the pointer to the queued pwq. Once execution starts, the flag
654 * is cleared and the high bits contain OFFQ flags and pool ID.
655 *
656 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
657 * and clear_work_data() can be used to set the pwq, pool or clear
658 * work->data. These functions should only be called while the work is
659 * owned - ie. while the PENDING bit is set.
660 *
661 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
662 * corresponding to a work. Pool is available once the work has been
663 * queued anywhere after initialization until it is sync canceled. pwq is
664 * available only while the work item is queued.
665 *
666 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
667 * canceled. While being canceled, a work item may have its PENDING set
668 * but stay off timer and worklist for arbitrarily long and nobody should
669 * try to steal the PENDING bit.
670 */
671static inline void set_work_data(struct work_struct *work, unsigned long data,
672 unsigned long flags)
673{
674 WARN_ON_ONCE(!work_pending(work));
675 atomic_long_set(&work->data, data | flags | work_static(work));
676}
677
678static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
679 unsigned long extra_flags)
680{
681 set_work_data(work, (unsigned long)pwq,
682 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
683}
684
685static void set_work_pool_and_keep_pending(struct work_struct *work,
686 int pool_id)
687{
688 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
689 WORK_STRUCT_PENDING);
690}
691
692static void set_work_pool_and_clear_pending(struct work_struct *work,
693 int pool_id)
694{
695 /*
696 * The following wmb is paired with the implied mb in
697 * test_and_set_bit(PENDING) and ensures all updates to @work made
698 * here are visible to and precede any updates by the next PENDING
699 * owner.
700 */
701 smp_wmb();
702 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
703 /*
704 * The following mb guarantees that previous clear of a PENDING bit
705 * will not be reordered with any speculative LOADS or STORES from
706 * work->current_func, which is executed afterwards. This possible
707 * reordering can lead to a missed execution on attempt to queue
708 * the same @work. E.g. consider this case:
709 *
710 * CPU#0 CPU#1
711 * ---------------------------- --------------------------------
712 *
713 * 1 STORE event_indicated
714 * 2 queue_work_on() {
715 * 3 test_and_set_bit(PENDING)
716 * 4 } set_..._and_clear_pending() {
717 * 5 set_work_data() # clear bit
718 * 6 smp_mb()
719 * 7 work->current_func() {
720 * 8 LOAD event_indicated
721 * }
722 *
723 * Without an explicit full barrier speculative LOAD on line 8 can
724 * be executed before CPU#0 does STORE on line 1. If that happens,
725 * CPU#0 observes the PENDING bit is still set and new execution of
726 * a @work is not queued in a hope, that CPU#1 will eventually
727 * finish the queued @work. Meanwhile CPU#1 does not see
728 * event_indicated is set, because speculative LOAD was executed
729 * before actual STORE.
730 */
731 smp_mb();
732}
733
734static void clear_work_data(struct work_struct *work)
735{
736 smp_wmb(); /* see set_work_pool_and_clear_pending() */
737 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
738}
739
740static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
741{
742 return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK);
743}
744
745static struct pool_workqueue *get_work_pwq(struct work_struct *work)
746{
747 unsigned long data = atomic_long_read(&work->data);
748
749 if (data & WORK_STRUCT_PWQ)
750 return work_struct_pwq(data);
751 else
752 return NULL;
753}
754
755/**
756 * get_work_pool - return the worker_pool a given work was associated with
757 * @work: the work item of interest
758 *
759 * Pools are created and destroyed under wq_pool_mutex, and allows read
760 * access under RCU read lock. As such, this function should be
761 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
762 *
763 * All fields of the returned pool are accessible as long as the above
764 * mentioned locking is in effect. If the returned pool needs to be used
765 * beyond the critical section, the caller is responsible for ensuring the
766 * returned pool is and stays online.
767 *
768 * Return: The worker_pool @work was last associated with. %NULL if none.
769 */
770static struct worker_pool *get_work_pool(struct work_struct *work)
771{
772 unsigned long data = atomic_long_read(&work->data);
773 int pool_id;
774
775 assert_rcu_or_pool_mutex();
776
777 if (data & WORK_STRUCT_PWQ)
778 return work_struct_pwq(data)->pool;
779
780 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
781 if (pool_id == WORK_OFFQ_POOL_NONE)
782 return NULL;
783
784 return idr_find(&worker_pool_idr, pool_id);
785}
786
787/**
788 * get_work_pool_id - return the worker pool ID a given work is associated with
789 * @work: the work item of interest
790 *
791 * Return: The worker_pool ID @work was last associated with.
792 * %WORK_OFFQ_POOL_NONE if none.
793 */
794static int get_work_pool_id(struct work_struct *work)
795{
796 unsigned long data = atomic_long_read(&work->data);
797
798 if (data & WORK_STRUCT_PWQ)
799 return work_struct_pwq(data)->pool->id;
800
801 return data >> WORK_OFFQ_POOL_SHIFT;
802}
803
804static void mark_work_canceling(struct work_struct *work)
805{
806 unsigned long pool_id = get_work_pool_id(work);
807
808 pool_id <<= WORK_OFFQ_POOL_SHIFT;
809 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
810}
811
812static bool work_is_canceling(struct work_struct *work)
813{
814 unsigned long data = atomic_long_read(&work->data);
815
816 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
817}
818
819/*
820 * Policy functions. These define the policies on how the global worker
821 * pools are managed. Unless noted otherwise, these functions assume that
822 * they're being called with pool->lock held.
823 */
824
825/*
826 * Need to wake up a worker? Called from anything but currently
827 * running workers.
828 *
829 * Note that, because unbound workers never contribute to nr_running, this
830 * function will always return %true for unbound pools as long as the
831 * worklist isn't empty.
832 */
833static bool need_more_worker(struct worker_pool *pool)
834{
835 return !list_empty(&pool->worklist) && !pool->nr_running;
836}
837
838/* Can I start working? Called from busy but !running workers. */
839static bool may_start_working(struct worker_pool *pool)
840{
841 return pool->nr_idle;
842}
843
844/* Do I need to keep working? Called from currently running workers. */
845static bool keep_working(struct worker_pool *pool)
846{
847 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
848}
849
850/* Do we need a new worker? Called from manager. */
851static bool need_to_create_worker(struct worker_pool *pool)
852{
853 return need_more_worker(pool) && !may_start_working(pool);
854}
855
856/* Do we have too many workers and should some go away? */
857static bool too_many_workers(struct worker_pool *pool)
858{
859 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
860 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
861 int nr_busy = pool->nr_workers - nr_idle;
862
863 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
864}
865
866/**
867 * worker_set_flags - set worker flags and adjust nr_running accordingly
868 * @worker: self
869 * @flags: flags to set
870 *
871 * Set @flags in @worker->flags and adjust nr_running accordingly.
872 */
873static inline void worker_set_flags(struct worker *worker, unsigned int flags)
874{
875 struct worker_pool *pool = worker->pool;
876
877 lockdep_assert_held(&pool->lock);
878
879 /* If transitioning into NOT_RUNNING, adjust nr_running. */
880 if ((flags & WORKER_NOT_RUNNING) &&
881 !(worker->flags & WORKER_NOT_RUNNING)) {
882 pool->nr_running--;
883 }
884
885 worker->flags |= flags;
886}
887
888/**
889 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
890 * @worker: self
891 * @flags: flags to clear
892 *
893 * Clear @flags in @worker->flags and adjust nr_running accordingly.
894 */
895static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
896{
897 struct worker_pool *pool = worker->pool;
898 unsigned int oflags = worker->flags;
899
900 lockdep_assert_held(&pool->lock);
901
902 worker->flags &= ~flags;
903
904 /*
905 * If transitioning out of NOT_RUNNING, increment nr_running. Note
906 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
907 * of multiple flags, not a single flag.
908 */
909 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
910 if (!(worker->flags & WORKER_NOT_RUNNING))
911 pool->nr_running++;
912}
913
914/* Return the first idle worker. Called with pool->lock held. */
915static struct worker *first_idle_worker(struct worker_pool *pool)
916{
917 if (unlikely(list_empty(&pool->idle_list)))
918 return NULL;
919
920 return list_first_entry(&pool->idle_list, struct worker, entry);
921}
922
923/**
924 * worker_enter_idle - enter idle state
925 * @worker: worker which is entering idle state
926 *
927 * @worker is entering idle state. Update stats and idle timer if
928 * necessary.
929 *
930 * LOCKING:
931 * raw_spin_lock_irq(pool->lock).
932 */
933static void worker_enter_idle(struct worker *worker)
934{
935 struct worker_pool *pool = worker->pool;
936
937 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
938 WARN_ON_ONCE(!list_empty(&worker->entry) &&
939 (worker->hentry.next || worker->hentry.pprev)))
940 return;
941
942 /* can't use worker_set_flags(), also called from create_worker() */
943 worker->flags |= WORKER_IDLE;
944 pool->nr_idle++;
945 worker->last_active = jiffies;
946
947 /* idle_list is LIFO */
948 list_add(&worker->entry, &pool->idle_list);
949
950 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
951 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
952
953 /* Sanity check nr_running. */
954 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
955}
956
957/**
958 * worker_leave_idle - leave idle state
959 * @worker: worker which is leaving idle state
960 *
961 * @worker is leaving idle state. Update stats.
962 *
963 * LOCKING:
964 * raw_spin_lock_irq(pool->lock).
965 */
966static void worker_leave_idle(struct worker *worker)
967{
968 struct worker_pool *pool = worker->pool;
969
970 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
971 return;
972 worker_clr_flags(worker, WORKER_IDLE);
973 pool->nr_idle--;
974 list_del_init(&worker->entry);
975}
976
977/**
978 * find_worker_executing_work - find worker which is executing a work
979 * @pool: pool of interest
980 * @work: work to find worker for
981 *
982 * Find a worker which is executing @work on @pool by searching
983 * @pool->busy_hash which is keyed by the address of @work. For a worker
984 * to match, its current execution should match the address of @work and
985 * its work function. This is to avoid unwanted dependency between
986 * unrelated work executions through a work item being recycled while still
987 * being executed.
988 *
989 * This is a bit tricky. A work item may be freed once its execution
990 * starts and nothing prevents the freed area from being recycled for
991 * another work item. If the same work item address ends up being reused
992 * before the original execution finishes, workqueue will identify the
993 * recycled work item as currently executing and make it wait until the
994 * current execution finishes, introducing an unwanted dependency.
995 *
996 * This function checks the work item address and work function to avoid
997 * false positives. Note that this isn't complete as one may construct a
998 * work function which can introduce dependency onto itself through a
999 * recycled work item. Well, if somebody wants to shoot oneself in the
1000 * foot that badly, there's only so much we can do, and if such deadlock
1001 * actually occurs, it should be easy to locate the culprit work function.
1002 *
1003 * CONTEXT:
1004 * raw_spin_lock_irq(pool->lock).
1005 *
1006 * Return:
1007 * Pointer to worker which is executing @work if found, %NULL
1008 * otherwise.
1009 */
1010static struct worker *find_worker_executing_work(struct worker_pool *pool,
1011 struct work_struct *work)
1012{
1013 struct worker *worker;
1014
1015 hash_for_each_possible(pool->busy_hash, worker, hentry,
1016 (unsigned long)work)
1017 if (worker->current_work == work &&
1018 worker->current_func == work->func)
1019 return worker;
1020
1021 return NULL;
1022}
1023
1024/**
1025 * move_linked_works - move linked works to a list
1026 * @work: start of series of works to be scheduled
1027 * @head: target list to append @work to
1028 * @nextp: out parameter for nested worklist walking
1029 *
1030 * Schedule linked works starting from @work to @head. Work series to be
1031 * scheduled starts at @work and includes any consecutive work with
1032 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1033 * @nextp.
1034 *
1035 * CONTEXT:
1036 * raw_spin_lock_irq(pool->lock).
1037 */
1038static void move_linked_works(struct work_struct *work, struct list_head *head,
1039 struct work_struct **nextp)
1040{
1041 struct work_struct *n;
1042
1043 /*
1044 * Linked worklist will always end before the end of the list,
1045 * use NULL for list head.
1046 */
1047 list_for_each_entry_safe_from(work, n, NULL, entry) {
1048 list_move_tail(&work->entry, head);
1049 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1050 break;
1051 }
1052
1053 /*
1054 * If we're already inside safe list traversal and have moved
1055 * multiple works to the scheduled queue, the next position
1056 * needs to be updated.
1057 */
1058 if (nextp)
1059 *nextp = n;
1060}
1061
1062/**
1063 * assign_work - assign a work item and its linked work items to a worker
1064 * @work: work to assign
1065 * @worker: worker to assign to
1066 * @nextp: out parameter for nested worklist walking
1067 *
1068 * Assign @work and its linked work items to @worker. If @work is already being
1069 * executed by another worker in the same pool, it'll be punted there.
1070 *
1071 * If @nextp is not NULL, it's updated to point to the next work of the last
1072 * scheduled work. This allows assign_work() to be nested inside
1073 * list_for_each_entry_safe().
1074 *
1075 * Returns %true if @work was successfully assigned to @worker. %false if @work
1076 * was punted to another worker already executing it.
1077 */
1078static bool assign_work(struct work_struct *work, struct worker *worker,
1079 struct work_struct **nextp)
1080{
1081 struct worker_pool *pool = worker->pool;
1082 struct worker *collision;
1083
1084 lockdep_assert_held(&pool->lock);
1085
1086 /*
1087 * A single work shouldn't be executed concurrently by multiple workers.
1088 * __queue_work() ensures that @work doesn't jump to a different pool
1089 * while still running in the previous pool. Here, we should ensure that
1090 * @work is not executed concurrently by multiple workers from the same
1091 * pool. Check whether anyone is already processing the work. If so,
1092 * defer the work to the currently executing one.
1093 */
1094 collision = find_worker_executing_work(pool, work);
1095 if (unlikely(collision)) {
1096 move_linked_works(work, &collision->scheduled, nextp);
1097 return false;
1098 }
1099
1100 move_linked_works(work, &worker->scheduled, nextp);
1101 return true;
1102}
1103
1104/**
1105 * kick_pool - wake up an idle worker if necessary
1106 * @pool: pool to kick
1107 *
1108 * @pool may have pending work items. Wake up worker if necessary. Returns
1109 * whether a worker was woken up.
1110 */
1111static bool kick_pool(struct worker_pool *pool)
1112{
1113 struct worker *worker = first_idle_worker(pool);
1114 struct task_struct *p;
1115
1116 lockdep_assert_held(&pool->lock);
1117
1118 if (!need_more_worker(pool) || !worker)
1119 return false;
1120
1121 p = worker->task;
1122
1123#ifdef CONFIG_SMP
1124 /*
1125 * Idle @worker is about to execute @work and waking up provides an
1126 * opportunity to migrate @worker at a lower cost by setting the task's
1127 * wake_cpu field. Let's see if we want to move @worker to improve
1128 * execution locality.
1129 *
1130 * We're waking the worker that went idle the latest and there's some
1131 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1132 * so, setting the wake_cpu won't do anything. As this is a best-effort
1133 * optimization and the race window is narrow, let's leave as-is for
1134 * now. If this becomes pronounced, we can skip over workers which are
1135 * still on cpu when picking an idle worker.
1136 *
1137 * If @pool has non-strict affinity, @worker might have ended up outside
1138 * its affinity scope. Repatriate.
1139 */
1140 if (!pool->attrs->affn_strict &&
1141 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1142 struct work_struct *work = list_first_entry(&pool->worklist,
1143 struct work_struct, entry);
1144 p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask);
1145 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1146 }
1147#endif
1148 wake_up_process(p);
1149 return true;
1150}
1151
1152#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1153
1154/*
1155 * Concurrency-managed per-cpu work items that hog CPU for longer than
1156 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1157 * which prevents them from stalling other concurrency-managed work items. If a
1158 * work function keeps triggering this mechanism, it's likely that the work item
1159 * should be using an unbound workqueue instead.
1160 *
1161 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1162 * and report them so that they can be examined and converted to use unbound
1163 * workqueues as appropriate. To avoid flooding the console, each violating work
1164 * function is tracked and reported with exponential backoff.
1165 */
1166#define WCI_MAX_ENTS 128
1167
1168struct wci_ent {
1169 work_func_t func;
1170 atomic64_t cnt;
1171 struct hlist_node hash_node;
1172};
1173
1174static struct wci_ent wci_ents[WCI_MAX_ENTS];
1175static int wci_nr_ents;
1176static DEFINE_RAW_SPINLOCK(wci_lock);
1177static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1178
1179static struct wci_ent *wci_find_ent(work_func_t func)
1180{
1181 struct wci_ent *ent;
1182
1183 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1184 (unsigned long)func) {
1185 if (ent->func == func)
1186 return ent;
1187 }
1188 return NULL;
1189}
1190
1191static void wq_cpu_intensive_report(work_func_t func)
1192{
1193 struct wci_ent *ent;
1194
1195restart:
1196 ent = wci_find_ent(func);
1197 if (ent) {
1198 u64 cnt;
1199
1200 /*
1201 * Start reporting from the fourth time and back off
1202 * exponentially.
1203 */
1204 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1205 if (cnt >= 4 && is_power_of_2(cnt))
1206 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1207 ent->func, wq_cpu_intensive_thresh_us,
1208 atomic64_read(&ent->cnt));
1209 return;
1210 }
1211
1212 /*
1213 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1214 * is exhausted, something went really wrong and we probably made enough
1215 * noise already.
1216 */
1217 if (wci_nr_ents >= WCI_MAX_ENTS)
1218 return;
1219
1220 raw_spin_lock(&wci_lock);
1221
1222 if (wci_nr_ents >= WCI_MAX_ENTS) {
1223 raw_spin_unlock(&wci_lock);
1224 return;
1225 }
1226
1227 if (wci_find_ent(func)) {
1228 raw_spin_unlock(&wci_lock);
1229 goto restart;
1230 }
1231
1232 ent = &wci_ents[wci_nr_ents++];
1233 ent->func = func;
1234 atomic64_set(&ent->cnt, 1);
1235 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1236
1237 raw_spin_unlock(&wci_lock);
1238}
1239
1240#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1241static void wq_cpu_intensive_report(work_func_t func) {}
1242#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1243
1244/**
1245 * wq_worker_running - a worker is running again
1246 * @task: task waking up
1247 *
1248 * This function is called when a worker returns from schedule()
1249 */
1250void wq_worker_running(struct task_struct *task)
1251{
1252 struct worker *worker = kthread_data(task);
1253
1254 if (!READ_ONCE(worker->sleeping))
1255 return;
1256
1257 /*
1258 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1259 * and the nr_running increment below, we may ruin the nr_running reset
1260 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1261 * pool. Protect against such race.
1262 */
1263 preempt_disable();
1264 if (!(worker->flags & WORKER_NOT_RUNNING))
1265 worker->pool->nr_running++;
1266 preempt_enable();
1267
1268 /*
1269 * CPU intensive auto-detection cares about how long a work item hogged
1270 * CPU without sleeping. Reset the starting timestamp on wakeup.
1271 */
1272 worker->current_at = worker->task->se.sum_exec_runtime;
1273
1274 WRITE_ONCE(worker->sleeping, 0);
1275}
1276
1277/**
1278 * wq_worker_sleeping - a worker is going to sleep
1279 * @task: task going to sleep
1280 *
1281 * This function is called from schedule() when a busy worker is
1282 * going to sleep.
1283 */
1284void wq_worker_sleeping(struct task_struct *task)
1285{
1286 struct worker *worker = kthread_data(task);
1287 struct worker_pool *pool;
1288
1289 /*
1290 * Rescuers, which may not have all the fields set up like normal
1291 * workers, also reach here, let's not access anything before
1292 * checking NOT_RUNNING.
1293 */
1294 if (worker->flags & WORKER_NOT_RUNNING)
1295 return;
1296
1297 pool = worker->pool;
1298
1299 /* Return if preempted before wq_worker_running() was reached */
1300 if (READ_ONCE(worker->sleeping))
1301 return;
1302
1303 WRITE_ONCE(worker->sleeping, 1);
1304 raw_spin_lock_irq(&pool->lock);
1305
1306 /*
1307 * Recheck in case unbind_workers() preempted us. We don't
1308 * want to decrement nr_running after the worker is unbound
1309 * and nr_running has been reset.
1310 */
1311 if (worker->flags & WORKER_NOT_RUNNING) {
1312 raw_spin_unlock_irq(&pool->lock);
1313 return;
1314 }
1315
1316 pool->nr_running--;
1317 if (kick_pool(pool))
1318 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1319
1320 raw_spin_unlock_irq(&pool->lock);
1321}
1322
1323/**
1324 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1325 * @task: task currently running
1326 *
1327 * Called from scheduler_tick(). We're in the IRQ context and the current
1328 * worker's fields which follow the 'K' locking rule can be accessed safely.
1329 */
1330void wq_worker_tick(struct task_struct *task)
1331{
1332 struct worker *worker = kthread_data(task);
1333 struct pool_workqueue *pwq = worker->current_pwq;
1334 struct worker_pool *pool = worker->pool;
1335
1336 if (!pwq)
1337 return;
1338
1339 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1340
1341 if (!wq_cpu_intensive_thresh_us)
1342 return;
1343
1344 /*
1345 * If the current worker is concurrency managed and hogged the CPU for
1346 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1347 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1348 *
1349 * Set @worker->sleeping means that @worker is in the process of
1350 * switching out voluntarily and won't be contributing to
1351 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1352 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1353 * double decrements. The task is releasing the CPU anyway. Let's skip.
1354 * We probably want to make this prettier in the future.
1355 */
1356 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1357 worker->task->se.sum_exec_runtime - worker->current_at <
1358 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1359 return;
1360
1361 raw_spin_lock(&pool->lock);
1362
1363 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1364 wq_cpu_intensive_report(worker->current_func);
1365 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1366
1367 if (kick_pool(pool))
1368 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1369
1370 raw_spin_unlock(&pool->lock);
1371}
1372
1373/**
1374 * wq_worker_last_func - retrieve worker's last work function
1375 * @task: Task to retrieve last work function of.
1376 *
1377 * Determine the last function a worker executed. This is called from
1378 * the scheduler to get a worker's last known identity.
1379 *
1380 * CONTEXT:
1381 * raw_spin_lock_irq(rq->lock)
1382 *
1383 * This function is called during schedule() when a kworker is going
1384 * to sleep. It's used by psi to identify aggregation workers during
1385 * dequeuing, to allow periodic aggregation to shut-off when that
1386 * worker is the last task in the system or cgroup to go to sleep.
1387 *
1388 * As this function doesn't involve any workqueue-related locking, it
1389 * only returns stable values when called from inside the scheduler's
1390 * queuing and dequeuing paths, when @task, which must be a kworker,
1391 * is guaranteed to not be processing any works.
1392 *
1393 * Return:
1394 * The last work function %current executed as a worker, NULL if it
1395 * hasn't executed any work yet.
1396 */
1397work_func_t wq_worker_last_func(struct task_struct *task)
1398{
1399 struct worker *worker = kthread_data(task);
1400
1401 return worker->last_func;
1402}
1403
1404/**
1405 * get_pwq - get an extra reference on the specified pool_workqueue
1406 * @pwq: pool_workqueue to get
1407 *
1408 * Obtain an extra reference on @pwq. The caller should guarantee that
1409 * @pwq has positive refcnt and be holding the matching pool->lock.
1410 */
1411static void get_pwq(struct pool_workqueue *pwq)
1412{
1413 lockdep_assert_held(&pwq->pool->lock);
1414 WARN_ON_ONCE(pwq->refcnt <= 0);
1415 pwq->refcnt++;
1416}
1417
1418/**
1419 * put_pwq - put a pool_workqueue reference
1420 * @pwq: pool_workqueue to put
1421 *
1422 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1423 * destruction. The caller should be holding the matching pool->lock.
1424 */
1425static void put_pwq(struct pool_workqueue *pwq)
1426{
1427 lockdep_assert_held(&pwq->pool->lock);
1428 if (likely(--pwq->refcnt))
1429 return;
1430 /*
1431 * @pwq can't be released under pool->lock, bounce to a dedicated
1432 * kthread_worker to avoid A-A deadlocks.
1433 */
1434 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1435}
1436
1437/**
1438 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1439 * @pwq: pool_workqueue to put (can be %NULL)
1440 *
1441 * put_pwq() with locking. This function also allows %NULL @pwq.
1442 */
1443static void put_pwq_unlocked(struct pool_workqueue *pwq)
1444{
1445 if (pwq) {
1446 /*
1447 * As both pwqs and pools are RCU protected, the
1448 * following lock operations are safe.
1449 */
1450 raw_spin_lock_irq(&pwq->pool->lock);
1451 put_pwq(pwq);
1452 raw_spin_unlock_irq(&pwq->pool->lock);
1453 }
1454}
1455
1456static void pwq_activate_inactive_work(struct work_struct *work)
1457{
1458 struct pool_workqueue *pwq = get_work_pwq(work);
1459
1460 trace_workqueue_activate_work(work);
1461 if (list_empty(&pwq->pool->worklist))
1462 pwq->pool->watchdog_ts = jiffies;
1463 move_linked_works(work, &pwq->pool->worklist, NULL);
1464 __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work));
1465 pwq->nr_active++;
1466}
1467
1468static void pwq_activate_first_inactive(struct pool_workqueue *pwq)
1469{
1470 struct work_struct *work = list_first_entry(&pwq->inactive_works,
1471 struct work_struct, entry);
1472
1473 pwq_activate_inactive_work(work);
1474}
1475
1476/**
1477 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1478 * @pwq: pwq of interest
1479 * @work_data: work_data of work which left the queue
1480 *
1481 * A work either has completed or is removed from pending queue,
1482 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1483 *
1484 * CONTEXT:
1485 * raw_spin_lock_irq(pool->lock).
1486 */
1487static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1488{
1489 int color = get_work_color(work_data);
1490
1491 if (!(work_data & WORK_STRUCT_INACTIVE)) {
1492 pwq->nr_active--;
1493 if (!list_empty(&pwq->inactive_works)) {
1494 /* one down, submit an inactive one */
1495 if (pwq->nr_active < pwq->max_active)
1496 pwq_activate_first_inactive(pwq);
1497 }
1498 }
1499
1500 pwq->nr_in_flight[color]--;
1501
1502 /* is flush in progress and are we at the flushing tip? */
1503 if (likely(pwq->flush_color != color))
1504 goto out_put;
1505
1506 /* are there still in-flight works? */
1507 if (pwq->nr_in_flight[color])
1508 goto out_put;
1509
1510 /* this pwq is done, clear flush_color */
1511 pwq->flush_color = -1;
1512
1513 /*
1514 * If this was the last pwq, wake up the first flusher. It
1515 * will handle the rest.
1516 */
1517 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1518 complete(&pwq->wq->first_flusher->done);
1519out_put:
1520 put_pwq(pwq);
1521}
1522
1523/**
1524 * try_to_grab_pending - steal work item from worklist and disable irq
1525 * @work: work item to steal
1526 * @is_dwork: @work is a delayed_work
1527 * @flags: place to store irq state
1528 *
1529 * Try to grab PENDING bit of @work. This function can handle @work in any
1530 * stable state - idle, on timer or on worklist.
1531 *
1532 * Return:
1533 *
1534 * ======== ================================================================
1535 * 1 if @work was pending and we successfully stole PENDING
1536 * 0 if @work was idle and we claimed PENDING
1537 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1538 * -ENOENT if someone else is canceling @work, this state may persist
1539 * for arbitrarily long
1540 * ======== ================================================================
1541 *
1542 * Note:
1543 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1544 * interrupted while holding PENDING and @work off queue, irq must be
1545 * disabled on entry. This, combined with delayed_work->timer being
1546 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1547 *
1548 * On successful return, >= 0, irq is disabled and the caller is
1549 * responsible for releasing it using local_irq_restore(*@flags).
1550 *
1551 * This function is safe to call from any context including IRQ handler.
1552 */
1553static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1554 unsigned long *flags)
1555{
1556 struct worker_pool *pool;
1557 struct pool_workqueue *pwq;
1558
1559 local_irq_save(*flags);
1560
1561 /* try to steal the timer if it exists */
1562 if (is_dwork) {
1563 struct delayed_work *dwork = to_delayed_work(work);
1564
1565 /*
1566 * dwork->timer is irqsafe. If del_timer() fails, it's
1567 * guaranteed that the timer is not queued anywhere and not
1568 * running on the local CPU.
1569 */
1570 if (likely(del_timer(&dwork->timer)))
1571 return 1;
1572 }
1573
1574 /* try to claim PENDING the normal way */
1575 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1576 return 0;
1577
1578 rcu_read_lock();
1579 /*
1580 * The queueing is in progress, or it is already queued. Try to
1581 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1582 */
1583 pool = get_work_pool(work);
1584 if (!pool)
1585 goto fail;
1586
1587 raw_spin_lock(&pool->lock);
1588 /*
1589 * work->data is guaranteed to point to pwq only while the work
1590 * item is queued on pwq->wq, and both updating work->data to point
1591 * to pwq on queueing and to pool on dequeueing are done under
1592 * pwq->pool->lock. This in turn guarantees that, if work->data
1593 * points to pwq which is associated with a locked pool, the work
1594 * item is currently queued on that pool.
1595 */
1596 pwq = get_work_pwq(work);
1597 if (pwq && pwq->pool == pool) {
1598 debug_work_deactivate(work);
1599
1600 /*
1601 * A cancelable inactive work item must be in the
1602 * pwq->inactive_works since a queued barrier can't be
1603 * canceled (see the comments in insert_wq_barrier()).
1604 *
1605 * An inactive work item cannot be grabbed directly because
1606 * it might have linked barrier work items which, if left
1607 * on the inactive_works list, will confuse pwq->nr_active
1608 * management later on and cause stall. Make sure the work
1609 * item is activated before grabbing.
1610 */
1611 if (*work_data_bits(work) & WORK_STRUCT_INACTIVE)
1612 pwq_activate_inactive_work(work);
1613
1614 list_del_init(&work->entry);
1615 pwq_dec_nr_in_flight(pwq, *work_data_bits(work));
1616
1617 /* work->data points to pwq iff queued, point to pool */
1618 set_work_pool_and_keep_pending(work, pool->id);
1619
1620 raw_spin_unlock(&pool->lock);
1621 rcu_read_unlock();
1622 return 1;
1623 }
1624 raw_spin_unlock(&pool->lock);
1625fail:
1626 rcu_read_unlock();
1627 local_irq_restore(*flags);
1628 if (work_is_canceling(work))
1629 return -ENOENT;
1630 cpu_relax();
1631 return -EAGAIN;
1632}
1633
1634/**
1635 * insert_work - insert a work into a pool
1636 * @pwq: pwq @work belongs to
1637 * @work: work to insert
1638 * @head: insertion point
1639 * @extra_flags: extra WORK_STRUCT_* flags to set
1640 *
1641 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1642 * work_struct flags.
1643 *
1644 * CONTEXT:
1645 * raw_spin_lock_irq(pool->lock).
1646 */
1647static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1648 struct list_head *head, unsigned int extra_flags)
1649{
1650 debug_work_activate(work);
1651
1652 /* record the work call stack in order to print it in KASAN reports */
1653 kasan_record_aux_stack_noalloc(work);
1654
1655 /* we own @work, set data and link */
1656 set_work_pwq(work, pwq, extra_flags);
1657 list_add_tail(&work->entry, head);
1658 get_pwq(pwq);
1659}
1660
1661/*
1662 * Test whether @work is being queued from another work executing on the
1663 * same workqueue.
1664 */
1665static bool is_chained_work(struct workqueue_struct *wq)
1666{
1667 struct worker *worker;
1668
1669 worker = current_wq_worker();
1670 /*
1671 * Return %true iff I'm a worker executing a work item on @wq. If
1672 * I'm @worker, it's safe to dereference it without locking.
1673 */
1674 return worker && worker->current_pwq->wq == wq;
1675}
1676
1677/*
1678 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1679 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1680 * avoid perturbing sensitive tasks.
1681 */
1682static int wq_select_unbound_cpu(int cpu)
1683{
1684 int new_cpu;
1685
1686 if (likely(!wq_debug_force_rr_cpu)) {
1687 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1688 return cpu;
1689 } else {
1690 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
1691 }
1692
1693 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1694 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1695 if (unlikely(new_cpu >= nr_cpu_ids)) {
1696 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1697 if (unlikely(new_cpu >= nr_cpu_ids))
1698 return cpu;
1699 }
1700 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1701
1702 return new_cpu;
1703}
1704
1705static void __queue_work(int cpu, struct workqueue_struct *wq,
1706 struct work_struct *work)
1707{
1708 struct pool_workqueue *pwq;
1709 struct worker_pool *last_pool, *pool;
1710 unsigned int work_flags;
1711 unsigned int req_cpu = cpu;
1712
1713 /*
1714 * While a work item is PENDING && off queue, a task trying to
1715 * steal the PENDING will busy-loop waiting for it to either get
1716 * queued or lose PENDING. Grabbing PENDING and queueing should
1717 * happen with IRQ disabled.
1718 */
1719 lockdep_assert_irqs_disabled();
1720
1721
1722 /*
1723 * For a draining wq, only works from the same workqueue are
1724 * allowed. The __WQ_DESTROYING helps to spot the issue that
1725 * queues a new work item to a wq after destroy_workqueue(wq).
1726 */
1727 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
1728 WARN_ON_ONCE(!is_chained_work(wq))))
1729 return;
1730 rcu_read_lock();
1731retry:
1732 /* pwq which will be used unless @work is executing elsewhere */
1733 if (req_cpu == WORK_CPU_UNBOUND) {
1734 if (wq->flags & WQ_UNBOUND)
1735 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1736 else
1737 cpu = raw_smp_processor_id();
1738 }
1739
1740 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
1741 pool = pwq->pool;
1742
1743 /*
1744 * If @work was previously on a different pool, it might still be
1745 * running there, in which case the work needs to be queued on that
1746 * pool to guarantee non-reentrancy.
1747 */
1748 last_pool = get_work_pool(work);
1749 if (last_pool && last_pool != pool) {
1750 struct worker *worker;
1751
1752 raw_spin_lock(&last_pool->lock);
1753
1754 worker = find_worker_executing_work(last_pool, work);
1755
1756 if (worker && worker->current_pwq->wq == wq) {
1757 pwq = worker->current_pwq;
1758 pool = pwq->pool;
1759 WARN_ON_ONCE(pool != last_pool);
1760 } else {
1761 /* meh... not running there, queue here */
1762 raw_spin_unlock(&last_pool->lock);
1763 raw_spin_lock(&pool->lock);
1764 }
1765 } else {
1766 raw_spin_lock(&pool->lock);
1767 }
1768
1769 /*
1770 * pwq is determined and locked. For unbound pools, we could have raced
1771 * with pwq release and it could already be dead. If its refcnt is zero,
1772 * repeat pwq selection. Note that unbound pwqs never die without
1773 * another pwq replacing it in cpu_pwq or while work items are executing
1774 * on it, so the retrying is guaranteed to make forward-progress.
1775 */
1776 if (unlikely(!pwq->refcnt)) {
1777 if (wq->flags & WQ_UNBOUND) {
1778 raw_spin_unlock(&pool->lock);
1779 cpu_relax();
1780 goto retry;
1781 }
1782 /* oops */
1783 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1784 wq->name, cpu);
1785 }
1786
1787 /* pwq determined, queue */
1788 trace_workqueue_queue_work(req_cpu, pwq, work);
1789
1790 if (WARN_ON(!list_empty(&work->entry)))
1791 goto out;
1792
1793 pwq->nr_in_flight[pwq->work_color]++;
1794 work_flags = work_color_to_flags(pwq->work_color);
1795
1796 if (likely(pwq->nr_active < pwq->max_active)) {
1797 if (list_empty(&pool->worklist))
1798 pool->watchdog_ts = jiffies;
1799
1800 trace_workqueue_activate_work(work);
1801 pwq->nr_active++;
1802 insert_work(pwq, work, &pool->worklist, work_flags);
1803 kick_pool(pool);
1804 } else {
1805 work_flags |= WORK_STRUCT_INACTIVE;
1806 insert_work(pwq, work, &pwq->inactive_works, work_flags);
1807 }
1808
1809out:
1810 raw_spin_unlock(&pool->lock);
1811 rcu_read_unlock();
1812}
1813
1814/**
1815 * queue_work_on - queue work on specific cpu
1816 * @cpu: CPU number to execute work on
1817 * @wq: workqueue to use
1818 * @work: work to queue
1819 *
1820 * We queue the work to a specific CPU, the caller must ensure it
1821 * can't go away. Callers that fail to ensure that the specified
1822 * CPU cannot go away will execute on a randomly chosen CPU.
1823 * But note well that callers specifying a CPU that never has been
1824 * online will get a splat.
1825 *
1826 * Return: %false if @work was already on a queue, %true otherwise.
1827 */
1828bool queue_work_on(int cpu, struct workqueue_struct *wq,
1829 struct work_struct *work)
1830{
1831 bool ret = false;
1832 unsigned long flags;
1833
1834 local_irq_save(flags);
1835
1836 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1837 __queue_work(cpu, wq, work);
1838 ret = true;
1839 }
1840
1841 local_irq_restore(flags);
1842 return ret;
1843}
1844EXPORT_SYMBOL(queue_work_on);
1845
1846/**
1847 * select_numa_node_cpu - Select a CPU based on NUMA node
1848 * @node: NUMA node ID that we want to select a CPU from
1849 *
1850 * This function will attempt to find a "random" cpu available on a given
1851 * node. If there are no CPUs available on the given node it will return
1852 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1853 * available CPU if we need to schedule this work.
1854 */
1855static int select_numa_node_cpu(int node)
1856{
1857 int cpu;
1858
1859 /* Delay binding to CPU if node is not valid or online */
1860 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1861 return WORK_CPU_UNBOUND;
1862
1863 /* Use local node/cpu if we are already there */
1864 cpu = raw_smp_processor_id();
1865 if (node == cpu_to_node(cpu))
1866 return cpu;
1867
1868 /* Use "random" otherwise know as "first" online CPU of node */
1869 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1870
1871 /* If CPU is valid return that, otherwise just defer */
1872 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1873}
1874
1875/**
1876 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1877 * @node: NUMA node that we are targeting the work for
1878 * @wq: workqueue to use
1879 * @work: work to queue
1880 *
1881 * We queue the work to a "random" CPU within a given NUMA node. The basic
1882 * idea here is to provide a way to somehow associate work with a given
1883 * NUMA node.
1884 *
1885 * This function will only make a best effort attempt at getting this onto
1886 * the right NUMA node. If no node is requested or the requested node is
1887 * offline then we just fall back to standard queue_work behavior.
1888 *
1889 * Currently the "random" CPU ends up being the first available CPU in the
1890 * intersection of cpu_online_mask and the cpumask of the node, unless we
1891 * are running on the node. In that case we just use the current CPU.
1892 *
1893 * Return: %false if @work was already on a queue, %true otherwise.
1894 */
1895bool queue_work_node(int node, struct workqueue_struct *wq,
1896 struct work_struct *work)
1897{
1898 unsigned long flags;
1899 bool ret = false;
1900
1901 /*
1902 * This current implementation is specific to unbound workqueues.
1903 * Specifically we only return the first available CPU for a given
1904 * node instead of cycling through individual CPUs within the node.
1905 *
1906 * If this is used with a per-cpu workqueue then the logic in
1907 * workqueue_select_cpu_near would need to be updated to allow for
1908 * some round robin type logic.
1909 */
1910 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1911
1912 local_irq_save(flags);
1913
1914 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1915 int cpu = select_numa_node_cpu(node);
1916
1917 __queue_work(cpu, wq, work);
1918 ret = true;
1919 }
1920
1921 local_irq_restore(flags);
1922 return ret;
1923}
1924EXPORT_SYMBOL_GPL(queue_work_node);
1925
1926void delayed_work_timer_fn(struct timer_list *t)
1927{
1928 struct delayed_work *dwork = from_timer(dwork, t, timer);
1929
1930 /* should have been called from irqsafe timer with irq already off */
1931 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1932}
1933EXPORT_SYMBOL(delayed_work_timer_fn);
1934
1935static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1936 struct delayed_work *dwork, unsigned long delay)
1937{
1938 struct timer_list *timer = &dwork->timer;
1939 struct work_struct *work = &dwork->work;
1940
1941 WARN_ON_ONCE(!wq);
1942 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1943 WARN_ON_ONCE(timer_pending(timer));
1944 WARN_ON_ONCE(!list_empty(&work->entry));
1945
1946 /*
1947 * If @delay is 0, queue @dwork->work immediately. This is for
1948 * both optimization and correctness. The earliest @timer can
1949 * expire is on the closest next tick and delayed_work users depend
1950 * on that there's no such delay when @delay is 0.
1951 */
1952 if (!delay) {
1953 __queue_work(cpu, wq, &dwork->work);
1954 return;
1955 }
1956
1957 dwork->wq = wq;
1958 dwork->cpu = cpu;
1959 timer->expires = jiffies + delay;
1960
1961 if (unlikely(cpu != WORK_CPU_UNBOUND))
1962 add_timer_on(timer, cpu);
1963 else
1964 add_timer(timer);
1965}
1966
1967/**
1968 * queue_delayed_work_on - queue work on specific CPU after delay
1969 * @cpu: CPU number to execute work on
1970 * @wq: workqueue to use
1971 * @dwork: work to queue
1972 * @delay: number of jiffies to wait before queueing
1973 *
1974 * Return: %false if @work was already on a queue, %true otherwise. If
1975 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1976 * execution.
1977 */
1978bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1979 struct delayed_work *dwork, unsigned long delay)
1980{
1981 struct work_struct *work = &dwork->work;
1982 bool ret = false;
1983 unsigned long flags;
1984
1985 /* read the comment in __queue_work() */
1986 local_irq_save(flags);
1987
1988 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1989 __queue_delayed_work(cpu, wq, dwork, delay);
1990 ret = true;
1991 }
1992
1993 local_irq_restore(flags);
1994 return ret;
1995}
1996EXPORT_SYMBOL(queue_delayed_work_on);
1997
1998/**
1999 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2000 * @cpu: CPU number to execute work on
2001 * @wq: workqueue to use
2002 * @dwork: work to queue
2003 * @delay: number of jiffies to wait before queueing
2004 *
2005 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2006 * modify @dwork's timer so that it expires after @delay. If @delay is
2007 * zero, @work is guaranteed to be scheduled immediately regardless of its
2008 * current state.
2009 *
2010 * Return: %false if @dwork was idle and queued, %true if @dwork was
2011 * pending and its timer was modified.
2012 *
2013 * This function is safe to call from any context including IRQ handler.
2014 * See try_to_grab_pending() for details.
2015 */
2016bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2017 struct delayed_work *dwork, unsigned long delay)
2018{
2019 unsigned long flags;
2020 int ret;
2021
2022 do {
2023 ret = try_to_grab_pending(&dwork->work, true, &flags);
2024 } while (unlikely(ret == -EAGAIN));
2025
2026 if (likely(ret >= 0)) {
2027 __queue_delayed_work(cpu, wq, dwork, delay);
2028 local_irq_restore(flags);
2029 }
2030
2031 /* -ENOENT from try_to_grab_pending() becomes %true */
2032 return ret;
2033}
2034EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2035
2036static void rcu_work_rcufn(struct rcu_head *rcu)
2037{
2038 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2039
2040 /* read the comment in __queue_work() */
2041 local_irq_disable();
2042 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2043 local_irq_enable();
2044}
2045
2046/**
2047 * queue_rcu_work - queue work after a RCU grace period
2048 * @wq: workqueue to use
2049 * @rwork: work to queue
2050 *
2051 * Return: %false if @rwork was already pending, %true otherwise. Note
2052 * that a full RCU grace period is guaranteed only after a %true return.
2053 * While @rwork is guaranteed to be executed after a %false return, the
2054 * execution may happen before a full RCU grace period has passed.
2055 */
2056bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2057{
2058 struct work_struct *work = &rwork->work;
2059
2060 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
2061 rwork->wq = wq;
2062 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2063 return true;
2064 }
2065
2066 return false;
2067}
2068EXPORT_SYMBOL(queue_rcu_work);
2069
2070static struct worker *alloc_worker(int node)
2071{
2072 struct worker *worker;
2073
2074 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2075 if (worker) {
2076 INIT_LIST_HEAD(&worker->entry);
2077 INIT_LIST_HEAD(&worker->scheduled);
2078 INIT_LIST_HEAD(&worker->node);
2079 /* on creation a worker is in !idle && prep state */
2080 worker->flags = WORKER_PREP;
2081 }
2082 return worker;
2083}
2084
2085static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2086{
2087 if (pool->cpu < 0 && pool->attrs->affn_strict)
2088 return pool->attrs->__pod_cpumask;
2089 else
2090 return pool->attrs->cpumask;
2091}
2092
2093/**
2094 * worker_attach_to_pool() - attach a worker to a pool
2095 * @worker: worker to be attached
2096 * @pool: the target pool
2097 *
2098 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2099 * cpu-binding of @worker are kept coordinated with the pool across
2100 * cpu-[un]hotplugs.
2101 */
2102static void worker_attach_to_pool(struct worker *worker,
2103 struct worker_pool *pool)
2104{
2105 mutex_lock(&wq_pool_attach_mutex);
2106
2107 /*
2108 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
2109 * stable across this function. See the comments above the flag
2110 * definition for details.
2111 */
2112 if (pool->flags & POOL_DISASSOCIATED)
2113 worker->flags |= WORKER_UNBOUND;
2114 else
2115 kthread_set_per_cpu(worker->task, pool->cpu);
2116
2117 if (worker->rescue_wq)
2118 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2119
2120 list_add_tail(&worker->node, &pool->workers);
2121 worker->pool = pool;
2122
2123 mutex_unlock(&wq_pool_attach_mutex);
2124}
2125
2126/**
2127 * worker_detach_from_pool() - detach a worker from its pool
2128 * @worker: worker which is attached to its pool
2129 *
2130 * Undo the attaching which had been done in worker_attach_to_pool(). The
2131 * caller worker shouldn't access to the pool after detached except it has
2132 * other reference to the pool.
2133 */
2134static void worker_detach_from_pool(struct worker *worker)
2135{
2136 struct worker_pool *pool = worker->pool;
2137 struct completion *detach_completion = NULL;
2138
2139 mutex_lock(&wq_pool_attach_mutex);
2140
2141 kthread_set_per_cpu(worker->task, -1);
2142 list_del(&worker->node);
2143 worker->pool = NULL;
2144
2145 if (list_empty(&pool->workers) && list_empty(&pool->dying_workers))
2146 detach_completion = pool->detach_completion;
2147 mutex_unlock(&wq_pool_attach_mutex);
2148
2149 /* clear leftover flags without pool->lock after it is detached */
2150 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2151
2152 if (detach_completion)
2153 complete(detach_completion);
2154}
2155
2156/**
2157 * create_worker - create a new workqueue worker
2158 * @pool: pool the new worker will belong to
2159 *
2160 * Create and start a new worker which is attached to @pool.
2161 *
2162 * CONTEXT:
2163 * Might sleep. Does GFP_KERNEL allocations.
2164 *
2165 * Return:
2166 * Pointer to the newly created worker.
2167 */
2168static struct worker *create_worker(struct worker_pool *pool)
2169{
2170 struct worker *worker;
2171 int id;
2172 char id_buf[23];
2173
2174 /* ID is needed to determine kthread name */
2175 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2176 if (id < 0) {
2177 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2178 ERR_PTR(id));
2179 return NULL;
2180 }
2181
2182 worker = alloc_worker(pool->node);
2183 if (!worker) {
2184 pr_err_once("workqueue: Failed to allocate a worker\n");
2185 goto fail;
2186 }
2187
2188 worker->id = id;
2189
2190 if (pool->cpu >= 0)
2191 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
2192 pool->attrs->nice < 0 ? "H" : "");
2193 else
2194 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
2195
2196 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
2197 "kworker/%s", id_buf);
2198 if (IS_ERR(worker->task)) {
2199 if (PTR_ERR(worker->task) == -EINTR) {
2200 pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n",
2201 id_buf);
2202 } else {
2203 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2204 worker->task);
2205 }
2206 goto fail;
2207 }
2208
2209 set_user_nice(worker->task, pool->attrs->nice);
2210 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2211
2212 /* successful, attach the worker to the pool */
2213 worker_attach_to_pool(worker, pool);
2214
2215 /* start the newly created worker */
2216 raw_spin_lock_irq(&pool->lock);
2217
2218 worker->pool->nr_workers++;
2219 worker_enter_idle(worker);
2220 kick_pool(pool);
2221
2222 /*
2223 * @worker is waiting on a completion in kthread() and will trigger hung
2224 * check if not woken up soon. As kick_pool() might not have waken it
2225 * up, wake it up explicitly once more.
2226 */
2227 wake_up_process(worker->task);
2228
2229 raw_spin_unlock_irq(&pool->lock);
2230
2231 return worker;
2232
2233fail:
2234 ida_free(&pool->worker_ida, id);
2235 kfree(worker);
2236 return NULL;
2237}
2238
2239static void unbind_worker(struct worker *worker)
2240{
2241 lockdep_assert_held(&wq_pool_attach_mutex);
2242
2243 kthread_set_per_cpu(worker->task, -1);
2244 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2245 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2246 else
2247 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2248}
2249
2250static void wake_dying_workers(struct list_head *cull_list)
2251{
2252 struct worker *worker, *tmp;
2253
2254 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2255 list_del_init(&worker->entry);
2256 unbind_worker(worker);
2257 /*
2258 * If the worker was somehow already running, then it had to be
2259 * in pool->idle_list when set_worker_dying() happened or we
2260 * wouldn't have gotten here.
2261 *
2262 * Thus, the worker must either have observed the WORKER_DIE
2263 * flag, or have set its state to TASK_IDLE. Either way, the
2264 * below will be observed by the worker and is safe to do
2265 * outside of pool->lock.
2266 */
2267 wake_up_process(worker->task);
2268 }
2269}
2270
2271/**
2272 * set_worker_dying - Tag a worker for destruction
2273 * @worker: worker to be destroyed
2274 * @list: transfer worker away from its pool->idle_list and into list
2275 *
2276 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2277 * should be idle.
2278 *
2279 * CONTEXT:
2280 * raw_spin_lock_irq(pool->lock).
2281 */
2282static void set_worker_dying(struct worker *worker, struct list_head *list)
2283{
2284 struct worker_pool *pool = worker->pool;
2285
2286 lockdep_assert_held(&pool->lock);
2287 lockdep_assert_held(&wq_pool_attach_mutex);
2288
2289 /* sanity check frenzy */
2290 if (WARN_ON(worker->current_work) ||
2291 WARN_ON(!list_empty(&worker->scheduled)) ||
2292 WARN_ON(!(worker->flags & WORKER_IDLE)))
2293 return;
2294
2295 pool->nr_workers--;
2296 pool->nr_idle--;
2297
2298 worker->flags |= WORKER_DIE;
2299
2300 list_move(&worker->entry, list);
2301 list_move(&worker->node, &pool->dying_workers);
2302}
2303
2304/**
2305 * idle_worker_timeout - check if some idle workers can now be deleted.
2306 * @t: The pool's idle_timer that just expired
2307 *
2308 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2309 * worker_leave_idle(), as a worker flicking between idle and active while its
2310 * pool is at the too_many_workers() tipping point would cause too much timer
2311 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2312 * it expire and re-evaluate things from there.
2313 */
2314static void idle_worker_timeout(struct timer_list *t)
2315{
2316 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2317 bool do_cull = false;
2318
2319 if (work_pending(&pool->idle_cull_work))
2320 return;
2321
2322 raw_spin_lock_irq(&pool->lock);
2323
2324 if (too_many_workers(pool)) {
2325 struct worker *worker;
2326 unsigned long expires;
2327
2328 /* idle_list is kept in LIFO order, check the last one */
2329 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2330 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2331 do_cull = !time_before(jiffies, expires);
2332
2333 if (!do_cull)
2334 mod_timer(&pool->idle_timer, expires);
2335 }
2336 raw_spin_unlock_irq(&pool->lock);
2337
2338 if (do_cull)
2339 queue_work(system_unbound_wq, &pool->idle_cull_work);
2340}
2341
2342/**
2343 * idle_cull_fn - cull workers that have been idle for too long.
2344 * @work: the pool's work for handling these idle workers
2345 *
2346 * This goes through a pool's idle workers and gets rid of those that have been
2347 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2348 *
2349 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2350 * culled, so this also resets worker affinity. This requires a sleepable
2351 * context, hence the split between timer callback and work item.
2352 */
2353static void idle_cull_fn(struct work_struct *work)
2354{
2355 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2356 LIST_HEAD(cull_list);
2357
2358 /*
2359 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2360 * cannot proceed beyong worker_detach_from_pool() in its self-destruct
2361 * path. This is required as a previously-preempted worker could run after
2362 * set_worker_dying() has happened but before wake_dying_workers() did.
2363 */
2364 mutex_lock(&wq_pool_attach_mutex);
2365 raw_spin_lock_irq(&pool->lock);
2366
2367 while (too_many_workers(pool)) {
2368 struct worker *worker;
2369 unsigned long expires;
2370
2371 worker = list_entry(pool->idle_list.prev, struct worker, entry);
2372 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2373
2374 if (time_before(jiffies, expires)) {
2375 mod_timer(&pool->idle_timer, expires);
2376 break;
2377 }
2378
2379 set_worker_dying(worker, &cull_list);
2380 }
2381
2382 raw_spin_unlock_irq(&pool->lock);
2383 wake_dying_workers(&cull_list);
2384 mutex_unlock(&wq_pool_attach_mutex);
2385}
2386
2387static void send_mayday(struct work_struct *work)
2388{
2389 struct pool_workqueue *pwq = get_work_pwq(work);
2390 struct workqueue_struct *wq = pwq->wq;
2391
2392 lockdep_assert_held(&wq_mayday_lock);
2393
2394 if (!wq->rescuer)
2395 return;
2396
2397 /* mayday mayday mayday */
2398 if (list_empty(&pwq->mayday_node)) {
2399 /*
2400 * If @pwq is for an unbound wq, its base ref may be put at
2401 * any time due to an attribute change. Pin @pwq until the
2402 * rescuer is done with it.
2403 */
2404 get_pwq(pwq);
2405 list_add_tail(&pwq->mayday_node, &wq->maydays);
2406 wake_up_process(wq->rescuer->task);
2407 pwq->stats[PWQ_STAT_MAYDAY]++;
2408 }
2409}
2410
2411static void pool_mayday_timeout(struct timer_list *t)
2412{
2413 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2414 struct work_struct *work;
2415
2416 raw_spin_lock_irq(&pool->lock);
2417 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
2418
2419 if (need_to_create_worker(pool)) {
2420 /*
2421 * We've been trying to create a new worker but
2422 * haven't been successful. We might be hitting an
2423 * allocation deadlock. Send distress signals to
2424 * rescuers.
2425 */
2426 list_for_each_entry(work, &pool->worklist, entry)
2427 send_mayday(work);
2428 }
2429
2430 raw_spin_unlock(&wq_mayday_lock);
2431 raw_spin_unlock_irq(&pool->lock);
2432
2433 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2434}
2435
2436/**
2437 * maybe_create_worker - create a new worker if necessary
2438 * @pool: pool to create a new worker for
2439 *
2440 * Create a new worker for @pool if necessary. @pool is guaranteed to
2441 * have at least one idle worker on return from this function. If
2442 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2443 * sent to all rescuers with works scheduled on @pool to resolve
2444 * possible allocation deadlock.
2445 *
2446 * On return, need_to_create_worker() is guaranteed to be %false and
2447 * may_start_working() %true.
2448 *
2449 * LOCKING:
2450 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2451 * multiple times. Does GFP_KERNEL allocations. Called only from
2452 * manager.
2453 */
2454static void maybe_create_worker(struct worker_pool *pool)
2455__releases(&pool->lock)
2456__acquires(&pool->lock)
2457{
2458restart:
2459 raw_spin_unlock_irq(&pool->lock);
2460
2461 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2462 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2463
2464 while (true) {
2465 if (create_worker(pool) || !need_to_create_worker(pool))
2466 break;
2467
2468 schedule_timeout_interruptible(CREATE_COOLDOWN);
2469
2470 if (!need_to_create_worker(pool))
2471 break;
2472 }
2473
2474 del_timer_sync(&pool->mayday_timer);
2475 raw_spin_lock_irq(&pool->lock);
2476 /*
2477 * This is necessary even after a new worker was just successfully
2478 * created as @pool->lock was dropped and the new worker might have
2479 * already become busy.
2480 */
2481 if (need_to_create_worker(pool))
2482 goto restart;
2483}
2484
2485/**
2486 * manage_workers - manage worker pool
2487 * @worker: self
2488 *
2489 * Assume the manager role and manage the worker pool @worker belongs
2490 * to. At any given time, there can be only zero or one manager per
2491 * pool. The exclusion is handled automatically by this function.
2492 *
2493 * The caller can safely start processing works on false return. On
2494 * true return, it's guaranteed that need_to_create_worker() is false
2495 * and may_start_working() is true.
2496 *
2497 * CONTEXT:
2498 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2499 * multiple times. Does GFP_KERNEL allocations.
2500 *
2501 * Return:
2502 * %false if the pool doesn't need management and the caller can safely
2503 * start processing works, %true if management function was performed and
2504 * the conditions that the caller verified before calling the function may
2505 * no longer be true.
2506 */
2507static bool manage_workers(struct worker *worker)
2508{
2509 struct worker_pool *pool = worker->pool;
2510
2511 if (pool->flags & POOL_MANAGER_ACTIVE)
2512 return false;
2513
2514 pool->flags |= POOL_MANAGER_ACTIVE;
2515 pool->manager = worker;
2516
2517 maybe_create_worker(pool);
2518
2519 pool->manager = NULL;
2520 pool->flags &= ~POOL_MANAGER_ACTIVE;
2521 rcuwait_wake_up(&manager_wait);
2522 return true;
2523}
2524
2525/**
2526 * process_one_work - process single work
2527 * @worker: self
2528 * @work: work to process
2529 *
2530 * Process @work. This function contains all the logics necessary to
2531 * process a single work including synchronization against and
2532 * interaction with other workers on the same cpu, queueing and
2533 * flushing. As long as context requirement is met, any worker can
2534 * call this function to process a work.
2535 *
2536 * CONTEXT:
2537 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
2538 */
2539static void process_one_work(struct worker *worker, struct work_struct *work)
2540__releases(&pool->lock)
2541__acquires(&pool->lock)
2542{
2543 struct pool_workqueue *pwq = get_work_pwq(work);
2544 struct worker_pool *pool = worker->pool;
2545 unsigned long work_data;
2546#ifdef CONFIG_LOCKDEP
2547 /*
2548 * It is permissible to free the struct work_struct from
2549 * inside the function that is called from it, this we need to
2550 * take into account for lockdep too. To avoid bogus "held
2551 * lock freed" warnings as well as problems when looking into
2552 * work->lockdep_map, make a copy and use that here.
2553 */
2554 struct lockdep_map lockdep_map;
2555
2556 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2557#endif
2558 /* ensure we're on the correct CPU */
2559 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2560 raw_smp_processor_id() != pool->cpu);
2561
2562 /* claim and dequeue */
2563 debug_work_deactivate(work);
2564 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2565 worker->current_work = work;
2566 worker->current_func = work->func;
2567 worker->current_pwq = pwq;
2568 worker->current_at = worker->task->se.sum_exec_runtime;
2569 work_data = *work_data_bits(work);
2570 worker->current_color = get_work_color(work_data);
2571
2572 /*
2573 * Record wq name for cmdline and debug reporting, may get
2574 * overridden through set_worker_desc().
2575 */
2576 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2577
2578 list_del_init(&work->entry);
2579
2580 /*
2581 * CPU intensive works don't participate in concurrency management.
2582 * They're the scheduler's responsibility. This takes @worker out
2583 * of concurrency management and the next code block will chain
2584 * execution of the pending work items.
2585 */
2586 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
2587 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2588
2589 /*
2590 * Kick @pool if necessary. It's always noop for per-cpu worker pools
2591 * since nr_running would always be >= 1 at this point. This is used to
2592 * chain execution of the pending work items for WORKER_NOT_RUNNING
2593 * workers such as the UNBOUND and CPU_INTENSIVE ones.
2594 */
2595 kick_pool(pool);
2596
2597 /*
2598 * Record the last pool and clear PENDING which should be the last
2599 * update to @work. Also, do this inside @pool->lock so that
2600 * PENDING and queued state changes happen together while IRQ is
2601 * disabled.
2602 */
2603 set_work_pool_and_clear_pending(work, pool->id);
2604
2605 pwq->stats[PWQ_STAT_STARTED]++;
2606 raw_spin_unlock_irq(&pool->lock);
2607
2608 lock_map_acquire(&pwq->wq->lockdep_map);
2609 lock_map_acquire(&lockdep_map);
2610 /*
2611 * Strictly speaking we should mark the invariant state without holding
2612 * any locks, that is, before these two lock_map_acquire()'s.
2613 *
2614 * However, that would result in:
2615 *
2616 * A(W1)
2617 * WFC(C)
2618 * A(W1)
2619 * C(C)
2620 *
2621 * Which would create W1->C->W1 dependencies, even though there is no
2622 * actual deadlock possible. There are two solutions, using a
2623 * read-recursive acquire on the work(queue) 'locks', but this will then
2624 * hit the lockdep limitation on recursive locks, or simply discard
2625 * these locks.
2626 *
2627 * AFAICT there is no possible deadlock scenario between the
2628 * flush_work() and complete() primitives (except for single-threaded
2629 * workqueues), so hiding them isn't a problem.
2630 */
2631 lockdep_invariant_state(true);
2632 trace_workqueue_execute_start(work);
2633 worker->current_func(work);
2634 /*
2635 * While we must be careful to not use "work" after this, the trace
2636 * point will only record its address.
2637 */
2638 trace_workqueue_execute_end(work, worker->current_func);
2639 pwq->stats[PWQ_STAT_COMPLETED]++;
2640 lock_map_release(&lockdep_map);
2641 lock_map_release(&pwq->wq->lockdep_map);
2642
2643 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2644 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2645 " last function: %ps\n",
2646 current->comm, preempt_count(), task_pid_nr(current),
2647 worker->current_func);
2648 debug_show_held_locks(current);
2649 dump_stack();
2650 }
2651
2652 /*
2653 * The following prevents a kworker from hogging CPU on !PREEMPTION
2654 * kernels, where a requeueing work item waiting for something to
2655 * happen could deadlock with stop_machine as such work item could
2656 * indefinitely requeue itself while all other CPUs are trapped in
2657 * stop_machine. At the same time, report a quiescent RCU state so
2658 * the same condition doesn't freeze RCU.
2659 */
2660 cond_resched();
2661
2662 raw_spin_lock_irq(&pool->lock);
2663
2664 /*
2665 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
2666 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
2667 * wq_cpu_intensive_thresh_us. Clear it.
2668 */
2669 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2670
2671 /* tag the worker for identification in schedule() */
2672 worker->last_func = worker->current_func;
2673
2674 /* we're done with it, release */
2675 hash_del(&worker->hentry);
2676 worker->current_work = NULL;
2677 worker->current_func = NULL;
2678 worker->current_pwq = NULL;
2679 worker->current_color = INT_MAX;
2680 pwq_dec_nr_in_flight(pwq, work_data);
2681}
2682
2683/**
2684 * process_scheduled_works - process scheduled works
2685 * @worker: self
2686 *
2687 * Process all scheduled works. Please note that the scheduled list
2688 * may change while processing a work, so this function repeatedly
2689 * fetches a work from the top and executes it.
2690 *
2691 * CONTEXT:
2692 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
2693 * multiple times.
2694 */
2695static void process_scheduled_works(struct worker *worker)
2696{
2697 struct work_struct *work;
2698 bool first = true;
2699
2700 while ((work = list_first_entry_or_null(&worker->scheduled,
2701 struct work_struct, entry))) {
2702 if (first) {
2703 worker->pool->watchdog_ts = jiffies;
2704 first = false;
2705 }
2706 process_one_work(worker, work);
2707 }
2708}
2709
2710static void set_pf_worker(bool val)
2711{
2712 mutex_lock(&wq_pool_attach_mutex);
2713 if (val)
2714 current->flags |= PF_WQ_WORKER;
2715 else
2716 current->flags &= ~PF_WQ_WORKER;
2717 mutex_unlock(&wq_pool_attach_mutex);
2718}
2719
2720/**
2721 * worker_thread - the worker thread function
2722 * @__worker: self
2723 *
2724 * The worker thread function. All workers belong to a worker_pool -
2725 * either a per-cpu one or dynamic unbound one. These workers process all
2726 * work items regardless of their specific target workqueue. The only
2727 * exception is work items which belong to workqueues with a rescuer which
2728 * will be explained in rescuer_thread().
2729 *
2730 * Return: 0
2731 */
2732static int worker_thread(void *__worker)
2733{
2734 struct worker *worker = __worker;
2735 struct worker_pool *pool = worker->pool;
2736
2737 /* tell the scheduler that this is a workqueue worker */
2738 set_pf_worker(true);
2739woke_up:
2740 raw_spin_lock_irq(&pool->lock);
2741
2742 /* am I supposed to die? */
2743 if (unlikely(worker->flags & WORKER_DIE)) {
2744 raw_spin_unlock_irq(&pool->lock);
2745 set_pf_worker(false);
2746
2747 set_task_comm(worker->task, "kworker/dying");
2748 ida_free(&pool->worker_ida, worker->id);
2749 worker_detach_from_pool(worker);
2750 WARN_ON_ONCE(!list_empty(&worker->entry));
2751 kfree(worker);
2752 return 0;
2753 }
2754
2755 worker_leave_idle(worker);
2756recheck:
2757 /* no more worker necessary? */
2758 if (!need_more_worker(pool))
2759 goto sleep;
2760
2761 /* do we need to manage? */
2762 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2763 goto recheck;
2764
2765 /*
2766 * ->scheduled list can only be filled while a worker is
2767 * preparing to process a work or actually processing it.
2768 * Make sure nobody diddled with it while I was sleeping.
2769 */
2770 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2771
2772 /*
2773 * Finish PREP stage. We're guaranteed to have at least one idle
2774 * worker or that someone else has already assumed the manager
2775 * role. This is where @worker starts participating in concurrency
2776 * management if applicable and concurrency management is restored
2777 * after being rebound. See rebind_workers() for details.
2778 */
2779 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2780
2781 do {
2782 struct work_struct *work =
2783 list_first_entry(&pool->worklist,
2784 struct work_struct, entry);
2785
2786 if (assign_work(work, worker, NULL))
2787 process_scheduled_works(worker);
2788 } while (keep_working(pool));
2789
2790 worker_set_flags(worker, WORKER_PREP);
2791sleep:
2792 /*
2793 * pool->lock is held and there's no work to process and no need to
2794 * manage, sleep. Workers are woken up only while holding
2795 * pool->lock or from local cpu, so setting the current state
2796 * before releasing pool->lock is enough to prevent losing any
2797 * event.
2798 */
2799 worker_enter_idle(worker);
2800 __set_current_state(TASK_IDLE);
2801 raw_spin_unlock_irq(&pool->lock);
2802 schedule();
2803 goto woke_up;
2804}
2805
2806/**
2807 * rescuer_thread - the rescuer thread function
2808 * @__rescuer: self
2809 *
2810 * Workqueue rescuer thread function. There's one rescuer for each
2811 * workqueue which has WQ_MEM_RECLAIM set.
2812 *
2813 * Regular work processing on a pool may block trying to create a new
2814 * worker which uses GFP_KERNEL allocation which has slight chance of
2815 * developing into deadlock if some works currently on the same queue
2816 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2817 * the problem rescuer solves.
2818 *
2819 * When such condition is possible, the pool summons rescuers of all
2820 * workqueues which have works queued on the pool and let them process
2821 * those works so that forward progress can be guaranteed.
2822 *
2823 * This should happen rarely.
2824 *
2825 * Return: 0
2826 */
2827static int rescuer_thread(void *__rescuer)
2828{
2829 struct worker *rescuer = __rescuer;
2830 struct workqueue_struct *wq = rescuer->rescue_wq;
2831 bool should_stop;
2832
2833 set_user_nice(current, RESCUER_NICE_LEVEL);
2834
2835 /*
2836 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2837 * doesn't participate in concurrency management.
2838 */
2839 set_pf_worker(true);
2840repeat:
2841 set_current_state(TASK_IDLE);
2842
2843 /*
2844 * By the time the rescuer is requested to stop, the workqueue
2845 * shouldn't have any work pending, but @wq->maydays may still have
2846 * pwq(s) queued. This can happen by non-rescuer workers consuming
2847 * all the work items before the rescuer got to them. Go through
2848 * @wq->maydays processing before acting on should_stop so that the
2849 * list is always empty on exit.
2850 */
2851 should_stop = kthread_should_stop();
2852
2853 /* see whether any pwq is asking for help */
2854 raw_spin_lock_irq(&wq_mayday_lock);
2855
2856 while (!list_empty(&wq->maydays)) {
2857 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2858 struct pool_workqueue, mayday_node);
2859 struct worker_pool *pool = pwq->pool;
2860 struct work_struct *work, *n;
2861
2862 __set_current_state(TASK_RUNNING);
2863 list_del_init(&pwq->mayday_node);
2864
2865 raw_spin_unlock_irq(&wq_mayday_lock);
2866
2867 worker_attach_to_pool(rescuer, pool);
2868
2869 raw_spin_lock_irq(&pool->lock);
2870
2871 /*
2872 * Slurp in all works issued via this workqueue and
2873 * process'em.
2874 */
2875 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2876 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2877 if (get_work_pwq(work) == pwq &&
2878 assign_work(work, rescuer, &n))
2879 pwq->stats[PWQ_STAT_RESCUED]++;
2880 }
2881
2882 if (!list_empty(&rescuer->scheduled)) {
2883 process_scheduled_works(rescuer);
2884
2885 /*
2886 * The above execution of rescued work items could
2887 * have created more to rescue through
2888 * pwq_activate_first_inactive() or chained
2889 * queueing. Let's put @pwq back on mayday list so
2890 * that such back-to-back work items, which may be
2891 * being used to relieve memory pressure, don't
2892 * incur MAYDAY_INTERVAL delay inbetween.
2893 */
2894 if (pwq->nr_active && need_to_create_worker(pool)) {
2895 raw_spin_lock(&wq_mayday_lock);
2896 /*
2897 * Queue iff we aren't racing destruction
2898 * and somebody else hasn't queued it already.
2899 */
2900 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
2901 get_pwq(pwq);
2902 list_add_tail(&pwq->mayday_node, &wq->maydays);
2903 }
2904 raw_spin_unlock(&wq_mayday_lock);
2905 }
2906 }
2907
2908 /*
2909 * Put the reference grabbed by send_mayday(). @pool won't
2910 * go away while we're still attached to it.
2911 */
2912 put_pwq(pwq);
2913
2914 /*
2915 * Leave this pool. Notify regular workers; otherwise, we end up
2916 * with 0 concurrency and stalling the execution.
2917 */
2918 kick_pool(pool);
2919
2920 raw_spin_unlock_irq(&pool->lock);
2921
2922 worker_detach_from_pool(rescuer);
2923
2924 raw_spin_lock_irq(&wq_mayday_lock);
2925 }
2926
2927 raw_spin_unlock_irq(&wq_mayday_lock);
2928
2929 if (should_stop) {
2930 __set_current_state(TASK_RUNNING);
2931 set_pf_worker(false);
2932 return 0;
2933 }
2934
2935 /* rescuers should never participate in concurrency management */
2936 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2937 schedule();
2938 goto repeat;
2939}
2940
2941/**
2942 * check_flush_dependency - check for flush dependency sanity
2943 * @target_wq: workqueue being flushed
2944 * @target_work: work item being flushed (NULL for workqueue flushes)
2945 *
2946 * %current is trying to flush the whole @target_wq or @target_work on it.
2947 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2948 * reclaiming memory or running on a workqueue which doesn't have
2949 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2950 * a deadlock.
2951 */
2952static void check_flush_dependency(struct workqueue_struct *target_wq,
2953 struct work_struct *target_work)
2954{
2955 work_func_t target_func = target_work ? target_work->func : NULL;
2956 struct worker *worker;
2957
2958 if (target_wq->flags & WQ_MEM_RECLAIM)
2959 return;
2960
2961 worker = current_wq_worker();
2962
2963 WARN_ONCE(current->flags & PF_MEMALLOC,
2964 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2965 current->pid, current->comm, target_wq->name, target_func);
2966 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2967 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2968 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2969 worker->current_pwq->wq->name, worker->current_func,
2970 target_wq->name, target_func);
2971}
2972
2973struct wq_barrier {
2974 struct work_struct work;
2975 struct completion done;
2976 struct task_struct *task; /* purely informational */
2977};
2978
2979static void wq_barrier_func(struct work_struct *work)
2980{
2981 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2982 complete(&barr->done);
2983}
2984
2985/**
2986 * insert_wq_barrier - insert a barrier work
2987 * @pwq: pwq to insert barrier into
2988 * @barr: wq_barrier to insert
2989 * @target: target work to attach @barr to
2990 * @worker: worker currently executing @target, NULL if @target is not executing
2991 *
2992 * @barr is linked to @target such that @barr is completed only after
2993 * @target finishes execution. Please note that the ordering
2994 * guarantee is observed only with respect to @target and on the local
2995 * cpu.
2996 *
2997 * Currently, a queued barrier can't be canceled. This is because
2998 * try_to_grab_pending() can't determine whether the work to be
2999 * grabbed is at the head of the queue and thus can't clear LINKED
3000 * flag of the previous work while there must be a valid next work
3001 * after a work with LINKED flag set.
3002 *
3003 * Note that when @worker is non-NULL, @target may be modified
3004 * underneath us, so we can't reliably determine pwq from @target.
3005 *
3006 * CONTEXT:
3007 * raw_spin_lock_irq(pool->lock).
3008 */
3009static void insert_wq_barrier(struct pool_workqueue *pwq,
3010 struct wq_barrier *barr,
3011 struct work_struct *target, struct worker *worker)
3012{
3013 unsigned int work_flags = 0;
3014 unsigned int work_color;
3015 struct list_head *head;
3016
3017 /*
3018 * debugobject calls are safe here even with pool->lock locked
3019 * as we know for sure that this will not trigger any of the
3020 * checks and call back into the fixup functions where we
3021 * might deadlock.
3022 */
3023 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
3024 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3025
3026 init_completion_map(&barr->done, &target->lockdep_map);
3027
3028 barr->task = current;
3029
3030 /* The barrier work item does not participate in pwq->nr_active. */
3031 work_flags |= WORK_STRUCT_INACTIVE;
3032
3033 /*
3034 * If @target is currently being executed, schedule the
3035 * barrier to the worker; otherwise, put it after @target.
3036 */
3037 if (worker) {
3038 head = worker->scheduled.next;
3039 work_color = worker->current_color;
3040 } else {
3041 unsigned long *bits = work_data_bits(target);
3042
3043 head = target->entry.next;
3044 /* there can already be other linked works, inherit and set */
3045 work_flags |= *bits & WORK_STRUCT_LINKED;
3046 work_color = get_work_color(*bits);
3047 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3048 }
3049
3050 pwq->nr_in_flight[work_color]++;
3051 work_flags |= work_color_to_flags(work_color);
3052
3053 insert_work(pwq, &barr->work, head, work_flags);
3054}
3055
3056/**
3057 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3058 * @wq: workqueue being flushed
3059 * @flush_color: new flush color, < 0 for no-op
3060 * @work_color: new work color, < 0 for no-op
3061 *
3062 * Prepare pwqs for workqueue flushing.
3063 *
3064 * If @flush_color is non-negative, flush_color on all pwqs should be
3065 * -1. If no pwq has in-flight commands at the specified color, all
3066 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3067 * has in flight commands, its pwq->flush_color is set to
3068 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3069 * wakeup logic is armed and %true is returned.
3070 *
3071 * The caller should have initialized @wq->first_flusher prior to
3072 * calling this function with non-negative @flush_color. If
3073 * @flush_color is negative, no flush color update is done and %false
3074 * is returned.
3075 *
3076 * If @work_color is non-negative, all pwqs should have the same
3077 * work_color which is previous to @work_color and all will be
3078 * advanced to @work_color.
3079 *
3080 * CONTEXT:
3081 * mutex_lock(wq->mutex).
3082 *
3083 * Return:
3084 * %true if @flush_color >= 0 and there's something to flush. %false
3085 * otherwise.
3086 */
3087static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3088 int flush_color, int work_color)
3089{
3090 bool wait = false;
3091 struct pool_workqueue *pwq;
3092
3093 if (flush_color >= 0) {
3094 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3095 atomic_set(&wq->nr_pwqs_to_flush, 1);
3096 }
3097
3098 for_each_pwq(pwq, wq) {
3099 struct worker_pool *pool = pwq->pool;
3100
3101 raw_spin_lock_irq(&pool->lock);
3102
3103 if (flush_color >= 0) {
3104 WARN_ON_ONCE(pwq->flush_color != -1);
3105
3106 if (pwq->nr_in_flight[flush_color]) {
3107 pwq->flush_color = flush_color;
3108 atomic_inc(&wq->nr_pwqs_to_flush);
3109 wait = true;
3110 }
3111 }
3112
3113 if (work_color >= 0) {
3114 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3115 pwq->work_color = work_color;
3116 }
3117
3118 raw_spin_unlock_irq(&pool->lock);
3119 }
3120
3121 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3122 complete(&wq->first_flusher->done);
3123
3124 return wait;
3125}
3126
3127/**
3128 * __flush_workqueue - ensure that any scheduled work has run to completion.
3129 * @wq: workqueue to flush
3130 *
3131 * This function sleeps until all work items which were queued on entry
3132 * have finished execution, but it is not livelocked by new incoming ones.
3133 */
3134void __flush_workqueue(struct workqueue_struct *wq)
3135{
3136 struct wq_flusher this_flusher = {
3137 .list = LIST_HEAD_INIT(this_flusher.list),
3138 .flush_color = -1,
3139 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
3140 };
3141 int next_color;
3142
3143 if (WARN_ON(!wq_online))
3144 return;
3145
3146 lock_map_acquire(&wq->lockdep_map);
3147 lock_map_release(&wq->lockdep_map);
3148
3149 mutex_lock(&wq->mutex);
3150
3151 /*
3152 * Start-to-wait phase
3153 */
3154 next_color = work_next_color(wq->work_color);
3155
3156 if (next_color != wq->flush_color) {
3157 /*
3158 * Color space is not full. The current work_color
3159 * becomes our flush_color and work_color is advanced
3160 * by one.
3161 */
3162 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3163 this_flusher.flush_color = wq->work_color;
3164 wq->work_color = next_color;
3165
3166 if (!wq->first_flusher) {
3167 /* no flush in progress, become the first flusher */
3168 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3169
3170 wq->first_flusher = &this_flusher;
3171
3172 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3173 wq->work_color)) {
3174 /* nothing to flush, done */
3175 wq->flush_color = next_color;
3176 wq->first_flusher = NULL;
3177 goto out_unlock;
3178 }
3179 } else {
3180 /* wait in queue */
3181 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3182 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3183 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3184 }
3185 } else {
3186 /*
3187 * Oops, color space is full, wait on overflow queue.
3188 * The next flush completion will assign us
3189 * flush_color and transfer to flusher_queue.
3190 */
3191 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3192 }
3193
3194 check_flush_dependency(wq, NULL);
3195
3196 mutex_unlock(&wq->mutex);
3197
3198 wait_for_completion(&this_flusher.done);
3199
3200 /*
3201 * Wake-up-and-cascade phase
3202 *
3203 * First flushers are responsible for cascading flushes and
3204 * handling overflow. Non-first flushers can simply return.
3205 */
3206 if (READ_ONCE(wq->first_flusher) != &this_flusher)
3207 return;
3208
3209 mutex_lock(&wq->mutex);
3210
3211 /* we might have raced, check again with mutex held */
3212 if (wq->first_flusher != &this_flusher)
3213 goto out_unlock;
3214
3215 WRITE_ONCE(wq->first_flusher, NULL);
3216
3217 WARN_ON_ONCE(!list_empty(&this_flusher.list));
3218 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3219
3220 while (true) {
3221 struct wq_flusher *next, *tmp;
3222
3223 /* complete all the flushers sharing the current flush color */
3224 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
3225 if (next->flush_color != wq->flush_color)
3226 break;
3227 list_del_init(&next->list);
3228 complete(&next->done);
3229 }
3230
3231 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
3232 wq->flush_color != work_next_color(wq->work_color));
3233
3234 /* this flush_color is finished, advance by one */
3235 wq->flush_color = work_next_color(wq->flush_color);
3236
3237 /* one color has been freed, handle overflow queue */
3238 if (!list_empty(&wq->flusher_overflow)) {
3239 /*
3240 * Assign the same color to all overflowed
3241 * flushers, advance work_color and append to
3242 * flusher_queue. This is the start-to-wait
3243 * phase for these overflowed flushers.
3244 */
3245 list_for_each_entry(tmp, &wq->flusher_overflow, list)
3246 tmp->flush_color = wq->work_color;
3247
3248 wq->work_color = work_next_color(wq->work_color);
3249
3250 list_splice_tail_init(&wq->flusher_overflow,
3251 &wq->flusher_queue);
3252 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3253 }
3254
3255 if (list_empty(&wq->flusher_queue)) {
3256 WARN_ON_ONCE(wq->flush_color != wq->work_color);
3257 break;
3258 }
3259
3260 /*
3261 * Need to flush more colors. Make the next flusher
3262 * the new first flusher and arm pwqs.
3263 */
3264 WARN_ON_ONCE(wq->flush_color == wq->work_color);
3265 WARN_ON_ONCE(wq->flush_color != next->flush_color);
3266
3267 list_del_init(&next->list);
3268 wq->first_flusher = next;
3269
3270 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
3271 break;
3272
3273 /*
3274 * Meh... this color is already done, clear first
3275 * flusher and repeat cascading.
3276 */
3277 wq->first_flusher = NULL;
3278 }
3279
3280out_unlock:
3281 mutex_unlock(&wq->mutex);
3282}
3283EXPORT_SYMBOL(__flush_workqueue);
3284
3285/**
3286 * drain_workqueue - drain a workqueue
3287 * @wq: workqueue to drain
3288 *
3289 * Wait until the workqueue becomes empty. While draining is in progress,
3290 * only chain queueing is allowed. IOW, only currently pending or running
3291 * work items on @wq can queue further work items on it. @wq is flushed
3292 * repeatedly until it becomes empty. The number of flushing is determined
3293 * by the depth of chaining and should be relatively short. Whine if it
3294 * takes too long.
3295 */
3296void drain_workqueue(struct workqueue_struct *wq)
3297{
3298 unsigned int flush_cnt = 0;
3299 struct pool_workqueue *pwq;
3300
3301 /*
3302 * __queue_work() needs to test whether there are drainers, is much
3303 * hotter than drain_workqueue() and already looks at @wq->flags.
3304 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
3305 */
3306 mutex_lock(&wq->mutex);
3307 if (!wq->nr_drainers++)
3308 wq->flags |= __WQ_DRAINING;
3309 mutex_unlock(&wq->mutex);
3310reflush:
3311 __flush_workqueue(wq);
3312
3313 mutex_lock(&wq->mutex);
3314
3315 for_each_pwq(pwq, wq) {
3316 bool drained;
3317
3318 raw_spin_lock_irq(&pwq->pool->lock);
3319 drained = !pwq->nr_active && list_empty(&pwq->inactive_works);
3320 raw_spin_unlock_irq(&pwq->pool->lock);
3321
3322 if (drained)
3323 continue;
3324
3325 if (++flush_cnt == 10 ||
3326 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3327 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
3328 wq->name, __func__, flush_cnt);
3329
3330 mutex_unlock(&wq->mutex);
3331 goto reflush;
3332 }
3333
3334 if (!--wq->nr_drainers)
3335 wq->flags &= ~__WQ_DRAINING;
3336 mutex_unlock(&wq->mutex);
3337}
3338EXPORT_SYMBOL_GPL(drain_workqueue);
3339
3340static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
3341 bool from_cancel)
3342{
3343 struct worker *worker = NULL;
3344 struct worker_pool *pool;
3345 struct pool_workqueue *pwq;
3346
3347 might_sleep();
3348
3349 rcu_read_lock();
3350 pool = get_work_pool(work);
3351 if (!pool) {
3352 rcu_read_unlock();
3353 return false;
3354 }
3355
3356 raw_spin_lock_irq(&pool->lock);
3357 /* see the comment in try_to_grab_pending() with the same code */
3358 pwq = get_work_pwq(work);
3359 if (pwq) {
3360 if (unlikely(pwq->pool != pool))
3361 goto already_gone;
3362 } else {
3363 worker = find_worker_executing_work(pool, work);
3364 if (!worker)
3365 goto already_gone;
3366 pwq = worker->current_pwq;
3367 }
3368
3369 check_flush_dependency(pwq->wq, work);
3370
3371 insert_wq_barrier(pwq, barr, work, worker);
3372 raw_spin_unlock_irq(&pool->lock);
3373
3374 /*
3375 * Force a lock recursion deadlock when using flush_work() inside a
3376 * single-threaded or rescuer equipped workqueue.
3377 *
3378 * For single threaded workqueues the deadlock happens when the work
3379 * is after the work issuing the flush_work(). For rescuer equipped
3380 * workqueues the deadlock happens when the rescuer stalls, blocking
3381 * forward progress.
3382 */
3383 if (!from_cancel &&
3384 (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3385 lock_map_acquire(&pwq->wq->lockdep_map);
3386 lock_map_release(&pwq->wq->lockdep_map);
3387 }
3388 rcu_read_unlock();
3389 return true;
3390already_gone:
3391 raw_spin_unlock_irq(&pool->lock);
3392 rcu_read_unlock();
3393 return false;
3394}
3395
3396static bool __flush_work(struct work_struct *work, bool from_cancel)
3397{
3398 struct wq_barrier barr;
3399
3400 if (WARN_ON(!wq_online))
3401 return false;
3402
3403 if (WARN_ON(!work->func))
3404 return false;
3405
3406 lock_map_acquire(&work->lockdep_map);
3407 lock_map_release(&work->lockdep_map);
3408
3409 if (start_flush_work(work, &barr, from_cancel)) {
3410 wait_for_completion(&barr.done);
3411 destroy_work_on_stack(&barr.work);
3412 return true;
3413 } else {
3414 return false;
3415 }
3416}
3417
3418/**
3419 * flush_work - wait for a work to finish executing the last queueing instance
3420 * @work: the work to flush
3421 *
3422 * Wait until @work has finished execution. @work is guaranteed to be idle
3423 * on return if it hasn't been requeued since flush started.
3424 *
3425 * Return:
3426 * %true if flush_work() waited for the work to finish execution,
3427 * %false if it was already idle.
3428 */
3429bool flush_work(struct work_struct *work)
3430{
3431 return __flush_work(work, false);
3432}
3433EXPORT_SYMBOL_GPL(flush_work);
3434
3435struct cwt_wait {
3436 wait_queue_entry_t wait;
3437 struct work_struct *work;
3438};
3439
3440static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3441{
3442 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3443
3444 if (cwait->work != key)
3445 return 0;
3446 return autoremove_wake_function(wait, mode, sync, key);
3447}
3448
3449static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3450{
3451 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3452 unsigned long flags;
3453 int ret;
3454
3455 do {
3456 ret = try_to_grab_pending(work, is_dwork, &flags);
3457 /*
3458 * If someone else is already canceling, wait for it to
3459 * finish. flush_work() doesn't work for PREEMPT_NONE
3460 * because we may get scheduled between @work's completion
3461 * and the other canceling task resuming and clearing
3462 * CANCELING - flush_work() will return false immediately
3463 * as @work is no longer busy, try_to_grab_pending() will
3464 * return -ENOENT as @work is still being canceled and the
3465 * other canceling task won't be able to clear CANCELING as
3466 * we're hogging the CPU.
3467 *
3468 * Let's wait for completion using a waitqueue. As this
3469 * may lead to the thundering herd problem, use a custom
3470 * wake function which matches @work along with exclusive
3471 * wait and wakeup.
3472 */
3473 if (unlikely(ret == -ENOENT)) {
3474 struct cwt_wait cwait;
3475
3476 init_wait(&cwait.wait);
3477 cwait.wait.func = cwt_wakefn;
3478 cwait.work = work;
3479
3480 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3481 TASK_UNINTERRUPTIBLE);
3482 if (work_is_canceling(work))
3483 schedule();
3484 finish_wait(&cancel_waitq, &cwait.wait);
3485 }
3486 } while (unlikely(ret < 0));
3487
3488 /* tell other tasks trying to grab @work to back off */
3489 mark_work_canceling(work);
3490 local_irq_restore(flags);
3491
3492 /*
3493 * This allows canceling during early boot. We know that @work
3494 * isn't executing.
3495 */
3496 if (wq_online)
3497 __flush_work(work, true);
3498
3499 clear_work_data(work);
3500
3501 /*
3502 * Paired with prepare_to_wait() above so that either
3503 * waitqueue_active() is visible here or !work_is_canceling() is
3504 * visible there.
3505 */
3506 smp_mb();
3507 if (waitqueue_active(&cancel_waitq))
3508 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3509
3510 return ret;
3511}
3512
3513/**
3514 * cancel_work_sync - cancel a work and wait for it to finish
3515 * @work: the work to cancel
3516 *
3517 * Cancel @work and wait for its execution to finish. This function
3518 * can be used even if the work re-queues itself or migrates to
3519 * another workqueue. On return from this function, @work is
3520 * guaranteed to be not pending or executing on any CPU.
3521 *
3522 * cancel_work_sync(&delayed_work->work) must not be used for
3523 * delayed_work's. Use cancel_delayed_work_sync() instead.
3524 *
3525 * The caller must ensure that the workqueue on which @work was last
3526 * queued can't be destroyed before this function returns.
3527 *
3528 * Return:
3529 * %true if @work was pending, %false otherwise.
3530 */
3531bool cancel_work_sync(struct work_struct *work)
3532{
3533 return __cancel_work_timer(work, false);
3534}
3535EXPORT_SYMBOL_GPL(cancel_work_sync);
3536
3537/**
3538 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3539 * @dwork: the delayed work to flush
3540 *
3541 * Delayed timer is cancelled and the pending work is queued for
3542 * immediate execution. Like flush_work(), this function only
3543 * considers the last queueing instance of @dwork.
3544 *
3545 * Return:
3546 * %true if flush_work() waited for the work to finish execution,
3547 * %false if it was already idle.
3548 */
3549bool flush_delayed_work(struct delayed_work *dwork)
3550{
3551 local_irq_disable();
3552 if (del_timer_sync(&dwork->timer))
3553 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
3554 local_irq_enable();
3555 return flush_work(&dwork->work);
3556}
3557EXPORT_SYMBOL(flush_delayed_work);
3558
3559/**
3560 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3561 * @rwork: the rcu work to flush
3562 *
3563 * Return:
3564 * %true if flush_rcu_work() waited for the work to finish execution,
3565 * %false if it was already idle.
3566 */
3567bool flush_rcu_work(struct rcu_work *rwork)
3568{
3569 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3570 rcu_barrier();
3571 flush_work(&rwork->work);
3572 return true;
3573 } else {
3574 return flush_work(&rwork->work);
3575 }
3576}
3577EXPORT_SYMBOL(flush_rcu_work);
3578
3579static bool __cancel_work(struct work_struct *work, bool is_dwork)
3580{
3581 unsigned long flags;
3582 int ret;
3583
3584 do {
3585 ret = try_to_grab_pending(work, is_dwork, &flags);
3586 } while (unlikely(ret == -EAGAIN));
3587
3588 if (unlikely(ret < 0))
3589 return false;
3590
3591 set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3592 local_irq_restore(flags);
3593 return ret;
3594}
3595
3596/*
3597 * See cancel_delayed_work()
3598 */
3599bool cancel_work(struct work_struct *work)
3600{
3601 return __cancel_work(work, false);
3602}
3603EXPORT_SYMBOL(cancel_work);
3604
3605/**
3606 * cancel_delayed_work - cancel a delayed work
3607 * @dwork: delayed_work to cancel
3608 *
3609 * Kill off a pending delayed_work.
3610 *
3611 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3612 * pending.
3613 *
3614 * Note:
3615 * The work callback function may still be running on return, unless
3616 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3617 * use cancel_delayed_work_sync() to wait on it.
3618 *
3619 * This function is safe to call from any context including IRQ handler.
3620 */
3621bool cancel_delayed_work(struct delayed_work *dwork)
3622{
3623 return __cancel_work(&dwork->work, true);
3624}
3625EXPORT_SYMBOL(cancel_delayed_work);
3626
3627/**
3628 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3629 * @dwork: the delayed work cancel
3630 *
3631 * This is cancel_work_sync() for delayed works.
3632 *
3633 * Return:
3634 * %true if @dwork was pending, %false otherwise.
3635 */
3636bool cancel_delayed_work_sync(struct delayed_work *dwork)
3637{
3638 return __cancel_work_timer(&dwork->work, true);
3639}
3640EXPORT_SYMBOL(cancel_delayed_work_sync);
3641
3642/**
3643 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3644 * @func: the function to call
3645 *
3646 * schedule_on_each_cpu() executes @func on each online CPU using the
3647 * system workqueue and blocks until all CPUs have completed.
3648 * schedule_on_each_cpu() is very slow.
3649 *
3650 * Return:
3651 * 0 on success, -errno on failure.
3652 */
3653int schedule_on_each_cpu(work_func_t func)
3654{
3655 int cpu;
3656 struct work_struct __percpu *works;
3657
3658 works = alloc_percpu(struct work_struct);
3659 if (!works)
3660 return -ENOMEM;
3661
3662 cpus_read_lock();
3663
3664 for_each_online_cpu(cpu) {
3665 struct work_struct *work = per_cpu_ptr(works, cpu);
3666
3667 INIT_WORK(work, func);
3668 schedule_work_on(cpu, work);
3669 }
3670
3671 for_each_online_cpu(cpu)
3672 flush_work(per_cpu_ptr(works, cpu));
3673
3674 cpus_read_unlock();
3675 free_percpu(works);
3676 return 0;
3677}
3678
3679/**
3680 * execute_in_process_context - reliably execute the routine with user context
3681 * @fn: the function to execute
3682 * @ew: guaranteed storage for the execute work structure (must
3683 * be available when the work executes)
3684 *
3685 * Executes the function immediately if process context is available,
3686 * otherwise schedules the function for delayed execution.
3687 *
3688 * Return: 0 - function was executed
3689 * 1 - function was scheduled for execution
3690 */
3691int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3692{
3693 if (!in_interrupt()) {
3694 fn(&ew->work);
3695 return 0;
3696 }
3697
3698 INIT_WORK(&ew->work, fn);
3699 schedule_work(&ew->work);
3700
3701 return 1;
3702}
3703EXPORT_SYMBOL_GPL(execute_in_process_context);
3704
3705/**
3706 * free_workqueue_attrs - free a workqueue_attrs
3707 * @attrs: workqueue_attrs to free
3708 *
3709 * Undo alloc_workqueue_attrs().
3710 */
3711void free_workqueue_attrs(struct workqueue_attrs *attrs)
3712{
3713 if (attrs) {
3714 free_cpumask_var(attrs->cpumask);
3715 free_cpumask_var(attrs->__pod_cpumask);
3716 kfree(attrs);
3717 }
3718}
3719
3720/**
3721 * alloc_workqueue_attrs - allocate a workqueue_attrs
3722 *
3723 * Allocate a new workqueue_attrs, initialize with default settings and
3724 * return it.
3725 *
3726 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3727 */
3728struct workqueue_attrs *alloc_workqueue_attrs(void)
3729{
3730 struct workqueue_attrs *attrs;
3731
3732 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3733 if (!attrs)
3734 goto fail;
3735 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3736 goto fail;
3737 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
3738 goto fail;
3739
3740 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3741 attrs->affn_scope = WQ_AFFN_DFL;
3742 return attrs;
3743fail:
3744 free_workqueue_attrs(attrs);
3745 return NULL;
3746}
3747
3748static void copy_workqueue_attrs(struct workqueue_attrs *to,
3749 const struct workqueue_attrs *from)
3750{
3751 to->nice = from->nice;
3752 cpumask_copy(to->cpumask, from->cpumask);
3753 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
3754 to->affn_strict = from->affn_strict;
3755
3756 /*
3757 * Unlike hash and equality test, copying shouldn't ignore wq-only
3758 * fields as copying is used for both pool and wq attrs. Instead,
3759 * get_unbound_pool() explicitly clears the fields.
3760 */
3761 to->affn_scope = from->affn_scope;
3762 to->ordered = from->ordered;
3763}
3764
3765/*
3766 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
3767 * comments in 'struct workqueue_attrs' definition.
3768 */
3769static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
3770{
3771 attrs->affn_scope = WQ_AFFN_NR_TYPES;
3772 attrs->ordered = false;
3773}
3774
3775/* hash value of the content of @attr */
3776static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3777{
3778 u32 hash = 0;
3779
3780 hash = jhash_1word(attrs->nice, hash);
3781 hash = jhash(cpumask_bits(attrs->cpumask),
3782 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3783 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
3784 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3785 hash = jhash_1word(attrs->affn_strict, hash);
3786 return hash;
3787}
3788
3789/* content equality test */
3790static bool wqattrs_equal(const struct workqueue_attrs *a,
3791 const struct workqueue_attrs *b)
3792{
3793 if (a->nice != b->nice)
3794 return false;
3795 if (!cpumask_equal(a->cpumask, b->cpumask))
3796 return false;
3797 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
3798 return false;
3799 if (a->affn_strict != b->affn_strict)
3800 return false;
3801 return true;
3802}
3803
3804/* Update @attrs with actually available CPUs */
3805static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
3806 const cpumask_t *unbound_cpumask)
3807{
3808 /*
3809 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
3810 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
3811 * @unbound_cpumask.
3812 */
3813 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
3814 if (unlikely(cpumask_empty(attrs->cpumask)))
3815 cpumask_copy(attrs->cpumask, unbound_cpumask);
3816}
3817
3818/* find wq_pod_type to use for @attrs */
3819static const struct wq_pod_type *
3820wqattrs_pod_type(const struct workqueue_attrs *attrs)
3821{
3822 enum wq_affn_scope scope;
3823 struct wq_pod_type *pt;
3824
3825 /* to synchronize access to wq_affn_dfl */
3826 lockdep_assert_held(&wq_pool_mutex);
3827
3828 if (attrs->affn_scope == WQ_AFFN_DFL)
3829 scope = wq_affn_dfl;
3830 else
3831 scope = attrs->affn_scope;
3832
3833 pt = &wq_pod_types[scope];
3834
3835 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
3836 likely(pt->nr_pods))
3837 return pt;
3838
3839 /*
3840 * Before workqueue_init_topology(), only SYSTEM is available which is
3841 * initialized in workqueue_init_early().
3842 */
3843 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
3844 BUG_ON(!pt->nr_pods);
3845 return pt;
3846}
3847
3848/**
3849 * init_worker_pool - initialize a newly zalloc'd worker_pool
3850 * @pool: worker_pool to initialize
3851 *
3852 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3853 *
3854 * Return: 0 on success, -errno on failure. Even on failure, all fields
3855 * inside @pool proper are initialized and put_unbound_pool() can be called
3856 * on @pool safely to release it.
3857 */
3858static int init_worker_pool(struct worker_pool *pool)
3859{
3860 raw_spin_lock_init(&pool->lock);
3861 pool->id = -1;
3862 pool->cpu = -1;
3863 pool->node = NUMA_NO_NODE;
3864 pool->flags |= POOL_DISASSOCIATED;
3865 pool->watchdog_ts = jiffies;
3866 INIT_LIST_HEAD(&pool->worklist);
3867 INIT_LIST_HEAD(&pool->idle_list);
3868 hash_init(pool->busy_hash);
3869
3870 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
3871 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
3872
3873 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
3874
3875 INIT_LIST_HEAD(&pool->workers);
3876 INIT_LIST_HEAD(&pool->dying_workers);
3877
3878 ida_init(&pool->worker_ida);
3879 INIT_HLIST_NODE(&pool->hash_node);
3880 pool->refcnt = 1;
3881
3882 /* shouldn't fail above this point */
3883 pool->attrs = alloc_workqueue_attrs();
3884 if (!pool->attrs)
3885 return -ENOMEM;
3886
3887 wqattrs_clear_for_pool(pool->attrs);
3888
3889 return 0;
3890}
3891
3892#ifdef CONFIG_LOCKDEP
3893static void wq_init_lockdep(struct workqueue_struct *wq)
3894{
3895 char *lock_name;
3896
3897 lockdep_register_key(&wq->key);
3898 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3899 if (!lock_name)
3900 lock_name = wq->name;
3901
3902 wq->lock_name = lock_name;
3903 lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3904}
3905
3906static void wq_unregister_lockdep(struct workqueue_struct *wq)
3907{
3908 lockdep_unregister_key(&wq->key);
3909}
3910
3911static void wq_free_lockdep(struct workqueue_struct *wq)
3912{
3913 if (wq->lock_name != wq->name)
3914 kfree(wq->lock_name);
3915}
3916#else
3917static void wq_init_lockdep(struct workqueue_struct *wq)
3918{
3919}
3920
3921static void wq_unregister_lockdep(struct workqueue_struct *wq)
3922{
3923}
3924
3925static void wq_free_lockdep(struct workqueue_struct *wq)
3926{
3927}
3928#endif
3929
3930static void rcu_free_wq(struct rcu_head *rcu)
3931{
3932 struct workqueue_struct *wq =
3933 container_of(rcu, struct workqueue_struct, rcu);
3934
3935 wq_free_lockdep(wq);
3936 free_percpu(wq->cpu_pwq);
3937 free_workqueue_attrs(wq->unbound_attrs);
3938 kfree(wq);
3939}
3940
3941static void rcu_free_pool(struct rcu_head *rcu)
3942{
3943 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3944
3945 ida_destroy(&pool->worker_ida);
3946 free_workqueue_attrs(pool->attrs);
3947 kfree(pool);
3948}
3949
3950/**
3951 * put_unbound_pool - put a worker_pool
3952 * @pool: worker_pool to put
3953 *
3954 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
3955 * safe manner. get_unbound_pool() calls this function on its failure path
3956 * and this function should be able to release pools which went through,
3957 * successfully or not, init_worker_pool().
3958 *
3959 * Should be called with wq_pool_mutex held.
3960 */
3961static void put_unbound_pool(struct worker_pool *pool)
3962{
3963 DECLARE_COMPLETION_ONSTACK(detach_completion);
3964 struct worker *worker;
3965 LIST_HEAD(cull_list);
3966
3967 lockdep_assert_held(&wq_pool_mutex);
3968
3969 if (--pool->refcnt)
3970 return;
3971
3972 /* sanity checks */
3973 if (WARN_ON(!(pool->cpu < 0)) ||
3974 WARN_ON(!list_empty(&pool->worklist)))
3975 return;
3976
3977 /* release id and unhash */
3978 if (pool->id >= 0)
3979 idr_remove(&worker_pool_idr, pool->id);
3980 hash_del(&pool->hash_node);
3981
3982 /*
3983 * Become the manager and destroy all workers. This prevents
3984 * @pool's workers from blocking on attach_mutex. We're the last
3985 * manager and @pool gets freed with the flag set.
3986 *
3987 * Having a concurrent manager is quite unlikely to happen as we can
3988 * only get here with
3989 * pwq->refcnt == pool->refcnt == 0
3990 * which implies no work queued to the pool, which implies no worker can
3991 * become the manager. However a worker could have taken the role of
3992 * manager before the refcnts dropped to 0, since maybe_create_worker()
3993 * drops pool->lock
3994 */
3995 while (true) {
3996 rcuwait_wait_event(&manager_wait,
3997 !(pool->flags & POOL_MANAGER_ACTIVE),
3998 TASK_UNINTERRUPTIBLE);
3999
4000 mutex_lock(&wq_pool_attach_mutex);
4001 raw_spin_lock_irq(&pool->lock);
4002 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4003 pool->flags |= POOL_MANAGER_ACTIVE;
4004 break;
4005 }
4006 raw_spin_unlock_irq(&pool->lock);
4007 mutex_unlock(&wq_pool_attach_mutex);
4008 }
4009
4010 while ((worker = first_idle_worker(pool)))
4011 set_worker_dying(worker, &cull_list);
4012 WARN_ON(pool->nr_workers || pool->nr_idle);
4013 raw_spin_unlock_irq(&pool->lock);
4014
4015 wake_dying_workers(&cull_list);
4016
4017 if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers))
4018 pool->detach_completion = &detach_completion;
4019 mutex_unlock(&wq_pool_attach_mutex);
4020
4021 if (pool->detach_completion)
4022 wait_for_completion(pool->detach_completion);
4023
4024 /* shut down the timers */
4025 del_timer_sync(&pool->idle_timer);
4026 cancel_work_sync(&pool->idle_cull_work);
4027 del_timer_sync(&pool->mayday_timer);
4028
4029 /* RCU protected to allow dereferences from get_work_pool() */
4030 call_rcu(&pool->rcu, rcu_free_pool);
4031}
4032
4033/**
4034 * get_unbound_pool - get a worker_pool with the specified attributes
4035 * @attrs: the attributes of the worker_pool to get
4036 *
4037 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4038 * reference count and return it. If there already is a matching
4039 * worker_pool, it will be used; otherwise, this function attempts to
4040 * create a new one.
4041 *
4042 * Should be called with wq_pool_mutex held.
4043 *
4044 * Return: On success, a worker_pool with the same attributes as @attrs.
4045 * On failure, %NULL.
4046 */
4047static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
4048{
4049 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
4050 u32 hash = wqattrs_hash(attrs);
4051 struct worker_pool *pool;
4052 int pod, node = NUMA_NO_NODE;
4053
4054 lockdep_assert_held(&wq_pool_mutex);
4055
4056 /* do we already have a matching pool? */
4057 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
4058 if (wqattrs_equal(pool->attrs, attrs)) {
4059 pool->refcnt++;
4060 return pool;
4061 }
4062 }
4063
4064 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
4065 for (pod = 0; pod < pt->nr_pods; pod++) {
4066 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
4067 node = pt->pod_node[pod];
4068 break;
4069 }
4070 }
4071
4072 /* nope, create a new one */
4073 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
4074 if (!pool || init_worker_pool(pool) < 0)
4075 goto fail;
4076
4077 pool->node = node;
4078 copy_workqueue_attrs(pool->attrs, attrs);
4079 wqattrs_clear_for_pool(pool->attrs);
4080
4081 if (worker_pool_assign_id(pool) < 0)
4082 goto fail;
4083
4084 /* create and start the initial worker */
4085 if (wq_online && !create_worker(pool))
4086 goto fail;
4087
4088 /* install */
4089 hash_add(unbound_pool_hash, &pool->hash_node, hash);
4090
4091 return pool;
4092fail:
4093 if (pool)
4094 put_unbound_pool(pool);
4095 return NULL;
4096}
4097
4098static void rcu_free_pwq(struct rcu_head *rcu)
4099{
4100 kmem_cache_free(pwq_cache,
4101 container_of(rcu, struct pool_workqueue, rcu));
4102}
4103
4104/*
4105 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
4106 * refcnt and needs to be destroyed.
4107 */
4108static void pwq_release_workfn(struct kthread_work *work)
4109{
4110 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
4111 release_work);
4112 struct workqueue_struct *wq = pwq->wq;
4113 struct worker_pool *pool = pwq->pool;
4114 bool is_last = false;
4115
4116 /*
4117 * When @pwq is not linked, it doesn't hold any reference to the
4118 * @wq, and @wq is invalid to access.
4119 */
4120 if (!list_empty(&pwq->pwqs_node)) {
4121 mutex_lock(&wq->mutex);
4122 list_del_rcu(&pwq->pwqs_node);
4123 is_last = list_empty(&wq->pwqs);
4124 mutex_unlock(&wq->mutex);
4125 }
4126
4127 if (wq->flags & WQ_UNBOUND) {
4128 mutex_lock(&wq_pool_mutex);
4129 put_unbound_pool(pool);
4130 mutex_unlock(&wq_pool_mutex);
4131 }
4132
4133 call_rcu(&pwq->rcu, rcu_free_pwq);
4134
4135 /*
4136 * If we're the last pwq going away, @wq is already dead and no one
4137 * is gonna access it anymore. Schedule RCU free.
4138 */
4139 if (is_last) {
4140 wq_unregister_lockdep(wq);
4141 call_rcu(&wq->rcu, rcu_free_wq);
4142 }
4143}
4144
4145/**
4146 * pwq_adjust_max_active - update a pwq's max_active to the current setting
4147 * @pwq: target pool_workqueue
4148 *
4149 * If @pwq isn't freezing, set @pwq->max_active to the associated
4150 * workqueue's saved_max_active and activate inactive work items
4151 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
4152 */
4153static void pwq_adjust_max_active(struct pool_workqueue *pwq)
4154{
4155 struct workqueue_struct *wq = pwq->wq;
4156 bool freezable = wq->flags & WQ_FREEZABLE;
4157 unsigned long flags;
4158
4159 /* for @wq->saved_max_active */
4160 lockdep_assert_held(&wq->mutex);
4161
4162 /* fast exit for non-freezable wqs */
4163 if (!freezable && pwq->max_active == wq->saved_max_active)
4164 return;
4165
4166 /* this function can be called during early boot w/ irq disabled */
4167 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
4168
4169 /*
4170 * During [un]freezing, the caller is responsible for ensuring that
4171 * this function is called at least once after @workqueue_freezing
4172 * is updated and visible.
4173 */
4174 if (!freezable || !workqueue_freezing) {
4175 pwq->max_active = wq->saved_max_active;
4176
4177 while (!list_empty(&pwq->inactive_works) &&
4178 pwq->nr_active < pwq->max_active)
4179 pwq_activate_first_inactive(pwq);
4180
4181 kick_pool(pwq->pool);
4182 } else {
4183 pwq->max_active = 0;
4184 }
4185
4186 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
4187}
4188
4189/* initialize newly allocated @pwq which is associated with @wq and @pool */
4190static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
4191 struct worker_pool *pool)
4192{
4193 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
4194
4195 memset(pwq, 0, sizeof(*pwq));
4196
4197 pwq->pool = pool;
4198 pwq->wq = wq;
4199 pwq->flush_color = -1;
4200 pwq->refcnt = 1;
4201 INIT_LIST_HEAD(&pwq->inactive_works);
4202 INIT_LIST_HEAD(&pwq->pwqs_node);
4203 INIT_LIST_HEAD(&pwq->mayday_node);
4204 kthread_init_work(&pwq->release_work, pwq_release_workfn);
4205}
4206
4207/* sync @pwq with the current state of its associated wq and link it */
4208static void link_pwq(struct pool_workqueue *pwq)
4209{
4210 struct workqueue_struct *wq = pwq->wq;
4211
4212 lockdep_assert_held(&wq->mutex);
4213
4214 /* may be called multiple times, ignore if already linked */
4215 if (!list_empty(&pwq->pwqs_node))
4216 return;
4217
4218 /* set the matching work_color */
4219 pwq->work_color = wq->work_color;
4220
4221 /* sync max_active to the current setting */
4222 pwq_adjust_max_active(pwq);
4223
4224 /* link in @pwq */
4225 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
4226}
4227
4228/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
4229static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
4230 const struct workqueue_attrs *attrs)
4231{
4232 struct worker_pool *pool;
4233 struct pool_workqueue *pwq;
4234
4235 lockdep_assert_held(&wq_pool_mutex);
4236
4237 pool = get_unbound_pool(attrs);
4238 if (!pool)
4239 return NULL;
4240
4241 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
4242 if (!pwq) {
4243 put_unbound_pool(pool);
4244 return NULL;
4245 }
4246
4247 init_pwq(pwq, wq, pool);
4248 return pwq;
4249}
4250
4251/**
4252 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
4253 * @attrs: the wq_attrs of the default pwq of the target workqueue
4254 * @cpu: the target CPU
4255 * @cpu_going_down: if >= 0, the CPU to consider as offline
4256 *
4257 * Calculate the cpumask a workqueue with @attrs should use on @pod. If
4258 * @cpu_going_down is >= 0, that cpu is considered offline during calculation.
4259 * The result is stored in @attrs->__pod_cpumask.
4260 *
4261 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
4262 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
4263 * intersection of the possible CPUs of @pod and @attrs->cpumask.
4264 *
4265 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
4266 */
4267static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu,
4268 int cpu_going_down)
4269{
4270 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
4271 int pod = pt->cpu_pod[cpu];
4272
4273 /* does @pod have any online CPUs @attrs wants? */
4274 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
4275 cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask);
4276 if (cpu_going_down >= 0)
4277 cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask);
4278
4279 if (cpumask_empty(attrs->__pod_cpumask)) {
4280 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
4281 return;
4282 }
4283
4284 /* yeap, return possible CPUs in @pod that @attrs wants */
4285 cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]);
4286
4287 if (cpumask_empty(attrs->__pod_cpumask))
4288 pr_warn_once("WARNING: workqueue cpumask: online intersect > "
4289 "possible intersect\n");
4290}
4291
4292/* install @pwq into @wq's cpu_pwq and return the old pwq */
4293static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
4294 int cpu, struct pool_workqueue *pwq)
4295{
4296 struct pool_workqueue *old_pwq;
4297
4298 lockdep_assert_held(&wq_pool_mutex);
4299 lockdep_assert_held(&wq->mutex);
4300
4301 /* link_pwq() can handle duplicate calls */
4302 link_pwq(pwq);
4303
4304 old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4305 rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq);
4306 return old_pwq;
4307}
4308
4309/* context to store the prepared attrs & pwqs before applying */
4310struct apply_wqattrs_ctx {
4311 struct workqueue_struct *wq; /* target workqueue */
4312 struct workqueue_attrs *attrs; /* attrs to apply */
4313 struct list_head list; /* queued for batching commit */
4314 struct pool_workqueue *dfl_pwq;
4315 struct pool_workqueue *pwq_tbl[];
4316};
4317
4318/* free the resources after success or abort */
4319static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
4320{
4321 if (ctx) {
4322 int cpu;
4323
4324 for_each_possible_cpu(cpu)
4325 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
4326 put_pwq_unlocked(ctx->dfl_pwq);
4327
4328 free_workqueue_attrs(ctx->attrs);
4329
4330 kfree(ctx);
4331 }
4332}
4333
4334/* allocate the attrs and pwqs for later installation */
4335static struct apply_wqattrs_ctx *
4336apply_wqattrs_prepare(struct workqueue_struct *wq,
4337 const struct workqueue_attrs *attrs,
4338 const cpumask_var_t unbound_cpumask)
4339{
4340 struct apply_wqattrs_ctx *ctx;
4341 struct workqueue_attrs *new_attrs;
4342 int cpu;
4343
4344 lockdep_assert_held(&wq_pool_mutex);
4345
4346 if (WARN_ON(attrs->affn_scope < 0 ||
4347 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
4348 return ERR_PTR(-EINVAL);
4349
4350 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
4351
4352 new_attrs = alloc_workqueue_attrs();
4353 if (!ctx || !new_attrs)
4354 goto out_free;
4355
4356 /*
4357 * If something goes wrong during CPU up/down, we'll fall back to
4358 * the default pwq covering whole @attrs->cpumask. Always create
4359 * it even if we don't use it immediately.
4360 */
4361 copy_workqueue_attrs(new_attrs, attrs);
4362 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
4363 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4364 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
4365 if (!ctx->dfl_pwq)
4366 goto out_free;
4367
4368 for_each_possible_cpu(cpu) {
4369 if (new_attrs->ordered) {
4370 ctx->dfl_pwq->refcnt++;
4371 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
4372 } else {
4373 wq_calc_pod_cpumask(new_attrs, cpu, -1);
4374 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
4375 if (!ctx->pwq_tbl[cpu])
4376 goto out_free;
4377 }
4378 }
4379
4380 /* save the user configured attrs and sanitize it. */
4381 copy_workqueue_attrs(new_attrs, attrs);
4382 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
4383 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
4384 ctx->attrs = new_attrs;
4385
4386 ctx->wq = wq;
4387 return ctx;
4388
4389out_free:
4390 free_workqueue_attrs(new_attrs);
4391 apply_wqattrs_cleanup(ctx);
4392 return ERR_PTR(-ENOMEM);
4393}
4394
4395/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
4396static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
4397{
4398 int cpu;
4399
4400 /* all pwqs have been created successfully, let's install'em */
4401 mutex_lock(&ctx->wq->mutex);
4402
4403 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
4404
4405 /* save the previous pwq and install the new one */
4406 for_each_possible_cpu(cpu)
4407 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
4408 ctx->pwq_tbl[cpu]);
4409
4410 /* @dfl_pwq might not have been used, ensure it's linked */
4411 link_pwq(ctx->dfl_pwq);
4412 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
4413
4414 mutex_unlock(&ctx->wq->mutex);
4415}
4416
4417static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
4418 const struct workqueue_attrs *attrs)
4419{
4420 struct apply_wqattrs_ctx *ctx;
4421
4422 /* only unbound workqueues can change attributes */
4423 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
4424 return -EINVAL;
4425
4426 /* creating multiple pwqs breaks ordering guarantee */
4427 if (!list_empty(&wq->pwqs)) {
4428 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4429 return -EINVAL;
4430
4431 wq->flags &= ~__WQ_ORDERED;
4432 }
4433
4434 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
4435 if (IS_ERR(ctx))
4436 return PTR_ERR(ctx);
4437
4438 /* the ctx has been prepared successfully, let's commit it */
4439 apply_wqattrs_commit(ctx);
4440 apply_wqattrs_cleanup(ctx);
4441
4442 return 0;
4443}
4444
4445/**
4446 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4447 * @wq: the target workqueue
4448 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4449 *
4450 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
4451 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
4452 * work items are affine to the pod it was issued on. Older pwqs are released as
4453 * in-flight work items finish. Note that a work item which repeatedly requeues
4454 * itself back-to-back will stay on its current pwq.
4455 *
4456 * Performs GFP_KERNEL allocations.
4457 *
4458 * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock().
4459 *
4460 * Return: 0 on success and -errno on failure.
4461 */
4462int apply_workqueue_attrs(struct workqueue_struct *wq,
4463 const struct workqueue_attrs *attrs)
4464{
4465 int ret;
4466
4467 lockdep_assert_cpus_held();
4468
4469 mutex_lock(&wq_pool_mutex);
4470 ret = apply_workqueue_attrs_locked(wq, attrs);
4471 mutex_unlock(&wq_pool_mutex);
4472
4473 return ret;
4474}
4475
4476/**
4477 * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug
4478 * @wq: the target workqueue
4479 * @cpu: the CPU to update pool association for
4480 * @hotplug_cpu: the CPU coming up or going down
4481 * @online: whether @cpu is coming up or going down
4482 *
4483 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4484 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of
4485 * @wq accordingly.
4486 *
4487 *
4488 * If pod affinity can't be adjusted due to memory allocation failure, it falls
4489 * back to @wq->dfl_pwq which may not be optimal but is always correct.
4490 *
4491 * Note that when the last allowed CPU of a pod goes offline for a workqueue
4492 * with a cpumask spanning multiple pods, the workers which were already
4493 * executing the work items for the workqueue will lose their CPU affinity and
4494 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
4495 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
4496 * responsibility to flush the work item from CPU_DOWN_PREPARE.
4497 */
4498static void wq_update_pod(struct workqueue_struct *wq, int cpu,
4499 int hotplug_cpu, bool online)
4500{
4501 int off_cpu = online ? -1 : hotplug_cpu;
4502 struct pool_workqueue *old_pwq = NULL, *pwq;
4503 struct workqueue_attrs *target_attrs;
4504
4505 lockdep_assert_held(&wq_pool_mutex);
4506
4507 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
4508 return;
4509
4510 /*
4511 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4512 * Let's use a preallocated one. The following buf is protected by
4513 * CPU hotplug exclusion.
4514 */
4515 target_attrs = wq_update_pod_attrs_buf;
4516
4517 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4518 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
4519
4520 /* nothing to do if the target cpumask matches the current pwq */
4521 wq_calc_pod_cpumask(target_attrs, cpu, off_cpu);
4522 pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu),
4523 lockdep_is_held(&wq_pool_mutex));
4524 if (wqattrs_equal(target_attrs, pwq->pool->attrs))
4525 return;
4526
4527 /* create a new pwq */
4528 pwq = alloc_unbound_pwq(wq, target_attrs);
4529 if (!pwq) {
4530 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
4531 wq->name);
4532 goto use_dfl_pwq;
4533 }
4534
4535 /* Install the new pwq. */
4536 mutex_lock(&wq->mutex);
4537 old_pwq = install_unbound_pwq(wq, cpu, pwq);
4538 goto out_unlock;
4539
4540use_dfl_pwq:
4541 mutex_lock(&wq->mutex);
4542 raw_spin_lock_irq(&wq->dfl_pwq->pool->lock);
4543 get_pwq(wq->dfl_pwq);
4544 raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4545 old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq);
4546out_unlock:
4547 mutex_unlock(&wq->mutex);
4548 put_pwq_unlocked(old_pwq);
4549}
4550
4551static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4552{
4553 bool highpri = wq->flags & WQ_HIGHPRI;
4554 int cpu, ret;
4555
4556 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
4557 if (!wq->cpu_pwq)
4558 goto enomem;
4559
4560 if (!(wq->flags & WQ_UNBOUND)) {
4561 for_each_possible_cpu(cpu) {
4562 struct pool_workqueue **pwq_p =
4563 per_cpu_ptr(wq->cpu_pwq, cpu);
4564 struct worker_pool *pool =
4565 &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]);
4566
4567 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
4568 pool->node);
4569 if (!*pwq_p)
4570 goto enomem;
4571
4572 init_pwq(*pwq_p, wq, pool);
4573
4574 mutex_lock(&wq->mutex);
4575 link_pwq(*pwq_p);
4576 mutex_unlock(&wq->mutex);
4577 }
4578 return 0;
4579 }
4580
4581 cpus_read_lock();
4582 if (wq->flags & __WQ_ORDERED) {
4583 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4584 /* there should only be single pwq for ordering guarantee */
4585 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4586 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4587 "ordering guarantee broken for workqueue %s\n", wq->name);
4588 } else {
4589 ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4590 }
4591 cpus_read_unlock();
4592
4593 /* for unbound pwq, flush the pwq_release_worker ensures that the
4594 * pwq_release_workfn() completes before calling kfree(wq).
4595 */
4596 if (ret)
4597 kthread_flush_worker(pwq_release_worker);
4598
4599 return ret;
4600
4601enomem:
4602 if (wq->cpu_pwq) {
4603 for_each_possible_cpu(cpu) {
4604 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4605
4606 if (pwq)
4607 kmem_cache_free(pwq_cache, pwq);
4608 }
4609 free_percpu(wq->cpu_pwq);
4610 wq->cpu_pwq = NULL;
4611 }
4612 return -ENOMEM;
4613}
4614
4615static int wq_clamp_max_active(int max_active, unsigned int flags,
4616 const char *name)
4617{
4618 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
4619 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4620 max_active, name, 1, WQ_MAX_ACTIVE);
4621
4622 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
4623}
4624
4625/*
4626 * Workqueues which may be used during memory reclaim should have a rescuer
4627 * to guarantee forward progress.
4628 */
4629static int init_rescuer(struct workqueue_struct *wq)
4630{
4631 struct worker *rescuer;
4632 int ret;
4633
4634 if (!(wq->flags & WQ_MEM_RECLAIM))
4635 return 0;
4636
4637 rescuer = alloc_worker(NUMA_NO_NODE);
4638 if (!rescuer) {
4639 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
4640 wq->name);
4641 return -ENOMEM;
4642 }
4643
4644 rescuer->rescue_wq = wq;
4645 rescuer->task = kthread_create(rescuer_thread, rescuer, "kworker/R-%s", wq->name);
4646 if (IS_ERR(rescuer->task)) {
4647 ret = PTR_ERR(rescuer->task);
4648 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
4649 wq->name, ERR_PTR(ret));
4650 kfree(rescuer);
4651 return ret;
4652 }
4653
4654 wq->rescuer = rescuer;
4655 kthread_bind_mask(rescuer->task, cpu_possible_mask);
4656 wake_up_process(rescuer->task);
4657
4658 return 0;
4659}
4660
4661__printf(1, 4)
4662struct workqueue_struct *alloc_workqueue(const char *fmt,
4663 unsigned int flags,
4664 int max_active, ...)
4665{
4666 va_list args;
4667 struct workqueue_struct *wq;
4668 struct pool_workqueue *pwq;
4669
4670 /*
4671 * Unbound && max_active == 1 used to imply ordered, which is no longer
4672 * the case on many machines due to per-pod pools. While
4673 * alloc_ordered_workqueue() is the right way to create an ordered
4674 * workqueue, keep the previous behavior to avoid subtle breakages.
4675 */
4676 if ((flags & WQ_UNBOUND) && max_active == 1)
4677 flags |= __WQ_ORDERED;
4678
4679 /* see the comment above the definition of WQ_POWER_EFFICIENT */
4680 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4681 flags |= WQ_UNBOUND;
4682
4683 /* allocate wq and format name */
4684 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
4685 if (!wq)
4686 return NULL;
4687
4688 if (flags & WQ_UNBOUND) {
4689 wq->unbound_attrs = alloc_workqueue_attrs();
4690 if (!wq->unbound_attrs)
4691 goto err_free_wq;
4692 }
4693
4694 va_start(args, max_active);
4695 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4696 va_end(args);
4697
4698 max_active = max_active ?: WQ_DFL_ACTIVE;
4699 max_active = wq_clamp_max_active(max_active, flags, wq->name);
4700
4701 /* init wq */
4702 wq->flags = flags;
4703 wq->saved_max_active = max_active;
4704 mutex_init(&wq->mutex);
4705 atomic_set(&wq->nr_pwqs_to_flush, 0);
4706 INIT_LIST_HEAD(&wq->pwqs);
4707 INIT_LIST_HEAD(&wq->flusher_queue);
4708 INIT_LIST_HEAD(&wq->flusher_overflow);
4709 INIT_LIST_HEAD(&wq->maydays);
4710
4711 wq_init_lockdep(wq);
4712 INIT_LIST_HEAD(&wq->list);
4713
4714 if (alloc_and_link_pwqs(wq) < 0)
4715 goto err_unreg_lockdep;
4716
4717 if (wq_online && init_rescuer(wq) < 0)
4718 goto err_destroy;
4719
4720 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4721 goto err_destroy;
4722
4723 /*
4724 * wq_pool_mutex protects global freeze state and workqueues list.
4725 * Grab it, adjust max_active and add the new @wq to workqueues
4726 * list.
4727 */
4728 mutex_lock(&wq_pool_mutex);
4729
4730 mutex_lock(&wq->mutex);
4731 for_each_pwq(pwq, wq)
4732 pwq_adjust_max_active(pwq);
4733 mutex_unlock(&wq->mutex);
4734
4735 list_add_tail_rcu(&wq->list, &workqueues);
4736
4737 mutex_unlock(&wq_pool_mutex);
4738
4739 return wq;
4740
4741err_unreg_lockdep:
4742 wq_unregister_lockdep(wq);
4743 wq_free_lockdep(wq);
4744err_free_wq:
4745 free_workqueue_attrs(wq->unbound_attrs);
4746 kfree(wq);
4747 return NULL;
4748err_destroy:
4749 destroy_workqueue(wq);
4750 return NULL;
4751}
4752EXPORT_SYMBOL_GPL(alloc_workqueue);
4753
4754static bool pwq_busy(struct pool_workqueue *pwq)
4755{
4756 int i;
4757
4758 for (i = 0; i < WORK_NR_COLORS; i++)
4759 if (pwq->nr_in_flight[i])
4760 return true;
4761
4762 if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1))
4763 return true;
4764 if (pwq->nr_active || !list_empty(&pwq->inactive_works))
4765 return true;
4766
4767 return false;
4768}
4769
4770/**
4771 * destroy_workqueue - safely terminate a workqueue
4772 * @wq: target workqueue
4773 *
4774 * Safely destroy a workqueue. All work currently pending will be done first.
4775 */
4776void destroy_workqueue(struct workqueue_struct *wq)
4777{
4778 struct pool_workqueue *pwq;
4779 int cpu;
4780
4781 /*
4782 * Remove it from sysfs first so that sanity check failure doesn't
4783 * lead to sysfs name conflicts.
4784 */
4785 workqueue_sysfs_unregister(wq);
4786
4787 /* mark the workqueue destruction is in progress */
4788 mutex_lock(&wq->mutex);
4789 wq->flags |= __WQ_DESTROYING;
4790 mutex_unlock(&wq->mutex);
4791
4792 /* drain it before proceeding with destruction */
4793 drain_workqueue(wq);
4794
4795 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
4796 if (wq->rescuer) {
4797 struct worker *rescuer = wq->rescuer;
4798
4799 /* this prevents new queueing */
4800 raw_spin_lock_irq(&wq_mayday_lock);
4801 wq->rescuer = NULL;
4802 raw_spin_unlock_irq(&wq_mayday_lock);
4803
4804 /* rescuer will empty maydays list before exiting */
4805 kthread_stop(rescuer->task);
4806 kfree(rescuer);
4807 }
4808
4809 /*
4810 * Sanity checks - grab all the locks so that we wait for all
4811 * in-flight operations which may do put_pwq().
4812 */
4813 mutex_lock(&wq_pool_mutex);
4814 mutex_lock(&wq->mutex);
4815 for_each_pwq(pwq, wq) {
4816 raw_spin_lock_irq(&pwq->pool->lock);
4817 if (WARN_ON(pwq_busy(pwq))) {
4818 pr_warn("%s: %s has the following busy pwq\n",
4819 __func__, wq->name);
4820 show_pwq(pwq);
4821 raw_spin_unlock_irq(&pwq->pool->lock);
4822 mutex_unlock(&wq->mutex);
4823 mutex_unlock(&wq_pool_mutex);
4824 show_one_workqueue(wq);
4825 return;
4826 }
4827 raw_spin_unlock_irq(&pwq->pool->lock);
4828 }
4829 mutex_unlock(&wq->mutex);
4830
4831 /*
4832 * wq list is used to freeze wq, remove from list after
4833 * flushing is complete in case freeze races us.
4834 */
4835 list_del_rcu(&wq->list);
4836 mutex_unlock(&wq_pool_mutex);
4837
4838 /*
4839 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
4840 * to put the base refs. @wq will be auto-destroyed from the last
4841 * pwq_put. RCU read lock prevents @wq from going away from under us.
4842 */
4843 rcu_read_lock();
4844
4845 for_each_possible_cpu(cpu) {
4846 pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu));
4847 RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL);
4848 put_pwq_unlocked(pwq);
4849 }
4850
4851 put_pwq_unlocked(wq->dfl_pwq);
4852 wq->dfl_pwq = NULL;
4853
4854 rcu_read_unlock();
4855}
4856EXPORT_SYMBOL_GPL(destroy_workqueue);
4857
4858/**
4859 * workqueue_set_max_active - adjust max_active of a workqueue
4860 * @wq: target workqueue
4861 * @max_active: new max_active value.
4862 *
4863 * Set max_active of @wq to @max_active.
4864 *
4865 * CONTEXT:
4866 * Don't call from IRQ context.
4867 */
4868void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4869{
4870 struct pool_workqueue *pwq;
4871
4872 /* disallow meddling with max_active for ordered workqueues */
4873 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4874 return;
4875
4876 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4877
4878 mutex_lock(&wq->mutex);
4879
4880 wq->flags &= ~__WQ_ORDERED;
4881 wq->saved_max_active = max_active;
4882
4883 for_each_pwq(pwq, wq)
4884 pwq_adjust_max_active(pwq);
4885
4886 mutex_unlock(&wq->mutex);
4887}
4888EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4889
4890/**
4891 * current_work - retrieve %current task's work struct
4892 *
4893 * Determine if %current task is a workqueue worker and what it's working on.
4894 * Useful to find out the context that the %current task is running in.
4895 *
4896 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4897 */
4898struct work_struct *current_work(void)
4899{
4900 struct worker *worker = current_wq_worker();
4901
4902 return worker ? worker->current_work : NULL;
4903}
4904EXPORT_SYMBOL(current_work);
4905
4906/**
4907 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4908 *
4909 * Determine whether %current is a workqueue rescuer. Can be used from
4910 * work functions to determine whether it's being run off the rescuer task.
4911 *
4912 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4913 */
4914bool current_is_workqueue_rescuer(void)
4915{
4916 struct worker *worker = current_wq_worker();
4917
4918 return worker && worker->rescue_wq;
4919}
4920
4921/**
4922 * workqueue_congested - test whether a workqueue is congested
4923 * @cpu: CPU in question
4924 * @wq: target workqueue
4925 *
4926 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4927 * no synchronization around this function and the test result is
4928 * unreliable and only useful as advisory hints or for debugging.
4929 *
4930 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4931 *
4932 * With the exception of ordered workqueues, all workqueues have per-cpu
4933 * pool_workqueues, each with its own congested state. A workqueue being
4934 * congested on one CPU doesn't mean that the workqueue is contested on any
4935 * other CPUs.
4936 *
4937 * Return:
4938 * %true if congested, %false otherwise.
4939 */
4940bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4941{
4942 struct pool_workqueue *pwq;
4943 bool ret;
4944
4945 rcu_read_lock();
4946 preempt_disable();
4947
4948 if (cpu == WORK_CPU_UNBOUND)
4949 cpu = smp_processor_id();
4950
4951 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
4952 ret = !list_empty(&pwq->inactive_works);
4953
4954 preempt_enable();
4955 rcu_read_unlock();
4956
4957 return ret;
4958}
4959EXPORT_SYMBOL_GPL(workqueue_congested);
4960
4961/**
4962 * work_busy - test whether a work is currently pending or running
4963 * @work: the work to be tested
4964 *
4965 * Test whether @work is currently pending or running. There is no
4966 * synchronization around this function and the test result is
4967 * unreliable and only useful as advisory hints or for debugging.
4968 *
4969 * Return:
4970 * OR'd bitmask of WORK_BUSY_* bits.
4971 */
4972unsigned int work_busy(struct work_struct *work)
4973{
4974 struct worker_pool *pool;
4975 unsigned long flags;
4976 unsigned int ret = 0;
4977
4978 if (work_pending(work))
4979 ret |= WORK_BUSY_PENDING;
4980
4981 rcu_read_lock();
4982 pool = get_work_pool(work);
4983 if (pool) {
4984 raw_spin_lock_irqsave(&pool->lock, flags);
4985 if (find_worker_executing_work(pool, work))
4986 ret |= WORK_BUSY_RUNNING;
4987 raw_spin_unlock_irqrestore(&pool->lock, flags);
4988 }
4989 rcu_read_unlock();
4990
4991 return ret;
4992}
4993EXPORT_SYMBOL_GPL(work_busy);
4994
4995/**
4996 * set_worker_desc - set description for the current work item
4997 * @fmt: printf-style format string
4998 * @...: arguments for the format string
4999 *
5000 * This function can be called by a running work function to describe what
5001 * the work item is about. If the worker task gets dumped, this
5002 * information will be printed out together to help debugging. The
5003 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
5004 */
5005void set_worker_desc(const char *fmt, ...)
5006{
5007 struct worker *worker = current_wq_worker();
5008 va_list args;
5009
5010 if (worker) {
5011 va_start(args, fmt);
5012 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
5013 va_end(args);
5014 }
5015}
5016EXPORT_SYMBOL_GPL(set_worker_desc);
5017
5018/**
5019 * print_worker_info - print out worker information and description
5020 * @log_lvl: the log level to use when printing
5021 * @task: target task
5022 *
5023 * If @task is a worker and currently executing a work item, print out the
5024 * name of the workqueue being serviced and worker description set with
5025 * set_worker_desc() by the currently executing work item.
5026 *
5027 * This function can be safely called on any task as long as the
5028 * task_struct itself is accessible. While safe, this function isn't
5029 * synchronized and may print out mixups or garbages of limited length.
5030 */
5031void print_worker_info(const char *log_lvl, struct task_struct *task)
5032{
5033 work_func_t *fn = NULL;
5034 char name[WQ_NAME_LEN] = { };
5035 char desc[WORKER_DESC_LEN] = { };
5036 struct pool_workqueue *pwq = NULL;
5037 struct workqueue_struct *wq = NULL;
5038 struct worker *worker;
5039
5040 if (!(task->flags & PF_WQ_WORKER))
5041 return;
5042
5043 /*
5044 * This function is called without any synchronization and @task
5045 * could be in any state. Be careful with dereferences.
5046 */
5047 worker = kthread_probe_data(task);
5048
5049 /*
5050 * Carefully copy the associated workqueue's workfn, name and desc.
5051 * Keep the original last '\0' in case the original is garbage.
5052 */
5053 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
5054 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
5055 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
5056 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
5057 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
5058
5059 if (fn || name[0] || desc[0]) {
5060 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
5061 if (strcmp(name, desc))
5062 pr_cont(" (%s)", desc);
5063 pr_cont("\n");
5064 }
5065}
5066
5067static void pr_cont_pool_info(struct worker_pool *pool)
5068{
5069 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
5070 if (pool->node != NUMA_NO_NODE)
5071 pr_cont(" node=%d", pool->node);
5072 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
5073}
5074
5075struct pr_cont_work_struct {
5076 bool comma;
5077 work_func_t func;
5078 long ctr;
5079};
5080
5081static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
5082{
5083 if (!pcwsp->ctr)
5084 goto out_record;
5085 if (func == pcwsp->func) {
5086 pcwsp->ctr++;
5087 return;
5088 }
5089 if (pcwsp->ctr == 1)
5090 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
5091 else
5092 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
5093 pcwsp->ctr = 0;
5094out_record:
5095 if ((long)func == -1L)
5096 return;
5097 pcwsp->comma = comma;
5098 pcwsp->func = func;
5099 pcwsp->ctr = 1;
5100}
5101
5102static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
5103{
5104 if (work->func == wq_barrier_func) {
5105 struct wq_barrier *barr;
5106
5107 barr = container_of(work, struct wq_barrier, work);
5108
5109 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5110 pr_cont("%s BAR(%d)", comma ? "," : "",
5111 task_pid_nr(barr->task));
5112 } else {
5113 if (!comma)
5114 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
5115 pr_cont_work_flush(comma, work->func, pcwsp);
5116 }
5117}
5118
5119static void show_pwq(struct pool_workqueue *pwq)
5120{
5121 struct pr_cont_work_struct pcws = { .ctr = 0, };
5122 struct worker_pool *pool = pwq->pool;
5123 struct work_struct *work;
5124 struct worker *worker;
5125 bool has_in_flight = false, has_pending = false;
5126 int bkt;
5127
5128 pr_info(" pwq %d:", pool->id);
5129 pr_cont_pool_info(pool);
5130
5131 pr_cont(" active=%d/%d refcnt=%d%s\n",
5132 pwq->nr_active, pwq->max_active, pwq->refcnt,
5133 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
5134
5135 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5136 if (worker->current_pwq == pwq) {
5137 has_in_flight = true;
5138 break;
5139 }
5140 }
5141 if (has_in_flight) {
5142 bool comma = false;
5143
5144 pr_info(" in-flight:");
5145 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
5146 if (worker->current_pwq != pwq)
5147 continue;
5148
5149 pr_cont("%s %d%s:%ps", comma ? "," : "",
5150 task_pid_nr(worker->task),
5151 worker->rescue_wq ? "(RESCUER)" : "",
5152 worker->current_func);
5153 list_for_each_entry(work, &worker->scheduled, entry)
5154 pr_cont_work(false, work, &pcws);
5155 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5156 comma = true;
5157 }
5158 pr_cont("\n");
5159 }
5160
5161 list_for_each_entry(work, &pool->worklist, entry) {
5162 if (get_work_pwq(work) == pwq) {
5163 has_pending = true;
5164 break;
5165 }
5166 }
5167 if (has_pending) {
5168 bool comma = false;
5169
5170 pr_info(" pending:");
5171 list_for_each_entry(work, &pool->worklist, entry) {
5172 if (get_work_pwq(work) != pwq)
5173 continue;
5174
5175 pr_cont_work(comma, work, &pcws);
5176 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5177 }
5178 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5179 pr_cont("\n");
5180 }
5181
5182 if (!list_empty(&pwq->inactive_works)) {
5183 bool comma = false;
5184
5185 pr_info(" inactive:");
5186 list_for_each_entry(work, &pwq->inactive_works, entry) {
5187 pr_cont_work(comma, work, &pcws);
5188 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
5189 }
5190 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
5191 pr_cont("\n");
5192 }
5193}
5194
5195/**
5196 * show_one_workqueue - dump state of specified workqueue
5197 * @wq: workqueue whose state will be printed
5198 */
5199void show_one_workqueue(struct workqueue_struct *wq)
5200{
5201 struct pool_workqueue *pwq;
5202 bool idle = true;
5203 unsigned long flags;
5204
5205 for_each_pwq(pwq, wq) {
5206 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5207 idle = false;
5208 break;
5209 }
5210 }
5211 if (idle) /* Nothing to print for idle workqueue */
5212 return;
5213
5214 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
5215
5216 for_each_pwq(pwq, wq) {
5217 raw_spin_lock_irqsave(&pwq->pool->lock, flags);
5218 if (pwq->nr_active || !list_empty(&pwq->inactive_works)) {
5219 /*
5220 * Defer printing to avoid deadlocks in console
5221 * drivers that queue work while holding locks
5222 * also taken in their write paths.
5223 */
5224 printk_deferred_enter();
5225 show_pwq(pwq);
5226 printk_deferred_exit();
5227 }
5228 raw_spin_unlock_irqrestore(&pwq->pool->lock, flags);
5229 /*
5230 * We could be printing a lot from atomic context, e.g.
5231 * sysrq-t -> show_all_workqueues(). Avoid triggering
5232 * hard lockup.
5233 */
5234 touch_nmi_watchdog();
5235 }
5236
5237}
5238
5239/**
5240 * show_one_worker_pool - dump state of specified worker pool
5241 * @pool: worker pool whose state will be printed
5242 */
5243static void show_one_worker_pool(struct worker_pool *pool)
5244{
5245 struct worker *worker;
5246 bool first = true;
5247 unsigned long flags;
5248 unsigned long hung = 0;
5249
5250 raw_spin_lock_irqsave(&pool->lock, flags);
5251 if (pool->nr_workers == pool->nr_idle)
5252 goto next_pool;
5253
5254 /* How long the first pending work is waiting for a worker. */
5255 if (!list_empty(&pool->worklist))
5256 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
5257
5258 /*
5259 * Defer printing to avoid deadlocks in console drivers that
5260 * queue work while holding locks also taken in their write
5261 * paths.
5262 */
5263 printk_deferred_enter();
5264 pr_info("pool %d:", pool->id);
5265 pr_cont_pool_info(pool);
5266 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
5267 if (pool->manager)
5268 pr_cont(" manager: %d",
5269 task_pid_nr(pool->manager->task));
5270 list_for_each_entry(worker, &pool->idle_list, entry) {
5271 pr_cont(" %s%d", first ? "idle: " : "",
5272 task_pid_nr(worker->task));
5273 first = false;
5274 }
5275 pr_cont("\n");
5276 printk_deferred_exit();
5277next_pool:
5278 raw_spin_unlock_irqrestore(&pool->lock, flags);
5279 /*
5280 * We could be printing a lot from atomic context, e.g.
5281 * sysrq-t -> show_all_workqueues(). Avoid triggering
5282 * hard lockup.
5283 */
5284 touch_nmi_watchdog();
5285
5286}
5287
5288/**
5289 * show_all_workqueues - dump workqueue state
5290 *
5291 * Called from a sysrq handler and prints out all busy workqueues and pools.
5292 */
5293void show_all_workqueues(void)
5294{
5295 struct workqueue_struct *wq;
5296 struct worker_pool *pool;
5297 int pi;
5298
5299 rcu_read_lock();
5300
5301 pr_info("Showing busy workqueues and worker pools:\n");
5302
5303 list_for_each_entry_rcu(wq, &workqueues, list)
5304 show_one_workqueue(wq);
5305
5306 for_each_pool(pool, pi)
5307 show_one_worker_pool(pool);
5308
5309 rcu_read_unlock();
5310}
5311
5312/**
5313 * show_freezable_workqueues - dump freezable workqueue state
5314 *
5315 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
5316 * still busy.
5317 */
5318void show_freezable_workqueues(void)
5319{
5320 struct workqueue_struct *wq;
5321
5322 rcu_read_lock();
5323
5324 pr_info("Showing freezable workqueues that are still busy:\n");
5325
5326 list_for_each_entry_rcu(wq, &workqueues, list) {
5327 if (!(wq->flags & WQ_FREEZABLE))
5328 continue;
5329 show_one_workqueue(wq);
5330 }
5331
5332 rcu_read_unlock();
5333}
5334
5335/* used to show worker information through /proc/PID/{comm,stat,status} */
5336void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
5337{
5338 int off;
5339
5340 /* always show the actual comm */
5341 off = strscpy(buf, task->comm, size);
5342 if (off < 0)
5343 return;
5344
5345 /* stabilize PF_WQ_WORKER and worker pool association */
5346 mutex_lock(&wq_pool_attach_mutex);
5347
5348 if (task->flags & PF_WQ_WORKER) {
5349 struct worker *worker = kthread_data(task);
5350 struct worker_pool *pool = worker->pool;
5351
5352 if (pool) {
5353 raw_spin_lock_irq(&pool->lock);
5354 /*
5355 * ->desc tracks information (wq name or
5356 * set_worker_desc()) for the latest execution. If
5357 * current, prepend '+', otherwise '-'.
5358 */
5359 if (worker->desc[0] != '\0') {
5360 if (worker->current_work)
5361 scnprintf(buf + off, size - off, "+%s",
5362 worker->desc);
5363 else
5364 scnprintf(buf + off, size - off, "-%s",
5365 worker->desc);
5366 }
5367 raw_spin_unlock_irq(&pool->lock);
5368 }
5369 }
5370
5371 mutex_unlock(&wq_pool_attach_mutex);
5372}
5373
5374#ifdef CONFIG_SMP
5375
5376/*
5377 * CPU hotplug.
5378 *
5379 * There are two challenges in supporting CPU hotplug. Firstly, there
5380 * are a lot of assumptions on strong associations among work, pwq and
5381 * pool which make migrating pending and scheduled works very
5382 * difficult to implement without impacting hot paths. Secondly,
5383 * worker pools serve mix of short, long and very long running works making
5384 * blocked draining impractical.
5385 *
5386 * This is solved by allowing the pools to be disassociated from the CPU
5387 * running as an unbound one and allowing it to be reattached later if the
5388 * cpu comes back online.
5389 */
5390
5391static void unbind_workers(int cpu)
5392{
5393 struct worker_pool *pool;
5394 struct worker *worker;
5395
5396 for_each_cpu_worker_pool(pool, cpu) {
5397 mutex_lock(&wq_pool_attach_mutex);
5398 raw_spin_lock_irq(&pool->lock);
5399
5400 /*
5401 * We've blocked all attach/detach operations. Make all workers
5402 * unbound and set DISASSOCIATED. Before this, all workers
5403 * must be on the cpu. After this, they may become diasporas.
5404 * And the preemption disabled section in their sched callbacks
5405 * are guaranteed to see WORKER_UNBOUND since the code here
5406 * is on the same cpu.
5407 */
5408 for_each_pool_worker(worker, pool)
5409 worker->flags |= WORKER_UNBOUND;
5410
5411 pool->flags |= POOL_DISASSOCIATED;
5412
5413 /*
5414 * The handling of nr_running in sched callbacks are disabled
5415 * now. Zap nr_running. After this, nr_running stays zero and
5416 * need_more_worker() and keep_working() are always true as
5417 * long as the worklist is not empty. This pool now behaves as
5418 * an unbound (in terms of concurrency management) pool which
5419 * are served by workers tied to the pool.
5420 */
5421 pool->nr_running = 0;
5422
5423 /*
5424 * With concurrency management just turned off, a busy
5425 * worker blocking could lead to lengthy stalls. Kick off
5426 * unbound chain execution of currently pending work items.
5427 */
5428 kick_pool(pool);
5429
5430 raw_spin_unlock_irq(&pool->lock);
5431
5432 for_each_pool_worker(worker, pool)
5433 unbind_worker(worker);
5434
5435 mutex_unlock(&wq_pool_attach_mutex);
5436 }
5437}
5438
5439/**
5440 * rebind_workers - rebind all workers of a pool to the associated CPU
5441 * @pool: pool of interest
5442 *
5443 * @pool->cpu is coming online. Rebind all workers to the CPU.
5444 */
5445static void rebind_workers(struct worker_pool *pool)
5446{
5447 struct worker *worker;
5448
5449 lockdep_assert_held(&wq_pool_attach_mutex);
5450
5451 /*
5452 * Restore CPU affinity of all workers. As all idle workers should
5453 * be on the run-queue of the associated CPU before any local
5454 * wake-ups for concurrency management happen, restore CPU affinity
5455 * of all workers first and then clear UNBOUND. As we're called
5456 * from CPU_ONLINE, the following shouldn't fail.
5457 */
5458 for_each_pool_worker(worker, pool) {
5459 kthread_set_per_cpu(worker->task, pool->cpu);
5460 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
5461 pool_allowed_cpus(pool)) < 0);
5462 }
5463
5464 raw_spin_lock_irq(&pool->lock);
5465
5466 pool->flags &= ~POOL_DISASSOCIATED;
5467
5468 for_each_pool_worker(worker, pool) {
5469 unsigned int worker_flags = worker->flags;
5470
5471 /*
5472 * We want to clear UNBOUND but can't directly call
5473 * worker_clr_flags() or adjust nr_running. Atomically
5474 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
5475 * @worker will clear REBOUND using worker_clr_flags() when
5476 * it initiates the next execution cycle thus restoring
5477 * concurrency management. Note that when or whether
5478 * @worker clears REBOUND doesn't affect correctness.
5479 *
5480 * WRITE_ONCE() is necessary because @worker->flags may be
5481 * tested without holding any lock in
5482 * wq_worker_running(). Without it, NOT_RUNNING test may
5483 * fail incorrectly leading to premature concurrency
5484 * management operations.
5485 */
5486 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
5487 worker_flags |= WORKER_REBOUND;
5488 worker_flags &= ~WORKER_UNBOUND;
5489 WRITE_ONCE(worker->flags, worker_flags);
5490 }
5491
5492 raw_spin_unlock_irq(&pool->lock);
5493}
5494
5495/**
5496 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
5497 * @pool: unbound pool of interest
5498 * @cpu: the CPU which is coming up
5499 *
5500 * An unbound pool may end up with a cpumask which doesn't have any online
5501 * CPUs. When a worker of such pool get scheduled, the scheduler resets
5502 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
5503 * online CPU before, cpus_allowed of all its workers should be restored.
5504 */
5505static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
5506{
5507 static cpumask_t cpumask;
5508 struct worker *worker;
5509
5510 lockdep_assert_held(&wq_pool_attach_mutex);
5511
5512 /* is @cpu allowed for @pool? */
5513 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
5514 return;
5515
5516 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
5517
5518 /* as we're called from CPU_ONLINE, the following shouldn't fail */
5519 for_each_pool_worker(worker, pool)
5520 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
5521}
5522
5523int workqueue_prepare_cpu(unsigned int cpu)
5524{
5525 struct worker_pool *pool;
5526
5527 for_each_cpu_worker_pool(pool, cpu) {
5528 if (pool->nr_workers)
5529 continue;
5530 if (!create_worker(pool))
5531 return -ENOMEM;
5532 }
5533 return 0;
5534}
5535
5536int workqueue_online_cpu(unsigned int cpu)
5537{
5538 struct worker_pool *pool;
5539 struct workqueue_struct *wq;
5540 int pi;
5541
5542 mutex_lock(&wq_pool_mutex);
5543
5544 for_each_pool(pool, pi) {
5545 mutex_lock(&wq_pool_attach_mutex);
5546
5547 if (pool->cpu == cpu)
5548 rebind_workers(pool);
5549 else if (pool->cpu < 0)
5550 restore_unbound_workers_cpumask(pool, cpu);
5551
5552 mutex_unlock(&wq_pool_attach_mutex);
5553 }
5554
5555 /* update pod affinity of unbound workqueues */
5556 list_for_each_entry(wq, &workqueues, list) {
5557 struct workqueue_attrs *attrs = wq->unbound_attrs;
5558
5559 if (attrs) {
5560 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5561 int tcpu;
5562
5563 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5564 wq_update_pod(wq, tcpu, cpu, true);
5565 }
5566 }
5567
5568 mutex_unlock(&wq_pool_mutex);
5569 return 0;
5570}
5571
5572int workqueue_offline_cpu(unsigned int cpu)
5573{
5574 struct workqueue_struct *wq;
5575
5576 /* unbinding per-cpu workers should happen on the local CPU */
5577 if (WARN_ON(cpu != smp_processor_id()))
5578 return -1;
5579
5580 unbind_workers(cpu);
5581
5582 /* update pod affinity of unbound workqueues */
5583 mutex_lock(&wq_pool_mutex);
5584 list_for_each_entry(wq, &workqueues, list) {
5585 struct workqueue_attrs *attrs = wq->unbound_attrs;
5586
5587 if (attrs) {
5588 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5589 int tcpu;
5590
5591 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
5592 wq_update_pod(wq, tcpu, cpu, false);
5593 }
5594 }
5595 mutex_unlock(&wq_pool_mutex);
5596
5597 return 0;
5598}
5599
5600struct work_for_cpu {
5601 struct work_struct work;
5602 long (*fn)(void *);
5603 void *arg;
5604 long ret;
5605};
5606
5607static void work_for_cpu_fn(struct work_struct *work)
5608{
5609 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5610
5611 wfc->ret = wfc->fn(wfc->arg);
5612}
5613
5614/**
5615 * work_on_cpu_key - run a function in thread context on a particular cpu
5616 * @cpu: the cpu to run on
5617 * @fn: the function to run
5618 * @arg: the function arg
5619 * @key: The lock class key for lock debugging purposes
5620 *
5621 * It is up to the caller to ensure that the cpu doesn't go offline.
5622 * The caller must not hold any locks which would prevent @fn from completing.
5623 *
5624 * Return: The value @fn returns.
5625 */
5626long work_on_cpu_key(int cpu, long (*fn)(void *),
5627 void *arg, struct lock_class_key *key)
5628{
5629 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5630
5631 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
5632 schedule_work_on(cpu, &wfc.work);
5633 flush_work(&wfc.work);
5634 destroy_work_on_stack(&wfc.work);
5635 return wfc.ret;
5636}
5637EXPORT_SYMBOL_GPL(work_on_cpu_key);
5638
5639/**
5640 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
5641 * @cpu: the cpu to run on
5642 * @fn: the function to run
5643 * @arg: the function argument
5644 * @key: The lock class key for lock debugging purposes
5645 *
5646 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5647 * any locks which would prevent @fn from completing.
5648 *
5649 * Return: The value @fn returns.
5650 */
5651long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
5652 void *arg, struct lock_class_key *key)
5653{
5654 long ret = -ENODEV;
5655
5656 cpus_read_lock();
5657 if (cpu_online(cpu))
5658 ret = work_on_cpu_key(cpu, fn, arg, key);
5659 cpus_read_unlock();
5660 return ret;
5661}
5662EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
5663#endif /* CONFIG_SMP */
5664
5665#ifdef CONFIG_FREEZER
5666
5667/**
5668 * freeze_workqueues_begin - begin freezing workqueues
5669 *
5670 * Start freezing workqueues. After this function returns, all freezable
5671 * workqueues will queue new works to their inactive_works list instead of
5672 * pool->worklist.
5673 *
5674 * CONTEXT:
5675 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5676 */
5677void freeze_workqueues_begin(void)
5678{
5679 struct workqueue_struct *wq;
5680 struct pool_workqueue *pwq;
5681
5682 mutex_lock(&wq_pool_mutex);
5683
5684 WARN_ON_ONCE(workqueue_freezing);
5685 workqueue_freezing = true;
5686
5687 list_for_each_entry(wq, &workqueues, list) {
5688 mutex_lock(&wq->mutex);
5689 for_each_pwq(pwq, wq)
5690 pwq_adjust_max_active(pwq);
5691 mutex_unlock(&wq->mutex);
5692 }
5693
5694 mutex_unlock(&wq_pool_mutex);
5695}
5696
5697/**
5698 * freeze_workqueues_busy - are freezable workqueues still busy?
5699 *
5700 * Check whether freezing is complete. This function must be called
5701 * between freeze_workqueues_begin() and thaw_workqueues().
5702 *
5703 * CONTEXT:
5704 * Grabs and releases wq_pool_mutex.
5705 *
5706 * Return:
5707 * %true if some freezable workqueues are still busy. %false if freezing
5708 * is complete.
5709 */
5710bool freeze_workqueues_busy(void)
5711{
5712 bool busy = false;
5713 struct workqueue_struct *wq;
5714 struct pool_workqueue *pwq;
5715
5716 mutex_lock(&wq_pool_mutex);
5717
5718 WARN_ON_ONCE(!workqueue_freezing);
5719
5720 list_for_each_entry(wq, &workqueues, list) {
5721 if (!(wq->flags & WQ_FREEZABLE))
5722 continue;
5723 /*
5724 * nr_active is monotonically decreasing. It's safe
5725 * to peek without lock.
5726 */
5727 rcu_read_lock();
5728 for_each_pwq(pwq, wq) {
5729 WARN_ON_ONCE(pwq->nr_active < 0);
5730 if (pwq->nr_active) {
5731 busy = true;
5732 rcu_read_unlock();
5733 goto out_unlock;
5734 }
5735 }
5736 rcu_read_unlock();
5737 }
5738out_unlock:
5739 mutex_unlock(&wq_pool_mutex);
5740 return busy;
5741}
5742
5743/**
5744 * thaw_workqueues - thaw workqueues
5745 *
5746 * Thaw workqueues. Normal queueing is restored and all collected
5747 * frozen works are transferred to their respective pool worklists.
5748 *
5749 * CONTEXT:
5750 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5751 */
5752void thaw_workqueues(void)
5753{
5754 struct workqueue_struct *wq;
5755 struct pool_workqueue *pwq;
5756
5757 mutex_lock(&wq_pool_mutex);
5758
5759 if (!workqueue_freezing)
5760 goto out_unlock;
5761
5762 workqueue_freezing = false;
5763
5764 /* restore max_active and repopulate worklist */
5765 list_for_each_entry(wq, &workqueues, list) {
5766 mutex_lock(&wq->mutex);
5767 for_each_pwq(pwq, wq)
5768 pwq_adjust_max_active(pwq);
5769 mutex_unlock(&wq->mutex);
5770 }
5771
5772out_unlock:
5773 mutex_unlock(&wq_pool_mutex);
5774}
5775#endif /* CONFIG_FREEZER */
5776
5777static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
5778{
5779 LIST_HEAD(ctxs);
5780 int ret = 0;
5781 struct workqueue_struct *wq;
5782 struct apply_wqattrs_ctx *ctx, *n;
5783
5784 lockdep_assert_held(&wq_pool_mutex);
5785
5786 list_for_each_entry(wq, &workqueues, list) {
5787 if (!(wq->flags & WQ_UNBOUND))
5788 continue;
5789 /* creating multiple pwqs breaks ordering guarantee */
5790 if (wq->flags & __WQ_ORDERED)
5791 continue;
5792
5793 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
5794 if (IS_ERR(ctx)) {
5795 ret = PTR_ERR(ctx);
5796 break;
5797 }
5798
5799 list_add_tail(&ctx->list, &ctxs);
5800 }
5801
5802 list_for_each_entry_safe(ctx, n, &ctxs, list) {
5803 if (!ret)
5804 apply_wqattrs_commit(ctx);
5805 apply_wqattrs_cleanup(ctx);
5806 }
5807
5808 if (!ret) {
5809 mutex_lock(&wq_pool_attach_mutex);
5810 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
5811 mutex_unlock(&wq_pool_attach_mutex);
5812 }
5813 return ret;
5814}
5815
5816/**
5817 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
5818 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
5819 *
5820 * This function can be called from cpuset code to provide a set of isolated
5821 * CPUs that should be excluded from wq_unbound_cpumask. The caller must hold
5822 * either cpus_read_lock or cpus_write_lock.
5823 */
5824int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
5825{
5826 cpumask_var_t cpumask;
5827 int ret = 0;
5828
5829 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5830 return -ENOMEM;
5831
5832 lockdep_assert_cpus_held();
5833 mutex_lock(&wq_pool_mutex);
5834
5835 /* Save the current isolated cpumask & export it via sysfs */
5836 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
5837
5838 /*
5839 * If the operation fails, it will fall back to
5840 * wq_requested_unbound_cpumask which is initially set to
5841 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
5842 * by any subsequent write to workqueue/cpumask sysfs file.
5843 */
5844 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
5845 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
5846 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
5847 ret = workqueue_apply_unbound_cpumask(cpumask);
5848
5849 mutex_unlock(&wq_pool_mutex);
5850 free_cpumask_var(cpumask);
5851 return ret;
5852}
5853
5854static int parse_affn_scope(const char *val)
5855{
5856 int i;
5857
5858 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
5859 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
5860 return i;
5861 }
5862 return -EINVAL;
5863}
5864
5865static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
5866{
5867 struct workqueue_struct *wq;
5868 int affn, cpu;
5869
5870 affn = parse_affn_scope(val);
5871 if (affn < 0)
5872 return affn;
5873 if (affn == WQ_AFFN_DFL)
5874 return -EINVAL;
5875
5876 cpus_read_lock();
5877 mutex_lock(&wq_pool_mutex);
5878
5879 wq_affn_dfl = affn;
5880
5881 list_for_each_entry(wq, &workqueues, list) {
5882 for_each_online_cpu(cpu) {
5883 wq_update_pod(wq, cpu, cpu, true);
5884 }
5885 }
5886
5887 mutex_unlock(&wq_pool_mutex);
5888 cpus_read_unlock();
5889
5890 return 0;
5891}
5892
5893static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
5894{
5895 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
5896}
5897
5898static const struct kernel_param_ops wq_affn_dfl_ops = {
5899 .set = wq_affn_dfl_set,
5900 .get = wq_affn_dfl_get,
5901};
5902
5903module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
5904
5905#ifdef CONFIG_SYSFS
5906/*
5907 * Workqueues with WQ_SYSFS flag set is visible to userland via
5908 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
5909 * following attributes.
5910 *
5911 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
5912 * max_active RW int : maximum number of in-flight work items
5913 *
5914 * Unbound workqueues have the following extra attributes.
5915 *
5916 * nice RW int : nice value of the workers
5917 * cpumask RW mask : bitmask of allowed CPUs for the workers
5918 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
5919 * affinity_strict RW bool : worker CPU affinity is strict
5920 */
5921struct wq_device {
5922 struct workqueue_struct *wq;
5923 struct device dev;
5924};
5925
5926static struct workqueue_struct *dev_to_wq(struct device *dev)
5927{
5928 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5929
5930 return wq_dev->wq;
5931}
5932
5933static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5934 char *buf)
5935{
5936 struct workqueue_struct *wq = dev_to_wq(dev);
5937
5938 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5939}
5940static DEVICE_ATTR_RO(per_cpu);
5941
5942static ssize_t max_active_show(struct device *dev,
5943 struct device_attribute *attr, char *buf)
5944{
5945 struct workqueue_struct *wq = dev_to_wq(dev);
5946
5947 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5948}
5949
5950static ssize_t max_active_store(struct device *dev,
5951 struct device_attribute *attr, const char *buf,
5952 size_t count)
5953{
5954 struct workqueue_struct *wq = dev_to_wq(dev);
5955 int val;
5956
5957 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5958 return -EINVAL;
5959
5960 workqueue_set_max_active(wq, val);
5961 return count;
5962}
5963static DEVICE_ATTR_RW(max_active);
5964
5965static struct attribute *wq_sysfs_attrs[] = {
5966 &dev_attr_per_cpu.attr,
5967 &dev_attr_max_active.attr,
5968 NULL,
5969};
5970ATTRIBUTE_GROUPS(wq_sysfs);
5971
5972static void apply_wqattrs_lock(void)
5973{
5974 /* CPUs should stay stable across pwq creations and installations */
5975 cpus_read_lock();
5976 mutex_lock(&wq_pool_mutex);
5977}
5978
5979static void apply_wqattrs_unlock(void)
5980{
5981 mutex_unlock(&wq_pool_mutex);
5982 cpus_read_unlock();
5983}
5984
5985static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5986 char *buf)
5987{
5988 struct workqueue_struct *wq = dev_to_wq(dev);
5989 int written;
5990
5991 mutex_lock(&wq->mutex);
5992 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5993 mutex_unlock(&wq->mutex);
5994
5995 return written;
5996}
5997
5998/* prepare workqueue_attrs for sysfs store operations */
5999static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
6000{
6001 struct workqueue_attrs *attrs;
6002
6003 lockdep_assert_held(&wq_pool_mutex);
6004
6005 attrs = alloc_workqueue_attrs();
6006 if (!attrs)
6007 return NULL;
6008
6009 copy_workqueue_attrs(attrs, wq->unbound_attrs);
6010 return attrs;
6011}
6012
6013static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
6014 const char *buf, size_t count)
6015{
6016 struct workqueue_struct *wq = dev_to_wq(dev);
6017 struct workqueue_attrs *attrs;
6018 int ret = -ENOMEM;
6019
6020 apply_wqattrs_lock();
6021
6022 attrs = wq_sysfs_prep_attrs(wq);
6023 if (!attrs)
6024 goto out_unlock;
6025
6026 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
6027 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
6028 ret = apply_workqueue_attrs_locked(wq, attrs);
6029 else
6030 ret = -EINVAL;
6031
6032out_unlock:
6033 apply_wqattrs_unlock();
6034 free_workqueue_attrs(attrs);
6035 return ret ?: count;
6036}
6037
6038static ssize_t wq_cpumask_show(struct device *dev,
6039 struct device_attribute *attr, char *buf)
6040{
6041 struct workqueue_struct *wq = dev_to_wq(dev);
6042 int written;
6043
6044 mutex_lock(&wq->mutex);
6045 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
6046 cpumask_pr_args(wq->unbound_attrs->cpumask));
6047 mutex_unlock(&wq->mutex);
6048 return written;
6049}
6050
6051static ssize_t wq_cpumask_store(struct device *dev,
6052 struct device_attribute *attr,
6053 const char *buf, size_t count)
6054{
6055 struct workqueue_struct *wq = dev_to_wq(dev);
6056 struct workqueue_attrs *attrs;
6057 int ret = -ENOMEM;
6058
6059 apply_wqattrs_lock();
6060
6061 attrs = wq_sysfs_prep_attrs(wq);
6062 if (!attrs)
6063 goto out_unlock;
6064
6065 ret = cpumask_parse(buf, attrs->cpumask);
6066 if (!ret)
6067 ret = apply_workqueue_attrs_locked(wq, attrs);
6068
6069out_unlock:
6070 apply_wqattrs_unlock();
6071 free_workqueue_attrs(attrs);
6072 return ret ?: count;
6073}
6074
6075static ssize_t wq_affn_scope_show(struct device *dev,
6076 struct device_attribute *attr, char *buf)
6077{
6078 struct workqueue_struct *wq = dev_to_wq(dev);
6079 int written;
6080
6081 mutex_lock(&wq->mutex);
6082 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
6083 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
6084 wq_affn_names[WQ_AFFN_DFL],
6085 wq_affn_names[wq_affn_dfl]);
6086 else
6087 written = scnprintf(buf, PAGE_SIZE, "%s\n",
6088 wq_affn_names[wq->unbound_attrs->affn_scope]);
6089 mutex_unlock(&wq->mutex);
6090
6091 return written;
6092}
6093
6094static ssize_t wq_affn_scope_store(struct device *dev,
6095 struct device_attribute *attr,
6096 const char *buf, size_t count)
6097{
6098 struct workqueue_struct *wq = dev_to_wq(dev);
6099 struct workqueue_attrs *attrs;
6100 int affn, ret = -ENOMEM;
6101
6102 affn = parse_affn_scope(buf);
6103 if (affn < 0)
6104 return affn;
6105
6106 apply_wqattrs_lock();
6107 attrs = wq_sysfs_prep_attrs(wq);
6108 if (attrs) {
6109 attrs->affn_scope = affn;
6110 ret = apply_workqueue_attrs_locked(wq, attrs);
6111 }
6112 apply_wqattrs_unlock();
6113 free_workqueue_attrs(attrs);
6114 return ret ?: count;
6115}
6116
6117static ssize_t wq_affinity_strict_show(struct device *dev,
6118 struct device_attribute *attr, char *buf)
6119{
6120 struct workqueue_struct *wq = dev_to_wq(dev);
6121
6122 return scnprintf(buf, PAGE_SIZE, "%d\n",
6123 wq->unbound_attrs->affn_strict);
6124}
6125
6126static ssize_t wq_affinity_strict_store(struct device *dev,
6127 struct device_attribute *attr,
6128 const char *buf, size_t count)
6129{
6130 struct workqueue_struct *wq = dev_to_wq(dev);
6131 struct workqueue_attrs *attrs;
6132 int v, ret = -ENOMEM;
6133
6134 if (sscanf(buf, "%d", &v) != 1)
6135 return -EINVAL;
6136
6137 apply_wqattrs_lock();
6138 attrs = wq_sysfs_prep_attrs(wq);
6139 if (attrs) {
6140 attrs->affn_strict = (bool)v;
6141 ret = apply_workqueue_attrs_locked(wq, attrs);
6142 }
6143 apply_wqattrs_unlock();
6144 free_workqueue_attrs(attrs);
6145 return ret ?: count;
6146}
6147
6148static struct device_attribute wq_sysfs_unbound_attrs[] = {
6149 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
6150 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
6151 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
6152 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
6153 __ATTR_NULL,
6154};
6155
6156static struct bus_type wq_subsys = {
6157 .name = "workqueue",
6158 .dev_groups = wq_sysfs_groups,
6159};
6160
6161/**
6162 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
6163 * @cpumask: the cpumask to set
6164 *
6165 * The low-level workqueues cpumask is a global cpumask that limits
6166 * the affinity of all unbound workqueues. This function check the @cpumask
6167 * and apply it to all unbound workqueues and updates all pwqs of them.
6168 *
6169 * Return: 0 - Success
6170 * -EINVAL - Invalid @cpumask
6171 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
6172 */
6173static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
6174{
6175 int ret = -EINVAL;
6176
6177 /*
6178 * Not excluding isolated cpus on purpose.
6179 * If the user wishes to include them, we allow that.
6180 */
6181 cpumask_and(cpumask, cpumask, cpu_possible_mask);
6182 if (!cpumask_empty(cpumask)) {
6183 apply_wqattrs_lock();
6184 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
6185 if (cpumask_equal(cpumask, wq_unbound_cpumask)) {
6186 ret = 0;
6187 goto out_unlock;
6188 }
6189
6190 ret = workqueue_apply_unbound_cpumask(cpumask);
6191
6192out_unlock:
6193 apply_wqattrs_unlock();
6194 }
6195
6196 return ret;
6197}
6198
6199static ssize_t __wq_cpumask_show(struct device *dev,
6200 struct device_attribute *attr, char *buf, cpumask_var_t mask)
6201{
6202 int written;
6203
6204 mutex_lock(&wq_pool_mutex);
6205 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
6206 mutex_unlock(&wq_pool_mutex);
6207
6208 return written;
6209}
6210
6211static ssize_t wq_unbound_cpumask_show(struct device *dev,
6212 struct device_attribute *attr, char *buf)
6213{
6214 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
6215}
6216
6217static ssize_t wq_requested_cpumask_show(struct device *dev,
6218 struct device_attribute *attr, char *buf)
6219{
6220 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
6221}
6222
6223static ssize_t wq_isolated_cpumask_show(struct device *dev,
6224 struct device_attribute *attr, char *buf)
6225{
6226 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
6227}
6228
6229static ssize_t wq_unbound_cpumask_store(struct device *dev,
6230 struct device_attribute *attr, const char *buf, size_t count)
6231{
6232 cpumask_var_t cpumask;
6233 int ret;
6234
6235 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6236 return -ENOMEM;
6237
6238 ret = cpumask_parse(buf, cpumask);
6239 if (!ret)
6240 ret = workqueue_set_unbound_cpumask(cpumask);
6241
6242 free_cpumask_var(cpumask);
6243 return ret ? ret : count;
6244}
6245
6246static struct device_attribute wq_sysfs_cpumask_attrs[] = {
6247 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
6248 wq_unbound_cpumask_store),
6249 __ATTR(cpumask_requested, 0444, wq_requested_cpumask_show, NULL),
6250 __ATTR(cpumask_isolated, 0444, wq_isolated_cpumask_show, NULL),
6251 __ATTR_NULL,
6252};
6253
6254static int __init wq_sysfs_init(void)
6255{
6256 struct device *dev_root;
6257 int err;
6258
6259 err = subsys_virtual_register(&wq_subsys, NULL);
6260 if (err)
6261 return err;
6262
6263 dev_root = bus_get_dev_root(&wq_subsys);
6264 if (dev_root) {
6265 struct device_attribute *attr;
6266
6267 for (attr = wq_sysfs_cpumask_attrs; attr->attr.name; attr++) {
6268 err = device_create_file(dev_root, attr);
6269 if (err)
6270 break;
6271 }
6272 put_device(dev_root);
6273 }
6274 return err;
6275}
6276core_initcall(wq_sysfs_init);
6277
6278static void wq_device_release(struct device *dev)
6279{
6280 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
6281
6282 kfree(wq_dev);
6283}
6284
6285/**
6286 * workqueue_sysfs_register - make a workqueue visible in sysfs
6287 * @wq: the workqueue to register
6288 *
6289 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
6290 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
6291 * which is the preferred method.
6292 *
6293 * Workqueue user should use this function directly iff it wants to apply
6294 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
6295 * apply_workqueue_attrs() may race against userland updating the
6296 * attributes.
6297 *
6298 * Return: 0 on success, -errno on failure.
6299 */
6300int workqueue_sysfs_register(struct workqueue_struct *wq)
6301{
6302 struct wq_device *wq_dev;
6303 int ret;
6304
6305 /*
6306 * Adjusting max_active or creating new pwqs by applying
6307 * attributes breaks ordering guarantee. Disallow exposing ordered
6308 * workqueues.
6309 */
6310 if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
6311 return -EINVAL;
6312
6313 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
6314 if (!wq_dev)
6315 return -ENOMEM;
6316
6317 wq_dev->wq = wq;
6318 wq_dev->dev.bus = &wq_subsys;
6319 wq_dev->dev.release = wq_device_release;
6320 dev_set_name(&wq_dev->dev, "%s", wq->name);
6321
6322 /*
6323 * unbound_attrs are created separately. Suppress uevent until
6324 * everything is ready.
6325 */
6326 dev_set_uevent_suppress(&wq_dev->dev, true);
6327
6328 ret = device_register(&wq_dev->dev);
6329 if (ret) {
6330 put_device(&wq_dev->dev);
6331 wq->wq_dev = NULL;
6332 return ret;
6333 }
6334
6335 if (wq->flags & WQ_UNBOUND) {
6336 struct device_attribute *attr;
6337
6338 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
6339 ret = device_create_file(&wq_dev->dev, attr);
6340 if (ret) {
6341 device_unregister(&wq_dev->dev);
6342 wq->wq_dev = NULL;
6343 return ret;
6344 }
6345 }
6346 }
6347
6348 dev_set_uevent_suppress(&wq_dev->dev, false);
6349 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
6350 return 0;
6351}
6352
6353/**
6354 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
6355 * @wq: the workqueue to unregister
6356 *
6357 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
6358 */
6359static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
6360{
6361 struct wq_device *wq_dev = wq->wq_dev;
6362
6363 if (!wq->wq_dev)
6364 return;
6365
6366 wq->wq_dev = NULL;
6367 device_unregister(&wq_dev->dev);
6368}
6369#else /* CONFIG_SYSFS */
6370static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
6371#endif /* CONFIG_SYSFS */
6372
6373/*
6374 * Workqueue watchdog.
6375 *
6376 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
6377 * flush dependency, a concurrency managed work item which stays RUNNING
6378 * indefinitely. Workqueue stalls can be very difficult to debug as the
6379 * usual warning mechanisms don't trigger and internal workqueue state is
6380 * largely opaque.
6381 *
6382 * Workqueue watchdog monitors all worker pools periodically and dumps
6383 * state if some pools failed to make forward progress for a while where
6384 * forward progress is defined as the first item on ->worklist changing.
6385 *
6386 * This mechanism is controlled through the kernel parameter
6387 * "workqueue.watchdog_thresh" which can be updated at runtime through the
6388 * corresponding sysfs parameter file.
6389 */
6390#ifdef CONFIG_WQ_WATCHDOG
6391
6392static unsigned long wq_watchdog_thresh = 30;
6393static struct timer_list wq_watchdog_timer;
6394
6395static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
6396static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
6397
6398/*
6399 * Show workers that might prevent the processing of pending work items.
6400 * The only candidates are CPU-bound workers in the running state.
6401 * Pending work items should be handled by another idle worker
6402 * in all other situations.
6403 */
6404static void show_cpu_pool_hog(struct worker_pool *pool)
6405{
6406 struct worker *worker;
6407 unsigned long flags;
6408 int bkt;
6409
6410 raw_spin_lock_irqsave(&pool->lock, flags);
6411
6412 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6413 if (task_is_running(worker->task)) {
6414 /*
6415 * Defer printing to avoid deadlocks in console
6416 * drivers that queue work while holding locks
6417 * also taken in their write paths.
6418 */
6419 printk_deferred_enter();
6420
6421 pr_info("pool %d:\n", pool->id);
6422 sched_show_task(worker->task);
6423
6424 printk_deferred_exit();
6425 }
6426 }
6427
6428 raw_spin_unlock_irqrestore(&pool->lock, flags);
6429}
6430
6431static void show_cpu_pools_hogs(void)
6432{
6433 struct worker_pool *pool;
6434 int pi;
6435
6436 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
6437
6438 rcu_read_lock();
6439
6440 for_each_pool(pool, pi) {
6441 if (pool->cpu_stall)
6442 show_cpu_pool_hog(pool);
6443
6444 }
6445
6446 rcu_read_unlock();
6447}
6448
6449static void wq_watchdog_reset_touched(void)
6450{
6451 int cpu;
6452
6453 wq_watchdog_touched = jiffies;
6454 for_each_possible_cpu(cpu)
6455 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6456}
6457
6458static void wq_watchdog_timer_fn(struct timer_list *unused)
6459{
6460 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
6461 bool lockup_detected = false;
6462 bool cpu_pool_stall = false;
6463 unsigned long now = jiffies;
6464 struct worker_pool *pool;
6465 int pi;
6466
6467 if (!thresh)
6468 return;
6469
6470 rcu_read_lock();
6471
6472 for_each_pool(pool, pi) {
6473 unsigned long pool_ts, touched, ts;
6474
6475 pool->cpu_stall = false;
6476 if (list_empty(&pool->worklist))
6477 continue;
6478
6479 /*
6480 * If a virtual machine is stopped by the host it can look to
6481 * the watchdog like a stall.
6482 */
6483 kvm_check_and_clear_guest_paused();
6484
6485 /* get the latest of pool and touched timestamps */
6486 if (pool->cpu >= 0)
6487 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
6488 else
6489 touched = READ_ONCE(wq_watchdog_touched);
6490 pool_ts = READ_ONCE(pool->watchdog_ts);
6491
6492 if (time_after(pool_ts, touched))
6493 ts = pool_ts;
6494 else
6495 ts = touched;
6496
6497 /* did we stall? */
6498 if (time_after(now, ts + thresh)) {
6499 lockup_detected = true;
6500 if (pool->cpu >= 0) {
6501 pool->cpu_stall = true;
6502 cpu_pool_stall = true;
6503 }
6504 pr_emerg("BUG: workqueue lockup - pool");
6505 pr_cont_pool_info(pool);
6506 pr_cont(" stuck for %us!\n",
6507 jiffies_to_msecs(now - pool_ts) / 1000);
6508 }
6509
6510
6511 }
6512
6513 rcu_read_unlock();
6514
6515 if (lockup_detected)
6516 show_all_workqueues();
6517
6518 if (cpu_pool_stall)
6519 show_cpu_pools_hogs();
6520
6521 wq_watchdog_reset_touched();
6522 mod_timer(&wq_watchdog_timer, jiffies + thresh);
6523}
6524
6525notrace void wq_watchdog_touch(int cpu)
6526{
6527 if (cpu >= 0)
6528 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
6529
6530 wq_watchdog_touched = jiffies;
6531}
6532
6533static void wq_watchdog_set_thresh(unsigned long thresh)
6534{
6535 wq_watchdog_thresh = 0;
6536 del_timer_sync(&wq_watchdog_timer);
6537
6538 if (thresh) {
6539 wq_watchdog_thresh = thresh;
6540 wq_watchdog_reset_touched();
6541 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
6542 }
6543}
6544
6545static int wq_watchdog_param_set_thresh(const char *val,
6546 const struct kernel_param *kp)
6547{
6548 unsigned long thresh;
6549 int ret;
6550
6551 ret = kstrtoul(val, 0, &thresh);
6552 if (ret)
6553 return ret;
6554
6555 if (system_wq)
6556 wq_watchdog_set_thresh(thresh);
6557 else
6558 wq_watchdog_thresh = thresh;
6559
6560 return 0;
6561}
6562
6563static const struct kernel_param_ops wq_watchdog_thresh_ops = {
6564 .set = wq_watchdog_param_set_thresh,
6565 .get = param_get_ulong,
6566};
6567
6568module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
6569 0644);
6570
6571static void wq_watchdog_init(void)
6572{
6573 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
6574 wq_watchdog_set_thresh(wq_watchdog_thresh);
6575}
6576
6577#else /* CONFIG_WQ_WATCHDOG */
6578
6579static inline void wq_watchdog_init(void) { }
6580
6581#endif /* CONFIG_WQ_WATCHDOG */
6582
6583static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
6584{
6585 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
6586 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
6587 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
6588 return;
6589 }
6590
6591 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
6592}
6593
6594/**
6595 * workqueue_init_early - early init for workqueue subsystem
6596 *
6597 * This is the first step of three-staged workqueue subsystem initialization and
6598 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
6599 * up. It sets up all the data structures and system workqueues and allows early
6600 * boot code to create workqueues and queue/cancel work items. Actual work item
6601 * execution starts only after kthreads can be created and scheduled right
6602 * before early initcalls.
6603 */
6604void __init workqueue_init_early(void)
6605{
6606 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
6607 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
6608 int i, cpu;
6609
6610 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
6611
6612 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
6613 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
6614 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
6615
6616 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
6617 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
6618 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
6619 if (!cpumask_empty(&wq_cmdline_cpumask))
6620 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
6621
6622 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
6623
6624 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
6625
6626 wq_update_pod_attrs_buf = alloc_workqueue_attrs();
6627 BUG_ON(!wq_update_pod_attrs_buf);
6628
6629 /* initialize WQ_AFFN_SYSTEM pods */
6630 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6631 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
6632 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6633 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
6634
6635 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
6636
6637 pt->nr_pods = 1;
6638 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
6639 pt->pod_node[0] = NUMA_NO_NODE;
6640 pt->cpu_pod[0] = 0;
6641
6642 /* initialize CPU pools */
6643 for_each_possible_cpu(cpu) {
6644 struct worker_pool *pool;
6645
6646 i = 0;
6647 for_each_cpu_worker_pool(pool, cpu) {
6648 BUG_ON(init_worker_pool(pool));
6649 pool->cpu = cpu;
6650 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
6651 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
6652 pool->attrs->nice = std_nice[i++];
6653 pool->attrs->affn_strict = true;
6654 pool->node = cpu_to_node(cpu);
6655
6656 /* alloc pool ID */
6657 mutex_lock(&wq_pool_mutex);
6658 BUG_ON(worker_pool_assign_id(pool));
6659 mutex_unlock(&wq_pool_mutex);
6660 }
6661 }
6662
6663 /* create default unbound and ordered wq attrs */
6664 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
6665 struct workqueue_attrs *attrs;
6666
6667 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6668 attrs->nice = std_nice[i];
6669 unbound_std_wq_attrs[i] = attrs;
6670
6671 /*
6672 * An ordered wq should have only one pwq as ordering is
6673 * guaranteed by max_active which is enforced by pwqs.
6674 */
6675 BUG_ON(!(attrs = alloc_workqueue_attrs()));
6676 attrs->nice = std_nice[i];
6677 attrs->ordered = true;
6678 ordered_wq_attrs[i] = attrs;
6679 }
6680
6681 system_wq = alloc_workqueue("events", 0, 0);
6682 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
6683 system_long_wq = alloc_workqueue("events_long", 0, 0);
6684 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
6685 WQ_MAX_ACTIVE);
6686 system_freezable_wq = alloc_workqueue("events_freezable",
6687 WQ_FREEZABLE, 0);
6688 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
6689 WQ_POWER_EFFICIENT, 0);
6690 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
6691 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
6692 0);
6693 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
6694 !system_unbound_wq || !system_freezable_wq ||
6695 !system_power_efficient_wq ||
6696 !system_freezable_power_efficient_wq);
6697}
6698
6699static void __init wq_cpu_intensive_thresh_init(void)
6700{
6701 unsigned long thresh;
6702 unsigned long bogo;
6703
6704 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
6705 BUG_ON(IS_ERR(pwq_release_worker));
6706
6707 /* if the user set it to a specific value, keep it */
6708 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
6709 return;
6710
6711 /*
6712 * The default of 10ms is derived from the fact that most modern (as of
6713 * 2023) processors can do a lot in 10ms and that it's just below what
6714 * most consider human-perceivable. However, the kernel also runs on a
6715 * lot slower CPUs including microcontrollers where the threshold is way
6716 * too low.
6717 *
6718 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
6719 * This is by no means accurate but it doesn't have to be. The mechanism
6720 * is still useful even when the threshold is fully scaled up. Also, as
6721 * the reports would usually be applicable to everyone, some machines
6722 * operating on longer thresholds won't significantly diminish their
6723 * usefulness.
6724 */
6725 thresh = 10 * USEC_PER_MSEC;
6726
6727 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
6728 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
6729 if (bogo < 4000)
6730 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
6731
6732 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
6733 loops_per_jiffy, bogo, thresh);
6734
6735 wq_cpu_intensive_thresh_us = thresh;
6736}
6737
6738/**
6739 * workqueue_init - bring workqueue subsystem fully online
6740 *
6741 * This is the second step of three-staged workqueue subsystem initialization
6742 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
6743 * been created and work items queued on them, but there are no kworkers
6744 * executing the work items yet. Populate the worker pools with the initial
6745 * workers and enable future kworker creations.
6746 */
6747void __init workqueue_init(void)
6748{
6749 struct workqueue_struct *wq;
6750 struct worker_pool *pool;
6751 int cpu, bkt;
6752
6753 wq_cpu_intensive_thresh_init();
6754
6755 mutex_lock(&wq_pool_mutex);
6756
6757 /*
6758 * Per-cpu pools created earlier could be missing node hint. Fix them
6759 * up. Also, create a rescuer for workqueues that requested it.
6760 */
6761 for_each_possible_cpu(cpu) {
6762 for_each_cpu_worker_pool(pool, cpu) {
6763 pool->node = cpu_to_node(cpu);
6764 }
6765 }
6766
6767 list_for_each_entry(wq, &workqueues, list) {
6768 WARN(init_rescuer(wq),
6769 "workqueue: failed to create early rescuer for %s",
6770 wq->name);
6771 }
6772
6773 mutex_unlock(&wq_pool_mutex);
6774
6775 /* create the initial workers */
6776 for_each_online_cpu(cpu) {
6777 for_each_cpu_worker_pool(pool, cpu) {
6778 pool->flags &= ~POOL_DISASSOCIATED;
6779 BUG_ON(!create_worker(pool));
6780 }
6781 }
6782
6783 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
6784 BUG_ON(!create_worker(pool));
6785
6786 wq_online = true;
6787 wq_watchdog_init();
6788}
6789
6790/*
6791 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
6792 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
6793 * and consecutive pod ID. The rest of @pt is initialized accordingly.
6794 */
6795static void __init init_pod_type(struct wq_pod_type *pt,
6796 bool (*cpus_share_pod)(int, int))
6797{
6798 int cur, pre, cpu, pod;
6799
6800 pt->nr_pods = 0;
6801
6802 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
6803 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
6804 BUG_ON(!pt->cpu_pod);
6805
6806 for_each_possible_cpu(cur) {
6807 for_each_possible_cpu(pre) {
6808 if (pre >= cur) {
6809 pt->cpu_pod[cur] = pt->nr_pods++;
6810 break;
6811 }
6812 if (cpus_share_pod(cur, pre)) {
6813 pt->cpu_pod[cur] = pt->cpu_pod[pre];
6814 break;
6815 }
6816 }
6817 }
6818
6819 /* init the rest to match @pt->cpu_pod[] */
6820 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
6821 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
6822 BUG_ON(!pt->pod_cpus || !pt->pod_node);
6823
6824 for (pod = 0; pod < pt->nr_pods; pod++)
6825 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
6826
6827 for_each_possible_cpu(cpu) {
6828 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
6829 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
6830 }
6831}
6832
6833static bool __init cpus_dont_share(int cpu0, int cpu1)
6834{
6835 return false;
6836}
6837
6838static bool __init cpus_share_smt(int cpu0, int cpu1)
6839{
6840#ifdef CONFIG_SCHED_SMT
6841 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
6842#else
6843 return false;
6844#endif
6845}
6846
6847static bool __init cpus_share_numa(int cpu0, int cpu1)
6848{
6849 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
6850}
6851
6852/**
6853 * workqueue_init_topology - initialize CPU pods for unbound workqueues
6854 *
6855 * This is the third step of there-staged workqueue subsystem initialization and
6856 * invoked after SMP and topology information are fully initialized. It
6857 * initializes the unbound CPU pods accordingly.
6858 */
6859void __init workqueue_init_topology(void)
6860{
6861 struct workqueue_struct *wq;
6862 int cpu;
6863
6864 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
6865 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
6866 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
6867 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
6868
6869 mutex_lock(&wq_pool_mutex);
6870
6871 /*
6872 * Workqueues allocated earlier would have all CPUs sharing the default
6873 * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU
6874 * combinations to apply per-pod sharing.
6875 */
6876 list_for_each_entry(wq, &workqueues, list) {
6877 for_each_online_cpu(cpu) {
6878 wq_update_pod(wq, cpu, cpu, true);
6879 }
6880 }
6881
6882 mutex_unlock(&wq_pool_mutex);
6883}
6884
6885void __warn_flushing_systemwide_wq(void)
6886{
6887 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
6888 dump_stack();
6889}
6890EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
6891
6892static int __init workqueue_unbound_cpus_setup(char *str)
6893{
6894 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
6895 cpumask_clear(&wq_cmdline_cpumask);
6896 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
6897 }
6898
6899 return 1;
6900}
6901__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);