Loading...
1/*
2 * kernel/workqueue.c - generic async execution with shared worker pool
3 *
4 * Copyright (C) 2002 Ingo Molnar
5 *
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
11 *
12 * Made to use alloc_percpu by Christoph Lameter.
13 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
23 *
24 * Please read Documentation/workqueue.txt for details.
25 */
26
27#include <linux/export.h>
28#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
38#include <linux/hardirq.h>
39#include <linux/mempolicy.h>
40#include <linux/freezer.h>
41#include <linux/kallsyms.h>
42#include <linux/debug_locks.h>
43#include <linux/lockdep.h>
44#include <linux/idr.h>
45#include <linux/jhash.h>
46#include <linux/hashtable.h>
47#include <linux/rculist.h>
48#include <linux/nodemask.h>
49#include <linux/moduleparam.h>
50#include <linux/uaccess.h>
51
52#include "workqueue_internal.h"
53
54enum {
55 /*
56 * worker_pool flags
57 *
58 * A bound pool is either associated or disassociated with its CPU.
59 * While associated (!DISASSOCIATED), all workers are bound to the
60 * CPU and none has %WORKER_UNBOUND set and concurrency management
61 * is in effect.
62 *
63 * While DISASSOCIATED, the cpu may be offline and all workers have
64 * %WORKER_UNBOUND set and concurrency management disabled, and may
65 * be executing on any CPU. The pool behaves as an unbound one.
66 *
67 * Note that DISASSOCIATED should be flipped only while holding
68 * attach_mutex to avoid changing binding state while
69 * worker_attach_to_pool() is in progress.
70 */
71 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
72
73 /* worker flags */
74 WORKER_DIE = 1 << 1, /* die die die */
75 WORKER_IDLE = 1 << 2, /* is idle */
76 WORKER_PREP = 1 << 3, /* preparing to run works */
77 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
78 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
79 WORKER_REBOUND = 1 << 8, /* worker was rebound */
80
81 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
82 WORKER_UNBOUND | WORKER_REBOUND,
83
84 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
85
86 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
87 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
88
89 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
90 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
91
92 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
93 /* call for help after 10ms
94 (min two ticks) */
95 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
96 CREATE_COOLDOWN = HZ, /* time to breath after fail */
97
98 /*
99 * Rescue workers are used only on emergencies and shared by
100 * all cpus. Give MIN_NICE.
101 */
102 RESCUER_NICE_LEVEL = MIN_NICE,
103 HIGHPRI_NICE_LEVEL = MIN_NICE,
104
105 WQ_NAME_LEN = 24,
106};
107
108/*
109 * Structure fields follow one of the following exclusion rules.
110 *
111 * I: Modifiable by initialization/destruction paths and read-only for
112 * everyone else.
113 *
114 * P: Preemption protected. Disabling preemption is enough and should
115 * only be modified and accessed from the local cpu.
116 *
117 * L: pool->lock protected. Access with pool->lock held.
118 *
119 * X: During normal operation, modification requires pool->lock and should
120 * be done only from local cpu. Either disabling preemption on local
121 * cpu or grabbing pool->lock is enough for read access. If
122 * POOL_DISASSOCIATED is set, it's identical to L.
123 *
124 * A: pool->attach_mutex protected.
125 *
126 * PL: wq_pool_mutex protected.
127 *
128 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
129 *
130 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
131 *
132 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
133 * sched-RCU for reads.
134 *
135 * WQ: wq->mutex protected.
136 *
137 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
138 *
139 * MD: wq_mayday_lock protected.
140 */
141
142/* struct worker is defined in workqueue_internal.h */
143
144struct worker_pool {
145 spinlock_t lock; /* the pool lock */
146 int cpu; /* I: the associated cpu */
147 int node; /* I: the associated node ID */
148 int id; /* I: pool ID */
149 unsigned int flags; /* X: flags */
150
151 unsigned long watchdog_ts; /* L: watchdog timestamp */
152
153 struct list_head worklist; /* L: list of pending works */
154 int nr_workers; /* L: total number of workers */
155
156 /* nr_idle includes the ones off idle_list for rebinding */
157 int nr_idle; /* L: currently idle ones */
158
159 struct list_head idle_list; /* X: list of idle workers */
160 struct timer_list idle_timer; /* L: worker idle timeout */
161 struct timer_list mayday_timer; /* L: SOS timer for workers */
162
163 /* a workers is either on busy_hash or idle_list, or the manager */
164 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
165 /* L: hash of busy workers */
166
167 /* see manage_workers() for details on the two manager mutexes */
168 struct mutex manager_arb; /* manager arbitration */
169 struct worker *manager; /* L: purely informational */
170 struct mutex attach_mutex; /* attach/detach exclusion */
171 struct list_head workers; /* A: attached workers */
172 struct completion *detach_completion; /* all workers detached */
173
174 struct ida worker_ida; /* worker IDs for task name */
175
176 struct workqueue_attrs *attrs; /* I: worker attributes */
177 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
178 int refcnt; /* PL: refcnt for unbound pools */
179
180 /*
181 * The current concurrency level. As it's likely to be accessed
182 * from other CPUs during try_to_wake_up(), put it in a separate
183 * cacheline.
184 */
185 atomic_t nr_running ____cacheline_aligned_in_smp;
186
187 /*
188 * Destruction of pool is sched-RCU protected to allow dereferences
189 * from get_work_pool().
190 */
191 struct rcu_head rcu;
192} ____cacheline_aligned_in_smp;
193
194/*
195 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
196 * of work_struct->data are used for flags and the remaining high bits
197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
198 * number of flag bits.
199 */
200struct pool_workqueue {
201 struct worker_pool *pool; /* I: the associated pool */
202 struct workqueue_struct *wq; /* I: the owning workqueue */
203 int work_color; /* L: current color */
204 int flush_color; /* L: flushing color */
205 int refcnt; /* L: reference count */
206 int nr_in_flight[WORK_NR_COLORS];
207 /* L: nr of in_flight works */
208 int nr_active; /* L: nr of active works */
209 int max_active; /* L: max active works */
210 struct list_head delayed_works; /* L: delayed works */
211 struct list_head pwqs_node; /* WR: node on wq->pwqs */
212 struct list_head mayday_node; /* MD: node on wq->maydays */
213
214 /*
215 * Release of unbound pwq is punted to system_wq. See put_pwq()
216 * and pwq_unbound_release_workfn() for details. pool_workqueue
217 * itself is also sched-RCU protected so that the first pwq can be
218 * determined without grabbing wq->mutex.
219 */
220 struct work_struct unbound_release_work;
221 struct rcu_head rcu;
222} __aligned(1 << WORK_STRUCT_FLAG_BITS);
223
224/*
225 * Structure used to wait for workqueue flush.
226 */
227struct wq_flusher {
228 struct list_head list; /* WQ: list of flushers */
229 int flush_color; /* WQ: flush color waiting for */
230 struct completion done; /* flush completion */
231};
232
233struct wq_device;
234
235/*
236 * The externally visible workqueue. It relays the issued work items to
237 * the appropriate worker_pool through its pool_workqueues.
238 */
239struct workqueue_struct {
240 struct list_head pwqs; /* WR: all pwqs of this wq */
241 struct list_head list; /* PR: list of all workqueues */
242
243 struct mutex mutex; /* protects this wq */
244 int work_color; /* WQ: current work color */
245 int flush_color; /* WQ: current flush color */
246 atomic_t nr_pwqs_to_flush; /* flush in progress */
247 struct wq_flusher *first_flusher; /* WQ: first flusher */
248 struct list_head flusher_queue; /* WQ: flush waiters */
249 struct list_head flusher_overflow; /* WQ: flush overflow list */
250
251 struct list_head maydays; /* MD: pwqs requesting rescue */
252 struct worker *rescuer; /* I: rescue worker */
253
254 int nr_drainers; /* WQ: drain in progress */
255 int saved_max_active; /* WQ: saved pwq max_active */
256
257 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
258 struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */
259
260#ifdef CONFIG_SYSFS
261 struct wq_device *wq_dev; /* I: for sysfs interface */
262#endif
263#ifdef CONFIG_LOCKDEP
264 struct lockdep_map lockdep_map;
265#endif
266 char name[WQ_NAME_LEN]; /* I: workqueue name */
267
268 /*
269 * Destruction of workqueue_struct is sched-RCU protected to allow
270 * walking the workqueues list without grabbing wq_pool_mutex.
271 * This is used to dump all workqueues from sysrq.
272 */
273 struct rcu_head rcu;
274
275 /* hot fields used during command issue, aligned to cacheline */
276 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
277 struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
278 struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
279};
280
281static struct kmem_cache *pwq_cache;
282
283static cpumask_var_t *wq_numa_possible_cpumask;
284 /* possible CPUs of each node */
285
286static bool wq_disable_numa;
287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
288
289/* see the comment above the definition of WQ_POWER_EFFICIENT */
290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
292
293static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
294
295/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
296static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
297
298static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
299static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
300
301static LIST_HEAD(workqueues); /* PR: list of all workqueues */
302static bool workqueue_freezing; /* PL: have wqs started freezing? */
303
304/* PL: allowable cpus for unbound wqs and work items */
305static cpumask_var_t wq_unbound_cpumask;
306
307/* CPU where unbound work was last round robin scheduled from this CPU */
308static DEFINE_PER_CPU(int, wq_rr_cpu_last);
309
310/*
311 * Local execution of unbound work items is no longer guaranteed. The
312 * following always forces round-robin CPU selection on unbound work items
313 * to uncover usages which depend on it.
314 */
315#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
316static bool wq_debug_force_rr_cpu = true;
317#else
318static bool wq_debug_force_rr_cpu = false;
319#endif
320module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
321
322/* the per-cpu worker pools */
323static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
324
325static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
326
327/* PL: hash of all unbound pools keyed by pool->attrs */
328static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
329
330/* I: attributes used when instantiating standard unbound pools on demand */
331static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
332
333/* I: attributes used when instantiating ordered pools on demand */
334static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
335
336struct workqueue_struct *system_wq __read_mostly;
337EXPORT_SYMBOL(system_wq);
338struct workqueue_struct *system_highpri_wq __read_mostly;
339EXPORT_SYMBOL_GPL(system_highpri_wq);
340struct workqueue_struct *system_long_wq __read_mostly;
341EXPORT_SYMBOL_GPL(system_long_wq);
342struct workqueue_struct *system_unbound_wq __read_mostly;
343EXPORT_SYMBOL_GPL(system_unbound_wq);
344struct workqueue_struct *system_freezable_wq __read_mostly;
345EXPORT_SYMBOL_GPL(system_freezable_wq);
346struct workqueue_struct *system_power_efficient_wq __read_mostly;
347EXPORT_SYMBOL_GPL(system_power_efficient_wq);
348struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
349EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
350
351static int worker_thread(void *__worker);
352static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
353
354#define CREATE_TRACE_POINTS
355#include <trace/events/workqueue.h>
356
357#define assert_rcu_or_pool_mutex() \
358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
359 !lockdep_is_held(&wq_pool_mutex), \
360 "sched RCU or wq_pool_mutex should be held")
361
362#define assert_rcu_or_wq_mutex(wq) \
363 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
364 !lockdep_is_held(&wq->mutex), \
365 "sched RCU or wq->mutex should be held")
366
367#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() && \
369 !lockdep_is_held(&wq->mutex) && \
370 !lockdep_is_held(&wq_pool_mutex), \
371 "sched RCU, wq->mutex or wq_pool_mutex should be held")
372
373#define for_each_cpu_worker_pool(pool, cpu) \
374 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
375 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
376 (pool)++)
377
378/**
379 * for_each_pool - iterate through all worker_pools in the system
380 * @pool: iteration cursor
381 * @pi: integer used for iteration
382 *
383 * This must be called either with wq_pool_mutex held or sched RCU read
384 * locked. If the pool needs to be used beyond the locking in effect, the
385 * caller is responsible for guaranteeing that the pool stays online.
386 *
387 * The if/else clause exists only for the lockdep assertion and can be
388 * ignored.
389 */
390#define for_each_pool(pool, pi) \
391 idr_for_each_entry(&worker_pool_idr, pool, pi) \
392 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
393 else
394
395/**
396 * for_each_pool_worker - iterate through all workers of a worker_pool
397 * @worker: iteration cursor
398 * @pool: worker_pool to iterate workers of
399 *
400 * This must be called with @pool->attach_mutex.
401 *
402 * The if/else clause exists only for the lockdep assertion and can be
403 * ignored.
404 */
405#define for_each_pool_worker(worker, pool) \
406 list_for_each_entry((worker), &(pool)->workers, node) \
407 if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
408 else
409
410/**
411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
412 * @pwq: iteration cursor
413 * @wq: the target workqueue
414 *
415 * This must be called either with wq->mutex held or sched RCU read locked.
416 * If the pwq needs to be used beyond the locking in effect, the caller is
417 * responsible for guaranteeing that the pwq stays online.
418 *
419 * The if/else clause exists only for the lockdep assertion and can be
420 * ignored.
421 */
422#define for_each_pwq(pwq, wq) \
423 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
424 if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
425 else
426
427#ifdef CONFIG_DEBUG_OBJECTS_WORK
428
429static struct debug_obj_descr work_debug_descr;
430
431static void *work_debug_hint(void *addr)
432{
433 return ((struct work_struct *) addr)->func;
434}
435
436/*
437 * fixup_init is called when:
438 * - an active object is initialized
439 */
440static int work_fixup_init(void *addr, enum debug_obj_state state)
441{
442 struct work_struct *work = addr;
443
444 switch (state) {
445 case ODEBUG_STATE_ACTIVE:
446 cancel_work_sync(work);
447 debug_object_init(work, &work_debug_descr);
448 return 1;
449 default:
450 return 0;
451 }
452}
453
454/*
455 * fixup_activate is called when:
456 * - an active object is activated
457 * - an unknown object is activated (might be a statically initialized object)
458 */
459static int work_fixup_activate(void *addr, enum debug_obj_state state)
460{
461 struct work_struct *work = addr;
462
463 switch (state) {
464
465 case ODEBUG_STATE_NOTAVAILABLE:
466 /*
467 * This is not really a fixup. The work struct was
468 * statically initialized. We just make sure that it
469 * is tracked in the object tracker.
470 */
471 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
472 debug_object_init(work, &work_debug_descr);
473 debug_object_activate(work, &work_debug_descr);
474 return 0;
475 }
476 WARN_ON_ONCE(1);
477 return 0;
478
479 case ODEBUG_STATE_ACTIVE:
480 WARN_ON(1);
481
482 default:
483 return 0;
484 }
485}
486
487/*
488 * fixup_free is called when:
489 * - an active object is freed
490 */
491static int work_fixup_free(void *addr, enum debug_obj_state state)
492{
493 struct work_struct *work = addr;
494
495 switch (state) {
496 case ODEBUG_STATE_ACTIVE:
497 cancel_work_sync(work);
498 debug_object_free(work, &work_debug_descr);
499 return 1;
500 default:
501 return 0;
502 }
503}
504
505static struct debug_obj_descr work_debug_descr = {
506 .name = "work_struct",
507 .debug_hint = work_debug_hint,
508 .fixup_init = work_fixup_init,
509 .fixup_activate = work_fixup_activate,
510 .fixup_free = work_fixup_free,
511};
512
513static inline void debug_work_activate(struct work_struct *work)
514{
515 debug_object_activate(work, &work_debug_descr);
516}
517
518static inline void debug_work_deactivate(struct work_struct *work)
519{
520 debug_object_deactivate(work, &work_debug_descr);
521}
522
523void __init_work(struct work_struct *work, int onstack)
524{
525 if (onstack)
526 debug_object_init_on_stack(work, &work_debug_descr);
527 else
528 debug_object_init(work, &work_debug_descr);
529}
530EXPORT_SYMBOL_GPL(__init_work);
531
532void destroy_work_on_stack(struct work_struct *work)
533{
534 debug_object_free(work, &work_debug_descr);
535}
536EXPORT_SYMBOL_GPL(destroy_work_on_stack);
537
538void destroy_delayed_work_on_stack(struct delayed_work *work)
539{
540 destroy_timer_on_stack(&work->timer);
541 debug_object_free(&work->work, &work_debug_descr);
542}
543EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
544
545#else
546static inline void debug_work_activate(struct work_struct *work) { }
547static inline void debug_work_deactivate(struct work_struct *work) { }
548#endif
549
550/**
551 * worker_pool_assign_id - allocate ID and assing it to @pool
552 * @pool: the pool pointer of interest
553 *
554 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
555 * successfully, -errno on failure.
556 */
557static int worker_pool_assign_id(struct worker_pool *pool)
558{
559 int ret;
560
561 lockdep_assert_held(&wq_pool_mutex);
562
563 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
564 GFP_KERNEL);
565 if (ret >= 0) {
566 pool->id = ret;
567 return 0;
568 }
569 return ret;
570}
571
572/**
573 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
574 * @wq: the target workqueue
575 * @node: the node ID
576 *
577 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
578 * read locked.
579 * If the pwq needs to be used beyond the locking in effect, the caller is
580 * responsible for guaranteeing that the pwq stays online.
581 *
582 * Return: The unbound pool_workqueue for @node.
583 */
584static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
585 int node)
586{
587 assert_rcu_or_wq_mutex_or_pool_mutex(wq);
588
589 /*
590 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
591 * delayed item is pending. The plan is to keep CPU -> NODE
592 * mapping valid and stable across CPU on/offlines. Once that
593 * happens, this workaround can be removed.
594 */
595 if (unlikely(node == NUMA_NO_NODE))
596 return wq->dfl_pwq;
597
598 return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
599}
600
601static unsigned int work_color_to_flags(int color)
602{
603 return color << WORK_STRUCT_COLOR_SHIFT;
604}
605
606static int get_work_color(struct work_struct *work)
607{
608 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
609 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
610}
611
612static int work_next_color(int color)
613{
614 return (color + 1) % WORK_NR_COLORS;
615}
616
617/*
618 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
619 * contain the pointer to the queued pwq. Once execution starts, the flag
620 * is cleared and the high bits contain OFFQ flags and pool ID.
621 *
622 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
623 * and clear_work_data() can be used to set the pwq, pool or clear
624 * work->data. These functions should only be called while the work is
625 * owned - ie. while the PENDING bit is set.
626 *
627 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
628 * corresponding to a work. Pool is available once the work has been
629 * queued anywhere after initialization until it is sync canceled. pwq is
630 * available only while the work item is queued.
631 *
632 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
633 * canceled. While being canceled, a work item may have its PENDING set
634 * but stay off timer and worklist for arbitrarily long and nobody should
635 * try to steal the PENDING bit.
636 */
637static inline void set_work_data(struct work_struct *work, unsigned long data,
638 unsigned long flags)
639{
640 WARN_ON_ONCE(!work_pending(work));
641 atomic_long_set(&work->data, data | flags | work_static(work));
642}
643
644static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
645 unsigned long extra_flags)
646{
647 set_work_data(work, (unsigned long)pwq,
648 WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
649}
650
651static void set_work_pool_and_keep_pending(struct work_struct *work,
652 int pool_id)
653{
654 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
655 WORK_STRUCT_PENDING);
656}
657
658static void set_work_pool_and_clear_pending(struct work_struct *work,
659 int pool_id)
660{
661 /*
662 * The following wmb is paired with the implied mb in
663 * test_and_set_bit(PENDING) and ensures all updates to @work made
664 * here are visible to and precede any updates by the next PENDING
665 * owner.
666 */
667 smp_wmb();
668 set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
669 /*
670 * The following mb guarantees that previous clear of a PENDING bit
671 * will not be reordered with any speculative LOADS or STORES from
672 * work->current_func, which is executed afterwards. This possible
673 * reordering can lead to a missed execution on attempt to qeueue
674 * the same @work. E.g. consider this case:
675 *
676 * CPU#0 CPU#1
677 * ---------------------------- --------------------------------
678 *
679 * 1 STORE event_indicated
680 * 2 queue_work_on() {
681 * 3 test_and_set_bit(PENDING)
682 * 4 } set_..._and_clear_pending() {
683 * 5 set_work_data() # clear bit
684 * 6 smp_mb()
685 * 7 work->current_func() {
686 * 8 LOAD event_indicated
687 * }
688 *
689 * Without an explicit full barrier speculative LOAD on line 8 can
690 * be executed before CPU#0 does STORE on line 1. If that happens,
691 * CPU#0 observes the PENDING bit is still set and new execution of
692 * a @work is not queued in a hope, that CPU#1 will eventually
693 * finish the queued @work. Meanwhile CPU#1 does not see
694 * event_indicated is set, because speculative LOAD was executed
695 * before actual STORE.
696 */
697 smp_mb();
698}
699
700static void clear_work_data(struct work_struct *work)
701{
702 smp_wmb(); /* see set_work_pool_and_clear_pending() */
703 set_work_data(work, WORK_STRUCT_NO_POOL, 0);
704}
705
706static struct pool_workqueue *get_work_pwq(struct work_struct *work)
707{
708 unsigned long data = atomic_long_read(&work->data);
709
710 if (data & WORK_STRUCT_PWQ)
711 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
712 else
713 return NULL;
714}
715
716/**
717 * get_work_pool - return the worker_pool a given work was associated with
718 * @work: the work item of interest
719 *
720 * Pools are created and destroyed under wq_pool_mutex, and allows read
721 * access under sched-RCU read lock. As such, this function should be
722 * called under wq_pool_mutex or with preemption disabled.
723 *
724 * All fields of the returned pool are accessible as long as the above
725 * mentioned locking is in effect. If the returned pool needs to be used
726 * beyond the critical section, the caller is responsible for ensuring the
727 * returned pool is and stays online.
728 *
729 * Return: The worker_pool @work was last associated with. %NULL if none.
730 */
731static struct worker_pool *get_work_pool(struct work_struct *work)
732{
733 unsigned long data = atomic_long_read(&work->data);
734 int pool_id;
735
736 assert_rcu_or_pool_mutex();
737
738 if (data & WORK_STRUCT_PWQ)
739 return ((struct pool_workqueue *)
740 (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
741
742 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
743 if (pool_id == WORK_OFFQ_POOL_NONE)
744 return NULL;
745
746 return idr_find(&worker_pool_idr, pool_id);
747}
748
749/**
750 * get_work_pool_id - return the worker pool ID a given work is associated with
751 * @work: the work item of interest
752 *
753 * Return: The worker_pool ID @work was last associated with.
754 * %WORK_OFFQ_POOL_NONE if none.
755 */
756static int get_work_pool_id(struct work_struct *work)
757{
758 unsigned long data = atomic_long_read(&work->data);
759
760 if (data & WORK_STRUCT_PWQ)
761 return ((struct pool_workqueue *)
762 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
763
764 return data >> WORK_OFFQ_POOL_SHIFT;
765}
766
767static void mark_work_canceling(struct work_struct *work)
768{
769 unsigned long pool_id = get_work_pool_id(work);
770
771 pool_id <<= WORK_OFFQ_POOL_SHIFT;
772 set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
773}
774
775static bool work_is_canceling(struct work_struct *work)
776{
777 unsigned long data = atomic_long_read(&work->data);
778
779 return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
780}
781
782/*
783 * Policy functions. These define the policies on how the global worker
784 * pools are managed. Unless noted otherwise, these functions assume that
785 * they're being called with pool->lock held.
786 */
787
788static bool __need_more_worker(struct worker_pool *pool)
789{
790 return !atomic_read(&pool->nr_running);
791}
792
793/*
794 * Need to wake up a worker? Called from anything but currently
795 * running workers.
796 *
797 * Note that, because unbound workers never contribute to nr_running, this
798 * function will always return %true for unbound pools as long as the
799 * worklist isn't empty.
800 */
801static bool need_more_worker(struct worker_pool *pool)
802{
803 return !list_empty(&pool->worklist) && __need_more_worker(pool);
804}
805
806/* Can I start working? Called from busy but !running workers. */
807static bool may_start_working(struct worker_pool *pool)
808{
809 return pool->nr_idle;
810}
811
812/* Do I need to keep working? Called from currently running workers. */
813static bool keep_working(struct worker_pool *pool)
814{
815 return !list_empty(&pool->worklist) &&
816 atomic_read(&pool->nr_running) <= 1;
817}
818
819/* Do we need a new worker? Called from manager. */
820static bool need_to_create_worker(struct worker_pool *pool)
821{
822 return need_more_worker(pool) && !may_start_working(pool);
823}
824
825/* Do we have too many workers and should some go away? */
826static bool too_many_workers(struct worker_pool *pool)
827{
828 bool managing = mutex_is_locked(&pool->manager_arb);
829 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
830 int nr_busy = pool->nr_workers - nr_idle;
831
832 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
833}
834
835/*
836 * Wake up functions.
837 */
838
839/* Return the first idle worker. Safe with preemption disabled */
840static struct worker *first_idle_worker(struct worker_pool *pool)
841{
842 if (unlikely(list_empty(&pool->idle_list)))
843 return NULL;
844
845 return list_first_entry(&pool->idle_list, struct worker, entry);
846}
847
848/**
849 * wake_up_worker - wake up an idle worker
850 * @pool: worker pool to wake worker from
851 *
852 * Wake up the first idle worker of @pool.
853 *
854 * CONTEXT:
855 * spin_lock_irq(pool->lock).
856 */
857static void wake_up_worker(struct worker_pool *pool)
858{
859 struct worker *worker = first_idle_worker(pool);
860
861 if (likely(worker))
862 wake_up_process(worker->task);
863}
864
865/**
866 * wq_worker_waking_up - a worker is waking up
867 * @task: task waking up
868 * @cpu: CPU @task is waking up to
869 *
870 * This function is called during try_to_wake_up() when a worker is
871 * being awoken.
872 *
873 * CONTEXT:
874 * spin_lock_irq(rq->lock)
875 */
876void wq_worker_waking_up(struct task_struct *task, int cpu)
877{
878 struct worker *worker = kthread_data(task);
879
880 if (!(worker->flags & WORKER_NOT_RUNNING)) {
881 WARN_ON_ONCE(worker->pool->cpu != cpu);
882 atomic_inc(&worker->pool->nr_running);
883 }
884}
885
886/**
887 * wq_worker_sleeping - a worker is going to sleep
888 * @task: task going to sleep
889 *
890 * This function is called during schedule() when a busy worker is
891 * going to sleep. Worker on the same cpu can be woken up by
892 * returning pointer to its task.
893 *
894 * CONTEXT:
895 * spin_lock_irq(rq->lock)
896 *
897 * Return:
898 * Worker task on @cpu to wake up, %NULL if none.
899 */
900struct task_struct *wq_worker_sleeping(struct task_struct *task)
901{
902 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
903 struct worker_pool *pool;
904
905 /*
906 * Rescuers, which may not have all the fields set up like normal
907 * workers, also reach here, let's not access anything before
908 * checking NOT_RUNNING.
909 */
910 if (worker->flags & WORKER_NOT_RUNNING)
911 return NULL;
912
913 pool = worker->pool;
914
915 /* this can only happen on the local cpu */
916 if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
917 return NULL;
918
919 /*
920 * The counterpart of the following dec_and_test, implied mb,
921 * worklist not empty test sequence is in insert_work().
922 * Please read comment there.
923 *
924 * NOT_RUNNING is clear. This means that we're bound to and
925 * running on the local cpu w/ rq lock held and preemption
926 * disabled, which in turn means that none else could be
927 * manipulating idle_list, so dereferencing idle_list without pool
928 * lock is safe.
929 */
930 if (atomic_dec_and_test(&pool->nr_running) &&
931 !list_empty(&pool->worklist))
932 to_wakeup = first_idle_worker(pool);
933 return to_wakeup ? to_wakeup->task : NULL;
934}
935
936/**
937 * worker_set_flags - set worker flags and adjust nr_running accordingly
938 * @worker: self
939 * @flags: flags to set
940 *
941 * Set @flags in @worker->flags and adjust nr_running accordingly.
942 *
943 * CONTEXT:
944 * spin_lock_irq(pool->lock)
945 */
946static inline void worker_set_flags(struct worker *worker, unsigned int flags)
947{
948 struct worker_pool *pool = worker->pool;
949
950 WARN_ON_ONCE(worker->task != current);
951
952 /* If transitioning into NOT_RUNNING, adjust nr_running. */
953 if ((flags & WORKER_NOT_RUNNING) &&
954 !(worker->flags & WORKER_NOT_RUNNING)) {
955 atomic_dec(&pool->nr_running);
956 }
957
958 worker->flags |= flags;
959}
960
961/**
962 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
963 * @worker: self
964 * @flags: flags to clear
965 *
966 * Clear @flags in @worker->flags and adjust nr_running accordingly.
967 *
968 * CONTEXT:
969 * spin_lock_irq(pool->lock)
970 */
971static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
972{
973 struct worker_pool *pool = worker->pool;
974 unsigned int oflags = worker->flags;
975
976 WARN_ON_ONCE(worker->task != current);
977
978 worker->flags &= ~flags;
979
980 /*
981 * If transitioning out of NOT_RUNNING, increment nr_running. Note
982 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
983 * of multiple flags, not a single flag.
984 */
985 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
986 if (!(worker->flags & WORKER_NOT_RUNNING))
987 atomic_inc(&pool->nr_running);
988}
989
990/**
991 * find_worker_executing_work - find worker which is executing a work
992 * @pool: pool of interest
993 * @work: work to find worker for
994 *
995 * Find a worker which is executing @work on @pool by searching
996 * @pool->busy_hash which is keyed by the address of @work. For a worker
997 * to match, its current execution should match the address of @work and
998 * its work function. This is to avoid unwanted dependency between
999 * unrelated work executions through a work item being recycled while still
1000 * being executed.
1001 *
1002 * This is a bit tricky. A work item may be freed once its execution
1003 * starts and nothing prevents the freed area from being recycled for
1004 * another work item. If the same work item address ends up being reused
1005 * before the original execution finishes, workqueue will identify the
1006 * recycled work item as currently executing and make it wait until the
1007 * current execution finishes, introducing an unwanted dependency.
1008 *
1009 * This function checks the work item address and work function to avoid
1010 * false positives. Note that this isn't complete as one may construct a
1011 * work function which can introduce dependency onto itself through a
1012 * recycled work item. Well, if somebody wants to shoot oneself in the
1013 * foot that badly, there's only so much we can do, and if such deadlock
1014 * actually occurs, it should be easy to locate the culprit work function.
1015 *
1016 * CONTEXT:
1017 * spin_lock_irq(pool->lock).
1018 *
1019 * Return:
1020 * Pointer to worker which is executing @work if found, %NULL
1021 * otherwise.
1022 */
1023static struct worker *find_worker_executing_work(struct worker_pool *pool,
1024 struct work_struct *work)
1025{
1026 struct worker *worker;
1027
1028 hash_for_each_possible(pool->busy_hash, worker, hentry,
1029 (unsigned long)work)
1030 if (worker->current_work == work &&
1031 worker->current_func == work->func)
1032 return worker;
1033
1034 return NULL;
1035}
1036
1037/**
1038 * move_linked_works - move linked works to a list
1039 * @work: start of series of works to be scheduled
1040 * @head: target list to append @work to
1041 * @nextp: out parameter for nested worklist walking
1042 *
1043 * Schedule linked works starting from @work to @head. Work series to
1044 * be scheduled starts at @work and includes any consecutive work with
1045 * WORK_STRUCT_LINKED set in its predecessor.
1046 *
1047 * If @nextp is not NULL, it's updated to point to the next work of
1048 * the last scheduled work. This allows move_linked_works() to be
1049 * nested inside outer list_for_each_entry_safe().
1050 *
1051 * CONTEXT:
1052 * spin_lock_irq(pool->lock).
1053 */
1054static void move_linked_works(struct work_struct *work, struct list_head *head,
1055 struct work_struct **nextp)
1056{
1057 struct work_struct *n;
1058
1059 /*
1060 * Linked worklist will always end before the end of the list,
1061 * use NULL for list head.
1062 */
1063 list_for_each_entry_safe_from(work, n, NULL, entry) {
1064 list_move_tail(&work->entry, head);
1065 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1066 break;
1067 }
1068
1069 /*
1070 * If we're already inside safe list traversal and have moved
1071 * multiple works to the scheduled queue, the next position
1072 * needs to be updated.
1073 */
1074 if (nextp)
1075 *nextp = n;
1076}
1077
1078/**
1079 * get_pwq - get an extra reference on the specified pool_workqueue
1080 * @pwq: pool_workqueue to get
1081 *
1082 * Obtain an extra reference on @pwq. The caller should guarantee that
1083 * @pwq has positive refcnt and be holding the matching pool->lock.
1084 */
1085static void get_pwq(struct pool_workqueue *pwq)
1086{
1087 lockdep_assert_held(&pwq->pool->lock);
1088 WARN_ON_ONCE(pwq->refcnt <= 0);
1089 pwq->refcnt++;
1090}
1091
1092/**
1093 * put_pwq - put a pool_workqueue reference
1094 * @pwq: pool_workqueue to put
1095 *
1096 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1097 * destruction. The caller should be holding the matching pool->lock.
1098 */
1099static void put_pwq(struct pool_workqueue *pwq)
1100{
1101 lockdep_assert_held(&pwq->pool->lock);
1102 if (likely(--pwq->refcnt))
1103 return;
1104 if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105 return;
1106 /*
1107 * @pwq can't be released under pool->lock, bounce to
1108 * pwq_unbound_release_workfn(). This never recurses on the same
1109 * pool->lock as this path is taken only for unbound workqueues and
1110 * the release work item is scheduled on a per-cpu workqueue. To
1111 * avoid lockdep warning, unbound pool->locks are given lockdep
1112 * subclass of 1 in get_unbound_pool().
1113 */
1114 schedule_work(&pwq->unbound_release_work);
1115}
1116
1117/**
1118 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1119 * @pwq: pool_workqueue to put (can be %NULL)
1120 *
1121 * put_pwq() with locking. This function also allows %NULL @pwq.
1122 */
1123static void put_pwq_unlocked(struct pool_workqueue *pwq)
1124{
1125 if (pwq) {
1126 /*
1127 * As both pwqs and pools are sched-RCU protected, the
1128 * following lock operations are safe.
1129 */
1130 spin_lock_irq(&pwq->pool->lock);
1131 put_pwq(pwq);
1132 spin_unlock_irq(&pwq->pool->lock);
1133 }
1134}
1135
1136static void pwq_activate_delayed_work(struct work_struct *work)
1137{
1138 struct pool_workqueue *pwq = get_work_pwq(work);
1139
1140 trace_workqueue_activate_work(work);
1141 if (list_empty(&pwq->pool->worklist))
1142 pwq->pool->watchdog_ts = jiffies;
1143 move_linked_works(work, &pwq->pool->worklist, NULL);
1144 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1145 pwq->nr_active++;
1146}
1147
1148static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1149{
1150 struct work_struct *work = list_first_entry(&pwq->delayed_works,
1151 struct work_struct, entry);
1152
1153 pwq_activate_delayed_work(work);
1154}
1155
1156/**
1157 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1158 * @pwq: pwq of interest
1159 * @color: color of work which left the queue
1160 *
1161 * A work either has completed or is removed from pending queue,
1162 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1163 *
1164 * CONTEXT:
1165 * spin_lock_irq(pool->lock).
1166 */
1167static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1168{
1169 /* uncolored work items don't participate in flushing or nr_active */
1170 if (color == WORK_NO_COLOR)
1171 goto out_put;
1172
1173 pwq->nr_in_flight[color]--;
1174
1175 pwq->nr_active--;
1176 if (!list_empty(&pwq->delayed_works)) {
1177 /* one down, submit a delayed one */
1178 if (pwq->nr_active < pwq->max_active)
1179 pwq_activate_first_delayed(pwq);
1180 }
1181
1182 /* is flush in progress and are we at the flushing tip? */
1183 if (likely(pwq->flush_color != color))
1184 goto out_put;
1185
1186 /* are there still in-flight works? */
1187 if (pwq->nr_in_flight[color])
1188 goto out_put;
1189
1190 /* this pwq is done, clear flush_color */
1191 pwq->flush_color = -1;
1192
1193 /*
1194 * If this was the last pwq, wake up the first flusher. It
1195 * will handle the rest.
1196 */
1197 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1198 complete(&pwq->wq->first_flusher->done);
1199out_put:
1200 put_pwq(pwq);
1201}
1202
1203/**
1204 * try_to_grab_pending - steal work item from worklist and disable irq
1205 * @work: work item to steal
1206 * @is_dwork: @work is a delayed_work
1207 * @flags: place to store irq state
1208 *
1209 * Try to grab PENDING bit of @work. This function can handle @work in any
1210 * stable state - idle, on timer or on worklist.
1211 *
1212 * Return:
1213 * 1 if @work was pending and we successfully stole PENDING
1214 * 0 if @work was idle and we claimed PENDING
1215 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1216 * -ENOENT if someone else is canceling @work, this state may persist
1217 * for arbitrarily long
1218 *
1219 * Note:
1220 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1221 * interrupted while holding PENDING and @work off queue, irq must be
1222 * disabled on entry. This, combined with delayed_work->timer being
1223 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1224 *
1225 * On successful return, >= 0, irq is disabled and the caller is
1226 * responsible for releasing it using local_irq_restore(*@flags).
1227 *
1228 * This function is safe to call from any context including IRQ handler.
1229 */
1230static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1231 unsigned long *flags)
1232{
1233 struct worker_pool *pool;
1234 struct pool_workqueue *pwq;
1235
1236 local_irq_save(*flags);
1237
1238 /* try to steal the timer if it exists */
1239 if (is_dwork) {
1240 struct delayed_work *dwork = to_delayed_work(work);
1241
1242 /*
1243 * dwork->timer is irqsafe. If del_timer() fails, it's
1244 * guaranteed that the timer is not queued anywhere and not
1245 * running on the local CPU.
1246 */
1247 if (likely(del_timer(&dwork->timer)))
1248 return 1;
1249 }
1250
1251 /* try to claim PENDING the normal way */
1252 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1253 return 0;
1254
1255 /*
1256 * The queueing is in progress, or it is already queued. Try to
1257 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258 */
1259 pool = get_work_pool(work);
1260 if (!pool)
1261 goto fail;
1262
1263 spin_lock(&pool->lock);
1264 /*
1265 * work->data is guaranteed to point to pwq only while the work
1266 * item is queued on pwq->wq, and both updating work->data to point
1267 * to pwq on queueing and to pool on dequeueing are done under
1268 * pwq->pool->lock. This in turn guarantees that, if work->data
1269 * points to pwq which is associated with a locked pool, the work
1270 * item is currently queued on that pool.
1271 */
1272 pwq = get_work_pwq(work);
1273 if (pwq && pwq->pool == pool) {
1274 debug_work_deactivate(work);
1275
1276 /*
1277 * A delayed work item cannot be grabbed directly because
1278 * it might have linked NO_COLOR work items which, if left
1279 * on the delayed_list, will confuse pwq->nr_active
1280 * management later on and cause stall. Make sure the work
1281 * item is activated before grabbing.
1282 */
1283 if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1284 pwq_activate_delayed_work(work);
1285
1286 list_del_init(&work->entry);
1287 pwq_dec_nr_in_flight(pwq, get_work_color(work));
1288
1289 /* work->data points to pwq iff queued, point to pool */
1290 set_work_pool_and_keep_pending(work, pool->id);
1291
1292 spin_unlock(&pool->lock);
1293 return 1;
1294 }
1295 spin_unlock(&pool->lock);
1296fail:
1297 local_irq_restore(*flags);
1298 if (work_is_canceling(work))
1299 return -ENOENT;
1300 cpu_relax();
1301 return -EAGAIN;
1302}
1303
1304/**
1305 * insert_work - insert a work into a pool
1306 * @pwq: pwq @work belongs to
1307 * @work: work to insert
1308 * @head: insertion point
1309 * @extra_flags: extra WORK_STRUCT_* flags to set
1310 *
1311 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1312 * work_struct flags.
1313 *
1314 * CONTEXT:
1315 * spin_lock_irq(pool->lock).
1316 */
1317static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1318 struct list_head *head, unsigned int extra_flags)
1319{
1320 struct worker_pool *pool = pwq->pool;
1321
1322 /* we own @work, set data and link */
1323 set_work_pwq(work, pwq, extra_flags);
1324 list_add_tail(&work->entry, head);
1325 get_pwq(pwq);
1326
1327 /*
1328 * Ensure either wq_worker_sleeping() sees the above
1329 * list_add_tail() or we see zero nr_running to avoid workers lying
1330 * around lazily while there are works to be processed.
1331 */
1332 smp_mb();
1333
1334 if (__need_more_worker(pool))
1335 wake_up_worker(pool);
1336}
1337
1338/*
1339 * Test whether @work is being queued from another work executing on the
1340 * same workqueue.
1341 */
1342static bool is_chained_work(struct workqueue_struct *wq)
1343{
1344 struct worker *worker;
1345
1346 worker = current_wq_worker();
1347 /*
1348 * Return %true iff I'm a worker execuing a work item on @wq. If
1349 * I'm @worker, it's safe to dereference it without locking.
1350 */
1351 return worker && worker->current_pwq->wq == wq;
1352}
1353
1354/*
1355 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1356 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
1357 * avoid perturbing sensitive tasks.
1358 */
1359static int wq_select_unbound_cpu(int cpu)
1360{
1361 static bool printed_dbg_warning;
1362 int new_cpu;
1363
1364 if (likely(!wq_debug_force_rr_cpu)) {
1365 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1366 return cpu;
1367 } else if (!printed_dbg_warning) {
1368 pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1369 printed_dbg_warning = true;
1370 }
1371
1372 if (cpumask_empty(wq_unbound_cpumask))
1373 return cpu;
1374
1375 new_cpu = __this_cpu_read(wq_rr_cpu_last);
1376 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1377 if (unlikely(new_cpu >= nr_cpu_ids)) {
1378 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1379 if (unlikely(new_cpu >= nr_cpu_ids))
1380 return cpu;
1381 }
1382 __this_cpu_write(wq_rr_cpu_last, new_cpu);
1383
1384 return new_cpu;
1385}
1386
1387static void __queue_work(int cpu, struct workqueue_struct *wq,
1388 struct work_struct *work)
1389{
1390 struct pool_workqueue *pwq;
1391 struct worker_pool *last_pool;
1392 struct list_head *worklist;
1393 unsigned int work_flags;
1394 unsigned int req_cpu = cpu;
1395
1396 /*
1397 * While a work item is PENDING && off queue, a task trying to
1398 * steal the PENDING will busy-loop waiting for it to either get
1399 * queued or lose PENDING. Grabbing PENDING and queueing should
1400 * happen with IRQ disabled.
1401 */
1402 WARN_ON_ONCE(!irqs_disabled());
1403
1404 debug_work_activate(work);
1405
1406 /* if draining, only works from the same workqueue are allowed */
1407 if (unlikely(wq->flags & __WQ_DRAINING) &&
1408 WARN_ON_ONCE(!is_chained_work(wq)))
1409 return;
1410retry:
1411 if (req_cpu == WORK_CPU_UNBOUND)
1412 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1413
1414 /* pwq which will be used unless @work is executing elsewhere */
1415 if (!(wq->flags & WQ_UNBOUND))
1416 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1417 else
1418 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419
1420 /*
1421 * If @work was previously on a different pool, it might still be
1422 * running there, in which case the work needs to be queued on that
1423 * pool to guarantee non-reentrancy.
1424 */
1425 last_pool = get_work_pool(work);
1426 if (last_pool && last_pool != pwq->pool) {
1427 struct worker *worker;
1428
1429 spin_lock(&last_pool->lock);
1430
1431 worker = find_worker_executing_work(last_pool, work);
1432
1433 if (worker && worker->current_pwq->wq == wq) {
1434 pwq = worker->current_pwq;
1435 } else {
1436 /* meh... not running there, queue here */
1437 spin_unlock(&last_pool->lock);
1438 spin_lock(&pwq->pool->lock);
1439 }
1440 } else {
1441 spin_lock(&pwq->pool->lock);
1442 }
1443
1444 /*
1445 * pwq is determined and locked. For unbound pools, we could have
1446 * raced with pwq release and it could already be dead. If its
1447 * refcnt is zero, repeat pwq selection. Note that pwqs never die
1448 * without another pwq replacing it in the numa_pwq_tbl or while
1449 * work items are executing on it, so the retrying is guaranteed to
1450 * make forward-progress.
1451 */
1452 if (unlikely(!pwq->refcnt)) {
1453 if (wq->flags & WQ_UNBOUND) {
1454 spin_unlock(&pwq->pool->lock);
1455 cpu_relax();
1456 goto retry;
1457 }
1458 /* oops */
1459 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1460 wq->name, cpu);
1461 }
1462
1463 /* pwq determined, queue */
1464 trace_workqueue_queue_work(req_cpu, pwq, work);
1465
1466 if (WARN_ON(!list_empty(&work->entry))) {
1467 spin_unlock(&pwq->pool->lock);
1468 return;
1469 }
1470
1471 pwq->nr_in_flight[pwq->work_color]++;
1472 work_flags = work_color_to_flags(pwq->work_color);
1473
1474 if (likely(pwq->nr_active < pwq->max_active)) {
1475 trace_workqueue_activate_work(work);
1476 pwq->nr_active++;
1477 worklist = &pwq->pool->worklist;
1478 if (list_empty(worklist))
1479 pwq->pool->watchdog_ts = jiffies;
1480 } else {
1481 work_flags |= WORK_STRUCT_DELAYED;
1482 worklist = &pwq->delayed_works;
1483 }
1484
1485 insert_work(pwq, work, worklist, work_flags);
1486
1487 spin_unlock(&pwq->pool->lock);
1488}
1489
1490/**
1491 * queue_work_on - queue work on specific cpu
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @work: work to queue
1495 *
1496 * We queue the work to a specific CPU, the caller must ensure it
1497 * can't go away.
1498 *
1499 * Return: %false if @work was already on a queue, %true otherwise.
1500 */
1501bool queue_work_on(int cpu, struct workqueue_struct *wq,
1502 struct work_struct *work)
1503{
1504 bool ret = false;
1505 unsigned long flags;
1506
1507 local_irq_save(flags);
1508
1509 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1510 __queue_work(cpu, wq, work);
1511 ret = true;
1512 }
1513
1514 local_irq_restore(flags);
1515 return ret;
1516}
1517EXPORT_SYMBOL(queue_work_on);
1518
1519void delayed_work_timer_fn(unsigned long __data)
1520{
1521 struct delayed_work *dwork = (struct delayed_work *)__data;
1522
1523 /* should have been called from irqsafe timer with irq already off */
1524 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1525}
1526EXPORT_SYMBOL(delayed_work_timer_fn);
1527
1528static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1529 struct delayed_work *dwork, unsigned long delay)
1530{
1531 struct timer_list *timer = &dwork->timer;
1532 struct work_struct *work = &dwork->work;
1533
1534 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1535 timer->data != (unsigned long)dwork);
1536 WARN_ON_ONCE(timer_pending(timer));
1537 WARN_ON_ONCE(!list_empty(&work->entry));
1538
1539 /*
1540 * If @delay is 0, queue @dwork->work immediately. This is for
1541 * both optimization and correctness. The earliest @timer can
1542 * expire is on the closest next tick and delayed_work users depend
1543 * on that there's no such delay when @delay is 0.
1544 */
1545 if (!delay) {
1546 __queue_work(cpu, wq, &dwork->work);
1547 return;
1548 }
1549
1550 timer_stats_timer_set_start_info(&dwork->timer);
1551
1552 dwork->wq = wq;
1553 dwork->cpu = cpu;
1554 timer->expires = jiffies + delay;
1555
1556 if (unlikely(cpu != WORK_CPU_UNBOUND))
1557 add_timer_on(timer, cpu);
1558 else
1559 add_timer(timer);
1560}
1561
1562/**
1563 * queue_delayed_work_on - queue work on specific CPU after delay
1564 * @cpu: CPU number to execute work on
1565 * @wq: workqueue to use
1566 * @dwork: work to queue
1567 * @delay: number of jiffies to wait before queueing
1568 *
1569 * Return: %false if @work was already on a queue, %true otherwise. If
1570 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1571 * execution.
1572 */
1573bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1574 struct delayed_work *dwork, unsigned long delay)
1575{
1576 struct work_struct *work = &dwork->work;
1577 bool ret = false;
1578 unsigned long flags;
1579
1580 /* read the comment in __queue_work() */
1581 local_irq_save(flags);
1582
1583 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1584 __queue_delayed_work(cpu, wq, dwork, delay);
1585 ret = true;
1586 }
1587
1588 local_irq_restore(flags);
1589 return ret;
1590}
1591EXPORT_SYMBOL(queue_delayed_work_on);
1592
1593/**
1594 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1595 * @cpu: CPU number to execute work on
1596 * @wq: workqueue to use
1597 * @dwork: work to queue
1598 * @delay: number of jiffies to wait before queueing
1599 *
1600 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1601 * modify @dwork's timer so that it expires after @delay. If @delay is
1602 * zero, @work is guaranteed to be scheduled immediately regardless of its
1603 * current state.
1604 *
1605 * Return: %false if @dwork was idle and queued, %true if @dwork was
1606 * pending and its timer was modified.
1607 *
1608 * This function is safe to call from any context including IRQ handler.
1609 * See try_to_grab_pending() for details.
1610 */
1611bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1612 struct delayed_work *dwork, unsigned long delay)
1613{
1614 unsigned long flags;
1615 int ret;
1616
1617 do {
1618 ret = try_to_grab_pending(&dwork->work, true, &flags);
1619 } while (unlikely(ret == -EAGAIN));
1620
1621 if (likely(ret >= 0)) {
1622 __queue_delayed_work(cpu, wq, dwork, delay);
1623 local_irq_restore(flags);
1624 }
1625
1626 /* -ENOENT from try_to_grab_pending() becomes %true */
1627 return ret;
1628}
1629EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1630
1631/**
1632 * worker_enter_idle - enter idle state
1633 * @worker: worker which is entering idle state
1634 *
1635 * @worker is entering idle state. Update stats and idle timer if
1636 * necessary.
1637 *
1638 * LOCKING:
1639 * spin_lock_irq(pool->lock).
1640 */
1641static void worker_enter_idle(struct worker *worker)
1642{
1643 struct worker_pool *pool = worker->pool;
1644
1645 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1646 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1647 (worker->hentry.next || worker->hentry.pprev)))
1648 return;
1649
1650 /* can't use worker_set_flags(), also called from create_worker() */
1651 worker->flags |= WORKER_IDLE;
1652 pool->nr_idle++;
1653 worker->last_active = jiffies;
1654
1655 /* idle_list is LIFO */
1656 list_add(&worker->entry, &pool->idle_list);
1657
1658 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1659 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1660
1661 /*
1662 * Sanity check nr_running. Because wq_unbind_fn() releases
1663 * pool->lock between setting %WORKER_UNBOUND and zapping
1664 * nr_running, the warning may trigger spuriously. Check iff
1665 * unbind is not in progress.
1666 */
1667 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1668 pool->nr_workers == pool->nr_idle &&
1669 atomic_read(&pool->nr_running));
1670}
1671
1672/**
1673 * worker_leave_idle - leave idle state
1674 * @worker: worker which is leaving idle state
1675 *
1676 * @worker is leaving idle state. Update stats.
1677 *
1678 * LOCKING:
1679 * spin_lock_irq(pool->lock).
1680 */
1681static void worker_leave_idle(struct worker *worker)
1682{
1683 struct worker_pool *pool = worker->pool;
1684
1685 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1686 return;
1687 worker_clr_flags(worker, WORKER_IDLE);
1688 pool->nr_idle--;
1689 list_del_init(&worker->entry);
1690}
1691
1692static struct worker *alloc_worker(int node)
1693{
1694 struct worker *worker;
1695
1696 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1697 if (worker) {
1698 INIT_LIST_HEAD(&worker->entry);
1699 INIT_LIST_HEAD(&worker->scheduled);
1700 INIT_LIST_HEAD(&worker->node);
1701 /* on creation a worker is in !idle && prep state */
1702 worker->flags = WORKER_PREP;
1703 }
1704 return worker;
1705}
1706
1707/**
1708 * worker_attach_to_pool() - attach a worker to a pool
1709 * @worker: worker to be attached
1710 * @pool: the target pool
1711 *
1712 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
1713 * cpu-binding of @worker are kept coordinated with the pool across
1714 * cpu-[un]hotplugs.
1715 */
1716static void worker_attach_to_pool(struct worker *worker,
1717 struct worker_pool *pool)
1718{
1719 mutex_lock(&pool->attach_mutex);
1720
1721 /*
1722 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723 * online CPUs. It'll be re-applied when any of the CPUs come up.
1724 */
1725 set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1726
1727 /*
1728 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1729 * stable across this function. See the comments above the
1730 * flag definition for details.
1731 */
1732 if (pool->flags & POOL_DISASSOCIATED)
1733 worker->flags |= WORKER_UNBOUND;
1734
1735 list_add_tail(&worker->node, &pool->workers);
1736
1737 mutex_unlock(&pool->attach_mutex);
1738}
1739
1740/**
1741 * worker_detach_from_pool() - detach a worker from its pool
1742 * @worker: worker which is attached to its pool
1743 * @pool: the pool @worker is attached to
1744 *
1745 * Undo the attaching which had been done in worker_attach_to_pool(). The
1746 * caller worker shouldn't access to the pool after detached except it has
1747 * other reference to the pool.
1748 */
1749static void worker_detach_from_pool(struct worker *worker,
1750 struct worker_pool *pool)
1751{
1752 struct completion *detach_completion = NULL;
1753
1754 mutex_lock(&pool->attach_mutex);
1755 list_del(&worker->node);
1756 if (list_empty(&pool->workers))
1757 detach_completion = pool->detach_completion;
1758 mutex_unlock(&pool->attach_mutex);
1759
1760 /* clear leftover flags without pool->lock after it is detached */
1761 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1762
1763 if (detach_completion)
1764 complete(detach_completion);
1765}
1766
1767/**
1768 * create_worker - create a new workqueue worker
1769 * @pool: pool the new worker will belong to
1770 *
1771 * Create and start a new worker which is attached to @pool.
1772 *
1773 * CONTEXT:
1774 * Might sleep. Does GFP_KERNEL allocations.
1775 *
1776 * Return:
1777 * Pointer to the newly created worker.
1778 */
1779static struct worker *create_worker(struct worker_pool *pool)
1780{
1781 struct worker *worker = NULL;
1782 int id = -1;
1783 char id_buf[16];
1784
1785 /* ID is needed to determine kthread name */
1786 id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1787 if (id < 0)
1788 goto fail;
1789
1790 worker = alloc_worker(pool->node);
1791 if (!worker)
1792 goto fail;
1793
1794 worker->pool = pool;
1795 worker->id = id;
1796
1797 if (pool->cpu >= 0)
1798 snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1799 pool->attrs->nice < 0 ? "H" : "");
1800 else
1801 snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1802
1803 worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1804 "kworker/%s", id_buf);
1805 if (IS_ERR(worker->task))
1806 goto fail;
1807
1808 set_user_nice(worker->task, pool->attrs->nice);
1809 kthread_bind_mask(worker->task, pool->attrs->cpumask);
1810
1811 /* successful, attach the worker to the pool */
1812 worker_attach_to_pool(worker, pool);
1813
1814 /* start the newly created worker */
1815 spin_lock_irq(&pool->lock);
1816 worker->pool->nr_workers++;
1817 worker_enter_idle(worker);
1818 wake_up_process(worker->task);
1819 spin_unlock_irq(&pool->lock);
1820
1821 return worker;
1822
1823fail:
1824 if (id >= 0)
1825 ida_simple_remove(&pool->worker_ida, id);
1826 kfree(worker);
1827 return NULL;
1828}
1829
1830/**
1831 * destroy_worker - destroy a workqueue worker
1832 * @worker: worker to be destroyed
1833 *
1834 * Destroy @worker and adjust @pool stats accordingly. The worker should
1835 * be idle.
1836 *
1837 * CONTEXT:
1838 * spin_lock_irq(pool->lock).
1839 */
1840static void destroy_worker(struct worker *worker)
1841{
1842 struct worker_pool *pool = worker->pool;
1843
1844 lockdep_assert_held(&pool->lock);
1845
1846 /* sanity check frenzy */
1847 if (WARN_ON(worker->current_work) ||
1848 WARN_ON(!list_empty(&worker->scheduled)) ||
1849 WARN_ON(!(worker->flags & WORKER_IDLE)))
1850 return;
1851
1852 pool->nr_workers--;
1853 pool->nr_idle--;
1854
1855 list_del_init(&worker->entry);
1856 worker->flags |= WORKER_DIE;
1857 wake_up_process(worker->task);
1858}
1859
1860static void idle_worker_timeout(unsigned long __pool)
1861{
1862 struct worker_pool *pool = (void *)__pool;
1863
1864 spin_lock_irq(&pool->lock);
1865
1866 while (too_many_workers(pool)) {
1867 struct worker *worker;
1868 unsigned long expires;
1869
1870 /* idle_list is kept in LIFO order, check the last one */
1871 worker = list_entry(pool->idle_list.prev, struct worker, entry);
1872 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873
1874 if (time_before(jiffies, expires)) {
1875 mod_timer(&pool->idle_timer, expires);
1876 break;
1877 }
1878
1879 destroy_worker(worker);
1880 }
1881
1882 spin_unlock_irq(&pool->lock);
1883}
1884
1885static void send_mayday(struct work_struct *work)
1886{
1887 struct pool_workqueue *pwq = get_work_pwq(work);
1888 struct workqueue_struct *wq = pwq->wq;
1889
1890 lockdep_assert_held(&wq_mayday_lock);
1891
1892 if (!wq->rescuer)
1893 return;
1894
1895 /* mayday mayday mayday */
1896 if (list_empty(&pwq->mayday_node)) {
1897 /*
1898 * If @pwq is for an unbound wq, its base ref may be put at
1899 * any time due to an attribute change. Pin @pwq until the
1900 * rescuer is done with it.
1901 */
1902 get_pwq(pwq);
1903 list_add_tail(&pwq->mayday_node, &wq->maydays);
1904 wake_up_process(wq->rescuer->task);
1905 }
1906}
1907
1908static void pool_mayday_timeout(unsigned long __pool)
1909{
1910 struct worker_pool *pool = (void *)__pool;
1911 struct work_struct *work;
1912
1913 spin_lock_irq(&pool->lock);
1914 spin_lock(&wq_mayday_lock); /* for wq->maydays */
1915
1916 if (need_to_create_worker(pool)) {
1917 /*
1918 * We've been trying to create a new worker but
1919 * haven't been successful. We might be hitting an
1920 * allocation deadlock. Send distress signals to
1921 * rescuers.
1922 */
1923 list_for_each_entry(work, &pool->worklist, entry)
1924 send_mayday(work);
1925 }
1926
1927 spin_unlock(&wq_mayday_lock);
1928 spin_unlock_irq(&pool->lock);
1929
1930 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1931}
1932
1933/**
1934 * maybe_create_worker - create a new worker if necessary
1935 * @pool: pool to create a new worker for
1936 *
1937 * Create a new worker for @pool if necessary. @pool is guaranteed to
1938 * have at least one idle worker on return from this function. If
1939 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1940 * sent to all rescuers with works scheduled on @pool to resolve
1941 * possible allocation deadlock.
1942 *
1943 * On return, need_to_create_worker() is guaranteed to be %false and
1944 * may_start_working() %true.
1945 *
1946 * LOCKING:
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times. Does GFP_KERNEL allocations. Called only from
1949 * manager.
1950 */
1951static void maybe_create_worker(struct worker_pool *pool)
1952__releases(&pool->lock)
1953__acquires(&pool->lock)
1954{
1955restart:
1956 spin_unlock_irq(&pool->lock);
1957
1958 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1959 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1960
1961 while (true) {
1962 if (create_worker(pool) || !need_to_create_worker(pool))
1963 break;
1964
1965 schedule_timeout_interruptible(CREATE_COOLDOWN);
1966
1967 if (!need_to_create_worker(pool))
1968 break;
1969 }
1970
1971 del_timer_sync(&pool->mayday_timer);
1972 spin_lock_irq(&pool->lock);
1973 /*
1974 * This is necessary even after a new worker was just successfully
1975 * created as @pool->lock was dropped and the new worker might have
1976 * already become busy.
1977 */
1978 if (need_to_create_worker(pool))
1979 goto restart;
1980}
1981
1982/**
1983 * manage_workers - manage worker pool
1984 * @worker: self
1985 *
1986 * Assume the manager role and manage the worker pool @worker belongs
1987 * to. At any given time, there can be only zero or one manager per
1988 * pool. The exclusion is handled automatically by this function.
1989 *
1990 * The caller can safely start processing works on false return. On
1991 * true return, it's guaranteed that need_to_create_worker() is false
1992 * and may_start_working() is true.
1993 *
1994 * CONTEXT:
1995 * spin_lock_irq(pool->lock) which may be released and regrabbed
1996 * multiple times. Does GFP_KERNEL allocations.
1997 *
1998 * Return:
1999 * %false if the pool doesn't need management and the caller can safely
2000 * start processing works, %true if management function was performed and
2001 * the conditions that the caller verified before calling the function may
2002 * no longer be true.
2003 */
2004static bool manage_workers(struct worker *worker)
2005{
2006 struct worker_pool *pool = worker->pool;
2007
2008 /*
2009 * Anyone who successfully grabs manager_arb wins the arbitration
2010 * and becomes the manager. mutex_trylock() on pool->manager_arb
2011 * failure while holding pool->lock reliably indicates that someone
2012 * else is managing the pool and the worker which failed trylock
2013 * can proceed to executing work items. This means that anyone
2014 * grabbing manager_arb is responsible for actually performing
2015 * manager duties. If manager_arb is grabbed and released without
2016 * actual management, the pool may stall indefinitely.
2017 */
2018 if (!mutex_trylock(&pool->manager_arb))
2019 return false;
2020 pool->manager = worker;
2021
2022 maybe_create_worker(pool);
2023
2024 pool->manager = NULL;
2025 mutex_unlock(&pool->manager_arb);
2026 return true;
2027}
2028
2029/**
2030 * process_one_work - process single work
2031 * @worker: self
2032 * @work: work to process
2033 *
2034 * Process @work. This function contains all the logics necessary to
2035 * process a single work including synchronization against and
2036 * interaction with other workers on the same cpu, queueing and
2037 * flushing. As long as context requirement is met, any worker can
2038 * call this function to process a work.
2039 *
2040 * CONTEXT:
2041 * spin_lock_irq(pool->lock) which is released and regrabbed.
2042 */
2043static void process_one_work(struct worker *worker, struct work_struct *work)
2044__releases(&pool->lock)
2045__acquires(&pool->lock)
2046{
2047 struct pool_workqueue *pwq = get_work_pwq(work);
2048 struct worker_pool *pool = worker->pool;
2049 bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2050 int work_color;
2051 struct worker *collision;
2052#ifdef CONFIG_LOCKDEP
2053 /*
2054 * It is permissible to free the struct work_struct from
2055 * inside the function that is called from it, this we need to
2056 * take into account for lockdep too. To avoid bogus "held
2057 * lock freed" warnings as well as problems when looking into
2058 * work->lockdep_map, make a copy and use that here.
2059 */
2060 struct lockdep_map lockdep_map;
2061
2062 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2063#endif
2064 /* ensure we're on the correct CPU */
2065 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2066 raw_smp_processor_id() != pool->cpu);
2067
2068 /*
2069 * A single work shouldn't be executed concurrently by
2070 * multiple workers on a single cpu. Check whether anyone is
2071 * already processing the work. If so, defer the work to the
2072 * currently executing one.
2073 */
2074 collision = find_worker_executing_work(pool, work);
2075 if (unlikely(collision)) {
2076 move_linked_works(work, &collision->scheduled, NULL);
2077 return;
2078 }
2079
2080 /* claim and dequeue */
2081 debug_work_deactivate(work);
2082 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2083 worker->current_work = work;
2084 worker->current_func = work->func;
2085 worker->current_pwq = pwq;
2086 work_color = get_work_color(work);
2087
2088 list_del_init(&work->entry);
2089
2090 /*
2091 * CPU intensive works don't participate in concurrency management.
2092 * They're the scheduler's responsibility. This takes @worker out
2093 * of concurrency management and the next code block will chain
2094 * execution of the pending work items.
2095 */
2096 if (unlikely(cpu_intensive))
2097 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2098
2099 /*
2100 * Wake up another worker if necessary. The condition is always
2101 * false for normal per-cpu workers since nr_running would always
2102 * be >= 1 at this point. This is used to chain execution of the
2103 * pending work items for WORKER_NOT_RUNNING workers such as the
2104 * UNBOUND and CPU_INTENSIVE ones.
2105 */
2106 if (need_more_worker(pool))
2107 wake_up_worker(pool);
2108
2109 /*
2110 * Record the last pool and clear PENDING which should be the last
2111 * update to @work. Also, do this inside @pool->lock so that
2112 * PENDING and queued state changes happen together while IRQ is
2113 * disabled.
2114 */
2115 set_work_pool_and_clear_pending(work, pool->id);
2116
2117 spin_unlock_irq(&pool->lock);
2118
2119 lock_map_acquire_read(&pwq->wq->lockdep_map);
2120 lock_map_acquire(&lockdep_map);
2121 trace_workqueue_execute_start(work);
2122 worker->current_func(work);
2123 /*
2124 * While we must be careful to not use "work" after this, the trace
2125 * point will only record its address.
2126 */
2127 trace_workqueue_execute_end(work);
2128 lock_map_release(&lockdep_map);
2129 lock_map_release(&pwq->wq->lockdep_map);
2130
2131 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2132 pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2133 " last function: %pf\n",
2134 current->comm, preempt_count(), task_pid_nr(current),
2135 worker->current_func);
2136 debug_show_held_locks(current);
2137 dump_stack();
2138 }
2139
2140 /*
2141 * The following prevents a kworker from hogging CPU on !PREEMPT
2142 * kernels, where a requeueing work item waiting for something to
2143 * happen could deadlock with stop_machine as such work item could
2144 * indefinitely requeue itself while all other CPUs are trapped in
2145 * stop_machine. At the same time, report a quiescent RCU state so
2146 * the same condition doesn't freeze RCU.
2147 */
2148 cond_resched_rcu_qs();
2149
2150 spin_lock_irq(&pool->lock);
2151
2152 /* clear cpu intensive status */
2153 if (unlikely(cpu_intensive))
2154 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2155
2156 /* we're done with it, release */
2157 hash_del(&worker->hentry);
2158 worker->current_work = NULL;
2159 worker->current_func = NULL;
2160 worker->current_pwq = NULL;
2161 worker->desc_valid = false;
2162 pwq_dec_nr_in_flight(pwq, work_color);
2163}
2164
2165/**
2166 * process_scheduled_works - process scheduled works
2167 * @worker: self
2168 *
2169 * Process all scheduled works. Please note that the scheduled list
2170 * may change while processing a work, so this function repeatedly
2171 * fetches a work from the top and executes it.
2172 *
2173 * CONTEXT:
2174 * spin_lock_irq(pool->lock) which may be released and regrabbed
2175 * multiple times.
2176 */
2177static void process_scheduled_works(struct worker *worker)
2178{
2179 while (!list_empty(&worker->scheduled)) {
2180 struct work_struct *work = list_first_entry(&worker->scheduled,
2181 struct work_struct, entry);
2182 process_one_work(worker, work);
2183 }
2184}
2185
2186/**
2187 * worker_thread - the worker thread function
2188 * @__worker: self
2189 *
2190 * The worker thread function. All workers belong to a worker_pool -
2191 * either a per-cpu one or dynamic unbound one. These workers process all
2192 * work items regardless of their specific target workqueue. The only
2193 * exception is work items which belong to workqueues with a rescuer which
2194 * will be explained in rescuer_thread().
2195 *
2196 * Return: 0
2197 */
2198static int worker_thread(void *__worker)
2199{
2200 struct worker *worker = __worker;
2201 struct worker_pool *pool = worker->pool;
2202
2203 /* tell the scheduler that this is a workqueue worker */
2204 worker->task->flags |= PF_WQ_WORKER;
2205woke_up:
2206 spin_lock_irq(&pool->lock);
2207
2208 /* am I supposed to die? */
2209 if (unlikely(worker->flags & WORKER_DIE)) {
2210 spin_unlock_irq(&pool->lock);
2211 WARN_ON_ONCE(!list_empty(&worker->entry));
2212 worker->task->flags &= ~PF_WQ_WORKER;
2213
2214 set_task_comm(worker->task, "kworker/dying");
2215 ida_simple_remove(&pool->worker_ida, worker->id);
2216 worker_detach_from_pool(worker, pool);
2217 kfree(worker);
2218 return 0;
2219 }
2220
2221 worker_leave_idle(worker);
2222recheck:
2223 /* no more worker necessary? */
2224 if (!need_more_worker(pool))
2225 goto sleep;
2226
2227 /* do we need to manage? */
2228 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2229 goto recheck;
2230
2231 /*
2232 * ->scheduled list can only be filled while a worker is
2233 * preparing to process a work or actually processing it.
2234 * Make sure nobody diddled with it while I was sleeping.
2235 */
2236 WARN_ON_ONCE(!list_empty(&worker->scheduled));
2237
2238 /*
2239 * Finish PREP stage. We're guaranteed to have at least one idle
2240 * worker or that someone else has already assumed the manager
2241 * role. This is where @worker starts participating in concurrency
2242 * management if applicable and concurrency management is restored
2243 * after being rebound. See rebind_workers() for details.
2244 */
2245 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2246
2247 do {
2248 struct work_struct *work =
2249 list_first_entry(&pool->worklist,
2250 struct work_struct, entry);
2251
2252 pool->watchdog_ts = jiffies;
2253
2254 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2255 /* optimization path, not strictly necessary */
2256 process_one_work(worker, work);
2257 if (unlikely(!list_empty(&worker->scheduled)))
2258 process_scheduled_works(worker);
2259 } else {
2260 move_linked_works(work, &worker->scheduled, NULL);
2261 process_scheduled_works(worker);
2262 }
2263 } while (keep_working(pool));
2264
2265 worker_set_flags(worker, WORKER_PREP);
2266sleep:
2267 /*
2268 * pool->lock is held and there's no work to process and no need to
2269 * manage, sleep. Workers are woken up only while holding
2270 * pool->lock or from local cpu, so setting the current state
2271 * before releasing pool->lock is enough to prevent losing any
2272 * event.
2273 */
2274 worker_enter_idle(worker);
2275 __set_current_state(TASK_INTERRUPTIBLE);
2276 spin_unlock_irq(&pool->lock);
2277 schedule();
2278 goto woke_up;
2279}
2280
2281/**
2282 * rescuer_thread - the rescuer thread function
2283 * @__rescuer: self
2284 *
2285 * Workqueue rescuer thread function. There's one rescuer for each
2286 * workqueue which has WQ_MEM_RECLAIM set.
2287 *
2288 * Regular work processing on a pool may block trying to create a new
2289 * worker which uses GFP_KERNEL allocation which has slight chance of
2290 * developing into deadlock if some works currently on the same queue
2291 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2292 * the problem rescuer solves.
2293 *
2294 * When such condition is possible, the pool summons rescuers of all
2295 * workqueues which have works queued on the pool and let them process
2296 * those works so that forward progress can be guaranteed.
2297 *
2298 * This should happen rarely.
2299 *
2300 * Return: 0
2301 */
2302static int rescuer_thread(void *__rescuer)
2303{
2304 struct worker *rescuer = __rescuer;
2305 struct workqueue_struct *wq = rescuer->rescue_wq;
2306 struct list_head *scheduled = &rescuer->scheduled;
2307 bool should_stop;
2308
2309 set_user_nice(current, RESCUER_NICE_LEVEL);
2310
2311 /*
2312 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2313 * doesn't participate in concurrency management.
2314 */
2315 rescuer->task->flags |= PF_WQ_WORKER;
2316repeat:
2317 set_current_state(TASK_INTERRUPTIBLE);
2318
2319 /*
2320 * By the time the rescuer is requested to stop, the workqueue
2321 * shouldn't have any work pending, but @wq->maydays may still have
2322 * pwq(s) queued. This can happen by non-rescuer workers consuming
2323 * all the work items before the rescuer got to them. Go through
2324 * @wq->maydays processing before acting on should_stop so that the
2325 * list is always empty on exit.
2326 */
2327 should_stop = kthread_should_stop();
2328
2329 /* see whether any pwq is asking for help */
2330 spin_lock_irq(&wq_mayday_lock);
2331
2332 while (!list_empty(&wq->maydays)) {
2333 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2334 struct pool_workqueue, mayday_node);
2335 struct worker_pool *pool = pwq->pool;
2336 struct work_struct *work, *n;
2337 bool first = true;
2338
2339 __set_current_state(TASK_RUNNING);
2340 list_del_init(&pwq->mayday_node);
2341
2342 spin_unlock_irq(&wq_mayday_lock);
2343
2344 worker_attach_to_pool(rescuer, pool);
2345
2346 spin_lock_irq(&pool->lock);
2347 rescuer->pool = pool;
2348
2349 /*
2350 * Slurp in all works issued via this workqueue and
2351 * process'em.
2352 */
2353 WARN_ON_ONCE(!list_empty(scheduled));
2354 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2355 if (get_work_pwq(work) == pwq) {
2356 if (first)
2357 pool->watchdog_ts = jiffies;
2358 move_linked_works(work, scheduled, &n);
2359 }
2360 first = false;
2361 }
2362
2363 if (!list_empty(scheduled)) {
2364 process_scheduled_works(rescuer);
2365
2366 /*
2367 * The above execution of rescued work items could
2368 * have created more to rescue through
2369 * pwq_activate_first_delayed() or chained
2370 * queueing. Let's put @pwq back on mayday list so
2371 * that such back-to-back work items, which may be
2372 * being used to relieve memory pressure, don't
2373 * incur MAYDAY_INTERVAL delay inbetween.
2374 */
2375 if (need_to_create_worker(pool)) {
2376 spin_lock(&wq_mayday_lock);
2377 get_pwq(pwq);
2378 list_move_tail(&pwq->mayday_node, &wq->maydays);
2379 spin_unlock(&wq_mayday_lock);
2380 }
2381 }
2382
2383 /*
2384 * Put the reference grabbed by send_mayday(). @pool won't
2385 * go away while we're still attached to it.
2386 */
2387 put_pwq(pwq);
2388
2389 /*
2390 * Leave this pool. If need_more_worker() is %true, notify a
2391 * regular worker; otherwise, we end up with 0 concurrency
2392 * and stalling the execution.
2393 */
2394 if (need_more_worker(pool))
2395 wake_up_worker(pool);
2396
2397 rescuer->pool = NULL;
2398 spin_unlock_irq(&pool->lock);
2399
2400 worker_detach_from_pool(rescuer, pool);
2401
2402 spin_lock_irq(&wq_mayday_lock);
2403 }
2404
2405 spin_unlock_irq(&wq_mayday_lock);
2406
2407 if (should_stop) {
2408 __set_current_state(TASK_RUNNING);
2409 rescuer->task->flags &= ~PF_WQ_WORKER;
2410 return 0;
2411 }
2412
2413 /* rescuers should never participate in concurrency management */
2414 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2415 schedule();
2416 goto repeat;
2417}
2418
2419/**
2420 * check_flush_dependency - check for flush dependency sanity
2421 * @target_wq: workqueue being flushed
2422 * @target_work: work item being flushed (NULL for workqueue flushes)
2423 *
2424 * %current is trying to flush the whole @target_wq or @target_work on it.
2425 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2426 * reclaiming memory or running on a workqueue which doesn't have
2427 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2428 * a deadlock.
2429 */
2430static void check_flush_dependency(struct workqueue_struct *target_wq,
2431 struct work_struct *target_work)
2432{
2433 work_func_t target_func = target_work ? target_work->func : NULL;
2434 struct worker *worker;
2435
2436 if (target_wq->flags & WQ_MEM_RECLAIM)
2437 return;
2438
2439 worker = current_wq_worker();
2440
2441 WARN_ONCE(current->flags & PF_MEMALLOC,
2442 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2443 current->pid, current->comm, target_wq->name, target_func);
2444 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2445 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2446 "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2447 worker->current_pwq->wq->name, worker->current_func,
2448 target_wq->name, target_func);
2449}
2450
2451struct wq_barrier {
2452 struct work_struct work;
2453 struct completion done;
2454 struct task_struct *task; /* purely informational */
2455};
2456
2457static void wq_barrier_func(struct work_struct *work)
2458{
2459 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2460 complete(&barr->done);
2461}
2462
2463/**
2464 * insert_wq_barrier - insert a barrier work
2465 * @pwq: pwq to insert barrier into
2466 * @barr: wq_barrier to insert
2467 * @target: target work to attach @barr to
2468 * @worker: worker currently executing @target, NULL if @target is not executing
2469 *
2470 * @barr is linked to @target such that @barr is completed only after
2471 * @target finishes execution. Please note that the ordering
2472 * guarantee is observed only with respect to @target and on the local
2473 * cpu.
2474 *
2475 * Currently, a queued barrier can't be canceled. This is because
2476 * try_to_grab_pending() can't determine whether the work to be
2477 * grabbed is at the head of the queue and thus can't clear LINKED
2478 * flag of the previous work while there must be a valid next work
2479 * after a work with LINKED flag set.
2480 *
2481 * Note that when @worker is non-NULL, @target may be modified
2482 * underneath us, so we can't reliably determine pwq from @target.
2483 *
2484 * CONTEXT:
2485 * spin_lock_irq(pool->lock).
2486 */
2487static void insert_wq_barrier(struct pool_workqueue *pwq,
2488 struct wq_barrier *barr,
2489 struct work_struct *target, struct worker *worker)
2490{
2491 struct list_head *head;
2492 unsigned int linked = 0;
2493
2494 /*
2495 * debugobject calls are safe here even with pool->lock locked
2496 * as we know for sure that this will not trigger any of the
2497 * checks and call back into the fixup functions where we
2498 * might deadlock.
2499 */
2500 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2501 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2502 init_completion(&barr->done);
2503 barr->task = current;
2504
2505 /*
2506 * If @target is currently being executed, schedule the
2507 * barrier to the worker; otherwise, put it after @target.
2508 */
2509 if (worker)
2510 head = worker->scheduled.next;
2511 else {
2512 unsigned long *bits = work_data_bits(target);
2513
2514 head = target->entry.next;
2515 /* there can already be other linked works, inherit and set */
2516 linked = *bits & WORK_STRUCT_LINKED;
2517 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2518 }
2519
2520 debug_work_activate(&barr->work);
2521 insert_work(pwq, &barr->work, head,
2522 work_color_to_flags(WORK_NO_COLOR) | linked);
2523}
2524
2525/**
2526 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2527 * @wq: workqueue being flushed
2528 * @flush_color: new flush color, < 0 for no-op
2529 * @work_color: new work color, < 0 for no-op
2530 *
2531 * Prepare pwqs for workqueue flushing.
2532 *
2533 * If @flush_color is non-negative, flush_color on all pwqs should be
2534 * -1. If no pwq has in-flight commands at the specified color, all
2535 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2536 * has in flight commands, its pwq->flush_color is set to
2537 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2538 * wakeup logic is armed and %true is returned.
2539 *
2540 * The caller should have initialized @wq->first_flusher prior to
2541 * calling this function with non-negative @flush_color. If
2542 * @flush_color is negative, no flush color update is done and %false
2543 * is returned.
2544 *
2545 * If @work_color is non-negative, all pwqs should have the same
2546 * work_color which is previous to @work_color and all will be
2547 * advanced to @work_color.
2548 *
2549 * CONTEXT:
2550 * mutex_lock(wq->mutex).
2551 *
2552 * Return:
2553 * %true if @flush_color >= 0 and there's something to flush. %false
2554 * otherwise.
2555 */
2556static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2557 int flush_color, int work_color)
2558{
2559 bool wait = false;
2560 struct pool_workqueue *pwq;
2561
2562 if (flush_color >= 0) {
2563 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2564 atomic_set(&wq->nr_pwqs_to_flush, 1);
2565 }
2566
2567 for_each_pwq(pwq, wq) {
2568 struct worker_pool *pool = pwq->pool;
2569
2570 spin_lock_irq(&pool->lock);
2571
2572 if (flush_color >= 0) {
2573 WARN_ON_ONCE(pwq->flush_color != -1);
2574
2575 if (pwq->nr_in_flight[flush_color]) {
2576 pwq->flush_color = flush_color;
2577 atomic_inc(&wq->nr_pwqs_to_flush);
2578 wait = true;
2579 }
2580 }
2581
2582 if (work_color >= 0) {
2583 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2584 pwq->work_color = work_color;
2585 }
2586
2587 spin_unlock_irq(&pool->lock);
2588 }
2589
2590 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2591 complete(&wq->first_flusher->done);
2592
2593 return wait;
2594}
2595
2596/**
2597 * flush_workqueue - ensure that any scheduled work has run to completion.
2598 * @wq: workqueue to flush
2599 *
2600 * This function sleeps until all work items which were queued on entry
2601 * have finished execution, but it is not livelocked by new incoming ones.
2602 */
2603void flush_workqueue(struct workqueue_struct *wq)
2604{
2605 struct wq_flusher this_flusher = {
2606 .list = LIST_HEAD_INIT(this_flusher.list),
2607 .flush_color = -1,
2608 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2609 };
2610 int next_color;
2611
2612 lock_map_acquire(&wq->lockdep_map);
2613 lock_map_release(&wq->lockdep_map);
2614
2615 mutex_lock(&wq->mutex);
2616
2617 /*
2618 * Start-to-wait phase
2619 */
2620 next_color = work_next_color(wq->work_color);
2621
2622 if (next_color != wq->flush_color) {
2623 /*
2624 * Color space is not full. The current work_color
2625 * becomes our flush_color and work_color is advanced
2626 * by one.
2627 */
2628 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2629 this_flusher.flush_color = wq->work_color;
2630 wq->work_color = next_color;
2631
2632 if (!wq->first_flusher) {
2633 /* no flush in progress, become the first flusher */
2634 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2635
2636 wq->first_flusher = &this_flusher;
2637
2638 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2639 wq->work_color)) {
2640 /* nothing to flush, done */
2641 wq->flush_color = next_color;
2642 wq->first_flusher = NULL;
2643 goto out_unlock;
2644 }
2645 } else {
2646 /* wait in queue */
2647 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2648 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2649 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2650 }
2651 } else {
2652 /*
2653 * Oops, color space is full, wait on overflow queue.
2654 * The next flush completion will assign us
2655 * flush_color and transfer to flusher_queue.
2656 */
2657 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2658 }
2659
2660 check_flush_dependency(wq, NULL);
2661
2662 mutex_unlock(&wq->mutex);
2663
2664 wait_for_completion(&this_flusher.done);
2665
2666 /*
2667 * Wake-up-and-cascade phase
2668 *
2669 * First flushers are responsible for cascading flushes and
2670 * handling overflow. Non-first flushers can simply return.
2671 */
2672 if (wq->first_flusher != &this_flusher)
2673 return;
2674
2675 mutex_lock(&wq->mutex);
2676
2677 /* we might have raced, check again with mutex held */
2678 if (wq->first_flusher != &this_flusher)
2679 goto out_unlock;
2680
2681 wq->first_flusher = NULL;
2682
2683 WARN_ON_ONCE(!list_empty(&this_flusher.list));
2684 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2685
2686 while (true) {
2687 struct wq_flusher *next, *tmp;
2688
2689 /* complete all the flushers sharing the current flush color */
2690 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2691 if (next->flush_color != wq->flush_color)
2692 break;
2693 list_del_init(&next->list);
2694 complete(&next->done);
2695 }
2696
2697 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2698 wq->flush_color != work_next_color(wq->work_color));
2699
2700 /* this flush_color is finished, advance by one */
2701 wq->flush_color = work_next_color(wq->flush_color);
2702
2703 /* one color has been freed, handle overflow queue */
2704 if (!list_empty(&wq->flusher_overflow)) {
2705 /*
2706 * Assign the same color to all overflowed
2707 * flushers, advance work_color and append to
2708 * flusher_queue. This is the start-to-wait
2709 * phase for these overflowed flushers.
2710 */
2711 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2712 tmp->flush_color = wq->work_color;
2713
2714 wq->work_color = work_next_color(wq->work_color);
2715
2716 list_splice_tail_init(&wq->flusher_overflow,
2717 &wq->flusher_queue);
2718 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2719 }
2720
2721 if (list_empty(&wq->flusher_queue)) {
2722 WARN_ON_ONCE(wq->flush_color != wq->work_color);
2723 break;
2724 }
2725
2726 /*
2727 * Need to flush more colors. Make the next flusher
2728 * the new first flusher and arm pwqs.
2729 */
2730 WARN_ON_ONCE(wq->flush_color == wq->work_color);
2731 WARN_ON_ONCE(wq->flush_color != next->flush_color);
2732
2733 list_del_init(&next->list);
2734 wq->first_flusher = next;
2735
2736 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2737 break;
2738
2739 /*
2740 * Meh... this color is already done, clear first
2741 * flusher and repeat cascading.
2742 */
2743 wq->first_flusher = NULL;
2744 }
2745
2746out_unlock:
2747 mutex_unlock(&wq->mutex);
2748}
2749EXPORT_SYMBOL(flush_workqueue);
2750
2751/**
2752 * drain_workqueue - drain a workqueue
2753 * @wq: workqueue to drain
2754 *
2755 * Wait until the workqueue becomes empty. While draining is in progress,
2756 * only chain queueing is allowed. IOW, only currently pending or running
2757 * work items on @wq can queue further work items on it. @wq is flushed
2758 * repeatedly until it becomes empty. The number of flushing is determined
2759 * by the depth of chaining and should be relatively short. Whine if it
2760 * takes too long.
2761 */
2762void drain_workqueue(struct workqueue_struct *wq)
2763{
2764 unsigned int flush_cnt = 0;
2765 struct pool_workqueue *pwq;
2766
2767 /*
2768 * __queue_work() needs to test whether there are drainers, is much
2769 * hotter than drain_workqueue() and already looks at @wq->flags.
2770 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2771 */
2772 mutex_lock(&wq->mutex);
2773 if (!wq->nr_drainers++)
2774 wq->flags |= __WQ_DRAINING;
2775 mutex_unlock(&wq->mutex);
2776reflush:
2777 flush_workqueue(wq);
2778
2779 mutex_lock(&wq->mutex);
2780
2781 for_each_pwq(pwq, wq) {
2782 bool drained;
2783
2784 spin_lock_irq(&pwq->pool->lock);
2785 drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2786 spin_unlock_irq(&pwq->pool->lock);
2787
2788 if (drained)
2789 continue;
2790
2791 if (++flush_cnt == 10 ||
2792 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2793 pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2794 wq->name, flush_cnt);
2795
2796 mutex_unlock(&wq->mutex);
2797 goto reflush;
2798 }
2799
2800 if (!--wq->nr_drainers)
2801 wq->flags &= ~__WQ_DRAINING;
2802 mutex_unlock(&wq->mutex);
2803}
2804EXPORT_SYMBOL_GPL(drain_workqueue);
2805
2806static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2807{
2808 struct worker *worker = NULL;
2809 struct worker_pool *pool;
2810 struct pool_workqueue *pwq;
2811
2812 might_sleep();
2813
2814 local_irq_disable();
2815 pool = get_work_pool(work);
2816 if (!pool) {
2817 local_irq_enable();
2818 return false;
2819 }
2820
2821 spin_lock(&pool->lock);
2822 /* see the comment in try_to_grab_pending() with the same code */
2823 pwq = get_work_pwq(work);
2824 if (pwq) {
2825 if (unlikely(pwq->pool != pool))
2826 goto already_gone;
2827 } else {
2828 worker = find_worker_executing_work(pool, work);
2829 if (!worker)
2830 goto already_gone;
2831 pwq = worker->current_pwq;
2832 }
2833
2834 check_flush_dependency(pwq->wq, work);
2835
2836 insert_wq_barrier(pwq, barr, work, worker);
2837 spin_unlock_irq(&pool->lock);
2838
2839 /*
2840 * If @max_active is 1 or rescuer is in use, flushing another work
2841 * item on the same workqueue may lead to deadlock. Make sure the
2842 * flusher is not running on the same workqueue by verifying write
2843 * access.
2844 */
2845 if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2846 lock_map_acquire(&pwq->wq->lockdep_map);
2847 else
2848 lock_map_acquire_read(&pwq->wq->lockdep_map);
2849 lock_map_release(&pwq->wq->lockdep_map);
2850
2851 return true;
2852already_gone:
2853 spin_unlock_irq(&pool->lock);
2854 return false;
2855}
2856
2857/**
2858 * flush_work - wait for a work to finish executing the last queueing instance
2859 * @work: the work to flush
2860 *
2861 * Wait until @work has finished execution. @work is guaranteed to be idle
2862 * on return if it hasn't been requeued since flush started.
2863 *
2864 * Return:
2865 * %true if flush_work() waited for the work to finish execution,
2866 * %false if it was already idle.
2867 */
2868bool flush_work(struct work_struct *work)
2869{
2870 struct wq_barrier barr;
2871
2872 lock_map_acquire(&work->lockdep_map);
2873 lock_map_release(&work->lockdep_map);
2874
2875 if (start_flush_work(work, &barr)) {
2876 wait_for_completion(&barr.done);
2877 destroy_work_on_stack(&barr.work);
2878 return true;
2879 } else {
2880 return false;
2881 }
2882}
2883EXPORT_SYMBOL_GPL(flush_work);
2884
2885struct cwt_wait {
2886 wait_queue_t wait;
2887 struct work_struct *work;
2888};
2889
2890static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2891{
2892 struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
2893
2894 if (cwait->work != key)
2895 return 0;
2896 return autoremove_wake_function(wait, mode, sync, key);
2897}
2898
2899static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2900{
2901 static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2902 unsigned long flags;
2903 int ret;
2904
2905 do {
2906 ret = try_to_grab_pending(work, is_dwork, &flags);
2907 /*
2908 * If someone else is already canceling, wait for it to
2909 * finish. flush_work() doesn't work for PREEMPT_NONE
2910 * because we may get scheduled between @work's completion
2911 * and the other canceling task resuming and clearing
2912 * CANCELING - flush_work() will return false immediately
2913 * as @work is no longer busy, try_to_grab_pending() will
2914 * return -ENOENT as @work is still being canceled and the
2915 * other canceling task won't be able to clear CANCELING as
2916 * we're hogging the CPU.
2917 *
2918 * Let's wait for completion using a waitqueue. As this
2919 * may lead to the thundering herd problem, use a custom
2920 * wake function which matches @work along with exclusive
2921 * wait and wakeup.
2922 */
2923 if (unlikely(ret == -ENOENT)) {
2924 struct cwt_wait cwait;
2925
2926 init_wait(&cwait.wait);
2927 cwait.wait.func = cwt_wakefn;
2928 cwait.work = work;
2929
2930 prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2931 TASK_UNINTERRUPTIBLE);
2932 if (work_is_canceling(work))
2933 schedule();
2934 finish_wait(&cancel_waitq, &cwait.wait);
2935 }
2936 } while (unlikely(ret < 0));
2937
2938 /* tell other tasks trying to grab @work to back off */
2939 mark_work_canceling(work);
2940 local_irq_restore(flags);
2941
2942 flush_work(work);
2943 clear_work_data(work);
2944
2945 /*
2946 * Paired with prepare_to_wait() above so that either
2947 * waitqueue_active() is visible here or !work_is_canceling() is
2948 * visible there.
2949 */
2950 smp_mb();
2951 if (waitqueue_active(&cancel_waitq))
2952 __wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
2953
2954 return ret;
2955}
2956
2957/**
2958 * cancel_work_sync - cancel a work and wait for it to finish
2959 * @work: the work to cancel
2960 *
2961 * Cancel @work and wait for its execution to finish. This function
2962 * can be used even if the work re-queues itself or migrates to
2963 * another workqueue. On return from this function, @work is
2964 * guaranteed to be not pending or executing on any CPU.
2965 *
2966 * cancel_work_sync(&delayed_work->work) must not be used for
2967 * delayed_work's. Use cancel_delayed_work_sync() instead.
2968 *
2969 * The caller must ensure that the workqueue on which @work was last
2970 * queued can't be destroyed before this function returns.
2971 *
2972 * Return:
2973 * %true if @work was pending, %false otherwise.
2974 */
2975bool cancel_work_sync(struct work_struct *work)
2976{
2977 return __cancel_work_timer(work, false);
2978}
2979EXPORT_SYMBOL_GPL(cancel_work_sync);
2980
2981/**
2982 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2983 * @dwork: the delayed work to flush
2984 *
2985 * Delayed timer is cancelled and the pending work is queued for
2986 * immediate execution. Like flush_work(), this function only
2987 * considers the last queueing instance of @dwork.
2988 *
2989 * Return:
2990 * %true if flush_work() waited for the work to finish execution,
2991 * %false if it was already idle.
2992 */
2993bool flush_delayed_work(struct delayed_work *dwork)
2994{
2995 local_irq_disable();
2996 if (del_timer_sync(&dwork->timer))
2997 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2998 local_irq_enable();
2999 return flush_work(&dwork->work);
3000}
3001EXPORT_SYMBOL(flush_delayed_work);
3002
3003/**
3004 * cancel_delayed_work - cancel a delayed work
3005 * @dwork: delayed_work to cancel
3006 *
3007 * Kill off a pending delayed_work.
3008 *
3009 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3010 * pending.
3011 *
3012 * Note:
3013 * The work callback function may still be running on return, unless
3014 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3015 * use cancel_delayed_work_sync() to wait on it.
3016 *
3017 * This function is safe to call from any context including IRQ handler.
3018 */
3019bool cancel_delayed_work(struct delayed_work *dwork)
3020{
3021 unsigned long flags;
3022 int ret;
3023
3024 do {
3025 ret = try_to_grab_pending(&dwork->work, true, &flags);
3026 } while (unlikely(ret == -EAGAIN));
3027
3028 if (unlikely(ret < 0))
3029 return false;
3030
3031 set_work_pool_and_clear_pending(&dwork->work,
3032 get_work_pool_id(&dwork->work));
3033 local_irq_restore(flags);
3034 return ret;
3035}
3036EXPORT_SYMBOL(cancel_delayed_work);
3037
3038/**
3039 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3040 * @dwork: the delayed work cancel
3041 *
3042 * This is cancel_work_sync() for delayed works.
3043 *
3044 * Return:
3045 * %true if @dwork was pending, %false otherwise.
3046 */
3047bool cancel_delayed_work_sync(struct delayed_work *dwork)
3048{
3049 return __cancel_work_timer(&dwork->work, true);
3050}
3051EXPORT_SYMBOL(cancel_delayed_work_sync);
3052
3053/**
3054 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3055 * @func: the function to call
3056 *
3057 * schedule_on_each_cpu() executes @func on each online CPU using the
3058 * system workqueue and blocks until all CPUs have completed.
3059 * schedule_on_each_cpu() is very slow.
3060 *
3061 * Return:
3062 * 0 on success, -errno on failure.
3063 */
3064int schedule_on_each_cpu(work_func_t func)
3065{
3066 int cpu;
3067 struct work_struct __percpu *works;
3068
3069 works = alloc_percpu(struct work_struct);
3070 if (!works)
3071 return -ENOMEM;
3072
3073 get_online_cpus();
3074
3075 for_each_online_cpu(cpu) {
3076 struct work_struct *work = per_cpu_ptr(works, cpu);
3077
3078 INIT_WORK(work, func);
3079 schedule_work_on(cpu, work);
3080 }
3081
3082 for_each_online_cpu(cpu)
3083 flush_work(per_cpu_ptr(works, cpu));
3084
3085 put_online_cpus();
3086 free_percpu(works);
3087 return 0;
3088}
3089
3090/**
3091 * execute_in_process_context - reliably execute the routine with user context
3092 * @fn: the function to execute
3093 * @ew: guaranteed storage for the execute work structure (must
3094 * be available when the work executes)
3095 *
3096 * Executes the function immediately if process context is available,
3097 * otherwise schedules the function for delayed execution.
3098 *
3099 * Return: 0 - function was executed
3100 * 1 - function was scheduled for execution
3101 */
3102int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3103{
3104 if (!in_interrupt()) {
3105 fn(&ew->work);
3106 return 0;
3107 }
3108
3109 INIT_WORK(&ew->work, fn);
3110 schedule_work(&ew->work);
3111
3112 return 1;
3113}
3114EXPORT_SYMBOL_GPL(execute_in_process_context);
3115
3116/**
3117 * free_workqueue_attrs - free a workqueue_attrs
3118 * @attrs: workqueue_attrs to free
3119 *
3120 * Undo alloc_workqueue_attrs().
3121 */
3122void free_workqueue_attrs(struct workqueue_attrs *attrs)
3123{
3124 if (attrs) {
3125 free_cpumask_var(attrs->cpumask);
3126 kfree(attrs);
3127 }
3128}
3129
3130/**
3131 * alloc_workqueue_attrs - allocate a workqueue_attrs
3132 * @gfp_mask: allocation mask to use
3133 *
3134 * Allocate a new workqueue_attrs, initialize with default settings and
3135 * return it.
3136 *
3137 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3138 */
3139struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3140{
3141 struct workqueue_attrs *attrs;
3142
3143 attrs = kzalloc(sizeof(*attrs), gfp_mask);
3144 if (!attrs)
3145 goto fail;
3146 if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3147 goto fail;
3148
3149 cpumask_copy(attrs->cpumask, cpu_possible_mask);
3150 return attrs;
3151fail:
3152 free_workqueue_attrs(attrs);
3153 return NULL;
3154}
3155
3156static void copy_workqueue_attrs(struct workqueue_attrs *to,
3157 const struct workqueue_attrs *from)
3158{
3159 to->nice = from->nice;
3160 cpumask_copy(to->cpumask, from->cpumask);
3161 /*
3162 * Unlike hash and equality test, this function doesn't ignore
3163 * ->no_numa as it is used for both pool and wq attrs. Instead,
3164 * get_unbound_pool() explicitly clears ->no_numa after copying.
3165 */
3166 to->no_numa = from->no_numa;
3167}
3168
3169/* hash value of the content of @attr */
3170static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3171{
3172 u32 hash = 0;
3173
3174 hash = jhash_1word(attrs->nice, hash);
3175 hash = jhash(cpumask_bits(attrs->cpumask),
3176 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3177 return hash;
3178}
3179
3180/* content equality test */
3181static bool wqattrs_equal(const struct workqueue_attrs *a,
3182 const struct workqueue_attrs *b)
3183{
3184 if (a->nice != b->nice)
3185 return false;
3186 if (!cpumask_equal(a->cpumask, b->cpumask))
3187 return false;
3188 return true;
3189}
3190
3191/**
3192 * init_worker_pool - initialize a newly zalloc'd worker_pool
3193 * @pool: worker_pool to initialize
3194 *
3195 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
3196 *
3197 * Return: 0 on success, -errno on failure. Even on failure, all fields
3198 * inside @pool proper are initialized and put_unbound_pool() can be called
3199 * on @pool safely to release it.
3200 */
3201static int init_worker_pool(struct worker_pool *pool)
3202{
3203 spin_lock_init(&pool->lock);
3204 pool->id = -1;
3205 pool->cpu = -1;
3206 pool->node = NUMA_NO_NODE;
3207 pool->flags |= POOL_DISASSOCIATED;
3208 pool->watchdog_ts = jiffies;
3209 INIT_LIST_HEAD(&pool->worklist);
3210 INIT_LIST_HEAD(&pool->idle_list);
3211 hash_init(pool->busy_hash);
3212
3213 init_timer_deferrable(&pool->idle_timer);
3214 pool->idle_timer.function = idle_worker_timeout;
3215 pool->idle_timer.data = (unsigned long)pool;
3216
3217 setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3218 (unsigned long)pool);
3219
3220 mutex_init(&pool->manager_arb);
3221 mutex_init(&pool->attach_mutex);
3222 INIT_LIST_HEAD(&pool->workers);
3223
3224 ida_init(&pool->worker_ida);
3225 INIT_HLIST_NODE(&pool->hash_node);
3226 pool->refcnt = 1;
3227
3228 /* shouldn't fail above this point */
3229 pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3230 if (!pool->attrs)
3231 return -ENOMEM;
3232 return 0;
3233}
3234
3235static void rcu_free_wq(struct rcu_head *rcu)
3236{
3237 struct workqueue_struct *wq =
3238 container_of(rcu, struct workqueue_struct, rcu);
3239
3240 if (!(wq->flags & WQ_UNBOUND))
3241 free_percpu(wq->cpu_pwqs);
3242 else
3243 free_workqueue_attrs(wq->unbound_attrs);
3244
3245 kfree(wq->rescuer);
3246 kfree(wq);
3247}
3248
3249static void rcu_free_pool(struct rcu_head *rcu)
3250{
3251 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3252
3253 ida_destroy(&pool->worker_ida);
3254 free_workqueue_attrs(pool->attrs);
3255 kfree(pool);
3256}
3257
3258/**
3259 * put_unbound_pool - put a worker_pool
3260 * @pool: worker_pool to put
3261 *
3262 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3263 * safe manner. get_unbound_pool() calls this function on its failure path
3264 * and this function should be able to release pools which went through,
3265 * successfully or not, init_worker_pool().
3266 *
3267 * Should be called with wq_pool_mutex held.
3268 */
3269static void put_unbound_pool(struct worker_pool *pool)
3270{
3271 DECLARE_COMPLETION_ONSTACK(detach_completion);
3272 struct worker *worker;
3273
3274 lockdep_assert_held(&wq_pool_mutex);
3275
3276 if (--pool->refcnt)
3277 return;
3278
3279 /* sanity checks */
3280 if (WARN_ON(!(pool->cpu < 0)) ||
3281 WARN_ON(!list_empty(&pool->worklist)))
3282 return;
3283
3284 /* release id and unhash */
3285 if (pool->id >= 0)
3286 idr_remove(&worker_pool_idr, pool->id);
3287 hash_del(&pool->hash_node);
3288
3289 /*
3290 * Become the manager and destroy all workers. Grabbing
3291 * manager_arb prevents @pool's workers from blocking on
3292 * attach_mutex.
3293 */
3294 mutex_lock(&pool->manager_arb);
3295
3296 spin_lock_irq(&pool->lock);
3297 while ((worker = first_idle_worker(pool)))
3298 destroy_worker(worker);
3299 WARN_ON(pool->nr_workers || pool->nr_idle);
3300 spin_unlock_irq(&pool->lock);
3301
3302 mutex_lock(&pool->attach_mutex);
3303 if (!list_empty(&pool->workers))
3304 pool->detach_completion = &detach_completion;
3305 mutex_unlock(&pool->attach_mutex);
3306
3307 if (pool->detach_completion)
3308 wait_for_completion(pool->detach_completion);
3309
3310 mutex_unlock(&pool->manager_arb);
3311
3312 /* shut down the timers */
3313 del_timer_sync(&pool->idle_timer);
3314 del_timer_sync(&pool->mayday_timer);
3315
3316 /* sched-RCU protected to allow dereferences from get_work_pool() */
3317 call_rcu_sched(&pool->rcu, rcu_free_pool);
3318}
3319
3320/**
3321 * get_unbound_pool - get a worker_pool with the specified attributes
3322 * @attrs: the attributes of the worker_pool to get
3323 *
3324 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3325 * reference count and return it. If there already is a matching
3326 * worker_pool, it will be used; otherwise, this function attempts to
3327 * create a new one.
3328 *
3329 * Should be called with wq_pool_mutex held.
3330 *
3331 * Return: On success, a worker_pool with the same attributes as @attrs.
3332 * On failure, %NULL.
3333 */
3334static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3335{
3336 u32 hash = wqattrs_hash(attrs);
3337 struct worker_pool *pool;
3338 int node;
3339 int target_node = NUMA_NO_NODE;
3340
3341 lockdep_assert_held(&wq_pool_mutex);
3342
3343 /* do we already have a matching pool? */
3344 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3345 if (wqattrs_equal(pool->attrs, attrs)) {
3346 pool->refcnt++;
3347 return pool;
3348 }
3349 }
3350
3351 /* if cpumask is contained inside a NUMA node, we belong to that node */
3352 if (wq_numa_enabled) {
3353 for_each_node(node) {
3354 if (cpumask_subset(attrs->cpumask,
3355 wq_numa_possible_cpumask[node])) {
3356 target_node = node;
3357 break;
3358 }
3359 }
3360 }
3361
3362 /* nope, create a new one */
3363 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3364 if (!pool || init_worker_pool(pool) < 0)
3365 goto fail;
3366
3367 lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3368 copy_workqueue_attrs(pool->attrs, attrs);
3369 pool->node = target_node;
3370
3371 /*
3372 * no_numa isn't a worker_pool attribute, always clear it. See
3373 * 'struct workqueue_attrs' comments for detail.
3374 */
3375 pool->attrs->no_numa = false;
3376
3377 if (worker_pool_assign_id(pool) < 0)
3378 goto fail;
3379
3380 /* create and start the initial worker */
3381 if (!create_worker(pool))
3382 goto fail;
3383
3384 /* install */
3385 hash_add(unbound_pool_hash, &pool->hash_node, hash);
3386
3387 return pool;
3388fail:
3389 if (pool)
3390 put_unbound_pool(pool);
3391 return NULL;
3392}
3393
3394static void rcu_free_pwq(struct rcu_head *rcu)
3395{
3396 kmem_cache_free(pwq_cache,
3397 container_of(rcu, struct pool_workqueue, rcu));
3398}
3399
3400/*
3401 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3402 * and needs to be destroyed.
3403 */
3404static void pwq_unbound_release_workfn(struct work_struct *work)
3405{
3406 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3407 unbound_release_work);
3408 struct workqueue_struct *wq = pwq->wq;
3409 struct worker_pool *pool = pwq->pool;
3410 bool is_last;
3411
3412 if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3413 return;
3414
3415 mutex_lock(&wq->mutex);
3416 list_del_rcu(&pwq->pwqs_node);
3417 is_last = list_empty(&wq->pwqs);
3418 mutex_unlock(&wq->mutex);
3419
3420 mutex_lock(&wq_pool_mutex);
3421 put_unbound_pool(pool);
3422 mutex_unlock(&wq_pool_mutex);
3423
3424 call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3425
3426 /*
3427 * If we're the last pwq going away, @wq is already dead and no one
3428 * is gonna access it anymore. Schedule RCU free.
3429 */
3430 if (is_last)
3431 call_rcu_sched(&wq->rcu, rcu_free_wq);
3432}
3433
3434/**
3435 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3436 * @pwq: target pool_workqueue
3437 *
3438 * If @pwq isn't freezing, set @pwq->max_active to the associated
3439 * workqueue's saved_max_active and activate delayed work items
3440 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3441 */
3442static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3443{
3444 struct workqueue_struct *wq = pwq->wq;
3445 bool freezable = wq->flags & WQ_FREEZABLE;
3446
3447 /* for @wq->saved_max_active */
3448 lockdep_assert_held(&wq->mutex);
3449
3450 /* fast exit for non-freezable wqs */
3451 if (!freezable && pwq->max_active == wq->saved_max_active)
3452 return;
3453
3454 spin_lock_irq(&pwq->pool->lock);
3455
3456 /*
3457 * During [un]freezing, the caller is responsible for ensuring that
3458 * this function is called at least once after @workqueue_freezing
3459 * is updated and visible.
3460 */
3461 if (!freezable || !workqueue_freezing) {
3462 pwq->max_active = wq->saved_max_active;
3463
3464 while (!list_empty(&pwq->delayed_works) &&
3465 pwq->nr_active < pwq->max_active)
3466 pwq_activate_first_delayed(pwq);
3467
3468 /*
3469 * Need to kick a worker after thawed or an unbound wq's
3470 * max_active is bumped. It's a slow path. Do it always.
3471 */
3472 wake_up_worker(pwq->pool);
3473 } else {
3474 pwq->max_active = 0;
3475 }
3476
3477 spin_unlock_irq(&pwq->pool->lock);
3478}
3479
3480/* initialize newly alloced @pwq which is associated with @wq and @pool */
3481static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3482 struct worker_pool *pool)
3483{
3484 BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
3486 memset(pwq, 0, sizeof(*pwq));
3487
3488 pwq->pool = pool;
3489 pwq->wq = wq;
3490 pwq->flush_color = -1;
3491 pwq->refcnt = 1;
3492 INIT_LIST_HEAD(&pwq->delayed_works);
3493 INIT_LIST_HEAD(&pwq->pwqs_node);
3494 INIT_LIST_HEAD(&pwq->mayday_node);
3495 INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3496}
3497
3498/* sync @pwq with the current state of its associated wq and link it */
3499static void link_pwq(struct pool_workqueue *pwq)
3500{
3501 struct workqueue_struct *wq = pwq->wq;
3502
3503 lockdep_assert_held(&wq->mutex);
3504
3505 /* may be called multiple times, ignore if already linked */
3506 if (!list_empty(&pwq->pwqs_node))
3507 return;
3508
3509 /* set the matching work_color */
3510 pwq->work_color = wq->work_color;
3511
3512 /* sync max_active to the current setting */
3513 pwq_adjust_max_active(pwq);
3514
3515 /* link in @pwq */
3516 list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3517}
3518
3519/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3521 const struct workqueue_attrs *attrs)
3522{
3523 struct worker_pool *pool;
3524 struct pool_workqueue *pwq;
3525
3526 lockdep_assert_held(&wq_pool_mutex);
3527
3528 pool = get_unbound_pool(attrs);
3529 if (!pool)
3530 return NULL;
3531
3532 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3533 if (!pwq) {
3534 put_unbound_pool(pool);
3535 return NULL;
3536 }
3537
3538 init_pwq(pwq, wq, pool);
3539 return pwq;
3540}
3541
3542/**
3543 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3544 * @attrs: the wq_attrs of the default pwq of the target workqueue
3545 * @node: the target NUMA node
3546 * @cpu_going_down: if >= 0, the CPU to consider as offline
3547 * @cpumask: outarg, the resulting cpumask
3548 *
3549 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3550 * @cpu_going_down is >= 0, that cpu is considered offline during
3551 * calculation. The result is stored in @cpumask.
3552 *
3553 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3554 * enabled and @node has online CPUs requested by @attrs, the returned
3555 * cpumask is the intersection of the possible CPUs of @node and
3556 * @attrs->cpumask.
3557 *
3558 * The caller is responsible for ensuring that the cpumask of @node stays
3559 * stable.
3560 *
3561 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562 * %false if equal.
3563 */
3564static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565 int cpu_going_down, cpumask_t *cpumask)
3566{
3567 if (!wq_numa_enabled || attrs->no_numa)
3568 goto use_dfl;
3569
3570 /* does @node have any online CPUs @attrs wants? */
3571 cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572 if (cpu_going_down >= 0)
3573 cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575 if (cpumask_empty(cpumask))
3576 goto use_dfl;
3577
3578 /* yeap, return possible CPUs in @node that @attrs wants */
3579 cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580 return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582use_dfl:
3583 cpumask_copy(cpumask, attrs->cpumask);
3584 return false;
3585}
3586
3587/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589 int node,
3590 struct pool_workqueue *pwq)
3591{
3592 struct pool_workqueue *old_pwq;
3593
3594 lockdep_assert_held(&wq_pool_mutex);
3595 lockdep_assert_held(&wq->mutex);
3596
3597 /* link_pwq() can handle duplicate calls */
3598 link_pwq(pwq);
3599
3600 old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601 rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602 return old_pwq;
3603}
3604
3605/* context to store the prepared attrs & pwqs before applying */
3606struct apply_wqattrs_ctx {
3607 struct workqueue_struct *wq; /* target workqueue */
3608 struct workqueue_attrs *attrs; /* attrs to apply */
3609 struct list_head list; /* queued for batching commit */
3610 struct pool_workqueue *dfl_pwq;
3611 struct pool_workqueue *pwq_tbl[];
3612};
3613
3614/* free the resources after success or abort */
3615static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616{
3617 if (ctx) {
3618 int node;
3619
3620 for_each_node(node)
3621 put_pwq_unlocked(ctx->pwq_tbl[node]);
3622 put_pwq_unlocked(ctx->dfl_pwq);
3623
3624 free_workqueue_attrs(ctx->attrs);
3625
3626 kfree(ctx);
3627 }
3628}
3629
3630/* allocate the attrs and pwqs for later installation */
3631static struct apply_wqattrs_ctx *
3632apply_wqattrs_prepare(struct workqueue_struct *wq,
3633 const struct workqueue_attrs *attrs)
3634{
3635 struct apply_wqattrs_ctx *ctx;
3636 struct workqueue_attrs *new_attrs, *tmp_attrs;
3637 int node;
3638
3639 lockdep_assert_held(&wq_pool_mutex);
3640
3641 ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642 GFP_KERNEL);
3643
3644 new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645 tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646 if (!ctx || !new_attrs || !tmp_attrs)
3647 goto out_free;
3648
3649 /*
3650 * Calculate the attrs of the default pwq.
3651 * If the user configured cpumask doesn't overlap with the
3652 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653 */
3654 copy_workqueue_attrs(new_attrs, attrs);
3655 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656 if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657 cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658
3659 /*
3660 * We may create multiple pwqs with differing cpumasks. Make a
3661 * copy of @new_attrs which will be modified and used to obtain
3662 * pools.
3663 */
3664 copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666 /*
3667 * If something goes wrong during CPU up/down, we'll fall back to
3668 * the default pwq covering whole @attrs->cpumask. Always create
3669 * it even if we don't use it immediately.
3670 */
3671 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672 if (!ctx->dfl_pwq)
3673 goto out_free;
3674
3675 for_each_node(node) {
3676 if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677 ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678 if (!ctx->pwq_tbl[node])
3679 goto out_free;
3680 } else {
3681 ctx->dfl_pwq->refcnt++;
3682 ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683 }
3684 }
3685
3686 /* save the user configured attrs and sanitize it. */
3687 copy_workqueue_attrs(new_attrs, attrs);
3688 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689 ctx->attrs = new_attrs;
3690
3691 ctx->wq = wq;
3692 free_workqueue_attrs(tmp_attrs);
3693 return ctx;
3694
3695out_free:
3696 free_workqueue_attrs(tmp_attrs);
3697 free_workqueue_attrs(new_attrs);
3698 apply_wqattrs_cleanup(ctx);
3699 return NULL;
3700}
3701
3702/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704{
3705 int node;
3706
3707 /* all pwqs have been created successfully, let's install'em */
3708 mutex_lock(&ctx->wq->mutex);
3709
3710 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712 /* save the previous pwq and install the new one */
3713 for_each_node(node)
3714 ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715 ctx->pwq_tbl[node]);
3716
3717 /* @dfl_pwq might not have been used, ensure it's linked */
3718 link_pwq(ctx->dfl_pwq);
3719 swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721 mutex_unlock(&ctx->wq->mutex);
3722}
3723
3724static void apply_wqattrs_lock(void)
3725{
3726 /* CPUs should stay stable across pwq creations and installations */
3727 get_online_cpus();
3728 mutex_lock(&wq_pool_mutex);
3729}
3730
3731static void apply_wqattrs_unlock(void)
3732{
3733 mutex_unlock(&wq_pool_mutex);
3734 put_online_cpus();
3735}
3736
3737static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738 const struct workqueue_attrs *attrs)
3739{
3740 struct apply_wqattrs_ctx *ctx;
3741
3742 /* only unbound workqueues can change attributes */
3743 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744 return -EINVAL;
3745
3746 /* creating multiple pwqs breaks ordering guarantee */
3747 if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748 return -EINVAL;
3749
3750 ctx = apply_wqattrs_prepare(wq, attrs);
3751 if (!ctx)
3752 return -ENOMEM;
3753
3754 /* the ctx has been prepared successfully, let's commit it */
3755 apply_wqattrs_commit(ctx);
3756 apply_wqattrs_cleanup(ctx);
3757
3758 return 0;
3759}
3760
3761/**
3762 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763 * @wq: the target workqueue
3764 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765 *
3766 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3767 * machines, this function maps a separate pwq to each NUMA node with
3768 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3769 * NUMA node it was issued on. Older pwqs are released as in-flight work
3770 * items finish. Note that a work item which repeatedly requeues itself
3771 * back-to-back will stay on its current pwq.
3772 *
3773 * Performs GFP_KERNEL allocations.
3774 *
3775 * Return: 0 on success and -errno on failure.
3776 */
3777int apply_workqueue_attrs(struct workqueue_struct *wq,
3778 const struct workqueue_attrs *attrs)
3779{
3780 int ret;
3781
3782 apply_wqattrs_lock();
3783 ret = apply_workqueue_attrs_locked(wq, attrs);
3784 apply_wqattrs_unlock();
3785
3786 return ret;
3787}
3788
3789/**
3790 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791 * @wq: the target workqueue
3792 * @cpu: the CPU coming up or going down
3793 * @online: whether @cpu is coming up or going down
3794 *
3795 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
3797 * @wq accordingly.
3798 *
3799 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3800 * falls back to @wq->dfl_pwq which may not be optimal but is always
3801 * correct.
3802 *
3803 * Note that when the last allowed CPU of a NUMA node goes offline for a
3804 * workqueue with a cpumask spanning multiple nodes, the workers which were
3805 * already executing the work items for the workqueue will lose their CPU
3806 * affinity and may execute on any CPU. This is similar to how per-cpu
3807 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
3808 * affinity, it's the user's responsibility to flush the work item from
3809 * CPU_DOWN_PREPARE.
3810 */
3811static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812 bool online)
3813{
3814 int node = cpu_to_node(cpu);
3815 int cpu_off = online ? -1 : cpu;
3816 struct pool_workqueue *old_pwq = NULL, *pwq;
3817 struct workqueue_attrs *target_attrs;
3818 cpumask_t *cpumask;
3819
3820 lockdep_assert_held(&wq_pool_mutex);
3821
3822 if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3823 wq->unbound_attrs->no_numa)
3824 return;
3825
3826 /*
3827 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828 * Let's use a preallocated one. The following buf is protected by
3829 * CPU hotplug exclusion.
3830 */
3831 target_attrs = wq_update_unbound_numa_attrs_buf;
3832 cpumask = target_attrs->cpumask;
3833
3834 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835 pwq = unbound_pwq_by_node(wq, node);
3836
3837 /*
3838 * Let's determine what needs to be done. If the target cpumask is
3839 * different from the default pwq's, we need to compare it to @pwq's
3840 * and create a new one if they don't match. If the target cpumask
3841 * equals the default pwq's, the default pwq should be used.
3842 */
3843 if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3844 if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3845 return;
3846 } else {
3847 goto use_dfl_pwq;
3848 }
3849
3850 /* create a new pwq */
3851 pwq = alloc_unbound_pwq(wq, target_attrs);
3852 if (!pwq) {
3853 pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854 wq->name);
3855 goto use_dfl_pwq;
3856 }
3857
3858 /* Install the new pwq. */
3859 mutex_lock(&wq->mutex);
3860 old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861 goto out_unlock;
3862
3863use_dfl_pwq:
3864 mutex_lock(&wq->mutex);
3865 spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866 get_pwq(wq->dfl_pwq);
3867 spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868 old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869out_unlock:
3870 mutex_unlock(&wq->mutex);
3871 put_pwq_unlocked(old_pwq);
3872}
3873
3874static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3875{
3876 bool highpri = wq->flags & WQ_HIGHPRI;
3877 int cpu, ret;
3878
3879 if (!(wq->flags & WQ_UNBOUND)) {
3880 wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881 if (!wq->cpu_pwqs)
3882 return -ENOMEM;
3883
3884 for_each_possible_cpu(cpu) {
3885 struct pool_workqueue *pwq =
3886 per_cpu_ptr(wq->cpu_pwqs, cpu);
3887 struct worker_pool *cpu_pools =
3888 per_cpu(cpu_worker_pools, cpu);
3889
3890 init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892 mutex_lock(&wq->mutex);
3893 link_pwq(pwq);
3894 mutex_unlock(&wq->mutex);
3895 }
3896 return 0;
3897 } else if (wq->flags & __WQ_ORDERED) {
3898 ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899 /* there should only be single pwq for ordering guarantee */
3900 WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3901 wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902 "ordering guarantee broken for workqueue %s\n", wq->name);
3903 return ret;
3904 } else {
3905 return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3906 }
3907}
3908
3909static int wq_clamp_max_active(int max_active, unsigned int flags,
3910 const char *name)
3911{
3912 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914 if (max_active < 1 || max_active > lim)
3915 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916 max_active, name, 1, lim);
3917
3918 return clamp_val(max_active, 1, lim);
3919}
3920
3921struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3922 unsigned int flags,
3923 int max_active,
3924 struct lock_class_key *key,
3925 const char *lock_name, ...)
3926{
3927 size_t tbl_size = 0;
3928 va_list args;
3929 struct workqueue_struct *wq;
3930 struct pool_workqueue *pwq;
3931
3932 /* see the comment above the definition of WQ_POWER_EFFICIENT */
3933 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934 flags |= WQ_UNBOUND;
3935
3936 /* allocate wq and format name */
3937 if (flags & WQ_UNBOUND)
3938 tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
3939
3940 wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3941 if (!wq)
3942 return NULL;
3943
3944 if (flags & WQ_UNBOUND) {
3945 wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946 if (!wq->unbound_attrs)
3947 goto err_free_wq;
3948 }
3949
3950 va_start(args, lock_name);
3951 vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3952 va_end(args);
3953
3954 max_active = max_active ?: WQ_DFL_ACTIVE;
3955 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3956
3957 /* init wq */
3958 wq->flags = flags;
3959 wq->saved_max_active = max_active;
3960 mutex_init(&wq->mutex);
3961 atomic_set(&wq->nr_pwqs_to_flush, 0);
3962 INIT_LIST_HEAD(&wq->pwqs);
3963 INIT_LIST_HEAD(&wq->flusher_queue);
3964 INIT_LIST_HEAD(&wq->flusher_overflow);
3965 INIT_LIST_HEAD(&wq->maydays);
3966
3967 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3968 INIT_LIST_HEAD(&wq->list);
3969
3970 if (alloc_and_link_pwqs(wq) < 0)
3971 goto err_free_wq;
3972
3973 /*
3974 * Workqueues which may be used during memory reclaim should
3975 * have a rescuer to guarantee forward progress.
3976 */
3977 if (flags & WQ_MEM_RECLAIM) {
3978 struct worker *rescuer;
3979
3980 rescuer = alloc_worker(NUMA_NO_NODE);
3981 if (!rescuer)
3982 goto err_destroy;
3983
3984 rescuer->rescue_wq = wq;
3985 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3986 wq->name);
3987 if (IS_ERR(rescuer->task)) {
3988 kfree(rescuer);
3989 goto err_destroy;
3990 }
3991
3992 wq->rescuer = rescuer;
3993 kthread_bind_mask(rescuer->task, cpu_possible_mask);
3994 wake_up_process(rescuer->task);
3995 }
3996
3997 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998 goto err_destroy;
3999
4000 /*
4001 * wq_pool_mutex protects global freeze state and workqueues list.
4002 * Grab it, adjust max_active and add the new @wq to workqueues
4003 * list.
4004 */
4005 mutex_lock(&wq_pool_mutex);
4006
4007 mutex_lock(&wq->mutex);
4008 for_each_pwq(pwq, wq)
4009 pwq_adjust_max_active(pwq);
4010 mutex_unlock(&wq->mutex);
4011
4012 list_add_tail_rcu(&wq->list, &workqueues);
4013
4014 mutex_unlock(&wq_pool_mutex);
4015
4016 return wq;
4017
4018err_free_wq:
4019 free_workqueue_attrs(wq->unbound_attrs);
4020 kfree(wq);
4021 return NULL;
4022err_destroy:
4023 destroy_workqueue(wq);
4024 return NULL;
4025}
4026EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4027
4028/**
4029 * destroy_workqueue - safely terminate a workqueue
4030 * @wq: target workqueue
4031 *
4032 * Safely destroy a workqueue. All work currently pending will be done first.
4033 */
4034void destroy_workqueue(struct workqueue_struct *wq)
4035{
4036 struct pool_workqueue *pwq;
4037 int node;
4038
4039 /* drain it before proceeding with destruction */
4040 drain_workqueue(wq);
4041
4042 /* sanity checks */
4043 mutex_lock(&wq->mutex);
4044 for_each_pwq(pwq, wq) {
4045 int i;
4046
4047 for (i = 0; i < WORK_NR_COLORS; i++) {
4048 if (WARN_ON(pwq->nr_in_flight[i])) {
4049 mutex_unlock(&wq->mutex);
4050 return;
4051 }
4052 }
4053
4054 if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4055 WARN_ON(pwq->nr_active) ||
4056 WARN_ON(!list_empty(&pwq->delayed_works))) {
4057 mutex_unlock(&wq->mutex);
4058 return;
4059 }
4060 }
4061 mutex_unlock(&wq->mutex);
4062
4063 /*
4064 * wq list is used to freeze wq, remove from list after
4065 * flushing is complete in case freeze races us.
4066 */
4067 mutex_lock(&wq_pool_mutex);
4068 list_del_rcu(&wq->list);
4069 mutex_unlock(&wq_pool_mutex);
4070
4071 workqueue_sysfs_unregister(wq);
4072
4073 if (wq->rescuer)
4074 kthread_stop(wq->rescuer->task);
4075
4076 if (!(wq->flags & WQ_UNBOUND)) {
4077 /*
4078 * The base ref is never dropped on per-cpu pwqs. Directly
4079 * schedule RCU free.
4080 */
4081 call_rcu_sched(&wq->rcu, rcu_free_wq);
4082 } else {
4083 /*
4084 * We're the sole accessor of @wq at this point. Directly
4085 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4086 * @wq will be freed when the last pwq is released.
4087 */
4088 for_each_node(node) {
4089 pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4090 RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4091 put_pwq_unlocked(pwq);
4092 }
4093
4094 /*
4095 * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4096 * put. Don't access it afterwards.
4097 */
4098 pwq = wq->dfl_pwq;
4099 wq->dfl_pwq = NULL;
4100 put_pwq_unlocked(pwq);
4101 }
4102}
4103EXPORT_SYMBOL_GPL(destroy_workqueue);
4104
4105/**
4106 * workqueue_set_max_active - adjust max_active of a workqueue
4107 * @wq: target workqueue
4108 * @max_active: new max_active value.
4109 *
4110 * Set max_active of @wq to @max_active.
4111 *
4112 * CONTEXT:
4113 * Don't call from IRQ context.
4114 */
4115void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4116{
4117 struct pool_workqueue *pwq;
4118
4119 /* disallow meddling with max_active for ordered workqueues */
4120 if (WARN_ON(wq->flags & __WQ_ORDERED))
4121 return;
4122
4123 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4124
4125 mutex_lock(&wq->mutex);
4126
4127 wq->saved_max_active = max_active;
4128
4129 for_each_pwq(pwq, wq)
4130 pwq_adjust_max_active(pwq);
4131
4132 mutex_unlock(&wq->mutex);
4133}
4134EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4135
4136/**
4137 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4138 *
4139 * Determine whether %current is a workqueue rescuer. Can be used from
4140 * work functions to determine whether it's being run off the rescuer task.
4141 *
4142 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4143 */
4144bool current_is_workqueue_rescuer(void)
4145{
4146 struct worker *worker = current_wq_worker();
4147
4148 return worker && worker->rescue_wq;
4149}
4150
4151/**
4152 * workqueue_congested - test whether a workqueue is congested
4153 * @cpu: CPU in question
4154 * @wq: target workqueue
4155 *
4156 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4157 * no synchronization around this function and the test result is
4158 * unreliable and only useful as advisory hints or for debugging.
4159 *
4160 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4161 * Note that both per-cpu and unbound workqueues may be associated with
4162 * multiple pool_workqueues which have separate congested states. A
4163 * workqueue being congested on one CPU doesn't mean the workqueue is also
4164 * contested on other CPUs / NUMA nodes.
4165 *
4166 * Return:
4167 * %true if congested, %false otherwise.
4168 */
4169bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4170{
4171 struct pool_workqueue *pwq;
4172 bool ret;
4173
4174 rcu_read_lock_sched();
4175
4176 if (cpu == WORK_CPU_UNBOUND)
4177 cpu = smp_processor_id();
4178
4179 if (!(wq->flags & WQ_UNBOUND))
4180 pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4181 else
4182 pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4183
4184 ret = !list_empty(&pwq->delayed_works);
4185 rcu_read_unlock_sched();
4186
4187 return ret;
4188}
4189EXPORT_SYMBOL_GPL(workqueue_congested);
4190
4191/**
4192 * work_busy - test whether a work is currently pending or running
4193 * @work: the work to be tested
4194 *
4195 * Test whether @work is currently pending or running. There is no
4196 * synchronization around this function and the test result is
4197 * unreliable and only useful as advisory hints or for debugging.
4198 *
4199 * Return:
4200 * OR'd bitmask of WORK_BUSY_* bits.
4201 */
4202unsigned int work_busy(struct work_struct *work)
4203{
4204 struct worker_pool *pool;
4205 unsigned long flags;
4206 unsigned int ret = 0;
4207
4208 if (work_pending(work))
4209 ret |= WORK_BUSY_PENDING;
4210
4211 local_irq_save(flags);
4212 pool = get_work_pool(work);
4213 if (pool) {
4214 spin_lock(&pool->lock);
4215 if (find_worker_executing_work(pool, work))
4216 ret |= WORK_BUSY_RUNNING;
4217 spin_unlock(&pool->lock);
4218 }
4219 local_irq_restore(flags);
4220
4221 return ret;
4222}
4223EXPORT_SYMBOL_GPL(work_busy);
4224
4225/**
4226 * set_worker_desc - set description for the current work item
4227 * @fmt: printf-style format string
4228 * @...: arguments for the format string
4229 *
4230 * This function can be called by a running work function to describe what
4231 * the work item is about. If the worker task gets dumped, this
4232 * information will be printed out together to help debugging. The
4233 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4234 */
4235void set_worker_desc(const char *fmt, ...)
4236{
4237 struct worker *worker = current_wq_worker();
4238 va_list args;
4239
4240 if (worker) {
4241 va_start(args, fmt);
4242 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4243 va_end(args);
4244 worker->desc_valid = true;
4245 }
4246}
4247
4248/**
4249 * print_worker_info - print out worker information and description
4250 * @log_lvl: the log level to use when printing
4251 * @task: target task
4252 *
4253 * If @task is a worker and currently executing a work item, print out the
4254 * name of the workqueue being serviced and worker description set with
4255 * set_worker_desc() by the currently executing work item.
4256 *
4257 * This function can be safely called on any task as long as the
4258 * task_struct itself is accessible. While safe, this function isn't
4259 * synchronized and may print out mixups or garbages of limited length.
4260 */
4261void print_worker_info(const char *log_lvl, struct task_struct *task)
4262{
4263 work_func_t *fn = NULL;
4264 char name[WQ_NAME_LEN] = { };
4265 char desc[WORKER_DESC_LEN] = { };
4266 struct pool_workqueue *pwq = NULL;
4267 struct workqueue_struct *wq = NULL;
4268 bool desc_valid = false;
4269 struct worker *worker;
4270
4271 if (!(task->flags & PF_WQ_WORKER))
4272 return;
4273
4274 /*
4275 * This function is called without any synchronization and @task
4276 * could be in any state. Be careful with dereferences.
4277 */
4278 worker = probe_kthread_data(task);
4279
4280 /*
4281 * Carefully copy the associated workqueue's workfn and name. Keep
4282 * the original last '\0' in case the original contains garbage.
4283 */
4284 probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4285 probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4286 probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4287 probe_kernel_read(name, wq->name, sizeof(name) - 1);
4288
4289 /* copy worker description */
4290 probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4291 if (desc_valid)
4292 probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4293
4294 if (fn || name[0] || desc[0]) {
4295 printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4296 if (desc[0])
4297 pr_cont(" (%s)", desc);
4298 pr_cont("\n");
4299 }
4300}
4301
4302static void pr_cont_pool_info(struct worker_pool *pool)
4303{
4304 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4305 if (pool->node != NUMA_NO_NODE)
4306 pr_cont(" node=%d", pool->node);
4307 pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4308}
4309
4310static void pr_cont_work(bool comma, struct work_struct *work)
4311{
4312 if (work->func == wq_barrier_func) {
4313 struct wq_barrier *barr;
4314
4315 barr = container_of(work, struct wq_barrier, work);
4316
4317 pr_cont("%s BAR(%d)", comma ? "," : "",
4318 task_pid_nr(barr->task));
4319 } else {
4320 pr_cont("%s %pf", comma ? "," : "", work->func);
4321 }
4322}
4323
4324static void show_pwq(struct pool_workqueue *pwq)
4325{
4326 struct worker_pool *pool = pwq->pool;
4327 struct work_struct *work;
4328 struct worker *worker;
4329 bool has_in_flight = false, has_pending = false;
4330 int bkt;
4331
4332 pr_info(" pwq %d:", pool->id);
4333 pr_cont_pool_info(pool);
4334
4335 pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4336 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4337
4338 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4339 if (worker->current_pwq == pwq) {
4340 has_in_flight = true;
4341 break;
4342 }
4343 }
4344 if (has_in_flight) {
4345 bool comma = false;
4346
4347 pr_info(" in-flight:");
4348 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4349 if (worker->current_pwq != pwq)
4350 continue;
4351
4352 pr_cont("%s %d%s:%pf", comma ? "," : "",
4353 task_pid_nr(worker->task),
4354 worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4355 worker->current_func);
4356 list_for_each_entry(work, &worker->scheduled, entry)
4357 pr_cont_work(false, work);
4358 comma = true;
4359 }
4360 pr_cont("\n");
4361 }
4362
4363 list_for_each_entry(work, &pool->worklist, entry) {
4364 if (get_work_pwq(work) == pwq) {
4365 has_pending = true;
4366 break;
4367 }
4368 }
4369 if (has_pending) {
4370 bool comma = false;
4371
4372 pr_info(" pending:");
4373 list_for_each_entry(work, &pool->worklist, entry) {
4374 if (get_work_pwq(work) != pwq)
4375 continue;
4376
4377 pr_cont_work(comma, work);
4378 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4379 }
4380 pr_cont("\n");
4381 }
4382
4383 if (!list_empty(&pwq->delayed_works)) {
4384 bool comma = false;
4385
4386 pr_info(" delayed:");
4387 list_for_each_entry(work, &pwq->delayed_works, entry) {
4388 pr_cont_work(comma, work);
4389 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4390 }
4391 pr_cont("\n");
4392 }
4393}
4394
4395/**
4396 * show_workqueue_state - dump workqueue state
4397 *
4398 * Called from a sysrq handler and prints out all busy workqueues and
4399 * pools.
4400 */
4401void show_workqueue_state(void)
4402{
4403 struct workqueue_struct *wq;
4404 struct worker_pool *pool;
4405 unsigned long flags;
4406 int pi;
4407
4408 rcu_read_lock_sched();
4409
4410 pr_info("Showing busy workqueues and worker pools:\n");
4411
4412 list_for_each_entry_rcu(wq, &workqueues, list) {
4413 struct pool_workqueue *pwq;
4414 bool idle = true;
4415
4416 for_each_pwq(pwq, wq) {
4417 if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4418 idle = false;
4419 break;
4420 }
4421 }
4422 if (idle)
4423 continue;
4424
4425 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4426
4427 for_each_pwq(pwq, wq) {
4428 spin_lock_irqsave(&pwq->pool->lock, flags);
4429 if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4430 show_pwq(pwq);
4431 spin_unlock_irqrestore(&pwq->pool->lock, flags);
4432 }
4433 }
4434
4435 for_each_pool(pool, pi) {
4436 struct worker *worker;
4437 bool first = true;
4438
4439 spin_lock_irqsave(&pool->lock, flags);
4440 if (pool->nr_workers == pool->nr_idle)
4441 goto next_pool;
4442
4443 pr_info("pool %d:", pool->id);
4444 pr_cont_pool_info(pool);
4445 pr_cont(" hung=%us workers=%d",
4446 jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4447 pool->nr_workers);
4448 if (pool->manager)
4449 pr_cont(" manager: %d",
4450 task_pid_nr(pool->manager->task));
4451 list_for_each_entry(worker, &pool->idle_list, entry) {
4452 pr_cont(" %s%d", first ? "idle: " : "",
4453 task_pid_nr(worker->task));
4454 first = false;
4455 }
4456 pr_cont("\n");
4457 next_pool:
4458 spin_unlock_irqrestore(&pool->lock, flags);
4459 }
4460
4461 rcu_read_unlock_sched();
4462}
4463
4464/*
4465 * CPU hotplug.
4466 *
4467 * There are two challenges in supporting CPU hotplug. Firstly, there
4468 * are a lot of assumptions on strong associations among work, pwq and
4469 * pool which make migrating pending and scheduled works very
4470 * difficult to implement without impacting hot paths. Secondly,
4471 * worker pools serve mix of short, long and very long running works making
4472 * blocked draining impractical.
4473 *
4474 * This is solved by allowing the pools to be disassociated from the CPU
4475 * running as an unbound one and allowing it to be reattached later if the
4476 * cpu comes back online.
4477 */
4478
4479static void wq_unbind_fn(struct work_struct *work)
4480{
4481 int cpu = smp_processor_id();
4482 struct worker_pool *pool;
4483 struct worker *worker;
4484
4485 for_each_cpu_worker_pool(pool, cpu) {
4486 mutex_lock(&pool->attach_mutex);
4487 spin_lock_irq(&pool->lock);
4488
4489 /*
4490 * We've blocked all attach/detach operations. Make all workers
4491 * unbound and set DISASSOCIATED. Before this, all workers
4492 * except for the ones which are still executing works from
4493 * before the last CPU down must be on the cpu. After
4494 * this, they may become diasporas.
4495 */
4496 for_each_pool_worker(worker, pool)
4497 worker->flags |= WORKER_UNBOUND;
4498
4499 pool->flags |= POOL_DISASSOCIATED;
4500
4501 spin_unlock_irq(&pool->lock);
4502 mutex_unlock(&pool->attach_mutex);
4503
4504 /*
4505 * Call schedule() so that we cross rq->lock and thus can
4506 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4507 * This is necessary as scheduler callbacks may be invoked
4508 * from other cpus.
4509 */
4510 schedule();
4511
4512 /*
4513 * Sched callbacks are disabled now. Zap nr_running.
4514 * After this, nr_running stays zero and need_more_worker()
4515 * and keep_working() are always true as long as the
4516 * worklist is not empty. This pool now behaves as an
4517 * unbound (in terms of concurrency management) pool which
4518 * are served by workers tied to the pool.
4519 */
4520 atomic_set(&pool->nr_running, 0);
4521
4522 /*
4523 * With concurrency management just turned off, a busy
4524 * worker blocking could lead to lengthy stalls. Kick off
4525 * unbound chain execution of currently pending work items.
4526 */
4527 spin_lock_irq(&pool->lock);
4528 wake_up_worker(pool);
4529 spin_unlock_irq(&pool->lock);
4530 }
4531}
4532
4533/**
4534 * rebind_workers - rebind all workers of a pool to the associated CPU
4535 * @pool: pool of interest
4536 *
4537 * @pool->cpu is coming online. Rebind all workers to the CPU.
4538 */
4539static void rebind_workers(struct worker_pool *pool)
4540{
4541 struct worker *worker;
4542
4543 lockdep_assert_held(&pool->attach_mutex);
4544
4545 /*
4546 * Restore CPU affinity of all workers. As all idle workers should
4547 * be on the run-queue of the associated CPU before any local
4548 * wake-ups for concurrency management happen, restore CPU affinity
4549 * of all workers first and then clear UNBOUND. As we're called
4550 * from CPU_ONLINE, the following shouldn't fail.
4551 */
4552 for_each_pool_worker(worker, pool)
4553 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4554 pool->attrs->cpumask) < 0);
4555
4556 spin_lock_irq(&pool->lock);
4557
4558 /*
4559 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4560 * w/o preceding DOWN_PREPARE. Work around it. CPU hotplug is
4561 * being reworked and this can go away in time.
4562 */
4563 if (!(pool->flags & POOL_DISASSOCIATED)) {
4564 spin_unlock_irq(&pool->lock);
4565 return;
4566 }
4567
4568 pool->flags &= ~POOL_DISASSOCIATED;
4569
4570 for_each_pool_worker(worker, pool) {
4571 unsigned int worker_flags = worker->flags;
4572
4573 /*
4574 * A bound idle worker should actually be on the runqueue
4575 * of the associated CPU for local wake-ups targeting it to
4576 * work. Kick all idle workers so that they migrate to the
4577 * associated CPU. Doing this in the same loop as
4578 * replacing UNBOUND with REBOUND is safe as no worker will
4579 * be bound before @pool->lock is released.
4580 */
4581 if (worker_flags & WORKER_IDLE)
4582 wake_up_process(worker->task);
4583
4584 /*
4585 * We want to clear UNBOUND but can't directly call
4586 * worker_clr_flags() or adjust nr_running. Atomically
4587 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4588 * @worker will clear REBOUND using worker_clr_flags() when
4589 * it initiates the next execution cycle thus restoring
4590 * concurrency management. Note that when or whether
4591 * @worker clears REBOUND doesn't affect correctness.
4592 *
4593 * ACCESS_ONCE() is necessary because @worker->flags may be
4594 * tested without holding any lock in
4595 * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4596 * fail incorrectly leading to premature concurrency
4597 * management operations.
4598 */
4599 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4600 worker_flags |= WORKER_REBOUND;
4601 worker_flags &= ~WORKER_UNBOUND;
4602 ACCESS_ONCE(worker->flags) = worker_flags;
4603 }
4604
4605 spin_unlock_irq(&pool->lock);
4606}
4607
4608/**
4609 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4610 * @pool: unbound pool of interest
4611 * @cpu: the CPU which is coming up
4612 *
4613 * An unbound pool may end up with a cpumask which doesn't have any online
4614 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4615 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4616 * online CPU before, cpus_allowed of all its workers should be restored.
4617 */
4618static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4619{
4620 static cpumask_t cpumask;
4621 struct worker *worker;
4622
4623 lockdep_assert_held(&pool->attach_mutex);
4624
4625 /* is @cpu allowed for @pool? */
4626 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4627 return;
4628
4629 /* is @cpu the only online CPU? */
4630 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4631 if (cpumask_weight(&cpumask) != 1)
4632 return;
4633
4634 /* as we're called from CPU_ONLINE, the following shouldn't fail */
4635 for_each_pool_worker(worker, pool)
4636 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4637 pool->attrs->cpumask) < 0);
4638}
4639
4640/*
4641 * Workqueues should be brought up before normal priority CPU notifiers.
4642 * This will be registered high priority CPU notifier.
4643 */
4644static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4645 unsigned long action,
4646 void *hcpu)
4647{
4648 int cpu = (unsigned long)hcpu;
4649 struct worker_pool *pool;
4650 struct workqueue_struct *wq;
4651 int pi;
4652
4653 switch (action & ~CPU_TASKS_FROZEN) {
4654 case CPU_UP_PREPARE:
4655 for_each_cpu_worker_pool(pool, cpu) {
4656 if (pool->nr_workers)
4657 continue;
4658 if (!create_worker(pool))
4659 return NOTIFY_BAD;
4660 }
4661 break;
4662
4663 case CPU_DOWN_FAILED:
4664 case CPU_ONLINE:
4665 mutex_lock(&wq_pool_mutex);
4666
4667 for_each_pool(pool, pi) {
4668 mutex_lock(&pool->attach_mutex);
4669
4670 if (pool->cpu == cpu)
4671 rebind_workers(pool);
4672 else if (pool->cpu < 0)
4673 restore_unbound_workers_cpumask(pool, cpu);
4674
4675 mutex_unlock(&pool->attach_mutex);
4676 }
4677
4678 /* update NUMA affinity of unbound workqueues */
4679 list_for_each_entry(wq, &workqueues, list)
4680 wq_update_unbound_numa(wq, cpu, true);
4681
4682 mutex_unlock(&wq_pool_mutex);
4683 break;
4684 }
4685 return NOTIFY_OK;
4686}
4687
4688/*
4689 * Workqueues should be brought down after normal priority CPU notifiers.
4690 * This will be registered as low priority CPU notifier.
4691 */
4692static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4693 unsigned long action,
4694 void *hcpu)
4695{
4696 int cpu = (unsigned long)hcpu;
4697 struct work_struct unbind_work;
4698 struct workqueue_struct *wq;
4699
4700 switch (action & ~CPU_TASKS_FROZEN) {
4701 case CPU_DOWN_PREPARE:
4702 /* unbinding per-cpu workers should happen on the local CPU */
4703 INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4704 queue_work_on(cpu, system_highpri_wq, &unbind_work);
4705
4706 /* update NUMA affinity of unbound workqueues */
4707 mutex_lock(&wq_pool_mutex);
4708 list_for_each_entry(wq, &workqueues, list)
4709 wq_update_unbound_numa(wq, cpu, false);
4710 mutex_unlock(&wq_pool_mutex);
4711
4712 /* wait for per-cpu unbinding to finish */
4713 flush_work(&unbind_work);
4714 destroy_work_on_stack(&unbind_work);
4715 break;
4716 }
4717 return NOTIFY_OK;
4718}
4719
4720#ifdef CONFIG_SMP
4721
4722struct work_for_cpu {
4723 struct work_struct work;
4724 long (*fn)(void *);
4725 void *arg;
4726 long ret;
4727};
4728
4729static void work_for_cpu_fn(struct work_struct *work)
4730{
4731 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4732
4733 wfc->ret = wfc->fn(wfc->arg);
4734}
4735
4736/**
4737 * work_on_cpu - run a function in thread context on a particular cpu
4738 * @cpu: the cpu to run on
4739 * @fn: the function to run
4740 * @arg: the function arg
4741 *
4742 * It is up to the caller to ensure that the cpu doesn't go offline.
4743 * The caller must not hold any locks which would prevent @fn from completing.
4744 *
4745 * Return: The value @fn returns.
4746 */
4747long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4748{
4749 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4750
4751 INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4752 schedule_work_on(cpu, &wfc.work);
4753 flush_work(&wfc.work);
4754 destroy_work_on_stack(&wfc.work);
4755 return wfc.ret;
4756}
4757EXPORT_SYMBOL_GPL(work_on_cpu);
4758#endif /* CONFIG_SMP */
4759
4760#ifdef CONFIG_FREEZER
4761
4762/**
4763 * freeze_workqueues_begin - begin freezing workqueues
4764 *
4765 * Start freezing workqueues. After this function returns, all freezable
4766 * workqueues will queue new works to their delayed_works list instead of
4767 * pool->worklist.
4768 *
4769 * CONTEXT:
4770 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4771 */
4772void freeze_workqueues_begin(void)
4773{
4774 struct workqueue_struct *wq;
4775 struct pool_workqueue *pwq;
4776
4777 mutex_lock(&wq_pool_mutex);
4778
4779 WARN_ON_ONCE(workqueue_freezing);
4780 workqueue_freezing = true;
4781
4782 list_for_each_entry(wq, &workqueues, list) {
4783 mutex_lock(&wq->mutex);
4784 for_each_pwq(pwq, wq)
4785 pwq_adjust_max_active(pwq);
4786 mutex_unlock(&wq->mutex);
4787 }
4788
4789 mutex_unlock(&wq_pool_mutex);
4790}
4791
4792/**
4793 * freeze_workqueues_busy - are freezable workqueues still busy?
4794 *
4795 * Check whether freezing is complete. This function must be called
4796 * between freeze_workqueues_begin() and thaw_workqueues().
4797 *
4798 * CONTEXT:
4799 * Grabs and releases wq_pool_mutex.
4800 *
4801 * Return:
4802 * %true if some freezable workqueues are still busy. %false if freezing
4803 * is complete.
4804 */
4805bool freeze_workqueues_busy(void)
4806{
4807 bool busy = false;
4808 struct workqueue_struct *wq;
4809 struct pool_workqueue *pwq;
4810
4811 mutex_lock(&wq_pool_mutex);
4812
4813 WARN_ON_ONCE(!workqueue_freezing);
4814
4815 list_for_each_entry(wq, &workqueues, list) {
4816 if (!(wq->flags & WQ_FREEZABLE))
4817 continue;
4818 /*
4819 * nr_active is monotonically decreasing. It's safe
4820 * to peek without lock.
4821 */
4822 rcu_read_lock_sched();
4823 for_each_pwq(pwq, wq) {
4824 WARN_ON_ONCE(pwq->nr_active < 0);
4825 if (pwq->nr_active) {
4826 busy = true;
4827 rcu_read_unlock_sched();
4828 goto out_unlock;
4829 }
4830 }
4831 rcu_read_unlock_sched();
4832 }
4833out_unlock:
4834 mutex_unlock(&wq_pool_mutex);
4835 return busy;
4836}
4837
4838/**
4839 * thaw_workqueues - thaw workqueues
4840 *
4841 * Thaw workqueues. Normal queueing is restored and all collected
4842 * frozen works are transferred to their respective pool worklists.
4843 *
4844 * CONTEXT:
4845 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4846 */
4847void thaw_workqueues(void)
4848{
4849 struct workqueue_struct *wq;
4850 struct pool_workqueue *pwq;
4851
4852 mutex_lock(&wq_pool_mutex);
4853
4854 if (!workqueue_freezing)
4855 goto out_unlock;
4856
4857 workqueue_freezing = false;
4858
4859 /* restore max_active and repopulate worklist */
4860 list_for_each_entry(wq, &workqueues, list) {
4861 mutex_lock(&wq->mutex);
4862 for_each_pwq(pwq, wq)
4863 pwq_adjust_max_active(pwq);
4864 mutex_unlock(&wq->mutex);
4865 }
4866
4867out_unlock:
4868 mutex_unlock(&wq_pool_mutex);
4869}
4870#endif /* CONFIG_FREEZER */
4871
4872static int workqueue_apply_unbound_cpumask(void)
4873{
4874 LIST_HEAD(ctxs);
4875 int ret = 0;
4876 struct workqueue_struct *wq;
4877 struct apply_wqattrs_ctx *ctx, *n;
4878
4879 lockdep_assert_held(&wq_pool_mutex);
4880
4881 list_for_each_entry(wq, &workqueues, list) {
4882 if (!(wq->flags & WQ_UNBOUND))
4883 continue;
4884 /* creating multiple pwqs breaks ordering guarantee */
4885 if (wq->flags & __WQ_ORDERED)
4886 continue;
4887
4888 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4889 if (!ctx) {
4890 ret = -ENOMEM;
4891 break;
4892 }
4893
4894 list_add_tail(&ctx->list, &ctxs);
4895 }
4896
4897 list_for_each_entry_safe(ctx, n, &ctxs, list) {
4898 if (!ret)
4899 apply_wqattrs_commit(ctx);
4900 apply_wqattrs_cleanup(ctx);
4901 }
4902
4903 return ret;
4904}
4905
4906/**
4907 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4908 * @cpumask: the cpumask to set
4909 *
4910 * The low-level workqueues cpumask is a global cpumask that limits
4911 * the affinity of all unbound workqueues. This function check the @cpumask
4912 * and apply it to all unbound workqueues and updates all pwqs of them.
4913 *
4914 * Retun: 0 - Success
4915 * -EINVAL - Invalid @cpumask
4916 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
4917 */
4918int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4919{
4920 int ret = -EINVAL;
4921 cpumask_var_t saved_cpumask;
4922
4923 if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4924 return -ENOMEM;
4925
4926 cpumask_and(cpumask, cpumask, cpu_possible_mask);
4927 if (!cpumask_empty(cpumask)) {
4928 apply_wqattrs_lock();
4929
4930 /* save the old wq_unbound_cpumask. */
4931 cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4932
4933 /* update wq_unbound_cpumask at first and apply it to wqs. */
4934 cpumask_copy(wq_unbound_cpumask, cpumask);
4935 ret = workqueue_apply_unbound_cpumask();
4936
4937 /* restore the wq_unbound_cpumask when failed. */
4938 if (ret < 0)
4939 cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4940
4941 apply_wqattrs_unlock();
4942 }
4943
4944 free_cpumask_var(saved_cpumask);
4945 return ret;
4946}
4947
4948#ifdef CONFIG_SYSFS
4949/*
4950 * Workqueues with WQ_SYSFS flag set is visible to userland via
4951 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
4952 * following attributes.
4953 *
4954 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
4955 * max_active RW int : maximum number of in-flight work items
4956 *
4957 * Unbound workqueues have the following extra attributes.
4958 *
4959 * id RO int : the associated pool ID
4960 * nice RW int : nice value of the workers
4961 * cpumask RW mask : bitmask of allowed CPUs for the workers
4962 */
4963struct wq_device {
4964 struct workqueue_struct *wq;
4965 struct device dev;
4966};
4967
4968static struct workqueue_struct *dev_to_wq(struct device *dev)
4969{
4970 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4971
4972 return wq_dev->wq;
4973}
4974
4975static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4976 char *buf)
4977{
4978 struct workqueue_struct *wq = dev_to_wq(dev);
4979
4980 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4981}
4982static DEVICE_ATTR_RO(per_cpu);
4983
4984static ssize_t max_active_show(struct device *dev,
4985 struct device_attribute *attr, char *buf)
4986{
4987 struct workqueue_struct *wq = dev_to_wq(dev);
4988
4989 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4990}
4991
4992static ssize_t max_active_store(struct device *dev,
4993 struct device_attribute *attr, const char *buf,
4994 size_t count)
4995{
4996 struct workqueue_struct *wq = dev_to_wq(dev);
4997 int val;
4998
4999 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5000 return -EINVAL;
5001
5002 workqueue_set_max_active(wq, val);
5003 return count;
5004}
5005static DEVICE_ATTR_RW(max_active);
5006
5007static struct attribute *wq_sysfs_attrs[] = {
5008 &dev_attr_per_cpu.attr,
5009 &dev_attr_max_active.attr,
5010 NULL,
5011};
5012ATTRIBUTE_GROUPS(wq_sysfs);
5013
5014static ssize_t wq_pool_ids_show(struct device *dev,
5015 struct device_attribute *attr, char *buf)
5016{
5017 struct workqueue_struct *wq = dev_to_wq(dev);
5018 const char *delim = "";
5019 int node, written = 0;
5020
5021 rcu_read_lock_sched();
5022 for_each_node(node) {
5023 written += scnprintf(buf + written, PAGE_SIZE - written,
5024 "%s%d:%d", delim, node,
5025 unbound_pwq_by_node(wq, node)->pool->id);
5026 delim = " ";
5027 }
5028 written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5029 rcu_read_unlock_sched();
5030
5031 return written;
5032}
5033
5034static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5035 char *buf)
5036{
5037 struct workqueue_struct *wq = dev_to_wq(dev);
5038 int written;
5039
5040 mutex_lock(&wq->mutex);
5041 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5042 mutex_unlock(&wq->mutex);
5043
5044 return written;
5045}
5046
5047/* prepare workqueue_attrs for sysfs store operations */
5048static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5049{
5050 struct workqueue_attrs *attrs;
5051
5052 lockdep_assert_held(&wq_pool_mutex);
5053
5054 attrs = alloc_workqueue_attrs(GFP_KERNEL);
5055 if (!attrs)
5056 return NULL;
5057
5058 copy_workqueue_attrs(attrs, wq->unbound_attrs);
5059 return attrs;
5060}
5061
5062static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5063 const char *buf, size_t count)
5064{
5065 struct workqueue_struct *wq = dev_to_wq(dev);
5066 struct workqueue_attrs *attrs;
5067 int ret = -ENOMEM;
5068
5069 apply_wqattrs_lock();
5070
5071 attrs = wq_sysfs_prep_attrs(wq);
5072 if (!attrs)
5073 goto out_unlock;
5074
5075 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5076 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5077 ret = apply_workqueue_attrs_locked(wq, attrs);
5078 else
5079 ret = -EINVAL;
5080
5081out_unlock:
5082 apply_wqattrs_unlock();
5083 free_workqueue_attrs(attrs);
5084 return ret ?: count;
5085}
5086
5087static ssize_t wq_cpumask_show(struct device *dev,
5088 struct device_attribute *attr, char *buf)
5089{
5090 struct workqueue_struct *wq = dev_to_wq(dev);
5091 int written;
5092
5093 mutex_lock(&wq->mutex);
5094 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5095 cpumask_pr_args(wq->unbound_attrs->cpumask));
5096 mutex_unlock(&wq->mutex);
5097 return written;
5098}
5099
5100static ssize_t wq_cpumask_store(struct device *dev,
5101 struct device_attribute *attr,
5102 const char *buf, size_t count)
5103{
5104 struct workqueue_struct *wq = dev_to_wq(dev);
5105 struct workqueue_attrs *attrs;
5106 int ret = -ENOMEM;
5107
5108 apply_wqattrs_lock();
5109
5110 attrs = wq_sysfs_prep_attrs(wq);
5111 if (!attrs)
5112 goto out_unlock;
5113
5114 ret = cpumask_parse(buf, attrs->cpumask);
5115 if (!ret)
5116 ret = apply_workqueue_attrs_locked(wq, attrs);
5117
5118out_unlock:
5119 apply_wqattrs_unlock();
5120 free_workqueue_attrs(attrs);
5121 return ret ?: count;
5122}
5123
5124static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5125 char *buf)
5126{
5127 struct workqueue_struct *wq = dev_to_wq(dev);
5128 int written;
5129
5130 mutex_lock(&wq->mutex);
5131 written = scnprintf(buf, PAGE_SIZE, "%d\n",
5132 !wq->unbound_attrs->no_numa);
5133 mutex_unlock(&wq->mutex);
5134
5135 return written;
5136}
5137
5138static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5139 const char *buf, size_t count)
5140{
5141 struct workqueue_struct *wq = dev_to_wq(dev);
5142 struct workqueue_attrs *attrs;
5143 int v, ret = -ENOMEM;
5144
5145 apply_wqattrs_lock();
5146
5147 attrs = wq_sysfs_prep_attrs(wq);
5148 if (!attrs)
5149 goto out_unlock;
5150
5151 ret = -EINVAL;
5152 if (sscanf(buf, "%d", &v) == 1) {
5153 attrs->no_numa = !v;
5154 ret = apply_workqueue_attrs_locked(wq, attrs);
5155 }
5156
5157out_unlock:
5158 apply_wqattrs_unlock();
5159 free_workqueue_attrs(attrs);
5160 return ret ?: count;
5161}
5162
5163static struct device_attribute wq_sysfs_unbound_attrs[] = {
5164 __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5165 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5166 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5167 __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5168 __ATTR_NULL,
5169};
5170
5171static struct bus_type wq_subsys = {
5172 .name = "workqueue",
5173 .dev_groups = wq_sysfs_groups,
5174};
5175
5176static ssize_t wq_unbound_cpumask_show(struct device *dev,
5177 struct device_attribute *attr, char *buf)
5178{
5179 int written;
5180
5181 mutex_lock(&wq_pool_mutex);
5182 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5183 cpumask_pr_args(wq_unbound_cpumask));
5184 mutex_unlock(&wq_pool_mutex);
5185
5186 return written;
5187}
5188
5189static ssize_t wq_unbound_cpumask_store(struct device *dev,
5190 struct device_attribute *attr, const char *buf, size_t count)
5191{
5192 cpumask_var_t cpumask;
5193 int ret;
5194
5195 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5196 return -ENOMEM;
5197
5198 ret = cpumask_parse(buf, cpumask);
5199 if (!ret)
5200 ret = workqueue_set_unbound_cpumask(cpumask);
5201
5202 free_cpumask_var(cpumask);
5203 return ret ? ret : count;
5204}
5205
5206static struct device_attribute wq_sysfs_cpumask_attr =
5207 __ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5208 wq_unbound_cpumask_store);
5209
5210static int __init wq_sysfs_init(void)
5211{
5212 int err;
5213
5214 err = subsys_virtual_register(&wq_subsys, NULL);
5215 if (err)
5216 return err;
5217
5218 return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5219}
5220core_initcall(wq_sysfs_init);
5221
5222static void wq_device_release(struct device *dev)
5223{
5224 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5225
5226 kfree(wq_dev);
5227}
5228
5229/**
5230 * workqueue_sysfs_register - make a workqueue visible in sysfs
5231 * @wq: the workqueue to register
5232 *
5233 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5234 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5235 * which is the preferred method.
5236 *
5237 * Workqueue user should use this function directly iff it wants to apply
5238 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5239 * apply_workqueue_attrs() may race against userland updating the
5240 * attributes.
5241 *
5242 * Return: 0 on success, -errno on failure.
5243 */
5244int workqueue_sysfs_register(struct workqueue_struct *wq)
5245{
5246 struct wq_device *wq_dev;
5247 int ret;
5248
5249 /*
5250 * Adjusting max_active or creating new pwqs by applying
5251 * attributes breaks ordering guarantee. Disallow exposing ordered
5252 * workqueues.
5253 */
5254 if (WARN_ON(wq->flags & __WQ_ORDERED))
5255 return -EINVAL;
5256
5257 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5258 if (!wq_dev)
5259 return -ENOMEM;
5260
5261 wq_dev->wq = wq;
5262 wq_dev->dev.bus = &wq_subsys;
5263 wq_dev->dev.release = wq_device_release;
5264 dev_set_name(&wq_dev->dev, "%s", wq->name);
5265
5266 /*
5267 * unbound_attrs are created separately. Suppress uevent until
5268 * everything is ready.
5269 */
5270 dev_set_uevent_suppress(&wq_dev->dev, true);
5271
5272 ret = device_register(&wq_dev->dev);
5273 if (ret) {
5274 kfree(wq_dev);
5275 wq->wq_dev = NULL;
5276 return ret;
5277 }
5278
5279 if (wq->flags & WQ_UNBOUND) {
5280 struct device_attribute *attr;
5281
5282 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5283 ret = device_create_file(&wq_dev->dev, attr);
5284 if (ret) {
5285 device_unregister(&wq_dev->dev);
5286 wq->wq_dev = NULL;
5287 return ret;
5288 }
5289 }
5290 }
5291
5292 dev_set_uevent_suppress(&wq_dev->dev, false);
5293 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5294 return 0;
5295}
5296
5297/**
5298 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5299 * @wq: the workqueue to unregister
5300 *
5301 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5302 */
5303static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5304{
5305 struct wq_device *wq_dev = wq->wq_dev;
5306
5307 if (!wq->wq_dev)
5308 return;
5309
5310 wq->wq_dev = NULL;
5311 device_unregister(&wq_dev->dev);
5312}
5313#else /* CONFIG_SYSFS */
5314static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
5315#endif /* CONFIG_SYSFS */
5316
5317/*
5318 * Workqueue watchdog.
5319 *
5320 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5321 * flush dependency, a concurrency managed work item which stays RUNNING
5322 * indefinitely. Workqueue stalls can be very difficult to debug as the
5323 * usual warning mechanisms don't trigger and internal workqueue state is
5324 * largely opaque.
5325 *
5326 * Workqueue watchdog monitors all worker pools periodically and dumps
5327 * state if some pools failed to make forward progress for a while where
5328 * forward progress is defined as the first item on ->worklist changing.
5329 *
5330 * This mechanism is controlled through the kernel parameter
5331 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5332 * corresponding sysfs parameter file.
5333 */
5334#ifdef CONFIG_WQ_WATCHDOG
5335
5336static void wq_watchdog_timer_fn(unsigned long data);
5337
5338static unsigned long wq_watchdog_thresh = 30;
5339static struct timer_list wq_watchdog_timer =
5340 TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5341
5342static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5343static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5344
5345static void wq_watchdog_reset_touched(void)
5346{
5347 int cpu;
5348
5349 wq_watchdog_touched = jiffies;
5350 for_each_possible_cpu(cpu)
5351 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5352}
5353
5354static void wq_watchdog_timer_fn(unsigned long data)
5355{
5356 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5357 bool lockup_detected = false;
5358 struct worker_pool *pool;
5359 int pi;
5360
5361 if (!thresh)
5362 return;
5363
5364 rcu_read_lock();
5365
5366 for_each_pool(pool, pi) {
5367 unsigned long pool_ts, touched, ts;
5368
5369 if (list_empty(&pool->worklist))
5370 continue;
5371
5372 /* get the latest of pool and touched timestamps */
5373 pool_ts = READ_ONCE(pool->watchdog_ts);
5374 touched = READ_ONCE(wq_watchdog_touched);
5375
5376 if (time_after(pool_ts, touched))
5377 ts = pool_ts;
5378 else
5379 ts = touched;
5380
5381 if (pool->cpu >= 0) {
5382 unsigned long cpu_touched =
5383 READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5384 pool->cpu));
5385 if (time_after(cpu_touched, ts))
5386 ts = cpu_touched;
5387 }
5388
5389 /* did we stall? */
5390 if (time_after(jiffies, ts + thresh)) {
5391 lockup_detected = true;
5392 pr_emerg("BUG: workqueue lockup - pool");
5393 pr_cont_pool_info(pool);
5394 pr_cont(" stuck for %us!\n",
5395 jiffies_to_msecs(jiffies - pool_ts) / 1000);
5396 }
5397 }
5398
5399 rcu_read_unlock();
5400
5401 if (lockup_detected)
5402 show_workqueue_state();
5403
5404 wq_watchdog_reset_touched();
5405 mod_timer(&wq_watchdog_timer, jiffies + thresh);
5406}
5407
5408void wq_watchdog_touch(int cpu)
5409{
5410 if (cpu >= 0)
5411 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5412 else
5413 wq_watchdog_touched = jiffies;
5414}
5415
5416static void wq_watchdog_set_thresh(unsigned long thresh)
5417{
5418 wq_watchdog_thresh = 0;
5419 del_timer_sync(&wq_watchdog_timer);
5420
5421 if (thresh) {
5422 wq_watchdog_thresh = thresh;
5423 wq_watchdog_reset_touched();
5424 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5425 }
5426}
5427
5428static int wq_watchdog_param_set_thresh(const char *val,
5429 const struct kernel_param *kp)
5430{
5431 unsigned long thresh;
5432 int ret;
5433
5434 ret = kstrtoul(val, 0, &thresh);
5435 if (ret)
5436 return ret;
5437
5438 if (system_wq)
5439 wq_watchdog_set_thresh(thresh);
5440 else
5441 wq_watchdog_thresh = thresh;
5442
5443 return 0;
5444}
5445
5446static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5447 .set = wq_watchdog_param_set_thresh,
5448 .get = param_get_ulong,
5449};
5450
5451module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5452 0644);
5453
5454static void wq_watchdog_init(void)
5455{
5456 wq_watchdog_set_thresh(wq_watchdog_thresh);
5457}
5458
5459#else /* CONFIG_WQ_WATCHDOG */
5460
5461static inline void wq_watchdog_init(void) { }
5462
5463#endif /* CONFIG_WQ_WATCHDOG */
5464
5465static void __init wq_numa_init(void)
5466{
5467 cpumask_var_t *tbl;
5468 int node, cpu;
5469
5470 if (num_possible_nodes() <= 1)
5471 return;
5472
5473 if (wq_disable_numa) {
5474 pr_info("workqueue: NUMA affinity support disabled\n");
5475 return;
5476 }
5477
5478 wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5479 BUG_ON(!wq_update_unbound_numa_attrs_buf);
5480
5481 /*
5482 * We want masks of possible CPUs of each node which isn't readily
5483 * available. Build one from cpu_to_node() which should have been
5484 * fully initialized by now.
5485 */
5486 tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5487 BUG_ON(!tbl);
5488
5489 for_each_node(node)
5490 BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5491 node_online(node) ? node : NUMA_NO_NODE));
5492
5493 for_each_possible_cpu(cpu) {
5494 node = cpu_to_node(cpu);
5495 if (WARN_ON(node == NUMA_NO_NODE)) {
5496 pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5497 /* happens iff arch is bonkers, let's just proceed */
5498 return;
5499 }
5500 cpumask_set_cpu(cpu, tbl[node]);
5501 }
5502
5503 wq_numa_possible_cpumask = tbl;
5504 wq_numa_enabled = true;
5505}
5506
5507static int __init init_workqueues(void)
5508{
5509 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5510 int i, cpu;
5511
5512 WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5513
5514 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5515 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
5516
5517 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5518
5519 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5520 hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5521
5522 wq_numa_init();
5523
5524 /* initialize CPU pools */
5525 for_each_possible_cpu(cpu) {
5526 struct worker_pool *pool;
5527
5528 i = 0;
5529 for_each_cpu_worker_pool(pool, cpu) {
5530 BUG_ON(init_worker_pool(pool));
5531 pool->cpu = cpu;
5532 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5533 pool->attrs->nice = std_nice[i++];
5534 pool->node = cpu_to_node(cpu);
5535
5536 /* alloc pool ID */
5537 mutex_lock(&wq_pool_mutex);
5538 BUG_ON(worker_pool_assign_id(pool));
5539 mutex_unlock(&wq_pool_mutex);
5540 }
5541 }
5542
5543 /* create the initial worker */
5544 for_each_online_cpu(cpu) {
5545 struct worker_pool *pool;
5546
5547 for_each_cpu_worker_pool(pool, cpu) {
5548 pool->flags &= ~POOL_DISASSOCIATED;
5549 BUG_ON(!create_worker(pool));
5550 }
5551 }
5552
5553 /* create default unbound and ordered wq attrs */
5554 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5555 struct workqueue_attrs *attrs;
5556
5557 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5558 attrs->nice = std_nice[i];
5559 unbound_std_wq_attrs[i] = attrs;
5560
5561 /*
5562 * An ordered wq should have only one pwq as ordering is
5563 * guaranteed by max_active which is enforced by pwqs.
5564 * Turn off NUMA so that dfl_pwq is used for all nodes.
5565 */
5566 BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5567 attrs->nice = std_nice[i];
5568 attrs->no_numa = true;
5569 ordered_wq_attrs[i] = attrs;
5570 }
5571
5572 system_wq = alloc_workqueue("events", 0, 0);
5573 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5574 system_long_wq = alloc_workqueue("events_long", 0, 0);
5575 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5576 WQ_UNBOUND_MAX_ACTIVE);
5577 system_freezable_wq = alloc_workqueue("events_freezable",
5578 WQ_FREEZABLE, 0);
5579 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5580 WQ_POWER_EFFICIENT, 0);
5581 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5582 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5583 0);
5584 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5585 !system_unbound_wq || !system_freezable_wq ||
5586 !system_power_efficient_wq ||
5587 !system_freezable_power_efficient_wq);
5588
5589 wq_watchdog_init();
5590
5591 return 0;
5592}
5593early_initcall(init_workqueues);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * kernel/workqueue.c - generic async execution with shared worker pool
4 *
5 * Copyright (C) 2002 Ingo Molnar
6 *
7 * Derived from the taskqueue/keventd code by:
8 * David Woodhouse <dwmw2@infradead.org>
9 * Andrew Morton
10 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
11 * Theodore Ts'o <tytso@mit.edu>
12 *
13 * Made to use alloc_percpu by Christoph Lameter.
14 *
15 * Copyright (C) 2010 SUSE Linux Products GmbH
16 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
17 *
18 * This is the generic async execution mechanism. Work items as are
19 * executed in process context. The worker pool is shared and
20 * automatically managed. There are two worker pools for each CPU (one for
21 * normal work items and the other for high priority ones) and some extra
22 * pools for workqueues which are not bound to any specific CPU - the
23 * number of these backing pools is dynamic.
24 *
25 * Please read Documentation/core-api/workqueue.rst for details.
26 */
27
28#include <linux/export.h>
29#include <linux/kernel.h>
30#include <linux/sched.h>
31#include <linux/init.h>
32#include <linux/interrupt.h>
33#include <linux/signal.h>
34#include <linux/completion.h>
35#include <linux/workqueue.h>
36#include <linux/slab.h>
37#include <linux/cpu.h>
38#include <linux/notifier.h>
39#include <linux/kthread.h>
40#include <linux/hardirq.h>
41#include <linux/mempolicy.h>
42#include <linux/freezer.h>
43#include <linux/debug_locks.h>
44#include <linux/lockdep.h>
45#include <linux/idr.h>
46#include <linux/jhash.h>
47#include <linux/hashtable.h>
48#include <linux/rculist.h>
49#include <linux/nodemask.h>
50#include <linux/moduleparam.h>
51#include <linux/uaccess.h>
52#include <linux/sched/isolation.h>
53#include <linux/sched/debug.h>
54#include <linux/nmi.h>
55#include <linux/kvm_para.h>
56#include <linux/delay.h>
57#include <linux/irq_work.h>
58
59#include "workqueue_internal.h"
60
61enum worker_pool_flags {
62 /*
63 * worker_pool flags
64 *
65 * A bound pool is either associated or disassociated with its CPU.
66 * While associated (!DISASSOCIATED), all workers are bound to the
67 * CPU and none has %WORKER_UNBOUND set and concurrency management
68 * is in effect.
69 *
70 * While DISASSOCIATED, the cpu may be offline and all workers have
71 * %WORKER_UNBOUND set and concurrency management disabled, and may
72 * be executing on any CPU. The pool behaves as an unbound one.
73 *
74 * Note that DISASSOCIATED should be flipped only while holding
75 * wq_pool_attach_mutex to avoid changing binding state while
76 * worker_attach_to_pool() is in progress.
77 *
78 * As there can only be one concurrent BH execution context per CPU, a
79 * BH pool is per-CPU and always DISASSOCIATED.
80 */
81 POOL_BH = 1 << 0, /* is a BH pool */
82 POOL_MANAGER_ACTIVE = 1 << 1, /* being managed */
83 POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
84 POOL_BH_DRAINING = 1 << 3, /* draining after CPU offline */
85};
86
87enum worker_flags {
88 /* worker flags */
89 WORKER_DIE = 1 << 1, /* die die die */
90 WORKER_IDLE = 1 << 2, /* is idle */
91 WORKER_PREP = 1 << 3, /* preparing to run works */
92 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
93 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
94 WORKER_REBOUND = 1 << 8, /* worker was rebound */
95
96 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
97 WORKER_UNBOUND | WORKER_REBOUND,
98};
99
100enum work_cancel_flags {
101 WORK_CANCEL_DELAYED = 1 << 0, /* canceling a delayed_work */
102 WORK_CANCEL_DISABLE = 1 << 1, /* canceling to disable */
103};
104
105enum wq_internal_consts {
106 NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
107
108 UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
109 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
110
111 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
112 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
113
114 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
115 /* call for help after 10ms
116 (min two ticks) */
117 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
118 CREATE_COOLDOWN = HZ, /* time to breath after fail */
119
120 /*
121 * Rescue workers are used only on emergencies and shared by
122 * all cpus. Give MIN_NICE.
123 */
124 RESCUER_NICE_LEVEL = MIN_NICE,
125 HIGHPRI_NICE_LEVEL = MIN_NICE,
126
127 WQ_NAME_LEN = 32,
128 WORKER_ID_LEN = 10 + WQ_NAME_LEN, /* "kworker/R-" + WQ_NAME_LEN */
129};
130
131/*
132 * We don't want to trap softirq for too long. See MAX_SOFTIRQ_TIME and
133 * MAX_SOFTIRQ_RESTART in kernel/softirq.c. These are macros because
134 * msecs_to_jiffies() can't be an initializer.
135 */
136#define BH_WORKER_JIFFIES msecs_to_jiffies(2)
137#define BH_WORKER_RESTARTS 10
138
139/*
140 * Structure fields follow one of the following exclusion rules.
141 *
142 * I: Modifiable by initialization/destruction paths and read-only for
143 * everyone else.
144 *
145 * P: Preemption protected. Disabling preemption is enough and should
146 * only be modified and accessed from the local cpu.
147 *
148 * L: pool->lock protected. Access with pool->lock held.
149 *
150 * LN: pool->lock and wq_node_nr_active->lock protected for writes. Either for
151 * reads.
152 *
153 * K: Only modified by worker while holding pool->lock. Can be safely read by
154 * self, while holding pool->lock or from IRQ context if %current is the
155 * kworker.
156 *
157 * S: Only modified by worker self.
158 *
159 * A: wq_pool_attach_mutex protected.
160 *
161 * PL: wq_pool_mutex protected.
162 *
163 * PR: wq_pool_mutex protected for writes. RCU protected for reads.
164 *
165 * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads.
166 *
167 * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or
168 * RCU for reads.
169 *
170 * WQ: wq->mutex protected.
171 *
172 * WR: wq->mutex protected for writes. RCU protected for reads.
173 *
174 * WO: wq->mutex protected for writes. Updated with WRITE_ONCE() and can be read
175 * with READ_ONCE() without locking.
176 *
177 * MD: wq_mayday_lock protected.
178 *
179 * WD: Used internally by the watchdog.
180 */
181
182/* struct worker is defined in workqueue_internal.h */
183
184struct worker_pool {
185 raw_spinlock_t lock; /* the pool lock */
186 int cpu; /* I: the associated cpu */
187 int node; /* I: the associated node ID */
188 int id; /* I: pool ID */
189 unsigned int flags; /* L: flags */
190
191 unsigned long watchdog_ts; /* L: watchdog timestamp */
192 bool cpu_stall; /* WD: stalled cpu bound pool */
193
194 /*
195 * The counter is incremented in a process context on the associated CPU
196 * w/ preemption disabled, and decremented or reset in the same context
197 * but w/ pool->lock held. The readers grab pool->lock and are
198 * guaranteed to see if the counter reached zero.
199 */
200 int nr_running;
201
202 struct list_head worklist; /* L: list of pending works */
203
204 int nr_workers; /* L: total number of workers */
205 int nr_idle; /* L: currently idle workers */
206
207 struct list_head idle_list; /* L: list of idle workers */
208 struct timer_list idle_timer; /* L: worker idle timeout */
209 struct work_struct idle_cull_work; /* L: worker idle cleanup */
210
211 struct timer_list mayday_timer; /* L: SOS timer for workers */
212
213 /* a workers is either on busy_hash or idle_list, or the manager */
214 DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
215 /* L: hash of busy workers */
216
217 struct worker *manager; /* L: purely informational */
218 struct list_head workers; /* A: attached workers */
219
220 struct ida worker_ida; /* worker IDs for task name */
221
222 struct workqueue_attrs *attrs; /* I: worker attributes */
223 struct hlist_node hash_node; /* PL: unbound_pool_hash node */
224 int refcnt; /* PL: refcnt for unbound pools */
225
226 /*
227 * Destruction of pool is RCU protected to allow dereferences
228 * from get_work_pool().
229 */
230 struct rcu_head rcu;
231};
232
233/*
234 * Per-pool_workqueue statistics. These can be monitored using
235 * tools/workqueue/wq_monitor.py.
236 */
237enum pool_workqueue_stats {
238 PWQ_STAT_STARTED, /* work items started execution */
239 PWQ_STAT_COMPLETED, /* work items completed execution */
240 PWQ_STAT_CPU_TIME, /* total CPU time consumed */
241 PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */
242 PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */
243 PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */
244 PWQ_STAT_MAYDAY, /* maydays to rescuer */
245 PWQ_STAT_RESCUED, /* linked work items executed by rescuer */
246
247 PWQ_NR_STATS,
248};
249
250/*
251 * The per-pool workqueue. While queued, bits below WORK_PWQ_SHIFT
252 * of work_struct->data are used for flags and the remaining high bits
253 * point to the pwq; thus, pwqs need to be aligned at two's power of the
254 * number of flag bits.
255 */
256struct pool_workqueue {
257 struct worker_pool *pool; /* I: the associated pool */
258 struct workqueue_struct *wq; /* I: the owning workqueue */
259 int work_color; /* L: current color */
260 int flush_color; /* L: flushing color */
261 int refcnt; /* L: reference count */
262 int nr_in_flight[WORK_NR_COLORS];
263 /* L: nr of in_flight works */
264 bool plugged; /* L: execution suspended */
265
266 /*
267 * nr_active management and WORK_STRUCT_INACTIVE:
268 *
269 * When pwq->nr_active >= max_active, new work item is queued to
270 * pwq->inactive_works instead of pool->worklist and marked with
271 * WORK_STRUCT_INACTIVE.
272 *
273 * All work items marked with WORK_STRUCT_INACTIVE do not participate in
274 * nr_active and all work items in pwq->inactive_works are marked with
275 * WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE work items are
276 * in pwq->inactive_works. Some of them are ready to run in
277 * pool->worklist or worker->scheduled. Those work itmes are only struct
278 * wq_barrier which is used for flush_work() and should not participate
279 * in nr_active. For non-barrier work item, it is marked with
280 * WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works.
281 */
282 int nr_active; /* L: nr of active works */
283 struct list_head inactive_works; /* L: inactive works */
284 struct list_head pending_node; /* LN: node on wq_node_nr_active->pending_pwqs */
285 struct list_head pwqs_node; /* WR: node on wq->pwqs */
286 struct list_head mayday_node; /* MD: node on wq->maydays */
287
288 u64 stats[PWQ_NR_STATS];
289
290 /*
291 * Release of unbound pwq is punted to a kthread_worker. See put_pwq()
292 * and pwq_release_workfn() for details. pool_workqueue itself is also
293 * RCU protected so that the first pwq can be determined without
294 * grabbing wq->mutex.
295 */
296 struct kthread_work release_work;
297 struct rcu_head rcu;
298} __aligned(1 << WORK_STRUCT_PWQ_SHIFT);
299
300/*
301 * Structure used to wait for workqueue flush.
302 */
303struct wq_flusher {
304 struct list_head list; /* WQ: list of flushers */
305 int flush_color; /* WQ: flush color waiting for */
306 struct completion done; /* flush completion */
307};
308
309struct wq_device;
310
311/*
312 * Unlike in a per-cpu workqueue where max_active limits its concurrency level
313 * on each CPU, in an unbound workqueue, max_active applies to the whole system.
314 * As sharing a single nr_active across multiple sockets can be very expensive,
315 * the counting and enforcement is per NUMA node.
316 *
317 * The following struct is used to enforce per-node max_active. When a pwq wants
318 * to start executing a work item, it should increment ->nr using
319 * tryinc_node_nr_active(). If acquisition fails due to ->nr already being over
320 * ->max, the pwq is queued on ->pending_pwqs. As in-flight work items finish
321 * and decrement ->nr, node_activate_pending_pwq() activates the pending pwqs in
322 * round-robin order.
323 */
324struct wq_node_nr_active {
325 int max; /* per-node max_active */
326 atomic_t nr; /* per-node nr_active */
327 raw_spinlock_t lock; /* nests inside pool locks */
328 struct list_head pending_pwqs; /* LN: pwqs with inactive works */
329};
330
331/*
332 * The externally visible workqueue. It relays the issued work items to
333 * the appropriate worker_pool through its pool_workqueues.
334 */
335struct workqueue_struct {
336 struct list_head pwqs; /* WR: all pwqs of this wq */
337 struct list_head list; /* PR: list of all workqueues */
338
339 struct mutex mutex; /* protects this wq */
340 int work_color; /* WQ: current work color */
341 int flush_color; /* WQ: current flush color */
342 atomic_t nr_pwqs_to_flush; /* flush in progress */
343 struct wq_flusher *first_flusher; /* WQ: first flusher */
344 struct list_head flusher_queue; /* WQ: flush waiters */
345 struct list_head flusher_overflow; /* WQ: flush overflow list */
346
347 struct list_head maydays; /* MD: pwqs requesting rescue */
348 struct worker *rescuer; /* MD: rescue worker */
349
350 int nr_drainers; /* WQ: drain in progress */
351
352 /* See alloc_workqueue() function comment for info on min/max_active */
353 int max_active; /* WO: max active works */
354 int min_active; /* WO: min active works */
355 int saved_max_active; /* WQ: saved max_active */
356 int saved_min_active; /* WQ: saved min_active */
357
358 struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */
359 struct pool_workqueue __rcu *dfl_pwq; /* PW: only for unbound wqs */
360
361#ifdef CONFIG_SYSFS
362 struct wq_device *wq_dev; /* I: for sysfs interface */
363#endif
364#ifdef CONFIG_LOCKDEP
365 char *lock_name;
366 struct lock_class_key key;
367 struct lockdep_map __lockdep_map;
368 struct lockdep_map *lockdep_map;
369#endif
370 char name[WQ_NAME_LEN]; /* I: workqueue name */
371
372 /*
373 * Destruction of workqueue_struct is RCU protected to allow walking
374 * the workqueues list without grabbing wq_pool_mutex.
375 * This is used to dump all workqueues from sysrq.
376 */
377 struct rcu_head rcu;
378
379 /* hot fields used during command issue, aligned to cacheline */
380 unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
381 struct pool_workqueue __rcu * __percpu *cpu_pwq; /* I: per-cpu pwqs */
382 struct wq_node_nr_active *node_nr_active[]; /* I: per-node nr_active */
383};
384
385/*
386 * Each pod type describes how CPUs should be grouped for unbound workqueues.
387 * See the comment above workqueue_attrs->affn_scope.
388 */
389struct wq_pod_type {
390 int nr_pods; /* number of pods */
391 cpumask_var_t *pod_cpus; /* pod -> cpus */
392 int *pod_node; /* pod -> node */
393 int *cpu_pod; /* cpu -> pod */
394};
395
396struct work_offq_data {
397 u32 pool_id;
398 u32 disable;
399 u32 flags;
400};
401
402static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = {
403 [WQ_AFFN_DFL] = "default",
404 [WQ_AFFN_CPU] = "cpu",
405 [WQ_AFFN_SMT] = "smt",
406 [WQ_AFFN_CACHE] = "cache",
407 [WQ_AFFN_NUMA] = "numa",
408 [WQ_AFFN_SYSTEM] = "system",
409};
410
411/*
412 * Per-cpu work items which run for longer than the following threshold are
413 * automatically considered CPU intensive and excluded from concurrency
414 * management to prevent them from noticeably delaying other per-cpu work items.
415 * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter.
416 * The actual value is initialized in wq_cpu_intensive_thresh_init().
417 */
418static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX;
419module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644);
420#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
421static unsigned int wq_cpu_intensive_warning_thresh = 4;
422module_param_named(cpu_intensive_warning_thresh, wq_cpu_intensive_warning_thresh, uint, 0644);
423#endif
424
425/* see the comment above the definition of WQ_POWER_EFFICIENT */
426static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
427module_param_named(power_efficient, wq_power_efficient, bool, 0444);
428
429static bool wq_online; /* can kworkers be created yet? */
430static bool wq_topo_initialized __read_mostly = false;
431
432static struct kmem_cache *pwq_cache;
433
434static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES];
435static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE;
436
437/* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */
438static struct workqueue_attrs *unbound_wq_update_pwq_attrs_buf;
439
440static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
441static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
442static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
443/* wait for manager to go away */
444static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait);
445
446static LIST_HEAD(workqueues); /* PR: list of all workqueues */
447static bool workqueue_freezing; /* PL: have wqs started freezing? */
448
449/* PL: mirror the cpu_online_mask excluding the CPU in the midst of hotplugging */
450static cpumask_var_t wq_online_cpumask;
451
452/* PL&A: allowable cpus for unbound wqs and work items */
453static cpumask_var_t wq_unbound_cpumask;
454
455/* PL: user requested unbound cpumask via sysfs */
456static cpumask_var_t wq_requested_unbound_cpumask;
457
458/* PL: isolated cpumask to be excluded from unbound cpumask */
459static cpumask_var_t wq_isolated_cpumask;
460
461/* for further constrain wq_unbound_cpumask by cmdline parameter*/
462static struct cpumask wq_cmdline_cpumask __initdata;
463
464/* CPU where unbound work was last round robin scheduled from this CPU */
465static DEFINE_PER_CPU(int, wq_rr_cpu_last);
466
467/*
468 * Local execution of unbound work items is no longer guaranteed. The
469 * following always forces round-robin CPU selection on unbound work items
470 * to uncover usages which depend on it.
471 */
472#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
473static bool wq_debug_force_rr_cpu = true;
474#else
475static bool wq_debug_force_rr_cpu = false;
476#endif
477module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
478
479/* to raise softirq for the BH worker pools on other CPUs */
480static DEFINE_PER_CPU_SHARED_ALIGNED(struct irq_work [NR_STD_WORKER_POOLS], bh_pool_irq_works);
481
482/* the BH worker pools */
483static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], bh_worker_pools);
484
485/* the per-cpu worker pools */
486static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
487
488static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
489
490/* PL: hash of all unbound pools keyed by pool->attrs */
491static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
492
493/* I: attributes used when instantiating standard unbound pools on demand */
494static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
495
496/* I: attributes used when instantiating ordered pools on demand */
497static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
498
499/*
500 * I: kthread_worker to release pwq's. pwq release needs to be bounced to a
501 * process context while holding a pool lock. Bounce to a dedicated kthread
502 * worker to avoid A-A deadlocks.
503 */
504static struct kthread_worker *pwq_release_worker __ro_after_init;
505
506struct workqueue_struct *system_wq __ro_after_init;
507EXPORT_SYMBOL(system_wq);
508struct workqueue_struct *system_highpri_wq __ro_after_init;
509EXPORT_SYMBOL_GPL(system_highpri_wq);
510struct workqueue_struct *system_long_wq __ro_after_init;
511EXPORT_SYMBOL_GPL(system_long_wq);
512struct workqueue_struct *system_unbound_wq __ro_after_init;
513EXPORT_SYMBOL_GPL(system_unbound_wq);
514struct workqueue_struct *system_freezable_wq __ro_after_init;
515EXPORT_SYMBOL_GPL(system_freezable_wq);
516struct workqueue_struct *system_power_efficient_wq __ro_after_init;
517EXPORT_SYMBOL_GPL(system_power_efficient_wq);
518struct workqueue_struct *system_freezable_power_efficient_wq __ro_after_init;
519EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
520struct workqueue_struct *system_bh_wq;
521EXPORT_SYMBOL_GPL(system_bh_wq);
522struct workqueue_struct *system_bh_highpri_wq;
523EXPORT_SYMBOL_GPL(system_bh_highpri_wq);
524
525static int worker_thread(void *__worker);
526static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
527static void show_pwq(struct pool_workqueue *pwq);
528static void show_one_worker_pool(struct worker_pool *pool);
529
530#define CREATE_TRACE_POINTS
531#include <trace/events/workqueue.h>
532
533#define assert_rcu_or_pool_mutex() \
534 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
535 !lockdep_is_held(&wq_pool_mutex), \
536 "RCU or wq_pool_mutex should be held")
537
538#define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
539 RCU_LOCKDEP_WARN(!rcu_read_lock_any_held() && \
540 !lockdep_is_held(&wq->mutex) && \
541 !lockdep_is_held(&wq_pool_mutex), \
542 "RCU, wq->mutex or wq_pool_mutex should be held")
543
544#define for_each_bh_worker_pool(pool, cpu) \
545 for ((pool) = &per_cpu(bh_worker_pools, cpu)[0]; \
546 (pool) < &per_cpu(bh_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
547 (pool)++)
548
549#define for_each_cpu_worker_pool(pool, cpu) \
550 for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
551 (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
552 (pool)++)
553
554/**
555 * for_each_pool - iterate through all worker_pools in the system
556 * @pool: iteration cursor
557 * @pi: integer used for iteration
558 *
559 * This must be called either with wq_pool_mutex held or RCU read
560 * locked. If the pool needs to be used beyond the locking in effect, the
561 * caller is responsible for guaranteeing that the pool stays online.
562 *
563 * The if/else clause exists only for the lockdep assertion and can be
564 * ignored.
565 */
566#define for_each_pool(pool, pi) \
567 idr_for_each_entry(&worker_pool_idr, pool, pi) \
568 if (({ assert_rcu_or_pool_mutex(); false; })) { } \
569 else
570
571/**
572 * for_each_pool_worker - iterate through all workers of a worker_pool
573 * @worker: iteration cursor
574 * @pool: worker_pool to iterate workers of
575 *
576 * This must be called with wq_pool_attach_mutex.
577 *
578 * The if/else clause exists only for the lockdep assertion and can be
579 * ignored.
580 */
581#define for_each_pool_worker(worker, pool) \
582 list_for_each_entry((worker), &(pool)->workers, node) \
583 if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
584 else
585
586/**
587 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
588 * @pwq: iteration cursor
589 * @wq: the target workqueue
590 *
591 * This must be called either with wq->mutex held or RCU read locked.
592 * If the pwq needs to be used beyond the locking in effect, the caller is
593 * responsible for guaranteeing that the pwq stays online.
594 *
595 * The if/else clause exists only for the lockdep assertion and can be
596 * ignored.
597 */
598#define for_each_pwq(pwq, wq) \
599 list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \
600 lockdep_is_held(&(wq->mutex)))
601
602#ifdef CONFIG_DEBUG_OBJECTS_WORK
603
604static const struct debug_obj_descr work_debug_descr;
605
606static void *work_debug_hint(void *addr)
607{
608 return ((struct work_struct *) addr)->func;
609}
610
611static bool work_is_static_object(void *addr)
612{
613 struct work_struct *work = addr;
614
615 return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
616}
617
618/*
619 * fixup_init is called when:
620 * - an active object is initialized
621 */
622static bool work_fixup_init(void *addr, enum debug_obj_state state)
623{
624 struct work_struct *work = addr;
625
626 switch (state) {
627 case ODEBUG_STATE_ACTIVE:
628 cancel_work_sync(work);
629 debug_object_init(work, &work_debug_descr);
630 return true;
631 default:
632 return false;
633 }
634}
635
636/*
637 * fixup_free is called when:
638 * - an active object is freed
639 */
640static bool work_fixup_free(void *addr, enum debug_obj_state state)
641{
642 struct work_struct *work = addr;
643
644 switch (state) {
645 case ODEBUG_STATE_ACTIVE:
646 cancel_work_sync(work);
647 debug_object_free(work, &work_debug_descr);
648 return true;
649 default:
650 return false;
651 }
652}
653
654static const struct debug_obj_descr work_debug_descr = {
655 .name = "work_struct",
656 .debug_hint = work_debug_hint,
657 .is_static_object = work_is_static_object,
658 .fixup_init = work_fixup_init,
659 .fixup_free = work_fixup_free,
660};
661
662static inline void debug_work_activate(struct work_struct *work)
663{
664 debug_object_activate(work, &work_debug_descr);
665}
666
667static inline void debug_work_deactivate(struct work_struct *work)
668{
669 debug_object_deactivate(work, &work_debug_descr);
670}
671
672void __init_work(struct work_struct *work, int onstack)
673{
674 if (onstack)
675 debug_object_init_on_stack(work, &work_debug_descr);
676 else
677 debug_object_init(work, &work_debug_descr);
678}
679EXPORT_SYMBOL_GPL(__init_work);
680
681void destroy_work_on_stack(struct work_struct *work)
682{
683 debug_object_free(work, &work_debug_descr);
684}
685EXPORT_SYMBOL_GPL(destroy_work_on_stack);
686
687void destroy_delayed_work_on_stack(struct delayed_work *work)
688{
689 destroy_timer_on_stack(&work->timer);
690 debug_object_free(&work->work, &work_debug_descr);
691}
692EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
693
694#else
695static inline void debug_work_activate(struct work_struct *work) { }
696static inline void debug_work_deactivate(struct work_struct *work) { }
697#endif
698
699/**
700 * worker_pool_assign_id - allocate ID and assign it to @pool
701 * @pool: the pool pointer of interest
702 *
703 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
704 * successfully, -errno on failure.
705 */
706static int worker_pool_assign_id(struct worker_pool *pool)
707{
708 int ret;
709
710 lockdep_assert_held(&wq_pool_mutex);
711
712 ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
713 GFP_KERNEL);
714 if (ret >= 0) {
715 pool->id = ret;
716 return 0;
717 }
718 return ret;
719}
720
721static struct pool_workqueue __rcu **
722unbound_pwq_slot(struct workqueue_struct *wq, int cpu)
723{
724 if (cpu >= 0)
725 return per_cpu_ptr(wq->cpu_pwq, cpu);
726 else
727 return &wq->dfl_pwq;
728}
729
730/* @cpu < 0 for dfl_pwq */
731static struct pool_workqueue *unbound_pwq(struct workqueue_struct *wq, int cpu)
732{
733 return rcu_dereference_check(*unbound_pwq_slot(wq, cpu),
734 lockdep_is_held(&wq_pool_mutex) ||
735 lockdep_is_held(&wq->mutex));
736}
737
738/**
739 * unbound_effective_cpumask - effective cpumask of an unbound workqueue
740 * @wq: workqueue of interest
741 *
742 * @wq->unbound_attrs->cpumask contains the cpumask requested by the user which
743 * is masked with wq_unbound_cpumask to determine the effective cpumask. The
744 * default pwq is always mapped to the pool with the current effective cpumask.
745 */
746static struct cpumask *unbound_effective_cpumask(struct workqueue_struct *wq)
747{
748 return unbound_pwq(wq, -1)->pool->attrs->__pod_cpumask;
749}
750
751static unsigned int work_color_to_flags(int color)
752{
753 return color << WORK_STRUCT_COLOR_SHIFT;
754}
755
756static int get_work_color(unsigned long work_data)
757{
758 return (work_data >> WORK_STRUCT_COLOR_SHIFT) &
759 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
760}
761
762static int work_next_color(int color)
763{
764 return (color + 1) % WORK_NR_COLORS;
765}
766
767static unsigned long pool_offq_flags(struct worker_pool *pool)
768{
769 return (pool->flags & POOL_BH) ? WORK_OFFQ_BH : 0;
770}
771
772/*
773 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
774 * contain the pointer to the queued pwq. Once execution starts, the flag
775 * is cleared and the high bits contain OFFQ flags and pool ID.
776 *
777 * set_work_pwq(), set_work_pool_and_clear_pending() and mark_work_canceling()
778 * can be used to set the pwq, pool or clear work->data. These functions should
779 * only be called while the work is owned - ie. while the PENDING bit is set.
780 *
781 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
782 * corresponding to a work. Pool is available once the work has been
783 * queued anywhere after initialization until it is sync canceled. pwq is
784 * available only while the work item is queued.
785 */
786static inline void set_work_data(struct work_struct *work, unsigned long data)
787{
788 WARN_ON_ONCE(!work_pending(work));
789 atomic_long_set(&work->data, data | work_static(work));
790}
791
792static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
793 unsigned long flags)
794{
795 set_work_data(work, (unsigned long)pwq | WORK_STRUCT_PENDING |
796 WORK_STRUCT_PWQ | flags);
797}
798
799static void set_work_pool_and_keep_pending(struct work_struct *work,
800 int pool_id, unsigned long flags)
801{
802 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
803 WORK_STRUCT_PENDING | flags);
804}
805
806static void set_work_pool_and_clear_pending(struct work_struct *work,
807 int pool_id, unsigned long flags)
808{
809 /*
810 * The following wmb is paired with the implied mb in
811 * test_and_set_bit(PENDING) and ensures all updates to @work made
812 * here are visible to and precede any updates by the next PENDING
813 * owner.
814 */
815 smp_wmb();
816 set_work_data(work, ((unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT) |
817 flags);
818 /*
819 * The following mb guarantees that previous clear of a PENDING bit
820 * will not be reordered with any speculative LOADS or STORES from
821 * work->current_func, which is executed afterwards. This possible
822 * reordering can lead to a missed execution on attempt to queue
823 * the same @work. E.g. consider this case:
824 *
825 * CPU#0 CPU#1
826 * ---------------------------- --------------------------------
827 *
828 * 1 STORE event_indicated
829 * 2 queue_work_on() {
830 * 3 test_and_set_bit(PENDING)
831 * 4 } set_..._and_clear_pending() {
832 * 5 set_work_data() # clear bit
833 * 6 smp_mb()
834 * 7 work->current_func() {
835 * 8 LOAD event_indicated
836 * }
837 *
838 * Without an explicit full barrier speculative LOAD on line 8 can
839 * be executed before CPU#0 does STORE on line 1. If that happens,
840 * CPU#0 observes the PENDING bit is still set and new execution of
841 * a @work is not queued in a hope, that CPU#1 will eventually
842 * finish the queued @work. Meanwhile CPU#1 does not see
843 * event_indicated is set, because speculative LOAD was executed
844 * before actual STORE.
845 */
846 smp_mb();
847}
848
849static inline struct pool_workqueue *work_struct_pwq(unsigned long data)
850{
851 return (struct pool_workqueue *)(data & WORK_STRUCT_PWQ_MASK);
852}
853
854static struct pool_workqueue *get_work_pwq(struct work_struct *work)
855{
856 unsigned long data = atomic_long_read(&work->data);
857
858 if (data & WORK_STRUCT_PWQ)
859 return work_struct_pwq(data);
860 else
861 return NULL;
862}
863
864/**
865 * get_work_pool - return the worker_pool a given work was associated with
866 * @work: the work item of interest
867 *
868 * Pools are created and destroyed under wq_pool_mutex, and allows read
869 * access under RCU read lock. As such, this function should be
870 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
871 *
872 * All fields of the returned pool are accessible as long as the above
873 * mentioned locking is in effect. If the returned pool needs to be used
874 * beyond the critical section, the caller is responsible for ensuring the
875 * returned pool is and stays online.
876 *
877 * Return: The worker_pool @work was last associated with. %NULL if none.
878 */
879static struct worker_pool *get_work_pool(struct work_struct *work)
880{
881 unsigned long data = atomic_long_read(&work->data);
882 int pool_id;
883
884 assert_rcu_or_pool_mutex();
885
886 if (data & WORK_STRUCT_PWQ)
887 return work_struct_pwq(data)->pool;
888
889 pool_id = data >> WORK_OFFQ_POOL_SHIFT;
890 if (pool_id == WORK_OFFQ_POOL_NONE)
891 return NULL;
892
893 return idr_find(&worker_pool_idr, pool_id);
894}
895
896static unsigned long shift_and_mask(unsigned long v, u32 shift, u32 bits)
897{
898 return (v >> shift) & ((1U << bits) - 1);
899}
900
901static void work_offqd_unpack(struct work_offq_data *offqd, unsigned long data)
902{
903 WARN_ON_ONCE(data & WORK_STRUCT_PWQ);
904
905 offqd->pool_id = shift_and_mask(data, WORK_OFFQ_POOL_SHIFT,
906 WORK_OFFQ_POOL_BITS);
907 offqd->disable = shift_and_mask(data, WORK_OFFQ_DISABLE_SHIFT,
908 WORK_OFFQ_DISABLE_BITS);
909 offqd->flags = data & WORK_OFFQ_FLAG_MASK;
910}
911
912static unsigned long work_offqd_pack_flags(struct work_offq_data *offqd)
913{
914 return ((unsigned long)offqd->disable << WORK_OFFQ_DISABLE_SHIFT) |
915 ((unsigned long)offqd->flags);
916}
917
918/*
919 * Policy functions. These define the policies on how the global worker
920 * pools are managed. Unless noted otherwise, these functions assume that
921 * they're being called with pool->lock held.
922 */
923
924/*
925 * Need to wake up a worker? Called from anything but currently
926 * running workers.
927 *
928 * Note that, because unbound workers never contribute to nr_running, this
929 * function will always return %true for unbound pools as long as the
930 * worklist isn't empty.
931 */
932static bool need_more_worker(struct worker_pool *pool)
933{
934 return !list_empty(&pool->worklist) && !pool->nr_running;
935}
936
937/* Can I start working? Called from busy but !running workers. */
938static bool may_start_working(struct worker_pool *pool)
939{
940 return pool->nr_idle;
941}
942
943/* Do I need to keep working? Called from currently running workers. */
944static bool keep_working(struct worker_pool *pool)
945{
946 return !list_empty(&pool->worklist) && (pool->nr_running <= 1);
947}
948
949/* Do we need a new worker? Called from manager. */
950static bool need_to_create_worker(struct worker_pool *pool)
951{
952 return need_more_worker(pool) && !may_start_working(pool);
953}
954
955/* Do we have too many workers and should some go away? */
956static bool too_many_workers(struct worker_pool *pool)
957{
958 bool managing = pool->flags & POOL_MANAGER_ACTIVE;
959 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
960 int nr_busy = pool->nr_workers - nr_idle;
961
962 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
963}
964
965/**
966 * worker_set_flags - set worker flags and adjust nr_running accordingly
967 * @worker: self
968 * @flags: flags to set
969 *
970 * Set @flags in @worker->flags and adjust nr_running accordingly.
971 */
972static inline void worker_set_flags(struct worker *worker, unsigned int flags)
973{
974 struct worker_pool *pool = worker->pool;
975
976 lockdep_assert_held(&pool->lock);
977
978 /* If transitioning into NOT_RUNNING, adjust nr_running. */
979 if ((flags & WORKER_NOT_RUNNING) &&
980 !(worker->flags & WORKER_NOT_RUNNING)) {
981 pool->nr_running--;
982 }
983
984 worker->flags |= flags;
985}
986
987/**
988 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
989 * @worker: self
990 * @flags: flags to clear
991 *
992 * Clear @flags in @worker->flags and adjust nr_running accordingly.
993 */
994static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
995{
996 struct worker_pool *pool = worker->pool;
997 unsigned int oflags = worker->flags;
998
999 lockdep_assert_held(&pool->lock);
1000
1001 worker->flags &= ~flags;
1002
1003 /*
1004 * If transitioning out of NOT_RUNNING, increment nr_running. Note
1005 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
1006 * of multiple flags, not a single flag.
1007 */
1008 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
1009 if (!(worker->flags & WORKER_NOT_RUNNING))
1010 pool->nr_running++;
1011}
1012
1013/* Return the first idle worker. Called with pool->lock held. */
1014static struct worker *first_idle_worker(struct worker_pool *pool)
1015{
1016 if (unlikely(list_empty(&pool->idle_list)))
1017 return NULL;
1018
1019 return list_first_entry(&pool->idle_list, struct worker, entry);
1020}
1021
1022/**
1023 * worker_enter_idle - enter idle state
1024 * @worker: worker which is entering idle state
1025 *
1026 * @worker is entering idle state. Update stats and idle timer if
1027 * necessary.
1028 *
1029 * LOCKING:
1030 * raw_spin_lock_irq(pool->lock).
1031 */
1032static void worker_enter_idle(struct worker *worker)
1033{
1034 struct worker_pool *pool = worker->pool;
1035
1036 if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1037 WARN_ON_ONCE(!list_empty(&worker->entry) &&
1038 (worker->hentry.next || worker->hentry.pprev)))
1039 return;
1040
1041 /* can't use worker_set_flags(), also called from create_worker() */
1042 worker->flags |= WORKER_IDLE;
1043 pool->nr_idle++;
1044 worker->last_active = jiffies;
1045
1046 /* idle_list is LIFO */
1047 list_add(&worker->entry, &pool->idle_list);
1048
1049 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1050 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1051
1052 /* Sanity check nr_running. */
1053 WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running);
1054}
1055
1056/**
1057 * worker_leave_idle - leave idle state
1058 * @worker: worker which is leaving idle state
1059 *
1060 * @worker is leaving idle state. Update stats.
1061 *
1062 * LOCKING:
1063 * raw_spin_lock_irq(pool->lock).
1064 */
1065static void worker_leave_idle(struct worker *worker)
1066{
1067 struct worker_pool *pool = worker->pool;
1068
1069 if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1070 return;
1071 worker_clr_flags(worker, WORKER_IDLE);
1072 pool->nr_idle--;
1073 list_del_init(&worker->entry);
1074}
1075
1076/**
1077 * find_worker_executing_work - find worker which is executing a work
1078 * @pool: pool of interest
1079 * @work: work to find worker for
1080 *
1081 * Find a worker which is executing @work on @pool by searching
1082 * @pool->busy_hash which is keyed by the address of @work. For a worker
1083 * to match, its current execution should match the address of @work and
1084 * its work function. This is to avoid unwanted dependency between
1085 * unrelated work executions through a work item being recycled while still
1086 * being executed.
1087 *
1088 * This is a bit tricky. A work item may be freed once its execution
1089 * starts and nothing prevents the freed area from being recycled for
1090 * another work item. If the same work item address ends up being reused
1091 * before the original execution finishes, workqueue will identify the
1092 * recycled work item as currently executing and make it wait until the
1093 * current execution finishes, introducing an unwanted dependency.
1094 *
1095 * This function checks the work item address and work function to avoid
1096 * false positives. Note that this isn't complete as one may construct a
1097 * work function which can introduce dependency onto itself through a
1098 * recycled work item. Well, if somebody wants to shoot oneself in the
1099 * foot that badly, there's only so much we can do, and if such deadlock
1100 * actually occurs, it should be easy to locate the culprit work function.
1101 *
1102 * CONTEXT:
1103 * raw_spin_lock_irq(pool->lock).
1104 *
1105 * Return:
1106 * Pointer to worker which is executing @work if found, %NULL
1107 * otherwise.
1108 */
1109static struct worker *find_worker_executing_work(struct worker_pool *pool,
1110 struct work_struct *work)
1111{
1112 struct worker *worker;
1113
1114 hash_for_each_possible(pool->busy_hash, worker, hentry,
1115 (unsigned long)work)
1116 if (worker->current_work == work &&
1117 worker->current_func == work->func)
1118 return worker;
1119
1120 return NULL;
1121}
1122
1123/**
1124 * move_linked_works - move linked works to a list
1125 * @work: start of series of works to be scheduled
1126 * @head: target list to append @work to
1127 * @nextp: out parameter for nested worklist walking
1128 *
1129 * Schedule linked works starting from @work to @head. Work series to be
1130 * scheduled starts at @work and includes any consecutive work with
1131 * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on
1132 * @nextp.
1133 *
1134 * CONTEXT:
1135 * raw_spin_lock_irq(pool->lock).
1136 */
1137static void move_linked_works(struct work_struct *work, struct list_head *head,
1138 struct work_struct **nextp)
1139{
1140 struct work_struct *n;
1141
1142 /*
1143 * Linked worklist will always end before the end of the list,
1144 * use NULL for list head.
1145 */
1146 list_for_each_entry_safe_from(work, n, NULL, entry) {
1147 list_move_tail(&work->entry, head);
1148 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1149 break;
1150 }
1151
1152 /*
1153 * If we're already inside safe list traversal and have moved
1154 * multiple works to the scheduled queue, the next position
1155 * needs to be updated.
1156 */
1157 if (nextp)
1158 *nextp = n;
1159}
1160
1161/**
1162 * assign_work - assign a work item and its linked work items to a worker
1163 * @work: work to assign
1164 * @worker: worker to assign to
1165 * @nextp: out parameter for nested worklist walking
1166 *
1167 * Assign @work and its linked work items to @worker. If @work is already being
1168 * executed by another worker in the same pool, it'll be punted there.
1169 *
1170 * If @nextp is not NULL, it's updated to point to the next work of the last
1171 * scheduled work. This allows assign_work() to be nested inside
1172 * list_for_each_entry_safe().
1173 *
1174 * Returns %true if @work was successfully assigned to @worker. %false if @work
1175 * was punted to another worker already executing it.
1176 */
1177static bool assign_work(struct work_struct *work, struct worker *worker,
1178 struct work_struct **nextp)
1179{
1180 struct worker_pool *pool = worker->pool;
1181 struct worker *collision;
1182
1183 lockdep_assert_held(&pool->lock);
1184
1185 /*
1186 * A single work shouldn't be executed concurrently by multiple workers.
1187 * __queue_work() ensures that @work doesn't jump to a different pool
1188 * while still running in the previous pool. Here, we should ensure that
1189 * @work is not executed concurrently by multiple workers from the same
1190 * pool. Check whether anyone is already processing the work. If so,
1191 * defer the work to the currently executing one.
1192 */
1193 collision = find_worker_executing_work(pool, work);
1194 if (unlikely(collision)) {
1195 move_linked_works(work, &collision->scheduled, nextp);
1196 return false;
1197 }
1198
1199 move_linked_works(work, &worker->scheduled, nextp);
1200 return true;
1201}
1202
1203static struct irq_work *bh_pool_irq_work(struct worker_pool *pool)
1204{
1205 int high = pool->attrs->nice == HIGHPRI_NICE_LEVEL ? 1 : 0;
1206
1207 return &per_cpu(bh_pool_irq_works, pool->cpu)[high];
1208}
1209
1210static void kick_bh_pool(struct worker_pool *pool)
1211{
1212#ifdef CONFIG_SMP
1213 /* see drain_dead_softirq_workfn() for BH_DRAINING */
1214 if (unlikely(pool->cpu != smp_processor_id() &&
1215 !(pool->flags & POOL_BH_DRAINING))) {
1216 irq_work_queue_on(bh_pool_irq_work(pool), pool->cpu);
1217 return;
1218 }
1219#endif
1220 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
1221 raise_softirq_irqoff(HI_SOFTIRQ);
1222 else
1223 raise_softirq_irqoff(TASKLET_SOFTIRQ);
1224}
1225
1226/**
1227 * kick_pool - wake up an idle worker if necessary
1228 * @pool: pool to kick
1229 *
1230 * @pool may have pending work items. Wake up worker if necessary. Returns
1231 * whether a worker was woken up.
1232 */
1233static bool kick_pool(struct worker_pool *pool)
1234{
1235 struct worker *worker = first_idle_worker(pool);
1236 struct task_struct *p;
1237
1238 lockdep_assert_held(&pool->lock);
1239
1240 if (!need_more_worker(pool) || !worker)
1241 return false;
1242
1243 if (pool->flags & POOL_BH) {
1244 kick_bh_pool(pool);
1245 return true;
1246 }
1247
1248 p = worker->task;
1249
1250#ifdef CONFIG_SMP
1251 /*
1252 * Idle @worker is about to execute @work and waking up provides an
1253 * opportunity to migrate @worker at a lower cost by setting the task's
1254 * wake_cpu field. Let's see if we want to move @worker to improve
1255 * execution locality.
1256 *
1257 * We're waking the worker that went idle the latest and there's some
1258 * chance that @worker is marked idle but hasn't gone off CPU yet. If
1259 * so, setting the wake_cpu won't do anything. As this is a best-effort
1260 * optimization and the race window is narrow, let's leave as-is for
1261 * now. If this becomes pronounced, we can skip over workers which are
1262 * still on cpu when picking an idle worker.
1263 *
1264 * If @pool has non-strict affinity, @worker might have ended up outside
1265 * its affinity scope. Repatriate.
1266 */
1267 if (!pool->attrs->affn_strict &&
1268 !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) {
1269 struct work_struct *work = list_first_entry(&pool->worklist,
1270 struct work_struct, entry);
1271 int wake_cpu = cpumask_any_and_distribute(pool->attrs->__pod_cpumask,
1272 cpu_online_mask);
1273 if (wake_cpu < nr_cpu_ids) {
1274 p->wake_cpu = wake_cpu;
1275 get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++;
1276 }
1277 }
1278#endif
1279 wake_up_process(p);
1280 return true;
1281}
1282
1283#ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT
1284
1285/*
1286 * Concurrency-managed per-cpu work items that hog CPU for longer than
1287 * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism,
1288 * which prevents them from stalling other concurrency-managed work items. If a
1289 * work function keeps triggering this mechanism, it's likely that the work item
1290 * should be using an unbound workqueue instead.
1291 *
1292 * wq_cpu_intensive_report() tracks work functions which trigger such conditions
1293 * and report them so that they can be examined and converted to use unbound
1294 * workqueues as appropriate. To avoid flooding the console, each violating work
1295 * function is tracked and reported with exponential backoff.
1296 */
1297#define WCI_MAX_ENTS 128
1298
1299struct wci_ent {
1300 work_func_t func;
1301 atomic64_t cnt;
1302 struct hlist_node hash_node;
1303};
1304
1305static struct wci_ent wci_ents[WCI_MAX_ENTS];
1306static int wci_nr_ents;
1307static DEFINE_RAW_SPINLOCK(wci_lock);
1308static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS));
1309
1310static struct wci_ent *wci_find_ent(work_func_t func)
1311{
1312 struct wci_ent *ent;
1313
1314 hash_for_each_possible_rcu(wci_hash, ent, hash_node,
1315 (unsigned long)func) {
1316 if (ent->func == func)
1317 return ent;
1318 }
1319 return NULL;
1320}
1321
1322static void wq_cpu_intensive_report(work_func_t func)
1323{
1324 struct wci_ent *ent;
1325
1326restart:
1327 ent = wci_find_ent(func);
1328 if (ent) {
1329 u64 cnt;
1330
1331 /*
1332 * Start reporting from the warning_thresh and back off
1333 * exponentially.
1334 */
1335 cnt = atomic64_inc_return_relaxed(&ent->cnt);
1336 if (wq_cpu_intensive_warning_thresh &&
1337 cnt >= wq_cpu_intensive_warning_thresh &&
1338 is_power_of_2(cnt + 1 - wq_cpu_intensive_warning_thresh))
1339 printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n",
1340 ent->func, wq_cpu_intensive_thresh_us,
1341 atomic64_read(&ent->cnt));
1342 return;
1343 }
1344
1345 /*
1346 * @func is a new violation. Allocate a new entry for it. If wcn_ents[]
1347 * is exhausted, something went really wrong and we probably made enough
1348 * noise already.
1349 */
1350 if (wci_nr_ents >= WCI_MAX_ENTS)
1351 return;
1352
1353 raw_spin_lock(&wci_lock);
1354
1355 if (wci_nr_ents >= WCI_MAX_ENTS) {
1356 raw_spin_unlock(&wci_lock);
1357 return;
1358 }
1359
1360 if (wci_find_ent(func)) {
1361 raw_spin_unlock(&wci_lock);
1362 goto restart;
1363 }
1364
1365 ent = &wci_ents[wci_nr_ents++];
1366 ent->func = func;
1367 atomic64_set(&ent->cnt, 0);
1368 hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func);
1369
1370 raw_spin_unlock(&wci_lock);
1371
1372 goto restart;
1373}
1374
1375#else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1376static void wq_cpu_intensive_report(work_func_t func) {}
1377#endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */
1378
1379/**
1380 * wq_worker_running - a worker is running again
1381 * @task: task waking up
1382 *
1383 * This function is called when a worker returns from schedule()
1384 */
1385void wq_worker_running(struct task_struct *task)
1386{
1387 struct worker *worker = kthread_data(task);
1388
1389 if (!READ_ONCE(worker->sleeping))
1390 return;
1391
1392 /*
1393 * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check
1394 * and the nr_running increment below, we may ruin the nr_running reset
1395 * and leave with an unexpected pool->nr_running == 1 on the newly unbound
1396 * pool. Protect against such race.
1397 */
1398 preempt_disable();
1399 if (!(worker->flags & WORKER_NOT_RUNNING))
1400 worker->pool->nr_running++;
1401 preempt_enable();
1402
1403 /*
1404 * CPU intensive auto-detection cares about how long a work item hogged
1405 * CPU without sleeping. Reset the starting timestamp on wakeup.
1406 */
1407 worker->current_at = worker->task->se.sum_exec_runtime;
1408
1409 WRITE_ONCE(worker->sleeping, 0);
1410}
1411
1412/**
1413 * wq_worker_sleeping - a worker is going to sleep
1414 * @task: task going to sleep
1415 *
1416 * This function is called from schedule() when a busy worker is
1417 * going to sleep.
1418 */
1419void wq_worker_sleeping(struct task_struct *task)
1420{
1421 struct worker *worker = kthread_data(task);
1422 struct worker_pool *pool;
1423
1424 /*
1425 * Rescuers, which may not have all the fields set up like normal
1426 * workers, also reach here, let's not access anything before
1427 * checking NOT_RUNNING.
1428 */
1429 if (worker->flags & WORKER_NOT_RUNNING)
1430 return;
1431
1432 pool = worker->pool;
1433
1434 /* Return if preempted before wq_worker_running() was reached */
1435 if (READ_ONCE(worker->sleeping))
1436 return;
1437
1438 WRITE_ONCE(worker->sleeping, 1);
1439 raw_spin_lock_irq(&pool->lock);
1440
1441 /*
1442 * Recheck in case unbind_workers() preempted us. We don't
1443 * want to decrement nr_running after the worker is unbound
1444 * and nr_running has been reset.
1445 */
1446 if (worker->flags & WORKER_NOT_RUNNING) {
1447 raw_spin_unlock_irq(&pool->lock);
1448 return;
1449 }
1450
1451 pool->nr_running--;
1452 if (kick_pool(pool))
1453 worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1454
1455 raw_spin_unlock_irq(&pool->lock);
1456}
1457
1458/**
1459 * wq_worker_tick - a scheduler tick occurred while a kworker is running
1460 * @task: task currently running
1461 *
1462 * Called from sched_tick(). We're in the IRQ context and the current
1463 * worker's fields which follow the 'K' locking rule can be accessed safely.
1464 */
1465void wq_worker_tick(struct task_struct *task)
1466{
1467 struct worker *worker = kthread_data(task);
1468 struct pool_workqueue *pwq = worker->current_pwq;
1469 struct worker_pool *pool = worker->pool;
1470
1471 if (!pwq)
1472 return;
1473
1474 pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC;
1475
1476 if (!wq_cpu_intensive_thresh_us)
1477 return;
1478
1479 /*
1480 * If the current worker is concurrency managed and hogged the CPU for
1481 * longer than wq_cpu_intensive_thresh_us, it's automatically marked
1482 * CPU_INTENSIVE to avoid stalling other concurrency-managed work items.
1483 *
1484 * Set @worker->sleeping means that @worker is in the process of
1485 * switching out voluntarily and won't be contributing to
1486 * @pool->nr_running until it wakes up. As wq_worker_sleeping() also
1487 * decrements ->nr_running, setting CPU_INTENSIVE here can lead to
1488 * double decrements. The task is releasing the CPU anyway. Let's skip.
1489 * We probably want to make this prettier in the future.
1490 */
1491 if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) ||
1492 worker->task->se.sum_exec_runtime - worker->current_at <
1493 wq_cpu_intensive_thresh_us * NSEC_PER_USEC)
1494 return;
1495
1496 raw_spin_lock(&pool->lock);
1497
1498 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
1499 wq_cpu_intensive_report(worker->current_func);
1500 pwq->stats[PWQ_STAT_CPU_INTENSIVE]++;
1501
1502 if (kick_pool(pool))
1503 pwq->stats[PWQ_STAT_CM_WAKEUP]++;
1504
1505 raw_spin_unlock(&pool->lock);
1506}
1507
1508/**
1509 * wq_worker_last_func - retrieve worker's last work function
1510 * @task: Task to retrieve last work function of.
1511 *
1512 * Determine the last function a worker executed. This is called from
1513 * the scheduler to get a worker's last known identity.
1514 *
1515 * CONTEXT:
1516 * raw_spin_lock_irq(rq->lock)
1517 *
1518 * This function is called during schedule() when a kworker is going
1519 * to sleep. It's used by psi to identify aggregation workers during
1520 * dequeuing, to allow periodic aggregation to shut-off when that
1521 * worker is the last task in the system or cgroup to go to sleep.
1522 *
1523 * As this function doesn't involve any workqueue-related locking, it
1524 * only returns stable values when called from inside the scheduler's
1525 * queuing and dequeuing paths, when @task, which must be a kworker,
1526 * is guaranteed to not be processing any works.
1527 *
1528 * Return:
1529 * The last work function %current executed as a worker, NULL if it
1530 * hasn't executed any work yet.
1531 */
1532work_func_t wq_worker_last_func(struct task_struct *task)
1533{
1534 struct worker *worker = kthread_data(task);
1535
1536 return worker->last_func;
1537}
1538
1539/**
1540 * wq_node_nr_active - Determine wq_node_nr_active to use
1541 * @wq: workqueue of interest
1542 * @node: NUMA node, can be %NUMA_NO_NODE
1543 *
1544 * Determine wq_node_nr_active to use for @wq on @node. Returns:
1545 *
1546 * - %NULL for per-cpu workqueues as they don't need to use shared nr_active.
1547 *
1548 * - node_nr_active[nr_node_ids] if @node is %NUMA_NO_NODE.
1549 *
1550 * - Otherwise, node_nr_active[@node].
1551 */
1552static struct wq_node_nr_active *wq_node_nr_active(struct workqueue_struct *wq,
1553 int node)
1554{
1555 if (!(wq->flags & WQ_UNBOUND))
1556 return NULL;
1557
1558 if (node == NUMA_NO_NODE)
1559 node = nr_node_ids;
1560
1561 return wq->node_nr_active[node];
1562}
1563
1564/**
1565 * wq_update_node_max_active - Update per-node max_actives to use
1566 * @wq: workqueue to update
1567 * @off_cpu: CPU that's going down, -1 if a CPU is not going down
1568 *
1569 * Update @wq->node_nr_active[]->max. @wq must be unbound. max_active is
1570 * distributed among nodes according to the proportions of numbers of online
1571 * cpus. The result is always between @wq->min_active and max_active.
1572 */
1573static void wq_update_node_max_active(struct workqueue_struct *wq, int off_cpu)
1574{
1575 struct cpumask *effective = unbound_effective_cpumask(wq);
1576 int min_active = READ_ONCE(wq->min_active);
1577 int max_active = READ_ONCE(wq->max_active);
1578 int total_cpus, node;
1579
1580 lockdep_assert_held(&wq->mutex);
1581
1582 if (!wq_topo_initialized)
1583 return;
1584
1585 if (off_cpu >= 0 && !cpumask_test_cpu(off_cpu, effective))
1586 off_cpu = -1;
1587
1588 total_cpus = cpumask_weight_and(effective, cpu_online_mask);
1589 if (off_cpu >= 0)
1590 total_cpus--;
1591
1592 /* If all CPUs of the wq get offline, use the default values */
1593 if (unlikely(!total_cpus)) {
1594 for_each_node(node)
1595 wq_node_nr_active(wq, node)->max = min_active;
1596
1597 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1598 return;
1599 }
1600
1601 for_each_node(node) {
1602 int node_cpus;
1603
1604 node_cpus = cpumask_weight_and(effective, cpumask_of_node(node));
1605 if (off_cpu >= 0 && cpu_to_node(off_cpu) == node)
1606 node_cpus--;
1607
1608 wq_node_nr_active(wq, node)->max =
1609 clamp(DIV_ROUND_UP(max_active * node_cpus, total_cpus),
1610 min_active, max_active);
1611 }
1612
1613 wq_node_nr_active(wq, NUMA_NO_NODE)->max = max_active;
1614}
1615
1616/**
1617 * get_pwq - get an extra reference on the specified pool_workqueue
1618 * @pwq: pool_workqueue to get
1619 *
1620 * Obtain an extra reference on @pwq. The caller should guarantee that
1621 * @pwq has positive refcnt and be holding the matching pool->lock.
1622 */
1623static void get_pwq(struct pool_workqueue *pwq)
1624{
1625 lockdep_assert_held(&pwq->pool->lock);
1626 WARN_ON_ONCE(pwq->refcnt <= 0);
1627 pwq->refcnt++;
1628}
1629
1630/**
1631 * put_pwq - put a pool_workqueue reference
1632 * @pwq: pool_workqueue to put
1633 *
1634 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1635 * destruction. The caller should be holding the matching pool->lock.
1636 */
1637static void put_pwq(struct pool_workqueue *pwq)
1638{
1639 lockdep_assert_held(&pwq->pool->lock);
1640 if (likely(--pwq->refcnt))
1641 return;
1642 /*
1643 * @pwq can't be released under pool->lock, bounce to a dedicated
1644 * kthread_worker to avoid A-A deadlocks.
1645 */
1646 kthread_queue_work(pwq_release_worker, &pwq->release_work);
1647}
1648
1649/**
1650 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1651 * @pwq: pool_workqueue to put (can be %NULL)
1652 *
1653 * put_pwq() with locking. This function also allows %NULL @pwq.
1654 */
1655static void put_pwq_unlocked(struct pool_workqueue *pwq)
1656{
1657 if (pwq) {
1658 /*
1659 * As both pwqs and pools are RCU protected, the
1660 * following lock operations are safe.
1661 */
1662 raw_spin_lock_irq(&pwq->pool->lock);
1663 put_pwq(pwq);
1664 raw_spin_unlock_irq(&pwq->pool->lock);
1665 }
1666}
1667
1668static bool pwq_is_empty(struct pool_workqueue *pwq)
1669{
1670 return !pwq->nr_active && list_empty(&pwq->inactive_works);
1671}
1672
1673static void __pwq_activate_work(struct pool_workqueue *pwq,
1674 struct work_struct *work)
1675{
1676 unsigned long *wdb = work_data_bits(work);
1677
1678 WARN_ON_ONCE(!(*wdb & WORK_STRUCT_INACTIVE));
1679 trace_workqueue_activate_work(work);
1680 if (list_empty(&pwq->pool->worklist))
1681 pwq->pool->watchdog_ts = jiffies;
1682 move_linked_works(work, &pwq->pool->worklist, NULL);
1683 __clear_bit(WORK_STRUCT_INACTIVE_BIT, wdb);
1684}
1685
1686static bool tryinc_node_nr_active(struct wq_node_nr_active *nna)
1687{
1688 int max = READ_ONCE(nna->max);
1689
1690 while (true) {
1691 int old, tmp;
1692
1693 old = atomic_read(&nna->nr);
1694 if (old >= max)
1695 return false;
1696 tmp = atomic_cmpxchg_relaxed(&nna->nr, old, old + 1);
1697 if (tmp == old)
1698 return true;
1699 }
1700}
1701
1702/**
1703 * pwq_tryinc_nr_active - Try to increment nr_active for a pwq
1704 * @pwq: pool_workqueue of interest
1705 * @fill: max_active may have increased, try to increase concurrency level
1706 *
1707 * Try to increment nr_active for @pwq. Returns %true if an nr_active count is
1708 * successfully obtained. %false otherwise.
1709 */
1710static bool pwq_tryinc_nr_active(struct pool_workqueue *pwq, bool fill)
1711{
1712 struct workqueue_struct *wq = pwq->wq;
1713 struct worker_pool *pool = pwq->pool;
1714 struct wq_node_nr_active *nna = wq_node_nr_active(wq, pool->node);
1715 bool obtained = false;
1716
1717 lockdep_assert_held(&pool->lock);
1718
1719 if (!nna) {
1720 /* BH or per-cpu workqueue, pwq->nr_active is sufficient */
1721 obtained = pwq->nr_active < READ_ONCE(wq->max_active);
1722 goto out;
1723 }
1724
1725 if (unlikely(pwq->plugged))
1726 return false;
1727
1728 /*
1729 * Unbound workqueue uses per-node shared nr_active $nna. If @pwq is
1730 * already waiting on $nna, pwq_dec_nr_active() will maintain the
1731 * concurrency level. Don't jump the line.
1732 *
1733 * We need to ignore the pending test after max_active has increased as
1734 * pwq_dec_nr_active() can only maintain the concurrency level but not
1735 * increase it. This is indicated by @fill.
1736 */
1737 if (!list_empty(&pwq->pending_node) && likely(!fill))
1738 goto out;
1739
1740 obtained = tryinc_node_nr_active(nna);
1741 if (obtained)
1742 goto out;
1743
1744 /*
1745 * Lockless acquisition failed. Lock, add ourself to $nna->pending_pwqs
1746 * and try again. The smp_mb() is paired with the implied memory barrier
1747 * of atomic_dec_return() in pwq_dec_nr_active() to ensure that either
1748 * we see the decremented $nna->nr or they see non-empty
1749 * $nna->pending_pwqs.
1750 */
1751 raw_spin_lock(&nna->lock);
1752
1753 if (list_empty(&pwq->pending_node))
1754 list_add_tail(&pwq->pending_node, &nna->pending_pwqs);
1755 else if (likely(!fill))
1756 goto out_unlock;
1757
1758 smp_mb();
1759
1760 obtained = tryinc_node_nr_active(nna);
1761
1762 /*
1763 * If @fill, @pwq might have already been pending. Being spuriously
1764 * pending in cold paths doesn't affect anything. Let's leave it be.
1765 */
1766 if (obtained && likely(!fill))
1767 list_del_init(&pwq->pending_node);
1768
1769out_unlock:
1770 raw_spin_unlock(&nna->lock);
1771out:
1772 if (obtained)
1773 pwq->nr_active++;
1774 return obtained;
1775}
1776
1777/**
1778 * pwq_activate_first_inactive - Activate the first inactive work item on a pwq
1779 * @pwq: pool_workqueue of interest
1780 * @fill: max_active may have increased, try to increase concurrency level
1781 *
1782 * Activate the first inactive work item of @pwq if available and allowed by
1783 * max_active limit.
1784 *
1785 * Returns %true if an inactive work item has been activated. %false if no
1786 * inactive work item is found or max_active limit is reached.
1787 */
1788static bool pwq_activate_first_inactive(struct pool_workqueue *pwq, bool fill)
1789{
1790 struct work_struct *work =
1791 list_first_entry_or_null(&pwq->inactive_works,
1792 struct work_struct, entry);
1793
1794 if (work && pwq_tryinc_nr_active(pwq, fill)) {
1795 __pwq_activate_work(pwq, work);
1796 return true;
1797 } else {
1798 return false;
1799 }
1800}
1801
1802/**
1803 * unplug_oldest_pwq - unplug the oldest pool_workqueue
1804 * @wq: workqueue_struct where its oldest pwq is to be unplugged
1805 *
1806 * This function should only be called for ordered workqueues where only the
1807 * oldest pwq is unplugged, the others are plugged to suspend execution to
1808 * ensure proper work item ordering::
1809 *
1810 * dfl_pwq --------------+ [P] - plugged
1811 * |
1812 * v
1813 * pwqs -> A -> B [P] -> C [P] (newest)
1814 * | | |
1815 * 1 3 5
1816 * | | |
1817 * 2 4 6
1818 *
1819 * When the oldest pwq is drained and removed, this function should be called
1820 * to unplug the next oldest one to start its work item execution. Note that
1821 * pwq's are linked into wq->pwqs with the oldest first, so the first one in
1822 * the list is the oldest.
1823 */
1824static void unplug_oldest_pwq(struct workqueue_struct *wq)
1825{
1826 struct pool_workqueue *pwq;
1827
1828 lockdep_assert_held(&wq->mutex);
1829
1830 /* Caller should make sure that pwqs isn't empty before calling */
1831 pwq = list_first_entry_or_null(&wq->pwqs, struct pool_workqueue,
1832 pwqs_node);
1833 raw_spin_lock_irq(&pwq->pool->lock);
1834 if (pwq->plugged) {
1835 pwq->plugged = false;
1836 if (pwq_activate_first_inactive(pwq, true))
1837 kick_pool(pwq->pool);
1838 }
1839 raw_spin_unlock_irq(&pwq->pool->lock);
1840}
1841
1842/**
1843 * node_activate_pending_pwq - Activate a pending pwq on a wq_node_nr_active
1844 * @nna: wq_node_nr_active to activate a pending pwq for
1845 * @caller_pool: worker_pool the caller is locking
1846 *
1847 * Activate a pwq in @nna->pending_pwqs. Called with @caller_pool locked.
1848 * @caller_pool may be unlocked and relocked to lock other worker_pools.
1849 */
1850static void node_activate_pending_pwq(struct wq_node_nr_active *nna,
1851 struct worker_pool *caller_pool)
1852{
1853 struct worker_pool *locked_pool = caller_pool;
1854 struct pool_workqueue *pwq;
1855 struct work_struct *work;
1856
1857 lockdep_assert_held(&caller_pool->lock);
1858
1859 raw_spin_lock(&nna->lock);
1860retry:
1861 pwq = list_first_entry_or_null(&nna->pending_pwqs,
1862 struct pool_workqueue, pending_node);
1863 if (!pwq)
1864 goto out_unlock;
1865
1866 /*
1867 * If @pwq is for a different pool than @locked_pool, we need to lock
1868 * @pwq->pool->lock. Let's trylock first. If unsuccessful, do the unlock
1869 * / lock dance. For that, we also need to release @nna->lock as it's
1870 * nested inside pool locks.
1871 */
1872 if (pwq->pool != locked_pool) {
1873 raw_spin_unlock(&locked_pool->lock);
1874 locked_pool = pwq->pool;
1875 if (!raw_spin_trylock(&locked_pool->lock)) {
1876 raw_spin_unlock(&nna->lock);
1877 raw_spin_lock(&locked_pool->lock);
1878 raw_spin_lock(&nna->lock);
1879 goto retry;
1880 }
1881 }
1882
1883 /*
1884 * $pwq may not have any inactive work items due to e.g. cancellations.
1885 * Drop it from pending_pwqs and see if there's another one.
1886 */
1887 work = list_first_entry_or_null(&pwq->inactive_works,
1888 struct work_struct, entry);
1889 if (!work) {
1890 list_del_init(&pwq->pending_node);
1891 goto retry;
1892 }
1893
1894 /*
1895 * Acquire an nr_active count and activate the inactive work item. If
1896 * $pwq still has inactive work items, rotate it to the end of the
1897 * pending_pwqs so that we round-robin through them. This means that
1898 * inactive work items are not activated in queueing order which is fine
1899 * given that there has never been any ordering across different pwqs.
1900 */
1901 if (likely(tryinc_node_nr_active(nna))) {
1902 pwq->nr_active++;
1903 __pwq_activate_work(pwq, work);
1904
1905 if (list_empty(&pwq->inactive_works))
1906 list_del_init(&pwq->pending_node);
1907 else
1908 list_move_tail(&pwq->pending_node, &nna->pending_pwqs);
1909
1910 /* if activating a foreign pool, make sure it's running */
1911 if (pwq->pool != caller_pool)
1912 kick_pool(pwq->pool);
1913 }
1914
1915out_unlock:
1916 raw_spin_unlock(&nna->lock);
1917 if (locked_pool != caller_pool) {
1918 raw_spin_unlock(&locked_pool->lock);
1919 raw_spin_lock(&caller_pool->lock);
1920 }
1921}
1922
1923/**
1924 * pwq_dec_nr_active - Retire an active count
1925 * @pwq: pool_workqueue of interest
1926 *
1927 * Decrement @pwq's nr_active and try to activate the first inactive work item.
1928 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock.
1929 */
1930static void pwq_dec_nr_active(struct pool_workqueue *pwq)
1931{
1932 struct worker_pool *pool = pwq->pool;
1933 struct wq_node_nr_active *nna = wq_node_nr_active(pwq->wq, pool->node);
1934
1935 lockdep_assert_held(&pool->lock);
1936
1937 /*
1938 * @pwq->nr_active should be decremented for both percpu and unbound
1939 * workqueues.
1940 */
1941 pwq->nr_active--;
1942
1943 /*
1944 * For a percpu workqueue, it's simple. Just need to kick the first
1945 * inactive work item on @pwq itself.
1946 */
1947 if (!nna) {
1948 pwq_activate_first_inactive(pwq, false);
1949 return;
1950 }
1951
1952 /*
1953 * If @pwq is for an unbound workqueue, it's more complicated because
1954 * multiple pwqs and pools may be sharing the nr_active count. When a
1955 * pwq needs to wait for an nr_active count, it puts itself on
1956 * $nna->pending_pwqs. The following atomic_dec_return()'s implied
1957 * memory barrier is paired with smp_mb() in pwq_tryinc_nr_active() to
1958 * guarantee that either we see non-empty pending_pwqs or they see
1959 * decremented $nna->nr.
1960 *
1961 * $nna->max may change as CPUs come online/offline and @pwq->wq's
1962 * max_active gets updated. However, it is guaranteed to be equal to or
1963 * larger than @pwq->wq->min_active which is above zero unless freezing.
1964 * This maintains the forward progress guarantee.
1965 */
1966 if (atomic_dec_return(&nna->nr) >= READ_ONCE(nna->max))
1967 return;
1968
1969 if (!list_empty(&nna->pending_pwqs))
1970 node_activate_pending_pwq(nna, pool);
1971}
1972
1973/**
1974 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1975 * @pwq: pwq of interest
1976 * @work_data: work_data of work which left the queue
1977 *
1978 * A work either has completed or is removed from pending queue,
1979 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1980 *
1981 * NOTE:
1982 * For unbound workqueues, this function may temporarily drop @pwq->pool->lock
1983 * and thus should be called after all other state updates for the in-flight
1984 * work item is complete.
1985 *
1986 * CONTEXT:
1987 * raw_spin_lock_irq(pool->lock).
1988 */
1989static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data)
1990{
1991 int color = get_work_color(work_data);
1992
1993 if (!(work_data & WORK_STRUCT_INACTIVE))
1994 pwq_dec_nr_active(pwq);
1995
1996 pwq->nr_in_flight[color]--;
1997
1998 /* is flush in progress and are we at the flushing tip? */
1999 if (likely(pwq->flush_color != color))
2000 goto out_put;
2001
2002 /* are there still in-flight works? */
2003 if (pwq->nr_in_flight[color])
2004 goto out_put;
2005
2006 /* this pwq is done, clear flush_color */
2007 pwq->flush_color = -1;
2008
2009 /*
2010 * If this was the last pwq, wake up the first flusher. It
2011 * will handle the rest.
2012 */
2013 if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
2014 complete(&pwq->wq->first_flusher->done);
2015out_put:
2016 put_pwq(pwq);
2017}
2018
2019/**
2020 * try_to_grab_pending - steal work item from worklist and disable irq
2021 * @work: work item to steal
2022 * @cflags: %WORK_CANCEL_ flags
2023 * @irq_flags: place to store irq state
2024 *
2025 * Try to grab PENDING bit of @work. This function can handle @work in any
2026 * stable state - idle, on timer or on worklist.
2027 *
2028 * Return:
2029 *
2030 * ======== ================================================================
2031 * 1 if @work was pending and we successfully stole PENDING
2032 * 0 if @work was idle and we claimed PENDING
2033 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
2034 * ======== ================================================================
2035 *
2036 * Note:
2037 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
2038 * interrupted while holding PENDING and @work off queue, irq must be
2039 * disabled on entry. This, combined with delayed_work->timer being
2040 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
2041 *
2042 * On successful return, >= 0, irq is disabled and the caller is
2043 * responsible for releasing it using local_irq_restore(*@irq_flags).
2044 *
2045 * This function is safe to call from any context including IRQ handler.
2046 */
2047static int try_to_grab_pending(struct work_struct *work, u32 cflags,
2048 unsigned long *irq_flags)
2049{
2050 struct worker_pool *pool;
2051 struct pool_workqueue *pwq;
2052
2053 local_irq_save(*irq_flags);
2054
2055 /* try to steal the timer if it exists */
2056 if (cflags & WORK_CANCEL_DELAYED) {
2057 struct delayed_work *dwork = to_delayed_work(work);
2058
2059 /*
2060 * dwork->timer is irqsafe. If del_timer() fails, it's
2061 * guaranteed that the timer is not queued anywhere and not
2062 * running on the local CPU.
2063 */
2064 if (likely(del_timer(&dwork->timer)))
2065 return 1;
2066 }
2067
2068 /* try to claim PENDING the normal way */
2069 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2070 return 0;
2071
2072 rcu_read_lock();
2073 /*
2074 * The queueing is in progress, or it is already queued. Try to
2075 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2076 */
2077 pool = get_work_pool(work);
2078 if (!pool)
2079 goto fail;
2080
2081 raw_spin_lock(&pool->lock);
2082 /*
2083 * work->data is guaranteed to point to pwq only while the work
2084 * item is queued on pwq->wq, and both updating work->data to point
2085 * to pwq on queueing and to pool on dequeueing are done under
2086 * pwq->pool->lock. This in turn guarantees that, if work->data
2087 * points to pwq which is associated with a locked pool, the work
2088 * item is currently queued on that pool.
2089 */
2090 pwq = get_work_pwq(work);
2091 if (pwq && pwq->pool == pool) {
2092 unsigned long work_data = *work_data_bits(work);
2093
2094 debug_work_deactivate(work);
2095
2096 /*
2097 * A cancelable inactive work item must be in the
2098 * pwq->inactive_works since a queued barrier can't be
2099 * canceled (see the comments in insert_wq_barrier()).
2100 *
2101 * An inactive work item cannot be deleted directly because
2102 * it might have linked barrier work items which, if left
2103 * on the inactive_works list, will confuse pwq->nr_active
2104 * management later on and cause stall. Move the linked
2105 * barrier work items to the worklist when deleting the grabbed
2106 * item. Also keep WORK_STRUCT_INACTIVE in work_data, so that
2107 * it doesn't participate in nr_active management in later
2108 * pwq_dec_nr_in_flight().
2109 */
2110 if (work_data & WORK_STRUCT_INACTIVE)
2111 move_linked_works(work, &pwq->pool->worklist, NULL);
2112
2113 list_del_init(&work->entry);
2114
2115 /*
2116 * work->data points to pwq iff queued. Let's point to pool. As
2117 * this destroys work->data needed by the next step, stash it.
2118 */
2119 set_work_pool_and_keep_pending(work, pool->id,
2120 pool_offq_flags(pool));
2121
2122 /* must be the last step, see the function comment */
2123 pwq_dec_nr_in_flight(pwq, work_data);
2124
2125 raw_spin_unlock(&pool->lock);
2126 rcu_read_unlock();
2127 return 1;
2128 }
2129 raw_spin_unlock(&pool->lock);
2130fail:
2131 rcu_read_unlock();
2132 local_irq_restore(*irq_flags);
2133 return -EAGAIN;
2134}
2135
2136/**
2137 * work_grab_pending - steal work item from worklist and disable irq
2138 * @work: work item to steal
2139 * @cflags: %WORK_CANCEL_ flags
2140 * @irq_flags: place to store IRQ state
2141 *
2142 * Grab PENDING bit of @work. @work can be in any stable state - idle, on timer
2143 * or on worklist.
2144 *
2145 * Can be called from any context. IRQ is disabled on return with IRQ state
2146 * stored in *@irq_flags. The caller is responsible for re-enabling it using
2147 * local_irq_restore().
2148 *
2149 * Returns %true if @work was pending. %false if idle.
2150 */
2151static bool work_grab_pending(struct work_struct *work, u32 cflags,
2152 unsigned long *irq_flags)
2153{
2154 int ret;
2155
2156 while (true) {
2157 ret = try_to_grab_pending(work, cflags, irq_flags);
2158 if (ret >= 0)
2159 return ret;
2160 cpu_relax();
2161 }
2162}
2163
2164/**
2165 * insert_work - insert a work into a pool
2166 * @pwq: pwq @work belongs to
2167 * @work: work to insert
2168 * @head: insertion point
2169 * @extra_flags: extra WORK_STRUCT_* flags to set
2170 *
2171 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
2172 * work_struct flags.
2173 *
2174 * CONTEXT:
2175 * raw_spin_lock_irq(pool->lock).
2176 */
2177static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
2178 struct list_head *head, unsigned int extra_flags)
2179{
2180 debug_work_activate(work);
2181
2182 /* record the work call stack in order to print it in KASAN reports */
2183 kasan_record_aux_stack_noalloc(work);
2184
2185 /* we own @work, set data and link */
2186 set_work_pwq(work, pwq, extra_flags);
2187 list_add_tail(&work->entry, head);
2188 get_pwq(pwq);
2189}
2190
2191/*
2192 * Test whether @work is being queued from another work executing on the
2193 * same workqueue.
2194 */
2195static bool is_chained_work(struct workqueue_struct *wq)
2196{
2197 struct worker *worker;
2198
2199 worker = current_wq_worker();
2200 /*
2201 * Return %true iff I'm a worker executing a work item on @wq. If
2202 * I'm @worker, it's safe to dereference it without locking.
2203 */
2204 return worker && worker->current_pwq->wq == wq;
2205}
2206
2207/*
2208 * When queueing an unbound work item to a wq, prefer local CPU if allowed
2209 * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to
2210 * avoid perturbing sensitive tasks.
2211 */
2212static int wq_select_unbound_cpu(int cpu)
2213{
2214 int new_cpu;
2215
2216 if (likely(!wq_debug_force_rr_cpu)) {
2217 if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
2218 return cpu;
2219 } else {
2220 pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n");
2221 }
2222
2223 new_cpu = __this_cpu_read(wq_rr_cpu_last);
2224 new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
2225 if (unlikely(new_cpu >= nr_cpu_ids)) {
2226 new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
2227 if (unlikely(new_cpu >= nr_cpu_ids))
2228 return cpu;
2229 }
2230 __this_cpu_write(wq_rr_cpu_last, new_cpu);
2231
2232 return new_cpu;
2233}
2234
2235static void __queue_work(int cpu, struct workqueue_struct *wq,
2236 struct work_struct *work)
2237{
2238 struct pool_workqueue *pwq;
2239 struct worker_pool *last_pool, *pool;
2240 unsigned int work_flags;
2241 unsigned int req_cpu = cpu;
2242
2243 /*
2244 * While a work item is PENDING && off queue, a task trying to
2245 * steal the PENDING will busy-loop waiting for it to either get
2246 * queued or lose PENDING. Grabbing PENDING and queueing should
2247 * happen with IRQ disabled.
2248 */
2249 lockdep_assert_irqs_disabled();
2250
2251 /*
2252 * For a draining wq, only works from the same workqueue are
2253 * allowed. The __WQ_DESTROYING helps to spot the issue that
2254 * queues a new work item to a wq after destroy_workqueue(wq).
2255 */
2256 if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) &&
2257 WARN_ON_ONCE(!is_chained_work(wq))))
2258 return;
2259 rcu_read_lock();
2260retry:
2261 /* pwq which will be used unless @work is executing elsewhere */
2262 if (req_cpu == WORK_CPU_UNBOUND) {
2263 if (wq->flags & WQ_UNBOUND)
2264 cpu = wq_select_unbound_cpu(raw_smp_processor_id());
2265 else
2266 cpu = raw_smp_processor_id();
2267 }
2268
2269 pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu));
2270 pool = pwq->pool;
2271
2272 /*
2273 * If @work was previously on a different pool, it might still be
2274 * running there, in which case the work needs to be queued on that
2275 * pool to guarantee non-reentrancy.
2276 *
2277 * For ordered workqueue, work items must be queued on the newest pwq
2278 * for accurate order management. Guaranteed order also guarantees
2279 * non-reentrancy. See the comments above unplug_oldest_pwq().
2280 */
2281 last_pool = get_work_pool(work);
2282 if (last_pool && last_pool != pool && !(wq->flags & __WQ_ORDERED)) {
2283 struct worker *worker;
2284
2285 raw_spin_lock(&last_pool->lock);
2286
2287 worker = find_worker_executing_work(last_pool, work);
2288
2289 if (worker && worker->current_pwq->wq == wq) {
2290 pwq = worker->current_pwq;
2291 pool = pwq->pool;
2292 WARN_ON_ONCE(pool != last_pool);
2293 } else {
2294 /* meh... not running there, queue here */
2295 raw_spin_unlock(&last_pool->lock);
2296 raw_spin_lock(&pool->lock);
2297 }
2298 } else {
2299 raw_spin_lock(&pool->lock);
2300 }
2301
2302 /*
2303 * pwq is determined and locked. For unbound pools, we could have raced
2304 * with pwq release and it could already be dead. If its refcnt is zero,
2305 * repeat pwq selection. Note that unbound pwqs never die without
2306 * another pwq replacing it in cpu_pwq or while work items are executing
2307 * on it, so the retrying is guaranteed to make forward-progress.
2308 */
2309 if (unlikely(!pwq->refcnt)) {
2310 if (wq->flags & WQ_UNBOUND) {
2311 raw_spin_unlock(&pool->lock);
2312 cpu_relax();
2313 goto retry;
2314 }
2315 /* oops */
2316 WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
2317 wq->name, cpu);
2318 }
2319
2320 /* pwq determined, queue */
2321 trace_workqueue_queue_work(req_cpu, pwq, work);
2322
2323 if (WARN_ON(!list_empty(&work->entry)))
2324 goto out;
2325
2326 pwq->nr_in_flight[pwq->work_color]++;
2327 work_flags = work_color_to_flags(pwq->work_color);
2328
2329 /*
2330 * Limit the number of concurrently active work items to max_active.
2331 * @work must also queue behind existing inactive work items to maintain
2332 * ordering when max_active changes. See wq_adjust_max_active().
2333 */
2334 if (list_empty(&pwq->inactive_works) && pwq_tryinc_nr_active(pwq, false)) {
2335 if (list_empty(&pool->worklist))
2336 pool->watchdog_ts = jiffies;
2337
2338 trace_workqueue_activate_work(work);
2339 insert_work(pwq, work, &pool->worklist, work_flags);
2340 kick_pool(pool);
2341 } else {
2342 work_flags |= WORK_STRUCT_INACTIVE;
2343 insert_work(pwq, work, &pwq->inactive_works, work_flags);
2344 }
2345
2346out:
2347 raw_spin_unlock(&pool->lock);
2348 rcu_read_unlock();
2349}
2350
2351static bool clear_pending_if_disabled(struct work_struct *work)
2352{
2353 unsigned long data = *work_data_bits(work);
2354 struct work_offq_data offqd;
2355
2356 if (likely((data & WORK_STRUCT_PWQ) ||
2357 !(data & WORK_OFFQ_DISABLE_MASK)))
2358 return false;
2359
2360 work_offqd_unpack(&offqd, data);
2361 set_work_pool_and_clear_pending(work, offqd.pool_id,
2362 work_offqd_pack_flags(&offqd));
2363 return true;
2364}
2365
2366/**
2367 * queue_work_on - queue work on specific cpu
2368 * @cpu: CPU number to execute work on
2369 * @wq: workqueue to use
2370 * @work: work to queue
2371 *
2372 * We queue the work to a specific CPU, the caller must ensure it
2373 * can't go away. Callers that fail to ensure that the specified
2374 * CPU cannot go away will execute on a randomly chosen CPU.
2375 * But note well that callers specifying a CPU that never has been
2376 * online will get a splat.
2377 *
2378 * Return: %false if @work was already on a queue, %true otherwise.
2379 */
2380bool queue_work_on(int cpu, struct workqueue_struct *wq,
2381 struct work_struct *work)
2382{
2383 bool ret = false;
2384 unsigned long irq_flags;
2385
2386 local_irq_save(irq_flags);
2387
2388 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2389 !clear_pending_if_disabled(work)) {
2390 __queue_work(cpu, wq, work);
2391 ret = true;
2392 }
2393
2394 local_irq_restore(irq_flags);
2395 return ret;
2396}
2397EXPORT_SYMBOL(queue_work_on);
2398
2399/**
2400 * select_numa_node_cpu - Select a CPU based on NUMA node
2401 * @node: NUMA node ID that we want to select a CPU from
2402 *
2403 * This function will attempt to find a "random" cpu available on a given
2404 * node. If there are no CPUs available on the given node it will return
2405 * WORK_CPU_UNBOUND indicating that we should just schedule to any
2406 * available CPU if we need to schedule this work.
2407 */
2408static int select_numa_node_cpu(int node)
2409{
2410 int cpu;
2411
2412 /* Delay binding to CPU if node is not valid or online */
2413 if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
2414 return WORK_CPU_UNBOUND;
2415
2416 /* Use local node/cpu if we are already there */
2417 cpu = raw_smp_processor_id();
2418 if (node == cpu_to_node(cpu))
2419 return cpu;
2420
2421 /* Use "random" otherwise know as "first" online CPU of node */
2422 cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
2423
2424 /* If CPU is valid return that, otherwise just defer */
2425 return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
2426}
2427
2428/**
2429 * queue_work_node - queue work on a "random" cpu for a given NUMA node
2430 * @node: NUMA node that we are targeting the work for
2431 * @wq: workqueue to use
2432 * @work: work to queue
2433 *
2434 * We queue the work to a "random" CPU within a given NUMA node. The basic
2435 * idea here is to provide a way to somehow associate work with a given
2436 * NUMA node.
2437 *
2438 * This function will only make a best effort attempt at getting this onto
2439 * the right NUMA node. If no node is requested or the requested node is
2440 * offline then we just fall back to standard queue_work behavior.
2441 *
2442 * Currently the "random" CPU ends up being the first available CPU in the
2443 * intersection of cpu_online_mask and the cpumask of the node, unless we
2444 * are running on the node. In that case we just use the current CPU.
2445 *
2446 * Return: %false if @work was already on a queue, %true otherwise.
2447 */
2448bool queue_work_node(int node, struct workqueue_struct *wq,
2449 struct work_struct *work)
2450{
2451 unsigned long irq_flags;
2452 bool ret = false;
2453
2454 /*
2455 * This current implementation is specific to unbound workqueues.
2456 * Specifically we only return the first available CPU for a given
2457 * node instead of cycling through individual CPUs within the node.
2458 *
2459 * If this is used with a per-cpu workqueue then the logic in
2460 * workqueue_select_cpu_near would need to be updated to allow for
2461 * some round robin type logic.
2462 */
2463 WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
2464
2465 local_irq_save(irq_flags);
2466
2467 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2468 !clear_pending_if_disabled(work)) {
2469 int cpu = select_numa_node_cpu(node);
2470
2471 __queue_work(cpu, wq, work);
2472 ret = true;
2473 }
2474
2475 local_irq_restore(irq_flags);
2476 return ret;
2477}
2478EXPORT_SYMBOL_GPL(queue_work_node);
2479
2480void delayed_work_timer_fn(struct timer_list *t)
2481{
2482 struct delayed_work *dwork = from_timer(dwork, t, timer);
2483
2484 /* should have been called from irqsafe timer with irq already off */
2485 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2486}
2487EXPORT_SYMBOL(delayed_work_timer_fn);
2488
2489static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
2490 struct delayed_work *dwork, unsigned long delay)
2491{
2492 struct timer_list *timer = &dwork->timer;
2493 struct work_struct *work = &dwork->work;
2494
2495 WARN_ON_ONCE(!wq);
2496 WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
2497 WARN_ON_ONCE(timer_pending(timer));
2498 WARN_ON_ONCE(!list_empty(&work->entry));
2499
2500 /*
2501 * If @delay is 0, queue @dwork->work immediately. This is for
2502 * both optimization and correctness. The earliest @timer can
2503 * expire is on the closest next tick and delayed_work users depend
2504 * on that there's no such delay when @delay is 0.
2505 */
2506 if (!delay) {
2507 __queue_work(cpu, wq, &dwork->work);
2508 return;
2509 }
2510
2511 WARN_ON_ONCE(cpu != WORK_CPU_UNBOUND && !cpu_online(cpu));
2512 dwork->wq = wq;
2513 dwork->cpu = cpu;
2514 timer->expires = jiffies + delay;
2515
2516 if (housekeeping_enabled(HK_TYPE_TIMER)) {
2517 /* If the current cpu is a housekeeping cpu, use it. */
2518 cpu = smp_processor_id();
2519 if (!housekeeping_test_cpu(cpu, HK_TYPE_TIMER))
2520 cpu = housekeeping_any_cpu(HK_TYPE_TIMER);
2521 add_timer_on(timer, cpu);
2522 } else {
2523 if (likely(cpu == WORK_CPU_UNBOUND))
2524 add_timer_global(timer);
2525 else
2526 add_timer_on(timer, cpu);
2527 }
2528}
2529
2530/**
2531 * queue_delayed_work_on - queue work on specific CPU after delay
2532 * @cpu: CPU number to execute work on
2533 * @wq: workqueue to use
2534 * @dwork: work to queue
2535 * @delay: number of jiffies to wait before queueing
2536 *
2537 * We queue the delayed_work to a specific CPU, for non-zero delays the
2538 * caller must ensure it is online and can't go away. Callers that fail
2539 * to ensure this, may get @dwork->timer queued to an offlined CPU and
2540 * this will prevent queueing of @dwork->work unless the offlined CPU
2541 * becomes online again.
2542 *
2543 * Return: %false if @work was already on a queue, %true otherwise. If
2544 * @delay is zero and @dwork is idle, it will be scheduled for immediate
2545 * execution.
2546 */
2547bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
2548 struct delayed_work *dwork, unsigned long delay)
2549{
2550 struct work_struct *work = &dwork->work;
2551 bool ret = false;
2552 unsigned long irq_flags;
2553
2554 /* read the comment in __queue_work() */
2555 local_irq_save(irq_flags);
2556
2557 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2558 !clear_pending_if_disabled(work)) {
2559 __queue_delayed_work(cpu, wq, dwork, delay);
2560 ret = true;
2561 }
2562
2563 local_irq_restore(irq_flags);
2564 return ret;
2565}
2566EXPORT_SYMBOL(queue_delayed_work_on);
2567
2568/**
2569 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
2570 * @cpu: CPU number to execute work on
2571 * @wq: workqueue to use
2572 * @dwork: work to queue
2573 * @delay: number of jiffies to wait before queueing
2574 *
2575 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
2576 * modify @dwork's timer so that it expires after @delay. If @delay is
2577 * zero, @work is guaranteed to be scheduled immediately regardless of its
2578 * current state.
2579 *
2580 * Return: %false if @dwork was idle and queued, %true if @dwork was
2581 * pending and its timer was modified.
2582 *
2583 * This function is safe to call from any context including IRQ handler.
2584 * See try_to_grab_pending() for details.
2585 */
2586bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
2587 struct delayed_work *dwork, unsigned long delay)
2588{
2589 unsigned long irq_flags;
2590 bool ret;
2591
2592 ret = work_grab_pending(&dwork->work, WORK_CANCEL_DELAYED, &irq_flags);
2593
2594 if (!clear_pending_if_disabled(&dwork->work))
2595 __queue_delayed_work(cpu, wq, dwork, delay);
2596
2597 local_irq_restore(irq_flags);
2598 return ret;
2599}
2600EXPORT_SYMBOL_GPL(mod_delayed_work_on);
2601
2602static void rcu_work_rcufn(struct rcu_head *rcu)
2603{
2604 struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
2605
2606 /* read the comment in __queue_work() */
2607 local_irq_disable();
2608 __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
2609 local_irq_enable();
2610}
2611
2612/**
2613 * queue_rcu_work - queue work after a RCU grace period
2614 * @wq: workqueue to use
2615 * @rwork: work to queue
2616 *
2617 * Return: %false if @rwork was already pending, %true otherwise. Note
2618 * that a full RCU grace period is guaranteed only after a %true return.
2619 * While @rwork is guaranteed to be executed after a %false return, the
2620 * execution may happen before a full RCU grace period has passed.
2621 */
2622bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
2623{
2624 struct work_struct *work = &rwork->work;
2625
2626 /*
2627 * rcu_work can't be canceled or disabled. Warn if the user reached
2628 * inside @rwork and disabled the inner work.
2629 */
2630 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)) &&
2631 !WARN_ON_ONCE(clear_pending_if_disabled(work))) {
2632 rwork->wq = wq;
2633 call_rcu_hurry(&rwork->rcu, rcu_work_rcufn);
2634 return true;
2635 }
2636
2637 return false;
2638}
2639EXPORT_SYMBOL(queue_rcu_work);
2640
2641static struct worker *alloc_worker(int node)
2642{
2643 struct worker *worker;
2644
2645 worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
2646 if (worker) {
2647 INIT_LIST_HEAD(&worker->entry);
2648 INIT_LIST_HEAD(&worker->scheduled);
2649 INIT_LIST_HEAD(&worker->node);
2650 /* on creation a worker is in !idle && prep state */
2651 worker->flags = WORKER_PREP;
2652 }
2653 return worker;
2654}
2655
2656static cpumask_t *pool_allowed_cpus(struct worker_pool *pool)
2657{
2658 if (pool->cpu < 0 && pool->attrs->affn_strict)
2659 return pool->attrs->__pod_cpumask;
2660 else
2661 return pool->attrs->cpumask;
2662}
2663
2664/**
2665 * worker_attach_to_pool() - attach a worker to a pool
2666 * @worker: worker to be attached
2667 * @pool: the target pool
2668 *
2669 * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and
2670 * cpu-binding of @worker are kept coordinated with the pool across
2671 * cpu-[un]hotplugs.
2672 */
2673static void worker_attach_to_pool(struct worker *worker,
2674 struct worker_pool *pool)
2675{
2676 mutex_lock(&wq_pool_attach_mutex);
2677
2678 /*
2679 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains stable
2680 * across this function. See the comments above the flag definition for
2681 * details. BH workers are, while per-CPU, always DISASSOCIATED.
2682 */
2683 if (pool->flags & POOL_DISASSOCIATED) {
2684 worker->flags |= WORKER_UNBOUND;
2685 } else {
2686 WARN_ON_ONCE(pool->flags & POOL_BH);
2687 kthread_set_per_cpu(worker->task, pool->cpu);
2688 }
2689
2690 if (worker->rescue_wq)
2691 set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool));
2692
2693 list_add_tail(&worker->node, &pool->workers);
2694 worker->pool = pool;
2695
2696 mutex_unlock(&wq_pool_attach_mutex);
2697}
2698
2699static void unbind_worker(struct worker *worker)
2700{
2701 lockdep_assert_held(&wq_pool_attach_mutex);
2702
2703 kthread_set_per_cpu(worker->task, -1);
2704 if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask))
2705 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0);
2706 else
2707 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0);
2708}
2709
2710
2711static void detach_worker(struct worker *worker)
2712{
2713 lockdep_assert_held(&wq_pool_attach_mutex);
2714
2715 unbind_worker(worker);
2716 list_del(&worker->node);
2717}
2718
2719/**
2720 * worker_detach_from_pool() - detach a worker from its pool
2721 * @worker: worker which is attached to its pool
2722 *
2723 * Undo the attaching which had been done in worker_attach_to_pool(). The
2724 * caller worker shouldn't access to the pool after detached except it has
2725 * other reference to the pool.
2726 */
2727static void worker_detach_from_pool(struct worker *worker)
2728{
2729 struct worker_pool *pool = worker->pool;
2730
2731 /* there is one permanent BH worker per CPU which should never detach */
2732 WARN_ON_ONCE(pool->flags & POOL_BH);
2733
2734 mutex_lock(&wq_pool_attach_mutex);
2735 detach_worker(worker);
2736 worker->pool = NULL;
2737 mutex_unlock(&wq_pool_attach_mutex);
2738
2739 /* clear leftover flags without pool->lock after it is detached */
2740 worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
2741}
2742
2743static int format_worker_id(char *buf, size_t size, struct worker *worker,
2744 struct worker_pool *pool)
2745{
2746 if (worker->rescue_wq)
2747 return scnprintf(buf, size, "kworker/R-%s",
2748 worker->rescue_wq->name);
2749
2750 if (pool) {
2751 if (pool->cpu >= 0)
2752 return scnprintf(buf, size, "kworker/%d:%d%s",
2753 pool->cpu, worker->id,
2754 pool->attrs->nice < 0 ? "H" : "");
2755 else
2756 return scnprintf(buf, size, "kworker/u%d:%d",
2757 pool->id, worker->id);
2758 } else {
2759 return scnprintf(buf, size, "kworker/dying");
2760 }
2761}
2762
2763/**
2764 * create_worker - create a new workqueue worker
2765 * @pool: pool the new worker will belong to
2766 *
2767 * Create and start a new worker which is attached to @pool.
2768 *
2769 * CONTEXT:
2770 * Might sleep. Does GFP_KERNEL allocations.
2771 *
2772 * Return:
2773 * Pointer to the newly created worker.
2774 */
2775static struct worker *create_worker(struct worker_pool *pool)
2776{
2777 struct worker *worker;
2778 int id;
2779
2780 /* ID is needed to determine kthread name */
2781 id = ida_alloc(&pool->worker_ida, GFP_KERNEL);
2782 if (id < 0) {
2783 pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n",
2784 ERR_PTR(id));
2785 return NULL;
2786 }
2787
2788 worker = alloc_worker(pool->node);
2789 if (!worker) {
2790 pr_err_once("workqueue: Failed to allocate a worker\n");
2791 goto fail;
2792 }
2793
2794 worker->id = id;
2795
2796 if (!(pool->flags & POOL_BH)) {
2797 char id_buf[WORKER_ID_LEN];
2798
2799 format_worker_id(id_buf, sizeof(id_buf), worker, pool);
2800 worker->task = kthread_create_on_node(worker_thread, worker,
2801 pool->node, "%s", id_buf);
2802 if (IS_ERR(worker->task)) {
2803 if (PTR_ERR(worker->task) == -EINTR) {
2804 pr_err("workqueue: Interrupted when creating a worker thread \"%s\"\n",
2805 id_buf);
2806 } else {
2807 pr_err_once("workqueue: Failed to create a worker thread: %pe",
2808 worker->task);
2809 }
2810 goto fail;
2811 }
2812
2813 set_user_nice(worker->task, pool->attrs->nice);
2814 kthread_bind_mask(worker->task, pool_allowed_cpus(pool));
2815 }
2816
2817 /* successful, attach the worker to the pool */
2818 worker_attach_to_pool(worker, pool);
2819
2820 /* start the newly created worker */
2821 raw_spin_lock_irq(&pool->lock);
2822
2823 worker->pool->nr_workers++;
2824 worker_enter_idle(worker);
2825
2826 /*
2827 * @worker is waiting on a completion in kthread() and will trigger hung
2828 * check if not woken up soon. As kick_pool() is noop if @pool is empty,
2829 * wake it up explicitly.
2830 */
2831 if (worker->task)
2832 wake_up_process(worker->task);
2833
2834 raw_spin_unlock_irq(&pool->lock);
2835
2836 return worker;
2837
2838fail:
2839 ida_free(&pool->worker_ida, id);
2840 kfree(worker);
2841 return NULL;
2842}
2843
2844static void detach_dying_workers(struct list_head *cull_list)
2845{
2846 struct worker *worker;
2847
2848 list_for_each_entry(worker, cull_list, entry)
2849 detach_worker(worker);
2850}
2851
2852static void reap_dying_workers(struct list_head *cull_list)
2853{
2854 struct worker *worker, *tmp;
2855
2856 list_for_each_entry_safe(worker, tmp, cull_list, entry) {
2857 list_del_init(&worker->entry);
2858 kthread_stop_put(worker->task);
2859 kfree(worker);
2860 }
2861}
2862
2863/**
2864 * set_worker_dying - Tag a worker for destruction
2865 * @worker: worker to be destroyed
2866 * @list: transfer worker away from its pool->idle_list and into list
2867 *
2868 * Tag @worker for destruction and adjust @pool stats accordingly. The worker
2869 * should be idle.
2870 *
2871 * CONTEXT:
2872 * raw_spin_lock_irq(pool->lock).
2873 */
2874static void set_worker_dying(struct worker *worker, struct list_head *list)
2875{
2876 struct worker_pool *pool = worker->pool;
2877
2878 lockdep_assert_held(&pool->lock);
2879 lockdep_assert_held(&wq_pool_attach_mutex);
2880
2881 /* sanity check frenzy */
2882 if (WARN_ON(worker->current_work) ||
2883 WARN_ON(!list_empty(&worker->scheduled)) ||
2884 WARN_ON(!(worker->flags & WORKER_IDLE)))
2885 return;
2886
2887 pool->nr_workers--;
2888 pool->nr_idle--;
2889
2890 worker->flags |= WORKER_DIE;
2891
2892 list_move(&worker->entry, list);
2893
2894 /* get an extra task struct reference for later kthread_stop_put() */
2895 get_task_struct(worker->task);
2896}
2897
2898/**
2899 * idle_worker_timeout - check if some idle workers can now be deleted.
2900 * @t: The pool's idle_timer that just expired
2901 *
2902 * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in
2903 * worker_leave_idle(), as a worker flicking between idle and active while its
2904 * pool is at the too_many_workers() tipping point would cause too much timer
2905 * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let
2906 * it expire and re-evaluate things from there.
2907 */
2908static void idle_worker_timeout(struct timer_list *t)
2909{
2910 struct worker_pool *pool = from_timer(pool, t, idle_timer);
2911 bool do_cull = false;
2912
2913 if (work_pending(&pool->idle_cull_work))
2914 return;
2915
2916 raw_spin_lock_irq(&pool->lock);
2917
2918 if (too_many_workers(pool)) {
2919 struct worker *worker;
2920 unsigned long expires;
2921
2922 /* idle_list is kept in LIFO order, check the last one */
2923 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2924 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2925 do_cull = !time_before(jiffies, expires);
2926
2927 if (!do_cull)
2928 mod_timer(&pool->idle_timer, expires);
2929 }
2930 raw_spin_unlock_irq(&pool->lock);
2931
2932 if (do_cull)
2933 queue_work(system_unbound_wq, &pool->idle_cull_work);
2934}
2935
2936/**
2937 * idle_cull_fn - cull workers that have been idle for too long.
2938 * @work: the pool's work for handling these idle workers
2939 *
2940 * This goes through a pool's idle workers and gets rid of those that have been
2941 * idle for at least IDLE_WORKER_TIMEOUT seconds.
2942 *
2943 * We don't want to disturb isolated CPUs because of a pcpu kworker being
2944 * culled, so this also resets worker affinity. This requires a sleepable
2945 * context, hence the split between timer callback and work item.
2946 */
2947static void idle_cull_fn(struct work_struct *work)
2948{
2949 struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work);
2950 LIST_HEAD(cull_list);
2951
2952 /*
2953 * Grabbing wq_pool_attach_mutex here ensures an already-running worker
2954 * cannot proceed beyong set_pf_worker() in its self-destruct path.
2955 * This is required as a previously-preempted worker could run after
2956 * set_worker_dying() has happened but before detach_dying_workers() did.
2957 */
2958 mutex_lock(&wq_pool_attach_mutex);
2959 raw_spin_lock_irq(&pool->lock);
2960
2961 while (too_many_workers(pool)) {
2962 struct worker *worker;
2963 unsigned long expires;
2964
2965 worker = list_last_entry(&pool->idle_list, struct worker, entry);
2966 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2967
2968 if (time_before(jiffies, expires)) {
2969 mod_timer(&pool->idle_timer, expires);
2970 break;
2971 }
2972
2973 set_worker_dying(worker, &cull_list);
2974 }
2975
2976 raw_spin_unlock_irq(&pool->lock);
2977 detach_dying_workers(&cull_list);
2978 mutex_unlock(&wq_pool_attach_mutex);
2979
2980 reap_dying_workers(&cull_list);
2981}
2982
2983static void send_mayday(struct work_struct *work)
2984{
2985 struct pool_workqueue *pwq = get_work_pwq(work);
2986 struct workqueue_struct *wq = pwq->wq;
2987
2988 lockdep_assert_held(&wq_mayday_lock);
2989
2990 if (!wq->rescuer)
2991 return;
2992
2993 /* mayday mayday mayday */
2994 if (list_empty(&pwq->mayday_node)) {
2995 /*
2996 * If @pwq is for an unbound wq, its base ref may be put at
2997 * any time due to an attribute change. Pin @pwq until the
2998 * rescuer is done with it.
2999 */
3000 get_pwq(pwq);
3001 list_add_tail(&pwq->mayday_node, &wq->maydays);
3002 wake_up_process(wq->rescuer->task);
3003 pwq->stats[PWQ_STAT_MAYDAY]++;
3004 }
3005}
3006
3007static void pool_mayday_timeout(struct timer_list *t)
3008{
3009 struct worker_pool *pool = from_timer(pool, t, mayday_timer);
3010 struct work_struct *work;
3011
3012 raw_spin_lock_irq(&pool->lock);
3013 raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */
3014
3015 if (need_to_create_worker(pool)) {
3016 /*
3017 * We've been trying to create a new worker but
3018 * haven't been successful. We might be hitting an
3019 * allocation deadlock. Send distress signals to
3020 * rescuers.
3021 */
3022 list_for_each_entry(work, &pool->worklist, entry)
3023 send_mayday(work);
3024 }
3025
3026 raw_spin_unlock(&wq_mayday_lock);
3027 raw_spin_unlock_irq(&pool->lock);
3028
3029 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
3030}
3031
3032/**
3033 * maybe_create_worker - create a new worker if necessary
3034 * @pool: pool to create a new worker for
3035 *
3036 * Create a new worker for @pool if necessary. @pool is guaranteed to
3037 * have at least one idle worker on return from this function. If
3038 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
3039 * sent to all rescuers with works scheduled on @pool to resolve
3040 * possible allocation deadlock.
3041 *
3042 * On return, need_to_create_worker() is guaranteed to be %false and
3043 * may_start_working() %true.
3044 *
3045 * LOCKING:
3046 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3047 * multiple times. Does GFP_KERNEL allocations. Called only from
3048 * manager.
3049 */
3050static void maybe_create_worker(struct worker_pool *pool)
3051__releases(&pool->lock)
3052__acquires(&pool->lock)
3053{
3054restart:
3055 raw_spin_unlock_irq(&pool->lock);
3056
3057 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
3058 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
3059
3060 while (true) {
3061 if (create_worker(pool) || !need_to_create_worker(pool))
3062 break;
3063
3064 schedule_timeout_interruptible(CREATE_COOLDOWN);
3065
3066 if (!need_to_create_worker(pool))
3067 break;
3068 }
3069
3070 del_timer_sync(&pool->mayday_timer);
3071 raw_spin_lock_irq(&pool->lock);
3072 /*
3073 * This is necessary even after a new worker was just successfully
3074 * created as @pool->lock was dropped and the new worker might have
3075 * already become busy.
3076 */
3077 if (need_to_create_worker(pool))
3078 goto restart;
3079}
3080
3081/**
3082 * manage_workers - manage worker pool
3083 * @worker: self
3084 *
3085 * Assume the manager role and manage the worker pool @worker belongs
3086 * to. At any given time, there can be only zero or one manager per
3087 * pool. The exclusion is handled automatically by this function.
3088 *
3089 * The caller can safely start processing works on false return. On
3090 * true return, it's guaranteed that need_to_create_worker() is false
3091 * and may_start_working() is true.
3092 *
3093 * CONTEXT:
3094 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3095 * multiple times. Does GFP_KERNEL allocations.
3096 *
3097 * Return:
3098 * %false if the pool doesn't need management and the caller can safely
3099 * start processing works, %true if management function was performed and
3100 * the conditions that the caller verified before calling the function may
3101 * no longer be true.
3102 */
3103static bool manage_workers(struct worker *worker)
3104{
3105 struct worker_pool *pool = worker->pool;
3106
3107 if (pool->flags & POOL_MANAGER_ACTIVE)
3108 return false;
3109
3110 pool->flags |= POOL_MANAGER_ACTIVE;
3111 pool->manager = worker;
3112
3113 maybe_create_worker(pool);
3114
3115 pool->manager = NULL;
3116 pool->flags &= ~POOL_MANAGER_ACTIVE;
3117 rcuwait_wake_up(&manager_wait);
3118 return true;
3119}
3120
3121/**
3122 * process_one_work - process single work
3123 * @worker: self
3124 * @work: work to process
3125 *
3126 * Process @work. This function contains all the logics necessary to
3127 * process a single work including synchronization against and
3128 * interaction with other workers on the same cpu, queueing and
3129 * flushing. As long as context requirement is met, any worker can
3130 * call this function to process a work.
3131 *
3132 * CONTEXT:
3133 * raw_spin_lock_irq(pool->lock) which is released and regrabbed.
3134 */
3135static void process_one_work(struct worker *worker, struct work_struct *work)
3136__releases(&pool->lock)
3137__acquires(&pool->lock)
3138{
3139 struct pool_workqueue *pwq = get_work_pwq(work);
3140 struct worker_pool *pool = worker->pool;
3141 unsigned long work_data;
3142 int lockdep_start_depth, rcu_start_depth;
3143 bool bh_draining = pool->flags & POOL_BH_DRAINING;
3144#ifdef CONFIG_LOCKDEP
3145 /*
3146 * It is permissible to free the struct work_struct from
3147 * inside the function that is called from it, this we need to
3148 * take into account for lockdep too. To avoid bogus "held
3149 * lock freed" warnings as well as problems when looking into
3150 * work->lockdep_map, make a copy and use that here.
3151 */
3152 struct lockdep_map lockdep_map;
3153
3154 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
3155#endif
3156 /* ensure we're on the correct CPU */
3157 WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
3158 raw_smp_processor_id() != pool->cpu);
3159
3160 /* claim and dequeue */
3161 debug_work_deactivate(work);
3162 hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
3163 worker->current_work = work;
3164 worker->current_func = work->func;
3165 worker->current_pwq = pwq;
3166 if (worker->task)
3167 worker->current_at = worker->task->se.sum_exec_runtime;
3168 work_data = *work_data_bits(work);
3169 worker->current_color = get_work_color(work_data);
3170
3171 /*
3172 * Record wq name for cmdline and debug reporting, may get
3173 * overridden through set_worker_desc().
3174 */
3175 strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
3176
3177 list_del_init(&work->entry);
3178
3179 /*
3180 * CPU intensive works don't participate in concurrency management.
3181 * They're the scheduler's responsibility. This takes @worker out
3182 * of concurrency management and the next code block will chain
3183 * execution of the pending work items.
3184 */
3185 if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE))
3186 worker_set_flags(worker, WORKER_CPU_INTENSIVE);
3187
3188 /*
3189 * Kick @pool if necessary. It's always noop for per-cpu worker pools
3190 * since nr_running would always be >= 1 at this point. This is used to
3191 * chain execution of the pending work items for WORKER_NOT_RUNNING
3192 * workers such as the UNBOUND and CPU_INTENSIVE ones.
3193 */
3194 kick_pool(pool);
3195
3196 /*
3197 * Record the last pool and clear PENDING which should be the last
3198 * update to @work. Also, do this inside @pool->lock so that
3199 * PENDING and queued state changes happen together while IRQ is
3200 * disabled.
3201 */
3202 set_work_pool_and_clear_pending(work, pool->id, pool_offq_flags(pool));
3203
3204 pwq->stats[PWQ_STAT_STARTED]++;
3205 raw_spin_unlock_irq(&pool->lock);
3206
3207 rcu_start_depth = rcu_preempt_depth();
3208 lockdep_start_depth = lockdep_depth(current);
3209 /* see drain_dead_softirq_workfn() */
3210 if (!bh_draining)
3211 lock_map_acquire(pwq->wq->lockdep_map);
3212 lock_map_acquire(&lockdep_map);
3213 /*
3214 * Strictly speaking we should mark the invariant state without holding
3215 * any locks, that is, before these two lock_map_acquire()'s.
3216 *
3217 * However, that would result in:
3218 *
3219 * A(W1)
3220 * WFC(C)
3221 * A(W1)
3222 * C(C)
3223 *
3224 * Which would create W1->C->W1 dependencies, even though there is no
3225 * actual deadlock possible. There are two solutions, using a
3226 * read-recursive acquire on the work(queue) 'locks', but this will then
3227 * hit the lockdep limitation on recursive locks, or simply discard
3228 * these locks.
3229 *
3230 * AFAICT there is no possible deadlock scenario between the
3231 * flush_work() and complete() primitives (except for single-threaded
3232 * workqueues), so hiding them isn't a problem.
3233 */
3234 lockdep_invariant_state(true);
3235 trace_workqueue_execute_start(work);
3236 worker->current_func(work);
3237 /*
3238 * While we must be careful to not use "work" after this, the trace
3239 * point will only record its address.
3240 */
3241 trace_workqueue_execute_end(work, worker->current_func);
3242 pwq->stats[PWQ_STAT_COMPLETED]++;
3243 lock_map_release(&lockdep_map);
3244 if (!bh_draining)
3245 lock_map_release(pwq->wq->lockdep_map);
3246
3247 if (unlikely((worker->task && in_atomic()) ||
3248 lockdep_depth(current) != lockdep_start_depth ||
3249 rcu_preempt_depth() != rcu_start_depth)) {
3250 pr_err("BUG: workqueue leaked atomic, lock or RCU: %s[%d]\n"
3251 " preempt=0x%08x lock=%d->%d RCU=%d->%d workfn=%ps\n",
3252 current->comm, task_pid_nr(current), preempt_count(),
3253 lockdep_start_depth, lockdep_depth(current),
3254 rcu_start_depth, rcu_preempt_depth(),
3255 worker->current_func);
3256 debug_show_held_locks(current);
3257 dump_stack();
3258 }
3259
3260 /*
3261 * The following prevents a kworker from hogging CPU on !PREEMPTION
3262 * kernels, where a requeueing work item waiting for something to
3263 * happen could deadlock with stop_machine as such work item could
3264 * indefinitely requeue itself while all other CPUs are trapped in
3265 * stop_machine. At the same time, report a quiescent RCU state so
3266 * the same condition doesn't freeze RCU.
3267 */
3268 if (worker->task)
3269 cond_resched();
3270
3271 raw_spin_lock_irq(&pool->lock);
3272
3273 /*
3274 * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked
3275 * CPU intensive by wq_worker_tick() if @work hogged CPU longer than
3276 * wq_cpu_intensive_thresh_us. Clear it.
3277 */
3278 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
3279
3280 /* tag the worker for identification in schedule() */
3281 worker->last_func = worker->current_func;
3282
3283 /* we're done with it, release */
3284 hash_del(&worker->hentry);
3285 worker->current_work = NULL;
3286 worker->current_func = NULL;
3287 worker->current_pwq = NULL;
3288 worker->current_color = INT_MAX;
3289
3290 /* must be the last step, see the function comment */
3291 pwq_dec_nr_in_flight(pwq, work_data);
3292}
3293
3294/**
3295 * process_scheduled_works - process scheduled works
3296 * @worker: self
3297 *
3298 * Process all scheduled works. Please note that the scheduled list
3299 * may change while processing a work, so this function repeatedly
3300 * fetches a work from the top and executes it.
3301 *
3302 * CONTEXT:
3303 * raw_spin_lock_irq(pool->lock) which may be released and regrabbed
3304 * multiple times.
3305 */
3306static void process_scheduled_works(struct worker *worker)
3307{
3308 struct work_struct *work;
3309 bool first = true;
3310
3311 while ((work = list_first_entry_or_null(&worker->scheduled,
3312 struct work_struct, entry))) {
3313 if (first) {
3314 worker->pool->watchdog_ts = jiffies;
3315 first = false;
3316 }
3317 process_one_work(worker, work);
3318 }
3319}
3320
3321static void set_pf_worker(bool val)
3322{
3323 mutex_lock(&wq_pool_attach_mutex);
3324 if (val)
3325 current->flags |= PF_WQ_WORKER;
3326 else
3327 current->flags &= ~PF_WQ_WORKER;
3328 mutex_unlock(&wq_pool_attach_mutex);
3329}
3330
3331/**
3332 * worker_thread - the worker thread function
3333 * @__worker: self
3334 *
3335 * The worker thread function. All workers belong to a worker_pool -
3336 * either a per-cpu one or dynamic unbound one. These workers process all
3337 * work items regardless of their specific target workqueue. The only
3338 * exception is work items which belong to workqueues with a rescuer which
3339 * will be explained in rescuer_thread().
3340 *
3341 * Return: 0
3342 */
3343static int worker_thread(void *__worker)
3344{
3345 struct worker *worker = __worker;
3346 struct worker_pool *pool = worker->pool;
3347
3348 /* tell the scheduler that this is a workqueue worker */
3349 set_pf_worker(true);
3350woke_up:
3351 raw_spin_lock_irq(&pool->lock);
3352
3353 /* am I supposed to die? */
3354 if (unlikely(worker->flags & WORKER_DIE)) {
3355 raw_spin_unlock_irq(&pool->lock);
3356 set_pf_worker(false);
3357 /*
3358 * The worker is dead and PF_WQ_WORKER is cleared, worker->pool
3359 * shouldn't be accessed, reset it to NULL in case otherwise.
3360 */
3361 worker->pool = NULL;
3362 ida_free(&pool->worker_ida, worker->id);
3363 return 0;
3364 }
3365
3366 worker_leave_idle(worker);
3367recheck:
3368 /* no more worker necessary? */
3369 if (!need_more_worker(pool))
3370 goto sleep;
3371
3372 /* do we need to manage? */
3373 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
3374 goto recheck;
3375
3376 /*
3377 * ->scheduled list can only be filled while a worker is
3378 * preparing to process a work or actually processing it.
3379 * Make sure nobody diddled with it while I was sleeping.
3380 */
3381 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3382
3383 /*
3384 * Finish PREP stage. We're guaranteed to have at least one idle
3385 * worker or that someone else has already assumed the manager
3386 * role. This is where @worker starts participating in concurrency
3387 * management if applicable and concurrency management is restored
3388 * after being rebound. See rebind_workers() for details.
3389 */
3390 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3391
3392 do {
3393 struct work_struct *work =
3394 list_first_entry(&pool->worklist,
3395 struct work_struct, entry);
3396
3397 if (assign_work(work, worker, NULL))
3398 process_scheduled_works(worker);
3399 } while (keep_working(pool));
3400
3401 worker_set_flags(worker, WORKER_PREP);
3402sleep:
3403 /*
3404 * pool->lock is held and there's no work to process and no need to
3405 * manage, sleep. Workers are woken up only while holding
3406 * pool->lock or from local cpu, so setting the current state
3407 * before releasing pool->lock is enough to prevent losing any
3408 * event.
3409 */
3410 worker_enter_idle(worker);
3411 __set_current_state(TASK_IDLE);
3412 raw_spin_unlock_irq(&pool->lock);
3413 schedule();
3414 goto woke_up;
3415}
3416
3417/**
3418 * rescuer_thread - the rescuer thread function
3419 * @__rescuer: self
3420 *
3421 * Workqueue rescuer thread function. There's one rescuer for each
3422 * workqueue which has WQ_MEM_RECLAIM set.
3423 *
3424 * Regular work processing on a pool may block trying to create a new
3425 * worker which uses GFP_KERNEL allocation which has slight chance of
3426 * developing into deadlock if some works currently on the same queue
3427 * need to be processed to satisfy the GFP_KERNEL allocation. This is
3428 * the problem rescuer solves.
3429 *
3430 * When such condition is possible, the pool summons rescuers of all
3431 * workqueues which have works queued on the pool and let them process
3432 * those works so that forward progress can be guaranteed.
3433 *
3434 * This should happen rarely.
3435 *
3436 * Return: 0
3437 */
3438static int rescuer_thread(void *__rescuer)
3439{
3440 struct worker *rescuer = __rescuer;
3441 struct workqueue_struct *wq = rescuer->rescue_wq;
3442 bool should_stop;
3443
3444 set_user_nice(current, RESCUER_NICE_LEVEL);
3445
3446 /*
3447 * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
3448 * doesn't participate in concurrency management.
3449 */
3450 set_pf_worker(true);
3451repeat:
3452 set_current_state(TASK_IDLE);
3453
3454 /*
3455 * By the time the rescuer is requested to stop, the workqueue
3456 * shouldn't have any work pending, but @wq->maydays may still have
3457 * pwq(s) queued. This can happen by non-rescuer workers consuming
3458 * all the work items before the rescuer got to them. Go through
3459 * @wq->maydays processing before acting on should_stop so that the
3460 * list is always empty on exit.
3461 */
3462 should_stop = kthread_should_stop();
3463
3464 /* see whether any pwq is asking for help */
3465 raw_spin_lock_irq(&wq_mayday_lock);
3466
3467 while (!list_empty(&wq->maydays)) {
3468 struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
3469 struct pool_workqueue, mayday_node);
3470 struct worker_pool *pool = pwq->pool;
3471 struct work_struct *work, *n;
3472
3473 __set_current_state(TASK_RUNNING);
3474 list_del_init(&pwq->mayday_node);
3475
3476 raw_spin_unlock_irq(&wq_mayday_lock);
3477
3478 worker_attach_to_pool(rescuer, pool);
3479
3480 raw_spin_lock_irq(&pool->lock);
3481
3482 /*
3483 * Slurp in all works issued via this workqueue and
3484 * process'em.
3485 */
3486 WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
3487 list_for_each_entry_safe(work, n, &pool->worklist, entry) {
3488 if (get_work_pwq(work) == pwq &&
3489 assign_work(work, rescuer, &n))
3490 pwq->stats[PWQ_STAT_RESCUED]++;
3491 }
3492
3493 if (!list_empty(&rescuer->scheduled)) {
3494 process_scheduled_works(rescuer);
3495
3496 /*
3497 * The above execution of rescued work items could
3498 * have created more to rescue through
3499 * pwq_activate_first_inactive() or chained
3500 * queueing. Let's put @pwq back on mayday list so
3501 * that such back-to-back work items, which may be
3502 * being used to relieve memory pressure, don't
3503 * incur MAYDAY_INTERVAL delay inbetween.
3504 */
3505 if (pwq->nr_active && need_to_create_worker(pool)) {
3506 raw_spin_lock(&wq_mayday_lock);
3507 /*
3508 * Queue iff we aren't racing destruction
3509 * and somebody else hasn't queued it already.
3510 */
3511 if (wq->rescuer && list_empty(&pwq->mayday_node)) {
3512 get_pwq(pwq);
3513 list_add_tail(&pwq->mayday_node, &wq->maydays);
3514 }
3515 raw_spin_unlock(&wq_mayday_lock);
3516 }
3517 }
3518
3519 /*
3520 * Leave this pool. Notify regular workers; otherwise, we end up
3521 * with 0 concurrency and stalling the execution.
3522 */
3523 kick_pool(pool);
3524
3525 raw_spin_unlock_irq(&pool->lock);
3526
3527 worker_detach_from_pool(rescuer);
3528
3529 /*
3530 * Put the reference grabbed by send_mayday(). @pool might
3531 * go away any time after it.
3532 */
3533 put_pwq_unlocked(pwq);
3534
3535 raw_spin_lock_irq(&wq_mayday_lock);
3536 }
3537
3538 raw_spin_unlock_irq(&wq_mayday_lock);
3539
3540 if (should_stop) {
3541 __set_current_state(TASK_RUNNING);
3542 set_pf_worker(false);
3543 return 0;
3544 }
3545
3546 /* rescuers should never participate in concurrency management */
3547 WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
3548 schedule();
3549 goto repeat;
3550}
3551
3552static void bh_worker(struct worker *worker)
3553{
3554 struct worker_pool *pool = worker->pool;
3555 int nr_restarts = BH_WORKER_RESTARTS;
3556 unsigned long end = jiffies + BH_WORKER_JIFFIES;
3557
3558 raw_spin_lock_irq(&pool->lock);
3559 worker_leave_idle(worker);
3560
3561 /*
3562 * This function follows the structure of worker_thread(). See there for
3563 * explanations on each step.
3564 */
3565 if (!need_more_worker(pool))
3566 goto done;
3567
3568 WARN_ON_ONCE(!list_empty(&worker->scheduled));
3569 worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
3570
3571 do {
3572 struct work_struct *work =
3573 list_first_entry(&pool->worklist,
3574 struct work_struct, entry);
3575
3576 if (assign_work(work, worker, NULL))
3577 process_scheduled_works(worker);
3578 } while (keep_working(pool) &&
3579 --nr_restarts && time_before(jiffies, end));
3580
3581 worker_set_flags(worker, WORKER_PREP);
3582done:
3583 worker_enter_idle(worker);
3584 kick_pool(pool);
3585 raw_spin_unlock_irq(&pool->lock);
3586}
3587
3588/*
3589 * TODO: Convert all tasklet users to workqueue and use softirq directly.
3590 *
3591 * This is currently called from tasklet[_hi]action() and thus is also called
3592 * whenever there are tasklets to run. Let's do an early exit if there's nothing
3593 * queued. Once conversion from tasklet is complete, the need_more_worker() test
3594 * can be dropped.
3595 *
3596 * After full conversion, we'll add worker->softirq_action, directly use the
3597 * softirq action and obtain the worker pointer from the softirq_action pointer.
3598 */
3599void workqueue_softirq_action(bool highpri)
3600{
3601 struct worker_pool *pool =
3602 &per_cpu(bh_worker_pools, smp_processor_id())[highpri];
3603 if (need_more_worker(pool))
3604 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3605}
3606
3607struct wq_drain_dead_softirq_work {
3608 struct work_struct work;
3609 struct worker_pool *pool;
3610 struct completion done;
3611};
3612
3613static void drain_dead_softirq_workfn(struct work_struct *work)
3614{
3615 struct wq_drain_dead_softirq_work *dead_work =
3616 container_of(work, struct wq_drain_dead_softirq_work, work);
3617 struct worker_pool *pool = dead_work->pool;
3618 bool repeat;
3619
3620 /*
3621 * @pool's CPU is dead and we want to execute its still pending work
3622 * items from this BH work item which is running on a different CPU. As
3623 * its CPU is dead, @pool can't be kicked and, as work execution path
3624 * will be nested, a lockdep annotation needs to be suppressed. Mark
3625 * @pool with %POOL_BH_DRAINING for the special treatments.
3626 */
3627 raw_spin_lock_irq(&pool->lock);
3628 pool->flags |= POOL_BH_DRAINING;
3629 raw_spin_unlock_irq(&pool->lock);
3630
3631 bh_worker(list_first_entry(&pool->workers, struct worker, node));
3632
3633 raw_spin_lock_irq(&pool->lock);
3634 pool->flags &= ~POOL_BH_DRAINING;
3635 repeat = need_more_worker(pool);
3636 raw_spin_unlock_irq(&pool->lock);
3637
3638 /*
3639 * bh_worker() might hit consecutive execution limit and bail. If there
3640 * still are pending work items, reschedule self and return so that we
3641 * don't hog this CPU's BH.
3642 */
3643 if (repeat) {
3644 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3645 queue_work(system_bh_highpri_wq, work);
3646 else
3647 queue_work(system_bh_wq, work);
3648 } else {
3649 complete(&dead_work->done);
3650 }
3651}
3652
3653/*
3654 * @cpu is dead. Drain the remaining BH work items on the current CPU. It's
3655 * possible to allocate dead_work per CPU and avoid flushing. However, then we
3656 * have to worry about draining overlapping with CPU coming back online or
3657 * nesting (one CPU's dead_work queued on another CPU which is also dead and so
3658 * on). Let's keep it simple and drain them synchronously. These are BH work
3659 * items which shouldn't be requeued on the same pool. Shouldn't take long.
3660 */
3661void workqueue_softirq_dead(unsigned int cpu)
3662{
3663 int i;
3664
3665 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
3666 struct worker_pool *pool = &per_cpu(bh_worker_pools, cpu)[i];
3667 struct wq_drain_dead_softirq_work dead_work;
3668
3669 if (!need_more_worker(pool))
3670 continue;
3671
3672 INIT_WORK_ONSTACK(&dead_work.work, drain_dead_softirq_workfn);
3673 dead_work.pool = pool;
3674 init_completion(&dead_work.done);
3675
3676 if (pool->attrs->nice == HIGHPRI_NICE_LEVEL)
3677 queue_work(system_bh_highpri_wq, &dead_work.work);
3678 else
3679 queue_work(system_bh_wq, &dead_work.work);
3680
3681 wait_for_completion(&dead_work.done);
3682 destroy_work_on_stack(&dead_work.work);
3683 }
3684}
3685
3686/**
3687 * check_flush_dependency - check for flush dependency sanity
3688 * @target_wq: workqueue being flushed
3689 * @target_work: work item being flushed (NULL for workqueue flushes)
3690 * @from_cancel: are we called from the work cancel path
3691 *
3692 * %current is trying to flush the whole @target_wq or @target_work on it.
3693 * If this is not the cancel path (which implies work being flushed is either
3694 * already running, or will not be at all), check if @target_wq doesn't have
3695 * %WQ_MEM_RECLAIM and verify that %current is not reclaiming memory or running
3696 * on a workqueue which doesn't have %WQ_MEM_RECLAIM as that can break forward-
3697 * progress guarantee leading to a deadlock.
3698 */
3699static void check_flush_dependency(struct workqueue_struct *target_wq,
3700 struct work_struct *target_work,
3701 bool from_cancel)
3702{
3703 work_func_t target_func;
3704 struct worker *worker;
3705
3706 if (from_cancel || target_wq->flags & WQ_MEM_RECLAIM)
3707 return;
3708
3709 worker = current_wq_worker();
3710 target_func = target_work ? target_work->func : NULL;
3711
3712 WARN_ONCE(current->flags & PF_MEMALLOC,
3713 "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
3714 current->pid, current->comm, target_wq->name, target_func);
3715 WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
3716 (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
3717 "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
3718 worker->current_pwq->wq->name, worker->current_func,
3719 target_wq->name, target_func);
3720}
3721
3722struct wq_barrier {
3723 struct work_struct work;
3724 struct completion done;
3725 struct task_struct *task; /* purely informational */
3726};
3727
3728static void wq_barrier_func(struct work_struct *work)
3729{
3730 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
3731 complete(&barr->done);
3732}
3733
3734/**
3735 * insert_wq_barrier - insert a barrier work
3736 * @pwq: pwq to insert barrier into
3737 * @barr: wq_barrier to insert
3738 * @target: target work to attach @barr to
3739 * @worker: worker currently executing @target, NULL if @target is not executing
3740 *
3741 * @barr is linked to @target such that @barr is completed only after
3742 * @target finishes execution. Please note that the ordering
3743 * guarantee is observed only with respect to @target and on the local
3744 * cpu.
3745 *
3746 * Currently, a queued barrier can't be canceled. This is because
3747 * try_to_grab_pending() can't determine whether the work to be
3748 * grabbed is at the head of the queue and thus can't clear LINKED
3749 * flag of the previous work while there must be a valid next work
3750 * after a work with LINKED flag set.
3751 *
3752 * Note that when @worker is non-NULL, @target may be modified
3753 * underneath us, so we can't reliably determine pwq from @target.
3754 *
3755 * CONTEXT:
3756 * raw_spin_lock_irq(pool->lock).
3757 */
3758static void insert_wq_barrier(struct pool_workqueue *pwq,
3759 struct wq_barrier *barr,
3760 struct work_struct *target, struct worker *worker)
3761{
3762 static __maybe_unused struct lock_class_key bh_key, thr_key;
3763 unsigned int work_flags = 0;
3764 unsigned int work_color;
3765 struct list_head *head;
3766
3767 /*
3768 * debugobject calls are safe here even with pool->lock locked
3769 * as we know for sure that this will not trigger any of the
3770 * checks and call back into the fixup functions where we
3771 * might deadlock.
3772 *
3773 * BH and threaded workqueues need separate lockdep keys to avoid
3774 * spuriously triggering "inconsistent {SOFTIRQ-ON-W} -> {IN-SOFTIRQ-W}
3775 * usage".
3776 */
3777 INIT_WORK_ONSTACK_KEY(&barr->work, wq_barrier_func,
3778 (pwq->wq->flags & WQ_BH) ? &bh_key : &thr_key);
3779 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
3780
3781 init_completion_map(&barr->done, &target->lockdep_map);
3782
3783 barr->task = current;
3784
3785 /* The barrier work item does not participate in nr_active. */
3786 work_flags |= WORK_STRUCT_INACTIVE;
3787
3788 /*
3789 * If @target is currently being executed, schedule the
3790 * barrier to the worker; otherwise, put it after @target.
3791 */
3792 if (worker) {
3793 head = worker->scheduled.next;
3794 work_color = worker->current_color;
3795 } else {
3796 unsigned long *bits = work_data_bits(target);
3797
3798 head = target->entry.next;
3799 /* there can already be other linked works, inherit and set */
3800 work_flags |= *bits & WORK_STRUCT_LINKED;
3801 work_color = get_work_color(*bits);
3802 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
3803 }
3804
3805 pwq->nr_in_flight[work_color]++;
3806 work_flags |= work_color_to_flags(work_color);
3807
3808 insert_work(pwq, &barr->work, head, work_flags);
3809}
3810
3811/**
3812 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
3813 * @wq: workqueue being flushed
3814 * @flush_color: new flush color, < 0 for no-op
3815 * @work_color: new work color, < 0 for no-op
3816 *
3817 * Prepare pwqs for workqueue flushing.
3818 *
3819 * If @flush_color is non-negative, flush_color on all pwqs should be
3820 * -1. If no pwq has in-flight commands at the specified color, all
3821 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
3822 * has in flight commands, its pwq->flush_color is set to
3823 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
3824 * wakeup logic is armed and %true is returned.
3825 *
3826 * The caller should have initialized @wq->first_flusher prior to
3827 * calling this function with non-negative @flush_color. If
3828 * @flush_color is negative, no flush color update is done and %false
3829 * is returned.
3830 *
3831 * If @work_color is non-negative, all pwqs should have the same
3832 * work_color which is previous to @work_color and all will be
3833 * advanced to @work_color.
3834 *
3835 * CONTEXT:
3836 * mutex_lock(wq->mutex).
3837 *
3838 * Return:
3839 * %true if @flush_color >= 0 and there's something to flush. %false
3840 * otherwise.
3841 */
3842static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
3843 int flush_color, int work_color)
3844{
3845 bool wait = false;
3846 struct pool_workqueue *pwq;
3847 struct worker_pool *current_pool = NULL;
3848
3849 if (flush_color >= 0) {
3850 WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
3851 atomic_set(&wq->nr_pwqs_to_flush, 1);
3852 }
3853
3854 /*
3855 * For unbound workqueue, pwqs will map to only a few pools.
3856 * Most of the time, pwqs within the same pool will be linked
3857 * sequentially to wq->pwqs by cpu index. So in the majority
3858 * of pwq iters, the pool is the same, only doing lock/unlock
3859 * if the pool has changed. This can largely reduce expensive
3860 * lock operations.
3861 */
3862 for_each_pwq(pwq, wq) {
3863 if (current_pool != pwq->pool) {
3864 if (likely(current_pool))
3865 raw_spin_unlock_irq(¤t_pool->lock);
3866 current_pool = pwq->pool;
3867 raw_spin_lock_irq(¤t_pool->lock);
3868 }
3869
3870 if (flush_color >= 0) {
3871 WARN_ON_ONCE(pwq->flush_color != -1);
3872
3873 if (pwq->nr_in_flight[flush_color]) {
3874 pwq->flush_color = flush_color;
3875 atomic_inc(&wq->nr_pwqs_to_flush);
3876 wait = true;
3877 }
3878 }
3879
3880 if (work_color >= 0) {
3881 WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
3882 pwq->work_color = work_color;
3883 }
3884
3885 }
3886
3887 if (current_pool)
3888 raw_spin_unlock_irq(¤t_pool->lock);
3889
3890 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
3891 complete(&wq->first_flusher->done);
3892
3893 return wait;
3894}
3895
3896static void touch_wq_lockdep_map(struct workqueue_struct *wq)
3897{
3898#ifdef CONFIG_LOCKDEP
3899 if (unlikely(!wq->lockdep_map))
3900 return;
3901
3902 if (wq->flags & WQ_BH)
3903 local_bh_disable();
3904
3905 lock_map_acquire(wq->lockdep_map);
3906 lock_map_release(wq->lockdep_map);
3907
3908 if (wq->flags & WQ_BH)
3909 local_bh_enable();
3910#endif
3911}
3912
3913static void touch_work_lockdep_map(struct work_struct *work,
3914 struct workqueue_struct *wq)
3915{
3916#ifdef CONFIG_LOCKDEP
3917 if (wq->flags & WQ_BH)
3918 local_bh_disable();
3919
3920 lock_map_acquire(&work->lockdep_map);
3921 lock_map_release(&work->lockdep_map);
3922
3923 if (wq->flags & WQ_BH)
3924 local_bh_enable();
3925#endif
3926}
3927
3928/**
3929 * __flush_workqueue - ensure that any scheduled work has run to completion.
3930 * @wq: workqueue to flush
3931 *
3932 * This function sleeps until all work items which were queued on entry
3933 * have finished execution, but it is not livelocked by new incoming ones.
3934 */
3935void __flush_workqueue(struct workqueue_struct *wq)
3936{
3937 struct wq_flusher this_flusher = {
3938 .list = LIST_HEAD_INIT(this_flusher.list),
3939 .flush_color = -1,
3940 .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, (*wq->lockdep_map)),
3941 };
3942 int next_color;
3943
3944 if (WARN_ON(!wq_online))
3945 return;
3946
3947 touch_wq_lockdep_map(wq);
3948
3949 mutex_lock(&wq->mutex);
3950
3951 /*
3952 * Start-to-wait phase
3953 */
3954 next_color = work_next_color(wq->work_color);
3955
3956 if (next_color != wq->flush_color) {
3957 /*
3958 * Color space is not full. The current work_color
3959 * becomes our flush_color and work_color is advanced
3960 * by one.
3961 */
3962 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
3963 this_flusher.flush_color = wq->work_color;
3964 wq->work_color = next_color;
3965
3966 if (!wq->first_flusher) {
3967 /* no flush in progress, become the first flusher */
3968 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
3969
3970 wq->first_flusher = &this_flusher;
3971
3972 if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
3973 wq->work_color)) {
3974 /* nothing to flush, done */
3975 wq->flush_color = next_color;
3976 wq->first_flusher = NULL;
3977 goto out_unlock;
3978 }
3979 } else {
3980 /* wait in queue */
3981 WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
3982 list_add_tail(&this_flusher.list, &wq->flusher_queue);
3983 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
3984 }
3985 } else {
3986 /*
3987 * Oops, color space is full, wait on overflow queue.
3988 * The next flush completion will assign us
3989 * flush_color and transfer to flusher_queue.
3990 */
3991 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
3992 }
3993
3994 check_flush_dependency(wq, NULL, false);
3995
3996 mutex_unlock(&wq->mutex);
3997
3998 wait_for_completion(&this_flusher.done);
3999
4000 /*
4001 * Wake-up-and-cascade phase
4002 *
4003 * First flushers are responsible for cascading flushes and
4004 * handling overflow. Non-first flushers can simply return.
4005 */
4006 if (READ_ONCE(wq->first_flusher) != &this_flusher)
4007 return;
4008
4009 mutex_lock(&wq->mutex);
4010
4011 /* we might have raced, check again with mutex held */
4012 if (wq->first_flusher != &this_flusher)
4013 goto out_unlock;
4014
4015 WRITE_ONCE(wq->first_flusher, NULL);
4016
4017 WARN_ON_ONCE(!list_empty(&this_flusher.list));
4018 WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
4019
4020 while (true) {
4021 struct wq_flusher *next, *tmp;
4022
4023 /* complete all the flushers sharing the current flush color */
4024 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
4025 if (next->flush_color != wq->flush_color)
4026 break;
4027 list_del_init(&next->list);
4028 complete(&next->done);
4029 }
4030
4031 WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
4032 wq->flush_color != work_next_color(wq->work_color));
4033
4034 /* this flush_color is finished, advance by one */
4035 wq->flush_color = work_next_color(wq->flush_color);
4036
4037 /* one color has been freed, handle overflow queue */
4038 if (!list_empty(&wq->flusher_overflow)) {
4039 /*
4040 * Assign the same color to all overflowed
4041 * flushers, advance work_color and append to
4042 * flusher_queue. This is the start-to-wait
4043 * phase for these overflowed flushers.
4044 */
4045 list_for_each_entry(tmp, &wq->flusher_overflow, list)
4046 tmp->flush_color = wq->work_color;
4047
4048 wq->work_color = work_next_color(wq->work_color);
4049
4050 list_splice_tail_init(&wq->flusher_overflow,
4051 &wq->flusher_queue);
4052 flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
4053 }
4054
4055 if (list_empty(&wq->flusher_queue)) {
4056 WARN_ON_ONCE(wq->flush_color != wq->work_color);
4057 break;
4058 }
4059
4060 /*
4061 * Need to flush more colors. Make the next flusher
4062 * the new first flusher and arm pwqs.
4063 */
4064 WARN_ON_ONCE(wq->flush_color == wq->work_color);
4065 WARN_ON_ONCE(wq->flush_color != next->flush_color);
4066
4067 list_del_init(&next->list);
4068 wq->first_flusher = next;
4069
4070 if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
4071 break;
4072
4073 /*
4074 * Meh... this color is already done, clear first
4075 * flusher and repeat cascading.
4076 */
4077 wq->first_flusher = NULL;
4078 }
4079
4080out_unlock:
4081 mutex_unlock(&wq->mutex);
4082}
4083EXPORT_SYMBOL(__flush_workqueue);
4084
4085/**
4086 * drain_workqueue - drain a workqueue
4087 * @wq: workqueue to drain
4088 *
4089 * Wait until the workqueue becomes empty. While draining is in progress,
4090 * only chain queueing is allowed. IOW, only currently pending or running
4091 * work items on @wq can queue further work items on it. @wq is flushed
4092 * repeatedly until it becomes empty. The number of flushing is determined
4093 * by the depth of chaining and should be relatively short. Whine if it
4094 * takes too long.
4095 */
4096void drain_workqueue(struct workqueue_struct *wq)
4097{
4098 unsigned int flush_cnt = 0;
4099 struct pool_workqueue *pwq;
4100
4101 /*
4102 * __queue_work() needs to test whether there are drainers, is much
4103 * hotter than drain_workqueue() and already looks at @wq->flags.
4104 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
4105 */
4106 mutex_lock(&wq->mutex);
4107 if (!wq->nr_drainers++)
4108 wq->flags |= __WQ_DRAINING;
4109 mutex_unlock(&wq->mutex);
4110reflush:
4111 __flush_workqueue(wq);
4112
4113 mutex_lock(&wq->mutex);
4114
4115 for_each_pwq(pwq, wq) {
4116 bool drained;
4117
4118 raw_spin_lock_irq(&pwq->pool->lock);
4119 drained = pwq_is_empty(pwq);
4120 raw_spin_unlock_irq(&pwq->pool->lock);
4121
4122 if (drained)
4123 continue;
4124
4125 if (++flush_cnt == 10 ||
4126 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
4127 pr_warn("workqueue %s: %s() isn't complete after %u tries\n",
4128 wq->name, __func__, flush_cnt);
4129
4130 mutex_unlock(&wq->mutex);
4131 goto reflush;
4132 }
4133
4134 if (!--wq->nr_drainers)
4135 wq->flags &= ~__WQ_DRAINING;
4136 mutex_unlock(&wq->mutex);
4137}
4138EXPORT_SYMBOL_GPL(drain_workqueue);
4139
4140static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
4141 bool from_cancel)
4142{
4143 struct worker *worker = NULL;
4144 struct worker_pool *pool;
4145 struct pool_workqueue *pwq;
4146 struct workqueue_struct *wq;
4147
4148 rcu_read_lock();
4149 pool = get_work_pool(work);
4150 if (!pool) {
4151 rcu_read_unlock();
4152 return false;
4153 }
4154
4155 raw_spin_lock_irq(&pool->lock);
4156 /* see the comment in try_to_grab_pending() with the same code */
4157 pwq = get_work_pwq(work);
4158 if (pwq) {
4159 if (unlikely(pwq->pool != pool))
4160 goto already_gone;
4161 } else {
4162 worker = find_worker_executing_work(pool, work);
4163 if (!worker)
4164 goto already_gone;
4165 pwq = worker->current_pwq;
4166 }
4167
4168 wq = pwq->wq;
4169 check_flush_dependency(wq, work, from_cancel);
4170
4171 insert_wq_barrier(pwq, barr, work, worker);
4172 raw_spin_unlock_irq(&pool->lock);
4173
4174 touch_work_lockdep_map(work, wq);
4175
4176 /*
4177 * Force a lock recursion deadlock when using flush_work() inside a
4178 * single-threaded or rescuer equipped workqueue.
4179 *
4180 * For single threaded workqueues the deadlock happens when the work
4181 * is after the work issuing the flush_work(). For rescuer equipped
4182 * workqueues the deadlock happens when the rescuer stalls, blocking
4183 * forward progress.
4184 */
4185 if (!from_cancel && (wq->saved_max_active == 1 || wq->rescuer))
4186 touch_wq_lockdep_map(wq);
4187
4188 rcu_read_unlock();
4189 return true;
4190already_gone:
4191 raw_spin_unlock_irq(&pool->lock);
4192 rcu_read_unlock();
4193 return false;
4194}
4195
4196static bool __flush_work(struct work_struct *work, bool from_cancel)
4197{
4198 struct wq_barrier barr;
4199
4200 if (WARN_ON(!wq_online))
4201 return false;
4202
4203 if (WARN_ON(!work->func))
4204 return false;
4205
4206 if (!start_flush_work(work, &barr, from_cancel))
4207 return false;
4208
4209 /*
4210 * start_flush_work() returned %true. If @from_cancel is set, we know
4211 * that @work must have been executing during start_flush_work() and
4212 * can't currently be queued. Its data must contain OFFQ bits. If @work
4213 * was queued on a BH workqueue, we also know that it was running in the
4214 * BH context and thus can be busy-waited.
4215 */
4216 if (from_cancel) {
4217 unsigned long data = *work_data_bits(work);
4218
4219 if (!WARN_ON_ONCE(data & WORK_STRUCT_PWQ) &&
4220 (data & WORK_OFFQ_BH)) {
4221 /*
4222 * On RT, prevent a live lock when %current preempted
4223 * soft interrupt processing or prevents ksoftirqd from
4224 * running by keeping flipping BH. If the BH work item
4225 * runs on a different CPU then this has no effect other
4226 * than doing the BH disable/enable dance for nothing.
4227 * This is copied from
4228 * kernel/softirq.c::tasklet_unlock_spin_wait().
4229 */
4230 while (!try_wait_for_completion(&barr.done)) {
4231 if (IS_ENABLED(CONFIG_PREEMPT_RT)) {
4232 local_bh_disable();
4233 local_bh_enable();
4234 } else {
4235 cpu_relax();
4236 }
4237 }
4238 goto out_destroy;
4239 }
4240 }
4241
4242 wait_for_completion(&barr.done);
4243
4244out_destroy:
4245 destroy_work_on_stack(&barr.work);
4246 return true;
4247}
4248
4249/**
4250 * flush_work - wait for a work to finish executing the last queueing instance
4251 * @work: the work to flush
4252 *
4253 * Wait until @work has finished execution. @work is guaranteed to be idle
4254 * on return if it hasn't been requeued since flush started.
4255 *
4256 * Return:
4257 * %true if flush_work() waited for the work to finish execution,
4258 * %false if it was already idle.
4259 */
4260bool flush_work(struct work_struct *work)
4261{
4262 might_sleep();
4263 return __flush_work(work, false);
4264}
4265EXPORT_SYMBOL_GPL(flush_work);
4266
4267/**
4268 * flush_delayed_work - wait for a dwork to finish executing the last queueing
4269 * @dwork: the delayed work to flush
4270 *
4271 * Delayed timer is cancelled and the pending work is queued for
4272 * immediate execution. Like flush_work(), this function only
4273 * considers the last queueing instance of @dwork.
4274 *
4275 * Return:
4276 * %true if flush_work() waited for the work to finish execution,
4277 * %false if it was already idle.
4278 */
4279bool flush_delayed_work(struct delayed_work *dwork)
4280{
4281 local_irq_disable();
4282 if (del_timer_sync(&dwork->timer))
4283 __queue_work(dwork->cpu, dwork->wq, &dwork->work);
4284 local_irq_enable();
4285 return flush_work(&dwork->work);
4286}
4287EXPORT_SYMBOL(flush_delayed_work);
4288
4289/**
4290 * flush_rcu_work - wait for a rwork to finish executing the last queueing
4291 * @rwork: the rcu work to flush
4292 *
4293 * Return:
4294 * %true if flush_rcu_work() waited for the work to finish execution,
4295 * %false if it was already idle.
4296 */
4297bool flush_rcu_work(struct rcu_work *rwork)
4298{
4299 if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
4300 rcu_barrier();
4301 flush_work(&rwork->work);
4302 return true;
4303 } else {
4304 return flush_work(&rwork->work);
4305 }
4306}
4307EXPORT_SYMBOL(flush_rcu_work);
4308
4309static void work_offqd_disable(struct work_offq_data *offqd)
4310{
4311 const unsigned long max = (1lu << WORK_OFFQ_DISABLE_BITS) - 1;
4312
4313 if (likely(offqd->disable < max))
4314 offqd->disable++;
4315 else
4316 WARN_ONCE(true, "workqueue: work disable count overflowed\n");
4317}
4318
4319static void work_offqd_enable(struct work_offq_data *offqd)
4320{
4321 if (likely(offqd->disable > 0))
4322 offqd->disable--;
4323 else
4324 WARN_ONCE(true, "workqueue: work disable count underflowed\n");
4325}
4326
4327static bool __cancel_work(struct work_struct *work, u32 cflags)
4328{
4329 struct work_offq_data offqd;
4330 unsigned long irq_flags;
4331 int ret;
4332
4333 ret = work_grab_pending(work, cflags, &irq_flags);
4334
4335 work_offqd_unpack(&offqd, *work_data_bits(work));
4336
4337 if (cflags & WORK_CANCEL_DISABLE)
4338 work_offqd_disable(&offqd);
4339
4340 set_work_pool_and_clear_pending(work, offqd.pool_id,
4341 work_offqd_pack_flags(&offqd));
4342 local_irq_restore(irq_flags);
4343 return ret;
4344}
4345
4346static bool __cancel_work_sync(struct work_struct *work, u32 cflags)
4347{
4348 bool ret;
4349
4350 ret = __cancel_work(work, cflags | WORK_CANCEL_DISABLE);
4351
4352 if (*work_data_bits(work) & WORK_OFFQ_BH)
4353 WARN_ON_ONCE(in_hardirq());
4354 else
4355 might_sleep();
4356
4357 /*
4358 * Skip __flush_work() during early boot when we know that @work isn't
4359 * executing. This allows canceling during early boot.
4360 */
4361 if (wq_online)
4362 __flush_work(work, true);
4363
4364 if (!(cflags & WORK_CANCEL_DISABLE))
4365 enable_work(work);
4366
4367 return ret;
4368}
4369
4370/*
4371 * See cancel_delayed_work()
4372 */
4373bool cancel_work(struct work_struct *work)
4374{
4375 return __cancel_work(work, 0);
4376}
4377EXPORT_SYMBOL(cancel_work);
4378
4379/**
4380 * cancel_work_sync - cancel a work and wait for it to finish
4381 * @work: the work to cancel
4382 *
4383 * Cancel @work and wait for its execution to finish. This function can be used
4384 * even if the work re-queues itself or migrates to another workqueue. On return
4385 * from this function, @work is guaranteed to be not pending or executing on any
4386 * CPU as long as there aren't racing enqueues.
4387 *
4388 * cancel_work_sync(&delayed_work->work) must not be used for delayed_work's.
4389 * Use cancel_delayed_work_sync() instead.
4390 *
4391 * Must be called from a sleepable context if @work was last queued on a non-BH
4392 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4393 * if @work was last queued on a BH workqueue.
4394 *
4395 * Returns %true if @work was pending, %false otherwise.
4396 */
4397bool cancel_work_sync(struct work_struct *work)
4398{
4399 return __cancel_work_sync(work, 0);
4400}
4401EXPORT_SYMBOL_GPL(cancel_work_sync);
4402
4403/**
4404 * cancel_delayed_work - cancel a delayed work
4405 * @dwork: delayed_work to cancel
4406 *
4407 * Kill off a pending delayed_work.
4408 *
4409 * Return: %true if @dwork was pending and canceled; %false if it wasn't
4410 * pending.
4411 *
4412 * Note:
4413 * The work callback function may still be running on return, unless
4414 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
4415 * use cancel_delayed_work_sync() to wait on it.
4416 *
4417 * This function is safe to call from any context including IRQ handler.
4418 */
4419bool cancel_delayed_work(struct delayed_work *dwork)
4420{
4421 return __cancel_work(&dwork->work, WORK_CANCEL_DELAYED);
4422}
4423EXPORT_SYMBOL(cancel_delayed_work);
4424
4425/**
4426 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
4427 * @dwork: the delayed work cancel
4428 *
4429 * This is cancel_work_sync() for delayed works.
4430 *
4431 * Return:
4432 * %true if @dwork was pending, %false otherwise.
4433 */
4434bool cancel_delayed_work_sync(struct delayed_work *dwork)
4435{
4436 return __cancel_work_sync(&dwork->work, WORK_CANCEL_DELAYED);
4437}
4438EXPORT_SYMBOL(cancel_delayed_work_sync);
4439
4440/**
4441 * disable_work - Disable and cancel a work item
4442 * @work: work item to disable
4443 *
4444 * Disable @work by incrementing its disable count and cancel it if currently
4445 * pending. As long as the disable count is non-zero, any attempt to queue @work
4446 * will fail and return %false. The maximum supported disable depth is 2 to the
4447 * power of %WORK_OFFQ_DISABLE_BITS, currently 65536.
4448 *
4449 * Can be called from any context. Returns %true if @work was pending, %false
4450 * otherwise.
4451 */
4452bool disable_work(struct work_struct *work)
4453{
4454 return __cancel_work(work, WORK_CANCEL_DISABLE);
4455}
4456EXPORT_SYMBOL_GPL(disable_work);
4457
4458/**
4459 * disable_work_sync - Disable, cancel and drain a work item
4460 * @work: work item to disable
4461 *
4462 * Similar to disable_work() but also wait for @work to finish if currently
4463 * executing.
4464 *
4465 * Must be called from a sleepable context if @work was last queued on a non-BH
4466 * workqueue. Can also be called from non-hardirq atomic contexts including BH
4467 * if @work was last queued on a BH workqueue.
4468 *
4469 * Returns %true if @work was pending, %false otherwise.
4470 */
4471bool disable_work_sync(struct work_struct *work)
4472{
4473 return __cancel_work_sync(work, WORK_CANCEL_DISABLE);
4474}
4475EXPORT_SYMBOL_GPL(disable_work_sync);
4476
4477/**
4478 * enable_work - Enable a work item
4479 * @work: work item to enable
4480 *
4481 * Undo disable_work[_sync]() by decrementing @work's disable count. @work can
4482 * only be queued if its disable count is 0.
4483 *
4484 * Can be called from any context. Returns %true if the disable count reached 0.
4485 * Otherwise, %false.
4486 */
4487bool enable_work(struct work_struct *work)
4488{
4489 struct work_offq_data offqd;
4490 unsigned long irq_flags;
4491
4492 work_grab_pending(work, 0, &irq_flags);
4493
4494 work_offqd_unpack(&offqd, *work_data_bits(work));
4495 work_offqd_enable(&offqd);
4496 set_work_pool_and_clear_pending(work, offqd.pool_id,
4497 work_offqd_pack_flags(&offqd));
4498 local_irq_restore(irq_flags);
4499
4500 return !offqd.disable;
4501}
4502EXPORT_SYMBOL_GPL(enable_work);
4503
4504/**
4505 * disable_delayed_work - Disable and cancel a delayed work item
4506 * @dwork: delayed work item to disable
4507 *
4508 * disable_work() for delayed work items.
4509 */
4510bool disable_delayed_work(struct delayed_work *dwork)
4511{
4512 return __cancel_work(&dwork->work,
4513 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4514}
4515EXPORT_SYMBOL_GPL(disable_delayed_work);
4516
4517/**
4518 * disable_delayed_work_sync - Disable, cancel and drain a delayed work item
4519 * @dwork: delayed work item to disable
4520 *
4521 * disable_work_sync() for delayed work items.
4522 */
4523bool disable_delayed_work_sync(struct delayed_work *dwork)
4524{
4525 return __cancel_work_sync(&dwork->work,
4526 WORK_CANCEL_DELAYED | WORK_CANCEL_DISABLE);
4527}
4528EXPORT_SYMBOL_GPL(disable_delayed_work_sync);
4529
4530/**
4531 * enable_delayed_work - Enable a delayed work item
4532 * @dwork: delayed work item to enable
4533 *
4534 * enable_work() for delayed work items.
4535 */
4536bool enable_delayed_work(struct delayed_work *dwork)
4537{
4538 return enable_work(&dwork->work);
4539}
4540EXPORT_SYMBOL_GPL(enable_delayed_work);
4541
4542/**
4543 * schedule_on_each_cpu - execute a function synchronously on each online CPU
4544 * @func: the function to call
4545 *
4546 * schedule_on_each_cpu() executes @func on each online CPU using the
4547 * system workqueue and blocks until all CPUs have completed.
4548 * schedule_on_each_cpu() is very slow.
4549 *
4550 * Return:
4551 * 0 on success, -errno on failure.
4552 */
4553int schedule_on_each_cpu(work_func_t func)
4554{
4555 int cpu;
4556 struct work_struct __percpu *works;
4557
4558 works = alloc_percpu(struct work_struct);
4559 if (!works)
4560 return -ENOMEM;
4561
4562 cpus_read_lock();
4563
4564 for_each_online_cpu(cpu) {
4565 struct work_struct *work = per_cpu_ptr(works, cpu);
4566
4567 INIT_WORK(work, func);
4568 schedule_work_on(cpu, work);
4569 }
4570
4571 for_each_online_cpu(cpu)
4572 flush_work(per_cpu_ptr(works, cpu));
4573
4574 cpus_read_unlock();
4575 free_percpu(works);
4576 return 0;
4577}
4578
4579/**
4580 * execute_in_process_context - reliably execute the routine with user context
4581 * @fn: the function to execute
4582 * @ew: guaranteed storage for the execute work structure (must
4583 * be available when the work executes)
4584 *
4585 * Executes the function immediately if process context is available,
4586 * otherwise schedules the function for delayed execution.
4587 *
4588 * Return: 0 - function was executed
4589 * 1 - function was scheduled for execution
4590 */
4591int execute_in_process_context(work_func_t fn, struct execute_work *ew)
4592{
4593 if (!in_interrupt()) {
4594 fn(&ew->work);
4595 return 0;
4596 }
4597
4598 INIT_WORK(&ew->work, fn);
4599 schedule_work(&ew->work);
4600
4601 return 1;
4602}
4603EXPORT_SYMBOL_GPL(execute_in_process_context);
4604
4605/**
4606 * free_workqueue_attrs - free a workqueue_attrs
4607 * @attrs: workqueue_attrs to free
4608 *
4609 * Undo alloc_workqueue_attrs().
4610 */
4611void free_workqueue_attrs(struct workqueue_attrs *attrs)
4612{
4613 if (attrs) {
4614 free_cpumask_var(attrs->cpumask);
4615 free_cpumask_var(attrs->__pod_cpumask);
4616 kfree(attrs);
4617 }
4618}
4619
4620/**
4621 * alloc_workqueue_attrs - allocate a workqueue_attrs
4622 *
4623 * Allocate a new workqueue_attrs, initialize with default settings and
4624 * return it.
4625 *
4626 * Return: The allocated new workqueue_attr on success. %NULL on failure.
4627 */
4628struct workqueue_attrs *alloc_workqueue_attrs(void)
4629{
4630 struct workqueue_attrs *attrs;
4631
4632 attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
4633 if (!attrs)
4634 goto fail;
4635 if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
4636 goto fail;
4637 if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL))
4638 goto fail;
4639
4640 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4641 attrs->affn_scope = WQ_AFFN_DFL;
4642 return attrs;
4643fail:
4644 free_workqueue_attrs(attrs);
4645 return NULL;
4646}
4647
4648static void copy_workqueue_attrs(struct workqueue_attrs *to,
4649 const struct workqueue_attrs *from)
4650{
4651 to->nice = from->nice;
4652 cpumask_copy(to->cpumask, from->cpumask);
4653 cpumask_copy(to->__pod_cpumask, from->__pod_cpumask);
4654 to->affn_strict = from->affn_strict;
4655
4656 /*
4657 * Unlike hash and equality test, copying shouldn't ignore wq-only
4658 * fields as copying is used for both pool and wq attrs. Instead,
4659 * get_unbound_pool() explicitly clears the fields.
4660 */
4661 to->affn_scope = from->affn_scope;
4662 to->ordered = from->ordered;
4663}
4664
4665/*
4666 * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the
4667 * comments in 'struct workqueue_attrs' definition.
4668 */
4669static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs)
4670{
4671 attrs->affn_scope = WQ_AFFN_NR_TYPES;
4672 attrs->ordered = false;
4673 if (attrs->affn_strict)
4674 cpumask_copy(attrs->cpumask, cpu_possible_mask);
4675}
4676
4677/* hash value of the content of @attr */
4678static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
4679{
4680 u32 hash = 0;
4681
4682 hash = jhash_1word(attrs->nice, hash);
4683 hash = jhash_1word(attrs->affn_strict, hash);
4684 hash = jhash(cpumask_bits(attrs->__pod_cpumask),
4685 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4686 if (!attrs->affn_strict)
4687 hash = jhash(cpumask_bits(attrs->cpumask),
4688 BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
4689 return hash;
4690}
4691
4692/* content equality test */
4693static bool wqattrs_equal(const struct workqueue_attrs *a,
4694 const struct workqueue_attrs *b)
4695{
4696 if (a->nice != b->nice)
4697 return false;
4698 if (a->affn_strict != b->affn_strict)
4699 return false;
4700 if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask))
4701 return false;
4702 if (!a->affn_strict && !cpumask_equal(a->cpumask, b->cpumask))
4703 return false;
4704 return true;
4705}
4706
4707/* Update @attrs with actually available CPUs */
4708static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs,
4709 const cpumask_t *unbound_cpumask)
4710{
4711 /*
4712 * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If
4713 * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to
4714 * @unbound_cpumask.
4715 */
4716 cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask);
4717 if (unlikely(cpumask_empty(attrs->cpumask)))
4718 cpumask_copy(attrs->cpumask, unbound_cpumask);
4719}
4720
4721/* find wq_pod_type to use for @attrs */
4722static const struct wq_pod_type *
4723wqattrs_pod_type(const struct workqueue_attrs *attrs)
4724{
4725 enum wq_affn_scope scope;
4726 struct wq_pod_type *pt;
4727
4728 /* to synchronize access to wq_affn_dfl */
4729 lockdep_assert_held(&wq_pool_mutex);
4730
4731 if (attrs->affn_scope == WQ_AFFN_DFL)
4732 scope = wq_affn_dfl;
4733 else
4734 scope = attrs->affn_scope;
4735
4736 pt = &wq_pod_types[scope];
4737
4738 if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) &&
4739 likely(pt->nr_pods))
4740 return pt;
4741
4742 /*
4743 * Before workqueue_init_topology(), only SYSTEM is available which is
4744 * initialized in workqueue_init_early().
4745 */
4746 pt = &wq_pod_types[WQ_AFFN_SYSTEM];
4747 BUG_ON(!pt->nr_pods);
4748 return pt;
4749}
4750
4751/**
4752 * init_worker_pool - initialize a newly zalloc'd worker_pool
4753 * @pool: worker_pool to initialize
4754 *
4755 * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs.
4756 *
4757 * Return: 0 on success, -errno on failure. Even on failure, all fields
4758 * inside @pool proper are initialized and put_unbound_pool() can be called
4759 * on @pool safely to release it.
4760 */
4761static int init_worker_pool(struct worker_pool *pool)
4762{
4763 raw_spin_lock_init(&pool->lock);
4764 pool->id = -1;
4765 pool->cpu = -1;
4766 pool->node = NUMA_NO_NODE;
4767 pool->flags |= POOL_DISASSOCIATED;
4768 pool->watchdog_ts = jiffies;
4769 INIT_LIST_HEAD(&pool->worklist);
4770 INIT_LIST_HEAD(&pool->idle_list);
4771 hash_init(pool->busy_hash);
4772
4773 timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
4774 INIT_WORK(&pool->idle_cull_work, idle_cull_fn);
4775
4776 timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
4777
4778 INIT_LIST_HEAD(&pool->workers);
4779
4780 ida_init(&pool->worker_ida);
4781 INIT_HLIST_NODE(&pool->hash_node);
4782 pool->refcnt = 1;
4783
4784 /* shouldn't fail above this point */
4785 pool->attrs = alloc_workqueue_attrs();
4786 if (!pool->attrs)
4787 return -ENOMEM;
4788
4789 wqattrs_clear_for_pool(pool->attrs);
4790
4791 return 0;
4792}
4793
4794#ifdef CONFIG_LOCKDEP
4795static void wq_init_lockdep(struct workqueue_struct *wq)
4796{
4797 char *lock_name;
4798
4799 lockdep_register_key(&wq->key);
4800 lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
4801 if (!lock_name)
4802 lock_name = wq->name;
4803
4804 wq->lock_name = lock_name;
4805 wq->lockdep_map = &wq->__lockdep_map;
4806 lockdep_init_map(wq->lockdep_map, lock_name, &wq->key, 0);
4807}
4808
4809static void wq_unregister_lockdep(struct workqueue_struct *wq)
4810{
4811 if (wq->lockdep_map != &wq->__lockdep_map)
4812 return;
4813
4814 lockdep_unregister_key(&wq->key);
4815}
4816
4817static void wq_free_lockdep(struct workqueue_struct *wq)
4818{
4819 if (wq->lockdep_map != &wq->__lockdep_map)
4820 return;
4821
4822 if (wq->lock_name != wq->name)
4823 kfree(wq->lock_name);
4824}
4825#else
4826static void wq_init_lockdep(struct workqueue_struct *wq)
4827{
4828}
4829
4830static void wq_unregister_lockdep(struct workqueue_struct *wq)
4831{
4832}
4833
4834static void wq_free_lockdep(struct workqueue_struct *wq)
4835{
4836}
4837#endif
4838
4839static void free_node_nr_active(struct wq_node_nr_active **nna_ar)
4840{
4841 int node;
4842
4843 for_each_node(node) {
4844 kfree(nna_ar[node]);
4845 nna_ar[node] = NULL;
4846 }
4847
4848 kfree(nna_ar[nr_node_ids]);
4849 nna_ar[nr_node_ids] = NULL;
4850}
4851
4852static void init_node_nr_active(struct wq_node_nr_active *nna)
4853{
4854 nna->max = WQ_DFL_MIN_ACTIVE;
4855 atomic_set(&nna->nr, 0);
4856 raw_spin_lock_init(&nna->lock);
4857 INIT_LIST_HEAD(&nna->pending_pwqs);
4858}
4859
4860/*
4861 * Each node's nr_active counter will be accessed mostly from its own node and
4862 * should be allocated in the node.
4863 */
4864static int alloc_node_nr_active(struct wq_node_nr_active **nna_ar)
4865{
4866 struct wq_node_nr_active *nna;
4867 int node;
4868
4869 for_each_node(node) {
4870 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, node);
4871 if (!nna)
4872 goto err_free;
4873 init_node_nr_active(nna);
4874 nna_ar[node] = nna;
4875 }
4876
4877 /* [nr_node_ids] is used as the fallback */
4878 nna = kzalloc_node(sizeof(*nna), GFP_KERNEL, NUMA_NO_NODE);
4879 if (!nna)
4880 goto err_free;
4881 init_node_nr_active(nna);
4882 nna_ar[nr_node_ids] = nna;
4883
4884 return 0;
4885
4886err_free:
4887 free_node_nr_active(nna_ar);
4888 return -ENOMEM;
4889}
4890
4891static void rcu_free_wq(struct rcu_head *rcu)
4892{
4893 struct workqueue_struct *wq =
4894 container_of(rcu, struct workqueue_struct, rcu);
4895
4896 if (wq->flags & WQ_UNBOUND)
4897 free_node_nr_active(wq->node_nr_active);
4898
4899 wq_free_lockdep(wq);
4900 free_percpu(wq->cpu_pwq);
4901 free_workqueue_attrs(wq->unbound_attrs);
4902 kfree(wq);
4903}
4904
4905static void rcu_free_pool(struct rcu_head *rcu)
4906{
4907 struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
4908
4909 ida_destroy(&pool->worker_ida);
4910 free_workqueue_attrs(pool->attrs);
4911 kfree(pool);
4912}
4913
4914/**
4915 * put_unbound_pool - put a worker_pool
4916 * @pool: worker_pool to put
4917 *
4918 * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU
4919 * safe manner. get_unbound_pool() calls this function on its failure path
4920 * and this function should be able to release pools which went through,
4921 * successfully or not, init_worker_pool().
4922 *
4923 * Should be called with wq_pool_mutex held.
4924 */
4925static void put_unbound_pool(struct worker_pool *pool)
4926{
4927 struct worker *worker;
4928 LIST_HEAD(cull_list);
4929
4930 lockdep_assert_held(&wq_pool_mutex);
4931
4932 if (--pool->refcnt)
4933 return;
4934
4935 /* sanity checks */
4936 if (WARN_ON(!(pool->cpu < 0)) ||
4937 WARN_ON(!list_empty(&pool->worklist)))
4938 return;
4939
4940 /* release id and unhash */
4941 if (pool->id >= 0)
4942 idr_remove(&worker_pool_idr, pool->id);
4943 hash_del(&pool->hash_node);
4944
4945 /*
4946 * Become the manager and destroy all workers. This prevents
4947 * @pool's workers from blocking on attach_mutex. We're the last
4948 * manager and @pool gets freed with the flag set.
4949 *
4950 * Having a concurrent manager is quite unlikely to happen as we can
4951 * only get here with
4952 * pwq->refcnt == pool->refcnt == 0
4953 * which implies no work queued to the pool, which implies no worker can
4954 * become the manager. However a worker could have taken the role of
4955 * manager before the refcnts dropped to 0, since maybe_create_worker()
4956 * drops pool->lock
4957 */
4958 while (true) {
4959 rcuwait_wait_event(&manager_wait,
4960 !(pool->flags & POOL_MANAGER_ACTIVE),
4961 TASK_UNINTERRUPTIBLE);
4962
4963 mutex_lock(&wq_pool_attach_mutex);
4964 raw_spin_lock_irq(&pool->lock);
4965 if (!(pool->flags & POOL_MANAGER_ACTIVE)) {
4966 pool->flags |= POOL_MANAGER_ACTIVE;
4967 break;
4968 }
4969 raw_spin_unlock_irq(&pool->lock);
4970 mutex_unlock(&wq_pool_attach_mutex);
4971 }
4972
4973 while ((worker = first_idle_worker(pool)))
4974 set_worker_dying(worker, &cull_list);
4975 WARN_ON(pool->nr_workers || pool->nr_idle);
4976 raw_spin_unlock_irq(&pool->lock);
4977
4978 detach_dying_workers(&cull_list);
4979
4980 mutex_unlock(&wq_pool_attach_mutex);
4981
4982 reap_dying_workers(&cull_list);
4983
4984 /* shut down the timers */
4985 del_timer_sync(&pool->idle_timer);
4986 cancel_work_sync(&pool->idle_cull_work);
4987 del_timer_sync(&pool->mayday_timer);
4988
4989 /* RCU protected to allow dereferences from get_work_pool() */
4990 call_rcu(&pool->rcu, rcu_free_pool);
4991}
4992
4993/**
4994 * get_unbound_pool - get a worker_pool with the specified attributes
4995 * @attrs: the attributes of the worker_pool to get
4996 *
4997 * Obtain a worker_pool which has the same attributes as @attrs, bump the
4998 * reference count and return it. If there already is a matching
4999 * worker_pool, it will be used; otherwise, this function attempts to
5000 * create a new one.
5001 *
5002 * Should be called with wq_pool_mutex held.
5003 *
5004 * Return: On success, a worker_pool with the same attributes as @attrs.
5005 * On failure, %NULL.
5006 */
5007static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
5008{
5009 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA];
5010 u32 hash = wqattrs_hash(attrs);
5011 struct worker_pool *pool;
5012 int pod, node = NUMA_NO_NODE;
5013
5014 lockdep_assert_held(&wq_pool_mutex);
5015
5016 /* do we already have a matching pool? */
5017 hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
5018 if (wqattrs_equal(pool->attrs, attrs)) {
5019 pool->refcnt++;
5020 return pool;
5021 }
5022 }
5023
5024 /* If __pod_cpumask is contained inside a NUMA pod, that's our node */
5025 for (pod = 0; pod < pt->nr_pods; pod++) {
5026 if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) {
5027 node = pt->pod_node[pod];
5028 break;
5029 }
5030 }
5031
5032 /* nope, create a new one */
5033 pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
5034 if (!pool || init_worker_pool(pool) < 0)
5035 goto fail;
5036
5037 pool->node = node;
5038 copy_workqueue_attrs(pool->attrs, attrs);
5039 wqattrs_clear_for_pool(pool->attrs);
5040
5041 if (worker_pool_assign_id(pool) < 0)
5042 goto fail;
5043
5044 /* create and start the initial worker */
5045 if (wq_online && !create_worker(pool))
5046 goto fail;
5047
5048 /* install */
5049 hash_add(unbound_pool_hash, &pool->hash_node, hash);
5050
5051 return pool;
5052fail:
5053 if (pool)
5054 put_unbound_pool(pool);
5055 return NULL;
5056}
5057
5058/*
5059 * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero
5060 * refcnt and needs to be destroyed.
5061 */
5062static void pwq_release_workfn(struct kthread_work *work)
5063{
5064 struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
5065 release_work);
5066 struct workqueue_struct *wq = pwq->wq;
5067 struct worker_pool *pool = pwq->pool;
5068 bool is_last = false;
5069
5070 /*
5071 * When @pwq is not linked, it doesn't hold any reference to the
5072 * @wq, and @wq is invalid to access.
5073 */
5074 if (!list_empty(&pwq->pwqs_node)) {
5075 mutex_lock(&wq->mutex);
5076 list_del_rcu(&pwq->pwqs_node);
5077 is_last = list_empty(&wq->pwqs);
5078
5079 /*
5080 * For ordered workqueue with a plugged dfl_pwq, restart it now.
5081 */
5082 if (!is_last && (wq->flags & __WQ_ORDERED))
5083 unplug_oldest_pwq(wq);
5084
5085 mutex_unlock(&wq->mutex);
5086 }
5087
5088 if (wq->flags & WQ_UNBOUND) {
5089 mutex_lock(&wq_pool_mutex);
5090 put_unbound_pool(pool);
5091 mutex_unlock(&wq_pool_mutex);
5092 }
5093
5094 if (!list_empty(&pwq->pending_node)) {
5095 struct wq_node_nr_active *nna =
5096 wq_node_nr_active(pwq->wq, pwq->pool->node);
5097
5098 raw_spin_lock_irq(&nna->lock);
5099 list_del_init(&pwq->pending_node);
5100 raw_spin_unlock_irq(&nna->lock);
5101 }
5102
5103 kfree_rcu(pwq, rcu);
5104
5105 /*
5106 * If we're the last pwq going away, @wq is already dead and no one
5107 * is gonna access it anymore. Schedule RCU free.
5108 */
5109 if (is_last) {
5110 wq_unregister_lockdep(wq);
5111 call_rcu(&wq->rcu, rcu_free_wq);
5112 }
5113}
5114
5115/* initialize newly allocated @pwq which is associated with @wq and @pool */
5116static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
5117 struct worker_pool *pool)
5118{
5119 BUG_ON((unsigned long)pwq & ~WORK_STRUCT_PWQ_MASK);
5120
5121 memset(pwq, 0, sizeof(*pwq));
5122
5123 pwq->pool = pool;
5124 pwq->wq = wq;
5125 pwq->flush_color = -1;
5126 pwq->refcnt = 1;
5127 INIT_LIST_HEAD(&pwq->inactive_works);
5128 INIT_LIST_HEAD(&pwq->pending_node);
5129 INIT_LIST_HEAD(&pwq->pwqs_node);
5130 INIT_LIST_HEAD(&pwq->mayday_node);
5131 kthread_init_work(&pwq->release_work, pwq_release_workfn);
5132}
5133
5134/* sync @pwq with the current state of its associated wq and link it */
5135static void link_pwq(struct pool_workqueue *pwq)
5136{
5137 struct workqueue_struct *wq = pwq->wq;
5138
5139 lockdep_assert_held(&wq->mutex);
5140
5141 /* may be called multiple times, ignore if already linked */
5142 if (!list_empty(&pwq->pwqs_node))
5143 return;
5144
5145 /* set the matching work_color */
5146 pwq->work_color = wq->work_color;
5147
5148 /* link in @pwq */
5149 list_add_tail_rcu(&pwq->pwqs_node, &wq->pwqs);
5150}
5151
5152/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
5153static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
5154 const struct workqueue_attrs *attrs)
5155{
5156 struct worker_pool *pool;
5157 struct pool_workqueue *pwq;
5158
5159 lockdep_assert_held(&wq_pool_mutex);
5160
5161 pool = get_unbound_pool(attrs);
5162 if (!pool)
5163 return NULL;
5164
5165 pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
5166 if (!pwq) {
5167 put_unbound_pool(pool);
5168 return NULL;
5169 }
5170
5171 init_pwq(pwq, wq, pool);
5172 return pwq;
5173}
5174
5175static void apply_wqattrs_lock(void)
5176{
5177 mutex_lock(&wq_pool_mutex);
5178}
5179
5180static void apply_wqattrs_unlock(void)
5181{
5182 mutex_unlock(&wq_pool_mutex);
5183}
5184
5185/**
5186 * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod
5187 * @attrs: the wq_attrs of the default pwq of the target workqueue
5188 * @cpu: the target CPU
5189 *
5190 * Calculate the cpumask a workqueue with @attrs should use on @pod.
5191 * The result is stored in @attrs->__pod_cpumask.
5192 *
5193 * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled
5194 * and @pod has online CPUs requested by @attrs, the returned cpumask is the
5195 * intersection of the possible CPUs of @pod and @attrs->cpumask.
5196 *
5197 * The caller is responsible for ensuring that the cpumask of @pod stays stable.
5198 */
5199static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu)
5200{
5201 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
5202 int pod = pt->cpu_pod[cpu];
5203
5204 /* calculate possible CPUs in @pod that @attrs wants */
5205 cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask);
5206 /* does @pod have any online CPUs @attrs wants? */
5207 if (!cpumask_intersects(attrs->__pod_cpumask, wq_online_cpumask)) {
5208 cpumask_copy(attrs->__pod_cpumask, attrs->cpumask);
5209 return;
5210 }
5211}
5212
5213/* install @pwq into @wq and return the old pwq, @cpu < 0 for dfl_pwq */
5214static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq,
5215 int cpu, struct pool_workqueue *pwq)
5216{
5217 struct pool_workqueue __rcu **slot = unbound_pwq_slot(wq, cpu);
5218 struct pool_workqueue *old_pwq;
5219
5220 lockdep_assert_held(&wq_pool_mutex);
5221 lockdep_assert_held(&wq->mutex);
5222
5223 /* link_pwq() can handle duplicate calls */
5224 link_pwq(pwq);
5225
5226 old_pwq = rcu_access_pointer(*slot);
5227 rcu_assign_pointer(*slot, pwq);
5228 return old_pwq;
5229}
5230
5231/* context to store the prepared attrs & pwqs before applying */
5232struct apply_wqattrs_ctx {
5233 struct workqueue_struct *wq; /* target workqueue */
5234 struct workqueue_attrs *attrs; /* attrs to apply */
5235 struct list_head list; /* queued for batching commit */
5236 struct pool_workqueue *dfl_pwq;
5237 struct pool_workqueue *pwq_tbl[];
5238};
5239
5240/* free the resources after success or abort */
5241static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
5242{
5243 if (ctx) {
5244 int cpu;
5245
5246 for_each_possible_cpu(cpu)
5247 put_pwq_unlocked(ctx->pwq_tbl[cpu]);
5248 put_pwq_unlocked(ctx->dfl_pwq);
5249
5250 free_workqueue_attrs(ctx->attrs);
5251
5252 kfree(ctx);
5253 }
5254}
5255
5256/* allocate the attrs and pwqs for later installation */
5257static struct apply_wqattrs_ctx *
5258apply_wqattrs_prepare(struct workqueue_struct *wq,
5259 const struct workqueue_attrs *attrs,
5260 const cpumask_var_t unbound_cpumask)
5261{
5262 struct apply_wqattrs_ctx *ctx;
5263 struct workqueue_attrs *new_attrs;
5264 int cpu;
5265
5266 lockdep_assert_held(&wq_pool_mutex);
5267
5268 if (WARN_ON(attrs->affn_scope < 0 ||
5269 attrs->affn_scope >= WQ_AFFN_NR_TYPES))
5270 return ERR_PTR(-EINVAL);
5271
5272 ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL);
5273
5274 new_attrs = alloc_workqueue_attrs();
5275 if (!ctx || !new_attrs)
5276 goto out_free;
5277
5278 /*
5279 * If something goes wrong during CPU up/down, we'll fall back to
5280 * the default pwq covering whole @attrs->cpumask. Always create
5281 * it even if we don't use it immediately.
5282 */
5283 copy_workqueue_attrs(new_attrs, attrs);
5284 wqattrs_actualize_cpumask(new_attrs, unbound_cpumask);
5285 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5286 ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
5287 if (!ctx->dfl_pwq)
5288 goto out_free;
5289
5290 for_each_possible_cpu(cpu) {
5291 if (new_attrs->ordered) {
5292 ctx->dfl_pwq->refcnt++;
5293 ctx->pwq_tbl[cpu] = ctx->dfl_pwq;
5294 } else {
5295 wq_calc_pod_cpumask(new_attrs, cpu);
5296 ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs);
5297 if (!ctx->pwq_tbl[cpu])
5298 goto out_free;
5299 }
5300 }
5301
5302 /* save the user configured attrs and sanitize it. */
5303 copy_workqueue_attrs(new_attrs, attrs);
5304 cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
5305 cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask);
5306 ctx->attrs = new_attrs;
5307
5308 /*
5309 * For initialized ordered workqueues, there should only be one pwq
5310 * (dfl_pwq). Set the plugged flag of ctx->dfl_pwq to suspend execution
5311 * of newly queued work items until execution of older work items in
5312 * the old pwq's have completed.
5313 */
5314 if ((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs))
5315 ctx->dfl_pwq->plugged = true;
5316
5317 ctx->wq = wq;
5318 return ctx;
5319
5320out_free:
5321 free_workqueue_attrs(new_attrs);
5322 apply_wqattrs_cleanup(ctx);
5323 return ERR_PTR(-ENOMEM);
5324}
5325
5326/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
5327static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
5328{
5329 int cpu;
5330
5331 /* all pwqs have been created successfully, let's install'em */
5332 mutex_lock(&ctx->wq->mutex);
5333
5334 copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
5335
5336 /* save the previous pwqs and install the new ones */
5337 for_each_possible_cpu(cpu)
5338 ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu,
5339 ctx->pwq_tbl[cpu]);
5340 ctx->dfl_pwq = install_unbound_pwq(ctx->wq, -1, ctx->dfl_pwq);
5341
5342 /* update node_nr_active->max */
5343 wq_update_node_max_active(ctx->wq, -1);
5344
5345 /* rescuer needs to respect wq cpumask changes */
5346 if (ctx->wq->rescuer)
5347 set_cpus_allowed_ptr(ctx->wq->rescuer->task,
5348 unbound_effective_cpumask(ctx->wq));
5349
5350 mutex_unlock(&ctx->wq->mutex);
5351}
5352
5353static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
5354 const struct workqueue_attrs *attrs)
5355{
5356 struct apply_wqattrs_ctx *ctx;
5357
5358 /* only unbound workqueues can change attributes */
5359 if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
5360 return -EINVAL;
5361
5362 ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask);
5363 if (IS_ERR(ctx))
5364 return PTR_ERR(ctx);
5365
5366 /* the ctx has been prepared successfully, let's commit it */
5367 apply_wqattrs_commit(ctx);
5368 apply_wqattrs_cleanup(ctx);
5369
5370 return 0;
5371}
5372
5373/**
5374 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
5375 * @wq: the target workqueue
5376 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
5377 *
5378 * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps
5379 * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that
5380 * work items are affine to the pod it was issued on. Older pwqs are released as
5381 * in-flight work items finish. Note that a work item which repeatedly requeues
5382 * itself back-to-back will stay on its current pwq.
5383 *
5384 * Performs GFP_KERNEL allocations.
5385 *
5386 * Return: 0 on success and -errno on failure.
5387 */
5388int apply_workqueue_attrs(struct workqueue_struct *wq,
5389 const struct workqueue_attrs *attrs)
5390{
5391 int ret;
5392
5393 mutex_lock(&wq_pool_mutex);
5394 ret = apply_workqueue_attrs_locked(wq, attrs);
5395 mutex_unlock(&wq_pool_mutex);
5396
5397 return ret;
5398}
5399
5400/**
5401 * unbound_wq_update_pwq - update a pwq slot for CPU hot[un]plug
5402 * @wq: the target workqueue
5403 * @cpu: the CPU to update the pwq slot for
5404 *
5405 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
5406 * %CPU_DOWN_FAILED. @cpu is in the same pod of the CPU being hot[un]plugged.
5407 *
5408 *
5409 * If pod affinity can't be adjusted due to memory allocation failure, it falls
5410 * back to @wq->dfl_pwq which may not be optimal but is always correct.
5411 *
5412 * Note that when the last allowed CPU of a pod goes offline for a workqueue
5413 * with a cpumask spanning multiple pods, the workers which were already
5414 * executing the work items for the workqueue will lose their CPU affinity and
5415 * may execute on any CPU. This is similar to how per-cpu workqueues behave on
5416 * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's
5417 * responsibility to flush the work item from CPU_DOWN_PREPARE.
5418 */
5419static void unbound_wq_update_pwq(struct workqueue_struct *wq, int cpu)
5420{
5421 struct pool_workqueue *old_pwq = NULL, *pwq;
5422 struct workqueue_attrs *target_attrs;
5423
5424 lockdep_assert_held(&wq_pool_mutex);
5425
5426 if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered)
5427 return;
5428
5429 /*
5430 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
5431 * Let's use a preallocated one. The following buf is protected by
5432 * CPU hotplug exclusion.
5433 */
5434 target_attrs = unbound_wq_update_pwq_attrs_buf;
5435
5436 copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
5437 wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask);
5438
5439 /* nothing to do if the target cpumask matches the current pwq */
5440 wq_calc_pod_cpumask(target_attrs, cpu);
5441 if (wqattrs_equal(target_attrs, unbound_pwq(wq, cpu)->pool->attrs))
5442 return;
5443
5444 /* create a new pwq */
5445 pwq = alloc_unbound_pwq(wq, target_attrs);
5446 if (!pwq) {
5447 pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n",
5448 wq->name);
5449 goto use_dfl_pwq;
5450 }
5451
5452 /* Install the new pwq. */
5453 mutex_lock(&wq->mutex);
5454 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5455 goto out_unlock;
5456
5457use_dfl_pwq:
5458 mutex_lock(&wq->mutex);
5459 pwq = unbound_pwq(wq, -1);
5460 raw_spin_lock_irq(&pwq->pool->lock);
5461 get_pwq(pwq);
5462 raw_spin_unlock_irq(&pwq->pool->lock);
5463 old_pwq = install_unbound_pwq(wq, cpu, pwq);
5464out_unlock:
5465 mutex_unlock(&wq->mutex);
5466 put_pwq_unlocked(old_pwq);
5467}
5468
5469static int alloc_and_link_pwqs(struct workqueue_struct *wq)
5470{
5471 bool highpri = wq->flags & WQ_HIGHPRI;
5472 int cpu, ret;
5473
5474 lockdep_assert_held(&wq_pool_mutex);
5475
5476 wq->cpu_pwq = alloc_percpu(struct pool_workqueue *);
5477 if (!wq->cpu_pwq)
5478 goto enomem;
5479
5480 if (!(wq->flags & WQ_UNBOUND)) {
5481 struct worker_pool __percpu *pools;
5482
5483 if (wq->flags & WQ_BH)
5484 pools = bh_worker_pools;
5485 else
5486 pools = cpu_worker_pools;
5487
5488 for_each_possible_cpu(cpu) {
5489 struct pool_workqueue **pwq_p;
5490 struct worker_pool *pool;
5491
5492 pool = &(per_cpu_ptr(pools, cpu)[highpri]);
5493 pwq_p = per_cpu_ptr(wq->cpu_pwq, cpu);
5494
5495 *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL,
5496 pool->node);
5497 if (!*pwq_p)
5498 goto enomem;
5499
5500 init_pwq(*pwq_p, wq, pool);
5501
5502 mutex_lock(&wq->mutex);
5503 link_pwq(*pwq_p);
5504 mutex_unlock(&wq->mutex);
5505 }
5506 return 0;
5507 }
5508
5509 if (wq->flags & __WQ_ORDERED) {
5510 struct pool_workqueue *dfl_pwq;
5511
5512 ret = apply_workqueue_attrs_locked(wq, ordered_wq_attrs[highpri]);
5513 /* there should only be single pwq for ordering guarantee */
5514 dfl_pwq = rcu_access_pointer(wq->dfl_pwq);
5515 WARN(!ret && (wq->pwqs.next != &dfl_pwq->pwqs_node ||
5516 wq->pwqs.prev != &dfl_pwq->pwqs_node),
5517 "ordering guarantee broken for workqueue %s\n", wq->name);
5518 } else {
5519 ret = apply_workqueue_attrs_locked(wq, unbound_std_wq_attrs[highpri]);
5520 }
5521
5522 return ret;
5523
5524enomem:
5525 if (wq->cpu_pwq) {
5526 for_each_possible_cpu(cpu) {
5527 struct pool_workqueue *pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
5528
5529 if (pwq)
5530 kmem_cache_free(pwq_cache, pwq);
5531 }
5532 free_percpu(wq->cpu_pwq);
5533 wq->cpu_pwq = NULL;
5534 }
5535 return -ENOMEM;
5536}
5537
5538static int wq_clamp_max_active(int max_active, unsigned int flags,
5539 const char *name)
5540{
5541 if (max_active < 1 || max_active > WQ_MAX_ACTIVE)
5542 pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
5543 max_active, name, 1, WQ_MAX_ACTIVE);
5544
5545 return clamp_val(max_active, 1, WQ_MAX_ACTIVE);
5546}
5547
5548/*
5549 * Workqueues which may be used during memory reclaim should have a rescuer
5550 * to guarantee forward progress.
5551 */
5552static int init_rescuer(struct workqueue_struct *wq)
5553{
5554 struct worker *rescuer;
5555 char id_buf[WORKER_ID_LEN];
5556 int ret;
5557
5558 lockdep_assert_held(&wq_pool_mutex);
5559
5560 if (!(wq->flags & WQ_MEM_RECLAIM))
5561 return 0;
5562
5563 rescuer = alloc_worker(NUMA_NO_NODE);
5564 if (!rescuer) {
5565 pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n",
5566 wq->name);
5567 return -ENOMEM;
5568 }
5569
5570 rescuer->rescue_wq = wq;
5571 format_worker_id(id_buf, sizeof(id_buf), rescuer, NULL);
5572
5573 rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", id_buf);
5574 if (IS_ERR(rescuer->task)) {
5575 ret = PTR_ERR(rescuer->task);
5576 pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe",
5577 wq->name, ERR_PTR(ret));
5578 kfree(rescuer);
5579 return ret;
5580 }
5581
5582 wq->rescuer = rescuer;
5583 if (wq->flags & WQ_UNBOUND)
5584 kthread_bind_mask(rescuer->task, unbound_effective_cpumask(wq));
5585 else
5586 kthread_bind_mask(rescuer->task, cpu_possible_mask);
5587 wake_up_process(rescuer->task);
5588
5589 return 0;
5590}
5591
5592/**
5593 * wq_adjust_max_active - update a wq's max_active to the current setting
5594 * @wq: target workqueue
5595 *
5596 * If @wq isn't freezing, set @wq->max_active to the saved_max_active and
5597 * activate inactive work items accordingly. If @wq is freezing, clear
5598 * @wq->max_active to zero.
5599 */
5600static void wq_adjust_max_active(struct workqueue_struct *wq)
5601{
5602 bool activated;
5603 int new_max, new_min;
5604
5605 lockdep_assert_held(&wq->mutex);
5606
5607 if ((wq->flags & WQ_FREEZABLE) && workqueue_freezing) {
5608 new_max = 0;
5609 new_min = 0;
5610 } else {
5611 new_max = wq->saved_max_active;
5612 new_min = wq->saved_min_active;
5613 }
5614
5615 if (wq->max_active == new_max && wq->min_active == new_min)
5616 return;
5617
5618 /*
5619 * Update @wq->max/min_active and then kick inactive work items if more
5620 * active work items are allowed. This doesn't break work item ordering
5621 * because new work items are always queued behind existing inactive
5622 * work items if there are any.
5623 */
5624 WRITE_ONCE(wq->max_active, new_max);
5625 WRITE_ONCE(wq->min_active, new_min);
5626
5627 if (wq->flags & WQ_UNBOUND)
5628 wq_update_node_max_active(wq, -1);
5629
5630 if (new_max == 0)
5631 return;
5632
5633 /*
5634 * Round-robin through pwq's activating the first inactive work item
5635 * until max_active is filled.
5636 */
5637 do {
5638 struct pool_workqueue *pwq;
5639
5640 activated = false;
5641 for_each_pwq(pwq, wq) {
5642 unsigned long irq_flags;
5643
5644 /* can be called during early boot w/ irq disabled */
5645 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
5646 if (pwq_activate_first_inactive(pwq, true)) {
5647 activated = true;
5648 kick_pool(pwq->pool);
5649 }
5650 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
5651 }
5652 } while (activated);
5653}
5654
5655__printf(1, 0)
5656static struct workqueue_struct *__alloc_workqueue(const char *fmt,
5657 unsigned int flags,
5658 int max_active, va_list args)
5659{
5660 struct workqueue_struct *wq;
5661 size_t wq_size;
5662 int name_len;
5663
5664 if (flags & WQ_BH) {
5665 if (WARN_ON_ONCE(flags & ~__WQ_BH_ALLOWS))
5666 return NULL;
5667 if (WARN_ON_ONCE(max_active))
5668 return NULL;
5669 }
5670
5671 /* see the comment above the definition of WQ_POWER_EFFICIENT */
5672 if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
5673 flags |= WQ_UNBOUND;
5674
5675 /* allocate wq and format name */
5676 if (flags & WQ_UNBOUND)
5677 wq_size = struct_size(wq, node_nr_active, nr_node_ids + 1);
5678 else
5679 wq_size = sizeof(*wq);
5680
5681 wq = kzalloc(wq_size, GFP_KERNEL);
5682 if (!wq)
5683 return NULL;
5684
5685 if (flags & WQ_UNBOUND) {
5686 wq->unbound_attrs = alloc_workqueue_attrs();
5687 if (!wq->unbound_attrs)
5688 goto err_free_wq;
5689 }
5690
5691 name_len = vsnprintf(wq->name, sizeof(wq->name), fmt, args);
5692
5693 if (name_len >= WQ_NAME_LEN)
5694 pr_warn_once("workqueue: name exceeds WQ_NAME_LEN. Truncating to: %s\n",
5695 wq->name);
5696
5697 if (flags & WQ_BH) {
5698 /*
5699 * BH workqueues always share a single execution context per CPU
5700 * and don't impose any max_active limit.
5701 */
5702 max_active = INT_MAX;
5703 } else {
5704 max_active = max_active ?: WQ_DFL_ACTIVE;
5705 max_active = wq_clamp_max_active(max_active, flags, wq->name);
5706 }
5707
5708 /* init wq */
5709 wq->flags = flags;
5710 wq->max_active = max_active;
5711 wq->min_active = min(max_active, WQ_DFL_MIN_ACTIVE);
5712 wq->saved_max_active = wq->max_active;
5713 wq->saved_min_active = wq->min_active;
5714 mutex_init(&wq->mutex);
5715 atomic_set(&wq->nr_pwqs_to_flush, 0);
5716 INIT_LIST_HEAD(&wq->pwqs);
5717 INIT_LIST_HEAD(&wq->flusher_queue);
5718 INIT_LIST_HEAD(&wq->flusher_overflow);
5719 INIT_LIST_HEAD(&wq->maydays);
5720
5721 INIT_LIST_HEAD(&wq->list);
5722
5723 if (flags & WQ_UNBOUND) {
5724 if (alloc_node_nr_active(wq->node_nr_active) < 0)
5725 goto err_free_wq;
5726 }
5727
5728 /*
5729 * wq_pool_mutex protects the workqueues list, allocations of PWQs,
5730 * and the global freeze state.
5731 */
5732 apply_wqattrs_lock();
5733
5734 if (alloc_and_link_pwqs(wq) < 0)
5735 goto err_unlock_free_node_nr_active;
5736
5737 mutex_lock(&wq->mutex);
5738 wq_adjust_max_active(wq);
5739 mutex_unlock(&wq->mutex);
5740
5741 list_add_tail_rcu(&wq->list, &workqueues);
5742
5743 if (wq_online && init_rescuer(wq) < 0)
5744 goto err_unlock_destroy;
5745
5746 apply_wqattrs_unlock();
5747
5748 if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
5749 goto err_destroy;
5750
5751 return wq;
5752
5753err_unlock_free_node_nr_active:
5754 apply_wqattrs_unlock();
5755 /*
5756 * Failed alloc_and_link_pwqs() may leave pending pwq->release_work,
5757 * flushing the pwq_release_worker ensures that the pwq_release_workfn()
5758 * completes before calling kfree(wq).
5759 */
5760 if (wq->flags & WQ_UNBOUND) {
5761 kthread_flush_worker(pwq_release_worker);
5762 free_node_nr_active(wq->node_nr_active);
5763 }
5764err_free_wq:
5765 free_workqueue_attrs(wq->unbound_attrs);
5766 kfree(wq);
5767 return NULL;
5768err_unlock_destroy:
5769 apply_wqattrs_unlock();
5770err_destroy:
5771 destroy_workqueue(wq);
5772 return NULL;
5773}
5774
5775__printf(1, 4)
5776struct workqueue_struct *alloc_workqueue(const char *fmt,
5777 unsigned int flags,
5778 int max_active, ...)
5779{
5780 struct workqueue_struct *wq;
5781 va_list args;
5782
5783 va_start(args, max_active);
5784 wq = __alloc_workqueue(fmt, flags, max_active, args);
5785 va_end(args);
5786 if (!wq)
5787 return NULL;
5788
5789 wq_init_lockdep(wq);
5790
5791 return wq;
5792}
5793EXPORT_SYMBOL_GPL(alloc_workqueue);
5794
5795#ifdef CONFIG_LOCKDEP
5796__printf(1, 5)
5797struct workqueue_struct *
5798alloc_workqueue_lockdep_map(const char *fmt, unsigned int flags,
5799 int max_active, struct lockdep_map *lockdep_map, ...)
5800{
5801 struct workqueue_struct *wq;
5802 va_list args;
5803
5804 va_start(args, lockdep_map);
5805 wq = __alloc_workqueue(fmt, flags, max_active, args);
5806 va_end(args);
5807 if (!wq)
5808 return NULL;
5809
5810 wq->lockdep_map = lockdep_map;
5811
5812 return wq;
5813}
5814EXPORT_SYMBOL_GPL(alloc_workqueue_lockdep_map);
5815#endif
5816
5817static bool pwq_busy(struct pool_workqueue *pwq)
5818{
5819 int i;
5820
5821 for (i = 0; i < WORK_NR_COLORS; i++)
5822 if (pwq->nr_in_flight[i])
5823 return true;
5824
5825 if ((pwq != rcu_access_pointer(pwq->wq->dfl_pwq)) && (pwq->refcnt > 1))
5826 return true;
5827 if (!pwq_is_empty(pwq))
5828 return true;
5829
5830 return false;
5831}
5832
5833/**
5834 * destroy_workqueue - safely terminate a workqueue
5835 * @wq: target workqueue
5836 *
5837 * Safely destroy a workqueue. All work currently pending will be done first.
5838 */
5839void destroy_workqueue(struct workqueue_struct *wq)
5840{
5841 struct pool_workqueue *pwq;
5842 int cpu;
5843
5844 /*
5845 * Remove it from sysfs first so that sanity check failure doesn't
5846 * lead to sysfs name conflicts.
5847 */
5848 workqueue_sysfs_unregister(wq);
5849
5850 /* mark the workqueue destruction is in progress */
5851 mutex_lock(&wq->mutex);
5852 wq->flags |= __WQ_DESTROYING;
5853 mutex_unlock(&wq->mutex);
5854
5855 /* drain it before proceeding with destruction */
5856 drain_workqueue(wq);
5857
5858 /* kill rescuer, if sanity checks fail, leave it w/o rescuer */
5859 if (wq->rescuer) {
5860 struct worker *rescuer = wq->rescuer;
5861
5862 /* this prevents new queueing */
5863 raw_spin_lock_irq(&wq_mayday_lock);
5864 wq->rescuer = NULL;
5865 raw_spin_unlock_irq(&wq_mayday_lock);
5866
5867 /* rescuer will empty maydays list before exiting */
5868 kthread_stop(rescuer->task);
5869 kfree(rescuer);
5870 }
5871
5872 /*
5873 * Sanity checks - grab all the locks so that we wait for all
5874 * in-flight operations which may do put_pwq().
5875 */
5876 mutex_lock(&wq_pool_mutex);
5877 mutex_lock(&wq->mutex);
5878 for_each_pwq(pwq, wq) {
5879 raw_spin_lock_irq(&pwq->pool->lock);
5880 if (WARN_ON(pwq_busy(pwq))) {
5881 pr_warn("%s: %s has the following busy pwq\n",
5882 __func__, wq->name);
5883 show_pwq(pwq);
5884 raw_spin_unlock_irq(&pwq->pool->lock);
5885 mutex_unlock(&wq->mutex);
5886 mutex_unlock(&wq_pool_mutex);
5887 show_one_workqueue(wq);
5888 return;
5889 }
5890 raw_spin_unlock_irq(&pwq->pool->lock);
5891 }
5892 mutex_unlock(&wq->mutex);
5893
5894 /*
5895 * wq list is used to freeze wq, remove from list after
5896 * flushing is complete in case freeze races us.
5897 */
5898 list_del_rcu(&wq->list);
5899 mutex_unlock(&wq_pool_mutex);
5900
5901 /*
5902 * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq
5903 * to put the base refs. @wq will be auto-destroyed from the last
5904 * pwq_put. RCU read lock prevents @wq from going away from under us.
5905 */
5906 rcu_read_lock();
5907
5908 for_each_possible_cpu(cpu) {
5909 put_pwq_unlocked(unbound_pwq(wq, cpu));
5910 RCU_INIT_POINTER(*unbound_pwq_slot(wq, cpu), NULL);
5911 }
5912
5913 put_pwq_unlocked(unbound_pwq(wq, -1));
5914 RCU_INIT_POINTER(*unbound_pwq_slot(wq, -1), NULL);
5915
5916 rcu_read_unlock();
5917}
5918EXPORT_SYMBOL_GPL(destroy_workqueue);
5919
5920/**
5921 * workqueue_set_max_active - adjust max_active of a workqueue
5922 * @wq: target workqueue
5923 * @max_active: new max_active value.
5924 *
5925 * Set max_active of @wq to @max_active. See the alloc_workqueue() function
5926 * comment.
5927 *
5928 * CONTEXT:
5929 * Don't call from IRQ context.
5930 */
5931void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
5932{
5933 /* max_active doesn't mean anything for BH workqueues */
5934 if (WARN_ON(wq->flags & WQ_BH))
5935 return;
5936 /* disallow meddling with max_active for ordered workqueues */
5937 if (WARN_ON(wq->flags & __WQ_ORDERED))
5938 return;
5939
5940 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
5941
5942 mutex_lock(&wq->mutex);
5943
5944 wq->saved_max_active = max_active;
5945 if (wq->flags & WQ_UNBOUND)
5946 wq->saved_min_active = min(wq->saved_min_active, max_active);
5947
5948 wq_adjust_max_active(wq);
5949
5950 mutex_unlock(&wq->mutex);
5951}
5952EXPORT_SYMBOL_GPL(workqueue_set_max_active);
5953
5954/**
5955 * workqueue_set_min_active - adjust min_active of an unbound workqueue
5956 * @wq: target unbound workqueue
5957 * @min_active: new min_active value
5958 *
5959 * Set min_active of an unbound workqueue. Unlike other types of workqueues, an
5960 * unbound workqueue is not guaranteed to be able to process max_active
5961 * interdependent work items. Instead, an unbound workqueue is guaranteed to be
5962 * able to process min_active number of interdependent work items which is
5963 * %WQ_DFL_MIN_ACTIVE by default.
5964 *
5965 * Use this function to adjust the min_active value between 0 and the current
5966 * max_active.
5967 */
5968void workqueue_set_min_active(struct workqueue_struct *wq, int min_active)
5969{
5970 /* min_active is only meaningful for non-ordered unbound workqueues */
5971 if (WARN_ON((wq->flags & (WQ_BH | WQ_UNBOUND | __WQ_ORDERED)) !=
5972 WQ_UNBOUND))
5973 return;
5974
5975 mutex_lock(&wq->mutex);
5976 wq->saved_min_active = clamp(min_active, 0, wq->saved_max_active);
5977 wq_adjust_max_active(wq);
5978 mutex_unlock(&wq->mutex);
5979}
5980
5981/**
5982 * current_work - retrieve %current task's work struct
5983 *
5984 * Determine if %current task is a workqueue worker and what it's working on.
5985 * Useful to find out the context that the %current task is running in.
5986 *
5987 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
5988 */
5989struct work_struct *current_work(void)
5990{
5991 struct worker *worker = current_wq_worker();
5992
5993 return worker ? worker->current_work : NULL;
5994}
5995EXPORT_SYMBOL(current_work);
5996
5997/**
5998 * current_is_workqueue_rescuer - is %current workqueue rescuer?
5999 *
6000 * Determine whether %current is a workqueue rescuer. Can be used from
6001 * work functions to determine whether it's being run off the rescuer task.
6002 *
6003 * Return: %true if %current is a workqueue rescuer. %false otherwise.
6004 */
6005bool current_is_workqueue_rescuer(void)
6006{
6007 struct worker *worker = current_wq_worker();
6008
6009 return worker && worker->rescue_wq;
6010}
6011
6012/**
6013 * workqueue_congested - test whether a workqueue is congested
6014 * @cpu: CPU in question
6015 * @wq: target workqueue
6016 *
6017 * Test whether @wq's cpu workqueue for @cpu is congested. There is
6018 * no synchronization around this function and the test result is
6019 * unreliable and only useful as advisory hints or for debugging.
6020 *
6021 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
6022 *
6023 * With the exception of ordered workqueues, all workqueues have per-cpu
6024 * pool_workqueues, each with its own congested state. A workqueue being
6025 * congested on one CPU doesn't mean that the workqueue is contested on any
6026 * other CPUs.
6027 *
6028 * Return:
6029 * %true if congested, %false otherwise.
6030 */
6031bool workqueue_congested(int cpu, struct workqueue_struct *wq)
6032{
6033 struct pool_workqueue *pwq;
6034 bool ret;
6035
6036 rcu_read_lock();
6037 preempt_disable();
6038
6039 if (cpu == WORK_CPU_UNBOUND)
6040 cpu = smp_processor_id();
6041
6042 pwq = *per_cpu_ptr(wq->cpu_pwq, cpu);
6043 ret = !list_empty(&pwq->inactive_works);
6044
6045 preempt_enable();
6046 rcu_read_unlock();
6047
6048 return ret;
6049}
6050EXPORT_SYMBOL_GPL(workqueue_congested);
6051
6052/**
6053 * work_busy - test whether a work is currently pending or running
6054 * @work: the work to be tested
6055 *
6056 * Test whether @work is currently pending or running. There is no
6057 * synchronization around this function and the test result is
6058 * unreliable and only useful as advisory hints or for debugging.
6059 *
6060 * Return:
6061 * OR'd bitmask of WORK_BUSY_* bits.
6062 */
6063unsigned int work_busy(struct work_struct *work)
6064{
6065 struct worker_pool *pool;
6066 unsigned long irq_flags;
6067 unsigned int ret = 0;
6068
6069 if (work_pending(work))
6070 ret |= WORK_BUSY_PENDING;
6071
6072 rcu_read_lock();
6073 pool = get_work_pool(work);
6074 if (pool) {
6075 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6076 if (find_worker_executing_work(pool, work))
6077 ret |= WORK_BUSY_RUNNING;
6078 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6079 }
6080 rcu_read_unlock();
6081
6082 return ret;
6083}
6084EXPORT_SYMBOL_GPL(work_busy);
6085
6086/**
6087 * set_worker_desc - set description for the current work item
6088 * @fmt: printf-style format string
6089 * @...: arguments for the format string
6090 *
6091 * This function can be called by a running work function to describe what
6092 * the work item is about. If the worker task gets dumped, this
6093 * information will be printed out together to help debugging. The
6094 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
6095 */
6096void set_worker_desc(const char *fmt, ...)
6097{
6098 struct worker *worker = current_wq_worker();
6099 va_list args;
6100
6101 if (worker) {
6102 va_start(args, fmt);
6103 vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
6104 va_end(args);
6105 }
6106}
6107EXPORT_SYMBOL_GPL(set_worker_desc);
6108
6109/**
6110 * print_worker_info - print out worker information and description
6111 * @log_lvl: the log level to use when printing
6112 * @task: target task
6113 *
6114 * If @task is a worker and currently executing a work item, print out the
6115 * name of the workqueue being serviced and worker description set with
6116 * set_worker_desc() by the currently executing work item.
6117 *
6118 * This function can be safely called on any task as long as the
6119 * task_struct itself is accessible. While safe, this function isn't
6120 * synchronized and may print out mixups or garbages of limited length.
6121 */
6122void print_worker_info(const char *log_lvl, struct task_struct *task)
6123{
6124 work_func_t *fn = NULL;
6125 char name[WQ_NAME_LEN] = { };
6126 char desc[WORKER_DESC_LEN] = { };
6127 struct pool_workqueue *pwq = NULL;
6128 struct workqueue_struct *wq = NULL;
6129 struct worker *worker;
6130
6131 if (!(task->flags & PF_WQ_WORKER))
6132 return;
6133
6134 /*
6135 * This function is called without any synchronization and @task
6136 * could be in any state. Be careful with dereferences.
6137 */
6138 worker = kthread_probe_data(task);
6139
6140 /*
6141 * Carefully copy the associated workqueue's workfn, name and desc.
6142 * Keep the original last '\0' in case the original is garbage.
6143 */
6144 copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn));
6145 copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq));
6146 copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq));
6147 copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1);
6148 copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1);
6149
6150 if (fn || name[0] || desc[0]) {
6151 printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
6152 if (strcmp(name, desc))
6153 pr_cont(" (%s)", desc);
6154 pr_cont("\n");
6155 }
6156}
6157
6158static void pr_cont_pool_info(struct worker_pool *pool)
6159{
6160 pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
6161 if (pool->node != NUMA_NO_NODE)
6162 pr_cont(" node=%d", pool->node);
6163 pr_cont(" flags=0x%x", pool->flags);
6164 if (pool->flags & POOL_BH)
6165 pr_cont(" bh%s",
6166 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6167 else
6168 pr_cont(" nice=%d", pool->attrs->nice);
6169}
6170
6171static void pr_cont_worker_id(struct worker *worker)
6172{
6173 struct worker_pool *pool = worker->pool;
6174
6175 if (pool->flags & WQ_BH)
6176 pr_cont("bh%s",
6177 pool->attrs->nice == HIGHPRI_NICE_LEVEL ? "-hi" : "");
6178 else
6179 pr_cont("%d%s", task_pid_nr(worker->task),
6180 worker->rescue_wq ? "(RESCUER)" : "");
6181}
6182
6183struct pr_cont_work_struct {
6184 bool comma;
6185 work_func_t func;
6186 long ctr;
6187};
6188
6189static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp)
6190{
6191 if (!pcwsp->ctr)
6192 goto out_record;
6193 if (func == pcwsp->func) {
6194 pcwsp->ctr++;
6195 return;
6196 }
6197 if (pcwsp->ctr == 1)
6198 pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func);
6199 else
6200 pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func);
6201 pcwsp->ctr = 0;
6202out_record:
6203 if ((long)func == -1L)
6204 return;
6205 pcwsp->comma = comma;
6206 pcwsp->func = func;
6207 pcwsp->ctr = 1;
6208}
6209
6210static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp)
6211{
6212 if (work->func == wq_barrier_func) {
6213 struct wq_barrier *barr;
6214
6215 barr = container_of(work, struct wq_barrier, work);
6216
6217 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6218 pr_cont("%s BAR(%d)", comma ? "," : "",
6219 task_pid_nr(barr->task));
6220 } else {
6221 if (!comma)
6222 pr_cont_work_flush(comma, (work_func_t)-1, pcwsp);
6223 pr_cont_work_flush(comma, work->func, pcwsp);
6224 }
6225}
6226
6227static void show_pwq(struct pool_workqueue *pwq)
6228{
6229 struct pr_cont_work_struct pcws = { .ctr = 0, };
6230 struct worker_pool *pool = pwq->pool;
6231 struct work_struct *work;
6232 struct worker *worker;
6233 bool has_in_flight = false, has_pending = false;
6234 int bkt;
6235
6236 pr_info(" pwq %d:", pool->id);
6237 pr_cont_pool_info(pool);
6238
6239 pr_cont(" active=%d refcnt=%d%s\n",
6240 pwq->nr_active, pwq->refcnt,
6241 !list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
6242
6243 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6244 if (worker->current_pwq == pwq) {
6245 has_in_flight = true;
6246 break;
6247 }
6248 }
6249 if (has_in_flight) {
6250 bool comma = false;
6251
6252 pr_info(" in-flight:");
6253 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
6254 if (worker->current_pwq != pwq)
6255 continue;
6256
6257 pr_cont(" %s", comma ? "," : "");
6258 pr_cont_worker_id(worker);
6259 pr_cont(":%ps", worker->current_func);
6260 list_for_each_entry(work, &worker->scheduled, entry)
6261 pr_cont_work(false, work, &pcws);
6262 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6263 comma = true;
6264 }
6265 pr_cont("\n");
6266 }
6267
6268 list_for_each_entry(work, &pool->worklist, entry) {
6269 if (get_work_pwq(work) == pwq) {
6270 has_pending = true;
6271 break;
6272 }
6273 }
6274 if (has_pending) {
6275 bool comma = false;
6276
6277 pr_info(" pending:");
6278 list_for_each_entry(work, &pool->worklist, entry) {
6279 if (get_work_pwq(work) != pwq)
6280 continue;
6281
6282 pr_cont_work(comma, work, &pcws);
6283 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6284 }
6285 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6286 pr_cont("\n");
6287 }
6288
6289 if (!list_empty(&pwq->inactive_works)) {
6290 bool comma = false;
6291
6292 pr_info(" inactive:");
6293 list_for_each_entry(work, &pwq->inactive_works, entry) {
6294 pr_cont_work(comma, work, &pcws);
6295 comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
6296 }
6297 pr_cont_work_flush(comma, (work_func_t)-1L, &pcws);
6298 pr_cont("\n");
6299 }
6300}
6301
6302/**
6303 * show_one_workqueue - dump state of specified workqueue
6304 * @wq: workqueue whose state will be printed
6305 */
6306void show_one_workqueue(struct workqueue_struct *wq)
6307{
6308 struct pool_workqueue *pwq;
6309 bool idle = true;
6310 unsigned long irq_flags;
6311
6312 for_each_pwq(pwq, wq) {
6313 if (!pwq_is_empty(pwq)) {
6314 idle = false;
6315 break;
6316 }
6317 }
6318 if (idle) /* Nothing to print for idle workqueue */
6319 return;
6320
6321 pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
6322
6323 for_each_pwq(pwq, wq) {
6324 raw_spin_lock_irqsave(&pwq->pool->lock, irq_flags);
6325 if (!pwq_is_empty(pwq)) {
6326 /*
6327 * Defer printing to avoid deadlocks in console
6328 * drivers that queue work while holding locks
6329 * also taken in their write paths.
6330 */
6331 printk_deferred_enter();
6332 show_pwq(pwq);
6333 printk_deferred_exit();
6334 }
6335 raw_spin_unlock_irqrestore(&pwq->pool->lock, irq_flags);
6336 /*
6337 * We could be printing a lot from atomic context, e.g.
6338 * sysrq-t -> show_all_workqueues(). Avoid triggering
6339 * hard lockup.
6340 */
6341 touch_nmi_watchdog();
6342 }
6343
6344}
6345
6346/**
6347 * show_one_worker_pool - dump state of specified worker pool
6348 * @pool: worker pool whose state will be printed
6349 */
6350static void show_one_worker_pool(struct worker_pool *pool)
6351{
6352 struct worker *worker;
6353 bool first = true;
6354 unsigned long irq_flags;
6355 unsigned long hung = 0;
6356
6357 raw_spin_lock_irqsave(&pool->lock, irq_flags);
6358 if (pool->nr_workers == pool->nr_idle)
6359 goto next_pool;
6360
6361 /* How long the first pending work is waiting for a worker. */
6362 if (!list_empty(&pool->worklist))
6363 hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000;
6364
6365 /*
6366 * Defer printing to avoid deadlocks in console drivers that
6367 * queue work while holding locks also taken in their write
6368 * paths.
6369 */
6370 printk_deferred_enter();
6371 pr_info("pool %d:", pool->id);
6372 pr_cont_pool_info(pool);
6373 pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers);
6374 if (pool->manager)
6375 pr_cont(" manager: %d",
6376 task_pid_nr(pool->manager->task));
6377 list_for_each_entry(worker, &pool->idle_list, entry) {
6378 pr_cont(" %s", first ? "idle: " : "");
6379 pr_cont_worker_id(worker);
6380 first = false;
6381 }
6382 pr_cont("\n");
6383 printk_deferred_exit();
6384next_pool:
6385 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
6386 /*
6387 * We could be printing a lot from atomic context, e.g.
6388 * sysrq-t -> show_all_workqueues(). Avoid triggering
6389 * hard lockup.
6390 */
6391 touch_nmi_watchdog();
6392
6393}
6394
6395/**
6396 * show_all_workqueues - dump workqueue state
6397 *
6398 * Called from a sysrq handler and prints out all busy workqueues and pools.
6399 */
6400void show_all_workqueues(void)
6401{
6402 struct workqueue_struct *wq;
6403 struct worker_pool *pool;
6404 int pi;
6405
6406 rcu_read_lock();
6407
6408 pr_info("Showing busy workqueues and worker pools:\n");
6409
6410 list_for_each_entry_rcu(wq, &workqueues, list)
6411 show_one_workqueue(wq);
6412
6413 for_each_pool(pool, pi)
6414 show_one_worker_pool(pool);
6415
6416 rcu_read_unlock();
6417}
6418
6419/**
6420 * show_freezable_workqueues - dump freezable workqueue state
6421 *
6422 * Called from try_to_freeze_tasks() and prints out all freezable workqueues
6423 * still busy.
6424 */
6425void show_freezable_workqueues(void)
6426{
6427 struct workqueue_struct *wq;
6428
6429 rcu_read_lock();
6430
6431 pr_info("Showing freezable workqueues that are still busy:\n");
6432
6433 list_for_each_entry_rcu(wq, &workqueues, list) {
6434 if (!(wq->flags & WQ_FREEZABLE))
6435 continue;
6436 show_one_workqueue(wq);
6437 }
6438
6439 rcu_read_unlock();
6440}
6441
6442/* used to show worker information through /proc/PID/{comm,stat,status} */
6443void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
6444{
6445 /* stabilize PF_WQ_WORKER and worker pool association */
6446 mutex_lock(&wq_pool_attach_mutex);
6447
6448 if (task->flags & PF_WQ_WORKER) {
6449 struct worker *worker = kthread_data(task);
6450 struct worker_pool *pool = worker->pool;
6451 int off;
6452
6453 off = format_worker_id(buf, size, worker, pool);
6454
6455 if (pool) {
6456 raw_spin_lock_irq(&pool->lock);
6457 /*
6458 * ->desc tracks information (wq name or
6459 * set_worker_desc()) for the latest execution. If
6460 * current, prepend '+', otherwise '-'.
6461 */
6462 if (worker->desc[0] != '\0') {
6463 if (worker->current_work)
6464 scnprintf(buf + off, size - off, "+%s",
6465 worker->desc);
6466 else
6467 scnprintf(buf + off, size - off, "-%s",
6468 worker->desc);
6469 }
6470 raw_spin_unlock_irq(&pool->lock);
6471 }
6472 } else {
6473 strscpy(buf, task->comm, size);
6474 }
6475
6476 mutex_unlock(&wq_pool_attach_mutex);
6477}
6478
6479#ifdef CONFIG_SMP
6480
6481/*
6482 * CPU hotplug.
6483 *
6484 * There are two challenges in supporting CPU hotplug. Firstly, there
6485 * are a lot of assumptions on strong associations among work, pwq and
6486 * pool which make migrating pending and scheduled works very
6487 * difficult to implement without impacting hot paths. Secondly,
6488 * worker pools serve mix of short, long and very long running works making
6489 * blocked draining impractical.
6490 *
6491 * This is solved by allowing the pools to be disassociated from the CPU
6492 * running as an unbound one and allowing it to be reattached later if the
6493 * cpu comes back online.
6494 */
6495
6496static void unbind_workers(int cpu)
6497{
6498 struct worker_pool *pool;
6499 struct worker *worker;
6500
6501 for_each_cpu_worker_pool(pool, cpu) {
6502 mutex_lock(&wq_pool_attach_mutex);
6503 raw_spin_lock_irq(&pool->lock);
6504
6505 /*
6506 * We've blocked all attach/detach operations. Make all workers
6507 * unbound and set DISASSOCIATED. Before this, all workers
6508 * must be on the cpu. After this, they may become diasporas.
6509 * And the preemption disabled section in their sched callbacks
6510 * are guaranteed to see WORKER_UNBOUND since the code here
6511 * is on the same cpu.
6512 */
6513 for_each_pool_worker(worker, pool)
6514 worker->flags |= WORKER_UNBOUND;
6515
6516 pool->flags |= POOL_DISASSOCIATED;
6517
6518 /*
6519 * The handling of nr_running in sched callbacks are disabled
6520 * now. Zap nr_running. After this, nr_running stays zero and
6521 * need_more_worker() and keep_working() are always true as
6522 * long as the worklist is not empty. This pool now behaves as
6523 * an unbound (in terms of concurrency management) pool which
6524 * are served by workers tied to the pool.
6525 */
6526 pool->nr_running = 0;
6527
6528 /*
6529 * With concurrency management just turned off, a busy
6530 * worker blocking could lead to lengthy stalls. Kick off
6531 * unbound chain execution of currently pending work items.
6532 */
6533 kick_pool(pool);
6534
6535 raw_spin_unlock_irq(&pool->lock);
6536
6537 for_each_pool_worker(worker, pool)
6538 unbind_worker(worker);
6539
6540 mutex_unlock(&wq_pool_attach_mutex);
6541 }
6542}
6543
6544/**
6545 * rebind_workers - rebind all workers of a pool to the associated CPU
6546 * @pool: pool of interest
6547 *
6548 * @pool->cpu is coming online. Rebind all workers to the CPU.
6549 */
6550static void rebind_workers(struct worker_pool *pool)
6551{
6552 struct worker *worker;
6553
6554 lockdep_assert_held(&wq_pool_attach_mutex);
6555
6556 /*
6557 * Restore CPU affinity of all workers. As all idle workers should
6558 * be on the run-queue of the associated CPU before any local
6559 * wake-ups for concurrency management happen, restore CPU affinity
6560 * of all workers first and then clear UNBOUND. As we're called
6561 * from CPU_ONLINE, the following shouldn't fail.
6562 */
6563 for_each_pool_worker(worker, pool) {
6564 kthread_set_per_cpu(worker->task, pool->cpu);
6565 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
6566 pool_allowed_cpus(pool)) < 0);
6567 }
6568
6569 raw_spin_lock_irq(&pool->lock);
6570
6571 pool->flags &= ~POOL_DISASSOCIATED;
6572
6573 for_each_pool_worker(worker, pool) {
6574 unsigned int worker_flags = worker->flags;
6575
6576 /*
6577 * We want to clear UNBOUND but can't directly call
6578 * worker_clr_flags() or adjust nr_running. Atomically
6579 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
6580 * @worker will clear REBOUND using worker_clr_flags() when
6581 * it initiates the next execution cycle thus restoring
6582 * concurrency management. Note that when or whether
6583 * @worker clears REBOUND doesn't affect correctness.
6584 *
6585 * WRITE_ONCE() is necessary because @worker->flags may be
6586 * tested without holding any lock in
6587 * wq_worker_running(). Without it, NOT_RUNNING test may
6588 * fail incorrectly leading to premature concurrency
6589 * management operations.
6590 */
6591 WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
6592 worker_flags |= WORKER_REBOUND;
6593 worker_flags &= ~WORKER_UNBOUND;
6594 WRITE_ONCE(worker->flags, worker_flags);
6595 }
6596
6597 raw_spin_unlock_irq(&pool->lock);
6598}
6599
6600/**
6601 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
6602 * @pool: unbound pool of interest
6603 * @cpu: the CPU which is coming up
6604 *
6605 * An unbound pool may end up with a cpumask which doesn't have any online
6606 * CPUs. When a worker of such pool get scheduled, the scheduler resets
6607 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
6608 * online CPU before, cpus_allowed of all its workers should be restored.
6609 */
6610static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
6611{
6612 static cpumask_t cpumask;
6613 struct worker *worker;
6614
6615 lockdep_assert_held(&wq_pool_attach_mutex);
6616
6617 /* is @cpu allowed for @pool? */
6618 if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
6619 return;
6620
6621 cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
6622
6623 /* as we're called from CPU_ONLINE, the following shouldn't fail */
6624 for_each_pool_worker(worker, pool)
6625 WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
6626}
6627
6628int workqueue_prepare_cpu(unsigned int cpu)
6629{
6630 struct worker_pool *pool;
6631
6632 for_each_cpu_worker_pool(pool, cpu) {
6633 if (pool->nr_workers)
6634 continue;
6635 if (!create_worker(pool))
6636 return -ENOMEM;
6637 }
6638 return 0;
6639}
6640
6641int workqueue_online_cpu(unsigned int cpu)
6642{
6643 struct worker_pool *pool;
6644 struct workqueue_struct *wq;
6645 int pi;
6646
6647 mutex_lock(&wq_pool_mutex);
6648
6649 cpumask_set_cpu(cpu, wq_online_cpumask);
6650
6651 for_each_pool(pool, pi) {
6652 /* BH pools aren't affected by hotplug */
6653 if (pool->flags & POOL_BH)
6654 continue;
6655
6656 mutex_lock(&wq_pool_attach_mutex);
6657 if (pool->cpu == cpu)
6658 rebind_workers(pool);
6659 else if (pool->cpu < 0)
6660 restore_unbound_workers_cpumask(pool, cpu);
6661 mutex_unlock(&wq_pool_attach_mutex);
6662 }
6663
6664 /* update pod affinity of unbound workqueues */
6665 list_for_each_entry(wq, &workqueues, list) {
6666 struct workqueue_attrs *attrs = wq->unbound_attrs;
6667
6668 if (attrs) {
6669 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6670 int tcpu;
6671
6672 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6673 unbound_wq_update_pwq(wq, tcpu);
6674
6675 mutex_lock(&wq->mutex);
6676 wq_update_node_max_active(wq, -1);
6677 mutex_unlock(&wq->mutex);
6678 }
6679 }
6680
6681 mutex_unlock(&wq_pool_mutex);
6682 return 0;
6683}
6684
6685int workqueue_offline_cpu(unsigned int cpu)
6686{
6687 struct workqueue_struct *wq;
6688
6689 /* unbinding per-cpu workers should happen on the local CPU */
6690 if (WARN_ON(cpu != smp_processor_id()))
6691 return -1;
6692
6693 unbind_workers(cpu);
6694
6695 /* update pod affinity of unbound workqueues */
6696 mutex_lock(&wq_pool_mutex);
6697
6698 cpumask_clear_cpu(cpu, wq_online_cpumask);
6699
6700 list_for_each_entry(wq, &workqueues, list) {
6701 struct workqueue_attrs *attrs = wq->unbound_attrs;
6702
6703 if (attrs) {
6704 const struct wq_pod_type *pt = wqattrs_pod_type(attrs);
6705 int tcpu;
6706
6707 for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]])
6708 unbound_wq_update_pwq(wq, tcpu);
6709
6710 mutex_lock(&wq->mutex);
6711 wq_update_node_max_active(wq, cpu);
6712 mutex_unlock(&wq->mutex);
6713 }
6714 }
6715 mutex_unlock(&wq_pool_mutex);
6716
6717 return 0;
6718}
6719
6720struct work_for_cpu {
6721 struct work_struct work;
6722 long (*fn)(void *);
6723 void *arg;
6724 long ret;
6725};
6726
6727static void work_for_cpu_fn(struct work_struct *work)
6728{
6729 struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
6730
6731 wfc->ret = wfc->fn(wfc->arg);
6732}
6733
6734/**
6735 * work_on_cpu_key - run a function in thread context on a particular cpu
6736 * @cpu: the cpu to run on
6737 * @fn: the function to run
6738 * @arg: the function arg
6739 * @key: The lock class key for lock debugging purposes
6740 *
6741 * It is up to the caller to ensure that the cpu doesn't go offline.
6742 * The caller must not hold any locks which would prevent @fn from completing.
6743 *
6744 * Return: The value @fn returns.
6745 */
6746long work_on_cpu_key(int cpu, long (*fn)(void *),
6747 void *arg, struct lock_class_key *key)
6748{
6749 struct work_for_cpu wfc = { .fn = fn, .arg = arg };
6750
6751 INIT_WORK_ONSTACK_KEY(&wfc.work, work_for_cpu_fn, key);
6752 schedule_work_on(cpu, &wfc.work);
6753 flush_work(&wfc.work);
6754 destroy_work_on_stack(&wfc.work);
6755 return wfc.ret;
6756}
6757EXPORT_SYMBOL_GPL(work_on_cpu_key);
6758
6759/**
6760 * work_on_cpu_safe_key - run a function in thread context on a particular cpu
6761 * @cpu: the cpu to run on
6762 * @fn: the function to run
6763 * @arg: the function argument
6764 * @key: The lock class key for lock debugging purposes
6765 *
6766 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
6767 * any locks which would prevent @fn from completing.
6768 *
6769 * Return: The value @fn returns.
6770 */
6771long work_on_cpu_safe_key(int cpu, long (*fn)(void *),
6772 void *arg, struct lock_class_key *key)
6773{
6774 long ret = -ENODEV;
6775
6776 cpus_read_lock();
6777 if (cpu_online(cpu))
6778 ret = work_on_cpu_key(cpu, fn, arg, key);
6779 cpus_read_unlock();
6780 return ret;
6781}
6782EXPORT_SYMBOL_GPL(work_on_cpu_safe_key);
6783#endif /* CONFIG_SMP */
6784
6785#ifdef CONFIG_FREEZER
6786
6787/**
6788 * freeze_workqueues_begin - begin freezing workqueues
6789 *
6790 * Start freezing workqueues. After this function returns, all freezable
6791 * workqueues will queue new works to their inactive_works list instead of
6792 * pool->worklist.
6793 *
6794 * CONTEXT:
6795 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6796 */
6797void freeze_workqueues_begin(void)
6798{
6799 struct workqueue_struct *wq;
6800
6801 mutex_lock(&wq_pool_mutex);
6802
6803 WARN_ON_ONCE(workqueue_freezing);
6804 workqueue_freezing = true;
6805
6806 list_for_each_entry(wq, &workqueues, list) {
6807 mutex_lock(&wq->mutex);
6808 wq_adjust_max_active(wq);
6809 mutex_unlock(&wq->mutex);
6810 }
6811
6812 mutex_unlock(&wq_pool_mutex);
6813}
6814
6815/**
6816 * freeze_workqueues_busy - are freezable workqueues still busy?
6817 *
6818 * Check whether freezing is complete. This function must be called
6819 * between freeze_workqueues_begin() and thaw_workqueues().
6820 *
6821 * CONTEXT:
6822 * Grabs and releases wq_pool_mutex.
6823 *
6824 * Return:
6825 * %true if some freezable workqueues are still busy. %false if freezing
6826 * is complete.
6827 */
6828bool freeze_workqueues_busy(void)
6829{
6830 bool busy = false;
6831 struct workqueue_struct *wq;
6832 struct pool_workqueue *pwq;
6833
6834 mutex_lock(&wq_pool_mutex);
6835
6836 WARN_ON_ONCE(!workqueue_freezing);
6837
6838 list_for_each_entry(wq, &workqueues, list) {
6839 if (!(wq->flags & WQ_FREEZABLE))
6840 continue;
6841 /*
6842 * nr_active is monotonically decreasing. It's safe
6843 * to peek without lock.
6844 */
6845 rcu_read_lock();
6846 for_each_pwq(pwq, wq) {
6847 WARN_ON_ONCE(pwq->nr_active < 0);
6848 if (pwq->nr_active) {
6849 busy = true;
6850 rcu_read_unlock();
6851 goto out_unlock;
6852 }
6853 }
6854 rcu_read_unlock();
6855 }
6856out_unlock:
6857 mutex_unlock(&wq_pool_mutex);
6858 return busy;
6859}
6860
6861/**
6862 * thaw_workqueues - thaw workqueues
6863 *
6864 * Thaw workqueues. Normal queueing is restored and all collected
6865 * frozen works are transferred to their respective pool worklists.
6866 *
6867 * CONTEXT:
6868 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
6869 */
6870void thaw_workqueues(void)
6871{
6872 struct workqueue_struct *wq;
6873
6874 mutex_lock(&wq_pool_mutex);
6875
6876 if (!workqueue_freezing)
6877 goto out_unlock;
6878
6879 workqueue_freezing = false;
6880
6881 /* restore max_active and repopulate worklist */
6882 list_for_each_entry(wq, &workqueues, list) {
6883 mutex_lock(&wq->mutex);
6884 wq_adjust_max_active(wq);
6885 mutex_unlock(&wq->mutex);
6886 }
6887
6888out_unlock:
6889 mutex_unlock(&wq_pool_mutex);
6890}
6891#endif /* CONFIG_FREEZER */
6892
6893static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask)
6894{
6895 LIST_HEAD(ctxs);
6896 int ret = 0;
6897 struct workqueue_struct *wq;
6898 struct apply_wqattrs_ctx *ctx, *n;
6899
6900 lockdep_assert_held(&wq_pool_mutex);
6901
6902 list_for_each_entry(wq, &workqueues, list) {
6903 if (!(wq->flags & WQ_UNBOUND) || (wq->flags & __WQ_DESTROYING))
6904 continue;
6905
6906 ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask);
6907 if (IS_ERR(ctx)) {
6908 ret = PTR_ERR(ctx);
6909 break;
6910 }
6911
6912 list_add_tail(&ctx->list, &ctxs);
6913 }
6914
6915 list_for_each_entry_safe(ctx, n, &ctxs, list) {
6916 if (!ret)
6917 apply_wqattrs_commit(ctx);
6918 apply_wqattrs_cleanup(ctx);
6919 }
6920
6921 if (!ret) {
6922 mutex_lock(&wq_pool_attach_mutex);
6923 cpumask_copy(wq_unbound_cpumask, unbound_cpumask);
6924 mutex_unlock(&wq_pool_attach_mutex);
6925 }
6926 return ret;
6927}
6928
6929/**
6930 * workqueue_unbound_exclude_cpumask - Exclude given CPUs from unbound cpumask
6931 * @exclude_cpumask: the cpumask to be excluded from wq_unbound_cpumask
6932 *
6933 * This function can be called from cpuset code to provide a set of isolated
6934 * CPUs that should be excluded from wq_unbound_cpumask.
6935 */
6936int workqueue_unbound_exclude_cpumask(cpumask_var_t exclude_cpumask)
6937{
6938 cpumask_var_t cpumask;
6939 int ret = 0;
6940
6941 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
6942 return -ENOMEM;
6943
6944 mutex_lock(&wq_pool_mutex);
6945
6946 /*
6947 * If the operation fails, it will fall back to
6948 * wq_requested_unbound_cpumask which is initially set to
6949 * (HK_TYPE_WQ ∩ HK_TYPE_DOMAIN) house keeping mask and rewritten
6950 * by any subsequent write to workqueue/cpumask sysfs file.
6951 */
6952 if (!cpumask_andnot(cpumask, wq_requested_unbound_cpumask, exclude_cpumask))
6953 cpumask_copy(cpumask, wq_requested_unbound_cpumask);
6954 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
6955 ret = workqueue_apply_unbound_cpumask(cpumask);
6956
6957 /* Save the current isolated cpumask & export it via sysfs */
6958 if (!ret)
6959 cpumask_copy(wq_isolated_cpumask, exclude_cpumask);
6960
6961 mutex_unlock(&wq_pool_mutex);
6962 free_cpumask_var(cpumask);
6963 return ret;
6964}
6965
6966static int parse_affn_scope(const char *val)
6967{
6968 int i;
6969
6970 for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) {
6971 if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i])))
6972 return i;
6973 }
6974 return -EINVAL;
6975}
6976
6977static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp)
6978{
6979 struct workqueue_struct *wq;
6980 int affn, cpu;
6981
6982 affn = parse_affn_scope(val);
6983 if (affn < 0)
6984 return affn;
6985 if (affn == WQ_AFFN_DFL)
6986 return -EINVAL;
6987
6988 cpus_read_lock();
6989 mutex_lock(&wq_pool_mutex);
6990
6991 wq_affn_dfl = affn;
6992
6993 list_for_each_entry(wq, &workqueues, list) {
6994 for_each_online_cpu(cpu)
6995 unbound_wq_update_pwq(wq, cpu);
6996 }
6997
6998 mutex_unlock(&wq_pool_mutex);
6999 cpus_read_unlock();
7000
7001 return 0;
7002}
7003
7004static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp)
7005{
7006 return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]);
7007}
7008
7009static const struct kernel_param_ops wq_affn_dfl_ops = {
7010 .set = wq_affn_dfl_set,
7011 .get = wq_affn_dfl_get,
7012};
7013
7014module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644);
7015
7016#ifdef CONFIG_SYSFS
7017/*
7018 * Workqueues with WQ_SYSFS flag set is visible to userland via
7019 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
7020 * following attributes.
7021 *
7022 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
7023 * max_active RW int : maximum number of in-flight work items
7024 *
7025 * Unbound workqueues have the following extra attributes.
7026 *
7027 * nice RW int : nice value of the workers
7028 * cpumask RW mask : bitmask of allowed CPUs for the workers
7029 * affinity_scope RW str : worker CPU affinity scope (cache, numa, none)
7030 * affinity_strict RW bool : worker CPU affinity is strict
7031 */
7032struct wq_device {
7033 struct workqueue_struct *wq;
7034 struct device dev;
7035};
7036
7037static struct workqueue_struct *dev_to_wq(struct device *dev)
7038{
7039 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7040
7041 return wq_dev->wq;
7042}
7043
7044static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
7045 char *buf)
7046{
7047 struct workqueue_struct *wq = dev_to_wq(dev);
7048
7049 return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
7050}
7051static DEVICE_ATTR_RO(per_cpu);
7052
7053static ssize_t max_active_show(struct device *dev,
7054 struct device_attribute *attr, char *buf)
7055{
7056 struct workqueue_struct *wq = dev_to_wq(dev);
7057
7058 return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
7059}
7060
7061static ssize_t max_active_store(struct device *dev,
7062 struct device_attribute *attr, const char *buf,
7063 size_t count)
7064{
7065 struct workqueue_struct *wq = dev_to_wq(dev);
7066 int val;
7067
7068 if (sscanf(buf, "%d", &val) != 1 || val <= 0)
7069 return -EINVAL;
7070
7071 workqueue_set_max_active(wq, val);
7072 return count;
7073}
7074static DEVICE_ATTR_RW(max_active);
7075
7076static struct attribute *wq_sysfs_attrs[] = {
7077 &dev_attr_per_cpu.attr,
7078 &dev_attr_max_active.attr,
7079 NULL,
7080};
7081ATTRIBUTE_GROUPS(wq_sysfs);
7082
7083static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
7084 char *buf)
7085{
7086 struct workqueue_struct *wq = dev_to_wq(dev);
7087 int written;
7088
7089 mutex_lock(&wq->mutex);
7090 written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
7091 mutex_unlock(&wq->mutex);
7092
7093 return written;
7094}
7095
7096/* prepare workqueue_attrs for sysfs store operations */
7097static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
7098{
7099 struct workqueue_attrs *attrs;
7100
7101 lockdep_assert_held(&wq_pool_mutex);
7102
7103 attrs = alloc_workqueue_attrs();
7104 if (!attrs)
7105 return NULL;
7106
7107 copy_workqueue_attrs(attrs, wq->unbound_attrs);
7108 return attrs;
7109}
7110
7111static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
7112 const char *buf, size_t count)
7113{
7114 struct workqueue_struct *wq = dev_to_wq(dev);
7115 struct workqueue_attrs *attrs;
7116 int ret = -ENOMEM;
7117
7118 apply_wqattrs_lock();
7119
7120 attrs = wq_sysfs_prep_attrs(wq);
7121 if (!attrs)
7122 goto out_unlock;
7123
7124 if (sscanf(buf, "%d", &attrs->nice) == 1 &&
7125 attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
7126 ret = apply_workqueue_attrs_locked(wq, attrs);
7127 else
7128 ret = -EINVAL;
7129
7130out_unlock:
7131 apply_wqattrs_unlock();
7132 free_workqueue_attrs(attrs);
7133 return ret ?: count;
7134}
7135
7136static ssize_t wq_cpumask_show(struct device *dev,
7137 struct device_attribute *attr, char *buf)
7138{
7139 struct workqueue_struct *wq = dev_to_wq(dev);
7140 int written;
7141
7142 mutex_lock(&wq->mutex);
7143 written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
7144 cpumask_pr_args(wq->unbound_attrs->cpumask));
7145 mutex_unlock(&wq->mutex);
7146 return written;
7147}
7148
7149static ssize_t wq_cpumask_store(struct device *dev,
7150 struct device_attribute *attr,
7151 const char *buf, size_t count)
7152{
7153 struct workqueue_struct *wq = dev_to_wq(dev);
7154 struct workqueue_attrs *attrs;
7155 int ret = -ENOMEM;
7156
7157 apply_wqattrs_lock();
7158
7159 attrs = wq_sysfs_prep_attrs(wq);
7160 if (!attrs)
7161 goto out_unlock;
7162
7163 ret = cpumask_parse(buf, attrs->cpumask);
7164 if (!ret)
7165 ret = apply_workqueue_attrs_locked(wq, attrs);
7166
7167out_unlock:
7168 apply_wqattrs_unlock();
7169 free_workqueue_attrs(attrs);
7170 return ret ?: count;
7171}
7172
7173static ssize_t wq_affn_scope_show(struct device *dev,
7174 struct device_attribute *attr, char *buf)
7175{
7176 struct workqueue_struct *wq = dev_to_wq(dev);
7177 int written;
7178
7179 mutex_lock(&wq->mutex);
7180 if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL)
7181 written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n",
7182 wq_affn_names[WQ_AFFN_DFL],
7183 wq_affn_names[wq_affn_dfl]);
7184 else
7185 written = scnprintf(buf, PAGE_SIZE, "%s\n",
7186 wq_affn_names[wq->unbound_attrs->affn_scope]);
7187 mutex_unlock(&wq->mutex);
7188
7189 return written;
7190}
7191
7192static ssize_t wq_affn_scope_store(struct device *dev,
7193 struct device_attribute *attr,
7194 const char *buf, size_t count)
7195{
7196 struct workqueue_struct *wq = dev_to_wq(dev);
7197 struct workqueue_attrs *attrs;
7198 int affn, ret = -ENOMEM;
7199
7200 affn = parse_affn_scope(buf);
7201 if (affn < 0)
7202 return affn;
7203
7204 apply_wqattrs_lock();
7205 attrs = wq_sysfs_prep_attrs(wq);
7206 if (attrs) {
7207 attrs->affn_scope = affn;
7208 ret = apply_workqueue_attrs_locked(wq, attrs);
7209 }
7210 apply_wqattrs_unlock();
7211 free_workqueue_attrs(attrs);
7212 return ret ?: count;
7213}
7214
7215static ssize_t wq_affinity_strict_show(struct device *dev,
7216 struct device_attribute *attr, char *buf)
7217{
7218 struct workqueue_struct *wq = dev_to_wq(dev);
7219
7220 return scnprintf(buf, PAGE_SIZE, "%d\n",
7221 wq->unbound_attrs->affn_strict);
7222}
7223
7224static ssize_t wq_affinity_strict_store(struct device *dev,
7225 struct device_attribute *attr,
7226 const char *buf, size_t count)
7227{
7228 struct workqueue_struct *wq = dev_to_wq(dev);
7229 struct workqueue_attrs *attrs;
7230 int v, ret = -ENOMEM;
7231
7232 if (sscanf(buf, "%d", &v) != 1)
7233 return -EINVAL;
7234
7235 apply_wqattrs_lock();
7236 attrs = wq_sysfs_prep_attrs(wq);
7237 if (attrs) {
7238 attrs->affn_strict = (bool)v;
7239 ret = apply_workqueue_attrs_locked(wq, attrs);
7240 }
7241 apply_wqattrs_unlock();
7242 free_workqueue_attrs(attrs);
7243 return ret ?: count;
7244}
7245
7246static struct device_attribute wq_sysfs_unbound_attrs[] = {
7247 __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
7248 __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
7249 __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store),
7250 __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store),
7251 __ATTR_NULL,
7252};
7253
7254static const struct bus_type wq_subsys = {
7255 .name = "workqueue",
7256 .dev_groups = wq_sysfs_groups,
7257};
7258
7259/**
7260 * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
7261 * @cpumask: the cpumask to set
7262 *
7263 * The low-level workqueues cpumask is a global cpumask that limits
7264 * the affinity of all unbound workqueues. This function check the @cpumask
7265 * and apply it to all unbound workqueues and updates all pwqs of them.
7266 *
7267 * Return: 0 - Success
7268 * -EINVAL - Invalid @cpumask
7269 * -ENOMEM - Failed to allocate memory for attrs or pwqs.
7270 */
7271static int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
7272{
7273 int ret = -EINVAL;
7274
7275 /*
7276 * Not excluding isolated cpus on purpose.
7277 * If the user wishes to include them, we allow that.
7278 */
7279 cpumask_and(cpumask, cpumask, cpu_possible_mask);
7280 if (!cpumask_empty(cpumask)) {
7281 ret = 0;
7282 apply_wqattrs_lock();
7283 if (!cpumask_equal(cpumask, wq_unbound_cpumask))
7284 ret = workqueue_apply_unbound_cpumask(cpumask);
7285 if (!ret)
7286 cpumask_copy(wq_requested_unbound_cpumask, cpumask);
7287 apply_wqattrs_unlock();
7288 }
7289
7290 return ret;
7291}
7292
7293static ssize_t __wq_cpumask_show(struct device *dev,
7294 struct device_attribute *attr, char *buf, cpumask_var_t mask)
7295{
7296 int written;
7297
7298 mutex_lock(&wq_pool_mutex);
7299 written = scnprintf(buf, PAGE_SIZE, "%*pb\n", cpumask_pr_args(mask));
7300 mutex_unlock(&wq_pool_mutex);
7301
7302 return written;
7303}
7304
7305static ssize_t cpumask_requested_show(struct device *dev,
7306 struct device_attribute *attr, char *buf)
7307{
7308 return __wq_cpumask_show(dev, attr, buf, wq_requested_unbound_cpumask);
7309}
7310static DEVICE_ATTR_RO(cpumask_requested);
7311
7312static ssize_t cpumask_isolated_show(struct device *dev,
7313 struct device_attribute *attr, char *buf)
7314{
7315 return __wq_cpumask_show(dev, attr, buf, wq_isolated_cpumask);
7316}
7317static DEVICE_ATTR_RO(cpumask_isolated);
7318
7319static ssize_t cpumask_show(struct device *dev,
7320 struct device_attribute *attr, char *buf)
7321{
7322 return __wq_cpumask_show(dev, attr, buf, wq_unbound_cpumask);
7323}
7324
7325static ssize_t cpumask_store(struct device *dev,
7326 struct device_attribute *attr, const char *buf, size_t count)
7327{
7328 cpumask_var_t cpumask;
7329 int ret;
7330
7331 if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
7332 return -ENOMEM;
7333
7334 ret = cpumask_parse(buf, cpumask);
7335 if (!ret)
7336 ret = workqueue_set_unbound_cpumask(cpumask);
7337
7338 free_cpumask_var(cpumask);
7339 return ret ? ret : count;
7340}
7341static DEVICE_ATTR_RW(cpumask);
7342
7343static struct attribute *wq_sysfs_cpumask_attrs[] = {
7344 &dev_attr_cpumask.attr,
7345 &dev_attr_cpumask_requested.attr,
7346 &dev_attr_cpumask_isolated.attr,
7347 NULL,
7348};
7349ATTRIBUTE_GROUPS(wq_sysfs_cpumask);
7350
7351static int __init wq_sysfs_init(void)
7352{
7353 return subsys_virtual_register(&wq_subsys, wq_sysfs_cpumask_groups);
7354}
7355core_initcall(wq_sysfs_init);
7356
7357static void wq_device_release(struct device *dev)
7358{
7359 struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
7360
7361 kfree(wq_dev);
7362}
7363
7364/**
7365 * workqueue_sysfs_register - make a workqueue visible in sysfs
7366 * @wq: the workqueue to register
7367 *
7368 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
7369 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
7370 * which is the preferred method.
7371 *
7372 * Workqueue user should use this function directly iff it wants to apply
7373 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
7374 * apply_workqueue_attrs() may race against userland updating the
7375 * attributes.
7376 *
7377 * Return: 0 on success, -errno on failure.
7378 */
7379int workqueue_sysfs_register(struct workqueue_struct *wq)
7380{
7381 struct wq_device *wq_dev;
7382 int ret;
7383
7384 /*
7385 * Adjusting max_active breaks ordering guarantee. Disallow exposing
7386 * ordered workqueues.
7387 */
7388 if (WARN_ON(wq->flags & __WQ_ORDERED))
7389 return -EINVAL;
7390
7391 wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
7392 if (!wq_dev)
7393 return -ENOMEM;
7394
7395 wq_dev->wq = wq;
7396 wq_dev->dev.bus = &wq_subsys;
7397 wq_dev->dev.release = wq_device_release;
7398 dev_set_name(&wq_dev->dev, "%s", wq->name);
7399
7400 /*
7401 * unbound_attrs are created separately. Suppress uevent until
7402 * everything is ready.
7403 */
7404 dev_set_uevent_suppress(&wq_dev->dev, true);
7405
7406 ret = device_register(&wq_dev->dev);
7407 if (ret) {
7408 put_device(&wq_dev->dev);
7409 wq->wq_dev = NULL;
7410 return ret;
7411 }
7412
7413 if (wq->flags & WQ_UNBOUND) {
7414 struct device_attribute *attr;
7415
7416 for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
7417 ret = device_create_file(&wq_dev->dev, attr);
7418 if (ret) {
7419 device_unregister(&wq_dev->dev);
7420 wq->wq_dev = NULL;
7421 return ret;
7422 }
7423 }
7424 }
7425
7426 dev_set_uevent_suppress(&wq_dev->dev, false);
7427 kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
7428 return 0;
7429}
7430
7431/**
7432 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
7433 * @wq: the workqueue to unregister
7434 *
7435 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
7436 */
7437static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
7438{
7439 struct wq_device *wq_dev = wq->wq_dev;
7440
7441 if (!wq->wq_dev)
7442 return;
7443
7444 wq->wq_dev = NULL;
7445 device_unregister(&wq_dev->dev);
7446}
7447#else /* CONFIG_SYSFS */
7448static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
7449#endif /* CONFIG_SYSFS */
7450
7451/*
7452 * Workqueue watchdog.
7453 *
7454 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
7455 * flush dependency, a concurrency managed work item which stays RUNNING
7456 * indefinitely. Workqueue stalls can be very difficult to debug as the
7457 * usual warning mechanisms don't trigger and internal workqueue state is
7458 * largely opaque.
7459 *
7460 * Workqueue watchdog monitors all worker pools periodically and dumps
7461 * state if some pools failed to make forward progress for a while where
7462 * forward progress is defined as the first item on ->worklist changing.
7463 *
7464 * This mechanism is controlled through the kernel parameter
7465 * "workqueue.watchdog_thresh" which can be updated at runtime through the
7466 * corresponding sysfs parameter file.
7467 */
7468#ifdef CONFIG_WQ_WATCHDOG
7469
7470static unsigned long wq_watchdog_thresh = 30;
7471static struct timer_list wq_watchdog_timer;
7472
7473static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
7474static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
7475
7476static unsigned int wq_panic_on_stall;
7477module_param_named(panic_on_stall, wq_panic_on_stall, uint, 0644);
7478
7479/*
7480 * Show workers that might prevent the processing of pending work items.
7481 * The only candidates are CPU-bound workers in the running state.
7482 * Pending work items should be handled by another idle worker
7483 * in all other situations.
7484 */
7485static void show_cpu_pool_hog(struct worker_pool *pool)
7486{
7487 struct worker *worker;
7488 unsigned long irq_flags;
7489 int bkt;
7490
7491 raw_spin_lock_irqsave(&pool->lock, irq_flags);
7492
7493 hash_for_each(pool->busy_hash, bkt, worker, hentry) {
7494 if (task_is_running(worker->task)) {
7495 /*
7496 * Defer printing to avoid deadlocks in console
7497 * drivers that queue work while holding locks
7498 * also taken in their write paths.
7499 */
7500 printk_deferred_enter();
7501
7502 pr_info("pool %d:\n", pool->id);
7503 sched_show_task(worker->task);
7504
7505 printk_deferred_exit();
7506 }
7507 }
7508
7509 raw_spin_unlock_irqrestore(&pool->lock, irq_flags);
7510}
7511
7512static void show_cpu_pools_hogs(void)
7513{
7514 struct worker_pool *pool;
7515 int pi;
7516
7517 pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n");
7518
7519 rcu_read_lock();
7520
7521 for_each_pool(pool, pi) {
7522 if (pool->cpu_stall)
7523 show_cpu_pool_hog(pool);
7524
7525 }
7526
7527 rcu_read_unlock();
7528}
7529
7530static void panic_on_wq_watchdog(void)
7531{
7532 static unsigned int wq_stall;
7533
7534 if (wq_panic_on_stall) {
7535 wq_stall++;
7536 BUG_ON(wq_stall >= wq_panic_on_stall);
7537 }
7538}
7539
7540static void wq_watchdog_reset_touched(void)
7541{
7542 int cpu;
7543
7544 wq_watchdog_touched = jiffies;
7545 for_each_possible_cpu(cpu)
7546 per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
7547}
7548
7549static void wq_watchdog_timer_fn(struct timer_list *unused)
7550{
7551 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7552 bool lockup_detected = false;
7553 bool cpu_pool_stall = false;
7554 unsigned long now = jiffies;
7555 struct worker_pool *pool;
7556 int pi;
7557
7558 if (!thresh)
7559 return;
7560
7561 rcu_read_lock();
7562
7563 for_each_pool(pool, pi) {
7564 unsigned long pool_ts, touched, ts;
7565
7566 pool->cpu_stall = false;
7567 if (list_empty(&pool->worklist))
7568 continue;
7569
7570 /*
7571 * If a virtual machine is stopped by the host it can look to
7572 * the watchdog like a stall.
7573 */
7574 kvm_check_and_clear_guest_paused();
7575
7576 /* get the latest of pool and touched timestamps */
7577 if (pool->cpu >= 0)
7578 touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu));
7579 else
7580 touched = READ_ONCE(wq_watchdog_touched);
7581 pool_ts = READ_ONCE(pool->watchdog_ts);
7582
7583 if (time_after(pool_ts, touched))
7584 ts = pool_ts;
7585 else
7586 ts = touched;
7587
7588 /* did we stall? */
7589 if (time_after(now, ts + thresh)) {
7590 lockup_detected = true;
7591 if (pool->cpu >= 0 && !(pool->flags & POOL_BH)) {
7592 pool->cpu_stall = true;
7593 cpu_pool_stall = true;
7594 }
7595 pr_emerg("BUG: workqueue lockup - pool");
7596 pr_cont_pool_info(pool);
7597 pr_cont(" stuck for %us!\n",
7598 jiffies_to_msecs(now - pool_ts) / 1000);
7599 }
7600
7601
7602 }
7603
7604 rcu_read_unlock();
7605
7606 if (lockup_detected)
7607 show_all_workqueues();
7608
7609 if (cpu_pool_stall)
7610 show_cpu_pools_hogs();
7611
7612 if (lockup_detected)
7613 panic_on_wq_watchdog();
7614
7615 wq_watchdog_reset_touched();
7616 mod_timer(&wq_watchdog_timer, jiffies + thresh);
7617}
7618
7619notrace void wq_watchdog_touch(int cpu)
7620{
7621 unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
7622 unsigned long touch_ts = READ_ONCE(wq_watchdog_touched);
7623 unsigned long now = jiffies;
7624
7625 if (cpu >= 0)
7626 per_cpu(wq_watchdog_touched_cpu, cpu) = now;
7627 else
7628 WARN_ONCE(1, "%s should be called with valid CPU", __func__);
7629
7630 /* Don't unnecessarily store to global cacheline */
7631 if (time_after(now, touch_ts + thresh / 4))
7632 WRITE_ONCE(wq_watchdog_touched, jiffies);
7633}
7634
7635static void wq_watchdog_set_thresh(unsigned long thresh)
7636{
7637 wq_watchdog_thresh = 0;
7638 del_timer_sync(&wq_watchdog_timer);
7639
7640 if (thresh) {
7641 wq_watchdog_thresh = thresh;
7642 wq_watchdog_reset_touched();
7643 mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
7644 }
7645}
7646
7647static int wq_watchdog_param_set_thresh(const char *val,
7648 const struct kernel_param *kp)
7649{
7650 unsigned long thresh;
7651 int ret;
7652
7653 ret = kstrtoul(val, 0, &thresh);
7654 if (ret)
7655 return ret;
7656
7657 if (system_wq)
7658 wq_watchdog_set_thresh(thresh);
7659 else
7660 wq_watchdog_thresh = thresh;
7661
7662 return 0;
7663}
7664
7665static const struct kernel_param_ops wq_watchdog_thresh_ops = {
7666 .set = wq_watchdog_param_set_thresh,
7667 .get = param_get_ulong,
7668};
7669
7670module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
7671 0644);
7672
7673static void wq_watchdog_init(void)
7674{
7675 timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
7676 wq_watchdog_set_thresh(wq_watchdog_thresh);
7677}
7678
7679#else /* CONFIG_WQ_WATCHDOG */
7680
7681static inline void wq_watchdog_init(void) { }
7682
7683#endif /* CONFIG_WQ_WATCHDOG */
7684
7685static void bh_pool_kick_normal(struct irq_work *irq_work)
7686{
7687 raise_softirq_irqoff(TASKLET_SOFTIRQ);
7688}
7689
7690static void bh_pool_kick_highpri(struct irq_work *irq_work)
7691{
7692 raise_softirq_irqoff(HI_SOFTIRQ);
7693}
7694
7695static void __init restrict_unbound_cpumask(const char *name, const struct cpumask *mask)
7696{
7697 if (!cpumask_intersects(wq_unbound_cpumask, mask)) {
7698 pr_warn("workqueue: Restricting unbound_cpumask (%*pb) with %s (%*pb) leaves no CPU, ignoring\n",
7699 cpumask_pr_args(wq_unbound_cpumask), name, cpumask_pr_args(mask));
7700 return;
7701 }
7702
7703 cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, mask);
7704}
7705
7706static void __init init_cpu_worker_pool(struct worker_pool *pool, int cpu, int nice)
7707{
7708 BUG_ON(init_worker_pool(pool));
7709 pool->cpu = cpu;
7710 cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
7711 cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu));
7712 pool->attrs->nice = nice;
7713 pool->attrs->affn_strict = true;
7714 pool->node = cpu_to_node(cpu);
7715
7716 /* alloc pool ID */
7717 mutex_lock(&wq_pool_mutex);
7718 BUG_ON(worker_pool_assign_id(pool));
7719 mutex_unlock(&wq_pool_mutex);
7720}
7721
7722/**
7723 * workqueue_init_early - early init for workqueue subsystem
7724 *
7725 * This is the first step of three-staged workqueue subsystem initialization and
7726 * invoked as soon as the bare basics - memory allocation, cpumasks and idr are
7727 * up. It sets up all the data structures and system workqueues and allows early
7728 * boot code to create workqueues and queue/cancel work items. Actual work item
7729 * execution starts only after kthreads can be created and scheduled right
7730 * before early initcalls.
7731 */
7732void __init workqueue_init_early(void)
7733{
7734 struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM];
7735 int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
7736 void (*irq_work_fns[2])(struct irq_work *) = { bh_pool_kick_normal,
7737 bh_pool_kick_highpri };
7738 int i, cpu;
7739
7740 BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
7741
7742 BUG_ON(!alloc_cpumask_var(&wq_online_cpumask, GFP_KERNEL));
7743 BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
7744 BUG_ON(!alloc_cpumask_var(&wq_requested_unbound_cpumask, GFP_KERNEL));
7745 BUG_ON(!zalloc_cpumask_var(&wq_isolated_cpumask, GFP_KERNEL));
7746
7747 cpumask_copy(wq_online_cpumask, cpu_online_mask);
7748 cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
7749 restrict_unbound_cpumask("HK_TYPE_WQ", housekeeping_cpumask(HK_TYPE_WQ));
7750 restrict_unbound_cpumask("HK_TYPE_DOMAIN", housekeeping_cpumask(HK_TYPE_DOMAIN));
7751 if (!cpumask_empty(&wq_cmdline_cpumask))
7752 restrict_unbound_cpumask("workqueue.unbound_cpus", &wq_cmdline_cpumask);
7753
7754 cpumask_copy(wq_requested_unbound_cpumask, wq_unbound_cpumask);
7755
7756 pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
7757
7758 unbound_wq_update_pwq_attrs_buf = alloc_workqueue_attrs();
7759 BUG_ON(!unbound_wq_update_pwq_attrs_buf);
7760
7761 /*
7762 * If nohz_full is enabled, set power efficient workqueue as unbound.
7763 * This allows workqueue items to be moved to HK CPUs.
7764 */
7765 if (housekeeping_enabled(HK_TYPE_TICK))
7766 wq_power_efficient = true;
7767
7768 /* initialize WQ_AFFN_SYSTEM pods */
7769 pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7770 pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL);
7771 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7772 BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod);
7773
7774 BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE));
7775
7776 pt->nr_pods = 1;
7777 cpumask_copy(pt->pod_cpus[0], cpu_possible_mask);
7778 pt->pod_node[0] = NUMA_NO_NODE;
7779 pt->cpu_pod[0] = 0;
7780
7781 /* initialize BH and CPU pools */
7782 for_each_possible_cpu(cpu) {
7783 struct worker_pool *pool;
7784
7785 i = 0;
7786 for_each_bh_worker_pool(pool, cpu) {
7787 init_cpu_worker_pool(pool, cpu, std_nice[i]);
7788 pool->flags |= POOL_BH;
7789 init_irq_work(bh_pool_irq_work(pool), irq_work_fns[i]);
7790 i++;
7791 }
7792
7793 i = 0;
7794 for_each_cpu_worker_pool(pool, cpu)
7795 init_cpu_worker_pool(pool, cpu, std_nice[i++]);
7796 }
7797
7798 /* create default unbound and ordered wq attrs */
7799 for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
7800 struct workqueue_attrs *attrs;
7801
7802 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7803 attrs->nice = std_nice[i];
7804 unbound_std_wq_attrs[i] = attrs;
7805
7806 /*
7807 * An ordered wq should have only one pwq as ordering is
7808 * guaranteed by max_active which is enforced by pwqs.
7809 */
7810 BUG_ON(!(attrs = alloc_workqueue_attrs()));
7811 attrs->nice = std_nice[i];
7812 attrs->ordered = true;
7813 ordered_wq_attrs[i] = attrs;
7814 }
7815
7816 system_wq = alloc_workqueue("events", 0, 0);
7817 system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
7818 system_long_wq = alloc_workqueue("events_long", 0, 0);
7819 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
7820 WQ_MAX_ACTIVE);
7821 system_freezable_wq = alloc_workqueue("events_freezable",
7822 WQ_FREEZABLE, 0);
7823 system_power_efficient_wq = alloc_workqueue("events_power_efficient",
7824 WQ_POWER_EFFICIENT, 0);
7825 system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_pwr_efficient",
7826 WQ_FREEZABLE | WQ_POWER_EFFICIENT,
7827 0);
7828 system_bh_wq = alloc_workqueue("events_bh", WQ_BH, 0);
7829 system_bh_highpri_wq = alloc_workqueue("events_bh_highpri",
7830 WQ_BH | WQ_HIGHPRI, 0);
7831 BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
7832 !system_unbound_wq || !system_freezable_wq ||
7833 !system_power_efficient_wq ||
7834 !system_freezable_power_efficient_wq ||
7835 !system_bh_wq || !system_bh_highpri_wq);
7836}
7837
7838static void __init wq_cpu_intensive_thresh_init(void)
7839{
7840 unsigned long thresh;
7841 unsigned long bogo;
7842
7843 pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release");
7844 BUG_ON(IS_ERR(pwq_release_worker));
7845
7846 /* if the user set it to a specific value, keep it */
7847 if (wq_cpu_intensive_thresh_us != ULONG_MAX)
7848 return;
7849
7850 /*
7851 * The default of 10ms is derived from the fact that most modern (as of
7852 * 2023) processors can do a lot in 10ms and that it's just below what
7853 * most consider human-perceivable. However, the kernel also runs on a
7854 * lot slower CPUs including microcontrollers where the threshold is way
7855 * too low.
7856 *
7857 * Let's scale up the threshold upto 1 second if BogoMips is below 4000.
7858 * This is by no means accurate but it doesn't have to be. The mechanism
7859 * is still useful even when the threshold is fully scaled up. Also, as
7860 * the reports would usually be applicable to everyone, some machines
7861 * operating on longer thresholds won't significantly diminish their
7862 * usefulness.
7863 */
7864 thresh = 10 * USEC_PER_MSEC;
7865
7866 /* see init/calibrate.c for lpj -> BogoMIPS calculation */
7867 bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1);
7868 if (bogo < 4000)
7869 thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC);
7870
7871 pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n",
7872 loops_per_jiffy, bogo, thresh);
7873
7874 wq_cpu_intensive_thresh_us = thresh;
7875}
7876
7877/**
7878 * workqueue_init - bring workqueue subsystem fully online
7879 *
7880 * This is the second step of three-staged workqueue subsystem initialization
7881 * and invoked as soon as kthreads can be created and scheduled. Workqueues have
7882 * been created and work items queued on them, but there are no kworkers
7883 * executing the work items yet. Populate the worker pools with the initial
7884 * workers and enable future kworker creations.
7885 */
7886void __init workqueue_init(void)
7887{
7888 struct workqueue_struct *wq;
7889 struct worker_pool *pool;
7890 int cpu, bkt;
7891
7892 wq_cpu_intensive_thresh_init();
7893
7894 mutex_lock(&wq_pool_mutex);
7895
7896 /*
7897 * Per-cpu pools created earlier could be missing node hint. Fix them
7898 * up. Also, create a rescuer for workqueues that requested it.
7899 */
7900 for_each_possible_cpu(cpu) {
7901 for_each_bh_worker_pool(pool, cpu)
7902 pool->node = cpu_to_node(cpu);
7903 for_each_cpu_worker_pool(pool, cpu)
7904 pool->node = cpu_to_node(cpu);
7905 }
7906
7907 list_for_each_entry(wq, &workqueues, list) {
7908 WARN(init_rescuer(wq),
7909 "workqueue: failed to create early rescuer for %s",
7910 wq->name);
7911 }
7912
7913 mutex_unlock(&wq_pool_mutex);
7914
7915 /*
7916 * Create the initial workers. A BH pool has one pseudo worker that
7917 * represents the shared BH execution context and thus doesn't get
7918 * affected by hotplug events. Create the BH pseudo workers for all
7919 * possible CPUs here.
7920 */
7921 for_each_possible_cpu(cpu)
7922 for_each_bh_worker_pool(pool, cpu)
7923 BUG_ON(!create_worker(pool));
7924
7925 for_each_online_cpu(cpu) {
7926 for_each_cpu_worker_pool(pool, cpu) {
7927 pool->flags &= ~POOL_DISASSOCIATED;
7928 BUG_ON(!create_worker(pool));
7929 }
7930 }
7931
7932 hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
7933 BUG_ON(!create_worker(pool));
7934
7935 wq_online = true;
7936 wq_watchdog_init();
7937}
7938
7939/*
7940 * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to
7941 * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique
7942 * and consecutive pod ID. The rest of @pt is initialized accordingly.
7943 */
7944static void __init init_pod_type(struct wq_pod_type *pt,
7945 bool (*cpus_share_pod)(int, int))
7946{
7947 int cur, pre, cpu, pod;
7948
7949 pt->nr_pods = 0;
7950
7951 /* init @pt->cpu_pod[] according to @cpus_share_pod() */
7952 pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL);
7953 BUG_ON(!pt->cpu_pod);
7954
7955 for_each_possible_cpu(cur) {
7956 for_each_possible_cpu(pre) {
7957 if (pre >= cur) {
7958 pt->cpu_pod[cur] = pt->nr_pods++;
7959 break;
7960 }
7961 if (cpus_share_pod(cur, pre)) {
7962 pt->cpu_pod[cur] = pt->cpu_pod[pre];
7963 break;
7964 }
7965 }
7966 }
7967
7968 /* init the rest to match @pt->cpu_pod[] */
7969 pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL);
7970 pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL);
7971 BUG_ON(!pt->pod_cpus || !pt->pod_node);
7972
7973 for (pod = 0; pod < pt->nr_pods; pod++)
7974 BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL));
7975
7976 for_each_possible_cpu(cpu) {
7977 cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]);
7978 pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu);
7979 }
7980}
7981
7982static bool __init cpus_dont_share(int cpu0, int cpu1)
7983{
7984 return false;
7985}
7986
7987static bool __init cpus_share_smt(int cpu0, int cpu1)
7988{
7989#ifdef CONFIG_SCHED_SMT
7990 return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1));
7991#else
7992 return false;
7993#endif
7994}
7995
7996static bool __init cpus_share_numa(int cpu0, int cpu1)
7997{
7998 return cpu_to_node(cpu0) == cpu_to_node(cpu1);
7999}
8000
8001/**
8002 * workqueue_init_topology - initialize CPU pods for unbound workqueues
8003 *
8004 * This is the third step of three-staged workqueue subsystem initialization and
8005 * invoked after SMP and topology information are fully initialized. It
8006 * initializes the unbound CPU pods accordingly.
8007 */
8008void __init workqueue_init_topology(void)
8009{
8010 struct workqueue_struct *wq;
8011 int cpu;
8012
8013 init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share);
8014 init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt);
8015 init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache);
8016 init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa);
8017
8018 wq_topo_initialized = true;
8019
8020 mutex_lock(&wq_pool_mutex);
8021
8022 /*
8023 * Workqueues allocated earlier would have all CPUs sharing the default
8024 * worker pool. Explicitly call unbound_wq_update_pwq() on all workqueue
8025 * and CPU combinations to apply per-pod sharing.
8026 */
8027 list_for_each_entry(wq, &workqueues, list) {
8028 for_each_online_cpu(cpu)
8029 unbound_wq_update_pwq(wq, cpu);
8030 if (wq->flags & WQ_UNBOUND) {
8031 mutex_lock(&wq->mutex);
8032 wq_update_node_max_active(wq, -1);
8033 mutex_unlock(&wq->mutex);
8034 }
8035 }
8036
8037 mutex_unlock(&wq_pool_mutex);
8038}
8039
8040void __warn_flushing_systemwide_wq(void)
8041{
8042 pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n");
8043 dump_stack();
8044}
8045EXPORT_SYMBOL(__warn_flushing_systemwide_wq);
8046
8047static int __init workqueue_unbound_cpus_setup(char *str)
8048{
8049 if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) {
8050 cpumask_clear(&wq_cmdline_cpumask);
8051 pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n");
8052 }
8053
8054 return 1;
8055}
8056__setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup);