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