Root/kernel/workqueue.c

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 is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
22 *
23 * Please read Documentation/workqueue.txt for details.
24 */
25
26#include <linux/export.h>
27#include <linux/kernel.h>
28#include <linux/sched.h>
29#include <linux/init.h>
30#include <linux/signal.h>
31#include <linux/completion.h>
32#include <linux/workqueue.h>
33#include <linux/slab.h>
34#include <linux/cpu.h>
35#include <linux/notifier.h>
36#include <linux/kthread.h>
37#include <linux/hardirq.h>
38#include <linux/mempolicy.h>
39#include <linux/freezer.h>
40#include <linux/kallsyms.h>
41#include <linux/debug_locks.h>
42#include <linux/lockdep.h>
43#include <linux/idr.h>
44
45#include "workqueue_sched.h"
46
47enum {
48    /* global_cwq flags */
49    GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50    GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51    GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
52    GCWQ_FREEZING = 1 << 3, /* freeze in progress */
53    GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
54
55    /* worker flags */
56    WORKER_STARTED = 1 << 0, /* started */
57    WORKER_DIE = 1 << 1, /* die die die */
58    WORKER_IDLE = 1 << 2, /* is idle */
59    WORKER_PREP = 1 << 3, /* preparing to run works */
60    WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
61    WORKER_REBIND = 1 << 5, /* mom is home, come back */
62    WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
63    WORKER_UNBOUND = 1 << 7, /* worker is unbound */
64
65    WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
66                  WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
67
68    /* gcwq->trustee_state */
69    TRUSTEE_START = 0, /* start */
70    TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71    TRUSTEE_BUTCHER = 2, /* butcher workers */
72    TRUSTEE_RELEASE = 3, /* release workers */
73    TRUSTEE_DONE = 4, /* trustee is done */
74
75    BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76    BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77    BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
78
79    MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80    IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
81
82    MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83                        /* call for help after 10ms
84                           (min two ticks) */
85    MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86    CREATE_COOLDOWN = HZ, /* time to breath after fail */
87    TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
88
89    /*
90     * Rescue workers are used only on emergencies and shared by
91     * all cpus. Give -20.
92     */
93    RESCUER_NICE_LEVEL = -20,
94};
95
96/*
97 * Structure fields follow one of the following exclusion rules.
98 *
99 * I: Modifiable by initialization/destruction paths and read-only for
100 * everyone else.
101 *
102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
104 *
105 * L: gcwq->lock protected. Access with gcwq->lock held.
106 *
107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
111 *
112 * F: wq->flush_mutex protected.
113 *
114 * W: workqueue_lock protected.
115 */
116
117struct global_cwq;
118
119/*
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
122 */
123struct worker {
124    /* on idle list while idle, on busy hash table while busy */
125    union {
126        struct list_head entry; /* L: while idle */
127        struct hlist_node hentry; /* L: while busy */
128    };
129
130    struct work_struct *current_work; /* L: work being processed */
131    struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
132    struct list_head scheduled; /* L: scheduled works */
133    struct task_struct *task; /* I: worker task */
134    struct global_cwq *gcwq; /* I: the associated gcwq */
135    /* 64 bytes boundary on 64bit, 32 on 32bit */
136    unsigned long last_active; /* L: last active timestamp */
137    unsigned int flags; /* X: flags */
138    int id; /* I: worker id */
139    struct work_struct rebind_work; /* L: rebind worker to cpu */
140};
141
142/*
143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
145 * target workqueues.
146 */
147struct global_cwq {
148    spinlock_t lock; /* the gcwq lock */
149    struct list_head worklist; /* L: list of pending works */
150    unsigned int cpu; /* I: the associated cpu */
151    unsigned int flags; /* L: GCWQ_* flags */
152
153    int nr_workers; /* L: total number of workers */
154    int nr_idle; /* L: currently idle ones */
155
156    /* workers are chained either in the idle_list or busy_hash */
157    struct list_head idle_list; /* X: list of idle workers */
158    struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159                        /* L: hash of busy workers */
160
161    struct timer_list idle_timer; /* L: worker idle timeout */
162    struct timer_list mayday_timer; /* L: SOS timer for dworkers */
163
164    struct ida worker_ida; /* L: for worker IDs */
165
166    struct task_struct *trustee; /* L: for gcwq shutdown */
167    unsigned int trustee_state; /* L: trustee state */
168    wait_queue_head_t trustee_wait; /* trustee wait */
169    struct worker *first_idle; /* L: first idle worker */
170} ____cacheline_aligned_in_smp;
171
172/*
173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
176 */
177struct cpu_workqueue_struct {
178    struct global_cwq *gcwq; /* I: the associated gcwq */
179    struct workqueue_struct *wq; /* I: the owning workqueue */
180    int work_color; /* L: current color */
181    int flush_color; /* L: flushing color */
182    int nr_in_flight[WORK_NR_COLORS];
183                        /* L: nr of in_flight works */
184    int nr_active; /* L: nr of active works */
185    int max_active; /* L: max active works */
186    struct list_head delayed_works; /* L: delayed works */
187};
188
189/*
190 * Structure used to wait for workqueue flush.
191 */
192struct wq_flusher {
193    struct list_head list; /* F: list of flushers */
194    int flush_color; /* F: flush color waiting for */
195    struct completion done; /* flush completion */
196};
197
198/*
199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
201 */
202#ifdef CONFIG_SMP
203typedef cpumask_var_t mayday_mask_t;
204#define mayday_test_and_set_cpu(cpu, mask) \
205    cpumask_test_and_set_cpu((cpu), (mask))
206#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
208#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
209#define free_mayday_mask(mask) free_cpumask_var((mask))
210#else
211typedef unsigned long mayday_mask_t;
212#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215#define alloc_mayday_mask(maskp, gfp) true
216#define free_mayday_mask(mask) do { } while (0)
217#endif
218
219/*
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
222 */
223struct workqueue_struct {
224    unsigned int flags; /* W: WQ_* flags */
225    union {
226        struct cpu_workqueue_struct __percpu *pcpu;
227        struct cpu_workqueue_struct *single;
228        unsigned long v;
229    } cpu_wq; /* I: cwq's */
230    struct list_head list; /* W: list of all workqueues */
231
232    struct mutex flush_mutex; /* protects wq flushing */
233    int work_color; /* F: current work color */
234    int flush_color; /* F: current flush color */
235    atomic_t nr_cwqs_to_flush; /* flush in progress */
236    struct wq_flusher *first_flusher; /* F: first flusher */
237    struct list_head flusher_queue; /* F: flush waiters */
238    struct list_head flusher_overflow; /* F: flush overflow list */
239
240    mayday_mask_t mayday_mask; /* cpus requesting rescue */
241    struct worker *rescuer; /* I: rescue worker */
242
243    int nr_drainers; /* W: drain in progress */
244    int saved_max_active; /* W: saved cwq max_active */
245#ifdef CONFIG_LOCKDEP
246    struct lockdep_map lockdep_map;
247#endif
248    char name[]; /* I: workqueue name */
249};
250
251struct workqueue_struct *system_wq __read_mostly;
252struct workqueue_struct *system_long_wq __read_mostly;
253struct workqueue_struct *system_nrt_wq __read_mostly;
254struct workqueue_struct *system_unbound_wq __read_mostly;
255struct workqueue_struct *system_freezable_wq __read_mostly;
256struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
257EXPORT_SYMBOL_GPL(system_wq);
258EXPORT_SYMBOL_GPL(system_long_wq);
259EXPORT_SYMBOL_GPL(system_nrt_wq);
260EXPORT_SYMBOL_GPL(system_unbound_wq);
261EXPORT_SYMBOL_GPL(system_freezable_wq);
262EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
263
264#define CREATE_TRACE_POINTS
265#include <trace/events/workqueue.h>
266
267#define for_each_busy_worker(worker, i, pos, gcwq) \
268    for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
269        hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
270
271static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
272                  unsigned int sw)
273{
274    if (cpu < nr_cpu_ids) {
275        if (sw & 1) {
276            cpu = cpumask_next(cpu, mask);
277            if (cpu < nr_cpu_ids)
278                return cpu;
279        }
280        if (sw & 2)
281            return WORK_CPU_UNBOUND;
282    }
283    return WORK_CPU_NONE;
284}
285
286static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
287                struct workqueue_struct *wq)
288{
289    return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
290}
291
292/*
293 * CPU iterators
294 *
295 * An extra gcwq is defined for an invalid cpu number
296 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
297 * specific CPU. The following iterators are similar to
298 * for_each_*_cpu() iterators but also considers the unbound gcwq.
299 *
300 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
301 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
302 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
303 * WORK_CPU_UNBOUND for unbound workqueues
304 */
305#define for_each_gcwq_cpu(cpu) \
306    for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
307         (cpu) < WORK_CPU_NONE; \
308         (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
309
310#define for_each_online_gcwq_cpu(cpu) \
311    for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
312         (cpu) < WORK_CPU_NONE; \
313         (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
314
315#define for_each_cwq_cpu(cpu, wq) \
316    for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
317         (cpu) < WORK_CPU_NONE; \
318         (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
319
320#ifdef CONFIG_DEBUG_OBJECTS_WORK
321
322static struct debug_obj_descr work_debug_descr;
323
324static void *work_debug_hint(void *addr)
325{
326    return ((struct work_struct *) addr)->func;
327}
328
329/*
330 * fixup_init is called when:
331 * - an active object is initialized
332 */
333static int work_fixup_init(void *addr, enum debug_obj_state state)
334{
335    struct work_struct *work = addr;
336
337    switch (state) {
338    case ODEBUG_STATE_ACTIVE:
339        cancel_work_sync(work);
340        debug_object_init(work, &work_debug_descr);
341        return 1;
342    default:
343        return 0;
344    }
345}
346
347/*
348 * fixup_activate is called when:
349 * - an active object is activated
350 * - an unknown object is activated (might be a statically initialized object)
351 */
352static int work_fixup_activate(void *addr, enum debug_obj_state state)
353{
354    struct work_struct *work = addr;
355
356    switch (state) {
357
358    case ODEBUG_STATE_NOTAVAILABLE:
359        /*
360         * This is not really a fixup. The work struct was
361         * statically initialized. We just make sure that it
362         * is tracked in the object tracker.
363         */
364        if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
365            debug_object_init(work, &work_debug_descr);
366            debug_object_activate(work, &work_debug_descr);
367            return 0;
368        }
369        WARN_ON_ONCE(1);
370        return 0;
371
372    case ODEBUG_STATE_ACTIVE:
373        WARN_ON(1);
374
375    default:
376        return 0;
377    }
378}
379
380/*
381 * fixup_free is called when:
382 * - an active object is freed
383 */
384static int work_fixup_free(void *addr, enum debug_obj_state state)
385{
386    struct work_struct *work = addr;
387
388    switch (state) {
389    case ODEBUG_STATE_ACTIVE:
390        cancel_work_sync(work);
391        debug_object_free(work, &work_debug_descr);
392        return 1;
393    default:
394        return 0;
395    }
396}
397
398static struct debug_obj_descr work_debug_descr = {
399    .name = "work_struct",
400    .debug_hint = work_debug_hint,
401    .fixup_init = work_fixup_init,
402    .fixup_activate = work_fixup_activate,
403    .fixup_free = work_fixup_free,
404};
405
406static inline void debug_work_activate(struct work_struct *work)
407{
408    debug_object_activate(work, &work_debug_descr);
409}
410
411static inline void debug_work_deactivate(struct work_struct *work)
412{
413    debug_object_deactivate(work, &work_debug_descr);
414}
415
416void __init_work(struct work_struct *work, int onstack)
417{
418    if (onstack)
419        debug_object_init_on_stack(work, &work_debug_descr);
420    else
421        debug_object_init(work, &work_debug_descr);
422}
423EXPORT_SYMBOL_GPL(__init_work);
424
425void destroy_work_on_stack(struct work_struct *work)
426{
427    debug_object_free(work, &work_debug_descr);
428}
429EXPORT_SYMBOL_GPL(destroy_work_on_stack);
430
431#else
432static inline void debug_work_activate(struct work_struct *work) { }
433static inline void debug_work_deactivate(struct work_struct *work) { }
434#endif
435
436/* Serializes the accesses to the list of workqueues. */
437static DEFINE_SPINLOCK(workqueue_lock);
438static LIST_HEAD(workqueues);
439static bool workqueue_freezing; /* W: have wqs started freezing? */
440
441/*
442 * The almighty global cpu workqueues. nr_running is the only field
443 * which is expected to be used frequently by other cpus via
444 * try_to_wake_up(). Put it in a separate cacheline.
445 */
446static DEFINE_PER_CPU(struct global_cwq, global_cwq);
447static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
448
449/*
450 * Global cpu workqueue and nr_running counter for unbound gcwq. The
451 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
452 * workers have WORKER_UNBOUND set.
453 */
454static struct global_cwq unbound_global_cwq;
455static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
456
457static int worker_thread(void *__worker);
458
459static struct global_cwq *get_gcwq(unsigned int cpu)
460{
461    if (cpu != WORK_CPU_UNBOUND)
462        return &per_cpu(global_cwq, cpu);
463    else
464        return &unbound_global_cwq;
465}
466
467static atomic_t *get_gcwq_nr_running(unsigned int cpu)
468{
469    if (cpu != WORK_CPU_UNBOUND)
470        return &per_cpu(gcwq_nr_running, cpu);
471    else
472        return &unbound_gcwq_nr_running;
473}
474
475static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
476                        struct workqueue_struct *wq)
477{
478    if (!(wq->flags & WQ_UNBOUND)) {
479        if (likely(cpu < nr_cpu_ids))
480            return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
481    } else if (likely(cpu == WORK_CPU_UNBOUND))
482        return wq->cpu_wq.single;
483    return NULL;
484}
485
486static unsigned int work_color_to_flags(int color)
487{
488    return color << WORK_STRUCT_COLOR_SHIFT;
489}
490
491static int get_work_color(struct work_struct *work)
492{
493    return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
494        ((1 << WORK_STRUCT_COLOR_BITS) - 1);
495}
496
497static int work_next_color(int color)
498{
499    return (color + 1) % WORK_NR_COLORS;
500}
501
502/*
503 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
504 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
505 * cleared and the work data contains the cpu number it was last on.
506 *
507 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
508 * cwq, cpu or clear work->data. These functions should only be
509 * called while the work is owned - ie. while the PENDING bit is set.
510 *
511 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
512 * corresponding to a work. gcwq is available once the work has been
513 * queued anywhere after initialization. cwq is available only from
514 * queueing until execution starts.
515 */
516static inline void set_work_data(struct work_struct *work, unsigned long data,
517                 unsigned long flags)
518{
519    BUG_ON(!work_pending(work));
520    atomic_long_set(&work->data, data | flags | work_static(work));
521}
522
523static void set_work_cwq(struct work_struct *work,
524             struct cpu_workqueue_struct *cwq,
525             unsigned long extra_flags)
526{
527    set_work_data(work, (unsigned long)cwq,
528              WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
529}
530
531static void set_work_cpu(struct work_struct *work, unsigned int cpu)
532{
533    set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
534}
535
536static void clear_work_data(struct work_struct *work)
537{
538    set_work_data(work, WORK_STRUCT_NO_CPU, 0);
539}
540
541static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
542{
543    unsigned long data = atomic_long_read(&work->data);
544
545    if (data & WORK_STRUCT_CWQ)
546        return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
547    else
548        return NULL;
549}
550
551static struct global_cwq *get_work_gcwq(struct work_struct *work)
552{
553    unsigned long data = atomic_long_read(&work->data);
554    unsigned int cpu;
555
556    if (data & WORK_STRUCT_CWQ)
557        return ((struct cpu_workqueue_struct *)
558            (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
559
560    cpu = data >> WORK_STRUCT_FLAG_BITS;
561    if (cpu == WORK_CPU_NONE)
562        return NULL;
563
564    BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
565    return get_gcwq(cpu);
566}
567
568/*
569 * Policy functions. These define the policies on how the global
570 * worker pool is managed. Unless noted otherwise, these functions
571 * assume that they're being called with gcwq->lock held.
572 */
573
574static bool __need_more_worker(struct global_cwq *gcwq)
575{
576    return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
577        gcwq->flags & GCWQ_HIGHPRI_PENDING;
578}
579
580/*
581 * Need to wake up a worker? Called from anything but currently
582 * running workers.
583 */
584static bool need_more_worker(struct global_cwq *gcwq)
585{
586    return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
587}
588
589/* Can I start working? Called from busy but !running workers. */
590static bool may_start_working(struct global_cwq *gcwq)
591{
592    return gcwq->nr_idle;
593}
594
595/* Do I need to keep working? Called from currently running workers. */
596static bool keep_working(struct global_cwq *gcwq)
597{
598    atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
599
600    return !list_empty(&gcwq->worklist) &&
601        (atomic_read(nr_running) <= 1 ||
602         gcwq->flags & GCWQ_HIGHPRI_PENDING);
603}
604
605/* Do we need a new worker? Called from manager. */
606static bool need_to_create_worker(struct global_cwq *gcwq)
607{
608    return need_more_worker(gcwq) && !may_start_working(gcwq);
609}
610
611/* Do I need to be the manager? */
612static bool need_to_manage_workers(struct global_cwq *gcwq)
613{
614    return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
615}
616
617/* Do we have too many workers and should some go away? */
618static bool too_many_workers(struct global_cwq *gcwq)
619{
620    bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
621    int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
622    int nr_busy = gcwq->nr_workers - nr_idle;
623
624    return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
625}
626
627/*
628 * Wake up functions.
629 */
630
631/* Return the first worker. Safe with preemption disabled */
632static struct worker *first_worker(struct global_cwq *gcwq)
633{
634    if (unlikely(list_empty(&gcwq->idle_list)))
635        return NULL;
636
637    return list_first_entry(&gcwq->idle_list, struct worker, entry);
638}
639
640/**
641 * wake_up_worker - wake up an idle worker
642 * @gcwq: gcwq to wake worker for
643 *
644 * Wake up the first idle worker of @gcwq.
645 *
646 * CONTEXT:
647 * spin_lock_irq(gcwq->lock).
648 */
649static void wake_up_worker(struct global_cwq *gcwq)
650{
651    struct worker *worker = first_worker(gcwq);
652
653    if (likely(worker))
654        wake_up_process(worker->task);
655}
656
657/**
658 * wq_worker_waking_up - a worker is waking up
659 * @task: task waking up
660 * @cpu: CPU @task is waking up to
661 *
662 * This function is called during try_to_wake_up() when a worker is
663 * being awoken.
664 *
665 * CONTEXT:
666 * spin_lock_irq(rq->lock)
667 */
668void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
669{
670    struct worker *worker = kthread_data(task);
671
672    if (!(worker->flags & WORKER_NOT_RUNNING))
673        atomic_inc(get_gcwq_nr_running(cpu));
674}
675
676/**
677 * wq_worker_sleeping - a worker is going to sleep
678 * @task: task going to sleep
679 * @cpu: CPU in question, must be the current CPU number
680 *
681 * This function is called during schedule() when a busy worker is
682 * going to sleep. Worker on the same cpu can be woken up by
683 * returning pointer to its task.
684 *
685 * CONTEXT:
686 * spin_lock_irq(rq->lock)
687 *
688 * RETURNS:
689 * Worker task on @cpu to wake up, %NULL if none.
690 */
691struct task_struct *wq_worker_sleeping(struct task_struct *task,
692                       unsigned int cpu)
693{
694    struct worker *worker = kthread_data(task), *to_wakeup = NULL;
695    struct global_cwq *gcwq = get_gcwq(cpu);
696    atomic_t *nr_running = get_gcwq_nr_running(cpu);
697
698    if (worker->flags & WORKER_NOT_RUNNING)
699        return NULL;
700
701    /* this can only happen on the local cpu */
702    BUG_ON(cpu != raw_smp_processor_id());
703
704    /*
705     * The counterpart of the following dec_and_test, implied mb,
706     * worklist not empty test sequence is in insert_work().
707     * Please read comment there.
708     *
709     * NOT_RUNNING is clear. This means that trustee is not in
710     * charge and we're running on the local cpu w/ rq lock held
711     * and preemption disabled, which in turn means that none else
712     * could be manipulating idle_list, so dereferencing idle_list
713     * without gcwq lock is safe.
714     */
715    if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
716        to_wakeup = first_worker(gcwq);
717    return to_wakeup ? to_wakeup->task : NULL;
718}
719
720/**
721 * worker_set_flags - set worker flags and adjust nr_running accordingly
722 * @worker: self
723 * @flags: flags to set
724 * @wakeup: wakeup an idle worker if necessary
725 *
726 * Set @flags in @worker->flags and adjust nr_running accordingly. If
727 * nr_running becomes zero and @wakeup is %true, an idle worker is
728 * woken up.
729 *
730 * CONTEXT:
731 * spin_lock_irq(gcwq->lock)
732 */
733static inline void worker_set_flags(struct worker *worker, unsigned int flags,
734                    bool wakeup)
735{
736    struct global_cwq *gcwq = worker->gcwq;
737
738    WARN_ON_ONCE(worker->task != current);
739
740    /*
741     * If transitioning into NOT_RUNNING, adjust nr_running and
742     * wake up an idle worker as necessary if requested by
743     * @wakeup.
744     */
745    if ((flags & WORKER_NOT_RUNNING) &&
746        !(worker->flags & WORKER_NOT_RUNNING)) {
747        atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
748
749        if (wakeup) {
750            if (atomic_dec_and_test(nr_running) &&
751                !list_empty(&gcwq->worklist))
752                wake_up_worker(gcwq);
753        } else
754            atomic_dec(nr_running);
755    }
756
757    worker->flags |= flags;
758}
759
760/**
761 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
762 * @worker: self
763 * @flags: flags to clear
764 *
765 * Clear @flags in @worker->flags and adjust nr_running accordingly.
766 *
767 * CONTEXT:
768 * spin_lock_irq(gcwq->lock)
769 */
770static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
771{
772    struct global_cwq *gcwq = worker->gcwq;
773    unsigned int oflags = worker->flags;
774
775    WARN_ON_ONCE(worker->task != current);
776
777    worker->flags &= ~flags;
778
779    /*
780     * If transitioning out of NOT_RUNNING, increment nr_running. Note
781     * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
782     * of multiple flags, not a single flag.
783     */
784    if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
785        if (!(worker->flags & WORKER_NOT_RUNNING))
786            atomic_inc(get_gcwq_nr_running(gcwq->cpu));
787}
788
789/**
790 * busy_worker_head - return the busy hash head for a work
791 * @gcwq: gcwq of interest
792 * @work: work to be hashed
793 *
794 * Return hash head of @gcwq for @work.
795 *
796 * CONTEXT:
797 * spin_lock_irq(gcwq->lock).
798 *
799 * RETURNS:
800 * Pointer to the hash head.
801 */
802static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
803                       struct work_struct *work)
804{
805    const int base_shift = ilog2(sizeof(struct work_struct));
806    unsigned long v = (unsigned long)work;
807
808    /* simple shift and fold hash, do we need something better? */
809    v >>= base_shift;
810    v += v >> BUSY_WORKER_HASH_ORDER;
811    v &= BUSY_WORKER_HASH_MASK;
812
813    return &gcwq->busy_hash[v];
814}
815
816/**
817 * __find_worker_executing_work - find worker which is executing a work
818 * @gcwq: gcwq of interest
819 * @bwh: hash head as returned by busy_worker_head()
820 * @work: work to find worker for
821 *
822 * Find a worker which is executing @work on @gcwq. @bwh should be
823 * the hash head obtained by calling busy_worker_head() with the same
824 * work.
825 *
826 * CONTEXT:
827 * spin_lock_irq(gcwq->lock).
828 *
829 * RETURNS:
830 * Pointer to worker which is executing @work if found, NULL
831 * otherwise.
832 */
833static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
834                           struct hlist_head *bwh,
835                           struct work_struct *work)
836{
837    struct worker *worker;
838    struct hlist_node *tmp;
839
840    hlist_for_each_entry(worker, tmp, bwh, hentry)
841        if (worker->current_work == work)
842            return worker;
843    return NULL;
844}
845
846/**
847 * find_worker_executing_work - find worker which is executing a work
848 * @gcwq: gcwq of interest
849 * @work: work to find worker for
850 *
851 * Find a worker which is executing @work on @gcwq. This function is
852 * identical to __find_worker_executing_work() except that this
853 * function calculates @bwh itself.
854 *
855 * CONTEXT:
856 * spin_lock_irq(gcwq->lock).
857 *
858 * RETURNS:
859 * Pointer to worker which is executing @work if found, NULL
860 * otherwise.
861 */
862static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
863                         struct work_struct *work)
864{
865    return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
866                        work);
867}
868
869/**
870 * gcwq_determine_ins_pos - find insertion position
871 * @gcwq: gcwq of interest
872 * @cwq: cwq a work is being queued for
873 *
874 * A work for @cwq is about to be queued on @gcwq, determine insertion
875 * position for the work. If @cwq is for HIGHPRI wq, the work is
876 * queued at the head of the queue but in FIFO order with respect to
877 * other HIGHPRI works; otherwise, at the end of the queue. This
878 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
879 * there are HIGHPRI works pending.
880 *
881 * CONTEXT:
882 * spin_lock_irq(gcwq->lock).
883 *
884 * RETURNS:
885 * Pointer to inserstion position.
886 */
887static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
888                           struct cpu_workqueue_struct *cwq)
889{
890    struct work_struct *twork;
891
892    if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
893        return &gcwq->worklist;
894
895    list_for_each_entry(twork, &gcwq->worklist, entry) {
896        struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
897
898        if (!(tcwq->wq->flags & WQ_HIGHPRI))
899            break;
900    }
901
902    gcwq->flags |= GCWQ_HIGHPRI_PENDING;
903    return &twork->entry;
904}
905
906/**
907 * insert_work - insert a work into gcwq
908 * @cwq: cwq @work belongs to
909 * @work: work to insert
910 * @head: insertion point
911 * @extra_flags: extra WORK_STRUCT_* flags to set
912 *
913 * Insert @work which belongs to @cwq into @gcwq after @head.
914 * @extra_flags is or'd to work_struct flags.
915 *
916 * CONTEXT:
917 * spin_lock_irq(gcwq->lock).
918 */
919static void insert_work(struct cpu_workqueue_struct *cwq,
920            struct work_struct *work, struct list_head *head,
921            unsigned int extra_flags)
922{
923    struct global_cwq *gcwq = cwq->gcwq;
924
925    /* we own @work, set data and link */
926    set_work_cwq(work, cwq, extra_flags);
927
928    /*
929     * Ensure that we get the right work->data if we see the
930     * result of list_add() below, see try_to_grab_pending().
931     */
932    smp_wmb();
933
934    list_add_tail(&work->entry, head);
935
936    /*
937     * Ensure either worker_sched_deactivated() sees the above
938     * list_add_tail() or we see zero nr_running to avoid workers
939     * lying around lazily while there are works to be processed.
940     */
941    smp_mb();
942
943    if (__need_more_worker(gcwq))
944        wake_up_worker(gcwq);
945}
946
947/*
948 * Test whether @work is being queued from another work executing on the
949 * same workqueue. This is rather expensive and should only be used from
950 * cold paths.
951 */
952static bool is_chained_work(struct workqueue_struct *wq)
953{
954    unsigned long flags;
955    unsigned int cpu;
956
957    for_each_gcwq_cpu(cpu) {
958        struct global_cwq *gcwq = get_gcwq(cpu);
959        struct worker *worker;
960        struct hlist_node *pos;
961        int i;
962
963        spin_lock_irqsave(&gcwq->lock, flags);
964        for_each_busy_worker(worker, i, pos, gcwq) {
965            if (worker->task != current)
966                continue;
967            spin_unlock_irqrestore(&gcwq->lock, flags);
968            /*
969             * I'm @worker, no locking necessary. See if @work
970             * is headed to the same workqueue.
971             */
972            return worker->current_cwq->wq == wq;
973        }
974        spin_unlock_irqrestore(&gcwq->lock, flags);
975    }
976    return false;
977}
978
979static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
980             struct work_struct *work)
981{
982    struct global_cwq *gcwq;
983    struct cpu_workqueue_struct *cwq;
984    struct list_head *worklist;
985    unsigned int work_flags;
986    unsigned long flags;
987
988    debug_work_activate(work);
989
990    /* if dying, only works from the same workqueue are allowed */
991    if (unlikely(wq->flags & WQ_DRAINING) &&
992        WARN_ON_ONCE(!is_chained_work(wq)))
993        return;
994
995    /* determine gcwq to use */
996    if (!(wq->flags & WQ_UNBOUND)) {
997        struct global_cwq *last_gcwq;
998
999        if (unlikely(cpu == WORK_CPU_UNBOUND))
1000            cpu = raw_smp_processor_id();
1001
1002        /*
1003         * It's multi cpu. If @wq is non-reentrant and @work
1004         * was previously on a different cpu, it might still
1005         * be running there, in which case the work needs to
1006         * be queued on that cpu to guarantee non-reentrance.
1007         */
1008        gcwq = get_gcwq(cpu);
1009        if (wq->flags & WQ_NON_REENTRANT &&
1010            (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1011            struct worker *worker;
1012
1013            spin_lock_irqsave(&last_gcwq->lock, flags);
1014
1015            worker = find_worker_executing_work(last_gcwq, work);
1016
1017            if (worker && worker->current_cwq->wq == wq)
1018                gcwq = last_gcwq;
1019            else {
1020                /* meh... not running there, queue here */
1021                spin_unlock_irqrestore(&last_gcwq->lock, flags);
1022                spin_lock_irqsave(&gcwq->lock, flags);
1023            }
1024        } else
1025            spin_lock_irqsave(&gcwq->lock, flags);
1026    } else {
1027        gcwq = get_gcwq(WORK_CPU_UNBOUND);
1028        spin_lock_irqsave(&gcwq->lock, flags);
1029    }
1030
1031    /* gcwq determined, get cwq and queue */
1032    cwq = get_cwq(gcwq->cpu, wq);
1033    trace_workqueue_queue_work(cpu, cwq, work);
1034
1035    if (WARN_ON(!list_empty(&work->entry))) {
1036        spin_unlock_irqrestore(&gcwq->lock, flags);
1037        return;
1038    }
1039
1040    cwq->nr_in_flight[cwq->work_color]++;
1041    work_flags = work_color_to_flags(cwq->work_color);
1042
1043    if (likely(cwq->nr_active < cwq->max_active)) {
1044        trace_workqueue_activate_work(work);
1045        cwq->nr_active++;
1046        worklist = gcwq_determine_ins_pos(gcwq, cwq);
1047    } else {
1048        work_flags |= WORK_STRUCT_DELAYED;
1049        worklist = &cwq->delayed_works;
1050    }
1051
1052    insert_work(cwq, work, worklist, work_flags);
1053
1054    spin_unlock_irqrestore(&gcwq->lock, flags);
1055}
1056
1057/**
1058 * queue_work - queue work on a workqueue
1059 * @wq: workqueue to use
1060 * @work: work to queue
1061 *
1062 * Returns 0 if @work was already on a queue, non-zero otherwise.
1063 *
1064 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1065 * it can be processed by another CPU.
1066 */
1067int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1068{
1069    int ret;
1070
1071    ret = queue_work_on(get_cpu(), wq, work);
1072    put_cpu();
1073
1074    return ret;
1075}
1076EXPORT_SYMBOL_GPL(queue_work);
1077
1078/**
1079 * queue_work_on - queue work on specific cpu
1080 * @cpu: CPU number to execute work on
1081 * @wq: workqueue to use
1082 * @work: work to queue
1083 *
1084 * Returns 0 if @work was already on a queue, non-zero otherwise.
1085 *
1086 * We queue the work to a specific CPU, the caller must ensure it
1087 * can't go away.
1088 */
1089int
1090queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1091{
1092    int ret = 0;
1093
1094    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1095        __queue_work(cpu, wq, work);
1096        ret = 1;
1097    }
1098    return ret;
1099}
1100EXPORT_SYMBOL_GPL(queue_work_on);
1101
1102static void delayed_work_timer_fn(unsigned long __data)
1103{
1104    struct delayed_work *dwork = (struct delayed_work *)__data;
1105    struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1106
1107    __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
1108}
1109
1110/**
1111 * queue_delayed_work - queue work on a workqueue after delay
1112 * @wq: workqueue to use
1113 * @dwork: delayable work to queue
1114 * @delay: number of jiffies to wait before queueing
1115 *
1116 * Returns 0 if @work was already on a queue, non-zero otherwise.
1117 */
1118int queue_delayed_work(struct workqueue_struct *wq,
1119            struct delayed_work *dwork, unsigned long delay)
1120{
1121    if (delay == 0)
1122        return queue_work(wq, &dwork->work);
1123
1124    return queue_delayed_work_on(-1, wq, dwork, delay);
1125}
1126EXPORT_SYMBOL_GPL(queue_delayed_work);
1127
1128/**
1129 * queue_delayed_work_on - queue work on specific CPU after delay
1130 * @cpu: CPU number to execute work on
1131 * @wq: workqueue to use
1132 * @dwork: work to queue
1133 * @delay: number of jiffies to wait before queueing
1134 *
1135 * Returns 0 if @work was already on a queue, non-zero otherwise.
1136 */
1137int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1138            struct delayed_work *dwork, unsigned long delay)
1139{
1140    int ret = 0;
1141    struct timer_list *timer = &dwork->timer;
1142    struct work_struct *work = &dwork->work;
1143
1144    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1145        unsigned int lcpu;
1146
1147        BUG_ON(timer_pending(timer));
1148        BUG_ON(!list_empty(&work->entry));
1149
1150        timer_stats_timer_set_start_info(&dwork->timer);
1151
1152        /*
1153         * This stores cwq for the moment, for the timer_fn.
1154         * Note that the work's gcwq is preserved to allow
1155         * reentrance detection for delayed works.
1156         */
1157        if (!(wq->flags & WQ_UNBOUND)) {
1158            struct global_cwq *gcwq = get_work_gcwq(work);
1159
1160            if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1161                lcpu = gcwq->cpu;
1162            else
1163                lcpu = raw_smp_processor_id();
1164        } else
1165            lcpu = WORK_CPU_UNBOUND;
1166
1167        set_work_cwq(work, get_cwq(lcpu, wq), 0);
1168
1169        timer->expires = jiffies + delay;
1170        timer->data = (unsigned long)dwork;
1171        timer->function = delayed_work_timer_fn;
1172
1173        if (unlikely(cpu >= 0))
1174            add_timer_on(timer, cpu);
1175        else
1176            add_timer(timer);
1177        ret = 1;
1178    }
1179    return ret;
1180}
1181EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1182
1183/**
1184 * worker_enter_idle - enter idle state
1185 * @worker: worker which is entering idle state
1186 *
1187 * @worker is entering idle state. Update stats and idle timer if
1188 * necessary.
1189 *
1190 * LOCKING:
1191 * spin_lock_irq(gcwq->lock).
1192 */
1193static void worker_enter_idle(struct worker *worker)
1194{
1195    struct global_cwq *gcwq = worker->gcwq;
1196
1197    BUG_ON(worker->flags & WORKER_IDLE);
1198    BUG_ON(!list_empty(&worker->entry) &&
1199           (worker->hentry.next || worker->hentry.pprev));
1200
1201    /* can't use worker_set_flags(), also called from start_worker() */
1202    worker->flags |= WORKER_IDLE;
1203    gcwq->nr_idle++;
1204    worker->last_active = jiffies;
1205
1206    /* idle_list is LIFO */
1207    list_add(&worker->entry, &gcwq->idle_list);
1208
1209    if (likely(!(worker->flags & WORKER_ROGUE))) {
1210        if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1211            mod_timer(&gcwq->idle_timer,
1212                  jiffies + IDLE_WORKER_TIMEOUT);
1213    } else
1214        wake_up_all(&gcwq->trustee_wait);
1215
1216    /*
1217     * Sanity check nr_running. Because trustee releases gcwq->lock
1218     * between setting %WORKER_ROGUE and zapping nr_running, the
1219     * warning may trigger spuriously. Check iff trustee is idle.
1220     */
1221    WARN_ON_ONCE(gcwq->trustee_state == TRUSTEE_DONE &&
1222             gcwq->nr_workers == gcwq->nr_idle &&
1223             atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1224}
1225
1226/**
1227 * worker_leave_idle - leave idle state
1228 * @worker: worker which is leaving idle state
1229 *
1230 * @worker is leaving idle state. Update stats.
1231 *
1232 * LOCKING:
1233 * spin_lock_irq(gcwq->lock).
1234 */
1235static void worker_leave_idle(struct worker *worker)
1236{
1237    struct global_cwq *gcwq = worker->gcwq;
1238
1239    BUG_ON(!(worker->flags & WORKER_IDLE));
1240    worker_clr_flags(worker, WORKER_IDLE);
1241    gcwq->nr_idle--;
1242    list_del_init(&worker->entry);
1243}
1244
1245/**
1246 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1247 * @worker: self
1248 *
1249 * Works which are scheduled while the cpu is online must at least be
1250 * scheduled to a worker which is bound to the cpu so that if they are
1251 * flushed from cpu callbacks while cpu is going down, they are
1252 * guaranteed to execute on the cpu.
1253 *
1254 * This function is to be used by rogue workers and rescuers to bind
1255 * themselves to the target cpu and may race with cpu going down or
1256 * coming online. kthread_bind() can't be used because it may put the
1257 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1258 * verbatim as it's best effort and blocking and gcwq may be
1259 * [dis]associated in the meantime.
1260 *
1261 * This function tries set_cpus_allowed() and locks gcwq and verifies
1262 * the binding against GCWQ_DISASSOCIATED which is set during
1263 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1264 * idle state or fetches works without dropping lock, it can guarantee
1265 * the scheduling requirement described in the first paragraph.
1266 *
1267 * CONTEXT:
1268 * Might sleep. Called without any lock but returns with gcwq->lock
1269 * held.
1270 *
1271 * RETURNS:
1272 * %true if the associated gcwq is online (@worker is successfully
1273 * bound), %false if offline.
1274 */
1275static bool worker_maybe_bind_and_lock(struct worker *worker)
1276__acquires(&gcwq->lock)
1277{
1278    struct global_cwq *gcwq = worker->gcwq;
1279    struct task_struct *task = worker->task;
1280
1281    while (true) {
1282        /*
1283         * The following call may fail, succeed or succeed
1284         * without actually migrating the task to the cpu if
1285         * it races with cpu hotunplug operation. Verify
1286         * against GCWQ_DISASSOCIATED.
1287         */
1288        if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1289            set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1290
1291        spin_lock_irq(&gcwq->lock);
1292        if (gcwq->flags & GCWQ_DISASSOCIATED)
1293            return false;
1294        if (task_cpu(task) == gcwq->cpu &&
1295            cpumask_equal(&current->cpus_allowed,
1296                  get_cpu_mask(gcwq->cpu)))
1297            return true;
1298        spin_unlock_irq(&gcwq->lock);
1299
1300        /*
1301         * We've raced with CPU hot[un]plug. Give it a breather
1302         * and retry migration. cond_resched() is required here;
1303         * otherwise, we might deadlock against cpu_stop trying to
1304         * bring down the CPU on non-preemptive kernel.
1305         */
1306        cpu_relax();
1307        cond_resched();
1308    }
1309}
1310
1311/*
1312 * Function for worker->rebind_work used to rebind rogue busy workers
1313 * to the associated cpu which is coming back online. This is
1314 * scheduled by cpu up but can race with other cpu hotplug operations
1315 * and may be executed twice without intervening cpu down.
1316 */
1317static void worker_rebind_fn(struct work_struct *work)
1318{
1319    struct worker *worker = container_of(work, struct worker, rebind_work);
1320    struct global_cwq *gcwq = worker->gcwq;
1321
1322    if (worker_maybe_bind_and_lock(worker))
1323        worker_clr_flags(worker, WORKER_REBIND);
1324
1325    spin_unlock_irq(&gcwq->lock);
1326}
1327
1328static struct worker *alloc_worker(void)
1329{
1330    struct worker *worker;
1331
1332    worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1333    if (worker) {
1334        INIT_LIST_HEAD(&worker->entry);
1335        INIT_LIST_HEAD(&worker->scheduled);
1336        INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1337        /* on creation a worker is in !idle && prep state */
1338        worker->flags = WORKER_PREP;
1339    }
1340    return worker;
1341}
1342
1343/**
1344 * create_worker - create a new workqueue worker
1345 * @gcwq: gcwq the new worker will belong to
1346 * @bind: whether to set affinity to @cpu or not
1347 *
1348 * Create a new worker which is bound to @gcwq. The returned worker
1349 * can be started by calling start_worker() or destroyed using
1350 * destroy_worker().
1351 *
1352 * CONTEXT:
1353 * Might sleep. Does GFP_KERNEL allocations.
1354 *
1355 * RETURNS:
1356 * Pointer to the newly created worker.
1357 */
1358static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1359{
1360    bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1361    struct worker *worker = NULL;
1362    int id = -1;
1363
1364    spin_lock_irq(&gcwq->lock);
1365    while (ida_get_new(&gcwq->worker_ida, &id)) {
1366        spin_unlock_irq(&gcwq->lock);
1367        if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1368            goto fail;
1369        spin_lock_irq(&gcwq->lock);
1370    }
1371    spin_unlock_irq(&gcwq->lock);
1372
1373    worker = alloc_worker();
1374    if (!worker)
1375        goto fail;
1376
1377    worker->gcwq = gcwq;
1378    worker->id = id;
1379
1380    if (!on_unbound_cpu)
1381        worker->task = kthread_create_on_node(worker_thread,
1382                              worker,
1383                              cpu_to_node(gcwq->cpu),
1384                              "kworker/%u:%d", gcwq->cpu, id);
1385    else
1386        worker->task = kthread_create(worker_thread, worker,
1387                          "kworker/u:%d", id);
1388    if (IS_ERR(worker->task))
1389        goto fail;
1390
1391    /*
1392     * A rogue worker will become a regular one if CPU comes
1393     * online later on. Make sure every worker has
1394     * PF_THREAD_BOUND set.
1395     */
1396    if (bind && !on_unbound_cpu)
1397        kthread_bind(worker->task, gcwq->cpu);
1398    else {
1399        worker->task->flags |= PF_THREAD_BOUND;
1400        if (on_unbound_cpu)
1401            worker->flags |= WORKER_UNBOUND;
1402    }
1403
1404    return worker;
1405fail:
1406    if (id >= 0) {
1407        spin_lock_irq(&gcwq->lock);
1408        ida_remove(&gcwq->worker_ida, id);
1409        spin_unlock_irq(&gcwq->lock);
1410    }
1411    kfree(worker);
1412    return NULL;
1413}
1414
1415/**
1416 * start_worker - start a newly created worker
1417 * @worker: worker to start
1418 *
1419 * Make the gcwq aware of @worker and start it.
1420 *
1421 * CONTEXT:
1422 * spin_lock_irq(gcwq->lock).
1423 */
1424static void start_worker(struct worker *worker)
1425{
1426    worker->flags |= WORKER_STARTED;
1427    worker->gcwq->nr_workers++;
1428    worker_enter_idle(worker);
1429    wake_up_process(worker->task);
1430}
1431
1432/**
1433 * destroy_worker - destroy a workqueue worker
1434 * @worker: worker to be destroyed
1435 *
1436 * Destroy @worker and adjust @gcwq stats accordingly.
1437 *
1438 * CONTEXT:
1439 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1440 */
1441static void destroy_worker(struct worker *worker)
1442{
1443    struct global_cwq *gcwq = worker->gcwq;
1444    int id = worker->id;
1445
1446    /* sanity check frenzy */
1447    BUG_ON(worker->current_work);
1448    BUG_ON(!list_empty(&worker->scheduled));
1449
1450    if (worker->flags & WORKER_STARTED)
1451        gcwq->nr_workers--;
1452    if (worker->flags & WORKER_IDLE)
1453        gcwq->nr_idle--;
1454
1455    list_del_init(&worker->entry);
1456    worker->flags |= WORKER_DIE;
1457
1458    spin_unlock_irq(&gcwq->lock);
1459
1460    kthread_stop(worker->task);
1461    kfree(worker);
1462
1463    spin_lock_irq(&gcwq->lock);
1464    ida_remove(&gcwq->worker_ida, id);
1465}
1466
1467static void idle_worker_timeout(unsigned long __gcwq)
1468{
1469    struct global_cwq *gcwq = (void *)__gcwq;
1470
1471    spin_lock_irq(&gcwq->lock);
1472
1473    if (too_many_workers(gcwq)) {
1474        struct worker *worker;
1475        unsigned long expires;
1476
1477        /* idle_list is kept in LIFO order, check the last one */
1478        worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1479        expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1480
1481        if (time_before(jiffies, expires))
1482            mod_timer(&gcwq->idle_timer, expires);
1483        else {
1484            /* it's been idle for too long, wake up manager */
1485            gcwq->flags |= GCWQ_MANAGE_WORKERS;
1486            wake_up_worker(gcwq);
1487        }
1488    }
1489
1490    spin_unlock_irq(&gcwq->lock);
1491}
1492
1493static bool send_mayday(struct work_struct *work)
1494{
1495    struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1496    struct workqueue_struct *wq = cwq->wq;
1497    unsigned int cpu;
1498
1499    if (!(wq->flags & WQ_RESCUER))
1500        return false;
1501
1502    /* mayday mayday mayday */
1503    cpu = cwq->gcwq->cpu;
1504    /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1505    if (cpu == WORK_CPU_UNBOUND)
1506        cpu = 0;
1507    if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1508        wake_up_process(wq->rescuer->task);
1509    return true;
1510}
1511
1512static void gcwq_mayday_timeout(unsigned long __gcwq)
1513{
1514    struct global_cwq *gcwq = (void *)__gcwq;
1515    struct work_struct *work;
1516
1517    spin_lock_irq(&gcwq->lock);
1518
1519    if (need_to_create_worker(gcwq)) {
1520        /*
1521         * We've been trying to create a new worker but
1522         * haven't been successful. We might be hitting an
1523         * allocation deadlock. Send distress signals to
1524         * rescuers.
1525         */
1526        list_for_each_entry(work, &gcwq->worklist, entry)
1527            send_mayday(work);
1528    }
1529
1530    spin_unlock_irq(&gcwq->lock);
1531
1532    mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1533}
1534
1535/**
1536 * maybe_create_worker - create a new worker if necessary
1537 * @gcwq: gcwq to create a new worker for
1538 *
1539 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1540 * have at least one idle worker on return from this function. If
1541 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1542 * sent to all rescuers with works scheduled on @gcwq to resolve
1543 * possible allocation deadlock.
1544 *
1545 * On return, need_to_create_worker() is guaranteed to be false and
1546 * may_start_working() true.
1547 *
1548 * LOCKING:
1549 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1550 * multiple times. Does GFP_KERNEL allocations. Called only from
1551 * manager.
1552 *
1553 * RETURNS:
1554 * false if no action was taken and gcwq->lock stayed locked, true
1555 * otherwise.
1556 */
1557static bool maybe_create_worker(struct global_cwq *gcwq)
1558__releases(&gcwq->lock)
1559__acquires(&gcwq->lock)
1560{
1561    if (!need_to_create_worker(gcwq))
1562        return false;
1563restart:
1564    spin_unlock_irq(&gcwq->lock);
1565
1566    /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1567    mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1568
1569    while (true) {
1570        struct worker *worker;
1571
1572        worker = create_worker(gcwq, true);
1573        if (worker) {
1574            del_timer_sync(&gcwq->mayday_timer);
1575            spin_lock_irq(&gcwq->lock);
1576            start_worker(worker);
1577            BUG_ON(need_to_create_worker(gcwq));
1578            return true;
1579        }
1580
1581        if (!need_to_create_worker(gcwq))
1582            break;
1583
1584        __set_current_state(TASK_INTERRUPTIBLE);
1585        schedule_timeout(CREATE_COOLDOWN);
1586
1587        if (!need_to_create_worker(gcwq))
1588            break;
1589    }
1590
1591    del_timer_sync(&gcwq->mayday_timer);
1592    spin_lock_irq(&gcwq->lock);
1593    if (need_to_create_worker(gcwq))
1594        goto restart;
1595    return true;
1596}
1597
1598/**
1599 * maybe_destroy_worker - destroy workers which have been idle for a while
1600 * @gcwq: gcwq to destroy workers for
1601 *
1602 * Destroy @gcwq workers which have been idle for longer than
1603 * IDLE_WORKER_TIMEOUT.
1604 *
1605 * LOCKING:
1606 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1607 * multiple times. Called only from manager.
1608 *
1609 * RETURNS:
1610 * false if no action was taken and gcwq->lock stayed locked, true
1611 * otherwise.
1612 */
1613static bool maybe_destroy_workers(struct global_cwq *gcwq)
1614{
1615    bool ret = false;
1616
1617    while (too_many_workers(gcwq)) {
1618        struct worker *worker;
1619        unsigned long expires;
1620
1621        worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1622        expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1623
1624        if (time_before(jiffies, expires)) {
1625            mod_timer(&gcwq->idle_timer, expires);
1626            break;
1627        }
1628
1629        destroy_worker(worker);
1630        ret = true;
1631    }
1632
1633    return ret;
1634}
1635
1636/**
1637 * manage_workers - manage worker pool
1638 * @worker: self
1639 *
1640 * Assume the manager role and manage gcwq worker pool @worker belongs
1641 * to. At any given time, there can be only zero or one manager per
1642 * gcwq. The exclusion is handled automatically by this function.
1643 *
1644 * The caller can safely start processing works on false return. On
1645 * true return, it's guaranteed that need_to_create_worker() is false
1646 * and may_start_working() is true.
1647 *
1648 * CONTEXT:
1649 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1650 * multiple times. Does GFP_KERNEL allocations.
1651 *
1652 * RETURNS:
1653 * false if no action was taken and gcwq->lock stayed locked, true if
1654 * some action was taken.
1655 */
1656static bool manage_workers(struct worker *worker)
1657{
1658    struct global_cwq *gcwq = worker->gcwq;
1659    bool ret = false;
1660
1661    if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1662        return ret;
1663
1664    gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1665    gcwq->flags |= GCWQ_MANAGING_WORKERS;
1666
1667    /*
1668     * Destroy and then create so that may_start_working() is true
1669     * on return.
1670     */
1671    ret |= maybe_destroy_workers(gcwq);
1672    ret |= maybe_create_worker(gcwq);
1673
1674    gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1675
1676    /*
1677     * The trustee might be waiting to take over the manager
1678     * position, tell it we're done.
1679     */
1680    if (unlikely(gcwq->trustee))
1681        wake_up_all(&gcwq->trustee_wait);
1682
1683    return ret;
1684}
1685
1686/**
1687 * move_linked_works - move linked works to a list
1688 * @work: start of series of works to be scheduled
1689 * @head: target list to append @work to
1690 * @nextp: out paramter for nested worklist walking
1691 *
1692 * Schedule linked works starting from @work to @head. Work series to
1693 * be scheduled starts at @work and includes any consecutive work with
1694 * WORK_STRUCT_LINKED set in its predecessor.
1695 *
1696 * If @nextp is not NULL, it's updated to point to the next work of
1697 * the last scheduled work. This allows move_linked_works() to be
1698 * nested inside outer list_for_each_entry_safe().
1699 *
1700 * CONTEXT:
1701 * spin_lock_irq(gcwq->lock).
1702 */
1703static void move_linked_works(struct work_struct *work, struct list_head *head,
1704                  struct work_struct **nextp)
1705{
1706    struct work_struct *n;
1707
1708    /*
1709     * Linked worklist will always end before the end of the list,
1710     * use NULL for list head.
1711     */
1712    list_for_each_entry_safe_from(work, n, NULL, entry) {
1713        list_move_tail(&work->entry, head);
1714        if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1715            break;
1716    }
1717
1718    /*
1719     * If we're already inside safe list traversal and have moved
1720     * multiple works to the scheduled queue, the next position
1721     * needs to be updated.
1722     */
1723    if (nextp)
1724        *nextp = n;
1725}
1726
1727static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1728{
1729    struct work_struct *work = list_first_entry(&cwq->delayed_works,
1730                            struct work_struct, entry);
1731    struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1732
1733    trace_workqueue_activate_work(work);
1734    move_linked_works(work, pos, NULL);
1735    __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1736    cwq->nr_active++;
1737}
1738
1739/**
1740 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1741 * @cwq: cwq of interest
1742 * @color: color of work which left the queue
1743 * @delayed: for a delayed work
1744 *
1745 * A work either has completed or is removed from pending queue,
1746 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1747 *
1748 * CONTEXT:
1749 * spin_lock_irq(gcwq->lock).
1750 */
1751static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1752                 bool delayed)
1753{
1754    /* ignore uncolored works */
1755    if (color == WORK_NO_COLOR)
1756        return;
1757
1758    cwq->nr_in_flight[color]--;
1759
1760    if (!delayed) {
1761        cwq->nr_active--;
1762        if (!list_empty(&cwq->delayed_works)) {
1763            /* one down, submit a delayed one */
1764            if (cwq->nr_active < cwq->max_active)
1765                cwq_activate_first_delayed(cwq);
1766        }
1767    }
1768
1769    /* is flush in progress and are we at the flushing tip? */
1770    if (likely(cwq->flush_color != color))
1771        return;
1772
1773    /* are there still in-flight works? */
1774    if (cwq->nr_in_flight[color])
1775        return;
1776
1777    /* this cwq is done, clear flush_color */
1778    cwq->flush_color = -1;
1779
1780    /*
1781     * If this was the last cwq, wake up the first flusher. It
1782     * will handle the rest.
1783     */
1784    if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1785        complete(&cwq->wq->first_flusher->done);
1786}
1787
1788/**
1789 * process_one_work - process single work
1790 * @worker: self
1791 * @work: work to process
1792 *
1793 * Process @work. This function contains all the logics necessary to
1794 * process a single work including synchronization against and
1795 * interaction with other workers on the same cpu, queueing and
1796 * flushing. As long as context requirement is met, any worker can
1797 * call this function to process a work.
1798 *
1799 * CONTEXT:
1800 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1801 */
1802static void process_one_work(struct worker *worker, struct work_struct *work)
1803__releases(&gcwq->lock)
1804__acquires(&gcwq->lock)
1805{
1806    struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1807    struct global_cwq *gcwq = cwq->gcwq;
1808    struct hlist_head *bwh = busy_worker_head(gcwq, work);
1809    bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1810    work_func_t f = work->func;
1811    int work_color;
1812    struct worker *collision;
1813#ifdef CONFIG_LOCKDEP
1814    /*
1815     * It is permissible to free the struct work_struct from
1816     * inside the function that is called from it, this we need to
1817     * take into account for lockdep too. To avoid bogus "held
1818     * lock freed" warnings as well as problems when looking into
1819     * work->lockdep_map, make a copy and use that here.
1820     */
1821    struct lockdep_map lockdep_map;
1822
1823    lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1824#endif
1825    /*
1826     * A single work shouldn't be executed concurrently by
1827     * multiple workers on a single cpu. Check whether anyone is
1828     * already processing the work. If so, defer the work to the
1829     * currently executing one.
1830     */
1831    collision = __find_worker_executing_work(gcwq, bwh, work);
1832    if (unlikely(collision)) {
1833        move_linked_works(work, &collision->scheduled, NULL);
1834        return;
1835    }
1836
1837    /* claim and process */
1838    debug_work_deactivate(work);
1839    hlist_add_head(&worker->hentry, bwh);
1840    worker->current_work = work;
1841    worker->current_cwq = cwq;
1842    work_color = get_work_color(work);
1843
1844    /* record the current cpu number in the work data and dequeue */
1845    set_work_cpu(work, gcwq->cpu);
1846    list_del_init(&work->entry);
1847
1848    /*
1849     * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1850     * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1851     */
1852    if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1853        struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1854                        struct work_struct, entry);
1855
1856        if (!list_empty(&gcwq->worklist) &&
1857            get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1858            wake_up_worker(gcwq);
1859        else
1860            gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1861    }
1862
1863    /*
1864     * CPU intensive works don't participate in concurrency
1865     * management. They're the scheduler's responsibility.
1866     */
1867    if (unlikely(cpu_intensive))
1868        worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1869
1870    spin_unlock_irq(&gcwq->lock);
1871
1872    work_clear_pending(work);
1873    lock_map_acquire_read(&cwq->wq->lockdep_map);
1874    lock_map_acquire(&lockdep_map);
1875    trace_workqueue_execute_start(work);
1876    f(work);
1877    /*
1878     * While we must be careful to not use "work" after this, the trace
1879     * point will only record its address.
1880     */
1881    trace_workqueue_execute_end(work);
1882    lock_map_release(&lockdep_map);
1883    lock_map_release(&cwq->wq->lockdep_map);
1884
1885    if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1886        printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1887               "%s/0x%08x/%d\n",
1888               current->comm, preempt_count(), task_pid_nr(current));
1889        printk(KERN_ERR " last function: ");
1890        print_symbol("%s\n", (unsigned long)f);
1891        debug_show_held_locks(current);
1892        dump_stack();
1893    }
1894
1895    spin_lock_irq(&gcwq->lock);
1896
1897    /* clear cpu intensive status */
1898    if (unlikely(cpu_intensive))
1899        worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1900
1901    /* we're done with it, release */
1902    hlist_del_init(&worker->hentry);
1903    worker->current_work = NULL;
1904    worker->current_cwq = NULL;
1905    cwq_dec_nr_in_flight(cwq, work_color, false);
1906}
1907
1908/**
1909 * process_scheduled_works - process scheduled works
1910 * @worker: self
1911 *
1912 * Process all scheduled works. Please note that the scheduled list
1913 * may change while processing a work, so this function repeatedly
1914 * fetches a work from the top and executes it.
1915 *
1916 * CONTEXT:
1917 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1918 * multiple times.
1919 */
1920static void process_scheduled_works(struct worker *worker)
1921{
1922    while (!list_empty(&worker->scheduled)) {
1923        struct work_struct *work = list_first_entry(&worker->scheduled,
1924                        struct work_struct, entry);
1925        process_one_work(worker, work);
1926    }
1927}
1928
1929/**
1930 * worker_thread - the worker thread function
1931 * @__worker: self
1932 *
1933 * The gcwq worker thread function. There's a single dynamic pool of
1934 * these per each cpu. These workers process all works regardless of
1935 * their specific target workqueue. The only exception is works which
1936 * belong to workqueues with a rescuer which will be explained in
1937 * rescuer_thread().
1938 */
1939static int worker_thread(void *__worker)
1940{
1941    struct worker *worker = __worker;
1942    struct global_cwq *gcwq = worker->gcwq;
1943
1944    /* tell the scheduler that this is a workqueue worker */
1945    worker->task->flags |= PF_WQ_WORKER;
1946woke_up:
1947    spin_lock_irq(&gcwq->lock);
1948
1949    /* DIE can be set only while we're idle, checking here is enough */
1950    if (worker->flags & WORKER_DIE) {
1951        spin_unlock_irq(&gcwq->lock);
1952        worker->task->flags &= ~PF_WQ_WORKER;
1953        return 0;
1954    }
1955
1956    worker_leave_idle(worker);
1957recheck:
1958    /* no more worker necessary? */
1959    if (!need_more_worker(gcwq))
1960        goto sleep;
1961
1962    /* do we need to manage? */
1963    if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1964        goto recheck;
1965
1966    /*
1967     * ->scheduled list can only be filled while a worker is
1968     * preparing to process a work or actually processing it.
1969     * Make sure nobody diddled with it while I was sleeping.
1970     */
1971    BUG_ON(!list_empty(&worker->scheduled));
1972
1973    /*
1974     * When control reaches this point, we're guaranteed to have
1975     * at least one idle worker or that someone else has already
1976     * assumed the manager role.
1977     */
1978    worker_clr_flags(worker, WORKER_PREP);
1979
1980    do {
1981        struct work_struct *work =
1982            list_first_entry(&gcwq->worklist,
1983                     struct work_struct, entry);
1984
1985        if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1986            /* optimization path, not strictly necessary */
1987            process_one_work(worker, work);
1988            if (unlikely(!list_empty(&worker->scheduled)))
1989                process_scheduled_works(worker);
1990        } else {
1991            move_linked_works(work, &worker->scheduled, NULL);
1992            process_scheduled_works(worker);
1993        }
1994    } while (keep_working(gcwq));
1995
1996    worker_set_flags(worker, WORKER_PREP, false);
1997sleep:
1998    if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1999        goto recheck;
2000
2001    /*
2002     * gcwq->lock is held and there's no work to process and no
2003     * need to manage, sleep. Workers are woken up only while
2004     * holding gcwq->lock or from local cpu, so setting the
2005     * current state before releasing gcwq->lock is enough to
2006     * prevent losing any event.
2007     */
2008    worker_enter_idle(worker);
2009    __set_current_state(TASK_INTERRUPTIBLE);
2010    spin_unlock_irq(&gcwq->lock);
2011    schedule();
2012    goto woke_up;
2013}
2014
2015/**
2016 * rescuer_thread - the rescuer thread function
2017 * @__wq: the associated workqueue
2018 *
2019 * Workqueue rescuer thread function. There's one rescuer for each
2020 * workqueue which has WQ_RESCUER set.
2021 *
2022 * Regular work processing on a gcwq may block trying to create a new
2023 * worker which uses GFP_KERNEL allocation which has slight chance of
2024 * developing into deadlock if some works currently on the same queue
2025 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2026 * the problem rescuer solves.
2027 *
2028 * When such condition is possible, the gcwq summons rescuers of all
2029 * workqueues which have works queued on the gcwq and let them process
2030 * those works so that forward progress can be guaranteed.
2031 *
2032 * This should happen rarely.
2033 */
2034static int rescuer_thread(void *__wq)
2035{
2036    struct workqueue_struct *wq = __wq;
2037    struct worker *rescuer = wq->rescuer;
2038    struct list_head *scheduled = &rescuer->scheduled;
2039    bool is_unbound = wq->flags & WQ_UNBOUND;
2040    unsigned int cpu;
2041
2042    set_user_nice(current, RESCUER_NICE_LEVEL);
2043repeat:
2044    set_current_state(TASK_INTERRUPTIBLE);
2045
2046    if (kthread_should_stop())
2047        return 0;
2048
2049    /*
2050     * See whether any cpu is asking for help. Unbounded
2051     * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2052     */
2053    for_each_mayday_cpu(cpu, wq->mayday_mask) {
2054        unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2055        struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2056        struct global_cwq *gcwq = cwq->gcwq;
2057        struct work_struct *work, *n;
2058
2059        __set_current_state(TASK_RUNNING);
2060        mayday_clear_cpu(cpu, wq->mayday_mask);
2061
2062        /* migrate to the target cpu if possible */
2063        rescuer->gcwq = gcwq;
2064        worker_maybe_bind_and_lock(rescuer);
2065
2066        /*
2067         * Slurp in all works issued via this workqueue and
2068         * process'em.
2069         */
2070        BUG_ON(!list_empty(&rescuer->scheduled));
2071        list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2072            if (get_work_cwq(work) == cwq)
2073                move_linked_works(work, scheduled, &n);
2074
2075        process_scheduled_works(rescuer);
2076
2077        /*
2078         * Leave this gcwq. If keep_working() is %true, notify a
2079         * regular worker; otherwise, we end up with 0 concurrency
2080         * and stalling the execution.
2081         */
2082        if (keep_working(gcwq))
2083            wake_up_worker(gcwq);
2084
2085        spin_unlock_irq(&gcwq->lock);
2086    }
2087
2088    schedule();
2089    goto repeat;
2090}
2091
2092struct wq_barrier {
2093    struct work_struct work;
2094    struct completion done;
2095};
2096
2097static void wq_barrier_func(struct work_struct *work)
2098{
2099    struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2100    complete(&barr->done);
2101}
2102
2103/**
2104 * insert_wq_barrier - insert a barrier work
2105 * @cwq: cwq to insert barrier into
2106 * @barr: wq_barrier to insert
2107 * @target: target work to attach @barr to
2108 * @worker: worker currently executing @target, NULL if @target is not executing
2109 *
2110 * @barr is linked to @target such that @barr is completed only after
2111 * @target finishes execution. Please note that the ordering
2112 * guarantee is observed only with respect to @target and on the local
2113 * cpu.
2114 *
2115 * Currently, a queued barrier can't be canceled. This is because
2116 * try_to_grab_pending() can't determine whether the work to be
2117 * grabbed is at the head of the queue and thus can't clear LINKED
2118 * flag of the previous work while there must be a valid next work
2119 * after a work with LINKED flag set.
2120 *
2121 * Note that when @worker is non-NULL, @target may be modified
2122 * underneath us, so we can't reliably determine cwq from @target.
2123 *
2124 * CONTEXT:
2125 * spin_lock_irq(gcwq->lock).
2126 */
2127static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2128                  struct wq_barrier *barr,
2129                  struct work_struct *target, struct worker *worker)
2130{
2131    struct list_head *head;
2132    unsigned int linked = 0;
2133
2134    /*
2135     * debugobject calls are safe here even with gcwq->lock locked
2136     * as we know for sure that this will not trigger any of the
2137     * checks and call back into the fixup functions where we
2138     * might deadlock.
2139     */
2140    INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2141    __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2142    init_completion(&barr->done);
2143
2144    /*
2145     * If @target is currently being executed, schedule the
2146     * barrier to the worker; otherwise, put it after @target.
2147     */
2148    if (worker)
2149        head = worker->scheduled.next;
2150    else {
2151        unsigned long *bits = work_data_bits(target);
2152
2153        head = target->entry.next;
2154        /* there can already be other linked works, inherit and set */
2155        linked = *bits & WORK_STRUCT_LINKED;
2156        __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2157    }
2158
2159    debug_work_activate(&barr->work);
2160    insert_work(cwq, &barr->work, head,
2161            work_color_to_flags(WORK_NO_COLOR) | linked);
2162}
2163
2164/**
2165 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2166 * @wq: workqueue being flushed
2167 * @flush_color: new flush color, < 0 for no-op
2168 * @work_color: new work color, < 0 for no-op
2169 *
2170 * Prepare cwqs for workqueue flushing.
2171 *
2172 * If @flush_color is non-negative, flush_color on all cwqs should be
2173 * -1. If no cwq has in-flight commands at the specified color, all
2174 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2175 * has in flight commands, its cwq->flush_color is set to
2176 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2177 * wakeup logic is armed and %true is returned.
2178 *
2179 * The caller should have initialized @wq->first_flusher prior to
2180 * calling this function with non-negative @flush_color. If
2181 * @flush_color is negative, no flush color update is done and %false
2182 * is returned.
2183 *
2184 * If @work_color is non-negative, all cwqs should have the same
2185 * work_color which is previous to @work_color and all will be
2186 * advanced to @work_color.
2187 *
2188 * CONTEXT:
2189 * mutex_lock(wq->flush_mutex).
2190 *
2191 * RETURNS:
2192 * %true if @flush_color >= 0 and there's something to flush. %false
2193 * otherwise.
2194 */
2195static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2196                      int flush_color, int work_color)
2197{
2198    bool wait = false;
2199    unsigned int cpu;
2200
2201    if (flush_color >= 0) {
2202        BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2203        atomic_set(&wq->nr_cwqs_to_flush, 1);
2204    }
2205
2206    for_each_cwq_cpu(cpu, wq) {
2207        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2208        struct global_cwq *gcwq = cwq->gcwq;
2209
2210        spin_lock_irq(&gcwq->lock);
2211
2212        if (flush_color >= 0) {
2213            BUG_ON(cwq->flush_color != -1);
2214
2215            if (cwq->nr_in_flight[flush_color]) {
2216                cwq->flush_color = flush_color;
2217                atomic_inc(&wq->nr_cwqs_to_flush);
2218                wait = true;
2219            }
2220        }
2221
2222        if (work_color >= 0) {
2223            BUG_ON(work_color != work_next_color(cwq->work_color));
2224            cwq->work_color = work_color;
2225        }
2226
2227        spin_unlock_irq(&gcwq->lock);
2228    }
2229
2230    if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2231        complete(&wq->first_flusher->done);
2232
2233    return wait;
2234}
2235
2236/**
2237 * flush_workqueue - ensure that any scheduled work has run to completion.
2238 * @wq: workqueue to flush
2239 *
2240 * Forces execution of the workqueue and blocks until its completion.
2241 * This is typically used in driver shutdown handlers.
2242 *
2243 * We sleep until all works which were queued on entry have been handled,
2244 * but we are not livelocked by new incoming ones.
2245 */
2246void flush_workqueue(struct workqueue_struct *wq)
2247{
2248    struct wq_flusher this_flusher = {
2249        .list = LIST_HEAD_INIT(this_flusher.list),
2250        .flush_color = -1,
2251        .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2252    };
2253    int next_color;
2254
2255    lock_map_acquire(&wq->lockdep_map);
2256    lock_map_release(&wq->lockdep_map);
2257
2258    mutex_lock(&wq->flush_mutex);
2259
2260    /*
2261     * Start-to-wait phase
2262     */
2263    next_color = work_next_color(wq->work_color);
2264
2265    if (next_color != wq->flush_color) {
2266        /*
2267         * Color space is not full. The current work_color
2268         * becomes our flush_color and work_color is advanced
2269         * by one.
2270         */
2271        BUG_ON(!list_empty(&wq->flusher_overflow));
2272        this_flusher.flush_color = wq->work_color;
2273        wq->work_color = next_color;
2274
2275        if (!wq->first_flusher) {
2276            /* no flush in progress, become the first flusher */
2277            BUG_ON(wq->flush_color != this_flusher.flush_color);
2278
2279            wq->first_flusher = &this_flusher;
2280
2281            if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2282                               wq->work_color)) {
2283                /* nothing to flush, done */
2284                wq->flush_color = next_color;
2285                wq->first_flusher = NULL;
2286                goto out_unlock;
2287            }
2288        } else {
2289            /* wait in queue */
2290            BUG_ON(wq->flush_color == this_flusher.flush_color);
2291            list_add_tail(&this_flusher.list, &wq->flusher_queue);
2292            flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2293        }
2294    } else {
2295        /*
2296         * Oops, color space is full, wait on overflow queue.
2297         * The next flush completion will assign us
2298         * flush_color and transfer to flusher_queue.
2299         */
2300        list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2301    }
2302
2303    mutex_unlock(&wq->flush_mutex);
2304
2305    wait_for_completion(&this_flusher.done);
2306
2307    /*
2308     * Wake-up-and-cascade phase
2309     *
2310     * First flushers are responsible for cascading flushes and
2311     * handling overflow. Non-first flushers can simply return.
2312     */
2313    if (wq->first_flusher != &this_flusher)
2314        return;
2315
2316    mutex_lock(&wq->flush_mutex);
2317
2318    /* we might have raced, check again with mutex held */
2319    if (wq->first_flusher != &this_flusher)
2320        goto out_unlock;
2321
2322    wq->first_flusher = NULL;
2323
2324    BUG_ON(!list_empty(&this_flusher.list));
2325    BUG_ON(wq->flush_color != this_flusher.flush_color);
2326
2327    while (true) {
2328        struct wq_flusher *next, *tmp;
2329
2330        /* complete all the flushers sharing the current flush color */
2331        list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2332            if (next->flush_color != wq->flush_color)
2333                break;
2334            list_del_init(&next->list);
2335            complete(&next->done);
2336        }
2337
2338        BUG_ON(!list_empty(&wq->flusher_overflow) &&
2339               wq->flush_color != work_next_color(wq->work_color));
2340
2341        /* this flush_color is finished, advance by one */
2342        wq->flush_color = work_next_color(wq->flush_color);
2343
2344        /* one color has been freed, handle overflow queue */
2345        if (!list_empty(&wq->flusher_overflow)) {
2346            /*
2347             * Assign the same color to all overflowed
2348             * flushers, advance work_color and append to
2349             * flusher_queue. This is the start-to-wait
2350             * phase for these overflowed flushers.
2351             */
2352            list_for_each_entry(tmp, &wq->flusher_overflow, list)
2353                tmp->flush_color = wq->work_color;
2354
2355            wq->work_color = work_next_color(wq->work_color);
2356
2357            list_splice_tail_init(&wq->flusher_overflow,
2358                          &wq->flusher_queue);
2359            flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2360        }
2361
2362        if (list_empty(&wq->flusher_queue)) {
2363            BUG_ON(wq->flush_color != wq->work_color);
2364            break;
2365        }
2366
2367        /*
2368         * Need to flush more colors. Make the next flusher
2369         * the new first flusher and arm cwqs.
2370         */
2371        BUG_ON(wq->flush_color == wq->work_color);
2372        BUG_ON(wq->flush_color != next->flush_color);
2373
2374        list_del_init(&next->list);
2375        wq->first_flusher = next;
2376
2377        if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2378            break;
2379
2380        /*
2381         * Meh... this color is already done, clear first
2382         * flusher and repeat cascading.
2383         */
2384        wq->first_flusher = NULL;
2385    }
2386
2387out_unlock:
2388    mutex_unlock(&wq->flush_mutex);
2389}
2390EXPORT_SYMBOL_GPL(flush_workqueue);
2391
2392/**
2393 * drain_workqueue - drain a workqueue
2394 * @wq: workqueue to drain
2395 *
2396 * Wait until the workqueue becomes empty. While draining is in progress,
2397 * only chain queueing is allowed. IOW, only currently pending or running
2398 * work items on @wq can queue further work items on it. @wq is flushed
2399 * repeatedly until it becomes empty. The number of flushing is detemined
2400 * by the depth of chaining and should be relatively short. Whine if it
2401 * takes too long.
2402 */
2403void drain_workqueue(struct workqueue_struct *wq)
2404{
2405    unsigned int flush_cnt = 0;
2406    unsigned int cpu;
2407
2408    /*
2409     * __queue_work() needs to test whether there are drainers, is much
2410     * hotter than drain_workqueue() and already looks at @wq->flags.
2411     * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2412     */
2413    spin_lock(&workqueue_lock);
2414    if (!wq->nr_drainers++)
2415        wq->flags |= WQ_DRAINING;
2416    spin_unlock(&workqueue_lock);
2417reflush:
2418    flush_workqueue(wq);
2419
2420    for_each_cwq_cpu(cpu, wq) {
2421        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2422        bool drained;
2423
2424        spin_lock_irq(&cwq->gcwq->lock);
2425        drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2426        spin_unlock_irq(&cwq->gcwq->lock);
2427
2428        if (drained)
2429            continue;
2430
2431        if (++flush_cnt == 10 ||
2432            (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2433            pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2434                   wq->name, flush_cnt);
2435        goto reflush;
2436    }
2437
2438    spin_lock(&workqueue_lock);
2439    if (!--wq->nr_drainers)
2440        wq->flags &= ~WQ_DRAINING;
2441    spin_unlock(&workqueue_lock);
2442}
2443EXPORT_SYMBOL_GPL(drain_workqueue);
2444
2445static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2446                 bool wait_executing)
2447{
2448    struct worker *worker = NULL;
2449    struct global_cwq *gcwq;
2450    struct cpu_workqueue_struct *cwq;
2451
2452    might_sleep();
2453    gcwq = get_work_gcwq(work);
2454    if (!gcwq)
2455        return false;
2456
2457    spin_lock_irq(&gcwq->lock);
2458    if (!list_empty(&work->entry)) {
2459        /*
2460         * See the comment near try_to_grab_pending()->smp_rmb().
2461         * If it was re-queued to a different gcwq under us, we
2462         * are not going to wait.
2463         */
2464        smp_rmb();
2465        cwq = get_work_cwq(work);
2466        if (unlikely(!cwq || gcwq != cwq->gcwq))
2467            goto already_gone;
2468    } else if (wait_executing) {
2469        worker = find_worker_executing_work(gcwq, work);
2470        if (!worker)
2471            goto already_gone;
2472        cwq = worker->current_cwq;
2473    } else
2474        goto already_gone;
2475
2476    insert_wq_barrier(cwq, barr, work, worker);
2477    spin_unlock_irq(&gcwq->lock);
2478
2479    /*
2480     * If @max_active is 1 or rescuer is in use, flushing another work
2481     * item on the same workqueue may lead to deadlock. Make sure the
2482     * flusher is not running on the same workqueue by verifying write
2483     * access.
2484     */
2485    if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2486        lock_map_acquire(&cwq->wq->lockdep_map);
2487    else
2488        lock_map_acquire_read(&cwq->wq->lockdep_map);
2489    lock_map_release(&cwq->wq->lockdep_map);
2490
2491    return true;
2492already_gone:
2493    spin_unlock_irq(&gcwq->lock);
2494    return false;
2495}
2496
2497/**
2498 * flush_work - wait for a work to finish executing the last queueing instance
2499 * @work: the work to flush
2500 *
2501 * Wait until @work has finished execution. This function considers
2502 * only the last queueing instance of @work. If @work has been
2503 * enqueued across different CPUs on a non-reentrant workqueue or on
2504 * multiple workqueues, @work might still be executing on return on
2505 * some of the CPUs from earlier queueing.
2506 *
2507 * If @work was queued only on a non-reentrant, ordered or unbound
2508 * workqueue, @work is guaranteed to be idle on return if it hasn't
2509 * been requeued since flush started.
2510 *
2511 * RETURNS:
2512 * %true if flush_work() waited for the work to finish execution,
2513 * %false if it was already idle.
2514 */
2515bool flush_work(struct work_struct *work)
2516{
2517    struct wq_barrier barr;
2518
2519    lock_map_acquire(&work->lockdep_map);
2520    lock_map_release(&work->lockdep_map);
2521
2522    if (start_flush_work(work, &barr, true)) {
2523        wait_for_completion(&barr.done);
2524        destroy_work_on_stack(&barr.work);
2525        return true;
2526    } else
2527        return false;
2528}
2529EXPORT_SYMBOL_GPL(flush_work);
2530
2531static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2532{
2533    struct wq_barrier barr;
2534    struct worker *worker;
2535
2536    spin_lock_irq(&gcwq->lock);
2537
2538    worker = find_worker_executing_work(gcwq, work);
2539    if (unlikely(worker))
2540        insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2541
2542    spin_unlock_irq(&gcwq->lock);
2543
2544    if (unlikely(worker)) {
2545        wait_for_completion(&barr.done);
2546        destroy_work_on_stack(&barr.work);
2547        return true;
2548    } else
2549        return false;
2550}
2551
2552static bool wait_on_work(struct work_struct *work)
2553{
2554    bool ret = false;
2555    int cpu;
2556
2557    might_sleep();
2558
2559    lock_map_acquire(&work->lockdep_map);
2560    lock_map_release(&work->lockdep_map);
2561
2562    for_each_gcwq_cpu(cpu)
2563        ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2564    return ret;
2565}
2566
2567/**
2568 * flush_work_sync - wait until a work has finished execution
2569 * @work: the work to flush
2570 *
2571 * Wait until @work has finished execution. On return, it's
2572 * guaranteed that all queueing instances of @work which happened
2573 * before this function is called are finished. In other words, if
2574 * @work hasn't been requeued since this function was called, @work is
2575 * guaranteed to be idle on return.
2576 *
2577 * RETURNS:
2578 * %true if flush_work_sync() waited for the work to finish execution,
2579 * %false if it was already idle.
2580 */
2581bool flush_work_sync(struct work_struct *work)
2582{
2583    struct wq_barrier barr;
2584    bool pending, waited;
2585
2586    /* we'll wait for executions separately, queue barr only if pending */
2587    pending = start_flush_work(work, &barr, false);
2588
2589    /* wait for executions to finish */
2590    waited = wait_on_work(work);
2591
2592    /* wait for the pending one */
2593    if (pending) {
2594        wait_for_completion(&barr.done);
2595        destroy_work_on_stack(&barr.work);
2596    }
2597
2598    return pending || waited;
2599}
2600EXPORT_SYMBOL_GPL(flush_work_sync);
2601
2602/*
2603 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2604 * so this work can't be re-armed in any way.
2605 */
2606static int try_to_grab_pending(struct work_struct *work)
2607{
2608    struct global_cwq *gcwq;
2609    int ret = -1;
2610
2611    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2612        return 0;
2613
2614    /*
2615     * The queueing is in progress, or it is already queued. Try to
2616     * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2617     */
2618    gcwq = get_work_gcwq(work);
2619    if (!gcwq)
2620        return ret;
2621
2622    spin_lock_irq(&gcwq->lock);
2623    if (!list_empty(&work->entry)) {
2624        /*
2625         * This work is queued, but perhaps we locked the wrong gcwq.
2626         * In that case we must see the new value after rmb(), see
2627         * insert_work()->wmb().
2628         */
2629        smp_rmb();
2630        if (gcwq == get_work_gcwq(work)) {
2631            debug_work_deactivate(work);
2632            list_del_init(&work->entry);
2633            cwq_dec_nr_in_flight(get_work_cwq(work),
2634                get_work_color(work),
2635                *work_data_bits(work) & WORK_STRUCT_DELAYED);
2636            ret = 1;
2637        }
2638    }
2639    spin_unlock_irq(&gcwq->lock);
2640
2641    return ret;
2642}
2643
2644static bool __cancel_work_timer(struct work_struct *work,
2645                struct timer_list* timer)
2646{
2647    int ret;
2648
2649    do {
2650        ret = (timer && likely(del_timer(timer)));
2651        if (!ret)
2652            ret = try_to_grab_pending(work);
2653        wait_on_work(work);
2654    } while (unlikely(ret < 0));
2655
2656    clear_work_data(work);
2657    return ret;
2658}
2659
2660/**
2661 * cancel_work_sync - cancel a work and wait for it to finish
2662 * @work: the work to cancel
2663 *
2664 * Cancel @work and wait for its execution to finish. This function
2665 * can be used even if the work re-queues itself or migrates to
2666 * another workqueue. On return from this function, @work is
2667 * guaranteed to be not pending or executing on any CPU.
2668 *
2669 * cancel_work_sync(&delayed_work->work) must not be used for
2670 * delayed_work's. Use cancel_delayed_work_sync() instead.
2671 *
2672 * The caller must ensure that the workqueue on which @work was last
2673 * queued can't be destroyed before this function returns.
2674 *
2675 * RETURNS:
2676 * %true if @work was pending, %false otherwise.
2677 */
2678bool cancel_work_sync(struct work_struct *work)
2679{
2680    return __cancel_work_timer(work, NULL);
2681}
2682EXPORT_SYMBOL_GPL(cancel_work_sync);
2683
2684/**
2685 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2686 * @dwork: the delayed work to flush
2687 *
2688 * Delayed timer is cancelled and the pending work is queued for
2689 * immediate execution. Like flush_work(), this function only
2690 * considers the last queueing instance of @dwork.
2691 *
2692 * RETURNS:
2693 * %true if flush_work() waited for the work to finish execution,
2694 * %false if it was already idle.
2695 */
2696bool flush_delayed_work(struct delayed_work *dwork)
2697{
2698    if (del_timer_sync(&dwork->timer))
2699        __queue_work(raw_smp_processor_id(),
2700                 get_work_cwq(&dwork->work)->wq, &dwork->work);
2701    return flush_work(&dwork->work);
2702}
2703EXPORT_SYMBOL(flush_delayed_work);
2704
2705/**
2706 * flush_delayed_work_sync - wait for a dwork to finish
2707 * @dwork: the delayed work to flush
2708 *
2709 * Delayed timer is cancelled and the pending work is queued for
2710 * execution immediately. Other than timer handling, its behavior
2711 * is identical to flush_work_sync().
2712 *
2713 * RETURNS:
2714 * %true if flush_work_sync() waited for the work to finish execution,
2715 * %false if it was already idle.
2716 */
2717bool flush_delayed_work_sync(struct delayed_work *dwork)
2718{
2719    if (del_timer_sync(&dwork->timer))
2720        __queue_work(raw_smp_processor_id(),
2721                 get_work_cwq(&dwork->work)->wq, &dwork->work);
2722    return flush_work_sync(&dwork->work);
2723}
2724EXPORT_SYMBOL(flush_delayed_work_sync);
2725
2726/**
2727 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2728 * @dwork: the delayed work cancel
2729 *
2730 * This is cancel_work_sync() for delayed works.
2731 *
2732 * RETURNS:
2733 * %true if @dwork was pending, %false otherwise.
2734 */
2735bool cancel_delayed_work_sync(struct delayed_work *dwork)
2736{
2737    return __cancel_work_timer(&dwork->work, &dwork->timer);
2738}
2739EXPORT_SYMBOL(cancel_delayed_work_sync);
2740
2741/**
2742 * schedule_work - put work task in global workqueue
2743 * @work: job to be done
2744 *
2745 * Returns zero if @work was already on the kernel-global workqueue and
2746 * non-zero otherwise.
2747 *
2748 * This puts a job in the kernel-global workqueue if it was not already
2749 * queued and leaves it in the same position on the kernel-global
2750 * workqueue otherwise.
2751 */
2752int schedule_work(struct work_struct *work)
2753{
2754    return queue_work(system_wq, work);
2755}
2756EXPORT_SYMBOL(schedule_work);
2757
2758/*
2759 * schedule_work_on - put work task on a specific cpu
2760 * @cpu: cpu to put the work task on
2761 * @work: job to be done
2762 *
2763 * This puts a job on a specific cpu
2764 */
2765int schedule_work_on(int cpu, struct work_struct *work)
2766{
2767    return queue_work_on(cpu, system_wq, work);
2768}
2769EXPORT_SYMBOL(schedule_work_on);
2770
2771/**
2772 * schedule_delayed_work - put work task in global workqueue after delay
2773 * @dwork: job to be done
2774 * @delay: number of jiffies to wait or 0 for immediate execution
2775 *
2776 * After waiting for a given time this puts a job in the kernel-global
2777 * workqueue.
2778 */
2779int schedule_delayed_work(struct delayed_work *dwork,
2780                    unsigned long delay)
2781{
2782    return queue_delayed_work(system_wq, dwork, delay);
2783}
2784EXPORT_SYMBOL(schedule_delayed_work);
2785
2786/**
2787 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2788 * @cpu: cpu to use
2789 * @dwork: job to be done
2790 * @delay: number of jiffies to wait
2791 *
2792 * After waiting for a given time this puts a job in the kernel-global
2793 * workqueue on the specified CPU.
2794 */
2795int schedule_delayed_work_on(int cpu,
2796            struct delayed_work *dwork, unsigned long delay)
2797{
2798    return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2799}
2800EXPORT_SYMBOL(schedule_delayed_work_on);
2801
2802/**
2803 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2804 * @func: the function to call
2805 *
2806 * schedule_on_each_cpu() executes @func on each online CPU using the
2807 * system workqueue and blocks until all CPUs have completed.
2808 * schedule_on_each_cpu() is very slow.
2809 *
2810 * RETURNS:
2811 * 0 on success, -errno on failure.
2812 */
2813int schedule_on_each_cpu(work_func_t func)
2814{
2815    int cpu;
2816    struct work_struct __percpu *works;
2817
2818    works = alloc_percpu(struct work_struct);
2819    if (!works)
2820        return -ENOMEM;
2821
2822    get_online_cpus();
2823
2824    for_each_online_cpu(cpu) {
2825        struct work_struct *work = per_cpu_ptr(works, cpu);
2826
2827        INIT_WORK(work, func);
2828        schedule_work_on(cpu, work);
2829    }
2830
2831    for_each_online_cpu(cpu)
2832        flush_work(per_cpu_ptr(works, cpu));
2833
2834    put_online_cpus();
2835    free_percpu(works);
2836    return 0;
2837}
2838
2839/**
2840 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2841 *
2842 * Forces execution of the kernel-global workqueue and blocks until its
2843 * completion.
2844 *
2845 * Think twice before calling this function! It's very easy to get into
2846 * trouble if you don't take great care. Either of the following situations
2847 * will lead to deadlock:
2848 *
2849 * One of the work items currently on the workqueue needs to acquire
2850 * a lock held by your code or its caller.
2851 *
2852 * Your code is running in the context of a work routine.
2853 *
2854 * They will be detected by lockdep when they occur, but the first might not
2855 * occur very often. It depends on what work items are on the workqueue and
2856 * what locks they need, which you have no control over.
2857 *
2858 * In most situations flushing the entire workqueue is overkill; you merely
2859 * need to know that a particular work item isn't queued and isn't running.
2860 * In such cases you should use cancel_delayed_work_sync() or
2861 * cancel_work_sync() instead.
2862 */
2863void flush_scheduled_work(void)
2864{
2865    flush_workqueue(system_wq);
2866}
2867EXPORT_SYMBOL(flush_scheduled_work);
2868
2869/**
2870 * execute_in_process_context - reliably execute the routine with user context
2871 * @fn: the function to execute
2872 * @ew: guaranteed storage for the execute work structure (must
2873 * be available when the work executes)
2874 *
2875 * Executes the function immediately if process context is available,
2876 * otherwise schedules the function for delayed execution.
2877 *
2878 * Returns: 0 - function was executed
2879 * 1 - function was scheduled for execution
2880 */
2881int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2882{
2883    if (!in_interrupt()) {
2884        fn(&ew->work);
2885        return 0;
2886    }
2887
2888    INIT_WORK(&ew->work, fn);
2889    schedule_work(&ew->work);
2890
2891    return 1;
2892}
2893EXPORT_SYMBOL_GPL(execute_in_process_context);
2894
2895int keventd_up(void)
2896{
2897    return system_wq != NULL;
2898}
2899
2900static int alloc_cwqs(struct workqueue_struct *wq)
2901{
2902    /*
2903     * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2904     * Make sure that the alignment isn't lower than that of
2905     * unsigned long long.
2906     */
2907    const size_t size = sizeof(struct cpu_workqueue_struct);
2908    const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2909                   __alignof__(unsigned long long));
2910
2911    if (!(wq->flags & WQ_UNBOUND))
2912        wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2913    else {
2914        void *ptr;
2915
2916        /*
2917         * Allocate enough room to align cwq and put an extra
2918         * pointer at the end pointing back to the originally
2919         * allocated pointer which will be used for free.
2920         */
2921        ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2922        if (ptr) {
2923            wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2924            *(void **)(wq->cpu_wq.single + 1) = ptr;
2925        }
2926    }
2927
2928    /* just in case, make sure it's actually aligned */
2929    BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2930    return wq->cpu_wq.v ? 0 : -ENOMEM;
2931}
2932
2933static void free_cwqs(struct workqueue_struct *wq)
2934{
2935    if (!(wq->flags & WQ_UNBOUND))
2936        free_percpu(wq->cpu_wq.pcpu);
2937    else if (wq->cpu_wq.single) {
2938        /* the pointer to free is stored right after the cwq */
2939        kfree(*(void **)(wq->cpu_wq.single + 1));
2940    }
2941}
2942
2943static int wq_clamp_max_active(int max_active, unsigned int flags,
2944                   const char *name)
2945{
2946    int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2947
2948    if (max_active < 1 || max_active > lim)
2949        printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2950               "is out of range, clamping between %d and %d\n",
2951               max_active, name, 1, lim);
2952
2953    return clamp_val(max_active, 1, lim);
2954}
2955
2956struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
2957                           unsigned int flags,
2958                           int max_active,
2959                           struct lock_class_key *key,
2960                           const char *lock_name, ...)
2961{
2962    va_list args, args1;
2963    struct workqueue_struct *wq;
2964    unsigned int cpu;
2965    size_t namelen;
2966
2967    /* determine namelen, allocate wq and format name */
2968    va_start(args, lock_name);
2969    va_copy(args1, args);
2970    namelen = vsnprintf(NULL, 0, fmt, args) + 1;
2971
2972    wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
2973    if (!wq)
2974        goto err;
2975
2976    vsnprintf(wq->name, namelen, fmt, args1);
2977    va_end(args);
2978    va_end(args1);
2979
2980    /*
2981     * Workqueues which may be used during memory reclaim should
2982     * have a rescuer to guarantee forward progress.
2983     */
2984    if (flags & WQ_MEM_RECLAIM)
2985        flags |= WQ_RESCUER;
2986
2987    /*
2988     * Unbound workqueues aren't concurrency managed and should be
2989     * dispatched to workers immediately.
2990     */
2991    if (flags & WQ_UNBOUND)
2992        flags |= WQ_HIGHPRI;
2993
2994    max_active = max_active ?: WQ_DFL_ACTIVE;
2995    max_active = wq_clamp_max_active(max_active, flags, wq->name);
2996
2997    /* init wq */
2998    wq->flags = flags;
2999    wq->saved_max_active = max_active;
3000    mutex_init(&wq->flush_mutex);
3001    atomic_set(&wq->nr_cwqs_to_flush, 0);
3002    INIT_LIST_HEAD(&wq->flusher_queue);
3003    INIT_LIST_HEAD(&wq->flusher_overflow);
3004
3005    lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3006    INIT_LIST_HEAD(&wq->list);
3007
3008    if (alloc_cwqs(wq) < 0)
3009        goto err;
3010
3011    for_each_cwq_cpu(cpu, wq) {
3012        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3013        struct global_cwq *gcwq = get_gcwq(cpu);
3014
3015        BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3016        cwq->gcwq = gcwq;
3017        cwq->wq = wq;
3018        cwq->flush_color = -1;
3019        cwq->max_active = max_active;
3020        INIT_LIST_HEAD(&cwq->delayed_works);
3021    }
3022
3023    if (flags & WQ_RESCUER) {
3024        struct worker *rescuer;
3025
3026        if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3027            goto err;
3028
3029        wq->rescuer = rescuer = alloc_worker();
3030        if (!rescuer)
3031            goto err;
3032
3033        rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3034                           wq->name);
3035        if (IS_ERR(rescuer->task))
3036            goto err;
3037
3038        rescuer->task->flags |= PF_THREAD_BOUND;
3039        wake_up_process(rescuer->task);
3040    }
3041
3042    /*
3043     * workqueue_lock protects global freeze state and workqueues
3044     * list. Grab it, set max_active accordingly and add the new
3045     * workqueue to workqueues list.
3046     */
3047    spin_lock(&workqueue_lock);
3048
3049    if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3050        for_each_cwq_cpu(cpu, wq)
3051            get_cwq(cpu, wq)->max_active = 0;
3052
3053    list_add(&wq->list, &workqueues);
3054
3055    spin_unlock(&workqueue_lock);
3056
3057    return wq;
3058err:
3059    if (wq) {
3060        free_cwqs(wq);
3061        free_mayday_mask(wq->mayday_mask);
3062        kfree(wq->rescuer);
3063        kfree(wq);
3064    }
3065    return NULL;
3066}
3067EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3068
3069/**
3070 * destroy_workqueue - safely terminate a workqueue
3071 * @wq: target workqueue
3072 *
3073 * Safely destroy a workqueue. All work currently pending will be done first.
3074 */
3075void destroy_workqueue(struct workqueue_struct *wq)
3076{
3077    unsigned int cpu;
3078
3079    /* drain it before proceeding with destruction */
3080    drain_workqueue(wq);
3081
3082    /*
3083     * wq list is used to freeze wq, remove from list after
3084     * flushing is complete in case freeze races us.
3085     */
3086    spin_lock(&workqueue_lock);
3087    list_del(&wq->list);
3088    spin_unlock(&workqueue_lock);
3089
3090    /* sanity check */
3091    for_each_cwq_cpu(cpu, wq) {
3092        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3093        int i;
3094
3095        for (i = 0; i < WORK_NR_COLORS; i++)
3096            BUG_ON(cwq->nr_in_flight[i]);
3097        BUG_ON(cwq->nr_active);
3098        BUG_ON(!list_empty(&cwq->delayed_works));
3099    }
3100
3101    if (wq->flags & WQ_RESCUER) {
3102        kthread_stop(wq->rescuer->task);
3103        free_mayday_mask(wq->mayday_mask);
3104        kfree(wq->rescuer);
3105    }
3106
3107    free_cwqs(wq);
3108    kfree(wq);
3109}
3110EXPORT_SYMBOL_GPL(destroy_workqueue);
3111
3112/**
3113 * workqueue_set_max_active - adjust max_active of a workqueue
3114 * @wq: target workqueue
3115 * @max_active: new max_active value.
3116 *
3117 * Set max_active of @wq to @max_active.
3118 *
3119 * CONTEXT:
3120 * Don't call from IRQ context.
3121 */
3122void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3123{
3124    unsigned int cpu;
3125
3126    max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3127
3128    spin_lock(&workqueue_lock);
3129
3130    wq->saved_max_active = max_active;
3131
3132    for_each_cwq_cpu(cpu, wq) {
3133        struct global_cwq *gcwq = get_gcwq(cpu);
3134
3135        spin_lock_irq(&gcwq->lock);
3136
3137        if (!(wq->flags & WQ_FREEZABLE) ||
3138            !(gcwq->flags & GCWQ_FREEZING))
3139            get_cwq(gcwq->cpu, wq)->max_active = max_active;
3140
3141        spin_unlock_irq(&gcwq->lock);
3142    }
3143
3144    spin_unlock(&workqueue_lock);
3145}
3146EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3147
3148/**
3149 * workqueue_congested - test whether a workqueue is congested
3150 * @cpu: CPU in question
3151 * @wq: target workqueue
3152 *
3153 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3154 * no synchronization around this function and the test result is
3155 * unreliable and only useful as advisory hints or for debugging.
3156 *
3157 * RETURNS:
3158 * %true if congested, %false otherwise.
3159 */
3160bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3161{
3162    struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3163
3164    return !list_empty(&cwq->delayed_works);
3165}
3166EXPORT_SYMBOL_GPL(workqueue_congested);
3167
3168/**
3169 * work_cpu - return the last known associated cpu for @work
3170 * @work: the work of interest
3171 *
3172 * RETURNS:
3173 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
3174 */
3175unsigned int work_cpu(struct work_struct *work)
3176{
3177    struct global_cwq *gcwq = get_work_gcwq(work);
3178
3179    return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3180}
3181EXPORT_SYMBOL_GPL(work_cpu);
3182
3183/**
3184 * work_busy - test whether a work is currently pending or running
3185 * @work: the work to be tested
3186 *
3187 * Test whether @work is currently pending or running. There is no
3188 * synchronization around this function and the test result is
3189 * unreliable and only useful as advisory hints or for debugging.
3190 * Especially for reentrant wqs, the pending state might hide the
3191 * running state.
3192 *
3193 * RETURNS:
3194 * OR'd bitmask of WORK_BUSY_* bits.
3195 */
3196unsigned int work_busy(struct work_struct *work)
3197{
3198    struct global_cwq *gcwq = get_work_gcwq(work);
3199    unsigned long flags;
3200    unsigned int ret = 0;
3201
3202    if (!gcwq)
3203        return false;
3204
3205    spin_lock_irqsave(&gcwq->lock, flags);
3206
3207    if (work_pending(work))
3208        ret |= WORK_BUSY_PENDING;
3209    if (find_worker_executing_work(gcwq, work))
3210        ret |= WORK_BUSY_RUNNING;
3211
3212    spin_unlock_irqrestore(&gcwq->lock, flags);
3213
3214    return ret;
3215}
3216EXPORT_SYMBOL_GPL(work_busy);
3217
3218/*
3219 * CPU hotplug.
3220 *
3221 * There are two challenges in supporting CPU hotplug. Firstly, there
3222 * are a lot of assumptions on strong associations among work, cwq and
3223 * gcwq which make migrating pending and scheduled works very
3224 * difficult to implement without impacting hot paths. Secondly,
3225 * gcwqs serve mix of short, long and very long running works making
3226 * blocked draining impractical.
3227 *
3228 * This is solved by allowing a gcwq to be detached from CPU, running
3229 * it with unbound (rogue) workers and allowing it to be reattached
3230 * later if the cpu comes back online. A separate thread is created
3231 * to govern a gcwq in such state and is called the trustee of the
3232 * gcwq.
3233 *
3234 * Trustee states and their descriptions.
3235 *
3236 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3237 * new trustee is started with this state.
3238 *
3239 * IN_CHARGE Once started, trustee will enter this state after
3240 * assuming the manager role and making all existing
3241 * workers rogue. DOWN_PREPARE waits for trustee to
3242 * enter this state. After reaching IN_CHARGE, trustee
3243 * tries to execute the pending worklist until it's empty
3244 * and the state is set to BUTCHER, or the state is set
3245 * to RELEASE.
3246 *
3247 * BUTCHER Command state which is set by the cpu callback after
3248 * the cpu has went down. Once this state is set trustee
3249 * knows that there will be no new works on the worklist
3250 * and once the worklist is empty it can proceed to
3251 * killing idle workers.
3252 *
3253 * RELEASE Command state which is set by the cpu callback if the
3254 * cpu down has been canceled or it has come online
3255 * again. After recognizing this state, trustee stops
3256 * trying to drain or butcher and clears ROGUE, rebinds
3257 * all remaining workers back to the cpu and releases
3258 * manager role.
3259 *
3260 * DONE Trustee will enter this state after BUTCHER or RELEASE
3261 * is complete.
3262 *
3263 * trustee CPU draining
3264 * took over down complete
3265 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3266 * | | ^
3267 * | CPU is back online v return workers |
3268 * ----------------> RELEASE --------------
3269 */
3270
3271/**
3272 * trustee_wait_event_timeout - timed event wait for trustee
3273 * @cond: condition to wait for
3274 * @timeout: timeout in jiffies
3275 *
3276 * wait_event_timeout() for trustee to use. Handles locking and
3277 * checks for RELEASE request.
3278 *
3279 * CONTEXT:
3280 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3281 * multiple times. To be used by trustee.
3282 *
3283 * RETURNS:
3284 * Positive indicating left time if @cond is satisfied, 0 if timed
3285 * out, -1 if canceled.
3286 */
3287#define trustee_wait_event_timeout(cond, timeout) ({ \
3288    long __ret = (timeout); \
3289    while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3290           __ret) { \
3291        spin_unlock_irq(&gcwq->lock); \
3292        __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3293            (gcwq->trustee_state == TRUSTEE_RELEASE), \
3294            __ret); \
3295        spin_lock_irq(&gcwq->lock); \
3296    } \
3297    gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3298})
3299
3300/**
3301 * trustee_wait_event - event wait for trustee
3302 * @cond: condition to wait for
3303 *
3304 * wait_event() for trustee to use. Automatically handles locking and
3305 * checks for CANCEL request.
3306 *
3307 * CONTEXT:
3308 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3309 * multiple times. To be used by trustee.
3310 *
3311 * RETURNS:
3312 * 0 if @cond is satisfied, -1 if canceled.
3313 */
3314#define trustee_wait_event(cond) ({ \
3315    long __ret1; \
3316    __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3317    __ret1 < 0 ? -1 : 0; \
3318})
3319
3320static int __cpuinit trustee_thread(void *__gcwq)
3321{
3322    struct global_cwq *gcwq = __gcwq;
3323    struct worker *worker;
3324    struct work_struct *work;
3325    struct hlist_node *pos;
3326    long rc;
3327    int i;
3328
3329    BUG_ON(gcwq->cpu != smp_processor_id());
3330
3331    spin_lock_irq(&gcwq->lock);
3332    /*
3333     * Claim the manager position and make all workers rogue.
3334     * Trustee must be bound to the target cpu and can't be
3335     * cancelled.
3336     */
3337    BUG_ON(gcwq->cpu != smp_processor_id());
3338    rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3339    BUG_ON(rc < 0);
3340
3341    gcwq->flags |= GCWQ_MANAGING_WORKERS;
3342
3343    list_for_each_entry(worker, &gcwq->idle_list, entry)
3344        worker->flags |= WORKER_ROGUE;
3345
3346    for_each_busy_worker(worker, i, pos, gcwq)
3347        worker->flags |= WORKER_ROGUE;
3348
3349    /*
3350     * Call schedule() so that we cross rq->lock and thus can
3351     * guarantee sched callbacks see the rogue flag. This is
3352     * necessary as scheduler callbacks may be invoked from other
3353     * cpus.
3354     */
3355    spin_unlock_irq(&gcwq->lock);
3356    schedule();
3357    spin_lock_irq(&gcwq->lock);
3358
3359    /*
3360     * Sched callbacks are disabled now. Zap nr_running. After
3361     * this, nr_running stays zero and need_more_worker() and
3362     * keep_working() are always true as long as the worklist is
3363     * not empty.
3364     */
3365    atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3366
3367    spin_unlock_irq(&gcwq->lock);
3368    del_timer_sync(&gcwq->idle_timer);
3369    spin_lock_irq(&gcwq->lock);
3370
3371    /*
3372     * We're now in charge. Notify and proceed to drain. We need
3373     * to keep the gcwq running during the whole CPU down
3374     * procedure as other cpu hotunplug callbacks may need to
3375     * flush currently running tasks.
3376     */
3377    gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3378    wake_up_all(&gcwq->trustee_wait);
3379
3380    /*
3381     * The original cpu is in the process of dying and may go away
3382     * anytime now. When that happens, we and all workers would
3383     * be migrated to other cpus. Try draining any left work. We
3384     * want to get it over with ASAP - spam rescuers, wake up as
3385     * many idlers as necessary and create new ones till the
3386     * worklist is empty. Note that if the gcwq is frozen, there
3387     * may be frozen works in freezable cwqs. Don't declare
3388     * completion while frozen.
3389     */
3390    while (gcwq->nr_workers != gcwq->nr_idle ||
3391           gcwq->flags & GCWQ_FREEZING ||
3392           gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3393        int nr_works = 0;
3394
3395        list_for_each_entry(work, &gcwq->worklist, entry) {
3396            send_mayday(work);
3397            nr_works++;
3398        }
3399
3400        list_for_each_entry(worker, &gcwq->idle_list, entry) {
3401            if (!nr_works--)
3402                break;
3403            wake_up_process(worker->task);
3404        }
3405
3406        if (need_to_create_worker(gcwq)) {
3407            spin_unlock_irq(&gcwq->lock);
3408            worker = create_worker(gcwq, false);
3409            spin_lock_irq(&gcwq->lock);
3410            if (worker) {
3411                worker->flags |= WORKER_ROGUE;
3412                start_worker(worker);
3413            }
3414        }
3415
3416        /* give a breather */
3417        if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3418            break;
3419    }
3420
3421    /*
3422     * Either all works have been scheduled and cpu is down, or
3423     * cpu down has already been canceled. Wait for and butcher
3424     * all workers till we're canceled.
3425     */
3426    do {
3427        rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3428        while (!list_empty(&gcwq->idle_list))
3429            destroy_worker(list_first_entry(&gcwq->idle_list,
3430                            struct worker, entry));
3431    } while (gcwq->nr_workers && rc >= 0);
3432
3433    /*
3434     * At this point, either draining has completed and no worker
3435     * is left, or cpu down has been canceled or the cpu is being
3436     * brought back up. There shouldn't be any idle one left.
3437     * Tell the remaining busy ones to rebind once it finishes the
3438     * currently scheduled works by scheduling the rebind_work.
3439     */
3440    WARN_ON(!list_empty(&gcwq->idle_list));
3441
3442    for_each_busy_worker(worker, i, pos, gcwq) {
3443        struct work_struct *rebind_work = &worker->rebind_work;
3444
3445        /*
3446         * Rebind_work may race with future cpu hotplug
3447         * operations. Use a separate flag to mark that
3448         * rebinding is scheduled.
3449         */
3450        worker->flags |= WORKER_REBIND;
3451        worker->flags &= ~WORKER_ROGUE;
3452
3453        /* queue rebind_work, wq doesn't matter, use the default one */
3454        if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3455                     work_data_bits(rebind_work)))
3456            continue;
3457
3458        debug_work_activate(rebind_work);
3459        insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3460                worker->scheduled.next,
3461                work_color_to_flags(WORK_NO_COLOR));
3462    }
3463
3464    /* relinquish manager role */
3465    gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3466
3467    /* notify completion */
3468    gcwq->trustee = NULL;
3469    gcwq->trustee_state = TRUSTEE_DONE;
3470    wake_up_all(&gcwq->trustee_wait);
3471    spin_unlock_irq(&gcwq->lock);
3472    return 0;
3473}
3474
3475/**
3476 * wait_trustee_state - wait for trustee to enter the specified state
3477 * @gcwq: gcwq the trustee of interest belongs to
3478 * @state: target state to wait for
3479 *
3480 * Wait for the trustee to reach @state. DONE is already matched.
3481 *
3482 * CONTEXT:
3483 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3484 * multiple times. To be used by cpu_callback.
3485 */
3486static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3487__releases(&gcwq->lock)
3488__acquires(&gcwq->lock)
3489{
3490    if (!(gcwq->trustee_state == state ||
3491          gcwq->trustee_state == TRUSTEE_DONE)) {
3492        spin_unlock_irq(&gcwq->lock);
3493        __wait_event(gcwq->trustee_wait,
3494                 gcwq->trustee_state == state ||
3495                 gcwq->trustee_state == TRUSTEE_DONE);
3496        spin_lock_irq(&gcwq->lock);
3497    }
3498}
3499
3500static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3501                        unsigned long action,
3502                        void *hcpu)
3503{
3504    unsigned int cpu = (unsigned long)hcpu;
3505    struct global_cwq *gcwq = get_gcwq(cpu);
3506    struct task_struct *new_trustee = NULL;
3507    struct worker *uninitialized_var(new_worker);
3508    unsigned long flags;
3509
3510    action &= ~CPU_TASKS_FROZEN;
3511
3512    switch (action) {
3513    case CPU_DOWN_PREPARE:
3514        new_trustee = kthread_create(trustee_thread, gcwq,
3515                         "workqueue_trustee/%d\n", cpu);
3516        if (IS_ERR(new_trustee))
3517            return notifier_from_errno(PTR_ERR(new_trustee));
3518        kthread_bind(new_trustee, cpu);
3519        /* fall through */
3520    case CPU_UP_PREPARE:
3521        BUG_ON(gcwq->first_idle);
3522        new_worker = create_worker(gcwq, false);
3523        if (!new_worker) {
3524            if (new_trustee)
3525                kthread_stop(new_trustee);
3526            return NOTIFY_BAD;
3527        }
3528    }
3529
3530    /* some are called w/ irq disabled, don't disturb irq status */
3531    spin_lock_irqsave(&gcwq->lock, flags);
3532
3533    switch (action) {
3534    case CPU_DOWN_PREPARE:
3535        /* initialize trustee and tell it to acquire the gcwq */
3536        BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3537        gcwq->trustee = new_trustee;
3538        gcwq->trustee_state = TRUSTEE_START;
3539        wake_up_process(gcwq->trustee);
3540        wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3541        /* fall through */
3542    case CPU_UP_PREPARE:
3543        BUG_ON(gcwq->first_idle);
3544        gcwq->first_idle = new_worker;
3545        break;
3546
3547    case CPU_DYING:
3548        /*
3549         * Before this, the trustee and all workers except for
3550         * the ones which are still executing works from
3551         * before the last CPU down must be on the cpu. After
3552         * this, they'll all be diasporas.
3553         */
3554        gcwq->flags |= GCWQ_DISASSOCIATED;
3555        break;
3556
3557    case CPU_POST_DEAD:
3558        gcwq->trustee_state = TRUSTEE_BUTCHER;
3559        /* fall through */
3560    case CPU_UP_CANCELED:
3561        destroy_worker(gcwq->first_idle);
3562        gcwq->first_idle = NULL;
3563        break;
3564
3565    case CPU_DOWN_FAILED:
3566    case CPU_ONLINE:
3567        gcwq->flags &= ~GCWQ_DISASSOCIATED;
3568        if (gcwq->trustee_state != TRUSTEE_DONE) {
3569            gcwq->trustee_state = TRUSTEE_RELEASE;
3570            wake_up_process(gcwq->trustee);
3571            wait_trustee_state(gcwq, TRUSTEE_DONE);
3572        }
3573
3574        /*
3575         * Trustee is done and there might be no worker left.
3576         * Put the first_idle in and request a real manager to
3577         * take a look.
3578         */
3579        spin_unlock_irq(&gcwq->lock);
3580        kthread_bind(gcwq->first_idle->task, cpu);
3581        spin_lock_irq(&gcwq->lock);
3582        gcwq->flags |= GCWQ_MANAGE_WORKERS;
3583        start_worker(gcwq->first_idle);
3584        gcwq->first_idle = NULL;
3585        break;
3586    }
3587
3588    spin_unlock_irqrestore(&gcwq->lock, flags);
3589
3590    return notifier_from_errno(0);
3591}
3592
3593#ifdef CONFIG_SMP
3594
3595struct work_for_cpu {
3596    struct completion completion;
3597    long (*fn)(void *);
3598    void *arg;
3599    long ret;
3600};
3601
3602static int do_work_for_cpu(void *_wfc)
3603{
3604    struct work_for_cpu *wfc = _wfc;
3605    wfc->ret = wfc->fn(wfc->arg);
3606    complete(&wfc->completion);
3607    return 0;
3608}
3609
3610/**
3611 * work_on_cpu - run a function in user context on a particular cpu
3612 * @cpu: the cpu to run on
3613 * @fn: the function to run
3614 * @arg: the function arg
3615 *
3616 * This will return the value @fn returns.
3617 * It is up to the caller to ensure that the cpu doesn't go offline.
3618 * The caller must not hold any locks which would prevent @fn from completing.
3619 */
3620long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3621{
3622    struct task_struct *sub_thread;
3623    struct work_for_cpu wfc = {
3624        .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3625        .fn = fn,
3626        .arg = arg,
3627    };
3628
3629    sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3630    if (IS_ERR(sub_thread))
3631        return PTR_ERR(sub_thread);
3632    kthread_bind(sub_thread, cpu);
3633    wake_up_process(sub_thread);
3634    wait_for_completion(&wfc.completion);
3635    return wfc.ret;
3636}
3637EXPORT_SYMBOL_GPL(work_on_cpu);
3638#endif /* CONFIG_SMP */
3639
3640#ifdef CONFIG_FREEZER
3641
3642/**
3643 * freeze_workqueues_begin - begin freezing workqueues
3644 *
3645 * Start freezing workqueues. After this function returns, all freezable
3646 * workqueues will queue new works to their frozen_works list instead of
3647 * gcwq->worklist.
3648 *
3649 * CONTEXT:
3650 * Grabs and releases workqueue_lock and gcwq->lock's.
3651 */
3652void freeze_workqueues_begin(void)
3653{
3654    unsigned int cpu;
3655
3656    spin_lock(&workqueue_lock);
3657
3658    BUG_ON(workqueue_freezing);
3659    workqueue_freezing = true;
3660
3661    for_each_gcwq_cpu(cpu) {
3662        struct global_cwq *gcwq = get_gcwq(cpu);
3663        struct workqueue_struct *wq;
3664
3665        spin_lock_irq(&gcwq->lock);
3666
3667        BUG_ON(gcwq->flags & GCWQ_FREEZING);
3668        gcwq->flags |= GCWQ_FREEZING;
3669
3670        list_for_each_entry(wq, &workqueues, list) {
3671            struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3672
3673            if (cwq && wq->flags & WQ_FREEZABLE)
3674                cwq->max_active = 0;
3675        }
3676
3677        spin_unlock_irq(&gcwq->lock);
3678    }
3679
3680    spin_unlock(&workqueue_lock);
3681}
3682
3683/**
3684 * freeze_workqueues_busy - are freezable workqueues still busy?
3685 *
3686 * Check whether freezing is complete. This function must be called
3687 * between freeze_workqueues_begin() and thaw_workqueues().
3688 *
3689 * CONTEXT:
3690 * Grabs and releases workqueue_lock.
3691 *
3692 * RETURNS:
3693 * %true if some freezable workqueues are still busy. %false if freezing
3694 * is complete.
3695 */
3696bool freeze_workqueues_busy(void)
3697{
3698    unsigned int cpu;
3699    bool busy = false;
3700
3701    spin_lock(&workqueue_lock);
3702
3703    BUG_ON(!workqueue_freezing);
3704
3705    for_each_gcwq_cpu(cpu) {
3706        struct workqueue_struct *wq;
3707        /*
3708         * nr_active is monotonically decreasing. It's safe
3709         * to peek without lock.
3710         */
3711        list_for_each_entry(wq, &workqueues, list) {
3712            struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3713
3714            if (!cwq || !(wq->flags & WQ_FREEZABLE))
3715                continue;
3716
3717            BUG_ON(cwq->nr_active < 0);
3718            if (cwq->nr_active) {
3719                busy = true;
3720                goto out_unlock;
3721            }
3722        }
3723    }
3724out_unlock:
3725    spin_unlock(&workqueue_lock);
3726    return busy;
3727}
3728
3729/**
3730 * thaw_workqueues - thaw workqueues
3731 *
3732 * Thaw workqueues. Normal queueing is restored and all collected
3733 * frozen works are transferred to their respective gcwq worklists.
3734 *
3735 * CONTEXT:
3736 * Grabs and releases workqueue_lock and gcwq->lock's.
3737 */
3738void thaw_workqueues(void)
3739{
3740    unsigned int cpu;
3741
3742    spin_lock(&workqueue_lock);
3743
3744    if (!workqueue_freezing)
3745        goto out_unlock;
3746
3747    for_each_gcwq_cpu(cpu) {
3748        struct global_cwq *gcwq = get_gcwq(cpu);
3749        struct workqueue_struct *wq;
3750
3751        spin_lock_irq(&gcwq->lock);
3752
3753        BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3754        gcwq->flags &= ~GCWQ_FREEZING;
3755
3756        list_for_each_entry(wq, &workqueues, list) {
3757            struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3758
3759            if (!cwq || !(wq->flags & WQ_FREEZABLE))
3760                continue;
3761
3762            /* restore max_active and repopulate worklist */
3763            cwq->max_active = wq->saved_max_active;
3764
3765            while (!list_empty(&cwq->delayed_works) &&
3766                   cwq->nr_active < cwq->max_active)
3767                cwq_activate_first_delayed(cwq);
3768        }
3769
3770        wake_up_worker(gcwq);
3771
3772        spin_unlock_irq(&gcwq->lock);
3773    }
3774
3775    workqueue_freezing = false;
3776out_unlock:
3777    spin_unlock(&workqueue_lock);
3778}
3779#endif /* CONFIG_FREEZER */
3780
3781static int __init init_workqueues(void)
3782{
3783    unsigned int cpu;
3784    int i;
3785
3786    cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
3787
3788    /* initialize gcwqs */
3789    for_each_gcwq_cpu(cpu) {
3790        struct global_cwq *gcwq = get_gcwq(cpu);
3791
3792        spin_lock_init(&gcwq->lock);
3793        INIT_LIST_HEAD(&gcwq->worklist);
3794        gcwq->cpu = cpu;
3795        gcwq->flags |= GCWQ_DISASSOCIATED;
3796
3797        INIT_LIST_HEAD(&gcwq->idle_list);
3798        for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3799            INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3800
3801        init_timer_deferrable(&gcwq->idle_timer);
3802        gcwq->idle_timer.function = idle_worker_timeout;
3803        gcwq->idle_timer.data = (unsigned long)gcwq;
3804
3805        setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3806                (unsigned long)gcwq);
3807
3808        ida_init(&gcwq->worker_ida);
3809
3810        gcwq->trustee_state = TRUSTEE_DONE;
3811        init_waitqueue_head(&gcwq->trustee_wait);
3812    }
3813
3814    /* create the initial worker */
3815    for_each_online_gcwq_cpu(cpu) {
3816        struct global_cwq *gcwq = get_gcwq(cpu);
3817        struct worker *worker;
3818
3819        if (cpu != WORK_CPU_UNBOUND)
3820            gcwq->flags &= ~GCWQ_DISASSOCIATED;
3821        worker = create_worker(gcwq, true);
3822        BUG_ON(!worker);
3823        spin_lock_irq(&gcwq->lock);
3824        start_worker(worker);
3825        spin_unlock_irq(&gcwq->lock);
3826    }
3827
3828    system_wq = alloc_workqueue("events", 0, 0);
3829    system_long_wq = alloc_workqueue("events_long", 0, 0);
3830    system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3831    system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3832                        WQ_UNBOUND_MAX_ACTIVE);
3833    system_freezable_wq = alloc_workqueue("events_freezable",
3834                          WQ_FREEZABLE, 0);
3835    system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3836            WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3837    BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
3838           !system_unbound_wq || !system_freezable_wq ||
3839        !system_nrt_freezable_wq);
3840    return 0;
3841}
3842early_initcall(init_workqueues);
3843

Archive Download this file



interactive