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 are two worker pools for each CPU (one for
20 * normal work items and the other for high priority ones) and some extra
21 * pools for workqueues which are not bound to any specific CPU - the
22 * number of these backing pools is dynamic.
23 *
24 * Please read Documentation/workqueue.txt for details.
25 */
26
27#include <linux/export.h>
28#include <linux/kernel.h>
29#include <linux/sched.h>
30#include <linux/init.h>
31#include <linux/signal.h>
32#include <linux/completion.h>
33#include <linux/workqueue.h>
34#include <linux/slab.h>
35#include <linux/cpu.h>
36#include <linux/notifier.h>
37#include <linux/kthread.h>
38#include <linux/hardirq.h>
39#include <linux/mempolicy.h>
40#include <linux/freezer.h>
41#include <linux/kallsyms.h>
42#include <linux/debug_locks.h>
43#include <linux/lockdep.h>
44#include <linux/idr.h>
45#include <linux/jhash.h>
46#include <linux/hashtable.h>
47#include <linux/rculist.h>
48#include <linux/nodemask.h>
49#include <linux/moduleparam.h>
50#include <linux/uaccess.h>
51
52#include "workqueue_internal.h"
53
54enum {
55    /*
56     * worker_pool flags
57     *
58     * A bound pool is either associated or disassociated with its CPU.
59     * While associated (!DISASSOCIATED), all workers are bound to the
60     * CPU and none has %WORKER_UNBOUND set and concurrency management
61     * is in effect.
62     *
63     * While DISASSOCIATED, the cpu may be offline and all workers have
64     * %WORKER_UNBOUND set and concurrency management disabled, and may
65     * be executing on any CPU. The pool behaves as an unbound one.
66     *
67     * Note that DISASSOCIATED should be flipped only while holding
68     * manager_mutex to avoid changing binding state while
69     * create_worker() is in progress.
70     */
71    POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
72    POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
73    POOL_FREEZING = 1 << 3, /* freeze in progress */
74
75    /* worker flags */
76    WORKER_STARTED = 1 << 0, /* started */
77    WORKER_DIE = 1 << 1, /* die die die */
78    WORKER_IDLE = 1 << 2, /* is idle */
79    WORKER_PREP = 1 << 3, /* preparing to run works */
80    WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
81    WORKER_UNBOUND = 1 << 7, /* worker is unbound */
82    WORKER_REBOUND = 1 << 8, /* worker was rebound */
83
84    WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE |
85                  WORKER_UNBOUND | WORKER_REBOUND,
86
87    NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */
88
89    UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */
90    BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
91
92    MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
93    IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
94
95    MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
96                        /* call for help after 10ms
97                           (min two ticks) */
98    MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
99    CREATE_COOLDOWN = HZ, /* time to breath after fail */
100
101    /*
102     * Rescue workers are used only on emergencies and shared by
103     * all cpus. Give -20.
104     */
105    RESCUER_NICE_LEVEL = -20,
106    HIGHPRI_NICE_LEVEL = -20,
107
108    WQ_NAME_LEN = 24,
109};
110
111/*
112 * Structure fields follow one of the following exclusion rules.
113 *
114 * I: Modifiable by initialization/destruction paths and read-only for
115 * everyone else.
116 *
117 * P: Preemption protected. Disabling preemption is enough and should
118 * only be modified and accessed from the local cpu.
119 *
120 * L: pool->lock protected. Access with pool->lock held.
121 *
122 * X: During normal operation, modification requires pool->lock and should
123 * be done only from local cpu. Either disabling preemption on local
124 * cpu or grabbing pool->lock is enough for read access. If
125 * POOL_DISASSOCIATED is set, it's identical to L.
126 *
127 * MG: pool->manager_mutex and pool->lock protected. Writes require both
128 * locks. Reads can happen under either lock.
129 *
130 * PL: wq_pool_mutex protected.
131 *
132 * PR: wq_pool_mutex protected for writes. Sched-RCU protected for reads.
133 *
134 * WQ: wq->mutex protected.
135 *
136 * WR: wq->mutex protected for writes. Sched-RCU protected for reads.
137 *
138 * MD: wq_mayday_lock protected.
139 */
140
141/* struct worker is defined in workqueue_internal.h */
142
143struct worker_pool {
144    spinlock_t lock; /* the pool lock */
145    int cpu; /* I: the associated cpu */
146    int node; /* I: the associated node ID */
147    int id; /* I: pool ID */
148    unsigned int flags; /* X: flags */
149
150    struct list_head worklist; /* L: list of pending works */
151    int nr_workers; /* L: total number of workers */
152
153    /* nr_idle includes the ones off idle_list for rebinding */
154    int nr_idle; /* L: currently idle ones */
155
156    struct list_head idle_list; /* X: list of idle workers */
157    struct timer_list idle_timer; /* L: worker idle timeout */
158    struct timer_list mayday_timer; /* L: SOS timer for workers */
159
160    /* a workers is either on busy_hash or idle_list, or the manager */
161    DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
162                        /* L: hash of busy workers */
163
164    /* see manage_workers() for details on the two manager mutexes */
165    struct mutex manager_arb; /* manager arbitration */
166    struct mutex manager_mutex; /* manager exclusion */
167    struct idr worker_idr; /* MG: worker IDs and iteration */
168
169    struct workqueue_attrs *attrs; /* I: worker attributes */
170    struct hlist_node hash_node; /* PL: unbound_pool_hash node */
171    int refcnt; /* PL: refcnt for unbound pools */
172
173    /*
174     * The current concurrency level. As it's likely to be accessed
175     * from other CPUs during try_to_wake_up(), put it in a separate
176     * cacheline.
177     */
178    atomic_t nr_running ____cacheline_aligned_in_smp;
179
180    /*
181     * Destruction of pool is sched-RCU protected to allow dereferences
182     * from get_work_pool().
183     */
184    struct rcu_head rcu;
185} ____cacheline_aligned_in_smp;
186
187/*
188 * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS
189 * of work_struct->data are used for flags and the remaining high bits
190 * point to the pwq; thus, pwqs need to be aligned at two's power of the
191 * number of flag bits.
192 */
193struct pool_workqueue {
194    struct worker_pool *pool; /* I: the associated pool */
195    struct workqueue_struct *wq; /* I: the owning workqueue */
196    int work_color; /* L: current color */
197    int flush_color; /* L: flushing color */
198    int refcnt; /* L: reference count */
199    int nr_in_flight[WORK_NR_COLORS];
200                        /* L: nr of in_flight works */
201    int nr_active; /* L: nr of active works */
202    int max_active; /* L: max active works */
203    struct list_head delayed_works; /* L: delayed works */
204    struct list_head pwqs_node; /* WR: node on wq->pwqs */
205    struct list_head mayday_node; /* MD: node on wq->maydays */
206
207    /*
208     * Release of unbound pwq is punted to system_wq. See put_pwq()
209     * and pwq_unbound_release_workfn() for details. pool_workqueue
210     * itself is also sched-RCU protected so that the first pwq can be
211     * determined without grabbing wq->mutex.
212     */
213    struct work_struct unbound_release_work;
214    struct rcu_head rcu;
215} __aligned(1 << WORK_STRUCT_FLAG_BITS);
216
217/*
218 * Structure used to wait for workqueue flush.
219 */
220struct wq_flusher {
221    struct list_head list; /* WQ: list of flushers */
222    int flush_color; /* WQ: flush color waiting for */
223    struct completion done; /* flush completion */
224};
225
226struct wq_device;
227
228/*
229 * The externally visible workqueue. It relays the issued work items to
230 * the appropriate worker_pool through its pool_workqueues.
231 */
232struct workqueue_struct {
233    struct list_head pwqs; /* WR: all pwqs of this wq */
234    struct list_head list; /* PL: list of all workqueues */
235
236    struct mutex mutex; /* protects this wq */
237    int work_color; /* WQ: current work color */
238    int flush_color; /* WQ: current flush color */
239    atomic_t nr_pwqs_to_flush; /* flush in progress */
240    struct wq_flusher *first_flusher; /* WQ: first flusher */
241    struct list_head flusher_queue; /* WQ: flush waiters */
242    struct list_head flusher_overflow; /* WQ: flush overflow list */
243
244    struct list_head maydays; /* MD: pwqs requesting rescue */
245    struct worker *rescuer; /* I: rescue worker */
246
247    int nr_drainers; /* WQ: drain in progress */
248    int saved_max_active; /* WQ: saved pwq max_active */
249
250    struct workqueue_attrs *unbound_attrs; /* WQ: only for unbound wqs */
251    struct pool_workqueue *dfl_pwq; /* WQ: only for unbound wqs */
252
253#ifdef CONFIG_SYSFS
254    struct wq_device *wq_dev; /* I: for sysfs interface */
255#endif
256#ifdef CONFIG_LOCKDEP
257    struct lockdep_map lockdep_map;
258#endif
259    char name[WQ_NAME_LEN]; /* I: workqueue name */
260
261    /* hot fields used during command issue, aligned to cacheline */
262    unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */
263    struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
264    struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
265};
266
267static struct kmem_cache *pwq_cache;
268
269static int wq_numa_tbl_len; /* highest possible NUMA node id + 1 */
270static cpumask_var_t *wq_numa_possible_cpumask;
271                    /* possible CPUs of each node */
272
273static bool wq_disable_numa;
274module_param_named(disable_numa, wq_disable_numa, bool, 0444);
275
276/* see the comment above the definition of WQ_POWER_EFFICIENT */
277#ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
278static bool wq_power_efficient = true;
279#else
280static bool wq_power_efficient;
281#endif
282
283module_param_named(power_efficient, wq_power_efficient, bool, 0444);
284
285static bool wq_numa_enabled; /* unbound NUMA affinity enabled */
286
287/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
288static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
289
290static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */
291static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
292
293static LIST_HEAD(workqueues); /* PL: list of all workqueues */
294static bool workqueue_freezing; /* PL: have wqs started freezing? */
295
296/* the per-cpu worker pools */
297static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
298                     cpu_worker_pools);
299
300static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */
301
302/* PL: hash of all unbound pools keyed by pool->attrs */
303static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
304
305/* I: attributes used when instantiating standard unbound pools on demand */
306static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
307
308/* I: attributes used when instantiating ordered pools on demand */
309static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
310
311struct workqueue_struct *system_wq __read_mostly;
312EXPORT_SYMBOL(system_wq);
313struct workqueue_struct *system_highpri_wq __read_mostly;
314EXPORT_SYMBOL_GPL(system_highpri_wq);
315struct workqueue_struct *system_long_wq __read_mostly;
316EXPORT_SYMBOL_GPL(system_long_wq);
317struct workqueue_struct *system_unbound_wq __read_mostly;
318EXPORT_SYMBOL_GPL(system_unbound_wq);
319struct workqueue_struct *system_freezable_wq __read_mostly;
320EXPORT_SYMBOL_GPL(system_freezable_wq);
321struct workqueue_struct *system_power_efficient_wq __read_mostly;
322EXPORT_SYMBOL_GPL(system_power_efficient_wq);
323struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
324EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
325
326static int worker_thread(void *__worker);
327static void copy_workqueue_attrs(struct workqueue_attrs *to,
328                 const struct workqueue_attrs *from);
329
330#define CREATE_TRACE_POINTS
331#include <trace/events/workqueue.h>
332
333#define assert_rcu_or_pool_mutex() \
334    rcu_lockdep_assert(rcu_read_lock_sched_held() || \
335               lockdep_is_held(&wq_pool_mutex), \
336               "sched RCU or wq_pool_mutex should be held")
337
338#define assert_rcu_or_wq_mutex(wq) \
339    rcu_lockdep_assert(rcu_read_lock_sched_held() || \
340               lockdep_is_held(&wq->mutex), \
341               "sched RCU or wq->mutex should be held")
342
343#ifdef CONFIG_LOCKDEP
344#define assert_manager_or_pool_lock(pool) \
345    WARN_ONCE(debug_locks && \
346          !lockdep_is_held(&(pool)->manager_mutex) && \
347          !lockdep_is_held(&(pool)->lock), \
348          "pool->manager_mutex or ->lock should be held")
349#else
350#define assert_manager_or_pool_lock(pool) do { } while (0)
351#endif
352
353#define for_each_cpu_worker_pool(pool, cpu) \
354    for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \
355         (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
356         (pool)++)
357
358/**
359 * for_each_pool - iterate through all worker_pools in the system
360 * @pool: iteration cursor
361 * @pi: integer used for iteration
362 *
363 * This must be called either with wq_pool_mutex held or sched RCU read
364 * locked. If the pool needs to be used beyond the locking in effect, the
365 * caller is responsible for guaranteeing that the pool stays online.
366 *
367 * The if/else clause exists only for the lockdep assertion and can be
368 * ignored.
369 */
370#define for_each_pool(pool, pi) \
371    idr_for_each_entry(&worker_pool_idr, pool, pi) \
372        if (({ assert_rcu_or_pool_mutex(); false; })) { } \
373        else
374
375/**
376 * for_each_pool_worker - iterate through all workers of a worker_pool
377 * @worker: iteration cursor
378 * @wi: integer used for iteration
379 * @pool: worker_pool to iterate workers of
380 *
381 * This must be called with either @pool->manager_mutex or ->lock held.
382 *
383 * The if/else clause exists only for the lockdep assertion and can be
384 * ignored.
385 */
386#define for_each_pool_worker(worker, wi, pool) \
387    idr_for_each_entry(&(pool)->worker_idr, (worker), (wi)) \
388        if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
389        else
390
391/**
392 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
393 * @pwq: iteration cursor
394 * @wq: the target workqueue
395 *
396 * This must be called either with wq->mutex held or sched RCU read locked.
397 * If the pwq needs to be used beyond the locking in effect, the caller is
398 * responsible for guaranteeing that the pwq stays online.
399 *
400 * The if/else clause exists only for the lockdep assertion and can be
401 * ignored.
402 */
403#define for_each_pwq(pwq, wq) \
404    list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node) \
405        if (({ assert_rcu_or_wq_mutex(wq); false; })) { } \
406        else
407
408#ifdef CONFIG_DEBUG_OBJECTS_WORK
409
410static struct debug_obj_descr work_debug_descr;
411
412static void *work_debug_hint(void *addr)
413{
414    return ((struct work_struct *) addr)->func;
415}
416
417/*
418 * fixup_init is called when:
419 * - an active object is initialized
420 */
421static int work_fixup_init(void *addr, enum debug_obj_state state)
422{
423    struct work_struct *work = addr;
424
425    switch (state) {
426    case ODEBUG_STATE_ACTIVE:
427        cancel_work_sync(work);
428        debug_object_init(work, &work_debug_descr);
429        return 1;
430    default:
431        return 0;
432    }
433}
434
435/*
436 * fixup_activate is called when:
437 * - an active object is activated
438 * - an unknown object is activated (might be a statically initialized object)
439 */
440static int work_fixup_activate(void *addr, enum debug_obj_state state)
441{
442    struct work_struct *work = addr;
443
444    switch (state) {
445
446    case ODEBUG_STATE_NOTAVAILABLE:
447        /*
448         * This is not really a fixup. The work struct was
449         * statically initialized. We just make sure that it
450         * is tracked in the object tracker.
451         */
452        if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
453            debug_object_init(work, &work_debug_descr);
454            debug_object_activate(work, &work_debug_descr);
455            return 0;
456        }
457        WARN_ON_ONCE(1);
458        return 0;
459
460    case ODEBUG_STATE_ACTIVE:
461        WARN_ON(1);
462
463    default:
464        return 0;
465    }
466}
467
468/*
469 * fixup_free is called when:
470 * - an active object is freed
471 */
472static int work_fixup_free(void *addr, enum debug_obj_state state)
473{
474    struct work_struct *work = addr;
475
476    switch (state) {
477    case ODEBUG_STATE_ACTIVE:
478        cancel_work_sync(work);
479        debug_object_free(work, &work_debug_descr);
480        return 1;
481    default:
482        return 0;
483    }
484}
485
486static struct debug_obj_descr work_debug_descr = {
487    .name = "work_struct",
488    .debug_hint = work_debug_hint,
489    .fixup_init = work_fixup_init,
490    .fixup_activate = work_fixup_activate,
491    .fixup_free = work_fixup_free,
492};
493
494static inline void debug_work_activate(struct work_struct *work)
495{
496    debug_object_activate(work, &work_debug_descr);
497}
498
499static inline void debug_work_deactivate(struct work_struct *work)
500{
501    debug_object_deactivate(work, &work_debug_descr);
502}
503
504void __init_work(struct work_struct *work, int onstack)
505{
506    if (onstack)
507        debug_object_init_on_stack(work, &work_debug_descr);
508    else
509        debug_object_init(work, &work_debug_descr);
510}
511EXPORT_SYMBOL_GPL(__init_work);
512
513void destroy_work_on_stack(struct work_struct *work)
514{
515    debug_object_free(work, &work_debug_descr);
516}
517EXPORT_SYMBOL_GPL(destroy_work_on_stack);
518
519void destroy_delayed_work_on_stack(struct delayed_work *work)
520{
521    destroy_timer_on_stack(&work->timer);
522    debug_object_free(&work->work, &work_debug_descr);
523}
524EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
525
526#else
527static inline void debug_work_activate(struct work_struct *work) { }
528static inline void debug_work_deactivate(struct work_struct *work) { }
529#endif
530
531/**
532 * worker_pool_assign_id - allocate ID and assing it to @pool
533 * @pool: the pool pointer of interest
534 *
535 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
536 * successfully, -errno on failure.
537 */
538static int worker_pool_assign_id(struct worker_pool *pool)
539{
540    int ret;
541
542    lockdep_assert_held(&wq_pool_mutex);
543
544    ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
545            GFP_KERNEL);
546    if (ret >= 0) {
547        pool->id = ret;
548        return 0;
549    }
550    return ret;
551}
552
553/**
554 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
555 * @wq: the target workqueue
556 * @node: the node ID
557 *
558 * This must be called either with pwq_lock held or sched RCU read locked.
559 * If the pwq needs to be used beyond the locking in effect, the caller is
560 * responsible for guaranteeing that the pwq stays online.
561 *
562 * Return: The unbound pool_workqueue for @node.
563 */
564static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
565                          int node)
566{
567    assert_rcu_or_wq_mutex(wq);
568    return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
569}
570
571static unsigned int work_color_to_flags(int color)
572{
573    return color << WORK_STRUCT_COLOR_SHIFT;
574}
575
576static int get_work_color(struct work_struct *work)
577{
578    return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
579        ((1 << WORK_STRUCT_COLOR_BITS) - 1);
580}
581
582static int work_next_color(int color)
583{
584    return (color + 1) % WORK_NR_COLORS;
585}
586
587/*
588 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
589 * contain the pointer to the queued pwq. Once execution starts, the flag
590 * is cleared and the high bits contain OFFQ flags and pool ID.
591 *
592 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
593 * and clear_work_data() can be used to set the pwq, pool or clear
594 * work->data. These functions should only be called while the work is
595 * owned - ie. while the PENDING bit is set.
596 *
597 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
598 * corresponding to a work. Pool is available once the work has been
599 * queued anywhere after initialization until it is sync canceled. pwq is
600 * available only while the work item is queued.
601 *
602 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
603 * canceled. While being canceled, a work item may have its PENDING set
604 * but stay off timer and worklist for arbitrarily long and nobody should
605 * try to steal the PENDING bit.
606 */
607static inline void set_work_data(struct work_struct *work, unsigned long data,
608                 unsigned long flags)
609{
610    WARN_ON_ONCE(!work_pending(work));
611    atomic_long_set(&work->data, data | flags | work_static(work));
612}
613
614static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
615             unsigned long extra_flags)
616{
617    set_work_data(work, (unsigned long)pwq,
618              WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
619}
620
621static void set_work_pool_and_keep_pending(struct work_struct *work,
622                       int pool_id)
623{
624    set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
625              WORK_STRUCT_PENDING);
626}
627
628static void set_work_pool_and_clear_pending(struct work_struct *work,
629                        int pool_id)
630{
631    /*
632     * The following wmb is paired with the implied mb in
633     * test_and_set_bit(PENDING) and ensures all updates to @work made
634     * here are visible to and precede any updates by the next PENDING
635     * owner.
636     */
637    smp_wmb();
638    set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
639}
640
641static void clear_work_data(struct work_struct *work)
642{
643    smp_wmb(); /* see set_work_pool_and_clear_pending() */
644    set_work_data(work, WORK_STRUCT_NO_POOL, 0);
645}
646
647static struct pool_workqueue *get_work_pwq(struct work_struct *work)
648{
649    unsigned long data = atomic_long_read(&work->data);
650
651    if (data & WORK_STRUCT_PWQ)
652        return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
653    else
654        return NULL;
655}
656
657/**
658 * get_work_pool - return the worker_pool a given work was associated with
659 * @work: the work item of interest
660 *
661 * Pools are created and destroyed under wq_pool_mutex, and allows read
662 * access under sched-RCU read lock. As such, this function should be
663 * called under wq_pool_mutex or with preemption disabled.
664 *
665 * All fields of the returned pool are accessible as long as the above
666 * mentioned locking is in effect. If the returned pool needs to be used
667 * beyond the critical section, the caller is responsible for ensuring the
668 * returned pool is and stays online.
669 *
670 * Return: The worker_pool @work was last associated with. %NULL if none.
671 */
672static struct worker_pool *get_work_pool(struct work_struct *work)
673{
674    unsigned long data = atomic_long_read(&work->data);
675    int pool_id;
676
677    assert_rcu_or_pool_mutex();
678
679    if (data & WORK_STRUCT_PWQ)
680        return ((struct pool_workqueue *)
681            (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
682
683    pool_id = data >> WORK_OFFQ_POOL_SHIFT;
684    if (pool_id == WORK_OFFQ_POOL_NONE)
685        return NULL;
686
687    return idr_find(&worker_pool_idr, pool_id);
688}
689
690/**
691 * get_work_pool_id - return the worker pool ID a given work is associated with
692 * @work: the work item of interest
693 *
694 * Return: The worker_pool ID @work was last associated with.
695 * %WORK_OFFQ_POOL_NONE if none.
696 */
697static int get_work_pool_id(struct work_struct *work)
698{
699    unsigned long data = atomic_long_read(&work->data);
700
701    if (data & WORK_STRUCT_PWQ)
702        return ((struct pool_workqueue *)
703            (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
704
705    return data >> WORK_OFFQ_POOL_SHIFT;
706}
707
708static void mark_work_canceling(struct work_struct *work)
709{
710    unsigned long pool_id = get_work_pool_id(work);
711
712    pool_id <<= WORK_OFFQ_POOL_SHIFT;
713    set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
714}
715
716static bool work_is_canceling(struct work_struct *work)
717{
718    unsigned long data = atomic_long_read(&work->data);
719
720    return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
721}
722
723/*
724 * Policy functions. These define the policies on how the global worker
725 * pools are managed. Unless noted otherwise, these functions assume that
726 * they're being called with pool->lock held.
727 */
728
729static bool __need_more_worker(struct worker_pool *pool)
730{
731    return !atomic_read(&pool->nr_running);
732}
733
734/*
735 * Need to wake up a worker? Called from anything but currently
736 * running workers.
737 *
738 * Note that, because unbound workers never contribute to nr_running, this
739 * function will always return %true for unbound pools as long as the
740 * worklist isn't empty.
741 */
742static bool need_more_worker(struct worker_pool *pool)
743{
744    return !list_empty(&pool->worklist) && __need_more_worker(pool);
745}
746
747/* Can I start working? Called from busy but !running workers. */
748static bool may_start_working(struct worker_pool *pool)
749{
750    return pool->nr_idle;
751}
752
753/* Do I need to keep working? Called from currently running workers. */
754static bool keep_working(struct worker_pool *pool)
755{
756    return !list_empty(&pool->worklist) &&
757        atomic_read(&pool->nr_running) <= 1;
758}
759
760/* Do we need a new worker? Called from manager. */
761static bool need_to_create_worker(struct worker_pool *pool)
762{
763    return need_more_worker(pool) && !may_start_working(pool);
764}
765
766/* Do I need to be the manager? */
767static bool need_to_manage_workers(struct worker_pool *pool)
768{
769    return need_to_create_worker(pool) ||
770        (pool->flags & POOL_MANAGE_WORKERS);
771}
772
773/* Do we have too many workers and should some go away? */
774static bool too_many_workers(struct worker_pool *pool)
775{
776    bool managing = mutex_is_locked(&pool->manager_arb);
777    int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
778    int nr_busy = pool->nr_workers - nr_idle;
779
780    /*
781     * nr_idle and idle_list may disagree if idle rebinding is in
782     * progress. Never return %true if idle_list is empty.
783     */
784    if (list_empty(&pool->idle_list))
785        return false;
786
787    return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
788}
789
790/*
791 * Wake up functions.
792 */
793
794/* Return the first worker. Safe with preemption disabled */
795static struct worker *first_worker(struct worker_pool *pool)
796{
797    if (unlikely(list_empty(&pool->idle_list)))
798        return NULL;
799
800    return list_first_entry(&pool->idle_list, struct worker, entry);
801}
802
803/**
804 * wake_up_worker - wake up an idle worker
805 * @pool: worker pool to wake worker from
806 *
807 * Wake up the first idle worker of @pool.
808 *
809 * CONTEXT:
810 * spin_lock_irq(pool->lock).
811 */
812static void wake_up_worker(struct worker_pool *pool)
813{
814    struct worker *worker = first_worker(pool);
815
816    if (likely(worker))
817        wake_up_process(worker->task);
818}
819
820/**
821 * wq_worker_waking_up - a worker is waking up
822 * @task: task waking up
823 * @cpu: CPU @task is waking up to
824 *
825 * This function is called during try_to_wake_up() when a worker is
826 * being awoken.
827 *
828 * CONTEXT:
829 * spin_lock_irq(rq->lock)
830 */
831void wq_worker_waking_up(struct task_struct *task, int cpu)
832{
833    struct worker *worker = kthread_data(task);
834
835    if (!(worker->flags & WORKER_NOT_RUNNING)) {
836        WARN_ON_ONCE(worker->pool->cpu != cpu);
837        atomic_inc(&worker->pool->nr_running);
838    }
839}
840
841/**
842 * wq_worker_sleeping - a worker is going to sleep
843 * @task: task going to sleep
844 * @cpu: CPU in question, must be the current CPU number
845 *
846 * This function is called during schedule() when a busy worker is
847 * going to sleep. Worker on the same cpu can be woken up by
848 * returning pointer to its task.
849 *
850 * CONTEXT:
851 * spin_lock_irq(rq->lock)
852 *
853 * Return:
854 * Worker task on @cpu to wake up, %NULL if none.
855 */
856struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
857{
858    struct worker *worker = kthread_data(task), *to_wakeup = NULL;
859    struct worker_pool *pool;
860
861    /*
862     * Rescuers, which may not have all the fields set up like normal
863     * workers, also reach here, let's not access anything before
864     * checking NOT_RUNNING.
865     */
866    if (worker->flags & WORKER_NOT_RUNNING)
867        return NULL;
868
869    pool = worker->pool;
870
871    /* this can only happen on the local cpu */
872    if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
873        return NULL;
874
875    /*
876     * The counterpart of the following dec_and_test, implied mb,
877     * worklist not empty test sequence is in insert_work().
878     * Please read comment there.
879     *
880     * NOT_RUNNING is clear. This means that we're bound to and
881     * running on the local cpu w/ rq lock held and preemption
882     * disabled, which in turn means that none else could be
883     * manipulating idle_list, so dereferencing idle_list without pool
884     * lock is safe.
885     */
886    if (atomic_dec_and_test(&pool->nr_running) &&
887        !list_empty(&pool->worklist))
888        to_wakeup = first_worker(pool);
889    return to_wakeup ? to_wakeup->task : NULL;
890}
891
892/**
893 * worker_set_flags - set worker flags and adjust nr_running accordingly
894 * @worker: self
895 * @flags: flags to set
896 * @wakeup: wakeup an idle worker if necessary
897 *
898 * Set @flags in @worker->flags and adjust nr_running accordingly. If
899 * nr_running becomes zero and @wakeup is %true, an idle worker is
900 * woken up.
901 *
902 * CONTEXT:
903 * spin_lock_irq(pool->lock)
904 */
905static inline void worker_set_flags(struct worker *worker, unsigned int flags,
906                    bool wakeup)
907{
908    struct worker_pool *pool = worker->pool;
909
910    WARN_ON_ONCE(worker->task != current);
911
912    /*
913     * If transitioning into NOT_RUNNING, adjust nr_running and
914     * wake up an idle worker as necessary if requested by
915     * @wakeup.
916     */
917    if ((flags & WORKER_NOT_RUNNING) &&
918        !(worker->flags & WORKER_NOT_RUNNING)) {
919        if (wakeup) {
920            if (atomic_dec_and_test(&pool->nr_running) &&
921                !list_empty(&pool->worklist))
922                wake_up_worker(pool);
923        } else
924            atomic_dec(&pool->nr_running);
925    }
926
927    worker->flags |= flags;
928}
929
930/**
931 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
932 * @worker: self
933 * @flags: flags to clear
934 *
935 * Clear @flags in @worker->flags and adjust nr_running accordingly.
936 *
937 * CONTEXT:
938 * spin_lock_irq(pool->lock)
939 */
940static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
941{
942    struct worker_pool *pool = worker->pool;
943    unsigned int oflags = worker->flags;
944
945    WARN_ON_ONCE(worker->task != current);
946
947    worker->flags &= ~flags;
948
949    /*
950     * If transitioning out of NOT_RUNNING, increment nr_running. Note
951     * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
952     * of multiple flags, not a single flag.
953     */
954    if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
955        if (!(worker->flags & WORKER_NOT_RUNNING))
956            atomic_inc(&pool->nr_running);
957}
958
959/**
960 * find_worker_executing_work - find worker which is executing a work
961 * @pool: pool of interest
962 * @work: work to find worker for
963 *
964 * Find a worker which is executing @work on @pool by searching
965 * @pool->busy_hash which is keyed by the address of @work. For a worker
966 * to match, its current execution should match the address of @work and
967 * its work function. This is to avoid unwanted dependency between
968 * unrelated work executions through a work item being recycled while still
969 * being executed.
970 *
971 * This is a bit tricky. A work item may be freed once its execution
972 * starts and nothing prevents the freed area from being recycled for
973 * another work item. If the same work item address ends up being reused
974 * before the original execution finishes, workqueue will identify the
975 * recycled work item as currently executing and make it wait until the
976 * current execution finishes, introducing an unwanted dependency.
977 *
978 * This function checks the work item address and work function to avoid
979 * false positives. Note that this isn't complete as one may construct a
980 * work function which can introduce dependency onto itself through a
981 * recycled work item. Well, if somebody wants to shoot oneself in the
982 * foot that badly, there's only so much we can do, and if such deadlock
983 * actually occurs, it should be easy to locate the culprit work function.
984 *
985 * CONTEXT:
986 * spin_lock_irq(pool->lock).
987 *
988 * Return:
989 * Pointer to worker which is executing @work if found, %NULL
990 * otherwise.
991 */
992static struct worker *find_worker_executing_work(struct worker_pool *pool,
993                         struct work_struct *work)
994{
995    struct worker *worker;
996
997    hash_for_each_possible(pool->busy_hash, worker, hentry,
998                   (unsigned long)work)
999        if (worker->current_work == work &&
1000            worker->current_func == work->func)
1001            return worker;
1002
1003    return NULL;
1004}
1005
1006/**
1007 * move_linked_works - move linked works to a list
1008 * @work: start of series of works to be scheduled
1009 * @head: target list to append @work to
1010 * @nextp: out paramter for nested worklist walking
1011 *
1012 * Schedule linked works starting from @work to @head. Work series to
1013 * be scheduled starts at @work and includes any consecutive work with
1014 * WORK_STRUCT_LINKED set in its predecessor.
1015 *
1016 * If @nextp is not NULL, it's updated to point to the next work of
1017 * the last scheduled work. This allows move_linked_works() to be
1018 * nested inside outer list_for_each_entry_safe().
1019 *
1020 * CONTEXT:
1021 * spin_lock_irq(pool->lock).
1022 */
1023static void move_linked_works(struct work_struct *work, struct list_head *head,
1024                  struct work_struct **nextp)
1025{
1026    struct work_struct *n;
1027
1028    /*
1029     * Linked worklist will always end before the end of the list,
1030     * use NULL for list head.
1031     */
1032    list_for_each_entry_safe_from(work, n, NULL, entry) {
1033        list_move_tail(&work->entry, head);
1034        if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1035            break;
1036    }
1037
1038    /*
1039     * If we're already inside safe list traversal and have moved
1040     * multiple works to the scheduled queue, the next position
1041     * needs to be updated.
1042     */
1043    if (nextp)
1044        *nextp = n;
1045}
1046
1047/**
1048 * get_pwq - get an extra reference on the specified pool_workqueue
1049 * @pwq: pool_workqueue to get
1050 *
1051 * Obtain an extra reference on @pwq. The caller should guarantee that
1052 * @pwq has positive refcnt and be holding the matching pool->lock.
1053 */
1054static void get_pwq(struct pool_workqueue *pwq)
1055{
1056    lockdep_assert_held(&pwq->pool->lock);
1057    WARN_ON_ONCE(pwq->refcnt <= 0);
1058    pwq->refcnt++;
1059}
1060
1061/**
1062 * put_pwq - put a pool_workqueue reference
1063 * @pwq: pool_workqueue to put
1064 *
1065 * Drop a reference of @pwq. If its refcnt reaches zero, schedule its
1066 * destruction. The caller should be holding the matching pool->lock.
1067 */
1068static void put_pwq(struct pool_workqueue *pwq)
1069{
1070    lockdep_assert_held(&pwq->pool->lock);
1071    if (likely(--pwq->refcnt))
1072        return;
1073    if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1074        return;
1075    /*
1076     * @pwq can't be released under pool->lock, bounce to
1077     * pwq_unbound_release_workfn(). This never recurses on the same
1078     * pool->lock as this path is taken only for unbound workqueues and
1079     * the release work item is scheduled on a per-cpu workqueue. To
1080     * avoid lockdep warning, unbound pool->locks are given lockdep
1081     * subclass of 1 in get_unbound_pool().
1082     */
1083    schedule_work(&pwq->unbound_release_work);
1084}
1085
1086/**
1087 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1088 * @pwq: pool_workqueue to put (can be %NULL)
1089 *
1090 * put_pwq() with locking. This function also allows %NULL @pwq.
1091 */
1092static void put_pwq_unlocked(struct pool_workqueue *pwq)
1093{
1094    if (pwq) {
1095        /*
1096         * As both pwqs and pools are sched-RCU protected, the
1097         * following lock operations are safe.
1098         */
1099        spin_lock_irq(&pwq->pool->lock);
1100        put_pwq(pwq);
1101        spin_unlock_irq(&pwq->pool->lock);
1102    }
1103}
1104
1105static void pwq_activate_delayed_work(struct work_struct *work)
1106{
1107    struct pool_workqueue *pwq = get_work_pwq(work);
1108
1109    trace_workqueue_activate_work(work);
1110    move_linked_works(work, &pwq->pool->worklist, NULL);
1111    __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1112    pwq->nr_active++;
1113}
1114
1115static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1116{
1117    struct work_struct *work = list_first_entry(&pwq->delayed_works,
1118                            struct work_struct, entry);
1119
1120    pwq_activate_delayed_work(work);
1121}
1122
1123/**
1124 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1125 * @pwq: pwq of interest
1126 * @color: color of work which left the queue
1127 *
1128 * A work either has completed or is removed from pending queue,
1129 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1130 *
1131 * CONTEXT:
1132 * spin_lock_irq(pool->lock).
1133 */
1134static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1135{
1136    /* uncolored work items don't participate in flushing or nr_active */
1137    if (color == WORK_NO_COLOR)
1138        goto out_put;
1139
1140    pwq->nr_in_flight[color]--;
1141
1142    pwq->nr_active--;
1143    if (!list_empty(&pwq->delayed_works)) {
1144        /* one down, submit a delayed one */
1145        if (pwq->nr_active < pwq->max_active)
1146            pwq_activate_first_delayed(pwq);
1147    }
1148
1149    /* is flush in progress and are we at the flushing tip? */
1150    if (likely(pwq->flush_color != color))
1151        goto out_put;
1152
1153    /* are there still in-flight works? */
1154    if (pwq->nr_in_flight[color])
1155        goto out_put;
1156
1157    /* this pwq is done, clear flush_color */
1158    pwq->flush_color = -1;
1159
1160    /*
1161     * If this was the last pwq, wake up the first flusher. It
1162     * will handle the rest.
1163     */
1164    if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1165        complete(&pwq->wq->first_flusher->done);
1166out_put:
1167    put_pwq(pwq);
1168}
1169
1170/**
1171 * try_to_grab_pending - steal work item from worklist and disable irq
1172 * @work: work item to steal
1173 * @is_dwork: @work is a delayed_work
1174 * @flags: place to store irq state
1175 *
1176 * Try to grab PENDING bit of @work. This function can handle @work in any
1177 * stable state - idle, on timer or on worklist.
1178 *
1179 * Return:
1180 * 1 if @work was pending and we successfully stole PENDING
1181 * 0 if @work was idle and we claimed PENDING
1182 * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry
1183 * -ENOENT if someone else is canceling @work, this state may persist
1184 * for arbitrarily long
1185 *
1186 * Note:
1187 * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting
1188 * interrupted while holding PENDING and @work off queue, irq must be
1189 * disabled on entry. This, combined with delayed_work->timer being
1190 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1191 *
1192 * On successful return, >= 0, irq is disabled and the caller is
1193 * responsible for releasing it using local_irq_restore(*@flags).
1194 *
1195 * This function is safe to call from any context including IRQ handler.
1196 */
1197static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1198                   unsigned long *flags)
1199{
1200    struct worker_pool *pool;
1201    struct pool_workqueue *pwq;
1202
1203    local_irq_save(*flags);
1204
1205    /* try to steal the timer if it exists */
1206    if (is_dwork) {
1207        struct delayed_work *dwork = to_delayed_work(work);
1208
1209        /*
1210         * dwork->timer is irqsafe. If del_timer() fails, it's
1211         * guaranteed that the timer is not queued anywhere and not
1212         * running on the local CPU.
1213         */
1214        if (likely(del_timer(&dwork->timer)))
1215            return 1;
1216    }
1217
1218    /* try to claim PENDING the normal way */
1219    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1220        return 0;
1221
1222    /*
1223     * The queueing is in progress, or it is already queued. Try to
1224     * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1225     */
1226    pool = get_work_pool(work);
1227    if (!pool)
1228        goto fail;
1229
1230    spin_lock(&pool->lock);
1231    /*
1232     * work->data is guaranteed to point to pwq only while the work
1233     * item is queued on pwq->wq, and both updating work->data to point
1234     * to pwq on queueing and to pool on dequeueing are done under
1235     * pwq->pool->lock. This in turn guarantees that, if work->data
1236     * points to pwq which is associated with a locked pool, the work
1237     * item is currently queued on that pool.
1238     */
1239    pwq = get_work_pwq(work);
1240    if (pwq && pwq->pool == pool) {
1241        debug_work_deactivate(work);
1242
1243        /*
1244         * A delayed work item cannot be grabbed directly because
1245         * it might have linked NO_COLOR work items which, if left
1246         * on the delayed_list, will confuse pwq->nr_active
1247         * management later on and cause stall. Make sure the work
1248         * item is activated before grabbing.
1249         */
1250        if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1251            pwq_activate_delayed_work(work);
1252
1253        list_del_init(&work->entry);
1254        pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1255
1256        /* work->data points to pwq iff queued, point to pool */
1257        set_work_pool_and_keep_pending(work, pool->id);
1258
1259        spin_unlock(&pool->lock);
1260        return 1;
1261    }
1262    spin_unlock(&pool->lock);
1263fail:
1264    local_irq_restore(*flags);
1265    if (work_is_canceling(work))
1266        return -ENOENT;
1267    cpu_relax();
1268    return -EAGAIN;
1269}
1270
1271/**
1272 * insert_work - insert a work into a pool
1273 * @pwq: pwq @work belongs to
1274 * @work: work to insert
1275 * @head: insertion point
1276 * @extra_flags: extra WORK_STRUCT_* flags to set
1277 *
1278 * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to
1279 * work_struct flags.
1280 *
1281 * CONTEXT:
1282 * spin_lock_irq(pool->lock).
1283 */
1284static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1285            struct list_head *head, unsigned int extra_flags)
1286{
1287    struct worker_pool *pool = pwq->pool;
1288
1289    /* we own @work, set data and link */
1290    set_work_pwq(work, pwq, extra_flags);
1291    list_add_tail(&work->entry, head);
1292    get_pwq(pwq);
1293
1294    /*
1295     * Ensure either wq_worker_sleeping() sees the above
1296     * list_add_tail() or we see zero nr_running to avoid workers lying
1297     * around lazily while there are works to be processed.
1298     */
1299    smp_mb();
1300
1301    if (__need_more_worker(pool))
1302        wake_up_worker(pool);
1303}
1304
1305/*
1306 * Test whether @work is being queued from another work executing on the
1307 * same workqueue.
1308 */
1309static bool is_chained_work(struct workqueue_struct *wq)
1310{
1311    struct worker *worker;
1312
1313    worker = current_wq_worker();
1314    /*
1315     * Return %true iff I'm a worker execuing a work item on @wq. If
1316     * I'm @worker, it's safe to dereference it without locking.
1317     */
1318    return worker && worker->current_pwq->wq == wq;
1319}
1320
1321static void __queue_work(int cpu, struct workqueue_struct *wq,
1322             struct work_struct *work)
1323{
1324    struct pool_workqueue *pwq;
1325    struct worker_pool *last_pool;
1326    struct list_head *worklist;
1327    unsigned int work_flags;
1328    unsigned int req_cpu = cpu;
1329
1330    /*
1331     * While a work item is PENDING && off queue, a task trying to
1332     * steal the PENDING will busy-loop waiting for it to either get
1333     * queued or lose PENDING. Grabbing PENDING and queueing should
1334     * happen with IRQ disabled.
1335     */
1336    WARN_ON_ONCE(!irqs_disabled());
1337
1338    debug_work_activate(work);
1339
1340    /* if draining, only works from the same workqueue are allowed */
1341    if (unlikely(wq->flags & __WQ_DRAINING) &&
1342        WARN_ON_ONCE(!is_chained_work(wq)))
1343        return;
1344retry:
1345    if (req_cpu == WORK_CPU_UNBOUND)
1346        cpu = raw_smp_processor_id();
1347
1348    /* pwq which will be used unless @work is executing elsewhere */
1349    if (!(wq->flags & WQ_UNBOUND))
1350        pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1351    else
1352        pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1353
1354    /*
1355     * If @work was previously on a different pool, it might still be
1356     * running there, in which case the work needs to be queued on that
1357     * pool to guarantee non-reentrancy.
1358     */
1359    last_pool = get_work_pool(work);
1360    if (last_pool && last_pool != pwq->pool) {
1361        struct worker *worker;
1362
1363        spin_lock(&last_pool->lock);
1364
1365        worker = find_worker_executing_work(last_pool, work);
1366
1367        if (worker && worker->current_pwq->wq == wq) {
1368            pwq = worker->current_pwq;
1369        } else {
1370            /* meh... not running there, queue here */
1371            spin_unlock(&last_pool->lock);
1372            spin_lock(&pwq->pool->lock);
1373        }
1374    } else {
1375        spin_lock(&pwq->pool->lock);
1376    }
1377
1378    /*
1379     * pwq is determined and locked. For unbound pools, we could have
1380     * raced with pwq release and it could already be dead. If its
1381     * refcnt is zero, repeat pwq selection. Note that pwqs never die
1382     * without another pwq replacing it in the numa_pwq_tbl or while
1383     * work items are executing on it, so the retrying is guaranteed to
1384     * make forward-progress.
1385     */
1386    if (unlikely(!pwq->refcnt)) {
1387        if (wq->flags & WQ_UNBOUND) {
1388            spin_unlock(&pwq->pool->lock);
1389            cpu_relax();
1390            goto retry;
1391        }
1392        /* oops */
1393        WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1394              wq->name, cpu);
1395    }
1396
1397    /* pwq determined, queue */
1398    trace_workqueue_queue_work(req_cpu, pwq, work);
1399
1400    if (WARN_ON(!list_empty(&work->entry))) {
1401        spin_unlock(&pwq->pool->lock);
1402        return;
1403    }
1404
1405    pwq->nr_in_flight[pwq->work_color]++;
1406    work_flags = work_color_to_flags(pwq->work_color);
1407
1408    if (likely(pwq->nr_active < pwq->max_active)) {
1409        trace_workqueue_activate_work(work);
1410        pwq->nr_active++;
1411        worklist = &pwq->pool->worklist;
1412    } else {
1413        work_flags |= WORK_STRUCT_DELAYED;
1414        worklist = &pwq->delayed_works;
1415    }
1416
1417    insert_work(pwq, work, worklist, work_flags);
1418
1419    spin_unlock(&pwq->pool->lock);
1420}
1421
1422/**
1423 * queue_work_on - queue work on specific cpu
1424 * @cpu: CPU number to execute work on
1425 * @wq: workqueue to use
1426 * @work: work to queue
1427 *
1428 * We queue the work to a specific CPU, the caller must ensure it
1429 * can't go away.
1430 *
1431 * Return: %false if @work was already on a queue, %true otherwise.
1432 */
1433bool queue_work_on(int cpu, struct workqueue_struct *wq,
1434           struct work_struct *work)
1435{
1436    bool ret = false;
1437    unsigned long flags;
1438
1439    local_irq_save(flags);
1440
1441    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1442        __queue_work(cpu, wq, work);
1443        ret = true;
1444    }
1445
1446    local_irq_restore(flags);
1447    return ret;
1448}
1449EXPORT_SYMBOL(queue_work_on);
1450
1451void delayed_work_timer_fn(unsigned long __data)
1452{
1453    struct delayed_work *dwork = (struct delayed_work *)__data;
1454
1455    /* should have been called from irqsafe timer with irq already off */
1456    __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1457}
1458EXPORT_SYMBOL(delayed_work_timer_fn);
1459
1460static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1461                struct delayed_work *dwork, unsigned long delay)
1462{
1463    struct timer_list *timer = &dwork->timer;
1464    struct work_struct *work = &dwork->work;
1465
1466    WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1467             timer->data != (unsigned long)dwork);
1468    WARN_ON_ONCE(timer_pending(timer));
1469    WARN_ON_ONCE(!list_empty(&work->entry));
1470
1471    /*
1472     * If @delay is 0, queue @dwork->work immediately. This is for
1473     * both optimization and correctness. The earliest @timer can
1474     * expire is on the closest next tick and delayed_work users depend
1475     * on that there's no such delay when @delay is 0.
1476     */
1477    if (!delay) {
1478        __queue_work(cpu, wq, &dwork->work);
1479        return;
1480    }
1481
1482    timer_stats_timer_set_start_info(&dwork->timer);
1483
1484    dwork->wq = wq;
1485    dwork->cpu = cpu;
1486    timer->expires = jiffies + delay;
1487
1488    if (unlikely(cpu != WORK_CPU_UNBOUND))
1489        add_timer_on(timer, cpu);
1490    else
1491        add_timer(timer);
1492}
1493
1494/**
1495 * queue_delayed_work_on - queue work on specific CPU after delay
1496 * @cpu: CPU number to execute work on
1497 * @wq: workqueue to use
1498 * @dwork: work to queue
1499 * @delay: number of jiffies to wait before queueing
1500 *
1501 * Return: %false if @work was already on a queue, %true otherwise. If
1502 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1503 * execution.
1504 */
1505bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1506               struct delayed_work *dwork, unsigned long delay)
1507{
1508    struct work_struct *work = &dwork->work;
1509    bool ret = false;
1510    unsigned long flags;
1511
1512    /* read the comment in __queue_work() */
1513    local_irq_save(flags);
1514
1515    if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1516        __queue_delayed_work(cpu, wq, dwork, delay);
1517        ret = true;
1518    }
1519
1520    local_irq_restore(flags);
1521    return ret;
1522}
1523EXPORT_SYMBOL(queue_delayed_work_on);
1524
1525/**
1526 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1527 * @cpu: CPU number to execute work on
1528 * @wq: workqueue to use
1529 * @dwork: work to queue
1530 * @delay: number of jiffies to wait before queueing
1531 *
1532 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1533 * modify @dwork's timer so that it expires after @delay. If @delay is
1534 * zero, @work is guaranteed to be scheduled immediately regardless of its
1535 * current state.
1536 *
1537 * Return: %false if @dwork was idle and queued, %true if @dwork was
1538 * pending and its timer was modified.
1539 *
1540 * This function is safe to call from any context including IRQ handler.
1541 * See try_to_grab_pending() for details.
1542 */
1543bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1544             struct delayed_work *dwork, unsigned long delay)
1545{
1546    unsigned long flags;
1547    int ret;
1548
1549    do {
1550        ret = try_to_grab_pending(&dwork->work, true, &flags);
1551    } while (unlikely(ret == -EAGAIN));
1552
1553    if (likely(ret >= 0)) {
1554        __queue_delayed_work(cpu, wq, dwork, delay);
1555        local_irq_restore(flags);
1556    }
1557
1558    /* -ENOENT from try_to_grab_pending() becomes %true */
1559    return ret;
1560}
1561EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1562
1563/**
1564 * worker_enter_idle - enter idle state
1565 * @worker: worker which is entering idle state
1566 *
1567 * @worker is entering idle state. Update stats and idle timer if
1568 * necessary.
1569 *
1570 * LOCKING:
1571 * spin_lock_irq(pool->lock).
1572 */
1573static void worker_enter_idle(struct worker *worker)
1574{
1575    struct worker_pool *pool = worker->pool;
1576
1577    if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1578        WARN_ON_ONCE(!list_empty(&worker->entry) &&
1579             (worker->hentry.next || worker->hentry.pprev)))
1580        return;
1581
1582    /* can't use worker_set_flags(), also called from start_worker() */
1583    worker->flags |= WORKER_IDLE;
1584    pool->nr_idle++;
1585    worker->last_active = jiffies;
1586
1587    /* idle_list is LIFO */
1588    list_add(&worker->entry, &pool->idle_list);
1589
1590    if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1591        mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1592
1593    /*
1594     * Sanity check nr_running. Because wq_unbind_fn() releases
1595     * pool->lock between setting %WORKER_UNBOUND and zapping
1596     * nr_running, the warning may trigger spuriously. Check iff
1597     * unbind is not in progress.
1598     */
1599    WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1600             pool->nr_workers == pool->nr_idle &&
1601             atomic_read(&pool->nr_running));
1602}
1603
1604/**
1605 * worker_leave_idle - leave idle state
1606 * @worker: worker which is leaving idle state
1607 *
1608 * @worker is leaving idle state. Update stats.
1609 *
1610 * LOCKING:
1611 * spin_lock_irq(pool->lock).
1612 */
1613static void worker_leave_idle(struct worker *worker)
1614{
1615    struct worker_pool *pool = worker->pool;
1616
1617    if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1618        return;
1619    worker_clr_flags(worker, WORKER_IDLE);
1620    pool->nr_idle--;
1621    list_del_init(&worker->entry);
1622}
1623
1624/**
1625 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1626 * @pool: target worker_pool
1627 *
1628 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1629 *
1630 * Works which are scheduled while the cpu is online must at least be
1631 * scheduled to a worker which is bound to the cpu so that if they are
1632 * flushed from cpu callbacks while cpu is going down, they are
1633 * guaranteed to execute on the cpu.
1634 *
1635 * This function is to be used by unbound workers and rescuers to bind
1636 * themselves to the target cpu and may race with cpu going down or
1637 * coming online. kthread_bind() can't be used because it may put the
1638 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1639 * verbatim as it's best effort and blocking and pool may be
1640 * [dis]associated in the meantime.
1641 *
1642 * This function tries set_cpus_allowed() and locks pool and verifies the
1643 * binding against %POOL_DISASSOCIATED which is set during
1644 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1645 * enters idle state or fetches works without dropping lock, it can
1646 * guarantee the scheduling requirement described in the first paragraph.
1647 *
1648 * CONTEXT:
1649 * Might sleep. Called without any lock but returns with pool->lock
1650 * held.
1651 *
1652 * Return:
1653 * %true if the associated pool is online (@worker is successfully
1654 * bound), %false if offline.
1655 */
1656static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1657__acquires(&pool->lock)
1658{
1659    while (true) {
1660        /*
1661         * The following call may fail, succeed or succeed
1662         * without actually migrating the task to the cpu if
1663         * it races with cpu hotunplug operation. Verify
1664         * against POOL_DISASSOCIATED.
1665         */
1666        if (!(pool->flags & POOL_DISASSOCIATED))
1667            set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1668
1669        spin_lock_irq(&pool->lock);
1670        if (pool->flags & POOL_DISASSOCIATED)
1671            return false;
1672        if (task_cpu(current) == pool->cpu &&
1673            cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
1674            return true;
1675        spin_unlock_irq(&pool->lock);
1676
1677        /*
1678         * We've raced with CPU hot[un]plug. Give it a breather
1679         * and retry migration. cond_resched() is required here;
1680         * otherwise, we might deadlock against cpu_stop trying to
1681         * bring down the CPU on non-preemptive kernel.
1682         */
1683        cpu_relax();
1684        cond_resched();
1685    }
1686}
1687
1688static struct worker *alloc_worker(void)
1689{
1690    struct worker *worker;
1691
1692    worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1693    if (worker) {
1694        INIT_LIST_HEAD(&worker->entry);
1695        INIT_LIST_HEAD(&worker->scheduled);
1696        /* on creation a worker is in !idle && prep state */
1697        worker->flags = WORKER_PREP;
1698    }
1699    return worker;
1700}
1701
1702/**
1703 * create_worker - create a new workqueue worker
1704 * @pool: pool the new worker will belong to
1705 *
1706 * Create a new worker which is bound to @pool. The returned worker
1707 * can be started by calling start_worker() or destroyed using
1708 * destroy_worker().
1709 *
1710 * CONTEXT:
1711 * Might sleep. Does GFP_KERNEL allocations.
1712 *
1713 * Return:
1714 * Pointer to the newly created worker.
1715 */
1716static struct worker *create_worker(struct worker_pool *pool)
1717{
1718    struct worker *worker = NULL;
1719    int id = -1;
1720    char id_buf[16];
1721
1722    lockdep_assert_held(&pool->manager_mutex);
1723
1724    /*
1725     * ID is needed to determine kthread name. Allocate ID first
1726     * without installing the pointer.
1727     */
1728    idr_preload(GFP_KERNEL);
1729    spin_lock_irq(&pool->lock);
1730
1731    id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1732
1733    spin_unlock_irq(&pool->lock);
1734    idr_preload_end();
1735    if (id < 0)
1736        goto fail;
1737
1738    worker = alloc_worker();
1739    if (!worker)
1740        goto fail;
1741
1742    worker->pool = pool;
1743    worker->id = id;
1744
1745    if (pool->cpu >= 0)
1746        snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1747             pool->attrs->nice < 0 ? "H" : "");
1748    else
1749        snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1750
1751    worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1752                          "kworker/%s", id_buf);
1753    if (IS_ERR(worker->task))
1754        goto fail;
1755
1756    set_user_nice(worker->task, pool->attrs->nice);
1757
1758    /* prevent userland from meddling with cpumask of workqueue workers */
1759    worker->task->flags |= PF_NO_SETAFFINITY;
1760
1761    /*
1762     * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1763     * online CPUs. It'll be re-applied when any of the CPUs come up.
1764     */
1765    set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1766
1767    /*
1768     * The caller is responsible for ensuring %POOL_DISASSOCIATED
1769     * remains stable across this function. See the comments above the
1770     * flag definition for details.
1771     */
1772    if (pool->flags & POOL_DISASSOCIATED)
1773        worker->flags |= WORKER_UNBOUND;
1774
1775    /* successful, commit the pointer to idr */
1776    spin_lock_irq(&pool->lock);
1777    idr_replace(&pool->worker_idr, worker, worker->id);
1778    spin_unlock_irq(&pool->lock);
1779
1780    return worker;
1781
1782fail:
1783    if (id >= 0) {
1784        spin_lock_irq(&pool->lock);
1785        idr_remove(&pool->worker_idr, id);
1786        spin_unlock_irq(&pool->lock);
1787    }
1788    kfree(worker);
1789    return NULL;
1790}
1791
1792/**
1793 * start_worker - start a newly created worker
1794 * @worker: worker to start
1795 *
1796 * Make the pool aware of @worker and start it.
1797 *
1798 * CONTEXT:
1799 * spin_lock_irq(pool->lock).
1800 */
1801static void start_worker(struct worker *worker)
1802{
1803    worker->flags |= WORKER_STARTED;
1804    worker->pool->nr_workers++;
1805    worker_enter_idle(worker);
1806    wake_up_process(worker->task);
1807}
1808
1809/**
1810 * create_and_start_worker - create and start a worker for a pool
1811 * @pool: the target pool
1812 *
1813 * Grab the managership of @pool and create and start a new worker for it.
1814 *
1815 * Return: 0 on success. A negative error code otherwise.
1816 */
1817static int create_and_start_worker(struct worker_pool *pool)
1818{
1819    struct worker *worker;
1820
1821    mutex_lock(&pool->manager_mutex);
1822
1823    worker = create_worker(pool);
1824    if (worker) {
1825        spin_lock_irq(&pool->lock);
1826        start_worker(worker);
1827        spin_unlock_irq(&pool->lock);
1828    }
1829
1830    mutex_unlock(&pool->manager_mutex);
1831
1832    return worker ? 0 : -ENOMEM;
1833}
1834
1835/**
1836 * destroy_worker - destroy a workqueue worker
1837 * @worker: worker to be destroyed
1838 *
1839 * Destroy @worker and adjust @pool stats accordingly.
1840 *
1841 * CONTEXT:
1842 * spin_lock_irq(pool->lock) which is released and regrabbed.
1843 */
1844static void destroy_worker(struct worker *worker)
1845{
1846    struct worker_pool *pool = worker->pool;
1847
1848    lockdep_assert_held(&pool->manager_mutex);
1849    lockdep_assert_held(&pool->lock);
1850
1851    /* sanity check frenzy */
1852    if (WARN_ON(worker->current_work) ||
1853        WARN_ON(!list_empty(&worker->scheduled)))
1854        return;
1855
1856    if (worker->flags & WORKER_STARTED)
1857        pool->nr_workers--;
1858    if (worker->flags & WORKER_IDLE)
1859        pool->nr_idle--;
1860
1861    /*
1862     * Once WORKER_DIE is set, the kworker may destroy itself at any
1863     * point. Pin to ensure the task stays until we're done with it.
1864     */
1865    get_task_struct(worker->task);
1866
1867    list_del_init(&worker->entry);
1868    worker->flags |= WORKER_DIE;
1869
1870    idr_remove(&pool->worker_idr, worker->id);
1871
1872    spin_unlock_irq(&pool->lock);
1873
1874    kthread_stop(worker->task);
1875    put_task_struct(worker->task);
1876    kfree(worker);
1877
1878    spin_lock_irq(&pool->lock);
1879}
1880
1881static void idle_worker_timeout(unsigned long __pool)
1882{
1883    struct worker_pool *pool = (void *)__pool;
1884
1885    spin_lock_irq(&pool->lock);
1886
1887    if (too_many_workers(pool)) {
1888        struct worker *worker;
1889        unsigned long expires;
1890
1891        /* idle_list is kept in LIFO order, check the last one */
1892        worker = list_entry(pool->idle_list.prev, struct worker, entry);
1893        expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1894
1895        if (time_before(jiffies, expires))
1896            mod_timer(&pool->idle_timer, expires);
1897        else {
1898            /* it's been idle for too long, wake up manager */
1899            pool->flags |= POOL_MANAGE_WORKERS;
1900            wake_up_worker(pool);
1901        }
1902    }
1903
1904    spin_unlock_irq(&pool->lock);
1905}
1906
1907static void send_mayday(struct work_struct *work)
1908{
1909    struct pool_workqueue *pwq = get_work_pwq(work);
1910    struct workqueue_struct *wq = pwq->wq;
1911
1912    lockdep_assert_held(&wq_mayday_lock);
1913
1914    if (!wq->rescuer)
1915        return;
1916
1917    /* mayday mayday mayday */
1918    if (list_empty(&pwq->mayday_node)) {
1919        /*
1920         * If @pwq is for an unbound wq, its base ref may be put at
1921         * any time due to an attribute change. Pin @pwq until the
1922         * rescuer is done with it.
1923         */
1924        get_pwq(pwq);
1925        list_add_tail(&pwq->mayday_node, &wq->maydays);
1926        wake_up_process(wq->rescuer->task);
1927    }
1928}
1929
1930static void pool_mayday_timeout(unsigned long __pool)
1931{
1932    struct worker_pool *pool = (void *)__pool;
1933    struct work_struct *work;
1934
1935    spin_lock_irq(&wq_mayday_lock); /* for wq->maydays */
1936    spin_lock(&pool->lock);
1937
1938    if (need_to_create_worker(pool)) {
1939        /*
1940         * We've been trying to create a new worker but
1941         * haven't been successful. We might be hitting an
1942         * allocation deadlock. Send distress signals to
1943         * rescuers.
1944         */
1945        list_for_each_entry(work, &pool->worklist, entry)
1946            send_mayday(work);
1947    }
1948
1949    spin_unlock(&pool->lock);
1950    spin_unlock_irq(&wq_mayday_lock);
1951
1952    mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1953}
1954
1955/**
1956 * maybe_create_worker - create a new worker if necessary
1957 * @pool: pool to create a new worker for
1958 *
1959 * Create a new worker for @pool if necessary. @pool is guaranteed to
1960 * have at least one idle worker on return from this function. If
1961 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1962 * sent to all rescuers with works scheduled on @pool to resolve
1963 * possible allocation deadlock.
1964 *
1965 * On return, need_to_create_worker() is guaranteed to be %false and
1966 * may_start_working() %true.
1967 *
1968 * LOCKING:
1969 * spin_lock_irq(pool->lock) which may be released and regrabbed
1970 * multiple times. Does GFP_KERNEL allocations. Called only from
1971 * manager.
1972 *
1973 * Return:
1974 * %false if no action was taken and pool->lock stayed locked, %true
1975 * otherwise.
1976 */
1977static bool maybe_create_worker(struct worker_pool *pool)
1978__releases(&pool->lock)
1979__acquires(&pool->lock)
1980{
1981    if (!need_to_create_worker(pool))
1982        return false;
1983restart:
1984    spin_unlock_irq(&pool->lock);
1985
1986    /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1987    mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1988
1989    while (true) {
1990        struct worker *worker;
1991
1992        worker = create_worker(pool);
1993        if (worker) {
1994            del_timer_sync(&pool->mayday_timer);
1995            spin_lock_irq(&pool->lock);
1996            start_worker(worker);
1997            if (WARN_ON_ONCE(need_to_create_worker(pool)))
1998                goto restart;
1999            return true;
2000        }
2001
2002        if (!need_to_create_worker(pool))
2003            break;
2004
2005        __set_current_state(TASK_INTERRUPTIBLE);
2006        schedule_timeout(CREATE_COOLDOWN);
2007
2008        if (!need_to_create_worker(pool))
2009            break;
2010    }
2011
2012    del_timer_sync(&pool->mayday_timer);
2013    spin_lock_irq(&pool->lock);
2014    if (need_to_create_worker(pool))
2015        goto restart;
2016    return true;
2017}
2018
2019/**
2020 * maybe_destroy_worker - destroy workers which have been idle for a while
2021 * @pool: pool to destroy workers for
2022 *
2023 * Destroy @pool workers which have been idle for longer than
2024 * IDLE_WORKER_TIMEOUT.
2025 *
2026 * LOCKING:
2027 * spin_lock_irq(pool->lock) which may be released and regrabbed
2028 * multiple times. Called only from manager.
2029 *
2030 * Return:
2031 * %false if no action was taken and pool->lock stayed locked, %true
2032 * otherwise.
2033 */
2034static bool maybe_destroy_workers(struct worker_pool *pool)
2035{
2036    bool ret = false;
2037
2038    while (too_many_workers(pool)) {
2039        struct worker *worker;
2040        unsigned long expires;
2041
2042        worker = list_entry(pool->idle_list.prev, struct worker, entry);
2043        expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2044
2045        if (time_before(jiffies, expires)) {
2046            mod_timer(&pool->idle_timer, expires);
2047            break;
2048        }
2049
2050        destroy_worker(worker);
2051        ret = true;
2052    }
2053
2054    return ret;
2055}
2056
2057/**
2058 * manage_workers - manage worker pool
2059 * @worker: self
2060 *
2061 * Assume the manager role and manage the worker pool @worker belongs
2062 * to. At any given time, there can be only zero or one manager per
2063 * pool. The exclusion is handled automatically by this function.
2064 *
2065 * The caller can safely start processing works on false return. On
2066 * true return, it's guaranteed that need_to_create_worker() is false
2067 * and may_start_working() is true.
2068 *
2069 * CONTEXT:
2070 * spin_lock_irq(pool->lock) which may be released and regrabbed
2071 * multiple times. Does GFP_KERNEL allocations.
2072 *
2073 * Return:
2074 * %false if the pool don't need management and the caller can safely start
2075 * processing works, %true indicates that the function released pool->lock
2076 * and reacquired it to perform some management function and that the
2077 * conditions that the caller verified while holding the lock before
2078 * calling the function might no longer be true.
2079 */
2080static bool manage_workers(struct worker *worker)
2081{
2082    struct worker_pool *pool = worker->pool;
2083    bool ret = false;
2084
2085    /*
2086     * Managership is governed by two mutexes - manager_arb and
2087     * manager_mutex. manager_arb handles arbitration of manager role.
2088     * Anyone who successfully grabs manager_arb wins the arbitration
2089     * and becomes the manager. mutex_trylock() on pool->manager_arb
2090     * failure while holding pool->lock reliably indicates that someone
2091     * else is managing the pool and the worker which failed trylock
2092     * can proceed to executing work items. This means that anyone
2093     * grabbing manager_arb is responsible for actually performing
2094     * manager duties. If manager_arb is grabbed and released without
2095     * actual management, the pool may stall indefinitely.
2096     *
2097     * manager_mutex is used for exclusion of actual management
2098     * operations. The holder of manager_mutex can be sure that none
2099     * of management operations, including creation and destruction of
2100     * workers, won't take place until the mutex is released. Because
2101     * manager_mutex doesn't interfere with manager role arbitration,
2102     * it is guaranteed that the pool's management, while may be
2103     * delayed, won't be disturbed by someone else grabbing
2104     * manager_mutex.
2105     */
2106    if (!mutex_trylock(&pool->manager_arb))
2107        return ret;
2108
2109    /*
2110     * With manager arbitration won, manager_mutex would be free in
2111     * most cases. trylock first without dropping @pool->lock.
2112     */
2113    if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2114        spin_unlock_irq(&pool->lock);
2115        mutex_lock(&pool->manager_mutex);
2116        spin_lock_irq(&pool->lock);
2117        ret = true;
2118    }
2119
2120    pool->flags &= ~POOL_MANAGE_WORKERS;
2121
2122    /*
2123     * Destroy and then create so that may_start_working() is true
2124     * on return.
2125     */
2126    ret |= maybe_destroy_workers(pool);
2127    ret |= maybe_create_worker(pool);
2128
2129    mutex_unlock(&pool->manager_mutex);
2130    mutex_unlock(&pool->manager_arb);
2131    return ret;
2132}
2133
2134/**
2135 * process_one_work - process single work
2136 * @worker: self
2137 * @work: work to process
2138 *
2139 * Process @work. This function contains all the logics necessary to
2140 * process a single work including synchronization against and
2141 * interaction with other workers on the same cpu, queueing and
2142 * flushing. As long as context requirement is met, any worker can
2143 * call this function to process a work.
2144 *
2145 * CONTEXT:
2146 * spin_lock_irq(pool->lock) which is released and regrabbed.
2147 */
2148static void process_one_work(struct worker *worker, struct work_struct *work)
2149__releases(&pool->lock)
2150__acquires(&pool->lock)
2151{
2152    struct pool_workqueue *pwq = get_work_pwq(work);
2153    struct worker_pool *pool = worker->pool;
2154    bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2155    int work_color;
2156    struct worker *collision;
2157#ifdef CONFIG_LOCKDEP
2158    /*
2159     * It is permissible to free the struct work_struct from
2160     * inside the function that is called from it, this we need to
2161     * take into account for lockdep too. To avoid bogus "held
2162     * lock freed" warnings as well as problems when looking into
2163     * work->lockdep_map, make a copy and use that here.
2164     */
2165    struct lockdep_map lockdep_map;
2166
2167    lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2168#endif
2169    /*
2170     * Ensure we're on the correct CPU. DISASSOCIATED test is
2171     * necessary to avoid spurious warnings from rescuers servicing the
2172     * unbound or a disassociated pool.
2173     */
2174    WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2175             !(pool->flags & POOL_DISASSOCIATED) &&
2176             raw_smp_processor_id() != pool->cpu);
2177
2178    /*
2179     * A single work shouldn't be executed concurrently by
2180     * multiple workers on a single cpu. Check whether anyone is
2181     * already processing the work. If so, defer the work to the
2182     * currently executing one.
2183     */
2184    collision = find_worker_executing_work(pool, work);
2185    if (unlikely(collision)) {
2186        move_linked_works(work, &collision->scheduled, NULL);
2187        return;
2188    }
2189
2190    /* claim and dequeue */
2191    debug_work_deactivate(work);
2192    hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2193    worker->current_work = work;
2194    worker->current_func = work->func;
2195    worker->current_pwq = pwq;
2196    work_color = get_work_color(work);
2197
2198    list_del_init(&work->entry);
2199
2200    /*
2201     * CPU intensive works don't participate in concurrency
2202     * management. They're the scheduler's responsibility.
2203     */
2204    if (unlikely(cpu_intensive))
2205        worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2206
2207    /*
2208     * Unbound pool isn't concurrency managed and work items should be
2209     * executed ASAP. Wake up another worker if necessary.
2210     */
2211    if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2212        wake_up_worker(pool);
2213
2214    /*
2215     * Record the last pool and clear PENDING which should be the last
2216     * update to @work. Also, do this inside @pool->lock so that
2217     * PENDING and queued state changes happen together while IRQ is
2218     * disabled.
2219     */
2220    set_work_pool_and_clear_pending(work, pool->id);
2221
2222    spin_unlock_irq(&pool->lock);
2223
2224    lock_map_acquire_read(&pwq->wq->lockdep_map);
2225    lock_map_acquire(&lockdep_map);
2226    trace_workqueue_execute_start(work);
2227    worker->current_func(work);
2228    /*
2229     * While we must be careful to not use "work" after this, the trace
2230     * point will only record its address.
2231     */
2232    trace_workqueue_execute_end(work);
2233    lock_map_release(&lockdep_map);
2234    lock_map_release(&pwq->wq->lockdep_map);
2235
2236    if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2237        pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2238               " last function: %pf\n",
2239               current->comm, preempt_count(), task_pid_nr(current),
2240               worker->current_func);
2241        debug_show_held_locks(current);
2242        dump_stack();
2243    }
2244
2245    /*
2246     * The following prevents a kworker from hogging CPU on !PREEMPT
2247     * kernels, where a requeueing work item waiting for something to
2248     * happen could deadlock with stop_machine as such work item could
2249     * indefinitely requeue itself while all other CPUs are trapped in
2250     * stop_machine.
2251     */
2252    cond_resched();
2253
2254    spin_lock_irq(&pool->lock);
2255
2256    /* clear cpu intensive status */
2257    if (unlikely(cpu_intensive))
2258        worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2259
2260    /* we're done with it, release */
2261    hash_del(&worker->hentry);
2262    worker->current_work = NULL;
2263    worker->current_func = NULL;
2264    worker->current_pwq = NULL;
2265    worker->desc_valid = false;
2266    pwq_dec_nr_in_flight(pwq, work_color);
2267}
2268
2269/**
2270 * process_scheduled_works - process scheduled works
2271 * @worker: self
2272 *
2273 * Process all scheduled works. Please note that the scheduled list
2274 * may change while processing a work, so this function repeatedly
2275 * fetches a work from the top and executes it.
2276 *
2277 * CONTEXT:
2278 * spin_lock_irq(pool->lock) which may be released and regrabbed
2279 * multiple times.
2280 */
2281static void process_scheduled_works(struct worker *worker)
2282{
2283    while (!list_empty(&worker->scheduled)) {
2284        struct work_struct *work = list_first_entry(&worker->scheduled,
2285                        struct work_struct, entry);
2286        process_one_work(worker, work);
2287    }
2288}
2289
2290/**
2291 * worker_thread - the worker thread function
2292 * @__worker: self
2293 *
2294 * The worker thread function. All workers belong to a worker_pool -
2295 * either a per-cpu one or dynamic unbound one. These workers process all
2296 * work items regardless of their specific target workqueue. The only
2297 * exception is work items which belong to workqueues with a rescuer which
2298 * will be explained in rescuer_thread().
2299 *
2300 * Return: 0
2301 */
2302static int worker_thread(void *__worker)
2303{
2304    struct worker *worker = __worker;
2305    struct worker_pool *pool = worker->pool;
2306
2307    /* tell the scheduler that this is a workqueue worker */
2308    worker->task->flags |= PF_WQ_WORKER;
2309woke_up:
2310    spin_lock_irq(&pool->lock);
2311
2312    /* am I supposed to die? */
2313    if (unlikely(worker->flags & WORKER_DIE)) {
2314        spin_unlock_irq(&pool->lock);
2315        WARN_ON_ONCE(!list_empty(&worker->entry));
2316        worker->task->flags &= ~PF_WQ_WORKER;
2317        return 0;
2318    }
2319
2320    worker_leave_idle(worker);
2321recheck:
2322    /* no more worker necessary? */
2323    if (!need_more_worker(pool))
2324        goto sleep;
2325
2326    /* do we need to manage? */
2327    if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2328        goto recheck;
2329
2330    /*
2331     * ->scheduled list can only be filled while a worker is
2332     * preparing to process a work or actually processing it.
2333     * Make sure nobody diddled with it while I was sleeping.
2334     */
2335    WARN_ON_ONCE(!list_empty(&worker->scheduled));
2336
2337    /*
2338     * Finish PREP stage. We're guaranteed to have at least one idle
2339     * worker or that someone else has already assumed the manager
2340     * role. This is where @worker starts participating in concurrency
2341     * management if applicable and concurrency management is restored
2342     * after being rebound. See rebind_workers() for details.
2343     */
2344    worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2345
2346    do {
2347        struct work_struct *work =
2348            list_first_entry(&pool->worklist,
2349                     struct work_struct, entry);
2350
2351        if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2352            /* optimization path, not strictly necessary */
2353            process_one_work(worker, work);
2354            if (unlikely(!list_empty(&worker->scheduled)))
2355                process_scheduled_works(worker);
2356        } else {
2357            move_linked_works(work, &worker->scheduled, NULL);
2358            process_scheduled_works(worker);
2359        }
2360    } while (keep_working(pool));
2361
2362    worker_set_flags(worker, WORKER_PREP, false);
2363sleep:
2364    if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2365        goto recheck;
2366
2367    /*
2368     * pool->lock is held and there's no work to process and no need to
2369     * manage, sleep. Workers are woken up only while holding
2370     * pool->lock or from local cpu, so setting the current state
2371     * before releasing pool->lock is enough to prevent losing any
2372     * event.
2373     */
2374    worker_enter_idle(worker);
2375    __set_current_state(TASK_INTERRUPTIBLE);
2376    spin_unlock_irq(&pool->lock);
2377    schedule();
2378    goto woke_up;
2379}
2380
2381/**
2382 * rescuer_thread - the rescuer thread function
2383 * @__rescuer: self
2384 *
2385 * Workqueue rescuer thread function. There's one rescuer for each
2386 * workqueue which has WQ_MEM_RECLAIM set.
2387 *
2388 * Regular work processing on a pool may block trying to create a new
2389 * worker which uses GFP_KERNEL allocation which has slight chance of
2390 * developing into deadlock if some works currently on the same queue
2391 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2392 * the problem rescuer solves.
2393 *
2394 * When such condition is possible, the pool summons rescuers of all
2395 * workqueues which have works queued on the pool and let them process
2396 * those works so that forward progress can be guaranteed.
2397 *
2398 * This should happen rarely.
2399 *
2400 * Return: 0
2401 */
2402static int rescuer_thread(void *__rescuer)
2403{
2404    struct worker *rescuer = __rescuer;
2405    struct workqueue_struct *wq = rescuer->rescue_wq;
2406    struct list_head *scheduled = &rescuer->scheduled;
2407    bool should_stop;
2408
2409    set_user_nice(current, RESCUER_NICE_LEVEL);
2410
2411    /*
2412     * Mark rescuer as worker too. As WORKER_PREP is never cleared, it
2413     * doesn't participate in concurrency management.
2414     */
2415    rescuer->task->flags |= PF_WQ_WORKER;
2416repeat:
2417    set_current_state(TASK_INTERRUPTIBLE);
2418
2419    /*
2420     * By the time the rescuer is requested to stop, the workqueue
2421     * shouldn't have any work pending, but @wq->maydays may still have
2422     * pwq(s) queued. This can happen by non-rescuer workers consuming
2423     * all the work items before the rescuer got to them. Go through
2424     * @wq->maydays processing before acting on should_stop so that the
2425     * list is always empty on exit.
2426     */
2427    should_stop = kthread_should_stop();
2428
2429    /* see whether any pwq is asking for help */
2430    spin_lock_irq(&wq_mayday_lock);
2431
2432    while (!list_empty(&wq->maydays)) {
2433        struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2434                    struct pool_workqueue, mayday_node);
2435        struct worker_pool *pool = pwq->pool;
2436        struct work_struct *work, *n;
2437
2438        __set_current_state(TASK_RUNNING);
2439        list_del_init(&pwq->mayday_node);
2440
2441        spin_unlock_irq(&wq_mayday_lock);
2442
2443        /* migrate to the target cpu if possible */
2444        worker_maybe_bind_and_lock(pool);
2445        rescuer->pool = pool;
2446
2447        /*
2448         * Slurp in all works issued via this workqueue and
2449         * process'em.
2450         */
2451        WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2452        list_for_each_entry_safe(work, n, &pool->worklist, entry)
2453            if (get_work_pwq(work) == pwq)
2454                move_linked_works(work, scheduled, &n);
2455
2456        process_scheduled_works(rescuer);
2457
2458        /*
2459         * Put the reference grabbed by send_mayday(). @pool won't
2460         * go away while we're holding its lock.
2461         */
2462        put_pwq(pwq);
2463
2464        /*
2465         * Leave this pool. If keep_working() is %true, notify a
2466         * regular worker; otherwise, we end up with 0 concurrency
2467         * and stalling the execution.
2468         */
2469        if (keep_working(pool))
2470            wake_up_worker(pool);
2471
2472        rescuer->pool = NULL;
2473        spin_unlock(&pool->lock);
2474        spin_lock(&wq_mayday_lock);
2475    }
2476
2477    spin_unlock_irq(&wq_mayday_lock);
2478
2479    if (should_stop) {
2480        __set_current_state(TASK_RUNNING);
2481        rescuer->task->flags &= ~PF_WQ_WORKER;
2482        return 0;
2483    }
2484
2485    /* rescuers should never participate in concurrency management */
2486    WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2487    schedule();
2488    goto repeat;
2489}
2490
2491struct wq_barrier {
2492    struct work_struct work;
2493    struct completion done;
2494};
2495
2496static void wq_barrier_func(struct work_struct *work)
2497{
2498    struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2499    complete(&barr->done);
2500}
2501
2502/**
2503 * insert_wq_barrier - insert a barrier work
2504 * @pwq: pwq to insert barrier into
2505 * @barr: wq_barrier to insert
2506 * @target: target work to attach @barr to
2507 * @worker: worker currently executing @target, NULL if @target is not executing
2508 *
2509 * @barr is linked to @target such that @barr is completed only after
2510 * @target finishes execution. Please note that the ordering
2511 * guarantee is observed only with respect to @target and on the local
2512 * cpu.
2513 *
2514 * Currently, a queued barrier can't be canceled. This is because
2515 * try_to_grab_pending() can't determine whether the work to be
2516 * grabbed is at the head of the queue and thus can't clear LINKED
2517 * flag of the previous work while there must be a valid next work
2518 * after a work with LINKED flag set.
2519 *
2520 * Note that when @worker is non-NULL, @target may be modified
2521 * underneath us, so we can't reliably determine pwq from @target.
2522 *
2523 * CONTEXT:
2524 * spin_lock_irq(pool->lock).
2525 */
2526static void insert_wq_barrier(struct pool_workqueue *pwq,
2527                  struct wq_barrier *barr,
2528                  struct work_struct *target, struct worker *worker)
2529{
2530    struct list_head *head;
2531    unsigned int linked = 0;
2532
2533    /*
2534     * debugobject calls are safe here even with pool->lock locked
2535     * as we know for sure that this will not trigger any of the
2536     * checks and call back into the fixup functions where we
2537     * might deadlock.
2538     */
2539    INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2540    __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2541    init_completion(&barr->done);
2542
2543    /*
2544     * If @target is currently being executed, schedule the
2545     * barrier to the worker; otherwise, put it after @target.
2546     */
2547    if (worker)
2548        head = worker->scheduled.next;
2549    else {
2550        unsigned long *bits = work_data_bits(target);
2551
2552        head = target->entry.next;
2553        /* there can already be other linked works, inherit and set */
2554        linked = *bits & WORK_STRUCT_LINKED;
2555        __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2556    }
2557
2558    debug_work_activate(&barr->work);
2559    insert_work(pwq, &barr->work, head,
2560            work_color_to_flags(WORK_NO_COLOR) | linked);
2561}
2562
2563/**
2564 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2565 * @wq: workqueue being flushed
2566 * @flush_color: new flush color, < 0 for no-op
2567 * @work_color: new work color, < 0 for no-op
2568 *
2569 * Prepare pwqs for workqueue flushing.
2570 *
2571 * If @flush_color is non-negative, flush_color on all pwqs should be
2572 * -1. If no pwq has in-flight commands at the specified color, all
2573 * pwq->flush_color's stay at -1 and %false is returned. If any pwq
2574 * has in flight commands, its pwq->flush_color is set to
2575 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2576 * wakeup logic is armed and %true is returned.
2577 *
2578 * The caller should have initialized @wq->first_flusher prior to
2579 * calling this function with non-negative @flush_color. If
2580 * @flush_color is negative, no flush color update is done and %false
2581 * is returned.
2582 *
2583 * If @work_color is non-negative, all pwqs should have the same
2584 * work_color which is previous to @work_color and all will be
2585 * advanced to @work_color.
2586 *
2587 * CONTEXT:
2588 * mutex_lock(wq->mutex).
2589 *
2590 * Return:
2591 * %true if @flush_color >= 0 and there's something to flush. %false
2592 * otherwise.
2593 */
2594static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2595                      int flush_color, int work_color)
2596{
2597    bool wait = false;
2598    struct pool_workqueue *pwq;
2599
2600    if (flush_color >= 0) {
2601        WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2602        atomic_set(&wq->nr_pwqs_to_flush, 1);
2603    }
2604
2605    for_each_pwq(pwq, wq) {
2606        struct worker_pool *pool = pwq->pool;
2607
2608        spin_lock_irq(&pool->lock);
2609
2610        if (flush_color >= 0) {
2611            WARN_ON_ONCE(pwq->flush_color != -1);
2612
2613            if (pwq->nr_in_flight[flush_color]) {
2614                pwq->flush_color = flush_color;
2615                atomic_inc(&wq->nr_pwqs_to_flush);
2616                wait = true;
2617            }
2618        }
2619
2620        if (work_color >= 0) {
2621            WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2622            pwq->work_color = work_color;
2623        }
2624
2625        spin_unlock_irq(&pool->lock);
2626    }
2627
2628    if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2629        complete(&wq->first_flusher->done);
2630
2631    return wait;
2632}
2633
2634/**
2635 * flush_workqueue - ensure that any scheduled work has run to completion.
2636 * @wq: workqueue to flush
2637 *
2638 * This function sleeps until all work items which were queued on entry
2639 * have finished execution, but it is not livelocked by new incoming ones.
2640 */
2641void flush_workqueue(struct workqueue_struct *wq)
2642{
2643    struct wq_flusher this_flusher = {
2644        .list = LIST_HEAD_INIT(this_flusher.list),
2645        .flush_color = -1,
2646        .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2647    };
2648    int next_color;
2649
2650    lock_map_acquire(&wq->lockdep_map);
2651    lock_map_release(&wq->lockdep_map);
2652
2653    mutex_lock(&wq->mutex);
2654
2655    /*
2656     * Start-to-wait phase
2657     */
2658    next_color = work_next_color(wq->work_color);
2659
2660    if (next_color != wq->flush_color) {
2661        /*
2662         * Color space is not full. The current work_color
2663         * becomes our flush_color and work_color is advanced
2664         * by one.
2665         */
2666        WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2667        this_flusher.flush_color = wq->work_color;
2668        wq->work_color = next_color;
2669
2670        if (!wq->first_flusher) {
2671            /* no flush in progress, become the first flusher */
2672            WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2673
2674            wq->first_flusher = &this_flusher;
2675
2676            if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2677                               wq->work_color)) {
2678                /* nothing to flush, done */
2679                wq->flush_color = next_color;
2680                wq->first_flusher = NULL;
2681                goto out_unlock;
2682            }
2683        } else {
2684            /* wait in queue */
2685            WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2686            list_add_tail(&this_flusher.list, &wq->flusher_queue);
2687            flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2688        }
2689    } else {
2690        /*
2691         * Oops, color space is full, wait on overflow queue.
2692         * The next flush completion will assign us
2693         * flush_color and transfer to flusher_queue.
2694         */
2695        list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2696    }
2697
2698    mutex_unlock(&wq->mutex);
2699
2700    wait_for_completion(&this_flusher.done);
2701
2702    /*
2703     * Wake-up-and-cascade phase
2704     *
2705     * First flushers are responsible for cascading flushes and
2706     * handling overflow. Non-first flushers can simply return.
2707     */
2708    if (wq->first_flusher != &this_flusher)
2709        return;
2710
2711    mutex_lock(&wq->mutex);
2712
2713    /* we might have raced, check again with mutex held */
2714    if (wq->first_flusher != &this_flusher)
2715        goto out_unlock;
2716
2717    wq->first_flusher = NULL;
2718
2719    WARN_ON_ONCE(!list_empty(&this_flusher.list));
2720    WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2721
2722    while (true) {
2723        struct wq_flusher *next, *tmp;
2724
2725        /* complete all the flushers sharing the current flush color */
2726        list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2727            if (next->flush_color != wq->flush_color)
2728                break;
2729            list_del_init(&next->list);
2730            complete(&next->done);
2731        }
2732
2733        WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2734                 wq->flush_color != work_next_color(wq->work_color));
2735
2736        /* this flush_color is finished, advance by one */
2737        wq->flush_color = work_next_color(wq->flush_color);
2738
2739        /* one color has been freed, handle overflow queue */
2740        if (!list_empty(&wq->flusher_overflow)) {
2741            /*
2742             * Assign the same color to all overflowed
2743             * flushers, advance work_color and append to
2744             * flusher_queue. This is the start-to-wait
2745             * phase for these overflowed flushers.
2746             */
2747            list_for_each_entry(tmp, &wq->flusher_overflow, list)
2748                tmp->flush_color = wq->work_color;
2749
2750            wq->work_color = work_next_color(wq->work_color);
2751
2752            list_splice_tail_init(&wq->flusher_overflow,
2753                          &wq->flusher_queue);
2754            flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2755        }
2756
2757        if (list_empty(&wq->flusher_queue)) {
2758            WARN_ON_ONCE(wq->flush_color != wq->work_color);
2759            break;
2760        }
2761
2762        /*
2763         * Need to flush more colors. Make the next flusher
2764         * the new first flusher and arm pwqs.
2765         */
2766        WARN_ON_ONCE(wq->flush_color == wq->work_color);
2767        WARN_ON_ONCE(wq->flush_color != next->flush_color);
2768
2769        list_del_init(&next->list);
2770        wq->first_flusher = next;
2771
2772        if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2773            break;
2774
2775        /*
2776         * Meh... this color is already done, clear first
2777         * flusher and repeat cascading.
2778         */
2779        wq->first_flusher = NULL;
2780    }
2781
2782out_unlock:
2783    mutex_unlock(&wq->mutex);
2784}
2785EXPORT_SYMBOL_GPL(flush_workqueue);
2786
2787/**
2788 * drain_workqueue - drain a workqueue
2789 * @wq: workqueue to drain
2790 *
2791 * Wait until the workqueue becomes empty. While draining is in progress,
2792 * only chain queueing is allowed. IOW, only currently pending or running
2793 * work items on @wq can queue further work items on it. @wq is flushed
2794 * repeatedly until it becomes empty. The number of flushing is detemined
2795 * by the depth of chaining and should be relatively short. Whine if it
2796 * takes too long.
2797 */
2798void drain_workqueue(struct workqueue_struct *wq)
2799{
2800    unsigned int flush_cnt = 0;
2801    struct pool_workqueue *pwq;
2802
2803    /*
2804     * __queue_work() needs to test whether there are drainers, is much
2805     * hotter than drain_workqueue() and already looks at @wq->flags.
2806     * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2807     */
2808    mutex_lock(&wq->mutex);
2809    if (!wq->nr_drainers++)
2810        wq->flags |= __WQ_DRAINING;
2811    mutex_unlock(&wq->mutex);
2812reflush:
2813    flush_workqueue(wq);
2814
2815    mutex_lock(&wq->mutex);
2816
2817    for_each_pwq(pwq, wq) {
2818        bool drained;
2819
2820        spin_lock_irq(&pwq->pool->lock);
2821        drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2822        spin_unlock_irq(&pwq->pool->lock);
2823
2824        if (drained)
2825            continue;
2826
2827        if (++flush_cnt == 10 ||
2828            (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2829            pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2830                wq->name, flush_cnt);
2831
2832        mutex_unlock(&wq->mutex);
2833        goto reflush;
2834    }
2835
2836    if (!--wq->nr_drainers)
2837        wq->flags &= ~__WQ_DRAINING;
2838    mutex_unlock(&wq->mutex);
2839}
2840EXPORT_SYMBOL_GPL(drain_workqueue);
2841
2842static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2843{
2844    struct worker *worker = NULL;
2845    struct worker_pool *pool;
2846    struct pool_workqueue *pwq;
2847
2848    might_sleep();
2849
2850    local_irq_disable();
2851    pool = get_work_pool(work);
2852    if (!pool) {
2853        local_irq_enable();
2854        return false;
2855    }
2856
2857    spin_lock(&pool->lock);
2858    /* see the comment in try_to_grab_pending() with the same code */
2859    pwq = get_work_pwq(work);
2860    if (pwq) {
2861        if (unlikely(pwq->pool != pool))
2862            goto already_gone;
2863    } else {
2864        worker = find_worker_executing_work(pool, work);
2865        if (!worker)
2866            goto already_gone;
2867        pwq = worker->current_pwq;
2868    }
2869
2870    insert_wq_barrier(pwq, barr, work, worker);
2871    spin_unlock_irq(&pool->lock);
2872
2873    /*
2874     * If @max_active is 1 or rescuer is in use, flushing another work
2875     * item on the same workqueue may lead to deadlock. Make sure the
2876     * flusher is not running on the same workqueue by verifying write
2877     * access.
2878     */
2879    if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2880        lock_map_acquire(&pwq->wq->lockdep_map);
2881    else
2882        lock_map_acquire_read(&pwq->wq->lockdep_map);
2883    lock_map_release(&pwq->wq->lockdep_map);
2884
2885    return true;
2886already_gone:
2887    spin_unlock_irq(&pool->lock);
2888    return false;
2889}
2890
2891/**
2892 * flush_work - wait for a work to finish executing the last queueing instance
2893 * @work: the work to flush
2894 *
2895 * Wait until @work has finished execution. @work is guaranteed to be idle
2896 * on return if it hasn't been requeued since flush started.
2897 *
2898 * Return:
2899 * %true if flush_work() waited for the work to finish execution,
2900 * %false if it was already idle.
2901 */
2902bool flush_work(struct work_struct *work)
2903{
2904    struct wq_barrier barr;
2905
2906    lock_map_acquire(&work->lockdep_map);
2907    lock_map_release(&work->lockdep_map);
2908
2909    if (start_flush_work(work, &barr)) {
2910        wait_for_completion(&barr.done);
2911        destroy_work_on_stack(&barr.work);
2912        return true;
2913    } else {
2914        return false;
2915    }
2916}
2917EXPORT_SYMBOL_GPL(flush_work);
2918
2919static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2920{
2921    unsigned long flags;
2922    int ret;
2923
2924    do {
2925        ret = try_to_grab_pending(work, is_dwork, &flags);
2926        /*
2927         * If someone else is canceling, wait for the same event it
2928         * would be waiting for before retrying.
2929         */
2930        if (unlikely(ret == -ENOENT))
2931            flush_work(work);
2932    } while (unlikely(ret < 0));
2933
2934    /* tell other tasks trying to grab @work to back off */
2935    mark_work_canceling(work);
2936    local_irq_restore(flags);
2937
2938    flush_work(work);
2939    clear_work_data(work);
2940    return ret;
2941}
2942
2943/**
2944 * cancel_work_sync - cancel a work and wait for it to finish
2945 * @work: the work to cancel
2946 *
2947 * Cancel @work and wait for its execution to finish. This function
2948 * can be used even if the work re-queues itself or migrates to
2949 * another workqueue. On return from this function, @work is
2950 * guaranteed to be not pending or executing on any CPU.
2951 *
2952 * cancel_work_sync(&delayed_work->work) must not be used for
2953 * delayed_work's. Use cancel_delayed_work_sync() instead.
2954 *
2955 * The caller must ensure that the workqueue on which @work was last
2956 * queued can't be destroyed before this function returns.
2957 *
2958 * Return:
2959 * %true if @work was pending, %false otherwise.
2960 */
2961bool cancel_work_sync(struct work_struct *work)
2962{
2963    return __cancel_work_timer(work, false);
2964}
2965EXPORT_SYMBOL_GPL(cancel_work_sync);
2966
2967/**
2968 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2969 * @dwork: the delayed work to flush
2970 *
2971 * Delayed timer is cancelled and the pending work is queued for
2972 * immediate execution. Like flush_work(), this function only
2973 * considers the last queueing instance of @dwork.
2974 *
2975 * Return:
2976 * %true if flush_work() waited for the work to finish execution,
2977 * %false if it was already idle.
2978 */
2979bool flush_delayed_work(struct delayed_work *dwork)
2980{
2981    local_irq_disable();
2982    if (del_timer_sync(&dwork->timer))
2983        __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2984    local_irq_enable();
2985    return flush_work(&dwork->work);
2986}
2987EXPORT_SYMBOL(flush_delayed_work);
2988
2989/**
2990 * cancel_delayed_work - cancel a delayed work
2991 * @dwork: delayed_work to cancel
2992 *
2993 * Kill off a pending delayed_work.
2994 *
2995 * Return: %true if @dwork was pending and canceled; %false if it wasn't
2996 * pending.
2997 *
2998 * Note:
2999 * The work callback function may still be running on return, unless
3000 * it returns %true and the work doesn't re-arm itself. Explicitly flush or
3001 * use cancel_delayed_work_sync() to wait on it.
3002 *
3003 * This function is safe to call from any context including IRQ handler.
3004 */
3005bool cancel_delayed_work(struct delayed_work *dwork)
3006{
3007    unsigned long flags;
3008    int ret;
3009
3010    do {
3011        ret = try_to_grab_pending(&dwork->work, true, &flags);
3012    } while (unlikely(ret == -EAGAIN));
3013
3014    if (unlikely(ret < 0))
3015        return false;
3016
3017    set_work_pool_and_clear_pending(&dwork->work,
3018                    get_work_pool_id(&dwork->work));
3019    local_irq_restore(flags);
3020    return ret;
3021}
3022EXPORT_SYMBOL(cancel_delayed_work);
3023
3024/**
3025 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3026 * @dwork: the delayed work cancel
3027 *
3028 * This is cancel_work_sync() for delayed works.
3029 *
3030 * Return:
3031 * %true if @dwork was pending, %false otherwise.
3032 */
3033bool cancel_delayed_work_sync(struct delayed_work *dwork)
3034{
3035    return __cancel_work_timer(&dwork->work, true);
3036}
3037EXPORT_SYMBOL(cancel_delayed_work_sync);
3038
3039/**
3040 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3041 * @func: the function to call
3042 *
3043 * schedule_on_each_cpu() executes @func on each online CPU using the
3044 * system workqueue and blocks until all CPUs have completed.
3045 * schedule_on_each_cpu() is very slow.
3046 *
3047 * Return:
3048 * 0 on success, -errno on failure.
3049 */
3050int schedule_on_each_cpu(work_func_t func)
3051{
3052    int cpu;
3053    struct work_struct __percpu *works;
3054
3055    works = alloc_percpu(struct work_struct);
3056    if (!works)
3057        return -ENOMEM;
3058
3059    get_online_cpus();
3060
3061    for_each_online_cpu(cpu) {
3062        struct work_struct *work = per_cpu_ptr(works, cpu);
3063
3064        INIT_WORK(work, func);
3065        schedule_work_on(cpu, work);
3066    }
3067
3068    for_each_online_cpu(cpu)
3069        flush_work(per_cpu_ptr(works, cpu));
3070
3071    put_online_cpus();
3072    free_percpu(works);
3073    return 0;
3074}
3075
3076/**
3077 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3078 *
3079 * Forces execution of the kernel-global workqueue and blocks until its
3080 * completion.
3081 *
3082 * Think twice before calling this function! It's very easy to get into
3083 * trouble if you don't take great care. Either of the following situations
3084 * will lead to deadlock:
3085 *
3086 * One of the work items currently on the workqueue needs to acquire
3087 * a lock held by your code or its caller.
3088 *
3089 * Your code is running in the context of a work routine.
3090 *
3091 * They will be detected by lockdep when they occur, but the first might not
3092 * occur very often. It depends on what work items are on the workqueue and
3093 * what locks they need, which you have no control over.
3094 *
3095 * In most situations flushing the entire workqueue is overkill; you merely
3096 * need to know that a particular work item isn't queued and isn't running.
3097 * In such cases you should use cancel_delayed_work_sync() or
3098 * cancel_work_sync() instead.
3099 */
3100void flush_scheduled_work(void)
3101{
3102    flush_workqueue(system_wq);
3103}
3104EXPORT_SYMBOL(flush_scheduled_work);
3105
3106/**
3107 * execute_in_process_context - reliably execute the routine with user context
3108 * @fn: the function to execute
3109 * @ew: guaranteed storage for the execute work structure (must
3110 * be available when the work executes)
3111 *
3112 * Executes the function immediately if process context is available,
3113 * otherwise schedules the function for delayed execution.
3114 *
3115 * Return: 0 - function was executed
3116 * 1 - function was scheduled for execution
3117 */
3118int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3119{
3120    if (!in_interrupt()) {
3121        fn(&ew->work);
3122        return 0;
3123    }
3124
3125    INIT_WORK(&ew->work, fn);
3126    schedule_work(&ew->work);
3127
3128    return 1;
3129}
3130EXPORT_SYMBOL_GPL(execute_in_process_context);
3131
3132#ifdef CONFIG_SYSFS
3133/*
3134 * Workqueues with WQ_SYSFS flag set is visible to userland via
3135 * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the
3136 * following attributes.
3137 *
3138 * per_cpu RO bool : whether the workqueue is per-cpu or unbound
3139 * max_active RW int : maximum number of in-flight work items
3140 *
3141 * Unbound workqueues have the following extra attributes.
3142 *
3143 * id RO int : the associated pool ID
3144 * nice RW int : nice value of the workers
3145 * cpumask RW mask : bitmask of allowed CPUs for the workers
3146 */
3147struct wq_device {
3148    struct workqueue_struct *wq;
3149    struct device dev;
3150};
3151
3152static struct workqueue_struct *dev_to_wq(struct device *dev)
3153{
3154    struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3155
3156    return wq_dev->wq;
3157}
3158
3159static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
3160                char *buf)
3161{
3162    struct workqueue_struct *wq = dev_to_wq(dev);
3163
3164    return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3165}
3166static DEVICE_ATTR_RO(per_cpu);
3167
3168static ssize_t max_active_show(struct device *dev,
3169                   struct device_attribute *attr, char *buf)
3170{
3171    struct workqueue_struct *wq = dev_to_wq(dev);
3172
3173    return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3174}
3175
3176static ssize_t max_active_store(struct device *dev,
3177                struct device_attribute *attr, const char *buf,
3178                size_t count)
3179{
3180    struct workqueue_struct *wq = dev_to_wq(dev);
3181    int val;
3182
3183    if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3184        return -EINVAL;
3185
3186    workqueue_set_max_active(wq, val);
3187    return count;
3188}
3189static DEVICE_ATTR_RW(max_active);
3190
3191static struct attribute *wq_sysfs_attrs[] = {
3192    &dev_attr_per_cpu.attr,
3193    &dev_attr_max_active.attr,
3194    NULL,
3195};
3196ATTRIBUTE_GROUPS(wq_sysfs);
3197
3198static ssize_t wq_pool_ids_show(struct device *dev,
3199                struct device_attribute *attr, char *buf)
3200{
3201    struct workqueue_struct *wq = dev_to_wq(dev);
3202    const char *delim = "";
3203    int node, written = 0;
3204
3205    rcu_read_lock_sched();
3206    for_each_node(node) {
3207        written += scnprintf(buf + written, PAGE_SIZE - written,
3208                     "%s%d:%d", delim, node,
3209                     unbound_pwq_by_node(wq, node)->pool->id);
3210        delim = " ";
3211    }
3212    written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3213    rcu_read_unlock_sched();
3214
3215    return written;
3216}
3217
3218static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3219                char *buf)
3220{
3221    struct workqueue_struct *wq = dev_to_wq(dev);
3222    int written;
3223
3224    mutex_lock(&wq->mutex);
3225    written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3226    mutex_unlock(&wq->mutex);
3227
3228    return written;
3229}
3230
3231/* prepare workqueue_attrs for sysfs store operations */
3232static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3233{
3234    struct workqueue_attrs *attrs;
3235
3236    attrs = alloc_workqueue_attrs(GFP_KERNEL);
3237    if (!attrs)
3238        return NULL;
3239
3240    mutex_lock(&wq->mutex);
3241    copy_workqueue_attrs(attrs, wq->unbound_attrs);
3242    mutex_unlock(&wq->mutex);
3243    return attrs;
3244}
3245
3246static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3247                 const char *buf, size_t count)
3248{
3249    struct workqueue_struct *wq = dev_to_wq(dev);
3250    struct workqueue_attrs *attrs;
3251    int ret;
3252
3253    attrs = wq_sysfs_prep_attrs(wq);
3254    if (!attrs)
3255        return -ENOMEM;
3256
3257    if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3258        attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
3259        ret = apply_workqueue_attrs(wq, attrs);
3260    else
3261        ret = -EINVAL;
3262
3263    free_workqueue_attrs(attrs);
3264    return ret ?: count;
3265}
3266
3267static ssize_t wq_cpumask_show(struct device *dev,
3268                   struct device_attribute *attr, char *buf)
3269{
3270    struct workqueue_struct *wq = dev_to_wq(dev);
3271    int written;
3272
3273    mutex_lock(&wq->mutex);
3274    written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3275    mutex_unlock(&wq->mutex);
3276
3277    written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3278    return written;
3279}
3280
3281static ssize_t wq_cpumask_store(struct device *dev,
3282                struct device_attribute *attr,
3283                const char *buf, size_t count)
3284{
3285    struct workqueue_struct *wq = dev_to_wq(dev);
3286    struct workqueue_attrs *attrs;
3287    int ret;
3288
3289    attrs = wq_sysfs_prep_attrs(wq);
3290    if (!attrs)
3291        return -ENOMEM;
3292
3293    ret = cpumask_parse(buf, attrs->cpumask);
3294    if (!ret)
3295        ret = apply_workqueue_attrs(wq, attrs);
3296
3297    free_workqueue_attrs(attrs);
3298    return ret ?: count;
3299}
3300
3301static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3302                char *buf)
3303{
3304    struct workqueue_struct *wq = dev_to_wq(dev);
3305    int written;
3306
3307    mutex_lock(&wq->mutex);
3308    written = scnprintf(buf, PAGE_SIZE, "%d\n",
3309                !wq->unbound_attrs->no_numa);
3310    mutex_unlock(&wq->mutex);
3311
3312    return written;
3313}
3314
3315static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3316                 const char *buf, size_t count)
3317{
3318    struct workqueue_struct *wq = dev_to_wq(dev);
3319    struct workqueue_attrs *attrs;
3320    int v, ret;
3321
3322    attrs = wq_sysfs_prep_attrs(wq);
3323    if (!attrs)
3324        return -ENOMEM;
3325
3326    ret = -EINVAL;
3327    if (sscanf(buf, "%d", &v) == 1) {
3328        attrs->no_numa = !v;
3329        ret = apply_workqueue_attrs(wq, attrs);
3330    }
3331
3332    free_workqueue_attrs(attrs);
3333    return ret ?: count;
3334}
3335
3336static struct device_attribute wq_sysfs_unbound_attrs[] = {
3337    __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3338    __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3339    __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3340    __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3341    __ATTR_NULL,
3342};
3343
3344static struct bus_type wq_subsys = {
3345    .name = "workqueue",
3346    .dev_groups = wq_sysfs_groups,
3347};
3348
3349static int __init wq_sysfs_init(void)
3350{
3351    return subsys_virtual_register(&wq_subsys, NULL);
3352}
3353core_initcall(wq_sysfs_init);
3354
3355static void wq_device_release(struct device *dev)
3356{
3357    struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3358
3359    kfree(wq_dev);
3360}
3361
3362/**
3363 * workqueue_sysfs_register - make a workqueue visible in sysfs
3364 * @wq: the workqueue to register
3365 *
3366 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3367 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3368 * which is the preferred method.
3369 *
3370 * Workqueue user should use this function directly iff it wants to apply
3371 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3372 * apply_workqueue_attrs() may race against userland updating the
3373 * attributes.
3374 *
3375 * Return: 0 on success, -errno on failure.
3376 */
3377int workqueue_sysfs_register(struct workqueue_struct *wq)
3378{
3379    struct wq_device *wq_dev;
3380    int ret;
3381
3382    /*
3383     * Adjusting max_active or creating new pwqs by applyting
3384     * attributes breaks ordering guarantee. Disallow exposing ordered
3385     * workqueues.
3386     */
3387    if (WARN_ON(wq->flags & __WQ_ORDERED))
3388        return -EINVAL;
3389
3390    wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3391    if (!wq_dev)
3392        return -ENOMEM;
3393
3394    wq_dev->wq = wq;
3395    wq_dev->dev.bus = &wq_subsys;
3396    wq_dev->dev.init_name = wq->name;
3397    wq_dev->dev.release = wq_device_release;
3398
3399    /*
3400     * unbound_attrs are created separately. Suppress uevent until
3401     * everything is ready.
3402     */
3403    dev_set_uevent_suppress(&wq_dev->dev, true);
3404
3405    ret = device_register(&wq_dev->dev);
3406    if (ret) {
3407        kfree(wq_dev);
3408        wq->wq_dev = NULL;
3409        return ret;
3410    }
3411
3412    if (wq->flags & WQ_UNBOUND) {
3413        struct device_attribute *attr;
3414
3415        for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3416            ret = device_create_file(&wq_dev->dev, attr);
3417            if (ret) {
3418                device_unregister(&wq_dev->dev);
3419                wq->wq_dev = NULL;
3420                return ret;
3421            }
3422        }
3423    }
3424
3425    kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3426    return 0;
3427}
3428
3429/**
3430 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3431 * @wq: the workqueue to unregister
3432 *
3433 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3434 */
3435static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3436{
3437    struct wq_device *wq_dev = wq->wq_dev;
3438
3439    if (!wq->wq_dev)
3440        return;
3441
3442    wq->wq_dev = NULL;
3443    device_unregister(&wq_dev->dev);
3444}
3445#else /* CONFIG_SYSFS */
3446static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { }
3447#endif /* CONFIG_SYSFS */
3448
3449/**
3450 * free_workqueue_attrs - free a workqueue_attrs
3451 * @attrs: workqueue_attrs to free
3452 *
3453 * Undo alloc_workqueue_attrs().
3454 */
3455void free_workqueue_attrs(struct workqueue_attrs *attrs)
3456{
3457    if (attrs) {
3458        free_cpumask_var(attrs->cpumask);
3459        kfree(attrs);
3460    }
3461}
3462
3463/**
3464 * alloc_workqueue_attrs - allocate a workqueue_attrs
3465 * @gfp_mask: allocation mask to use
3466 *
3467 * Allocate a new workqueue_attrs, initialize with default settings and
3468 * return it.
3469 *
3470 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3471 */
3472struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3473{
3474    struct workqueue_attrs *attrs;
3475
3476    attrs = kzalloc(sizeof(*attrs), gfp_mask);
3477    if (!attrs)
3478        goto fail;
3479    if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3480        goto fail;
3481
3482    cpumask_copy(attrs->cpumask, cpu_possible_mask);
3483    return attrs;
3484fail:
3485    free_workqueue_attrs(attrs);
3486    return NULL;
3487}
3488
3489static void copy_workqueue_attrs(struct workqueue_attrs *to,
3490                 const struct workqueue_attrs *from)
3491{
3492    to->nice = from->nice;
3493    cpumask_copy(to->cpumask, from->cpumask);
3494    /*
3495     * Unlike hash and equality test, this function doesn't ignore
3496     * ->no_numa as it is used for both pool and wq attrs. Instead,
3497     * get_unbound_pool() explicitly clears ->no_numa after copying.
3498     */
3499    to->no_numa = from->no_numa;
3500}
3501
3502/* hash value of the content of @attr */
3503static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3504{
3505    u32 hash = 0;
3506
3507    hash = jhash_1word(attrs->nice, hash);
3508    hash = jhash(cpumask_bits(attrs->cpumask),
3509             BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3510    return hash;
3511}
3512
3513/* content equality test */
3514static bool wqattrs_equal(const struct workqueue_attrs *a,
3515              const struct workqueue_attrs *b)
3516{
3517    if (a->nice != b->nice)
3518        return false;
3519    if (!cpumask_equal(a->cpumask, b->cpumask))
3520        return false;
3521    return true;
3522}
3523
3524/**
3525 * init_worker_pool - initialize a newly zalloc'd worker_pool
3526 * @pool: worker_pool to initialize
3527 *
3528 * Initiailize a newly zalloc'd @pool. It also allocates @pool->attrs.
3529 *
3530 * Return: 0 on success, -errno on failure. Even on failure, all fields
3531 * inside @pool proper are initialized and put_unbound_pool() can be called
3532 * on @pool safely to release it.
3533 */
3534static int init_worker_pool(struct worker_pool *pool)
3535{
3536    spin_lock_init(&pool->lock);
3537    pool->id = -1;
3538    pool->cpu = -1;
3539    pool->node = NUMA_NO_NODE;
3540    pool->flags |= POOL_DISASSOCIATED;
3541    INIT_LIST_HEAD(&pool->worklist);
3542    INIT_LIST_HEAD(&pool->idle_list);
3543    hash_init(pool->busy_hash);
3544
3545    init_timer_deferrable(&pool->idle_timer);
3546    pool->idle_timer.function = idle_worker_timeout;
3547    pool->idle_timer.data = (unsigned long)pool;
3548
3549    setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3550            (unsigned long)pool);
3551
3552    mutex_init(&pool->manager_arb);
3553    mutex_init(&pool->manager_mutex);
3554    idr_init(&pool->worker_idr);
3555
3556    INIT_HLIST_NODE(&pool->hash_node);
3557    pool->refcnt = 1;
3558
3559    /* shouldn't fail above this point */
3560    pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3561    if (!pool->attrs)
3562        return -ENOMEM;
3563    return 0;
3564}
3565
3566static void rcu_free_pool(struct rcu_head *rcu)
3567{
3568    struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3569
3570    idr_destroy(&pool->worker_idr);
3571    free_workqueue_attrs(pool->attrs);
3572    kfree(pool);
3573}
3574
3575/**
3576 * put_unbound_pool - put a worker_pool
3577 * @pool: worker_pool to put
3578 *
3579 * Put @pool. If its refcnt reaches zero, it gets destroyed in sched-RCU
3580 * safe manner. get_unbound_pool() calls this function on its failure path
3581 * and this function should be able to release pools which went through,
3582 * successfully or not, init_worker_pool().
3583 *
3584 * Should be called with wq_pool_mutex held.
3585 */
3586static void put_unbound_pool(struct worker_pool *pool)
3587{
3588    struct worker *worker;
3589
3590    lockdep_assert_held(&wq_pool_mutex);
3591
3592    if (--pool->refcnt)
3593        return;
3594
3595    /* sanity checks */
3596    if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3597        WARN_ON(!list_empty(&pool->worklist)))
3598        return;
3599
3600    /* release id and unhash */
3601    if (pool->id >= 0)
3602        idr_remove(&worker_pool_idr, pool->id);
3603    hash_del(&pool->hash_node);
3604
3605    /*
3606     * Become the manager and destroy all workers. Grabbing
3607     * manager_arb prevents @pool's workers from blocking on
3608     * manager_mutex.
3609     */
3610    mutex_lock(&pool->manager_arb);
3611    mutex_lock(&pool->manager_mutex);
3612    spin_lock_irq(&pool->lock);
3613
3614    while ((worker = first_worker(pool)))
3615        destroy_worker(worker);
3616    WARN_ON(pool->nr_workers || pool->nr_idle);
3617
3618    spin_unlock_irq(&pool->lock);
3619    mutex_unlock(&pool->manager_mutex);
3620    mutex_unlock(&pool->manager_arb);
3621
3622    /* shut down the timers */
3623    del_timer_sync(&pool->idle_timer);
3624    del_timer_sync(&pool->mayday_timer);
3625
3626    /* sched-RCU protected to allow dereferences from get_work_pool() */
3627    call_rcu_sched(&pool->rcu, rcu_free_pool);
3628}
3629
3630/**
3631 * get_unbound_pool - get a worker_pool with the specified attributes
3632 * @attrs: the attributes of the worker_pool to get
3633 *
3634 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3635 * reference count and return it. If there already is a matching
3636 * worker_pool, it will be used; otherwise, this function attempts to
3637 * create a new one.
3638 *
3639 * Should be called with wq_pool_mutex held.
3640 *
3641 * Return: On success, a worker_pool with the same attributes as @attrs.
3642 * On failure, %NULL.
3643 */
3644static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3645{
3646    u32 hash = wqattrs_hash(attrs);
3647    struct worker_pool *pool;
3648    int node;
3649
3650    lockdep_assert_held(&wq_pool_mutex);
3651
3652    /* do we already have a matching pool? */
3653    hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3654        if (wqattrs_equal(pool->attrs, attrs)) {
3655            pool->refcnt++;
3656            goto out_unlock;
3657        }
3658    }
3659
3660    /* nope, create a new one */
3661    pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3662    if (!pool || init_worker_pool(pool) < 0)
3663        goto fail;
3664
3665    if (workqueue_freezing)
3666        pool->flags |= POOL_FREEZING;
3667
3668    lockdep_set_subclass(&pool->lock, 1); /* see put_pwq() */
3669    copy_workqueue_attrs(pool->attrs, attrs);
3670
3671    /*
3672     * no_numa isn't a worker_pool attribute, always clear it. See
3673     * 'struct workqueue_attrs' comments for detail.
3674     */
3675    pool->attrs->no_numa = false;
3676
3677    /* if cpumask is contained inside a NUMA node, we belong to that node */
3678    if (wq_numa_enabled) {
3679        for_each_node(node) {
3680            if (cpumask_subset(pool->attrs->cpumask,
3681                       wq_numa_possible_cpumask[node])) {
3682                pool->node = node;
3683                break;
3684            }
3685        }
3686    }
3687
3688    if (worker_pool_assign_id(pool) < 0)
3689        goto fail;
3690
3691    /* create and start the initial worker */
3692    if (create_and_start_worker(pool) < 0)
3693        goto fail;
3694
3695    /* install */
3696    hash_add(unbound_pool_hash, &pool->hash_node, hash);
3697out_unlock:
3698    return pool;
3699fail:
3700    if (pool)
3701        put_unbound_pool(pool);
3702    return NULL;
3703}
3704
3705static void rcu_free_pwq(struct rcu_head *rcu)
3706{
3707    kmem_cache_free(pwq_cache,
3708            container_of(rcu, struct pool_workqueue, rcu));
3709}
3710
3711/*
3712 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3713 * and needs to be destroyed.
3714 */
3715static void pwq_unbound_release_workfn(struct work_struct *work)
3716{
3717    struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3718                          unbound_release_work);
3719    struct workqueue_struct *wq = pwq->wq;
3720    struct worker_pool *pool = pwq->pool;
3721    bool is_last;
3722
3723    if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3724        return;
3725
3726    /*
3727     * Unlink @pwq. Synchronization against wq->mutex isn't strictly
3728     * necessary on release but do it anyway. It's easier to verify
3729     * and consistent with the linking path.
3730     */
3731    mutex_lock(&wq->mutex);
3732    list_del_rcu(&pwq->pwqs_node);
3733    is_last = list_empty(&wq->pwqs);
3734    mutex_unlock(&wq->mutex);
3735
3736    mutex_lock(&wq_pool_mutex);
3737    put_unbound_pool(pool);
3738    mutex_unlock(&wq_pool_mutex);
3739
3740    call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3741
3742    /*
3743     * If we're the last pwq going away, @wq is already dead and no one
3744     * is gonna access it anymore. Free it.
3745     */
3746    if (is_last) {
3747        free_workqueue_attrs(wq->unbound_attrs);
3748        kfree(wq);
3749    }
3750}
3751
3752/**
3753 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3754 * @pwq: target pool_workqueue
3755 *
3756 * If @pwq isn't freezing, set @pwq->max_active to the associated
3757 * workqueue's saved_max_active and activate delayed work items
3758 * accordingly. If @pwq is freezing, clear @pwq->max_active to zero.
3759 */
3760static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3761{
3762    struct workqueue_struct *wq = pwq->wq;
3763    bool freezable = wq->flags & WQ_FREEZABLE;
3764
3765    /* for @wq->saved_max_active */
3766    lockdep_assert_held(&wq->mutex);
3767
3768    /* fast exit for non-freezable wqs */
3769    if (!freezable && pwq->max_active == wq->saved_max_active)
3770        return;
3771
3772    spin_lock_irq(&pwq->pool->lock);
3773
3774    if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3775        pwq->max_active = wq->saved_max_active;
3776
3777        while (!list_empty(&pwq->delayed_works) &&
3778               pwq->nr_active < pwq->max_active)
3779            pwq_activate_first_delayed(pwq);
3780
3781        /*
3782         * Need to kick a worker after thawed or an unbound wq's
3783         * max_active is bumped. It's a slow path. Do it always.
3784         */
3785        wake_up_worker(pwq->pool);
3786    } else {
3787        pwq->max_active = 0;
3788    }
3789
3790    spin_unlock_irq(&pwq->pool->lock);
3791}
3792
3793/* initialize newly alloced @pwq which is associated with @wq and @pool */
3794static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3795             struct worker_pool *pool)
3796{
3797    BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3798
3799    memset(pwq, 0, sizeof(*pwq));
3800
3801    pwq->pool = pool;
3802    pwq->wq = wq;
3803    pwq->flush_color = -1;
3804    pwq->refcnt = 1;
3805    INIT_LIST_HEAD(&pwq->delayed_works);
3806    INIT_LIST_HEAD(&pwq->pwqs_node);
3807    INIT_LIST_HEAD(&pwq->mayday_node);
3808    INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3809}
3810
3811/* sync @pwq with the current state of its associated wq and link it */
3812static void link_pwq(struct pool_workqueue *pwq)
3813{
3814    struct workqueue_struct *wq = pwq->wq;
3815
3816    lockdep_assert_held(&wq->mutex);
3817
3818    /* may be called multiple times, ignore if already linked */
3819    if (!list_empty(&pwq->pwqs_node))
3820        return;
3821
3822    /*
3823     * Set the matching work_color. This is synchronized with
3824     * wq->mutex to avoid confusing flush_workqueue().
3825     */
3826    pwq->work_color = wq->work_color;
3827
3828    /* sync max_active to the current setting */
3829    pwq_adjust_max_active(pwq);
3830
3831    /* link in @pwq */
3832    list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3833}
3834
3835/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3836static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3837                    const struct workqueue_attrs *attrs)
3838{
3839    struct worker_pool *pool;
3840    struct pool_workqueue *pwq;
3841
3842    lockdep_assert_held(&wq_pool_mutex);
3843
3844    pool = get_unbound_pool(attrs);
3845    if (!pool)
3846        return NULL;
3847
3848    pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3849    if (!pwq) {
3850        put_unbound_pool(pool);
3851        return NULL;
3852    }
3853
3854    init_pwq(pwq, wq, pool);
3855    return pwq;
3856}
3857
3858/* undo alloc_unbound_pwq(), used only in the error path */
3859static void free_unbound_pwq(struct pool_workqueue *pwq)
3860{
3861    lockdep_assert_held(&wq_pool_mutex);
3862
3863    if (pwq) {
3864        put_unbound_pool(pwq->pool);
3865        kmem_cache_free(pwq_cache, pwq);
3866    }
3867}
3868
3869/**
3870 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3871 * @attrs: the wq_attrs of interest
3872 * @node: the target NUMA node
3873 * @cpu_going_down: if >= 0, the CPU to consider as offline
3874 * @cpumask: outarg, the resulting cpumask
3875 *
3876 * Calculate the cpumask a workqueue with @attrs should use on @node. If
3877 * @cpu_going_down is >= 0, that cpu is considered offline during
3878 * calculation. The result is stored in @cpumask.
3879 *
3880 * If NUMA affinity is not enabled, @attrs->cpumask is always used. If
3881 * enabled and @node has online CPUs requested by @attrs, the returned
3882 * cpumask is the intersection of the possible CPUs of @node and
3883 * @attrs->cpumask.
3884 *
3885 * The caller is responsible for ensuring that the cpumask of @node stays
3886 * stable.
3887 *
3888 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3889 * %false if equal.
3890 */
3891static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3892                 int cpu_going_down, cpumask_t *cpumask)
3893{
3894    if (!wq_numa_enabled || attrs->no_numa)
3895        goto use_dfl;
3896
3897    /* does @node have any online CPUs @attrs wants? */
3898    cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3899    if (cpu_going_down >= 0)
3900        cpumask_clear_cpu(cpu_going_down, cpumask);
3901
3902    if (cpumask_empty(cpumask))
3903        goto use_dfl;
3904
3905    /* yeap, return possible CPUs in @node that @attrs wants */
3906    cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3907    return !cpumask_equal(cpumask, attrs->cpumask);
3908
3909use_dfl:
3910    cpumask_copy(cpumask, attrs->cpumask);
3911    return false;
3912}
3913
3914/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3915static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3916                           int node,
3917                           struct pool_workqueue *pwq)
3918{
3919    struct pool_workqueue *old_pwq;
3920
3921    lockdep_assert_held(&wq->mutex);
3922
3923    /* link_pwq() can handle duplicate calls */
3924    link_pwq(pwq);
3925
3926    old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3927    rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3928    return old_pwq;
3929}
3930
3931/**
3932 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3933 * @wq: the target workqueue
3934 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3935 *
3936 * Apply @attrs to an unbound workqueue @wq. Unless disabled, on NUMA
3937 * machines, this function maps a separate pwq to each NUMA node with
3938 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3939 * NUMA node it was issued on. Older pwqs are released as in-flight work
3940 * items finish. Note that a work item which repeatedly requeues itself
3941 * back-to-back will stay on its current pwq.
3942 *
3943 * Performs GFP_KERNEL allocations.
3944 *
3945 * Return: 0 on success and -errno on failure.
3946 */
3947int apply_workqueue_attrs(struct workqueue_struct *wq,
3948              const struct workqueue_attrs *attrs)
3949{
3950    struct workqueue_attrs *new_attrs, *tmp_attrs;
3951    struct pool_workqueue **pwq_tbl, *dfl_pwq;
3952    int node, ret;
3953
3954    /* only unbound workqueues can change attributes */
3955    if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3956        return -EINVAL;
3957
3958    /* creating multiple pwqs breaks ordering guarantee */
3959    if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3960        return -EINVAL;
3961
3962    pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3963    new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3964    tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3965    if (!pwq_tbl || !new_attrs || !tmp_attrs)
3966        goto enomem;
3967
3968    /* make a copy of @attrs and sanitize it */
3969    copy_workqueue_attrs(new_attrs, attrs);
3970    cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3971
3972    /*
3973     * We may create multiple pwqs with differing cpumasks. Make a
3974     * copy of @new_attrs which will be modified and used to obtain
3975     * pools.
3976     */
3977    copy_workqueue_attrs(tmp_attrs, new_attrs);
3978
3979    /*
3980     * CPUs should stay stable across pwq creations and installations.
3981     * Pin CPUs, determine the target cpumask for each node and create
3982     * pwqs accordingly.
3983     */
3984    get_online_cpus();
3985
3986    mutex_lock(&wq_pool_mutex);
3987
3988    /*
3989     * If something goes wrong during CPU up/down, we'll fall back to
3990     * the default pwq covering whole @attrs->cpumask. Always create
3991     * it even if we don't use it immediately.
3992     */
3993    dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3994    if (!dfl_pwq)
3995        goto enomem_pwq;
3996
3997    for_each_node(node) {
3998        if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3999            pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
4000            if (!pwq_tbl[node])
4001                goto enomem_pwq;
4002        } else {
4003            dfl_pwq->refcnt++;
4004            pwq_tbl[node] = dfl_pwq;
4005        }
4006    }
4007
4008    mutex_unlock(&wq_pool_mutex);
4009
4010    /* all pwqs have been created successfully, let's install'em */
4011    mutex_lock(&wq->mutex);
4012
4013    copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
4014
4015    /* save the previous pwq and install the new one */
4016    for_each_node(node)
4017        pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
4018
4019    /* @dfl_pwq might not have been used, ensure it's linked */
4020    link_pwq(dfl_pwq);
4021    swap(wq->dfl_pwq, dfl_pwq);
4022
4023    mutex_unlock(&wq->mutex);
4024
4025    /* put the old pwqs */
4026    for_each_node(node)
4027        put_pwq_unlocked(pwq_tbl[node]);
4028    put_pwq_unlocked(dfl_pwq);
4029
4030    put_online_cpus();
4031    ret = 0;
4032    /* fall through */
4033out_free:
4034    free_workqueue_attrs(tmp_attrs);
4035    free_workqueue_attrs(new_attrs);
4036    kfree(pwq_tbl);
4037    return ret;
4038
4039enomem_pwq:
4040    free_unbound_pwq(dfl_pwq);
4041    for_each_node(node)
4042        if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
4043            free_unbound_pwq(pwq_tbl[node]);
4044    mutex_unlock(&wq_pool_mutex);
4045    put_online_cpus();
4046enomem:
4047    ret = -ENOMEM;
4048    goto out_free;
4049}
4050
4051/**
4052 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4053 * @wq: the target workqueue
4054 * @cpu: the CPU coming up or going down
4055 * @online: whether @cpu is coming up or going down
4056 *
4057 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4058 * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update NUMA affinity of
4059 * @wq accordingly.
4060 *
4061 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4062 * falls back to @wq->dfl_pwq which may not be optimal but is always
4063 * correct.
4064 *
4065 * Note that when the last allowed CPU of a NUMA node goes offline for a
4066 * workqueue with a cpumask spanning multiple nodes, the workers which were
4067 * already executing the work items for the workqueue will lose their CPU
4068 * affinity and may execute on any CPU. This is similar to how per-cpu
4069 * workqueues behave on CPU_DOWN. If a workqueue user wants strict
4070 * affinity, it's the user's responsibility to flush the work item from
4071 * CPU_DOWN_PREPARE.
4072 */
4073static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4074                   bool online)
4075{
4076    int node = cpu_to_node(cpu);
4077    int cpu_off = online ? -1 : cpu;
4078    struct pool_workqueue *old_pwq = NULL, *pwq;
4079    struct workqueue_attrs *target_attrs;
4080    cpumask_t *cpumask;
4081
4082    lockdep_assert_held(&wq_pool_mutex);
4083
4084    if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4085        return;
4086
4087    /*
4088     * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4089     * Let's use a preallocated one. The following buf is protected by
4090     * CPU hotplug exclusion.
4091     */
4092    target_attrs = wq_update_unbound_numa_attrs_buf;
4093    cpumask = target_attrs->cpumask;
4094
4095    mutex_lock(&wq->mutex);
4096    if (wq->unbound_attrs->no_numa)
4097        goto out_unlock;
4098
4099    copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4100    pwq = unbound_pwq_by_node(wq, node);
4101
4102    /*
4103     * Let's determine what needs to be done. If the target cpumask is
4104     * different from wq's, we need to compare it to @pwq's and create
4105     * a new one if they don't match. If the target cpumask equals
4106     * wq's, the default pwq should be used. If @pwq is already the
4107     * default one, nothing to do; otherwise, install the default one.
4108     */
4109    if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4110        if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4111            goto out_unlock;
4112    } else {
4113        if (pwq == wq->dfl_pwq)
4114            goto out_unlock;
4115        else
4116            goto use_dfl_pwq;
4117    }
4118
4119    mutex_unlock(&wq->mutex);
4120
4121    /* create a new pwq */
4122    pwq = alloc_unbound_pwq(wq, target_attrs);
4123    if (!pwq) {
4124        pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4125               wq->name);
4126        mutex_lock(&wq->mutex);
4127        goto use_dfl_pwq;
4128    }
4129
4130    /*
4131     * Install the new pwq. As this function is called only from CPU
4132     * hotplug callbacks and applying a new attrs is wrapped with
4133     * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4134     * inbetween.
4135     */
4136    mutex_lock(&wq->mutex);
4137    old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4138    goto out_unlock;
4139
4140use_dfl_pwq:
4141    spin_lock_irq(&wq->dfl_pwq->pool->lock);
4142    get_pwq(wq->dfl_pwq);
4143    spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4144    old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4145out_unlock:
4146    mutex_unlock(&wq->mutex);
4147    put_pwq_unlocked(old_pwq);
4148}
4149
4150static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4151{
4152    bool highpri = wq->flags & WQ_HIGHPRI;
4153    int cpu, ret;
4154
4155    if (!(wq->flags & WQ_UNBOUND)) {
4156        wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4157        if (!wq->cpu_pwqs)
4158            return -ENOMEM;
4159
4160        for_each_possible_cpu(cpu) {
4161            struct pool_workqueue *pwq =
4162                per_cpu_ptr(wq->cpu_pwqs, cpu);
4163            struct worker_pool *cpu_pools =
4164                per_cpu(cpu_worker_pools, cpu);
4165
4166            init_pwq(pwq, wq, &cpu_pools[highpri]);
4167
4168            mutex_lock(&wq->mutex);
4169            link_pwq(pwq);
4170            mutex_unlock(&wq->mutex);
4171        }
4172        return 0;
4173    } else if (wq->flags & __WQ_ORDERED) {
4174        ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4175        /* there should only be single pwq for ordering guarantee */
4176        WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4177                  wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4178             "ordering guarantee broken for workqueue %s\n", wq->name);
4179        return ret;
4180    } else {
4181        return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4182    }
4183}
4184
4185static int wq_clamp_max_active(int max_active, unsigned int flags,
4186                   const char *name)
4187{
4188    int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4189
4190    if (max_active < 1 || max_active > lim)
4191        pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4192            max_active, name, 1, lim);
4193
4194    return clamp_val(max_active, 1, lim);
4195}
4196
4197struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4198                           unsigned int flags,
4199                           int max_active,
4200                           struct lock_class_key *key,
4201                           const char *lock_name, ...)
4202{
4203    size_t tbl_size = 0;
4204    va_list args;
4205    struct workqueue_struct *wq;
4206    struct pool_workqueue *pwq;
4207
4208    /* see the comment above the definition of WQ_POWER_EFFICIENT */
4209    if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4210        flags |= WQ_UNBOUND;
4211
4212    /* allocate wq and format name */
4213    if (flags & WQ_UNBOUND)
4214        tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4215
4216    wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4217    if (!wq)
4218        return NULL;
4219
4220    if (flags & WQ_UNBOUND) {
4221        wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4222        if (!wq->unbound_attrs)
4223            goto err_free_wq;
4224    }
4225
4226    va_start(args, lock_name);
4227    vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4228    va_end(args);
4229
4230    max_active = max_active ?: WQ_DFL_ACTIVE;
4231    max_active = wq_clamp_max_active(max_active, flags, wq->name);
4232
4233    /* init wq */
4234    wq->flags = flags;
4235    wq->saved_max_active = max_active;
4236    mutex_init(&wq->mutex);
4237    atomic_set(&wq->nr_pwqs_to_flush, 0);
4238    INIT_LIST_HEAD(&wq->pwqs);
4239    INIT_LIST_HEAD(&wq->flusher_queue);
4240    INIT_LIST_HEAD(&wq->flusher_overflow);
4241    INIT_LIST_HEAD(&wq->maydays);
4242
4243    lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4244    INIT_LIST_HEAD(&wq->list);
4245
4246    if (alloc_and_link_pwqs(wq) < 0)
4247        goto err_free_wq;
4248
4249    /*
4250     * Workqueues which may be used during memory reclaim should
4251     * have a rescuer to guarantee forward progress.
4252     */
4253    if (flags & WQ_MEM_RECLAIM) {
4254        struct worker *rescuer;
4255
4256        rescuer = alloc_worker();
4257        if (!rescuer)
4258            goto err_destroy;
4259
4260        rescuer->rescue_wq = wq;
4261        rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4262                           wq->name);
4263        if (IS_ERR(rescuer->task)) {
4264            kfree(rescuer);
4265            goto err_destroy;
4266        }
4267
4268        wq->rescuer = rescuer;
4269        rescuer->task->flags |= PF_NO_SETAFFINITY;
4270        wake_up_process(rescuer->task);
4271    }
4272
4273    if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4274        goto err_destroy;
4275
4276    /*
4277     * wq_pool_mutex protects global freeze state and workqueues list.
4278     * Grab it, adjust max_active and add the new @wq to workqueues
4279     * list.
4280     */
4281    mutex_lock(&wq_pool_mutex);
4282
4283    mutex_lock(&wq->mutex);
4284    for_each_pwq(pwq, wq)
4285        pwq_adjust_max_active(pwq);
4286    mutex_unlock(&wq->mutex);
4287
4288    list_add(&wq->list, &workqueues);
4289
4290    mutex_unlock(&wq_pool_mutex);
4291
4292    return wq;
4293
4294err_free_wq:
4295    free_workqueue_attrs(wq->unbound_attrs);
4296    kfree(wq);
4297    return NULL;
4298err_destroy:
4299    destroy_workqueue(wq);
4300    return NULL;
4301}
4302EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4303
4304/**
4305 * destroy_workqueue - safely terminate a workqueue
4306 * @wq: target workqueue
4307 *
4308 * Safely destroy a workqueue. All work currently pending will be done first.
4309 */
4310void destroy_workqueue(struct workqueue_struct *wq)
4311{
4312    struct pool_workqueue *pwq;
4313    int node;
4314
4315    /* drain it before proceeding with destruction */
4316    drain_workqueue(wq);
4317
4318    /* sanity checks */
4319    mutex_lock(&wq->mutex);
4320    for_each_pwq(pwq, wq) {
4321        int i;
4322
4323        for (i = 0; i < WORK_NR_COLORS; i++) {
4324            if (WARN_ON(pwq->nr_in_flight[i])) {
4325                mutex_unlock(&wq->mutex);
4326                return;
4327            }
4328        }
4329
4330        if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4331            WARN_ON(pwq->nr_active) ||
4332            WARN_ON(!list_empty(&pwq->delayed_works))) {
4333            mutex_unlock(&wq->mutex);
4334            return;
4335        }
4336    }
4337    mutex_unlock(&wq->mutex);
4338
4339    /*
4340     * wq list is used to freeze wq, remove from list after
4341     * flushing is complete in case freeze races us.
4342     */
4343    mutex_lock(&wq_pool_mutex);
4344    list_del_init(&wq->list);
4345    mutex_unlock(&wq_pool_mutex);
4346
4347    workqueue_sysfs_unregister(wq);
4348
4349    if (wq->rescuer) {
4350        kthread_stop(wq->rescuer->task);
4351        kfree(wq->rescuer);
4352        wq->rescuer = NULL;
4353    }
4354
4355    if (!(wq->flags & WQ_UNBOUND)) {
4356        /*
4357         * The base ref is never dropped on per-cpu pwqs. Directly
4358         * free the pwqs and wq.
4359         */
4360        free_percpu(wq->cpu_pwqs);
4361        kfree(wq);
4362    } else {
4363        /*
4364         * We're the sole accessor of @wq at this point. Directly
4365         * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4366         * @wq will be freed when the last pwq is released.
4367         */
4368        for_each_node(node) {
4369            pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4370            RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4371            put_pwq_unlocked(pwq);
4372        }
4373
4374        /*
4375         * Put dfl_pwq. @wq may be freed any time after dfl_pwq is
4376         * put. Don't access it afterwards.
4377         */
4378        pwq = wq->dfl_pwq;
4379        wq->dfl_pwq = NULL;
4380        put_pwq_unlocked(pwq);
4381    }
4382}
4383EXPORT_SYMBOL_GPL(destroy_workqueue);
4384
4385/**
4386 * workqueue_set_max_active - adjust max_active of a workqueue
4387 * @wq: target workqueue
4388 * @max_active: new max_active value.
4389 *
4390 * Set max_active of @wq to @max_active.
4391 *
4392 * CONTEXT:
4393 * Don't call from IRQ context.
4394 */
4395void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4396{
4397    struct pool_workqueue *pwq;
4398
4399    /* disallow meddling with max_active for ordered workqueues */
4400    if (WARN_ON(wq->flags & __WQ_ORDERED))
4401        return;
4402
4403    max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4404
4405    mutex_lock(&wq->mutex);
4406
4407    wq->saved_max_active = max_active;
4408
4409    for_each_pwq(pwq, wq)
4410        pwq_adjust_max_active(pwq);
4411
4412    mutex_unlock(&wq->mutex);
4413}
4414EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4415
4416/**
4417 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4418 *
4419 * Determine whether %current is a workqueue rescuer. Can be used from
4420 * work functions to determine whether it's being run off the rescuer task.
4421 *
4422 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4423 */
4424bool current_is_workqueue_rescuer(void)
4425{
4426    struct worker *worker = current_wq_worker();
4427
4428    return worker && worker->rescue_wq;
4429}
4430
4431/**
4432 * workqueue_congested - test whether a workqueue is congested
4433 * @cpu: CPU in question
4434 * @wq: target workqueue
4435 *
4436 * Test whether @wq's cpu workqueue for @cpu is congested. There is
4437 * no synchronization around this function and the test result is
4438 * unreliable and only useful as advisory hints or for debugging.
4439 *
4440 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4441 * Note that both per-cpu and unbound workqueues may be associated with
4442 * multiple pool_workqueues which have separate congested states. A
4443 * workqueue being congested on one CPU doesn't mean the workqueue is also
4444 * contested on other CPUs / NUMA nodes.
4445 *
4446 * Return:
4447 * %true if congested, %false otherwise.
4448 */
4449bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4450{
4451    struct pool_workqueue *pwq;
4452    bool ret;
4453
4454    rcu_read_lock_sched();
4455
4456    if (cpu == WORK_CPU_UNBOUND)
4457        cpu = smp_processor_id();
4458
4459    if (!(wq->flags & WQ_UNBOUND))
4460        pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4461    else
4462        pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4463
4464    ret = !list_empty(&pwq->delayed_works);
4465    rcu_read_unlock_sched();
4466
4467    return ret;
4468}
4469EXPORT_SYMBOL_GPL(workqueue_congested);
4470
4471/**
4472 * work_busy - test whether a work is currently pending or running
4473 * @work: the work to be tested
4474 *
4475 * Test whether @work is currently pending or running. There is no
4476 * synchronization around this function and the test result is
4477 * unreliable and only useful as advisory hints or for debugging.
4478 *
4479 * Return:
4480 * OR'd bitmask of WORK_BUSY_* bits.
4481 */
4482unsigned int work_busy(struct work_struct *work)
4483{
4484    struct worker_pool *pool;
4485    unsigned long flags;
4486    unsigned int ret = 0;
4487
4488    if (work_pending(work))
4489        ret |= WORK_BUSY_PENDING;
4490
4491    local_irq_save(flags);
4492    pool = get_work_pool(work);
4493    if (pool) {
4494        spin_lock(&pool->lock);
4495        if (find_worker_executing_work(pool, work))
4496            ret |= WORK_BUSY_RUNNING;
4497        spin_unlock(&pool->lock);
4498    }
4499    local_irq_restore(flags);
4500
4501    return ret;
4502}
4503EXPORT_SYMBOL_GPL(work_busy);
4504
4505/**
4506 * set_worker_desc - set description for the current work item
4507 * @fmt: printf-style format string
4508 * @...: arguments for the format string
4509 *
4510 * This function can be called by a running work function to describe what
4511 * the work item is about. If the worker task gets dumped, this
4512 * information will be printed out together to help debugging. The
4513 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4514 */
4515void set_worker_desc(const char *fmt, ...)
4516{
4517    struct worker *worker = current_wq_worker();
4518    va_list args;
4519
4520    if (worker) {
4521        va_start(args, fmt);
4522        vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4523        va_end(args);
4524        worker->desc_valid = true;
4525    }
4526}
4527
4528/**
4529 * print_worker_info - print out worker information and description
4530 * @log_lvl: the log level to use when printing
4531 * @task: target task
4532 *
4533 * If @task is a worker and currently executing a work item, print out the
4534 * name of the workqueue being serviced and worker description set with
4535 * set_worker_desc() by the currently executing work item.
4536 *
4537 * This function can be safely called on any task as long as the
4538 * task_struct itself is accessible. While safe, this function isn't
4539 * synchronized and may print out mixups or garbages of limited length.
4540 */
4541void print_worker_info(const char *log_lvl, struct task_struct *task)
4542{
4543    work_func_t *fn = NULL;
4544    char name[WQ_NAME_LEN] = { };
4545    char desc[WORKER_DESC_LEN] = { };
4546    struct pool_workqueue *pwq = NULL;
4547    struct workqueue_struct *wq = NULL;
4548    bool desc_valid = false;
4549    struct worker *worker;
4550
4551    if (!(task->flags & PF_WQ_WORKER))
4552        return;
4553
4554    /*
4555     * This function is called without any synchronization and @task
4556     * could be in any state. Be careful with dereferences.
4557     */
4558    worker = probe_kthread_data(task);
4559
4560    /*
4561     * Carefully copy the associated workqueue's workfn and name. Keep
4562     * the original last '\0' in case the original contains garbage.
4563     */
4564    probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4565    probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4566    probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4567    probe_kernel_read(name, wq->name, sizeof(name) - 1);
4568
4569    /* copy worker description */
4570    probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4571    if (desc_valid)
4572        probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4573
4574    if (fn || name[0] || desc[0]) {
4575        printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4576        if (desc[0])
4577            pr_cont(" (%s)", desc);
4578        pr_cont("\n");
4579    }
4580}
4581
4582/*
4583 * CPU hotplug.
4584 *
4585 * There are two challenges in supporting CPU hotplug. Firstly, there
4586 * are a lot of assumptions on strong associations among work, pwq and
4587 * pool which make migrating pending and scheduled works very
4588 * difficult to implement without impacting hot paths. Secondly,
4589 * worker pools serve mix of short, long and very long running works making
4590 * blocked draining impractical.
4591 *
4592 * This is solved by allowing the pools to be disassociated from the CPU
4593 * running as an unbound one and allowing it to be reattached later if the
4594 * cpu comes back online.
4595 */
4596
4597static void wq_unbind_fn(struct work_struct *work)
4598{
4599    int cpu = smp_processor_id();
4600    struct worker_pool *pool;
4601    struct worker *worker;
4602    int wi;
4603
4604    for_each_cpu_worker_pool(pool, cpu) {
4605        WARN_ON_ONCE(cpu != smp_processor_id());
4606
4607        mutex_lock(&pool->manager_mutex);
4608        spin_lock_irq(&pool->lock);
4609
4610        /*
4611         * We've blocked all manager operations. Make all workers
4612         * unbound and set DISASSOCIATED. Before this, all workers
4613         * except for the ones which are still executing works from
4614         * before the last CPU down must be on the cpu. After
4615         * this, they may become diasporas.
4616         */
4617        for_each_pool_worker(worker, wi, pool)
4618            worker->flags |= WORKER_UNBOUND;
4619
4620        pool->flags |= POOL_DISASSOCIATED;
4621
4622        spin_unlock_irq(&pool->lock);
4623        mutex_unlock(&pool->manager_mutex);
4624
4625        /*
4626         * Call schedule() so that we cross rq->lock and thus can
4627         * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4628         * This is necessary as scheduler callbacks may be invoked
4629         * from other cpus.
4630         */
4631        schedule();
4632
4633        /*
4634         * Sched callbacks are disabled now. Zap nr_running.
4635         * After this, nr_running stays zero and need_more_worker()
4636         * and keep_working() are always true as long as the
4637         * worklist is not empty. This pool now behaves as an
4638         * unbound (in terms of concurrency management) pool which
4639         * are served by workers tied to the pool.
4640         */
4641        atomic_set(&pool->nr_running, 0);
4642
4643        /*
4644         * With concurrency management just turned off, a busy
4645         * worker blocking could lead to lengthy stalls. Kick off
4646         * unbound chain execution of currently pending work items.
4647         */
4648        spin_lock_irq(&pool->lock);
4649        wake_up_worker(pool);
4650        spin_unlock_irq(&pool->lock);
4651    }
4652}
4653
4654/**
4655 * rebind_workers - rebind all workers of a pool to the associated CPU
4656 * @pool: pool of interest
4657 *
4658 * @pool->cpu is coming online. Rebind all workers to the CPU.
4659 */
4660static void rebind_workers(struct worker_pool *pool)
4661{
4662    struct worker *worker;
4663    int wi;
4664
4665    lockdep_assert_held(&pool->manager_mutex);
4666
4667    /*
4668     * Restore CPU affinity of all workers. As all idle workers should
4669     * be on the run-queue of the associated CPU before any local
4670     * wake-ups for concurrency management happen, restore CPU affinty
4671     * of all workers first and then clear UNBOUND. As we're called
4672     * from CPU_ONLINE, the following shouldn't fail.
4673     */
4674    for_each_pool_worker(worker, wi, pool)
4675        WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4676                          pool->attrs->cpumask) < 0);
4677
4678    spin_lock_irq(&pool->lock);
4679
4680    for_each_pool_worker(worker, wi, pool) {
4681        unsigned int worker_flags = worker->flags;
4682
4683        /*
4684         * A bound idle worker should actually be on the runqueue
4685         * of the associated CPU for local wake-ups targeting it to
4686         * work. Kick all idle workers so that they migrate to the
4687         * associated CPU. Doing this in the same loop as
4688         * replacing UNBOUND with REBOUND is safe as no worker will
4689         * be bound before @pool->lock is released.
4690         */
4691        if (worker_flags & WORKER_IDLE)
4692            wake_up_process(worker->task);
4693
4694        /*
4695         * We want to clear UNBOUND but can't directly call
4696         * worker_clr_flags() or adjust nr_running. Atomically
4697         * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4698         * @worker will clear REBOUND using worker_clr_flags() when
4699         * it initiates the next execution cycle thus restoring
4700         * concurrency management. Note that when or whether
4701         * @worker clears REBOUND doesn't affect correctness.
4702         *
4703         * ACCESS_ONCE() is necessary because @worker->flags may be
4704         * tested without holding any lock in
4705         * wq_worker_waking_up(). Without it, NOT_RUNNING test may
4706         * fail incorrectly leading to premature concurrency
4707         * management operations.
4708         */
4709        WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4710        worker_flags |= WORKER_REBOUND;
4711        worker_flags &= ~WORKER_UNBOUND;
4712        ACCESS_ONCE(worker->flags) = worker_flags;
4713    }
4714
4715    spin_unlock_irq(&pool->lock);
4716}
4717
4718/**
4719 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4720 * @pool: unbound pool of interest
4721 * @cpu: the CPU which is coming up
4722 *
4723 * An unbound pool may end up with a cpumask which doesn't have any online
4724 * CPUs. When a worker of such pool get scheduled, the scheduler resets
4725 * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any
4726 * online CPU before, cpus_allowed of all its workers should be restored.
4727 */
4728static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4729{
4730    static cpumask_t cpumask;
4731    struct worker *worker;
4732    int wi;
4733
4734    lockdep_assert_held(&pool->manager_mutex);
4735
4736    /* is @cpu allowed for @pool? */
4737    if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4738        return;
4739
4740    /* is @cpu the only online CPU? */
4741    cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4742    if (cpumask_weight(&cpumask) != 1)
4743        return;
4744
4745    /* as we're called from CPU_ONLINE, the following shouldn't fail */
4746    for_each_pool_worker(worker, wi, pool)
4747        WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4748                          pool->attrs->cpumask) < 0);
4749}
4750
4751/*
4752 * Workqueues should be brought up before normal priority CPU notifiers.
4753 * This will be registered high priority CPU notifier.
4754 */
4755static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4756                           unsigned long action,
4757                           void *hcpu)
4758{
4759    int cpu = (unsigned long)hcpu;
4760    struct worker_pool *pool;
4761    struct workqueue_struct *wq;
4762    int pi;
4763
4764    switch (action & ~CPU_TASKS_FROZEN) {
4765    case CPU_UP_PREPARE:
4766        for_each_cpu_worker_pool(pool, cpu) {
4767            if (pool->nr_workers)
4768                continue;
4769            if (create_and_start_worker(pool) < 0)
4770                return NOTIFY_BAD;
4771        }
4772        break;
4773
4774    case CPU_DOWN_FAILED:
4775    case CPU_ONLINE:
4776        mutex_lock(&wq_pool_mutex);
4777
4778        for_each_pool(pool, pi) {
4779            mutex_lock(&pool->manager_mutex);
4780
4781            if (pool->cpu == cpu) {
4782                spin_lock_irq(&pool->lock);
4783                pool->flags &= ~POOL_DISASSOCIATED;
4784                spin_unlock_irq(&pool->lock);
4785
4786                rebind_workers(pool);
4787            } else if (pool->cpu < 0) {
4788                restore_unbound_workers_cpumask(pool, cpu);
4789            }
4790
4791            mutex_unlock(&pool->manager_mutex);
4792        }
4793
4794        /* update NUMA affinity of unbound workqueues */
4795        list_for_each_entry(wq, &workqueues, list)
4796            wq_update_unbound_numa(wq, cpu, true);
4797
4798        mutex_unlock(&wq_pool_mutex);
4799        break;
4800    }
4801    return NOTIFY_OK;
4802}
4803
4804/*
4805 * Workqueues should be brought down after normal priority CPU notifiers.
4806 * This will be registered as low priority CPU notifier.
4807 */
4808static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4809                         unsigned long action,
4810                         void *hcpu)
4811{
4812    int cpu = (unsigned long)hcpu;
4813    struct work_struct unbind_work;
4814    struct workqueue_struct *wq;
4815
4816    switch (action & ~CPU_TASKS_FROZEN) {
4817    case CPU_DOWN_PREPARE:
4818        /* unbinding per-cpu workers should happen on the local CPU */
4819        INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4820        queue_work_on(cpu, system_highpri_wq, &unbind_work);
4821
4822        /* update NUMA affinity of unbound workqueues */
4823        mutex_lock(&wq_pool_mutex);
4824        list_for_each_entry(wq, &workqueues, list)
4825            wq_update_unbound_numa(wq, cpu, false);
4826        mutex_unlock(&wq_pool_mutex);
4827
4828        /* wait for per-cpu unbinding to finish */
4829        flush_work(&unbind_work);
4830        destroy_work_on_stack(&unbind_work);
4831        break;
4832    }
4833    return NOTIFY_OK;
4834}
4835
4836#ifdef CONFIG_SMP
4837
4838struct work_for_cpu {
4839    struct work_struct work;
4840    long (*fn)(void *);
4841    void *arg;
4842    long ret;
4843};
4844
4845static void work_for_cpu_fn(struct work_struct *work)
4846{
4847    struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4848
4849    wfc->ret = wfc->fn(wfc->arg);
4850}
4851
4852/**
4853 * work_on_cpu - run a function in user context on a particular cpu
4854 * @cpu: the cpu to run on
4855 * @fn: the function to run
4856 * @arg: the function arg
4857 *
4858 * It is up to the caller to ensure that the cpu doesn't go offline.
4859 * The caller must not hold any locks which would prevent @fn from completing.
4860 *
4861 * Return: The value @fn returns.
4862 */
4863long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4864{
4865    struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4866
4867    INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4868    schedule_work_on(cpu, &wfc.work);
4869    flush_work(&wfc.work);
4870    destroy_work_on_stack(&wfc.work);
4871    return wfc.ret;
4872}
4873EXPORT_SYMBOL_GPL(work_on_cpu);
4874#endif /* CONFIG_SMP */
4875
4876#ifdef CONFIG_FREEZER
4877
4878/**
4879 * freeze_workqueues_begin - begin freezing workqueues
4880 *
4881 * Start freezing workqueues. After this function returns, all freezable
4882 * workqueues will queue new works to their delayed_works list instead of
4883 * pool->worklist.
4884 *
4885 * CONTEXT:
4886 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4887 */
4888void freeze_workqueues_begin(void)
4889{
4890    struct worker_pool *pool;
4891    struct workqueue_struct *wq;
4892    struct pool_workqueue *pwq;
4893    int pi;
4894
4895    mutex_lock(&wq_pool_mutex);
4896
4897    WARN_ON_ONCE(workqueue_freezing);
4898    workqueue_freezing = true;
4899
4900    /* set FREEZING */
4901    for_each_pool(pool, pi) {
4902        spin_lock_irq(&pool->lock);
4903        WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4904        pool->flags |= POOL_FREEZING;
4905        spin_unlock_irq(&pool->lock);
4906    }
4907
4908    list_for_each_entry(wq, &workqueues, list) {
4909        mutex_lock(&wq->mutex);
4910        for_each_pwq(pwq, wq)
4911            pwq_adjust_max_active(pwq);
4912        mutex_unlock(&wq->mutex);
4913    }
4914
4915    mutex_unlock(&wq_pool_mutex);
4916}
4917
4918/**
4919 * freeze_workqueues_busy - are freezable workqueues still busy?
4920 *
4921 * Check whether freezing is complete. This function must be called
4922 * between freeze_workqueues_begin() and thaw_workqueues().
4923 *
4924 * CONTEXT:
4925 * Grabs and releases wq_pool_mutex.
4926 *
4927 * Return:
4928 * %true if some freezable workqueues are still busy. %false if freezing
4929 * is complete.
4930 */
4931bool freeze_workqueues_busy(void)
4932{
4933    bool busy = false;
4934    struct workqueue_struct *wq;
4935    struct pool_workqueue *pwq;
4936
4937    mutex_lock(&wq_pool_mutex);
4938
4939    WARN_ON_ONCE(!workqueue_freezing);
4940
4941    list_for_each_entry(wq, &workqueues, list) {
4942        if (!(wq->flags & WQ_FREEZABLE))
4943            continue;
4944        /*
4945         * nr_active is monotonically decreasing. It's safe
4946         * to peek without lock.
4947         */
4948        rcu_read_lock_sched();
4949        for_each_pwq(pwq, wq) {
4950            WARN_ON_ONCE(pwq->nr_active < 0);
4951            if (pwq->nr_active) {
4952                busy = true;
4953                rcu_read_unlock_sched();
4954                goto out_unlock;
4955            }
4956        }
4957        rcu_read_unlock_sched();
4958    }
4959out_unlock:
4960    mutex_unlock(&wq_pool_mutex);
4961    return busy;
4962}
4963
4964/**
4965 * thaw_workqueues - thaw workqueues
4966 *
4967 * Thaw workqueues. Normal queueing is restored and all collected
4968 * frozen works are transferred to their respective pool worklists.
4969 *
4970 * CONTEXT:
4971 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4972 */
4973void thaw_workqueues(void)
4974{
4975    struct workqueue_struct *wq;
4976    struct pool_workqueue *pwq;
4977    struct worker_pool *pool;
4978    int pi;
4979
4980    mutex_lock(&wq_pool_mutex);
4981
4982    if (!workqueue_freezing)
4983        goto out_unlock;
4984
4985    /* clear FREEZING */
4986    for_each_pool(pool, pi) {
4987        spin_lock_irq(&pool->lock);
4988        WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4989        pool->flags &= ~POOL_FREEZING;
4990        spin_unlock_irq(&pool->lock);
4991    }
4992
4993    /* restore max_active and repopulate worklist */
4994    list_for_each_entry(wq, &workqueues, list) {
4995        mutex_lock(&wq->mutex);
4996        for_each_pwq(pwq, wq)
4997            pwq_adjust_max_active(pwq);
4998        mutex_unlock(&wq->mutex);
4999    }
5000
5001    workqueue_freezing = false;
5002out_unlock:
5003    mutex_unlock(&wq_pool_mutex);
5004}
5005#endif /* CONFIG_FREEZER */
5006
5007static void __init wq_numa_init(void)
5008{
5009    cpumask_var_t *tbl;
5010    int node, cpu;
5011
5012    /* determine NUMA pwq table len - highest node id + 1 */
5013    for_each_node(node)
5014        wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
5015
5016    if (num_possible_nodes() <= 1)
5017        return;
5018
5019    if (wq_disable_numa) {
5020        pr_info("workqueue: NUMA affinity support disabled\n");
5021        return;
5022    }
5023
5024    wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5025    BUG_ON(!wq_update_unbound_numa_attrs_buf);
5026
5027    /*
5028     * We want masks of possible CPUs of each node which isn't readily
5029     * available. Build one from cpu_to_node() which should have been
5030     * fully initialized by now.
5031     */
5032    tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
5033    BUG_ON(!tbl);
5034
5035    for_each_node(node)
5036        BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5037                node_online(node) ? node : NUMA_NO_NODE));
5038
5039    for_each_possible_cpu(cpu) {
5040        node = cpu_to_node(cpu);
5041        if (WARN_ON(node == NUMA_NO_NODE)) {
5042            pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5043            /* happens iff arch is bonkers, let's just proceed */
5044            return;
5045        }
5046        cpumask_set_cpu(cpu, tbl[node]);
5047    }
5048
5049    wq_numa_possible_cpumask = tbl;
5050    wq_numa_enabled = true;
5051}
5052
5053static int __init init_workqueues(void)
5054{
5055    int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5056    int i, cpu;
5057
5058    WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5059
5060    pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5061
5062    cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5063    hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
5064
5065    wq_numa_init();
5066
5067    /* initialize CPU pools */
5068    for_each_possible_cpu(cpu) {
5069        struct worker_pool *pool;
5070
5071        i = 0;
5072        for_each_cpu_worker_pool(pool, cpu) {
5073            BUG_ON(init_worker_pool(pool));
5074            pool->cpu = cpu;
5075            cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5076            pool->attrs->nice = std_nice[i++];
5077            pool->node = cpu_to_node(cpu);
5078
5079            /* alloc pool ID */
5080            mutex_lock(&wq_pool_mutex);
5081            BUG_ON(worker_pool_assign_id(pool));
5082            mutex_unlock(&wq_pool_mutex);
5083        }
5084    }
5085
5086    /* create the initial worker */
5087    for_each_online_cpu(cpu) {
5088        struct worker_pool *pool;
5089
5090        for_each_cpu_worker_pool(pool, cpu) {
5091            pool->flags &= ~POOL_DISASSOCIATED;
5092            BUG_ON(create_and_start_worker(pool) < 0);
5093        }
5094    }
5095
5096    /* create default unbound and ordered wq attrs */
5097    for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5098        struct workqueue_attrs *attrs;
5099
5100        BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5101        attrs->nice = std_nice[i];
5102        unbound_std_wq_attrs[i] = attrs;
5103
5104        /*
5105         * An ordered wq should have only one pwq as ordering is
5106         * guaranteed by max_active which is enforced by pwqs.
5107         * Turn off NUMA so that dfl_pwq is used for all nodes.
5108         */
5109        BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5110        attrs->nice = std_nice[i];
5111        attrs->no_numa = true;
5112        ordered_wq_attrs[i] = attrs;
5113    }
5114
5115    system_wq = alloc_workqueue("events", 0, 0);
5116    system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5117    system_long_wq = alloc_workqueue("events_long", 0, 0);
5118    system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5119                        WQ_UNBOUND_MAX_ACTIVE);
5120    system_freezable_wq = alloc_workqueue("events_freezable",
5121                          WQ_FREEZABLE, 0);
5122    system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5123                          WQ_POWER_EFFICIENT, 0);
5124    system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5125                          WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5126                          0);
5127    BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5128           !system_unbound_wq || !system_freezable_wq ||
5129           !system_power_efficient_wq ||
5130           !system_freezable_power_efficient_wq);
5131    return 0;
5132}
5133early_initcall(init_workqueues);
5134

Archive Download this file



interactive