linux/kernel/workqueue.c
<<
>>
Prefs
   1/*
   2 * kernel/workqueue.c - generic async execution with shared worker pool
   3 *
   4 * Copyright (C) 2002           Ingo Molnar
   5 *
   6 *   Derived from the taskqueue/keventd code by:
   7 *     David Woodhouse <dwmw2@infradead.org>
   8 *     Andrew Morton
   9 *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
  10 *     Theodore Ts'o <tytso@mit.edu>
  11 *
  12 * Made to use alloc_percpu by Christoph Lameter.
  13 *
  14 * Copyright (C) 2010           SUSE Linux Products GmbH
  15 * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
  16 *
  17 * This is the generic async execution mechanism.  Work items as are
  18 * executed in process context.  The worker pool is shared and
  19 * automatically managed.  There is one worker pool for each CPU and
  20 * one extra for works which are better served by workers which are
  21 * not bound to any specific CPU.
  22 *
  23 * Please read Documentation/workqueue.txt for details.
  24 */
  25
  26#include <linux/export.h>
  27#include <linux/kernel.h>
  28#include <linux/sched.h>
  29#include <linux/init.h>
  30#include <linux/signal.h>
  31#include <linux/completion.h>
  32#include <linux/workqueue.h>
  33#include <linux/slab.h>
  34#include <linux/cpu.h>
  35#include <linux/notifier.h>
  36#include <linux/kthread.h>
  37#include <linux/hardirq.h>
  38#include <linux/mempolicy.h>
  39#include <linux/freezer.h>
  40#include <linux/kallsyms.h>
  41#include <linux/debug_locks.h>
  42#include <linux/lockdep.h>
  43#include <linux/idr.h>
  44
  45#include "workqueue_sched.h"
  46
  47enum {
  48        /*
  49         * global_cwq flags
  50         *
  51         * A bound gcwq is either associated or disassociated with its CPU.
  52         * While associated (!DISASSOCIATED), all workers are bound to the
  53         * CPU and none has %WORKER_UNBOUND set and concurrency management
  54         * is in effect.
  55         *
  56         * While DISASSOCIATED, the cpu may be offline and all workers have
  57         * %WORKER_UNBOUND set and concurrency management disabled, and may
  58         * be executing on any CPU.  The gcwq behaves as an unbound one.
  59         *
  60         * Note that DISASSOCIATED can be flipped only while holding
  61         * assoc_mutex of all pools on the gcwq to avoid changing binding
  62         * state while create_worker() is in progress.
  63         */
  64        GCWQ_DISASSOCIATED      = 1 << 0,       /* cpu can't serve workers */
  65        GCWQ_FREEZING           = 1 << 1,       /* freeze in progress */
  66
  67        /* pool flags */
  68        POOL_MANAGE_WORKERS     = 1 << 0,       /* need to manage workers */
  69        POOL_MANAGING_WORKERS   = 1 << 1,       /* managing workers */
  70
  71        /* worker flags */
  72        WORKER_STARTED          = 1 << 0,       /* started */
  73        WORKER_DIE              = 1 << 1,       /* die die die */
  74        WORKER_IDLE             = 1 << 2,       /* is idle */
  75        WORKER_PREP             = 1 << 3,       /* preparing to run works */
  76        WORKER_CPU_INTENSIVE    = 1 << 6,       /* cpu intensive */
  77        WORKER_UNBOUND          = 1 << 7,       /* worker is unbound */
  78
  79        WORKER_NOT_RUNNING      = WORKER_PREP | WORKER_UNBOUND |
  80                                  WORKER_CPU_INTENSIVE,
  81
  82        NR_WORKER_POOLS         = 2,            /* # worker pools per gcwq */
  83
  84        BUSY_WORKER_HASH_ORDER  = 6,            /* 64 pointers */
  85        BUSY_WORKER_HASH_SIZE   = 1 << BUSY_WORKER_HASH_ORDER,
  86        BUSY_WORKER_HASH_MASK   = BUSY_WORKER_HASH_SIZE - 1,
  87
  88        MAX_IDLE_WORKERS_RATIO  = 4,            /* 1/4 of busy can be idle */
  89        IDLE_WORKER_TIMEOUT     = 300 * HZ,     /* keep idle ones for 5 mins */
  90
  91        MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
  92                                                /* call for help after 10ms
  93                                                   (min two ticks) */
  94        MAYDAY_INTERVAL         = HZ / 10,      /* and then every 100ms */
  95        CREATE_COOLDOWN         = HZ,           /* time to breath after fail */
  96
  97        /*
  98         * Rescue workers are used only on emergencies and shared by
  99         * all cpus.  Give -20.
 100         */
 101        RESCUER_NICE_LEVEL      = -20,
 102        HIGHPRI_NICE_LEVEL      = -20,
 103};
 104
 105/*
 106 * Structure fields follow one of the following exclusion rules.
 107 *
 108 * I: Modifiable by initialization/destruction paths and read-only for
 109 *    everyone else.
 110 *
 111 * P: Preemption protected.  Disabling preemption is enough and should
 112 *    only be modified and accessed from the local cpu.
 113 *
 114 * L: gcwq->lock protected.  Access with gcwq->lock held.
 115 *
 116 * X: During normal operation, modification requires gcwq->lock and
 117 *    should be done only from local cpu.  Either disabling preemption
 118 *    on local cpu or grabbing gcwq->lock is enough for read access.
 119 *    If GCWQ_DISASSOCIATED is set, it's identical to L.
 120 *
 121 * F: wq->flush_mutex protected.
 122 *
 123 * W: workqueue_lock protected.
 124 */
 125
 126struct global_cwq;
 127struct worker_pool;
 128
 129/*
 130 * The poor guys doing the actual heavy lifting.  All on-duty workers
 131 * are either serving the manager role, on idle list or on busy hash.
 132 */
 133struct worker {
 134        /* on idle list while idle, on busy hash table while busy */
 135        union {
 136                struct list_head        entry;  /* L: while idle */
 137                struct hlist_node       hentry; /* L: while busy */
 138        };
 139
 140        struct work_struct      *current_work;  /* L: work being processed */
 141        struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
 142        struct list_head        scheduled;      /* L: scheduled works */
 143        struct task_struct      *task;          /* I: worker task */
 144        struct worker_pool      *pool;          /* I: the associated pool */
 145        /* 64 bytes boundary on 64bit, 32 on 32bit */
 146        unsigned long           last_active;    /* L: last active timestamp */
 147        unsigned int            flags;          /* X: flags */
 148        int                     id;             /* I: worker id */
 149
 150        /* for rebinding worker to CPU */
 151        struct work_struct      rebind_work;    /* L: for busy worker */
 152};
 153
 154struct worker_pool {
 155        struct global_cwq       *gcwq;          /* I: the owning gcwq */
 156        unsigned int            flags;          /* X: flags */
 157
 158        struct list_head        worklist;       /* L: list of pending works */
 159        int                     nr_workers;     /* L: total number of workers */
 160
 161        /* nr_idle includes the ones off idle_list for rebinding */
 162        int                     nr_idle;        /* L: currently idle ones */
 163
 164        struct list_head        idle_list;      /* X: list of idle workers */
 165        struct timer_list       idle_timer;     /* L: worker idle timeout */
 166        struct timer_list       mayday_timer;   /* L: SOS timer for workers */
 167
 168        struct mutex            assoc_mutex;    /* protect GCWQ_DISASSOCIATED */
 169        struct ida              worker_ida;     /* L: for worker IDs */
 170};
 171
 172/*
 173 * Global per-cpu workqueue.  There's one and only one for each cpu
 174 * and all works are queued and processed here regardless of their
 175 * target workqueues.
 176 */
 177struct global_cwq {
 178        spinlock_t              lock;           /* the gcwq lock */
 179        unsigned int            cpu;            /* I: the associated cpu */
 180        unsigned int            flags;          /* L: GCWQ_* flags */
 181
 182        /* workers are chained either in busy_hash or pool idle_list */
 183        struct hlist_head       busy_hash[BUSY_WORKER_HASH_SIZE];
 184                                                /* L: hash of busy workers */
 185
 186        struct worker_pool      pools[NR_WORKER_POOLS];
 187                                                /* normal and highpri pools */
 188} ____cacheline_aligned_in_smp;
 189
 190/*
 191 * The per-CPU workqueue.  The lower WORK_STRUCT_FLAG_BITS of
 192 * work_struct->data are used for flags and thus cwqs need to be
 193 * aligned at two's power of the number of flag bits.
 194 */
 195struct cpu_workqueue_struct {
 196        struct worker_pool      *pool;          /* I: the associated pool */
 197        struct workqueue_struct *wq;            /* I: the owning workqueue */
 198        int                     work_color;     /* L: current color */
 199        int                     flush_color;    /* L: flushing color */
 200        int                     nr_in_flight[WORK_NR_COLORS];
 201                                                /* L: nr of in_flight works */
 202        int                     nr_active;      /* L: nr of active works */
 203        int                     max_active;     /* L: max active works */
 204        struct list_head        delayed_works;  /* L: delayed works */
 205};
 206
 207/*
 208 * Structure used to wait for workqueue flush.
 209 */
 210struct wq_flusher {
 211        struct list_head        list;           /* F: list of flushers */
 212        int                     flush_color;    /* F: flush color waiting for */
 213        struct completion       done;           /* flush completion */
 214};
 215
 216/*
 217 * All cpumasks are assumed to be always set on UP and thus can't be
 218 * used to determine whether there's something to be done.
 219 */
 220#ifdef CONFIG_SMP
 221typedef cpumask_var_t mayday_mask_t;
 222#define mayday_test_and_set_cpu(cpu, mask)      \
 223        cpumask_test_and_set_cpu((cpu), (mask))
 224#define mayday_clear_cpu(cpu, mask)             cpumask_clear_cpu((cpu), (mask))
 225#define for_each_mayday_cpu(cpu, mask)          for_each_cpu((cpu), (mask))
 226#define alloc_mayday_mask(maskp, gfp)           zalloc_cpumask_var((maskp), (gfp))
 227#define free_mayday_mask(mask)                  free_cpumask_var((mask))
 228#else
 229typedef unsigned long mayday_mask_t;
 230#define mayday_test_and_set_cpu(cpu, mask)      test_and_set_bit(0, &(mask))
 231#define mayday_clear_cpu(cpu, mask)             clear_bit(0, &(mask))
 232#define for_each_mayday_cpu(cpu, mask)          if ((cpu) = 0, (mask))
 233#define alloc_mayday_mask(maskp, gfp)           true
 234#define free_mayday_mask(mask)                  do { } while (0)
 235#endif
 236
 237/*
 238 * The externally visible workqueue abstraction is an array of
 239 * per-CPU workqueues:
 240 */
 241struct workqueue_struct {
 242        unsigned int            flags;          /* W: WQ_* flags */
 243        union {
 244                struct cpu_workqueue_struct __percpu    *pcpu;
 245                struct cpu_workqueue_struct             *single;
 246                unsigned long                           v;
 247        } cpu_wq;                               /* I: cwq's */
 248        struct list_head        list;           /* W: list of all workqueues */
 249
 250        struct mutex            flush_mutex;    /* protects wq flushing */
 251        int                     work_color;     /* F: current work color */
 252        int                     flush_color;    /* F: current flush color */
 253        atomic_t                nr_cwqs_to_flush; /* flush in progress */
 254        struct wq_flusher       *first_flusher; /* F: first flusher */
 255        struct list_head        flusher_queue;  /* F: flush waiters */
 256        struct list_head        flusher_overflow; /* F: flush overflow list */
 257
 258        mayday_mask_t           mayday_mask;    /* cpus requesting rescue */
 259        struct worker           *rescuer;       /* I: rescue worker */
 260
 261        int                     nr_drainers;    /* W: drain in progress */
 262        int                     saved_max_active; /* W: saved cwq max_active */
 263#ifdef CONFIG_LOCKDEP
 264        struct lockdep_map      lockdep_map;
 265#endif
 266        char                    name[];         /* I: workqueue name */
 267};
 268
 269struct workqueue_struct *system_wq __read_mostly;
 270EXPORT_SYMBOL_GPL(system_wq);
 271struct workqueue_struct *system_highpri_wq __read_mostly;
 272EXPORT_SYMBOL_GPL(system_highpri_wq);
 273struct workqueue_struct *system_long_wq __read_mostly;
 274EXPORT_SYMBOL_GPL(system_long_wq);
 275struct workqueue_struct *system_unbound_wq __read_mostly;
 276EXPORT_SYMBOL_GPL(system_unbound_wq);
 277struct workqueue_struct *system_freezable_wq __read_mostly;
 278EXPORT_SYMBOL_GPL(system_freezable_wq);
 279
 280#define CREATE_TRACE_POINTS
 281#include <trace/events/workqueue.h>
 282
 283#define for_each_worker_pool(pool, gcwq)                                \
 284        for ((pool) = &(gcwq)->pools[0];                                \
 285             (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
 286
 287#define for_each_busy_worker(worker, i, pos, gcwq)                      \
 288        for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)                     \
 289                hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
 290
 291static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
 292                                  unsigned int sw)
 293{
 294        if (cpu < nr_cpu_ids) {
 295                if (sw & 1) {
 296                        cpu = cpumask_next(cpu, mask);
 297                        if (cpu < nr_cpu_ids)
 298                                return cpu;
 299                }
 300                if (sw & 2)
 301                        return WORK_CPU_UNBOUND;
 302        }
 303        return WORK_CPU_NONE;
 304}
 305
 306static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
 307                                struct workqueue_struct *wq)
 308{
 309        return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
 310}
 311
 312/*
 313 * CPU iterators
 314 *
 315 * An extra gcwq is defined for an invalid cpu number
 316 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
 317 * specific CPU.  The following iterators are similar to
 318 * for_each_*_cpu() iterators but also considers the unbound gcwq.
 319 *
 320 * for_each_gcwq_cpu()          : possible CPUs + WORK_CPU_UNBOUND
 321 * for_each_online_gcwq_cpu()   : online CPUs + WORK_CPU_UNBOUND
 322 * for_each_cwq_cpu()           : possible CPUs for bound workqueues,
 323 *                                WORK_CPU_UNBOUND for unbound workqueues
 324 */
 325#define for_each_gcwq_cpu(cpu)                                          \
 326        for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3);         \
 327             (cpu) < WORK_CPU_NONE;                                     \
 328             (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
 329
 330#define for_each_online_gcwq_cpu(cpu)                                   \
 331        for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3);           \
 332             (cpu) < WORK_CPU_NONE;                                     \
 333             (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
 334
 335#define for_each_cwq_cpu(cpu, wq)                                       \
 336        for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq));        \
 337             (cpu) < WORK_CPU_NONE;                                     \
 338             (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
 339
 340#ifdef CONFIG_DEBUG_OBJECTS_WORK
 341
 342static struct debug_obj_descr work_debug_descr;
 343
 344static void *work_debug_hint(void *addr)
 345{
 346        return ((struct work_struct *) addr)->func;
 347}
 348
 349/*
 350 * fixup_init is called when:
 351 * - an active object is initialized
 352 */
 353static int work_fixup_init(void *addr, enum debug_obj_state state)
 354{
 355        struct work_struct *work = addr;
 356
 357        switch (state) {
 358        case ODEBUG_STATE_ACTIVE:
 359                cancel_work_sync(work);
 360                debug_object_init(work, &work_debug_descr);
 361                return 1;
 362        default:
 363                return 0;
 364        }
 365}
 366
 367/*
 368 * fixup_activate is called when:
 369 * - an active object is activated
 370 * - an unknown object is activated (might be a statically initialized object)
 371 */
 372static int work_fixup_activate(void *addr, enum debug_obj_state state)
 373{
 374        struct work_struct *work = addr;
 375
 376        switch (state) {
 377
 378        case ODEBUG_STATE_NOTAVAILABLE:
 379                /*
 380                 * This is not really a fixup. The work struct was
 381                 * statically initialized. We just make sure that it
 382                 * is tracked in the object tracker.
 383                 */
 384                if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
 385                        debug_object_init(work, &work_debug_descr);
 386                        debug_object_activate(work, &work_debug_descr);
 387                        return 0;
 388                }
 389                WARN_ON_ONCE(1);
 390                return 0;
 391
 392        case ODEBUG_STATE_ACTIVE:
 393                WARN_ON(1);
 394
 395        default:
 396                return 0;
 397        }
 398}
 399
 400/*
 401 * fixup_free is called when:
 402 * - an active object is freed
 403 */
 404static int work_fixup_free(void *addr, enum debug_obj_state state)
 405{
 406        struct work_struct *work = addr;
 407
 408        switch (state) {
 409        case ODEBUG_STATE_ACTIVE:
 410                cancel_work_sync(work);
 411                debug_object_free(work, &work_debug_descr);
 412                return 1;
 413        default:
 414                return 0;
 415        }
 416}
 417
 418static struct debug_obj_descr work_debug_descr = {
 419        .name           = "work_struct",
 420        .debug_hint     = work_debug_hint,
 421        .fixup_init     = work_fixup_init,
 422        .fixup_activate = work_fixup_activate,
 423        .fixup_free     = work_fixup_free,
 424};
 425
 426static inline void debug_work_activate(struct work_struct *work)
 427{
 428        debug_object_activate(work, &work_debug_descr);
 429}
 430
 431static inline void debug_work_deactivate(struct work_struct *work)
 432{
 433        debug_object_deactivate(work, &work_debug_descr);
 434}
 435
 436void __init_work(struct work_struct *work, int onstack)
 437{
 438        if (onstack)
 439                debug_object_init_on_stack(work, &work_debug_descr);
 440        else
 441                debug_object_init(work, &work_debug_descr);
 442}
 443EXPORT_SYMBOL_GPL(__init_work);
 444
 445void destroy_work_on_stack(struct work_struct *work)
 446{
 447        debug_object_free(work, &work_debug_descr);
 448}
 449EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 450
 451#else
 452static inline void debug_work_activate(struct work_struct *work) { }
 453static inline void debug_work_deactivate(struct work_struct *work) { }
 454#endif
 455
 456/* Serializes the accesses to the list of workqueues. */
 457static DEFINE_SPINLOCK(workqueue_lock);
 458static LIST_HEAD(workqueues);
 459static bool workqueue_freezing;         /* W: have wqs started freezing? */
 460
 461/*
 462 * The almighty global cpu workqueues.  nr_running is the only field
 463 * which is expected to be used frequently by other cpus via
 464 * try_to_wake_up().  Put it in a separate cacheline.
 465 */
 466static DEFINE_PER_CPU(struct global_cwq, global_cwq);
 467static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
 468
 469/*
 470 * Global cpu workqueue and nr_running counter for unbound gcwq.  The
 471 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
 472 * workers have WORKER_UNBOUND set.
 473 */
 474static struct global_cwq unbound_global_cwq;
 475static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
 476        [0 ... NR_WORKER_POOLS - 1]     = ATOMIC_INIT(0),       /* always 0 */
 477};
 478
 479static int worker_thread(void *__worker);
 480
 481static int worker_pool_pri(struct worker_pool *pool)
 482{
 483        return pool - pool->gcwq->pools;
 484}
 485
 486static struct global_cwq *get_gcwq(unsigned int cpu)
 487{
 488        if (cpu != WORK_CPU_UNBOUND)
 489                return &per_cpu(global_cwq, cpu);
 490        else
 491                return &unbound_global_cwq;
 492}
 493
 494static atomic_t *get_pool_nr_running(struct worker_pool *pool)
 495{
 496        int cpu = pool->gcwq->cpu;
 497        int idx = worker_pool_pri(pool);
 498
 499        if (cpu != WORK_CPU_UNBOUND)
 500                return &per_cpu(pool_nr_running, cpu)[idx];
 501        else
 502                return &unbound_pool_nr_running[idx];
 503}
 504
 505static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
 506                                            struct workqueue_struct *wq)
 507{
 508        if (!(wq->flags & WQ_UNBOUND)) {
 509                if (likely(cpu < nr_cpu_ids))
 510                        return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
 511        } else if (likely(cpu == WORK_CPU_UNBOUND))
 512                return wq->cpu_wq.single;
 513        return NULL;
 514}
 515
 516static unsigned int work_color_to_flags(int color)
 517{
 518        return color << WORK_STRUCT_COLOR_SHIFT;
 519}
 520
 521static int get_work_color(struct work_struct *work)
 522{
 523        return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 524                ((1 << WORK_STRUCT_COLOR_BITS) - 1);
 525}
 526
 527static int work_next_color(int color)
 528{
 529        return (color + 1) % WORK_NR_COLORS;
 530}
 531
 532/*
 533 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
 534 * contain the pointer to the queued cwq.  Once execution starts, the flag
 535 * is cleared and the high bits contain OFFQ flags and CPU number.
 536 *
 537 * set_work_cwq(), set_work_cpu_and_clear_pending(), mark_work_canceling()
 538 * and clear_work_data() can be used to set the cwq, cpu or clear
 539 * work->data.  These functions should only be called while the work is
 540 * owned - ie. while the PENDING bit is set.
 541 *
 542 * get_work_[g]cwq() can be used to obtain the gcwq or cwq corresponding to
 543 * a work.  gcwq is available once the work has been queued anywhere after
 544 * initialization until it is sync canceled.  cwq is available only while
 545 * the work item is queued.
 546 *
 547 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
 548 * canceled.  While being canceled, a work item may have its PENDING set
 549 * but stay off timer and worklist for arbitrarily long and nobody should
 550 * try to steal the PENDING bit.
 551 */
 552static inline void set_work_data(struct work_struct *work, unsigned long data,
 553                                 unsigned long flags)
 554{
 555        BUG_ON(!work_pending(work));
 556        atomic_long_set(&work->data, data | flags | work_static(work));
 557}
 558
 559static void set_work_cwq(struct work_struct *work,
 560                         struct cpu_workqueue_struct *cwq,
 561                         unsigned long extra_flags)
 562{
 563        set_work_data(work, (unsigned long)cwq,
 564                      WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
 565}
 566
 567static void set_work_cpu_and_clear_pending(struct work_struct *work,
 568                                           unsigned int cpu)
 569{
 570        /*
 571         * The following wmb is paired with the implied mb in
 572         * test_and_set_bit(PENDING) and ensures all updates to @work made
 573         * here are visible to and precede any updates by the next PENDING
 574         * owner.
 575         */
 576        smp_wmb();
 577        set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
 578}
 579
 580static void clear_work_data(struct work_struct *work)
 581{
 582        smp_wmb();      /* see set_work_cpu_and_clear_pending() */
 583        set_work_data(work, WORK_STRUCT_NO_CPU, 0);
 584}
 585
 586static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
 587{
 588        unsigned long data = atomic_long_read(&work->data);
 589
 590        if (data & WORK_STRUCT_CWQ)
 591                return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 592        else
 593                return NULL;
 594}
 595
 596static struct global_cwq *get_work_gcwq(struct work_struct *work)
 597{
 598        unsigned long data = atomic_long_read(&work->data);
 599        unsigned int cpu;
 600
 601        if (data & WORK_STRUCT_CWQ)
 602                return ((struct cpu_workqueue_struct *)
 603                        (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
 604
 605        cpu = data >> WORK_OFFQ_CPU_SHIFT;
 606        if (cpu == WORK_CPU_NONE)
 607                return NULL;
 608
 609        BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
 610        return get_gcwq(cpu);
 611}
 612
 613static void mark_work_canceling(struct work_struct *work)
 614{
 615        struct global_cwq *gcwq = get_work_gcwq(work);
 616        unsigned long cpu = gcwq ? gcwq->cpu : WORK_CPU_NONE;
 617
 618        set_work_data(work, (cpu << WORK_OFFQ_CPU_SHIFT) | WORK_OFFQ_CANCELING,
 619                      WORK_STRUCT_PENDING);
 620}
 621
 622static bool work_is_canceling(struct work_struct *work)
 623{
 624        unsigned long data = atomic_long_read(&work->data);
 625
 626        return !(data & WORK_STRUCT_CWQ) && (data & WORK_OFFQ_CANCELING);
 627}
 628
 629/*
 630 * Policy functions.  These define the policies on how the global worker
 631 * pools are managed.  Unless noted otherwise, these functions assume that
 632 * they're being called with gcwq->lock held.
 633 */
 634
 635static bool __need_more_worker(struct worker_pool *pool)
 636{
 637        return !atomic_read(get_pool_nr_running(pool));
 638}
 639
 640/*
 641 * Need to wake up a worker?  Called from anything but currently
 642 * running workers.
 643 *
 644 * Note that, because unbound workers never contribute to nr_running, this
 645 * function will always return %true for unbound gcwq as long as the
 646 * worklist isn't empty.
 647 */
 648static bool need_more_worker(struct worker_pool *pool)
 649{
 650        return !list_empty(&pool->worklist) && __need_more_worker(pool);
 651}
 652
 653/* Can I start working?  Called from busy but !running workers. */
 654static bool may_start_working(struct worker_pool *pool)
 655{
 656        return pool->nr_idle;
 657}
 658
 659/* Do I need to keep working?  Called from currently running workers. */
 660static bool keep_working(struct worker_pool *pool)
 661{
 662        atomic_t *nr_running = get_pool_nr_running(pool);
 663
 664        return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
 665}
 666
 667/* Do we need a new worker?  Called from manager. */
 668static bool need_to_create_worker(struct worker_pool *pool)
 669{
 670        return need_more_worker(pool) && !may_start_working(pool);
 671}
 672
 673/* Do I need to be the manager? */
 674static bool need_to_manage_workers(struct worker_pool *pool)
 675{
 676        return need_to_create_worker(pool) ||
 677                (pool->flags & POOL_MANAGE_WORKERS);
 678}
 679
 680/* Do we have too many workers and should some go away? */
 681static bool too_many_workers(struct worker_pool *pool)
 682{
 683        bool managing = pool->flags & POOL_MANAGING_WORKERS;
 684        int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
 685        int nr_busy = pool->nr_workers - nr_idle;
 686
 687        /*
 688         * nr_idle and idle_list may disagree if idle rebinding is in
 689         * progress.  Never return %true if idle_list is empty.
 690         */
 691        if (list_empty(&pool->idle_list))
 692                return false;
 693
 694        return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 695}
 696
 697/*
 698 * Wake up functions.
 699 */
 700
 701/* Return the first worker.  Safe with preemption disabled */
 702static struct worker *first_worker(struct worker_pool *pool)
 703{
 704        if (unlikely(list_empty(&pool->idle_list)))
 705                return NULL;
 706
 707        return list_first_entry(&pool->idle_list, struct worker, entry);
 708}
 709
 710/**
 711 * wake_up_worker - wake up an idle worker
 712 * @pool: worker pool to wake worker from
 713 *
 714 * Wake up the first idle worker of @pool.
 715 *
 716 * CONTEXT:
 717 * spin_lock_irq(gcwq->lock).
 718 */
 719static void wake_up_worker(struct worker_pool *pool)
 720{
 721        struct worker *worker = first_worker(pool);
 722
 723        if (likely(worker))
 724                wake_up_process(worker->task);
 725}
 726
 727/**
 728 * wq_worker_waking_up - a worker is waking up
 729 * @task: task waking up
 730 * @cpu: CPU @task is waking up to
 731 *
 732 * This function is called during try_to_wake_up() when a worker is
 733 * being awoken.
 734 *
 735 * CONTEXT:
 736 * spin_lock_irq(rq->lock)
 737 */
 738void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
 739{
 740        struct worker *worker = kthread_data(task);
 741
 742        if (!(worker->flags & WORKER_NOT_RUNNING))
 743                atomic_inc(get_pool_nr_running(worker->pool));
 744}
 745
 746/**
 747 * wq_worker_sleeping - a worker is going to sleep
 748 * @task: task going to sleep
 749 * @cpu: CPU in question, must be the current CPU number
 750 *
 751 * This function is called during schedule() when a busy worker is
 752 * going to sleep.  Worker on the same cpu can be woken up by
 753 * returning pointer to its task.
 754 *
 755 * CONTEXT:
 756 * spin_lock_irq(rq->lock)
 757 *
 758 * RETURNS:
 759 * Worker task on @cpu to wake up, %NULL if none.
 760 */
 761struct task_struct *wq_worker_sleeping(struct task_struct *task,
 762                                       unsigned int cpu)
 763{
 764        struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 765        struct worker_pool *pool = worker->pool;
 766        atomic_t *nr_running = get_pool_nr_running(pool);
 767
 768        if (worker->flags & WORKER_NOT_RUNNING)
 769                return NULL;
 770
 771        /* this can only happen on the local cpu */
 772        BUG_ON(cpu != raw_smp_processor_id());
 773
 774        /*
 775         * The counterpart of the following dec_and_test, implied mb,
 776         * worklist not empty test sequence is in insert_work().
 777         * Please read comment there.
 778         *
 779         * NOT_RUNNING is clear.  This means that we're bound to and
 780         * running on the local cpu w/ rq lock held and preemption
 781         * disabled, which in turn means that none else could be
 782         * manipulating idle_list, so dereferencing idle_list without gcwq
 783         * lock is safe.
 784         */
 785        if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
 786                to_wakeup = first_worker(pool);
 787        return to_wakeup ? to_wakeup->task : NULL;
 788}
 789
 790/**
 791 * worker_set_flags - set worker flags and adjust nr_running accordingly
 792 * @worker: self
 793 * @flags: flags to set
 794 * @wakeup: wakeup an idle worker if necessary
 795 *
 796 * Set @flags in @worker->flags and adjust nr_running accordingly.  If
 797 * nr_running becomes zero and @wakeup is %true, an idle worker is
 798 * woken up.
 799 *
 800 * CONTEXT:
 801 * spin_lock_irq(gcwq->lock)
 802 */
 803static inline void worker_set_flags(struct worker *worker, unsigned int flags,
 804                                    bool wakeup)
 805{
 806        struct worker_pool *pool = worker->pool;
 807
 808        WARN_ON_ONCE(worker->task != current);
 809
 810        /*
 811         * If transitioning into NOT_RUNNING, adjust nr_running and
 812         * wake up an idle worker as necessary if requested by
 813         * @wakeup.
 814         */
 815        if ((flags & WORKER_NOT_RUNNING) &&
 816            !(worker->flags & WORKER_NOT_RUNNING)) {
 817                atomic_t *nr_running = get_pool_nr_running(pool);
 818
 819                if (wakeup) {
 820                        if (atomic_dec_and_test(nr_running) &&
 821                            !list_empty(&pool->worklist))
 822                                wake_up_worker(pool);
 823                } else
 824                        atomic_dec(nr_running);
 825        }
 826
 827        worker->flags |= flags;
 828}
 829
 830/**
 831 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 832 * @worker: self
 833 * @flags: flags to clear
 834 *
 835 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 836 *
 837 * CONTEXT:
 838 * spin_lock_irq(gcwq->lock)
 839 */
 840static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 841{
 842        struct worker_pool *pool = worker->pool;
 843        unsigned int oflags = worker->flags;
 844
 845        WARN_ON_ONCE(worker->task != current);
 846
 847        worker->flags &= ~flags;
 848
 849        /*
 850         * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 851         * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 852         * of multiple flags, not a single flag.
 853         */
 854        if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 855                if (!(worker->flags & WORKER_NOT_RUNNING))
 856                        atomic_inc(get_pool_nr_running(pool));
 857}
 858
 859/**
 860 * busy_worker_head - return the busy hash head for a work
 861 * @gcwq: gcwq of interest
 862 * @work: work to be hashed
 863 *
 864 * Return hash head of @gcwq for @work.
 865 *
 866 * CONTEXT:
 867 * spin_lock_irq(gcwq->lock).
 868 *
 869 * RETURNS:
 870 * Pointer to the hash head.
 871 */
 872static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
 873                                           struct work_struct *work)
 874{
 875        const int base_shift = ilog2(sizeof(struct work_struct));
 876        unsigned long v = (unsigned long)work;
 877
 878        /* simple shift and fold hash, do we need something better? */
 879        v >>= base_shift;
 880        v += v >> BUSY_WORKER_HASH_ORDER;
 881        v &= BUSY_WORKER_HASH_MASK;
 882
 883        return &gcwq->busy_hash[v];
 884}
 885
 886/**
 887 * __find_worker_executing_work - find worker which is executing a work
 888 * @gcwq: gcwq of interest
 889 * @bwh: hash head as returned by busy_worker_head()
 890 * @work: work to find worker for
 891 *
 892 * Find a worker which is executing @work on @gcwq.  @bwh should be
 893 * the hash head obtained by calling busy_worker_head() with the same
 894 * work.
 895 *
 896 * CONTEXT:
 897 * spin_lock_irq(gcwq->lock).
 898 *
 899 * RETURNS:
 900 * Pointer to worker which is executing @work if found, NULL
 901 * otherwise.
 902 */
 903static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
 904                                                   struct hlist_head *bwh,
 905                                                   struct work_struct *work)
 906{
 907        struct worker *worker;
 908        struct hlist_node *tmp;
 909
 910        hlist_for_each_entry(worker, tmp, bwh, hentry)
 911                if (worker->current_work == work)
 912                        return worker;
 913        return NULL;
 914}
 915
 916/**
 917 * find_worker_executing_work - find worker which is executing a work
 918 * @gcwq: gcwq of interest
 919 * @work: work to find worker for
 920 *
 921 * Find a worker which is executing @work on @gcwq.  This function is
 922 * identical to __find_worker_executing_work() except that this
 923 * function calculates @bwh itself.
 924 *
 925 * CONTEXT:
 926 * spin_lock_irq(gcwq->lock).
 927 *
 928 * RETURNS:
 929 * Pointer to worker which is executing @work if found, NULL
 930 * otherwise.
 931 */
 932static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
 933                                                 struct work_struct *work)
 934{
 935        return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
 936                                            work);
 937}
 938
 939/**
 940 * move_linked_works - move linked works to a list
 941 * @work: start of series of works to be scheduled
 942 * @head: target list to append @work to
 943 * @nextp: out paramter for nested worklist walking
 944 *
 945 * Schedule linked works starting from @work to @head.  Work series to
 946 * be scheduled starts at @work and includes any consecutive work with
 947 * WORK_STRUCT_LINKED set in its predecessor.
 948 *
 949 * If @nextp is not NULL, it's updated to point to the next work of
 950 * the last scheduled work.  This allows move_linked_works() to be
 951 * nested inside outer list_for_each_entry_safe().
 952 *
 953 * CONTEXT:
 954 * spin_lock_irq(gcwq->lock).
 955 */
 956static void move_linked_works(struct work_struct *work, struct list_head *head,
 957                              struct work_struct **nextp)
 958{
 959        struct work_struct *n;
 960
 961        /*
 962         * Linked worklist will always end before the end of the list,
 963         * use NULL for list head.
 964         */
 965        list_for_each_entry_safe_from(work, n, NULL, entry) {
 966                list_move_tail(&work->entry, head);
 967                if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
 968                        break;
 969        }
 970
 971        /*
 972         * If we're already inside safe list traversal and have moved
 973         * multiple works to the scheduled queue, the next position
 974         * needs to be updated.
 975         */
 976        if (nextp)
 977                *nextp = n;
 978}
 979
 980static void cwq_activate_delayed_work(struct work_struct *work)
 981{
 982        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
 983
 984        trace_workqueue_activate_work(work);
 985        move_linked_works(work, &cwq->pool->worklist, NULL);
 986        __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
 987        cwq->nr_active++;
 988}
 989
 990static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
 991{
 992        struct work_struct *work = list_first_entry(&cwq->delayed_works,
 993                                                    struct work_struct, entry);
 994
 995        cwq_activate_delayed_work(work);
 996}
 997
 998/**
 999 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1000 * @cwq: cwq of interest
1001 * @color: color of work which left the queue
1002 *
1003 * A work either has completed or is removed from pending queue,
1004 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1005 *
1006 * CONTEXT:
1007 * spin_lock_irq(gcwq->lock).
1008 */
1009static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1010{
1011        /* ignore uncolored works */
1012        if (color == WORK_NO_COLOR)
1013                return;
1014
1015        cwq->nr_in_flight[color]--;
1016
1017        cwq->nr_active--;
1018        if (!list_empty(&cwq->delayed_works)) {
1019                /* one down, submit a delayed one */
1020                if (cwq->nr_active < cwq->max_active)
1021                        cwq_activate_first_delayed(cwq);
1022        }
1023
1024        /* is flush in progress and are we at the flushing tip? */
1025        if (likely(cwq->flush_color != color))
1026                return;
1027
1028        /* are there still in-flight works? */
1029        if (cwq->nr_in_flight[color])
1030                return;
1031
1032        /* this cwq is done, clear flush_color */
1033        cwq->flush_color = -1;
1034
1035        /*
1036         * If this was the last cwq, wake up the first flusher.  It
1037         * will handle the rest.
1038         */
1039        if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1040                complete(&cwq->wq->first_flusher->done);
1041}
1042
1043/**
1044 * try_to_grab_pending - steal work item from worklist and disable irq
1045 * @work: work item to steal
1046 * @is_dwork: @work is a delayed_work
1047 * @flags: place to store irq state
1048 *
1049 * Try to grab PENDING bit of @work.  This function can handle @work in any
1050 * stable state - idle, on timer or on worklist.  Return values are
1051 *
1052 *  1           if @work was pending and we successfully stole PENDING
1053 *  0           if @work was idle and we claimed PENDING
1054 *  -EAGAIN     if PENDING couldn't be grabbed at the moment, safe to busy-retry
1055 *  -ENOENT     if someone else is canceling @work, this state may persist
1056 *              for arbitrarily long
1057 *
1058 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1059 * interrupted while holding PENDING and @work off queue, irq must be
1060 * disabled on entry.  This, combined with delayed_work->timer being
1061 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1062 *
1063 * On successful return, >= 0, irq is disabled and the caller is
1064 * responsible for releasing it using local_irq_restore(*@flags).
1065 *
1066 * This function is safe to call from any context including IRQ handler.
1067 */
1068static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1069                               unsigned long *flags)
1070{
1071        struct global_cwq *gcwq;
1072
1073        local_irq_save(*flags);
1074
1075        /* try to steal the timer if it exists */
1076        if (is_dwork) {
1077                struct delayed_work *dwork = to_delayed_work(work);
1078
1079                /*
1080                 * dwork->timer is irqsafe.  If del_timer() fails, it's
1081                 * guaranteed that the timer is not queued anywhere and not
1082                 * running on the local CPU.
1083                 */
1084                if (likely(del_timer(&dwork->timer)))
1085                        return 1;
1086        }
1087
1088        /* try to claim PENDING the normal way */
1089        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1090                return 0;
1091
1092        /*
1093         * The queueing is in progress, or it is already queued. Try to
1094         * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1095         */
1096        gcwq = get_work_gcwq(work);
1097        if (!gcwq)
1098                goto fail;
1099
1100        spin_lock(&gcwq->lock);
1101        if (!list_empty(&work->entry)) {
1102                /*
1103                 * This work is queued, but perhaps we locked the wrong gcwq.
1104                 * In that case we must see the new value after rmb(), see
1105                 * insert_work()->wmb().
1106                 */
1107                smp_rmb();
1108                if (gcwq == get_work_gcwq(work)) {
1109                        debug_work_deactivate(work);
1110
1111                        /*
1112                         * A delayed work item cannot be grabbed directly
1113                         * because it might have linked NO_COLOR work items
1114                         * which, if left on the delayed_list, will confuse
1115                         * cwq->nr_active management later on and cause
1116                         * stall.  Make sure the work item is activated
1117                         * before grabbing.
1118                         */
1119                        if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1120                                cwq_activate_delayed_work(work);
1121
1122                        list_del_init(&work->entry);
1123                        cwq_dec_nr_in_flight(get_work_cwq(work),
1124                                get_work_color(work));
1125
1126                        spin_unlock(&gcwq->lock);
1127                        return 1;
1128                }
1129        }
1130        spin_unlock(&gcwq->lock);
1131fail:
1132        local_irq_restore(*flags);
1133        if (work_is_canceling(work))
1134                return -ENOENT;
1135        cpu_relax();
1136        return -EAGAIN;
1137}
1138
1139/**
1140 * insert_work - insert a work into gcwq
1141 * @cwq: cwq @work belongs to
1142 * @work: work to insert
1143 * @head: insertion point
1144 * @extra_flags: extra WORK_STRUCT_* flags to set
1145 *
1146 * Insert @work which belongs to @cwq into @gcwq after @head.
1147 * @extra_flags is or'd to work_struct flags.
1148 *
1149 * CONTEXT:
1150 * spin_lock_irq(gcwq->lock).
1151 */
1152static void insert_work(struct cpu_workqueue_struct *cwq,
1153                        struct work_struct *work, struct list_head *head,
1154                        unsigned int extra_flags)
1155{
1156        struct worker_pool *pool = cwq->pool;
1157
1158        /* we own @work, set data and link */
1159        set_work_cwq(work, cwq, extra_flags);
1160
1161        /*
1162         * Ensure that we get the right work->data if we see the
1163         * result of list_add() below, see try_to_grab_pending().
1164         */
1165        smp_wmb();
1166
1167        list_add_tail(&work->entry, head);
1168
1169        /*
1170         * Ensure either worker_sched_deactivated() sees the above
1171         * list_add_tail() or we see zero nr_running to avoid workers
1172         * lying around lazily while there are works to be processed.
1173         */
1174        smp_mb();
1175
1176        if (__need_more_worker(pool))
1177                wake_up_worker(pool);
1178}
1179
1180/*
1181 * Test whether @work is being queued from another work executing on the
1182 * same workqueue.  This is rather expensive and should only be used from
1183 * cold paths.
1184 */
1185static bool is_chained_work(struct workqueue_struct *wq)
1186{
1187        unsigned long flags;
1188        unsigned int cpu;
1189
1190        for_each_gcwq_cpu(cpu) {
1191                struct global_cwq *gcwq = get_gcwq(cpu);
1192                struct worker *worker;
1193                struct hlist_node *pos;
1194                int i;
1195
1196                spin_lock_irqsave(&gcwq->lock, flags);
1197                for_each_busy_worker(worker, i, pos, gcwq) {
1198                        if (worker->task != current)
1199                                continue;
1200                        spin_unlock_irqrestore(&gcwq->lock, flags);
1201                        /*
1202                         * I'm @worker, no locking necessary.  See if @work
1203                         * is headed to the same workqueue.
1204                         */
1205                        return worker->current_cwq->wq == wq;
1206                }
1207                spin_unlock_irqrestore(&gcwq->lock, flags);
1208        }
1209        return false;
1210}
1211
1212static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1213                         struct work_struct *work)
1214{
1215        struct global_cwq *gcwq;
1216        struct cpu_workqueue_struct *cwq;
1217        struct list_head *worklist;
1218        unsigned int work_flags;
1219        unsigned int req_cpu = cpu;
1220
1221        /*
1222         * While a work item is PENDING && off queue, a task trying to
1223         * steal the PENDING will busy-loop waiting for it to either get
1224         * queued or lose PENDING.  Grabbing PENDING and queueing should
1225         * happen with IRQ disabled.
1226         */
1227        WARN_ON_ONCE(!irqs_disabled());
1228
1229        debug_work_activate(work);
1230
1231        /* if dying, only works from the same workqueue are allowed */
1232        if (unlikely(wq->flags & WQ_DRAINING) &&
1233            WARN_ON_ONCE(!is_chained_work(wq)))
1234                return;
1235
1236        /* determine gcwq to use */
1237        if (!(wq->flags & WQ_UNBOUND)) {
1238                struct global_cwq *last_gcwq;
1239
1240                if (cpu == WORK_CPU_UNBOUND)
1241                        cpu = raw_smp_processor_id();
1242
1243                /*
1244                 * It's multi cpu.  If @work was previously on a different
1245                 * cpu, it might still be running there, in which case the
1246                 * work needs to be queued on that cpu to guarantee
1247                 * non-reentrancy.
1248                 */
1249                gcwq = get_gcwq(cpu);
1250                last_gcwq = get_work_gcwq(work);
1251
1252                if (last_gcwq && last_gcwq != gcwq) {
1253                        struct worker *worker;
1254
1255                        spin_lock(&last_gcwq->lock);
1256
1257                        worker = find_worker_executing_work(last_gcwq, work);
1258
1259                        if (worker && worker->current_cwq->wq == wq)
1260                                gcwq = last_gcwq;
1261                        else {
1262                                /* meh... not running there, queue here */
1263                                spin_unlock(&last_gcwq->lock);
1264                                spin_lock(&gcwq->lock);
1265                        }
1266                } else {
1267                        spin_lock(&gcwq->lock);
1268                }
1269        } else {
1270                gcwq = get_gcwq(WORK_CPU_UNBOUND);
1271                spin_lock(&gcwq->lock);
1272        }
1273
1274        /* gcwq determined, get cwq and queue */
1275        cwq = get_cwq(gcwq->cpu, wq);
1276        trace_workqueue_queue_work(req_cpu, cwq, work);
1277
1278        if (WARN_ON(!list_empty(&work->entry))) {
1279                spin_unlock(&gcwq->lock);
1280                return;
1281        }
1282
1283        cwq->nr_in_flight[cwq->work_color]++;
1284        work_flags = work_color_to_flags(cwq->work_color);
1285
1286        if (likely(cwq->nr_active < cwq->max_active)) {
1287                trace_workqueue_activate_work(work);
1288                cwq->nr_active++;
1289                worklist = &cwq->pool->worklist;
1290        } else {
1291                work_flags |= WORK_STRUCT_DELAYED;
1292                worklist = &cwq->delayed_works;
1293        }
1294
1295        insert_work(cwq, work, worklist, work_flags);
1296
1297        spin_unlock(&gcwq->lock);
1298}
1299
1300/**
1301 * queue_work_on - queue work on specific cpu
1302 * @cpu: CPU number to execute work on
1303 * @wq: workqueue to use
1304 * @work: work to queue
1305 *
1306 * Returns %false if @work was already on a queue, %true otherwise.
1307 *
1308 * We queue the work to a specific CPU, the caller must ensure it
1309 * can't go away.
1310 */
1311bool queue_work_on(int cpu, struct workqueue_struct *wq,
1312                   struct work_struct *work)
1313{
1314        bool ret = false;
1315        unsigned long flags;
1316
1317        local_irq_save(flags);
1318
1319        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1320                __queue_work(cpu, wq, work);
1321                ret = true;
1322        }
1323
1324        local_irq_restore(flags);
1325        return ret;
1326}
1327EXPORT_SYMBOL_GPL(queue_work_on);
1328
1329/**
1330 * queue_work - queue work on a workqueue
1331 * @wq: workqueue to use
1332 * @work: work to queue
1333 *
1334 * Returns %false if @work was already on a queue, %true otherwise.
1335 *
1336 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1337 * it can be processed by another CPU.
1338 */
1339bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1340{
1341        return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1342}
1343EXPORT_SYMBOL_GPL(queue_work);
1344
1345void delayed_work_timer_fn(unsigned long __data)
1346{
1347        struct delayed_work *dwork = (struct delayed_work *)__data;
1348        struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1349
1350        /* should have been called from irqsafe timer with irq already off */
1351        __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1352}
1353EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1354
1355static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1356                                struct delayed_work *dwork, unsigned long delay)
1357{
1358        struct timer_list *timer = &dwork->timer;
1359        struct work_struct *work = &dwork->work;
1360        unsigned int lcpu;
1361
1362        WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1363                     timer->data != (unsigned long)dwork);
1364        WARN_ON_ONCE(timer_pending(timer));
1365        WARN_ON_ONCE(!list_empty(&work->entry));
1366
1367        /*
1368         * If @delay is 0, queue @dwork->work immediately.  This is for
1369         * both optimization and correctness.  The earliest @timer can
1370         * expire is on the closest next tick and delayed_work users depend
1371         * on that there's no such delay when @delay is 0.
1372         */
1373        if (!delay) {
1374                __queue_work(cpu, wq, &dwork->work);
1375                return;
1376        }
1377
1378        timer_stats_timer_set_start_info(&dwork->timer);
1379
1380        /*
1381         * This stores cwq for the moment, for the timer_fn.  Note that the
1382         * work's gcwq is preserved to allow reentrance detection for
1383         * delayed works.
1384         */
1385        if (!(wq->flags & WQ_UNBOUND)) {
1386                struct global_cwq *gcwq = get_work_gcwq(work);
1387
1388                /*
1389                 * If we cannot get the last gcwq from @work directly,
1390                 * select the last CPU such that it avoids unnecessarily
1391                 * triggering non-reentrancy check in __queue_work().
1392                 */
1393                lcpu = cpu;
1394                if (gcwq)
1395                        lcpu = gcwq->cpu;
1396                if (lcpu == WORK_CPU_UNBOUND)
1397                        lcpu = raw_smp_processor_id();
1398        } else {
1399                lcpu = WORK_CPU_UNBOUND;
1400        }
1401
1402        set_work_cwq(work, get_cwq(lcpu, wq), 0);
1403
1404        dwork->cpu = cpu;
1405        timer->expires = jiffies + delay;
1406
1407        if (unlikely(cpu != WORK_CPU_UNBOUND))
1408                add_timer_on(timer, cpu);
1409        else
1410                add_timer(timer);
1411}
1412
1413/**
1414 * queue_delayed_work_on - queue work on specific CPU after delay
1415 * @cpu: CPU number to execute work on
1416 * @wq: workqueue to use
1417 * @dwork: work to queue
1418 * @delay: number of jiffies to wait before queueing
1419 *
1420 * Returns %false if @work was already on a queue, %true otherwise.  If
1421 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1422 * execution.
1423 */
1424bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1425                           struct delayed_work *dwork, unsigned long delay)
1426{
1427        struct work_struct *work = &dwork->work;
1428        bool ret = false;
1429        unsigned long flags;
1430
1431        /* read the comment in __queue_work() */
1432        local_irq_save(flags);
1433
1434        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1435                __queue_delayed_work(cpu, wq, dwork, delay);
1436                ret = true;
1437        }
1438
1439        local_irq_restore(flags);
1440        return ret;
1441}
1442EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1443
1444/**
1445 * queue_delayed_work - queue work on a workqueue after delay
1446 * @wq: workqueue to use
1447 * @dwork: delayable work to queue
1448 * @delay: number of jiffies to wait before queueing
1449 *
1450 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1451 */
1452bool queue_delayed_work(struct workqueue_struct *wq,
1453                        struct delayed_work *dwork, unsigned long delay)
1454{
1455        return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1456}
1457EXPORT_SYMBOL_GPL(queue_delayed_work);
1458
1459/**
1460 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1461 * @cpu: CPU number to execute work on
1462 * @wq: workqueue to use
1463 * @dwork: work to queue
1464 * @delay: number of jiffies to wait before queueing
1465 *
1466 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1467 * modify @dwork's timer so that it expires after @delay.  If @delay is
1468 * zero, @work is guaranteed to be scheduled immediately regardless of its
1469 * current state.
1470 *
1471 * Returns %false if @dwork was idle and queued, %true if @dwork was
1472 * pending and its timer was modified.
1473 *
1474 * This function is safe to call from any context including IRQ handler.
1475 * See try_to_grab_pending() for details.
1476 */
1477bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1478                         struct delayed_work *dwork, unsigned long delay)
1479{
1480        unsigned long flags;
1481        int ret;
1482
1483        do {
1484                ret = try_to_grab_pending(&dwork->work, true, &flags);
1485        } while (unlikely(ret == -EAGAIN));
1486
1487        if (likely(ret >= 0)) {
1488                __queue_delayed_work(cpu, wq, dwork, delay);
1489                local_irq_restore(flags);
1490        }
1491
1492        /* -ENOENT from try_to_grab_pending() becomes %true */
1493        return ret;
1494}
1495EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1496
1497/**
1498 * mod_delayed_work - modify delay of or queue a delayed work
1499 * @wq: workqueue to use
1500 * @dwork: work to queue
1501 * @delay: number of jiffies to wait before queueing
1502 *
1503 * mod_delayed_work_on() on local CPU.
1504 */
1505bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1506                      unsigned long delay)
1507{
1508        return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1509}
1510EXPORT_SYMBOL_GPL(mod_delayed_work);
1511
1512/**
1513 * worker_enter_idle - enter idle state
1514 * @worker: worker which is entering idle state
1515 *
1516 * @worker is entering idle state.  Update stats and idle timer if
1517 * necessary.
1518 *
1519 * LOCKING:
1520 * spin_lock_irq(gcwq->lock).
1521 */
1522static void worker_enter_idle(struct worker *worker)
1523{
1524        struct worker_pool *pool = worker->pool;
1525        struct global_cwq *gcwq = pool->gcwq;
1526
1527        BUG_ON(worker->flags & WORKER_IDLE);
1528        BUG_ON(!list_empty(&worker->entry) &&
1529               (worker->hentry.next || worker->hentry.pprev));
1530
1531        /* can't use worker_set_flags(), also called from start_worker() */
1532        worker->flags |= WORKER_IDLE;
1533        pool->nr_idle++;
1534        worker->last_active = jiffies;
1535
1536        /* idle_list is LIFO */
1537        list_add(&worker->entry, &pool->idle_list);
1538
1539        if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1540                mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1541
1542        /*
1543         * Sanity check nr_running.  Because gcwq_unbind_fn() releases
1544         * gcwq->lock between setting %WORKER_UNBOUND and zapping
1545         * nr_running, the warning may trigger spuriously.  Check iff
1546         * unbind is not in progress.
1547         */
1548        WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1549                     pool->nr_workers == pool->nr_idle &&
1550                     atomic_read(get_pool_nr_running(pool)));
1551}
1552
1553/**
1554 * worker_leave_idle - leave idle state
1555 * @worker: worker which is leaving idle state
1556 *
1557 * @worker is leaving idle state.  Update stats.
1558 *
1559 * LOCKING:
1560 * spin_lock_irq(gcwq->lock).
1561 */
1562static void worker_leave_idle(struct worker *worker)
1563{
1564        struct worker_pool *pool = worker->pool;
1565
1566        BUG_ON(!(worker->flags & WORKER_IDLE));
1567        worker_clr_flags(worker, WORKER_IDLE);
1568        pool->nr_idle--;
1569        list_del_init(&worker->entry);
1570}
1571
1572/**
1573 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1574 * @worker: self
1575 *
1576 * Works which are scheduled while the cpu is online must at least be
1577 * scheduled to a worker which is bound to the cpu so that if they are
1578 * flushed from cpu callbacks while cpu is going down, they are
1579 * guaranteed to execute on the cpu.
1580 *
1581 * This function is to be used by rogue workers and rescuers to bind
1582 * themselves to the target cpu and may race with cpu going down or
1583 * coming online.  kthread_bind() can't be used because it may put the
1584 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1585 * verbatim as it's best effort and blocking and gcwq may be
1586 * [dis]associated in the meantime.
1587 *
1588 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1589 * binding against %GCWQ_DISASSOCIATED which is set during
1590 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1591 * enters idle state or fetches works without dropping lock, it can
1592 * guarantee the scheduling requirement described in the first paragraph.
1593 *
1594 * CONTEXT:
1595 * Might sleep.  Called without any lock but returns with gcwq->lock
1596 * held.
1597 *
1598 * RETURNS:
1599 * %true if the associated gcwq is online (@worker is successfully
1600 * bound), %false if offline.
1601 */
1602static bool worker_maybe_bind_and_lock(struct worker *worker)
1603__acquires(&gcwq->lock)
1604{
1605        struct global_cwq *gcwq = worker->pool->gcwq;
1606        struct task_struct *task = worker->task;
1607
1608        while (true) {
1609                /*
1610                 * The following call may fail, succeed or succeed
1611                 * without actually migrating the task to the cpu if
1612                 * it races with cpu hotunplug operation.  Verify
1613                 * against GCWQ_DISASSOCIATED.
1614                 */
1615                if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1616                        set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1617
1618                spin_lock_irq(&gcwq->lock);
1619                if (gcwq->flags & GCWQ_DISASSOCIATED)
1620                        return false;
1621                if (task_cpu(task) == gcwq->cpu &&
1622                    cpumask_equal(&current->cpus_allowed,
1623                                  get_cpu_mask(gcwq->cpu)))
1624                        return true;
1625                spin_unlock_irq(&gcwq->lock);
1626
1627                /*
1628                 * We've raced with CPU hot[un]plug.  Give it a breather
1629                 * and retry migration.  cond_resched() is required here;
1630                 * otherwise, we might deadlock against cpu_stop trying to
1631                 * bring down the CPU on non-preemptive kernel.
1632                 */
1633                cpu_relax();
1634                cond_resched();
1635        }
1636}
1637
1638/*
1639 * Rebind an idle @worker to its CPU.  worker_thread() will test
1640 * list_empty(@worker->entry) before leaving idle and call this function.
1641 */
1642static void idle_worker_rebind(struct worker *worker)
1643{
1644        struct global_cwq *gcwq = worker->pool->gcwq;
1645
1646        /* CPU may go down again inbetween, clear UNBOUND only on success */
1647        if (worker_maybe_bind_and_lock(worker))
1648                worker_clr_flags(worker, WORKER_UNBOUND);
1649
1650        /* rebind complete, become available again */
1651        list_add(&worker->entry, &worker->pool->idle_list);
1652        spin_unlock_irq(&gcwq->lock);
1653}
1654
1655/*
1656 * Function for @worker->rebind.work used to rebind unbound busy workers to
1657 * the associated cpu which is coming back online.  This is scheduled by
1658 * cpu up but can race with other cpu hotplug operations and may be
1659 * executed twice without intervening cpu down.
1660 */
1661static void busy_worker_rebind_fn(struct work_struct *work)
1662{
1663        struct worker *worker = container_of(work, struct worker, rebind_work);
1664        struct global_cwq *gcwq = worker->pool->gcwq;
1665
1666        if (worker_maybe_bind_and_lock(worker))
1667                worker_clr_flags(worker, WORKER_UNBOUND);
1668
1669        spin_unlock_irq(&gcwq->lock);
1670}
1671
1672/**
1673 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1674 * @gcwq: gcwq of interest
1675 *
1676 * @gcwq->cpu is coming online.  Rebind all workers to the CPU.  Rebinding
1677 * is different for idle and busy ones.
1678 *
1679 * Idle ones will be removed from the idle_list and woken up.  They will
1680 * add themselves back after completing rebind.  This ensures that the
1681 * idle_list doesn't contain any unbound workers when re-bound busy workers
1682 * try to perform local wake-ups for concurrency management.
1683 *
1684 * Busy workers can rebind after they finish their current work items.
1685 * Queueing the rebind work item at the head of the scheduled list is
1686 * enough.  Note that nr_running will be properly bumped as busy workers
1687 * rebind.
1688 *
1689 * On return, all non-manager workers are scheduled for rebind - see
1690 * manage_workers() for the manager special case.  Any idle worker
1691 * including the manager will not appear on @idle_list until rebind is
1692 * complete, making local wake-ups safe.
1693 */
1694static void rebind_workers(struct global_cwq *gcwq)
1695{
1696        struct worker_pool *pool;
1697        struct worker *worker, *n;
1698        struct hlist_node *pos;
1699        int i;
1700
1701        lockdep_assert_held(&gcwq->lock);
1702
1703        for_each_worker_pool(pool, gcwq)
1704                lockdep_assert_held(&pool->assoc_mutex);
1705
1706        /* dequeue and kick idle ones */
1707        for_each_worker_pool(pool, gcwq) {
1708                list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1709                        /*
1710                         * idle workers should be off @pool->idle_list
1711                         * until rebind is complete to avoid receiving
1712                         * premature local wake-ups.
1713                         */
1714                        list_del_init(&worker->entry);
1715
1716                        /*
1717                         * worker_thread() will see the above dequeuing
1718                         * and call idle_worker_rebind().
1719                         */
1720                        wake_up_process(worker->task);
1721                }
1722        }
1723
1724        /* rebind busy workers */
1725        for_each_busy_worker(worker, i, pos, gcwq) {
1726                struct work_struct *rebind_work = &worker->rebind_work;
1727                struct workqueue_struct *wq;
1728
1729                if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1730                                     work_data_bits(rebind_work)))
1731                        continue;
1732
1733                debug_work_activate(rebind_work);
1734
1735                /*
1736                 * wq doesn't really matter but let's keep @worker->pool
1737                 * and @cwq->pool consistent for sanity.
1738                 */
1739                if (worker_pool_pri(worker->pool))
1740                        wq = system_highpri_wq;
1741                else
1742                        wq = system_wq;
1743
1744                insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1745                        worker->scheduled.next,
1746                        work_color_to_flags(WORK_NO_COLOR));
1747        }
1748}
1749
1750static struct worker *alloc_worker(void)
1751{
1752        struct worker *worker;
1753
1754        worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1755        if (worker) {
1756                INIT_LIST_HEAD(&worker->entry);
1757                INIT_LIST_HEAD(&worker->scheduled);
1758                INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1759                /* on creation a worker is in !idle && prep state */
1760                worker->flags = WORKER_PREP;
1761        }
1762        return worker;
1763}
1764
1765/**
1766 * create_worker - create a new workqueue worker
1767 * @pool: pool the new worker will belong to
1768 *
1769 * Create a new worker which is bound to @pool.  The returned worker
1770 * can be started by calling start_worker() or destroyed using
1771 * destroy_worker().
1772 *
1773 * CONTEXT:
1774 * Might sleep.  Does GFP_KERNEL allocations.
1775 *
1776 * RETURNS:
1777 * Pointer to the newly created worker.
1778 */
1779static struct worker *create_worker(struct worker_pool *pool)
1780{
1781        struct global_cwq *gcwq = pool->gcwq;
1782        const char *pri = worker_pool_pri(pool) ? "H" : "";
1783        struct worker *worker = NULL;
1784        int id = -1;
1785
1786        spin_lock_irq(&gcwq->lock);
1787        while (ida_get_new(&pool->worker_ida, &id)) {
1788                spin_unlock_irq(&gcwq->lock);
1789                if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1790                        goto fail;
1791                spin_lock_irq(&gcwq->lock);
1792        }
1793        spin_unlock_irq(&gcwq->lock);
1794
1795        worker = alloc_worker();
1796        if (!worker)
1797                goto fail;
1798
1799        worker->pool = pool;
1800        worker->id = id;
1801
1802        if (gcwq->cpu != WORK_CPU_UNBOUND)
1803                worker->task = kthread_create_on_node(worker_thread,
1804                                        worker, cpu_to_node(gcwq->cpu),
1805                                        "kworker/%u:%d%s", gcwq->cpu, id, pri);
1806        else
1807                worker->task = kthread_create(worker_thread, worker,
1808                                              "kworker/u:%d%s", id, pri);
1809        if (IS_ERR(worker->task))
1810                goto fail;
1811
1812        if (worker_pool_pri(pool))
1813                set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1814
1815        /*
1816         * Determine CPU binding of the new worker depending on
1817         * %GCWQ_DISASSOCIATED.  The caller is responsible for ensuring the
1818         * flag remains stable across this function.  See the comments
1819         * above the flag definition for details.
1820         *
1821         * As an unbound worker may later become a regular one if CPU comes
1822         * online, make sure every worker has %PF_THREAD_BOUND set.
1823         */
1824        if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1825                kthread_bind(worker->task, gcwq->cpu);
1826        } else {
1827                worker->task->flags |= PF_THREAD_BOUND;
1828                worker->flags |= WORKER_UNBOUND;
1829        }
1830
1831        return worker;
1832fail:
1833        if (id >= 0) {
1834                spin_lock_irq(&gcwq->lock);
1835                ida_remove(&pool->worker_ida, id);
1836                spin_unlock_irq(&gcwq->lock);
1837        }
1838        kfree(worker);
1839        return NULL;
1840}
1841
1842/**
1843 * start_worker - start a newly created worker
1844 * @worker: worker to start
1845 *
1846 * Make the gcwq aware of @worker and start it.
1847 *
1848 * CONTEXT:
1849 * spin_lock_irq(gcwq->lock).
1850 */
1851static void start_worker(struct worker *worker)
1852{
1853        worker->flags |= WORKER_STARTED;
1854        worker->pool->nr_workers++;
1855        worker_enter_idle(worker);
1856        wake_up_process(worker->task);
1857}
1858
1859/**
1860 * destroy_worker - destroy a workqueue worker
1861 * @worker: worker to be destroyed
1862 *
1863 * Destroy @worker and adjust @gcwq stats accordingly.
1864 *
1865 * CONTEXT:
1866 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1867 */
1868static void destroy_worker(struct worker *worker)
1869{
1870        struct worker_pool *pool = worker->pool;
1871        struct global_cwq *gcwq = pool->gcwq;
1872        int id = worker->id;
1873
1874        /* sanity check frenzy */
1875        BUG_ON(worker->current_work);
1876        BUG_ON(!list_empty(&worker->scheduled));
1877
1878        if (worker->flags & WORKER_STARTED)
1879                pool->nr_workers--;
1880        if (worker->flags & WORKER_IDLE)
1881                pool->nr_idle--;
1882
1883        list_del_init(&worker->entry);
1884        worker->flags |= WORKER_DIE;
1885
1886        spin_unlock_irq(&gcwq->lock);
1887
1888        kthread_stop(worker->task);
1889        kfree(worker);
1890
1891        spin_lock_irq(&gcwq->lock);
1892        ida_remove(&pool->worker_ida, id);
1893}
1894
1895static void idle_worker_timeout(unsigned long __pool)
1896{
1897        struct worker_pool *pool = (void *)__pool;
1898        struct global_cwq *gcwq = pool->gcwq;
1899
1900        spin_lock_irq(&gcwq->lock);
1901
1902        if (too_many_workers(pool)) {
1903                struct worker *worker;
1904                unsigned long expires;
1905
1906                /* idle_list is kept in LIFO order, check the last one */
1907                worker = list_entry(pool->idle_list.prev, struct worker, entry);
1908                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1909
1910                if (time_before(jiffies, expires))
1911                        mod_timer(&pool->idle_timer, expires);
1912                else {
1913                        /* it's been idle for too long, wake up manager */
1914                        pool->flags |= POOL_MANAGE_WORKERS;
1915                        wake_up_worker(pool);
1916                }
1917        }
1918
1919        spin_unlock_irq(&gcwq->lock);
1920}
1921
1922static bool send_mayday(struct work_struct *work)
1923{
1924        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1925        struct workqueue_struct *wq = cwq->wq;
1926        unsigned int cpu;
1927
1928        if (!(wq->flags & WQ_RESCUER))
1929                return false;
1930
1931        /* mayday mayday mayday */
1932        cpu = cwq->pool->gcwq->cpu;
1933        /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1934        if (cpu == WORK_CPU_UNBOUND)
1935                cpu = 0;
1936        if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1937                wake_up_process(wq->rescuer->task);
1938        return true;
1939}
1940
1941static void gcwq_mayday_timeout(unsigned long __pool)
1942{
1943        struct worker_pool *pool = (void *)__pool;
1944        struct global_cwq *gcwq = pool->gcwq;
1945        struct work_struct *work;
1946
1947        spin_lock_irq(&gcwq->lock);
1948
1949        if (need_to_create_worker(pool)) {
1950                /*
1951                 * We've been trying to create a new worker but
1952                 * haven't been successful.  We might be hitting an
1953                 * allocation deadlock.  Send distress signals to
1954                 * rescuers.
1955                 */
1956                list_for_each_entry(work, &pool->worklist, entry)
1957                        send_mayday(work);
1958        }
1959
1960        spin_unlock_irq(&gcwq->lock);
1961
1962        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1963}
1964
1965/**
1966 * maybe_create_worker - create a new worker if necessary
1967 * @pool: pool to create a new worker for
1968 *
1969 * Create a new worker for @pool if necessary.  @pool is guaranteed to
1970 * have at least one idle worker on return from this function.  If
1971 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1972 * sent to all rescuers with works scheduled on @pool to resolve
1973 * possible allocation deadlock.
1974 *
1975 * On return, need_to_create_worker() is guaranteed to be false and
1976 * may_start_working() true.
1977 *
1978 * LOCKING:
1979 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1980 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1981 * manager.
1982 *
1983 * RETURNS:
1984 * false if no action was taken and gcwq->lock stayed locked, true
1985 * otherwise.
1986 */
1987static bool maybe_create_worker(struct worker_pool *pool)
1988__releases(&gcwq->lock)
1989__acquires(&gcwq->lock)
1990{
1991        struct global_cwq *gcwq = pool->gcwq;
1992
1993        if (!need_to_create_worker(pool))
1994                return false;
1995restart:
1996        spin_unlock_irq(&gcwq->lock);
1997
1998        /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1999        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2000
2001        while (true) {
2002                struct worker *worker;
2003
2004                worker = create_worker(pool);
2005                if (worker) {
2006                        del_timer_sync(&pool->mayday_timer);
2007                        spin_lock_irq(&gcwq->lock);
2008                        start_worker(worker);
2009                        BUG_ON(need_to_create_worker(pool));
2010                        return true;
2011                }
2012
2013                if (!need_to_create_worker(pool))
2014                        break;
2015
2016                __set_current_state(TASK_INTERRUPTIBLE);
2017                schedule_timeout(CREATE_COOLDOWN);
2018
2019                if (!need_to_create_worker(pool))
2020                        break;
2021        }
2022
2023        del_timer_sync(&pool->mayday_timer);
2024        spin_lock_irq(&gcwq->lock);
2025        if (need_to_create_worker(pool))
2026                goto restart;
2027        return true;
2028}
2029
2030/**
2031 * maybe_destroy_worker - destroy workers which have been idle for a while
2032 * @pool: pool to destroy workers for
2033 *
2034 * Destroy @pool workers which have been idle for longer than
2035 * IDLE_WORKER_TIMEOUT.
2036 *
2037 * LOCKING:
2038 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2039 * multiple times.  Called only from manager.
2040 *
2041 * RETURNS:
2042 * false if no action was taken and gcwq->lock stayed locked, true
2043 * otherwise.
2044 */
2045static bool maybe_destroy_workers(struct worker_pool *pool)
2046{
2047        bool ret = false;
2048
2049        while (too_many_workers(pool)) {
2050                struct worker *worker;
2051                unsigned long expires;
2052
2053                worker = list_entry(pool->idle_list.prev, struct worker, entry);
2054                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2055
2056                if (time_before(jiffies, expires)) {
2057                        mod_timer(&pool->idle_timer, expires);
2058                        break;
2059                }
2060
2061                destroy_worker(worker);
2062                ret = true;
2063        }
2064
2065        return ret;
2066}
2067
2068/**
2069 * manage_workers - manage worker pool
2070 * @worker: self
2071 *
2072 * Assume the manager role and manage gcwq worker pool @worker belongs
2073 * to.  At any given time, there can be only zero or one manager per
2074 * gcwq.  The exclusion is handled automatically by this function.
2075 *
2076 * The caller can safely start processing works on false return.  On
2077 * true return, it's guaranteed that need_to_create_worker() is false
2078 * and may_start_working() is true.
2079 *
2080 * CONTEXT:
2081 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2082 * multiple times.  Does GFP_KERNEL allocations.
2083 *
2084 * RETURNS:
2085 * false if no action was taken and gcwq->lock stayed locked, true if
2086 * some action was taken.
2087 */
2088static bool manage_workers(struct worker *worker)
2089{
2090        struct worker_pool *pool = worker->pool;
2091        bool ret = false;
2092
2093        if (pool->flags & POOL_MANAGING_WORKERS)
2094                return ret;
2095
2096        pool->flags |= POOL_MANAGING_WORKERS;
2097
2098        /*
2099         * To simplify both worker management and CPU hotplug, hold off
2100         * management while hotplug is in progress.  CPU hotplug path can't
2101         * grab %POOL_MANAGING_WORKERS to achieve this because that can
2102         * lead to idle worker depletion (all become busy thinking someone
2103         * else is managing) which in turn can result in deadlock under
2104         * extreme circumstances.  Use @pool->assoc_mutex to synchronize
2105         * manager against CPU hotplug.
2106         *
2107         * assoc_mutex would always be free unless CPU hotplug is in
2108         * progress.  trylock first without dropping @gcwq->lock.
2109         */
2110        if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2111                spin_unlock_irq(&pool->gcwq->lock);
2112                mutex_lock(&pool->assoc_mutex);
2113                /*
2114                 * CPU hotplug could have happened while we were waiting
2115                 * for assoc_mutex.  Hotplug itself can't handle us
2116                 * because manager isn't either on idle or busy list, and
2117                 * @gcwq's state and ours could have deviated.
2118                 *
2119                 * As hotplug is now excluded via assoc_mutex, we can
2120                 * simply try to bind.  It will succeed or fail depending
2121                 * on @gcwq's current state.  Try it and adjust
2122                 * %WORKER_UNBOUND accordingly.
2123                 */
2124                if (worker_maybe_bind_and_lock(worker))
2125                        worker->flags &= ~WORKER_UNBOUND;
2126                else
2127                        worker->flags |= WORKER_UNBOUND;
2128
2129                ret = true;
2130        }
2131
2132        pool->flags &= ~POOL_MANAGE_WORKERS;
2133
2134        /*
2135         * Destroy and then create so that may_start_working() is true
2136         * on return.
2137         */
2138        ret |= maybe_destroy_workers(pool);
2139        ret |= maybe_create_worker(pool);
2140
2141        pool->flags &= ~POOL_MANAGING_WORKERS;
2142        mutex_unlock(&pool->assoc_mutex);
2143        return ret;
2144}
2145
2146/**
2147 * process_one_work - process single work
2148 * @worker: self
2149 * @work: work to process
2150 *
2151 * Process @work.  This function contains all the logics necessary to
2152 * process a single work including synchronization against and
2153 * interaction with other workers on the same cpu, queueing and
2154 * flushing.  As long as context requirement is met, any worker can
2155 * call this function to process a work.
2156 *
2157 * CONTEXT:
2158 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2159 */
2160static void process_one_work(struct worker *worker, struct work_struct *work)
2161__releases(&gcwq->lock)
2162__acquires(&gcwq->lock)
2163{
2164        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2165        struct worker_pool *pool = worker->pool;
2166        struct global_cwq *gcwq = pool->gcwq;
2167        struct hlist_head *bwh = busy_worker_head(gcwq, work);
2168        bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2169        work_func_t f = work->func;
2170        int work_color;
2171        struct worker *collision;
2172#ifdef CONFIG_LOCKDEP
2173        /*
2174         * It is permissible to free the struct work_struct from
2175         * inside the function that is called from it, this we need to
2176         * take into account for lockdep too.  To avoid bogus "held
2177         * lock freed" warnings as well as problems when looking into
2178         * work->lockdep_map, make a copy and use that here.
2179         */
2180        struct lockdep_map lockdep_map;
2181
2182        lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2183#endif
2184        /*
2185         * Ensure we're on the correct CPU.  DISASSOCIATED test is
2186         * necessary to avoid spurious warnings from rescuers servicing the
2187         * unbound or a disassociated gcwq.
2188         */
2189        WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2190                     !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2191                     raw_smp_processor_id() != gcwq->cpu);
2192
2193        /*
2194         * A single work shouldn't be executed concurrently by
2195         * multiple workers on a single cpu.  Check whether anyone is
2196         * already processing the work.  If so, defer the work to the
2197         * currently executing one.
2198         */
2199        collision = __find_worker_executing_work(gcwq, bwh, work);
2200        if (unlikely(collision)) {
2201                move_linked_works(work, &collision->scheduled, NULL);
2202                return;
2203        }
2204
2205        /* claim and dequeue */
2206        debug_work_deactivate(work);
2207        hlist_add_head(&worker->hentry, bwh);
2208        worker->current_work = work;
2209        worker->current_cwq = cwq;
2210        work_color = get_work_color(work);
2211
2212        list_del_init(&work->entry);
2213
2214        /*
2215         * CPU intensive works don't participate in concurrency
2216         * management.  They're the scheduler's responsibility.
2217         */
2218        if (unlikely(cpu_intensive))
2219                worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2220
2221        /*
2222         * Unbound gcwq isn't concurrency managed and work items should be
2223         * executed ASAP.  Wake up another worker if necessary.
2224         */
2225        if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2226                wake_up_worker(pool);
2227
2228        /*
2229         * Record the last CPU and clear PENDING which should be the last
2230         * update to @work.  Also, do this inside @gcwq->lock so that
2231         * PENDING and queued state changes happen together while IRQ is
2232         * disabled.
2233         */
2234        set_work_cpu_and_clear_pending(work, gcwq->cpu);
2235
2236        spin_unlock_irq(&gcwq->lock);
2237
2238        lock_map_acquire_read(&cwq->wq->lockdep_map);
2239        lock_map_acquire(&lockdep_map);
2240        trace_workqueue_execute_start(work);
2241        f(work);
2242        /*
2243         * While we must be careful to not use "work" after this, the trace
2244         * point will only record its address.
2245         */
2246        trace_workqueue_execute_end(work);
2247        lock_map_release(&lockdep_map);
2248        lock_map_release(&cwq->wq->lockdep_map);
2249
2250        if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2251                pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2252                       "     last function: %pf\n",
2253                       current->comm, preempt_count(), task_pid_nr(current), f);
2254                debug_show_held_locks(current);
2255                dump_stack();
2256        }
2257
2258        spin_lock_irq(&gcwq->lock);
2259
2260        /* clear cpu intensive status */
2261        if (unlikely(cpu_intensive))
2262                worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2263
2264        /* we're done with it, release */
2265        hlist_del_init(&worker->hentry);
2266        worker->current_work = NULL;
2267        worker->current_cwq = NULL;
2268        cwq_dec_nr_in_flight(cwq, work_color);
2269}
2270
2271/**
2272 * process_scheduled_works - process scheduled works
2273 * @worker: self
2274 *
2275 * Process all scheduled works.  Please note that the scheduled list
2276 * may change while processing a work, so this function repeatedly
2277 * fetches a work from the top and executes it.
2278 *
2279 * CONTEXT:
2280 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2281 * multiple times.
2282 */
2283static void process_scheduled_works(struct worker *worker)
2284{
2285        while (!list_empty(&worker->scheduled)) {
2286                struct work_struct *work = list_first_entry(&worker->scheduled,
2287                                                struct work_struct, entry);
2288                process_one_work(worker, work);
2289        }
2290}
2291
2292/**
2293 * worker_thread - the worker thread function
2294 * @__worker: self
2295 *
2296 * The gcwq worker thread function.  There's a single dynamic pool of
2297 * these per each cpu.  These workers process all works regardless of
2298 * their specific target workqueue.  The only exception is works which
2299 * belong to workqueues with a rescuer which will be explained in
2300 * rescuer_thread().
2301 */
2302static int worker_thread(void *__worker)
2303{
2304        struct worker *worker = __worker;
2305        struct worker_pool *pool = worker->pool;
2306        struct global_cwq *gcwq = pool->gcwq;
2307
2308        /* tell the scheduler that this is a workqueue worker */
2309        worker->task->flags |= PF_WQ_WORKER;
2310woke_up:
2311        spin_lock_irq(&gcwq->lock);
2312
2313        /* we are off idle list if destruction or rebind is requested */
2314        if (unlikely(list_empty(&worker->entry))) {
2315                spin_unlock_irq(&gcwq->lock);
2316
2317                /* if DIE is set, destruction is requested */
2318                if (worker->flags & WORKER_DIE) {
2319                        worker->task->flags &= ~PF_WQ_WORKER;
2320                        return 0;
2321                }
2322
2323                /* otherwise, rebind */
2324                idle_worker_rebind(worker);
2325                goto woke_up;
2326        }
2327
2328        worker_leave_idle(worker);
2329recheck:
2330        /* no more worker necessary? */
2331        if (!need_more_worker(pool))
2332                goto sleep;
2333
2334        /* do we need to manage? */
2335        if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2336                goto recheck;
2337
2338        /*
2339         * ->scheduled list can only be filled while a worker is
2340         * preparing to process a work or actually processing it.
2341         * Make sure nobody diddled with it while I was sleeping.
2342         */
2343        BUG_ON(!list_empty(&worker->scheduled));
2344
2345        /*
2346         * When control reaches this point, we're guaranteed to have
2347         * at least one idle worker or that someone else has already
2348         * assumed the manager role.
2349         */
2350        worker_clr_flags(worker, WORKER_PREP);
2351
2352        do {
2353                struct work_struct *work =
2354                        list_first_entry(&pool->worklist,
2355                                         struct work_struct, entry);
2356
2357                if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2358                        /* optimization path, not strictly necessary */
2359                        process_one_work(worker, work);
2360                        if (unlikely(!list_empty(&worker->scheduled)))
2361                                process_scheduled_works(worker);
2362                } else {
2363                        move_linked_works(work, &worker->scheduled, NULL);
2364                        process_scheduled_works(worker);
2365                }
2366        } while (keep_working(pool));
2367
2368        worker_set_flags(worker, WORKER_PREP, false);
2369sleep:
2370        if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2371                goto recheck;
2372
2373        /*
2374         * gcwq->lock is held and there's no work to process and no
2375         * need to manage, sleep.  Workers are woken up only while
2376         * holding gcwq->lock or from local cpu, so setting the
2377         * current state before releasing gcwq->lock is enough to
2378         * prevent losing any event.
2379         */
2380        worker_enter_idle(worker);
2381        __set_current_state(TASK_INTERRUPTIBLE);
2382        spin_unlock_irq(&gcwq->lock);
2383        schedule();
2384        goto woke_up;
2385}
2386
2387/**
2388 * rescuer_thread - the rescuer thread function
2389 * @__wq: the associated workqueue
2390 *
2391 * Workqueue rescuer thread function.  There's one rescuer for each
2392 * workqueue which has WQ_RESCUER set.
2393 *
2394 * Regular work processing on a gcwq may block trying to create a new
2395 * worker which uses GFP_KERNEL allocation which has slight chance of
2396 * developing into deadlock if some works currently on the same queue
2397 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2398 * the problem rescuer solves.
2399 *
2400 * When such condition is possible, the gcwq summons rescuers of all
2401 * workqueues which have works queued on the gcwq and let them process
2402 * those works so that forward progress can be guaranteed.
2403 *
2404 * This should happen rarely.
2405 */
2406static int rescuer_thread(void *__wq)
2407{
2408        struct workqueue_struct *wq = __wq;
2409        struct worker *rescuer = wq->rescuer;
2410        struct list_head *scheduled = &rescuer->scheduled;
2411        bool is_unbound = wq->flags & WQ_UNBOUND;
2412        unsigned int cpu;
2413
2414        set_user_nice(current, RESCUER_NICE_LEVEL);
2415repeat:
2416        set_current_state(TASK_INTERRUPTIBLE);
2417
2418        if (kthread_should_stop()) {
2419                __set_current_state(TASK_RUNNING);
2420                return 0;
2421        }
2422
2423        /*
2424         * See whether any cpu is asking for help.  Unbounded
2425         * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2426         */
2427        for_each_mayday_cpu(cpu, wq->mayday_mask) {
2428                unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2429                struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2430                struct worker_pool *pool = cwq->pool;
2431                struct global_cwq *gcwq = pool->gcwq;
2432                struct work_struct *work, *n;
2433
2434                __set_current_state(TASK_RUNNING);
2435                mayday_clear_cpu(cpu, wq->mayday_mask);
2436
2437                /* migrate to the target cpu if possible */
2438                rescuer->pool = pool;
2439                worker_maybe_bind_and_lock(rescuer);
2440
2441                /*
2442                 * Slurp in all works issued via this workqueue and
2443                 * process'em.
2444                 */
2445                BUG_ON(!list_empty(&rescuer->scheduled));
2446                list_for_each_entry_safe(work, n, &pool->worklist, entry)
2447                        if (get_work_cwq(work) == cwq)
2448                                move_linked_works(work, scheduled, &n);
2449
2450                process_scheduled_works(rescuer);
2451
2452                /*
2453                 * Leave this gcwq.  If keep_working() is %true, notify a
2454                 * regular worker; otherwise, we end up with 0 concurrency
2455                 * and stalling the execution.
2456                 */
2457                if (keep_working(pool))
2458                        wake_up_worker(pool);
2459
2460                spin_unlock_irq(&gcwq->lock);
2461        }
2462
2463        schedule();
2464        goto repeat;
2465}
2466
2467struct wq_barrier {
2468        struct work_struct      work;
2469        struct completion       done;
2470};
2471
2472static void wq_barrier_func(struct work_struct *work)
2473{
2474        struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2475        complete(&barr->done);
2476}
2477
2478/**
2479 * insert_wq_barrier - insert a barrier work
2480 * @cwq: cwq to insert barrier into
2481 * @barr: wq_barrier to insert
2482 * @target: target work to attach @barr to
2483 * @worker: worker currently executing @target, NULL if @target is not executing
2484 *
2485 * @barr is linked to @target such that @barr is completed only after
2486 * @target finishes execution.  Please note that the ordering
2487 * guarantee is observed only with respect to @target and on the local
2488 * cpu.
2489 *
2490 * Currently, a queued barrier can't be canceled.  This is because
2491 * try_to_grab_pending() can't determine whether the work to be
2492 * grabbed is at the head of the queue and thus can't clear LINKED
2493 * flag of the previous work while there must be a valid next work
2494 * after a work with LINKED flag set.
2495 *
2496 * Note that when @worker is non-NULL, @target may be modified
2497 * underneath us, so we can't reliably determine cwq from @target.
2498 *
2499 * CONTEXT:
2500 * spin_lock_irq(gcwq->lock).
2501 */
2502static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2503                              struct wq_barrier *barr,
2504                              struct work_struct *target, struct worker *worker)
2505{
2506        struct list_head *head;
2507        unsigned int linked = 0;
2508
2509        /*
2510         * debugobject calls are safe here even with gcwq->lock locked
2511         * as we know for sure that this will not trigger any of the
2512         * checks and call back into the fixup functions where we
2513         * might deadlock.
2514         */
2515        INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2516        __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2517        init_completion(&barr->done);
2518
2519        /*
2520         * If @target is currently being executed, schedule the
2521         * barrier to the worker; otherwise, put it after @target.
2522         */
2523        if (worker)
2524                head = worker->scheduled.next;
2525        else {
2526                unsigned long *bits = work_data_bits(target);
2527
2528                head = target->entry.next;
2529                /* there can already be other linked works, inherit and set */
2530                linked = *bits & WORK_STRUCT_LINKED;
2531                __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2532        }
2533
2534        debug_work_activate(&barr->work);
2535        insert_work(cwq, &barr->work, head,
2536                    work_color_to_flags(WORK_NO_COLOR) | linked);
2537}
2538
2539/**
2540 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2541 * @wq: workqueue being flushed
2542 * @flush_color: new flush color, < 0 for no-op
2543 * @work_color: new work color, < 0 for no-op
2544 *
2545 * Prepare cwqs for workqueue flushing.
2546 *
2547 * If @flush_color is non-negative, flush_color on all cwqs should be
2548 * -1.  If no cwq has in-flight commands at the specified color, all
2549 * cwq->flush_color's stay at -1 and %false is returned.  If any cwq
2550 * has in flight commands, its cwq->flush_color is set to
2551 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2552 * wakeup logic is armed and %true is returned.
2553 *
2554 * The caller should have initialized @wq->first_flusher prior to
2555 * calling this function with non-negative @flush_color.  If
2556 * @flush_color is negative, no flush color update is done and %false
2557 * is returned.
2558 *
2559 * If @work_color is non-negative, all cwqs should have the same
2560 * work_color which is previous to @work_color and all will be
2561 * advanced to @work_color.
2562 *
2563 * CONTEXT:
2564 * mutex_lock(wq->flush_mutex).
2565 *
2566 * RETURNS:
2567 * %true if @flush_color >= 0 and there's something to flush.  %false
2568 * otherwise.
2569 */
2570static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2571                                      int flush_color, int work_color)
2572{
2573        bool wait = false;
2574        unsigned int cpu;
2575
2576        if (flush_color >= 0) {
2577                BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2578                atomic_set(&wq->nr_cwqs_to_flush, 1);
2579        }
2580
2581        for_each_cwq_cpu(cpu, wq) {
2582                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2583                struct global_cwq *gcwq = cwq->pool->gcwq;
2584
2585                spin_lock_irq(&gcwq->lock);
2586
2587                if (flush_color >= 0) {
2588                        BUG_ON(cwq->flush_color != -1);
2589
2590                        if (cwq->nr_in_flight[flush_color]) {
2591                                cwq->flush_color = flush_color;
2592                                atomic_inc(&wq->nr_cwqs_to_flush);
2593                                wait = true;
2594                        }
2595                }
2596
2597                if (work_color >= 0) {
2598                        BUG_ON(work_color != work_next_color(cwq->work_color));
2599                        cwq->work_color = work_color;
2600                }
2601
2602                spin_unlock_irq(&gcwq->lock);
2603        }
2604
2605        if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2606                complete(&wq->first_flusher->done);
2607
2608        return wait;
2609}
2610
2611/**
2612 * flush_workqueue - ensure that any scheduled work has run to completion.
2613 * @wq: workqueue to flush
2614 *
2615 * Forces execution of the workqueue and blocks until its completion.
2616 * This is typically used in driver shutdown handlers.
2617 *
2618 * We sleep until all works which were queued on entry have been handled,
2619 * but we are not livelocked by new incoming ones.
2620 */
2621void flush_workqueue(struct workqueue_struct *wq)
2622{
2623        struct wq_flusher this_flusher = {
2624                .list = LIST_HEAD_INIT(this_flusher.list),
2625                .flush_color = -1,
2626                .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2627        };
2628        int next_color;
2629
2630        lock_map_acquire(&wq->lockdep_map);
2631        lock_map_release(&wq->lockdep_map);
2632
2633        mutex_lock(&wq->flush_mutex);
2634
2635        /*
2636         * Start-to-wait phase
2637         */
2638        next_color = work_next_color(wq->work_color);
2639
2640        if (next_color != wq->flush_color) {
2641                /*
2642                 * Color space is not full.  The current work_color
2643                 * becomes our flush_color and work_color is advanced
2644                 * by one.
2645                 */
2646                BUG_ON(!list_empty(&wq->flusher_overflow));
2647                this_flusher.flush_color = wq->work_color;
2648                wq->work_color = next_color;
2649
2650                if (!wq->first_flusher) {
2651                        /* no flush in progress, become the first flusher */
2652                        BUG_ON(wq->flush_color != this_flusher.flush_color);
2653
2654                        wq->first_flusher = &this_flusher;
2655
2656                        if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2657                                                       wq->work_color)) {
2658                                /* nothing to flush, done */
2659                                wq->flush_color = next_color;
2660                                wq->first_flusher = NULL;
2661                                goto out_unlock;
2662                        }
2663                } else {
2664                        /* wait in queue */
2665                        BUG_ON(wq->flush_color == this_flusher.flush_color);
2666                        list_add_tail(&this_flusher.list, &wq->flusher_queue);
2667                        flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2668                }
2669        } else {
2670                /*
2671                 * Oops, color space is full, wait on overflow queue.
2672                 * The next flush completion will assign us
2673                 * flush_color and transfer to flusher_queue.
2674                 */
2675                list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2676        }
2677
2678        mutex_unlock(&wq->flush_mutex);
2679
2680        wait_for_completion(&this_flusher.done);
2681
2682        /*
2683         * Wake-up-and-cascade phase
2684         *
2685         * First flushers are responsible for cascading flushes and
2686         * handling overflow.  Non-first flushers can simply return.
2687         */
2688        if (wq->first_flusher != &this_flusher)
2689                return;
2690
2691        mutex_lock(&wq->flush_mutex);
2692
2693        /* we might have raced, check again with mutex held */
2694        if (wq->first_flusher != &this_flusher)
2695                goto out_unlock;
2696
2697        wq->first_flusher = NULL;
2698
2699        BUG_ON(!list_empty(&this_flusher.list));
2700        BUG_ON(wq->flush_color != this_flusher.flush_color);
2701
2702        while (true) {
2703                struct wq_flusher *next, *tmp;
2704
2705                /* complete all the flushers sharing the current flush color */
2706                list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2707                        if (next->flush_color != wq->flush_color)
2708                                break;
2709                        list_del_init(&next->list);
2710                        complete(&next->done);
2711                }
2712
2713                BUG_ON(!list_empty(&wq->flusher_overflow) &&
2714                       wq->flush_color != work_next_color(wq->work_color));
2715
2716                /* this flush_color is finished, advance by one */
2717                wq->flush_color = work_next_color(wq->flush_color);
2718
2719                /* one color has been freed, handle overflow queue */
2720                if (!list_empty(&wq->flusher_overflow)) {
2721                        /*
2722                         * Assign the same color to all overflowed
2723                         * flushers, advance work_color and append to
2724                         * flusher_queue.  This is the start-to-wait
2725                         * phase for these overflowed flushers.
2726                         */
2727                        list_for_each_entry(tmp, &wq->flusher_overflow, list)
2728                                tmp->flush_color = wq->work_color;
2729
2730                        wq->work_color = work_next_color(wq->work_color);
2731
2732                        list_splice_tail_init(&wq->flusher_overflow,
2733                                              &wq->flusher_queue);
2734                        flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2735                }
2736
2737                if (list_empty(&wq->flusher_queue)) {
2738                        BUG_ON(wq->flush_color != wq->work_color);
2739                        break;
2740                }
2741
2742                /*
2743                 * Need to flush more colors.  Make the next flusher
2744                 * the new first flusher and arm cwqs.
2745                 */
2746                BUG_ON(wq->flush_color == wq->work_color);
2747                BUG_ON(wq->flush_color != next->flush_color);
2748
2749                list_del_init(&next->list);
2750                wq->first_flusher = next;
2751
2752                if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2753                        break;
2754
2755                /*
2756                 * Meh... this color is already done, clear first
2757                 * flusher and repeat cascading.
2758                 */
2759                wq->first_flusher = NULL;
2760        }
2761
2762out_unlock:
2763        mutex_unlock(&wq->flush_mutex);
2764}
2765EXPORT_SYMBOL_GPL(flush_workqueue);
2766
2767/**
2768 * drain_workqueue - drain a workqueue
2769 * @wq: workqueue to drain
2770 *
2771 * Wait until the workqueue becomes empty.  While draining is in progress,
2772 * only chain queueing is allowed.  IOW, only currently pending or running
2773 * work items on @wq can queue further work items on it.  @wq is flushed
2774 * repeatedly until it becomes empty.  The number of flushing is detemined
2775 * by the depth of chaining and should be relatively short.  Whine if it
2776 * takes too long.
2777 */
2778void drain_workqueue(struct workqueue_struct *wq)
2779{
2780        unsigned int flush_cnt = 0;
2781        unsigned int cpu;
2782
2783        /*
2784         * __queue_work() needs to test whether there are drainers, is much
2785         * hotter than drain_workqueue() and already looks at @wq->flags.
2786         * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2787         */
2788        spin_lock(&workqueue_lock);
2789        if (!wq->nr_drainers++)
2790                wq->flags |= WQ_DRAINING;
2791        spin_unlock(&workqueue_lock);
2792reflush:
2793        flush_workqueue(wq);
2794
2795        for_each_cwq_cpu(cpu, wq) {
2796                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2797                bool drained;
2798
2799                spin_lock_irq(&cwq->pool->gcwq->lock);
2800                drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2801                spin_unlock_irq(&cwq->pool->gcwq->lock);
2802
2803                if (drained)
2804                        continue;
2805
2806                if (++flush_cnt == 10 ||
2807                    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2808                        pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2809                                wq->name, flush_cnt);
2810                goto reflush;
2811        }
2812
2813        spin_lock(&workqueue_lock);
2814        if (!--wq->nr_drainers)
2815                wq->flags &= ~WQ_DRAINING;
2816        spin_unlock(&workqueue_lock);
2817}
2818EXPORT_SYMBOL_GPL(drain_workqueue);
2819
2820static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2821{
2822        struct worker *worker = NULL;
2823        struct global_cwq *gcwq;
2824        struct cpu_workqueue_struct *cwq;
2825
2826        might_sleep();
2827        gcwq = get_work_gcwq(work);
2828        if (!gcwq)
2829                return false;
2830
2831        spin_lock_irq(&gcwq->lock);
2832        if (!list_empty(&work->entry)) {
2833                /*
2834                 * See the comment near try_to_grab_pending()->smp_rmb().
2835                 * If it was re-queued to a different gcwq under us, we
2836                 * are not going to wait.
2837                 */
2838                smp_rmb();
2839                cwq = get_work_cwq(work);
2840                if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2841                        goto already_gone;
2842        } else {
2843                worker = find_worker_executing_work(gcwq, work);
2844                if (!worker)
2845                        goto already_gone;
2846                cwq = worker->current_cwq;
2847        }
2848
2849        insert_wq_barrier(cwq, barr, work, worker);
2850        spin_unlock_irq(&gcwq->lock);
2851
2852        /*
2853         * If @max_active is 1 or rescuer is in use, flushing another work
2854         * item on the same workqueue may lead to deadlock.  Make sure the
2855         * flusher is not running on the same workqueue by verifying write
2856         * access.
2857         */
2858        if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2859                lock_map_acquire(&cwq->wq->lockdep_map);
2860        else
2861                lock_map_acquire_read(&cwq->wq->lockdep_map);
2862        lock_map_release(&cwq->wq->lockdep_map);
2863
2864        return true;
2865already_gone:
2866        spin_unlock_irq(&gcwq->lock);
2867        return false;
2868}
2869
2870/**
2871 * flush_work - wait for a work to finish executing the last queueing instance
2872 * @work: the work to flush
2873 *
2874 * Wait until @work has finished execution.  @work is guaranteed to be idle
2875 * on return if it hasn't been requeued since flush started.
2876 *
2877 * RETURNS:
2878 * %true if flush_work() waited for the work to finish execution,
2879 * %false if it was already idle.
2880 */
2881bool flush_work(struct work_struct *work)
2882{
2883        struct wq_barrier barr;
2884
2885        lock_map_acquire(&work->lockdep_map);
2886        lock_map_release(&work->lockdep_map);
2887
2888        if (start_flush_work(work, &barr)) {
2889                wait_for_completion(&barr.done);
2890                destroy_work_on_stack(&barr.work);
2891                return true;
2892        } else {
2893                return false;
2894        }
2895}
2896EXPORT_SYMBOL_GPL(flush_work);
2897
2898static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2899{
2900        unsigned long flags;
2901        int ret;
2902
2903        do {
2904                ret = try_to_grab_pending(work, is_dwork, &flags);
2905                /*
2906                 * If someone else is canceling, wait for the same event it
2907                 * would be waiting for before retrying.
2908                 */
2909                if (unlikely(ret == -ENOENT))
2910                        flush_work(work);
2911        } while (unlikely(ret < 0));
2912
2913        /* tell other tasks trying to grab @work to back off */
2914        mark_work_canceling(work);
2915        local_irq_restore(flags);
2916
2917        flush_work(work);
2918        clear_work_data(work);
2919        return ret;
2920}
2921
2922/**
2923 * cancel_work_sync - cancel a work and wait for it to finish
2924 * @work: the work to cancel
2925 *
2926 * Cancel @work and wait for its execution to finish.  This function
2927 * can be used even if the work re-queues itself or migrates to
2928 * another workqueue.  On return from this function, @work is
2929 * guaranteed to be not pending or executing on any CPU.
2930 *
2931 * cancel_work_sync(&delayed_work->work) must not be used for
2932 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2933 *
2934 * The caller must ensure that the workqueue on which @work was last
2935 * queued can't be destroyed before this function returns.
2936 *
2937 * RETURNS:
2938 * %true if @work was pending, %false otherwise.
2939 */
2940bool cancel_work_sync(struct work_struct *work)
2941{
2942        return __cancel_work_timer(work, false);
2943}
2944EXPORT_SYMBOL_GPL(cancel_work_sync);
2945
2946/**
2947 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2948 * @dwork: the delayed work to flush
2949 *
2950 * Delayed timer is cancelled and the pending work is queued for
2951 * immediate execution.  Like flush_work(), this function only
2952 * considers the last queueing instance of @dwork.
2953 *
2954 * RETURNS:
2955 * %true if flush_work() waited for the work to finish execution,
2956 * %false if it was already idle.
2957 */
2958bool flush_delayed_work(struct delayed_work *dwork)
2959{
2960        local_irq_disable();
2961        if (del_timer_sync(&dwork->timer))
2962                __queue_work(dwork->cpu,
2963                             get_work_cwq(&dwork->work)->wq, &dwork->work);
2964        local_irq_enable();
2965        return flush_work(&dwork->work);
2966}
2967EXPORT_SYMBOL(flush_delayed_work);
2968
2969/**
2970 * cancel_delayed_work - cancel a delayed work
2971 * @dwork: delayed_work to cancel
2972 *
2973 * Kill off a pending delayed_work.  Returns %true if @dwork was pending
2974 * and canceled; %false if wasn't pending.  Note that the work callback
2975 * function may still be running on return, unless it returns %true and the
2976 * work doesn't re-arm itself.  Explicitly flush or use
2977 * cancel_delayed_work_sync() to wait on it.
2978 *
2979 * This function is safe to call from any context including IRQ handler.
2980 */
2981bool cancel_delayed_work(struct delayed_work *dwork)
2982{
2983        unsigned long flags;
2984        int ret;
2985
2986        do {
2987                ret = try_to_grab_pending(&dwork->work, true, &flags);
2988        } while (unlikely(ret == -EAGAIN));
2989
2990        if (unlikely(ret < 0))
2991                return false;
2992
2993        set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2994        local_irq_restore(flags);
2995        return ret;
2996}
2997EXPORT_SYMBOL(cancel_delayed_work);
2998
2999/**
3000 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3001 * @dwork: the delayed work cancel
3002 *
3003 * This is cancel_work_sync() for delayed works.
3004 *
3005 * RETURNS:
3006 * %true if @dwork was pending, %false otherwise.
3007 */
3008bool cancel_delayed_work_sync(struct delayed_work *dwork)
3009{
3010        return __cancel_work_timer(&dwork->work, true);
3011}
3012EXPORT_SYMBOL(cancel_delayed_work_sync);
3013
3014/**
3015 * schedule_work_on - put work task on a specific cpu
3016 * @cpu: cpu to put the work task on
3017 * @work: job to be done
3018 *
3019 * This puts a job on a specific cpu
3020 */
3021bool schedule_work_on(int cpu, struct work_struct *work)
3022{
3023        return queue_work_on(cpu, system_wq, work);
3024}
3025EXPORT_SYMBOL(schedule_work_on);
3026
3027/**
3028 * schedule_work - put work task in global workqueue
3029 * @work: job to be done
3030 *
3031 * Returns %false if @work was already on the kernel-global workqueue and
3032 * %true otherwise.
3033 *
3034 * This puts a job in the kernel-global workqueue if it was not already
3035 * queued and leaves it in the same position on the kernel-global
3036 * workqueue otherwise.
3037 */
3038bool schedule_work(struct work_struct *work)
3039{
3040        return queue_work(system_wq, work);
3041}
3042EXPORT_SYMBOL(schedule_work);
3043
3044/**
3045 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3046 * @cpu: cpu to use
3047 * @dwork: job to be done
3048 * @delay: number of jiffies to wait
3049 *
3050 * After waiting for a given time this puts a job in the kernel-global
3051 * workqueue on the specified CPU.
3052 */
3053bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3054                              unsigned long delay)
3055{
3056        return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3057}
3058EXPORT_SYMBOL(schedule_delayed_work_on);
3059
3060/**
3061 * schedule_delayed_work - put work task in global workqueue after delay
3062 * @dwork: job to be done
3063 * @delay: number of jiffies to wait or 0 for immediate execution
3064 *
3065 * After waiting for a given time this puts a job in the kernel-global
3066 * workqueue.
3067 */
3068bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3069{
3070        return queue_delayed_work(system_wq, dwork, delay);
3071}
3072EXPORT_SYMBOL(schedule_delayed_work);
3073
3074/**
3075 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3076 * @func: the function to call
3077 *
3078 * schedule_on_each_cpu() executes @func on each online CPU using the
3079 * system workqueue and blocks until all CPUs have completed.
3080 * schedule_on_each_cpu() is very slow.
3081 *
3082 * RETURNS:
3083 * 0 on success, -errno on failure.
3084 */
3085int schedule_on_each_cpu(work_func_t func)
3086{
3087        int cpu;
3088        struct work_struct __percpu *works;
3089
3090        works = alloc_percpu(struct work_struct);
3091        if (!works)
3092                return -ENOMEM;
3093
3094        get_online_cpus();
3095
3096        for_each_online_cpu(cpu) {
3097                struct work_struct *work = per_cpu_ptr(works, cpu);
3098
3099                INIT_WORK(work, func);
3100                schedule_work_on(cpu, work);
3101        }
3102
3103        for_each_online_cpu(cpu)
3104                flush_work(per_cpu_ptr(works, cpu));
3105
3106        put_online_cpus();
3107        free_percpu(works);
3108        return 0;
3109}
3110
3111/**
3112 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3113 *
3114 * Forces execution of the kernel-global workqueue and blocks until its
3115 * completion.
3116 *
3117 * Think twice before calling this function!  It's very easy to get into
3118 * trouble if you don't take great care.  Either of the following situations
3119 * will lead to deadlock:
3120 *
3121 *      One of the work items currently on the workqueue needs to acquire
3122 *      a lock held by your code or its caller.
3123 *
3124 *      Your code is running in the context of a work routine.
3125 *
3126 * They will be detected by lockdep when they occur, but the first might not
3127 * occur very often.  It depends on what work items are on the workqueue and
3128 * what locks they need, which you have no control over.
3129 *
3130 * In most situations flushing the entire workqueue is overkill; you merely
3131 * need to know that a particular work item isn't queued and isn't running.
3132 * In such cases you should use cancel_delayed_work_sync() or
3133 * cancel_work_sync() instead.
3134 */
3135void flush_scheduled_work(void)
3136{
3137        flush_workqueue(system_wq);
3138}
3139EXPORT_SYMBOL(flush_scheduled_work);
3140
3141/**
3142 * execute_in_process_context - reliably execute the routine with user context
3143 * @fn:         the function to execute
3144 * @ew:         guaranteed storage for the execute work structure (must
3145 *              be available when the work executes)
3146 *
3147 * Executes the function immediately if process context is available,
3148 * otherwise schedules the function for delayed execution.
3149 *
3150 * Returns:     0 - function was executed
3151 *              1 - function was scheduled for execution
3152 */
3153int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3154{
3155        if (!in_interrupt()) {
3156                fn(&ew->work);
3157                return 0;
3158        }
3159
3160        INIT_WORK(&ew->work, fn);
3161        schedule_work(&ew->work);
3162
3163        return 1;
3164}
3165EXPORT_SYMBOL_GPL(execute_in_process_context);
3166
3167int keventd_up(void)
3168{
3169        return system_wq != NULL;
3170}
3171
3172static int alloc_cwqs(struct workqueue_struct *wq)
3173{
3174        /*
3175         * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3176         * Make sure that the alignment isn't lower than that of
3177         * unsigned long long.
3178         */
3179        const size_t size = sizeof(struct cpu_workqueue_struct);
3180        const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3181                                   __alignof__(unsigned long long));
3182
3183        if (!(wq->flags & WQ_UNBOUND))
3184                wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3185        else {
3186                void *ptr;
3187
3188                /*
3189                 * Allocate enough room to align cwq and put an extra
3190                 * pointer at the end pointing back to the originally
3191                 * allocated pointer which will be used for free.
3192                 */
3193                ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3194                if (ptr) {
3195                        wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3196                        *(void **)(wq->cpu_wq.single + 1) = ptr;
3197                }
3198        }
3199
3200        /* just in case, make sure it's actually aligned */
3201        BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3202        return wq->cpu_wq.v ? 0 : -ENOMEM;
3203}
3204
3205static void free_cwqs(struct workqueue_struct *wq)
3206{
3207        if (!(wq->flags & WQ_UNBOUND))
3208                free_percpu(wq->cpu_wq.pcpu);
3209        else if (wq->cpu_wq.single) {
3210                /* the pointer to free is stored right after the cwq */
3211                kfree(*(void **)(wq->cpu_wq.single + 1));
3212        }
3213}
3214
3215static int wq_clamp_max_active(int max_active, unsigned int flags,
3216                               const char *name)
3217{
3218        int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3219
3220        if (max_active < 1 || max_active > lim)
3221                pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3222                        max_active, name, 1, lim);
3223
3224        return clamp_val(max_active, 1, lim);
3225}
3226
3227struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3228                                               unsigned int flags,
3229                                               int max_active,
3230                                               struct lock_class_key *key,
3231                                               const char *lock_name, ...)
3232{
3233        va_list args, args1;
3234        struct workqueue_struct *wq;
3235        unsigned int cpu;
3236        size_t namelen;
3237
3238        /* determine namelen, allocate wq and format name */
3239        va_start(args, lock_name);
3240        va_copy(args1, args);
3241        namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3242
3243        wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3244        if (!wq)
3245                goto err;
3246
3247        vsnprintf(wq->name, namelen, fmt, args1);
3248        va_end(args);
3249        va_end(args1);
3250
3251        /*
3252         * Workqueues which may be used during memory reclaim should
3253         * have a rescuer to guarantee forward progress.
3254         */
3255        if (flags & WQ_MEM_RECLAIM)
3256                flags |= WQ_RESCUER;
3257
3258        max_active = max_active ?: WQ_DFL_ACTIVE;
3259        max_active = wq_clamp_max_active(max_active, flags, wq->name);
3260
3261        /* init wq */
3262        wq->flags = flags;
3263        wq->saved_max_active = max_active;
3264        mutex_init(&wq->flush_mutex);
3265        atomic_set(&wq->nr_cwqs_to_flush, 0);
3266        INIT_LIST_HEAD(&wq->flusher_queue);
3267        INIT_LIST_HEAD(&wq->flusher_overflow);
3268
3269        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3270        INIT_LIST_HEAD(&wq->list);
3271
3272        if (alloc_cwqs(wq) < 0)
3273                goto err;
3274
3275        for_each_cwq_cpu(cpu, wq) {
3276                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3277                struct global_cwq *gcwq = get_gcwq(cpu);
3278                int pool_idx = (bool)(flags & WQ_HIGHPRI);
3279
3280                BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3281                cwq->pool = &gcwq->pools[pool_idx];
3282                cwq->wq = wq;
3283                cwq->flush_color = -1;
3284                cwq->max_active = max_active;
3285                INIT_LIST_HEAD(&cwq->delayed_works);
3286        }
3287
3288        if (flags & WQ_RESCUER) {
3289                struct worker *rescuer;
3290
3291                if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3292                        goto err;
3293
3294                wq->rescuer = rescuer = alloc_worker();
3295                if (!rescuer)
3296                        goto err;
3297
3298                rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3299                                               wq->name);
3300                if (IS_ERR(rescuer->task))
3301                        goto err;
3302
3303                rescuer->task->flags |= PF_THREAD_BOUND;
3304                wake_up_process(rescuer->task);
3305        }
3306
3307        /*
3308         * workqueue_lock protects global freeze state and workqueues
3309         * list.  Grab it, set max_active accordingly and add the new
3310         * workqueue to workqueues list.
3311         */
3312        spin_lock(&workqueue_lock);
3313
3314        if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3315                for_each_cwq_cpu(cpu, wq)
3316                        get_cwq(cpu, wq)->max_active = 0;
3317
3318        list_add(&wq->list, &workqueues);
3319
3320        spin_unlock(&workqueue_lock);
3321
3322        return wq;
3323err:
3324        if (wq) {
3325                free_cwqs(wq);
3326                free_mayday_mask(wq->mayday_mask);
3327                kfree(wq->rescuer);
3328                kfree(wq);
3329        }
3330        return NULL;
3331}
3332EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3333
3334/**
3335 * destroy_workqueue - safely terminate a workqueue
3336 * @wq: target workqueue
3337 *
3338 * Safely destroy a workqueue. All work currently pending will be done first.
3339 */
3340void destroy_workqueue(struct workqueue_struct *wq)
3341{
3342        unsigned int cpu;
3343
3344        /* drain it before proceeding with destruction */
3345        drain_workqueue(wq);
3346
3347        /*
3348         * wq list is used to freeze wq, remove from list after
3349         * flushing is complete in case freeze races us.
3350         */
3351        spin_lock(&workqueue_lock);
3352        list_del(&wq->list);
3353        spin_unlock(&workqueue_lock);
3354
3355        /* sanity check */
3356        for_each_cwq_cpu(cpu, wq) {
3357                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3358                int i;
3359
3360                for (i = 0; i < WORK_NR_COLORS; i++)
3361                        BUG_ON(cwq->nr_in_flight[i]);
3362                BUG_ON(cwq->nr_active);
3363                BUG_ON(!list_empty(&cwq->delayed_works));
3364        }
3365
3366        if (wq->flags & WQ_RESCUER) {
3367                kthread_stop(wq->rescuer->task);
3368                free_mayday_mask(wq->mayday_mask);
3369                kfree(wq->rescuer);
3370        }
3371
3372        free_cwqs(wq);
3373        kfree(wq);
3374}
3375EXPORT_SYMBOL_GPL(destroy_workqueue);
3376
3377/**
3378 * cwq_set_max_active - adjust max_active of a cwq
3379 * @cwq: target cpu_workqueue_struct
3380 * @max_active: new max_active value.
3381 *
3382 * Set @cwq->max_active to @max_active and activate delayed works if
3383 * increased.
3384 *
3385 * CONTEXT:
3386 * spin_lock_irq(gcwq->lock).
3387 */
3388static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3389{
3390        cwq->max_active = max_active;
3391
3392        while (!list_empty(&cwq->delayed_works) &&
3393               cwq->nr_active < cwq->max_active)
3394                cwq_activate_first_delayed(cwq);
3395}
3396
3397/**
3398 * workqueue_set_max_active - adjust max_active of a workqueue
3399 * @wq: target workqueue
3400 * @max_active: new max_active value.
3401 *
3402 * Set max_active of @wq to @max_active.
3403 *
3404 * CONTEXT:
3405 * Don't call from IRQ context.
3406 */
3407void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3408{
3409        unsigned int cpu;
3410
3411        max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3412
3413        spin_lock(&workqueue_lock);
3414
3415        wq->saved_max_active = max_active;
3416
3417        for_each_cwq_cpu(cpu, wq) {
3418                struct global_cwq *gcwq = get_gcwq(cpu);
3419
3420                spin_lock_irq(&gcwq->lock);
3421
3422                if (!(wq->flags & WQ_FREEZABLE) ||
3423                    !(gcwq->flags & GCWQ_FREEZING))
3424                        cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3425
3426                spin_unlock_irq(&gcwq->lock);
3427        }
3428
3429        spin_unlock(&workqueue_lock);
3430}
3431EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3432
3433/**
3434 * workqueue_congested - test whether a workqueue is congested
3435 * @cpu: CPU in question
3436 * @wq: target workqueue
3437 *
3438 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
3439 * no synchronization around this function and the test result is
3440 * unreliable and only useful as advisory hints or for debugging.
3441 *
3442 * RETURNS:
3443 * %true if congested, %false otherwise.
3444 */
3445bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3446{
3447        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3448
3449        return !list_empty(&cwq->delayed_works);
3450}
3451EXPORT_SYMBOL_GPL(workqueue_congested);
3452
3453/**
3454 * work_cpu - return the last known associated cpu for @work
3455 * @work: the work of interest
3456 *
3457 * RETURNS:
3458 * CPU number if @work was ever queued.  WORK_CPU_NONE otherwise.
3459 */
3460unsigned int work_cpu(struct work_struct *work)
3461{
3462        struct global_cwq *gcwq = get_work_gcwq(work);
3463
3464        return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3465}
3466EXPORT_SYMBOL_GPL(work_cpu);
3467
3468/**
3469 * work_busy - test whether a work is currently pending or running
3470 * @work: the work to be tested
3471 *
3472 * Test whether @work is currently pending or running.  There is no
3473 * synchronization around this function and the test result is
3474 * unreliable and only useful as advisory hints or for debugging.
3475 * Especially for reentrant wqs, the pending state might hide the
3476 * running state.
3477 *
3478 * RETURNS:
3479 * OR'd bitmask of WORK_BUSY_* bits.
3480 */
3481unsigned int work_busy(struct work_struct *work)
3482{
3483        struct global_cwq *gcwq = get_work_gcwq(work);
3484        unsigned long flags;
3485        unsigned int ret = 0;
3486
3487        if (!gcwq)
3488                return false;
3489
3490        spin_lock_irqsave(&gcwq->lock, flags);
3491
3492        if (work_pending(work))
3493                ret |= WORK_BUSY_PENDING;
3494        if (find_worker_executing_work(gcwq, work))
3495                ret |= WORK_BUSY_RUNNING;
3496
3497        spin_unlock_irqrestore(&gcwq->lock, flags);
3498
3499        return ret;
3500}
3501EXPORT_SYMBOL_GPL(work_busy);
3502
3503/*
3504 * CPU hotplug.
3505 *
3506 * There are two challenges in supporting CPU hotplug.  Firstly, there
3507 * are a lot of assumptions on strong associations among work, cwq and
3508 * gcwq which make migrating pending and scheduled works very
3509 * difficult to implement without impacting hot paths.  Secondly,
3510 * gcwqs serve mix of short, long and very long running works making
3511 * blocked draining impractical.
3512 *
3513 * This is solved by allowing a gcwq to be disassociated from the CPU
3514 * running as an unbound one and allowing it to be reattached later if the
3515 * cpu comes back online.
3516 */
3517
3518/* claim manager positions of all pools */
3519static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3520{
3521        struct worker_pool *pool;
3522
3523        for_each_worker_pool(pool, gcwq)
3524                mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3525        spin_lock_irq(&gcwq->lock);
3526}
3527
3528/* release manager positions */
3529static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3530{
3531        struct worker_pool *pool;
3532
3533        spin_unlock_irq(&gcwq->lock);
3534        for_each_worker_pool(pool, gcwq)
3535                mutex_unlock(&pool->assoc_mutex);
3536}
3537
3538static void gcwq_unbind_fn(struct work_struct *work)
3539{
3540        struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3541        struct worker_pool *pool;
3542        struct worker *worker;
3543        struct hlist_node *pos;
3544        int i;
3545
3546        BUG_ON(gcwq->cpu != smp_processor_id());
3547
3548        gcwq_claim_assoc_and_lock(gcwq);
3549
3550        /*
3551         * We've claimed all manager positions.  Make all workers unbound
3552         * and set DISASSOCIATED.  Before this, all workers except for the
3553         * ones which are still executing works from before the last CPU
3554         * down must be on the cpu.  After this, they may become diasporas.
3555         */
3556        for_each_worker_pool(pool, gcwq)
3557                list_for_each_entry(worker, &pool->idle_list, entry)
3558                        worker->flags |= WORKER_UNBOUND;
3559
3560        for_each_busy_worker(worker, i, pos, gcwq)
3561                worker->flags |= WORKER_UNBOUND;
3562
3563        gcwq->flags |= GCWQ_DISASSOCIATED;
3564
3565        gcwq_release_assoc_and_unlock(gcwq);
3566
3567        /*
3568         * Call schedule() so that we cross rq->lock and thus can guarantee
3569         * sched callbacks see the %WORKER_UNBOUND flag.  This is necessary
3570         * as scheduler callbacks may be invoked from other cpus.
3571         */
3572        schedule();
3573
3574        /*
3575         * Sched callbacks are disabled now.  Zap nr_running.  After this,
3576         * nr_running stays zero and need_more_worker() and keep_working()
3577         * are always true as long as the worklist is not empty.  @gcwq now
3578         * behaves as unbound (in terms of concurrency management) gcwq
3579         * which is served by workers tied to the CPU.
3580         *
3581         * On return from this function, the current worker would trigger
3582         * unbound chain execution of pending work items if other workers
3583         * didn't already.
3584         */
3585        for_each_worker_pool(pool, gcwq)
3586                atomic_set(get_pool_nr_running(pool), 0);
3587}
3588
3589/*
3590 * Workqueues should be brought up before normal priority CPU notifiers.
3591 * This will be registered high priority CPU notifier.
3592 */
3593static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3594                                               unsigned long action,
3595                                               void *hcpu)
3596{
3597        unsigned int cpu = (unsigned long)hcpu;
3598        struct global_cwq *gcwq = get_gcwq(cpu);
3599        struct worker_pool *pool;
3600
3601        switch (action & ~CPU_TASKS_FROZEN) {
3602        case CPU_UP_PREPARE:
3603                for_each_worker_pool(pool, gcwq) {
3604                        struct worker *worker;
3605
3606                        if (pool->nr_workers)
3607                                continue;
3608
3609                        worker = create_worker(pool);
3610                        if (!worker)
3611                                return NOTIFY_BAD;
3612
3613                        spin_lock_irq(&gcwq->lock);
3614                        start_worker(worker);
3615                        spin_unlock_irq(&gcwq->lock);
3616                }
3617                break;
3618
3619        case CPU_DOWN_FAILED:
3620        case CPU_ONLINE:
3621                gcwq_claim_assoc_and_lock(gcwq);
3622                gcwq->flags &= ~GCWQ_DISASSOCIATED;
3623                rebind_workers(gcwq);
3624                gcwq_release_assoc_and_unlock(gcwq);
3625                break;
3626        }
3627        return NOTIFY_OK;
3628}
3629
3630/*
3631 * Workqueues should be brought down after normal priority CPU notifiers.
3632 * This will be registered as low priority CPU notifier.
3633 */
3634static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3635                                                 unsigned long action,
3636                                                 void *hcpu)
3637{
3638        unsigned int cpu = (unsigned long)hcpu;
3639        struct work_struct unbind_work;
3640
3641        switch (action & ~CPU_TASKS_FROZEN) {
3642        case CPU_DOWN_PREPARE:
3643                /* unbinding should happen on the local CPU */
3644                INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3645                queue_work_on(cpu, system_highpri_wq, &unbind_work);
3646                flush_work(&unbind_work);
3647                break;
3648        }
3649        return NOTIFY_OK;
3650}
3651
3652#ifdef CONFIG_SMP
3653
3654struct work_for_cpu {
3655        struct work_struct work;
3656        long (*fn)(void *);
3657        void *arg;
3658        long ret;
3659};
3660
3661static void work_for_cpu_fn(struct work_struct *work)
3662{
3663        struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3664
3665        wfc->ret = wfc->fn(wfc->arg);
3666}
3667
3668/**
3669 * work_on_cpu - run a function in user context on a particular cpu
3670 * @cpu: the cpu to run on
3671 * @fn: the function to run
3672 * @arg: the function arg
3673 *
3674 * This will return the value @fn returns.
3675 * It is up to the caller to ensure that the cpu doesn't go offline.
3676 * The caller must not hold any locks which would prevent @fn from completing.
3677 */
3678long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3679{
3680        struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3681
3682        INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3683        schedule_work_on(cpu, &wfc.work);
3684        flush_work(&wfc.work);
3685        return wfc.ret;
3686}
3687EXPORT_SYMBOL_GPL(work_on_cpu);
3688#endif /* CONFIG_SMP */
3689
3690#ifdef CONFIG_FREEZER
3691
3692/**
3693 * freeze_workqueues_begin - begin freezing workqueues
3694 *
3695 * Start freezing workqueues.  After this function returns, all freezable
3696 * workqueues will queue new works to their frozen_works list instead of
3697 * gcwq->worklist.
3698 *
3699 * CONTEXT:
3700 * Grabs and releases workqueue_lock and gcwq->lock's.
3701 */
3702void freeze_workqueues_begin(void)
3703{
3704        unsigned int cpu;
3705
3706        spin_lock(&workqueue_lock);
3707
3708        BUG_ON(workqueue_freezing);
3709        workqueue_freezing = true;
3710
3711        for_each_gcwq_cpu(cpu) {
3712                struct global_cwq *gcwq = get_gcwq(cpu);
3713                struct workqueue_struct *wq;
3714
3715                spin_lock_irq(&gcwq->lock);
3716
3717                BUG_ON(gcwq->flags & GCWQ_FREEZING);
3718                gcwq->flags |= GCWQ_FREEZING;
3719
3720                list_for_each_entry(wq, &workqueues, list) {
3721                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3722
3723                        if (cwq && wq->flags & WQ_FREEZABLE)
3724                                cwq->max_active = 0;
3725                }
3726
3727                spin_unlock_irq(&gcwq->lock);
3728        }
3729
3730        spin_unlock(&workqueue_lock);
3731}
3732
3733/**
3734 * freeze_workqueues_busy - are freezable workqueues still busy?
3735 *
3736 * Check whether freezing is complete.  This function must be called
3737 * between freeze_workqueues_begin() and thaw_workqueues().
3738 *
3739 * CONTEXT:
3740 * Grabs and releases workqueue_lock.
3741 *
3742 * RETURNS:
3743 * %true if some freezable workqueues are still busy.  %false if freezing
3744 * is complete.
3745 */
3746bool freeze_workqueues_busy(void)
3747{
3748        unsigned int cpu;
3749        bool busy = false;
3750
3751        spin_lock(&workqueue_lock);
3752
3753        BUG_ON(!workqueue_freezing);
3754
3755        for_each_gcwq_cpu(cpu) {
3756                struct workqueue_struct *wq;
3757                /*
3758                 * nr_active is monotonically decreasing.  It's safe
3759                 * to peek without lock.
3760                 */
3761                list_for_each_entry(wq, &workqueues, list) {
3762                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3763
3764                        if (!cwq || !(wq->flags & WQ_FREEZABLE))
3765                                continue;
3766
3767                        BUG_ON(cwq->nr_active < 0);
3768                        if (cwq->nr_active) {
3769                                busy = true;
3770                                goto out_unlock;
3771                        }
3772                }
3773        }
3774out_unlock:
3775        spin_unlock(&workqueue_lock);
3776        return busy;
3777}
3778
3779/**
3780 * thaw_workqueues - thaw workqueues
3781 *
3782 * Thaw workqueues.  Normal queueing is restored and all collected
3783 * frozen works are transferred to their respective gcwq worklists.
3784 *
3785 * CONTEXT:
3786 * Grabs and releases workqueue_lock and gcwq->lock's.
3787 */
3788void thaw_workqueues(void)
3789{
3790        unsigned int cpu;
3791
3792        spin_lock(&workqueue_lock);
3793
3794        if (!workqueue_freezing)
3795                goto out_unlock;
3796
3797        for_each_gcwq_cpu(cpu) {
3798                struct global_cwq *gcwq = get_gcwq(cpu);
3799                struct worker_pool *pool;
3800                struct workqueue_struct *wq;
3801
3802                spin_lock_irq(&gcwq->lock);
3803
3804                BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3805                gcwq->flags &= ~GCWQ_FREEZING;
3806
3807                list_for_each_entry(wq, &workqueues, list) {
3808                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3809
3810                        if (!cwq || !(wq->flags & WQ_FREEZABLE))
3811                                continue;
3812
3813                        /* restore max_active and repopulate worklist */
3814                        cwq_set_max_active(cwq, wq->saved_max_active);
3815                }
3816
3817                for_each_worker_pool(pool, gcwq)
3818                        wake_up_worker(pool);
3819
3820                spin_unlock_irq(&gcwq->lock);
3821        }
3822
3823        workqueue_freezing = false;
3824out_unlock:
3825        spin_unlock(&workqueue_lock);
3826}
3827#endif /* CONFIG_FREEZER */
3828
3829static int __init init_workqueues(void)
3830{
3831        unsigned int cpu;
3832        int i;
3833
3834        /* make sure we have enough bits for OFFQ CPU number */
3835        BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3836                     WORK_CPU_LAST);
3837
3838        cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3839        hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3840
3841        /* initialize gcwqs */
3842        for_each_gcwq_cpu(cpu) {
3843                struct global_cwq *gcwq = get_gcwq(cpu);
3844                struct worker_pool *pool;
3845
3846                spin_lock_init(&gcwq->lock);
3847                gcwq->cpu = cpu;
3848                gcwq->flags |= GCWQ_DISASSOCIATED;
3849
3850                for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3851                        INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3852
3853                for_each_worker_pool(pool, gcwq) {
3854                        pool->gcwq = gcwq;
3855                        INIT_LIST_HEAD(&pool->worklist);
3856                        INIT_LIST_HEAD(&pool->idle_list);
3857
3858                        init_timer_deferrable(&pool->idle_timer);
3859                        pool->idle_timer.function = idle_worker_timeout;
3860                        pool->idle_timer.data = (unsigned long)pool;
3861
3862                        setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3863                                    (unsigned long)pool);
3864
3865                        mutex_init(&pool->assoc_mutex);
3866                        ida_init(&pool->worker_ida);
3867                }
3868        }
3869
3870        /* create the initial worker */
3871        for_each_online_gcwq_cpu(cpu) {
3872                struct global_cwq *gcwq = get_gcwq(cpu);
3873                struct worker_pool *pool;
3874
3875                if (cpu != WORK_CPU_UNBOUND)
3876                        gcwq->flags &= ~GCWQ_DISASSOCIATED;
3877
3878                for_each_worker_pool(pool, gcwq) {
3879                        struct worker *worker;
3880
3881                        worker = create_worker(pool);
3882                        BUG_ON(!worker);
3883                        spin_lock_irq(&gcwq->lock);
3884                        start_worker(worker);
3885                        spin_unlock_irq(&gcwq->lock);
3886                }
3887        }
3888
3889        system_wq = alloc_workqueue("events", 0, 0);
3890        system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3891        system_long_wq = alloc_workqueue("events_long", 0, 0);
3892        system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3893                                            WQ_UNBOUND_MAX_ACTIVE);
3894        system_freezable_wq = alloc_workqueue("events_freezable",
3895                                              WQ_FREEZABLE, 0);
3896        BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3897               !system_unbound_wq || !system_freezable_wq);
3898        return 0;
3899}
3900early_initcall(init_workqueues);
3901
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.