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                WARN_ON_ONCE(worker->pool->gcwq->cpu != cpu);
 744                atomic_inc(get_pool_nr_running(worker->pool));
 745        }
 746}
 747
 748/**
 749 * wq_worker_sleeping - a worker is going to sleep
 750 * @task: task going to sleep
 751 * @cpu: CPU in question, must be the current CPU number
 752 *
 753 * This function is called during schedule() when a busy worker is
 754 * going to sleep.  Worker on the same cpu can be woken up by
 755 * returning pointer to its task.
 756 *
 757 * CONTEXT:
 758 * spin_lock_irq(rq->lock)
 759 *
 760 * RETURNS:
 761 * Worker task on @cpu to wake up, %NULL if none.
 762 */
 763struct task_struct *wq_worker_sleeping(struct task_struct *task,
 764                                       unsigned int cpu)
 765{
 766        struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 767        struct worker_pool *pool = worker->pool;
 768        atomic_t *nr_running = get_pool_nr_running(pool);
 769
 770        if (worker->flags & WORKER_NOT_RUNNING)
 771                return NULL;
 772
 773        /* this can only happen on the local cpu */
 774        BUG_ON(cpu != raw_smp_processor_id());
 775
 776        /*
 777         * The counterpart of the following dec_and_test, implied mb,
 778         * worklist not empty test sequence is in insert_work().
 779         * Please read comment there.
 780         *
 781         * NOT_RUNNING is clear.  This means that we're bound to and
 782         * running on the local cpu w/ rq lock held and preemption
 783         * disabled, which in turn means that none else could be
 784         * manipulating idle_list, so dereferencing idle_list without gcwq
 785         * lock is safe.
 786         */
 787        if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
 788                to_wakeup = first_worker(pool);
 789        return to_wakeup ? to_wakeup->task : NULL;
 790}
 791
 792/**
 793 * worker_set_flags - set worker flags and adjust nr_running accordingly
 794 * @worker: self
 795 * @flags: flags to set
 796 * @wakeup: wakeup an idle worker if necessary
 797 *
 798 * Set @flags in @worker->flags and adjust nr_running accordingly.  If
 799 * nr_running becomes zero and @wakeup is %true, an idle worker is
 800 * woken up.
 801 *
 802 * CONTEXT:
 803 * spin_lock_irq(gcwq->lock)
 804 */
 805static inline void worker_set_flags(struct worker *worker, unsigned int flags,
 806                                    bool wakeup)
 807{
 808        struct worker_pool *pool = worker->pool;
 809
 810        WARN_ON_ONCE(worker->task != current);
 811
 812        /*
 813         * If transitioning into NOT_RUNNING, adjust nr_running and
 814         * wake up an idle worker as necessary if requested by
 815         * @wakeup.
 816         */
 817        if ((flags & WORKER_NOT_RUNNING) &&
 818            !(worker->flags & WORKER_NOT_RUNNING)) {
 819                atomic_t *nr_running = get_pool_nr_running(pool);
 820
 821                if (wakeup) {
 822                        if (atomic_dec_and_test(nr_running) &&
 823                            !list_empty(&pool->worklist))
 824                                wake_up_worker(pool);
 825                } else
 826                        atomic_dec(nr_running);
 827        }
 828
 829        worker->flags |= flags;
 830}
 831
 832/**
 833 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 834 * @worker: self
 835 * @flags: flags to clear
 836 *
 837 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 838 *
 839 * CONTEXT:
 840 * spin_lock_irq(gcwq->lock)
 841 */
 842static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 843{
 844        struct worker_pool *pool = worker->pool;
 845        unsigned int oflags = worker->flags;
 846
 847        WARN_ON_ONCE(worker->task != current);
 848
 849        worker->flags &= ~flags;
 850
 851        /*
 852         * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 853         * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 854         * of multiple flags, not a single flag.
 855         */
 856        if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 857                if (!(worker->flags & WORKER_NOT_RUNNING))
 858                        atomic_inc(get_pool_nr_running(pool));
 859}
 860
 861/**
 862 * busy_worker_head - return the busy hash head for a work
 863 * @gcwq: gcwq of interest
 864 * @work: work to be hashed
 865 *
 866 * Return hash head of @gcwq for @work.
 867 *
 868 * CONTEXT:
 869 * spin_lock_irq(gcwq->lock).
 870 *
 871 * RETURNS:
 872 * Pointer to the hash head.
 873 */
 874static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
 875                                           struct work_struct *work)
 876{
 877        const int base_shift = ilog2(sizeof(struct work_struct));
 878        unsigned long v = (unsigned long)work;
 879
 880        /* simple shift and fold hash, do we need something better? */
 881        v >>= base_shift;
 882        v += v >> BUSY_WORKER_HASH_ORDER;
 883        v &= BUSY_WORKER_HASH_MASK;
 884
 885        return &gcwq->busy_hash[v];
 886}
 887
 888/**
 889 * __find_worker_executing_work - find worker which is executing a work
 890 * @gcwq: gcwq of interest
 891 * @bwh: hash head as returned by busy_worker_head()
 892 * @work: work to find worker for
 893 *
 894 * Find a worker which is executing @work on @gcwq.  @bwh should be
 895 * the hash head obtained by calling busy_worker_head() with the same
 896 * work.
 897 *
 898 * CONTEXT:
 899 * spin_lock_irq(gcwq->lock).
 900 *
 901 * RETURNS:
 902 * Pointer to worker which is executing @work if found, NULL
 903 * otherwise.
 904 */
 905static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
 906                                                   struct hlist_head *bwh,
 907                                                   struct work_struct *work)
 908{
 909        struct worker *worker;
 910        struct hlist_node *tmp;
 911
 912        hlist_for_each_entry(worker, tmp, bwh, hentry)
 913                if (worker->current_work == work)
 914                        return worker;
 915        return NULL;
 916}
 917
 918/**
 919 * find_worker_executing_work - find worker which is executing a work
 920 * @gcwq: gcwq of interest
 921 * @work: work to find worker for
 922 *
 923 * Find a worker which is executing @work on @gcwq.  This function is
 924 * identical to __find_worker_executing_work() except that this
 925 * function calculates @bwh itself.
 926 *
 927 * CONTEXT:
 928 * spin_lock_irq(gcwq->lock).
 929 *
 930 * RETURNS:
 931 * Pointer to worker which is executing @work if found, NULL
 932 * otherwise.
 933 */
 934static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
 935                                                 struct work_struct *work)
 936{
 937        return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
 938                                            work);
 939}
 940
 941/**
 942 * move_linked_works - move linked works to a list
 943 * @work: start of series of works to be scheduled
 944 * @head: target list to append @work to
 945 * @nextp: out paramter for nested worklist walking
 946 *
 947 * Schedule linked works starting from @work to @head.  Work series to
 948 * be scheduled starts at @work and includes any consecutive work with
 949 * WORK_STRUCT_LINKED set in its predecessor.
 950 *
 951 * If @nextp is not NULL, it's updated to point to the next work of
 952 * the last scheduled work.  This allows move_linked_works() to be
 953 * nested inside outer list_for_each_entry_safe().
 954 *
 955 * CONTEXT:
 956 * spin_lock_irq(gcwq->lock).
 957 */
 958static void move_linked_works(struct work_struct *work, struct list_head *head,
 959                              struct work_struct **nextp)
 960{
 961        struct work_struct *n;
 962
 963        /*
 964         * Linked worklist will always end before the end of the list,
 965         * use NULL for list head.
 966         */
 967        list_for_each_entry_safe_from(work, n, NULL, entry) {
 968                list_move_tail(&work->entry, head);
 969                if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
 970                        break;
 971        }
 972
 973        /*
 974         * If we're already inside safe list traversal and have moved
 975         * multiple works to the scheduled queue, the next position
 976         * needs to be updated.
 977         */
 978        if (nextp)
 979                *nextp = n;
 980}
 981
 982static void cwq_activate_delayed_work(struct work_struct *work)
 983{
 984        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
 985
 986        trace_workqueue_activate_work(work);
 987        move_linked_works(work, &cwq->pool->worklist, NULL);
 988        __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
 989        cwq->nr_active++;
 990}
 991
 992static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
 993{
 994        struct work_struct *work = list_first_entry(&cwq->delayed_works,
 995                                                    struct work_struct, entry);
 996
 997        cwq_activate_delayed_work(work);
 998}
 999
1000/**
1001 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1002 * @cwq: cwq of interest
1003 * @color: color of work which left the queue
1004 *
1005 * A work either has completed or is removed from pending queue,
1006 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1007 *
1008 * CONTEXT:
1009 * spin_lock_irq(gcwq->lock).
1010 */
1011static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color)
1012{
1013        /* ignore uncolored works */
1014        if (color == WORK_NO_COLOR)
1015                return;
1016
1017        cwq->nr_in_flight[color]--;
1018
1019        cwq->nr_active--;
1020        if (!list_empty(&cwq->delayed_works)) {
1021                /* one down, submit a delayed one */
1022                if (cwq->nr_active < cwq->max_active)
1023                        cwq_activate_first_delayed(cwq);
1024        }
1025
1026        /* is flush in progress and are we at the flushing tip? */
1027        if (likely(cwq->flush_color != color))
1028                return;
1029
1030        /* are there still in-flight works? */
1031        if (cwq->nr_in_flight[color])
1032                return;
1033
1034        /* this cwq is done, clear flush_color */
1035        cwq->flush_color = -1;
1036
1037        /*
1038         * If this was the last cwq, wake up the first flusher.  It
1039         * will handle the rest.
1040         */
1041        if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1042                complete(&cwq->wq->first_flusher->done);
1043}
1044
1045/**
1046 * try_to_grab_pending - steal work item from worklist and disable irq
1047 * @work: work item to steal
1048 * @is_dwork: @work is a delayed_work
1049 * @flags: place to store irq state
1050 *
1051 * Try to grab PENDING bit of @work.  This function can handle @work in any
1052 * stable state - idle, on timer or on worklist.  Return values are
1053 *
1054 *  1           if @work was pending and we successfully stole PENDING
1055 *  0           if @work was idle and we claimed PENDING
1056 *  -EAGAIN     if PENDING couldn't be grabbed at the moment, safe to busy-retry
1057 *  -ENOENT     if someone else is canceling @work, this state may persist
1058 *              for arbitrarily long
1059 *
1060 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1061 * interrupted while holding PENDING and @work off queue, irq must be
1062 * disabled on entry.  This, combined with delayed_work->timer being
1063 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1064 *
1065 * On successful return, >= 0, irq is disabled and the caller is
1066 * responsible for releasing it using local_irq_restore(*@flags).
1067 *
1068 * This function is safe to call from any context including IRQ handler.
1069 */
1070static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1071                               unsigned long *flags)
1072{
1073        struct global_cwq *gcwq;
1074
1075        local_irq_save(*flags);
1076
1077        /* try to steal the timer if it exists */
1078        if (is_dwork) {
1079                struct delayed_work *dwork = to_delayed_work(work);
1080
1081                /*
1082                 * dwork->timer is irqsafe.  If del_timer() fails, it's
1083                 * guaranteed that the timer is not queued anywhere and not
1084                 * running on the local CPU.
1085                 */
1086                if (likely(del_timer(&dwork->timer)))
1087                        return 1;
1088        }
1089
1090        /* try to claim PENDING the normal way */
1091        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1092                return 0;
1093
1094        /*
1095         * The queueing is in progress, or it is already queued. Try to
1096         * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1097         */
1098        gcwq = get_work_gcwq(work);
1099        if (!gcwq)
1100                goto fail;
1101
1102        spin_lock(&gcwq->lock);
1103        if (!list_empty(&work->entry)) {
1104                /*
1105                 * This work is queued, but perhaps we locked the wrong gcwq.
1106                 * In that case we must see the new value after rmb(), see
1107                 * insert_work()->wmb().
1108                 */
1109                smp_rmb();
1110                if (gcwq == get_work_gcwq(work)) {
1111                        debug_work_deactivate(work);
1112
1113                        /*
1114                         * A delayed work item cannot be grabbed directly
1115                         * because it might have linked NO_COLOR work items
1116                         * which, if left on the delayed_list, will confuse
1117                         * cwq->nr_active management later on and cause
1118                         * stall.  Make sure the work item is activated
1119                         * before grabbing.
1120                         */
1121                        if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1122                                cwq_activate_delayed_work(work);
1123
1124                        list_del_init(&work->entry);
1125                        cwq_dec_nr_in_flight(get_work_cwq(work),
1126                                get_work_color(work));
1127
1128                        spin_unlock(&gcwq->lock);
1129                        return 1;
1130                }
1131        }
1132        spin_unlock(&gcwq->lock);
1133fail:
1134        local_irq_restore(*flags);
1135        if (work_is_canceling(work))
1136                return -ENOENT;
1137        cpu_relax();
1138        return -EAGAIN;
1139}
1140
1141/**
1142 * insert_work - insert a work into gcwq
1143 * @cwq: cwq @work belongs to
1144 * @work: work to insert
1145 * @head: insertion point
1146 * @extra_flags: extra WORK_STRUCT_* flags to set
1147 *
1148 * Insert @work which belongs to @cwq into @gcwq after @head.
1149 * @extra_flags is or'd to work_struct flags.
1150 *
1151 * CONTEXT:
1152 * spin_lock_irq(gcwq->lock).
1153 */
1154static void insert_work(struct cpu_workqueue_struct *cwq,
1155                        struct work_struct *work, struct list_head *head,
1156                        unsigned int extra_flags)
1157{
1158        struct worker_pool *pool = cwq->pool;
1159
1160        /* we own @work, set data and link */
1161        set_work_cwq(work, cwq, extra_flags);
1162
1163        /*
1164         * Ensure that we get the right work->data if we see the
1165         * result of list_add() below, see try_to_grab_pending().
1166         */
1167        smp_wmb();
1168
1169        list_add_tail(&work->entry, head);
1170
1171        /*
1172         * Ensure either worker_sched_deactivated() sees the above
1173         * list_add_tail() or we see zero nr_running to avoid workers
1174         * lying around lazily while there are works to be processed.
1175         */
1176        smp_mb();
1177
1178        if (__need_more_worker(pool))
1179                wake_up_worker(pool);
1180}
1181
1182/*
1183 * Test whether @work is being queued from another work executing on the
1184 * same workqueue.  This is rather expensive and should only be used from
1185 * cold paths.
1186 */
1187static bool is_chained_work(struct workqueue_struct *wq)
1188{
1189        unsigned long flags;
1190        unsigned int cpu;
1191
1192        for_each_gcwq_cpu(cpu) {
1193                struct global_cwq *gcwq = get_gcwq(cpu);
1194                struct worker *worker;
1195                struct hlist_node *pos;
1196                int i;
1197
1198                spin_lock_irqsave(&gcwq->lock, flags);
1199                for_each_busy_worker(worker, i, pos, gcwq) {
1200                        if (worker->task != current)
1201                                continue;
1202                        spin_unlock_irqrestore(&gcwq->lock, flags);
1203                        /*
1204                         * I'm @worker, no locking necessary.  See if @work
1205                         * is headed to the same workqueue.
1206                         */
1207                        return worker->current_cwq->wq == wq;
1208                }
1209                spin_unlock_irqrestore(&gcwq->lock, flags);
1210        }
1211        return false;
1212}
1213
1214static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1215                         struct work_struct *work)
1216{
1217        struct global_cwq *gcwq;
1218        struct cpu_workqueue_struct *cwq;
1219        struct list_head *worklist;
1220        unsigned int work_flags;
1221        unsigned int req_cpu = cpu;
1222
1223        /*
1224         * While a work item is PENDING && off queue, a task trying to
1225         * steal the PENDING will busy-loop waiting for it to either get
1226         * queued or lose PENDING.  Grabbing PENDING and queueing should
1227         * happen with IRQ disabled.
1228         */
1229        WARN_ON_ONCE(!irqs_disabled());
1230
1231        debug_work_activate(work);
1232
1233        /* if dying, only works from the same workqueue are allowed */
1234        if (unlikely(wq->flags & WQ_DRAINING) &&
1235            WARN_ON_ONCE(!is_chained_work(wq)))
1236                return;
1237
1238        /* determine gcwq to use */
1239        if (!(wq->flags & WQ_UNBOUND)) {
1240                struct global_cwq *last_gcwq;
1241
1242                if (cpu == WORK_CPU_UNBOUND)
1243                        cpu = raw_smp_processor_id();
1244
1245                /*
1246                 * It's multi cpu.  If @work was previously on a different
1247                 * cpu, it might still be running there, in which case the
1248                 * work needs to be queued on that cpu to guarantee
1249                 * non-reentrancy.
1250                 */
1251                gcwq = get_gcwq(cpu);
1252                last_gcwq = get_work_gcwq(work);
1253
1254                if (last_gcwq && last_gcwq != gcwq) {
1255                        struct worker *worker;
1256
1257                        spin_lock(&last_gcwq->lock);
1258
1259                        worker = find_worker_executing_work(last_gcwq, work);
1260
1261                        if (worker && worker->current_cwq->wq == wq)
1262                                gcwq = last_gcwq;
1263                        else {
1264                                /* meh... not running there, queue here */
1265                                spin_unlock(&last_gcwq->lock);
1266                                spin_lock(&gcwq->lock);
1267                        }
1268                } else {
1269                        spin_lock(&gcwq->lock);
1270                }
1271        } else {
1272                gcwq = get_gcwq(WORK_CPU_UNBOUND);
1273                spin_lock(&gcwq->lock);
1274        }
1275
1276        /* gcwq determined, get cwq and queue */
1277        cwq = get_cwq(gcwq->cpu, wq);
1278        trace_workqueue_queue_work(req_cpu, cwq, work);
1279
1280        if (WARN_ON(!list_empty(&work->entry))) {
1281                spin_unlock(&gcwq->lock);
1282                return;
1283        }
1284
1285        cwq->nr_in_flight[cwq->work_color]++;
1286        work_flags = work_color_to_flags(cwq->work_color);
1287
1288        if (likely(cwq->nr_active < cwq->max_active)) {
1289                trace_workqueue_activate_work(work);
1290                cwq->nr_active++;
1291                worklist = &cwq->pool->worklist;
1292        } else {
1293                work_flags |= WORK_STRUCT_DELAYED;
1294                worklist = &cwq->delayed_works;
1295        }
1296
1297        insert_work(cwq, work, worklist, work_flags);
1298
1299        spin_unlock(&gcwq->lock);
1300}
1301
1302/**
1303 * queue_work_on - queue work on specific cpu
1304 * @cpu: CPU number to execute work on
1305 * @wq: workqueue to use
1306 * @work: work to queue
1307 *
1308 * Returns %false if @work was already on a queue, %true otherwise.
1309 *
1310 * We queue the work to a specific CPU, the caller must ensure it
1311 * can't go away.
1312 */
1313bool queue_work_on(int cpu, struct workqueue_struct *wq,
1314                   struct work_struct *work)
1315{
1316        bool ret = false;
1317        unsigned long flags;
1318
1319        local_irq_save(flags);
1320
1321        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1322                __queue_work(cpu, wq, work);
1323                ret = true;
1324        }
1325
1326        local_irq_restore(flags);
1327        return ret;
1328}
1329EXPORT_SYMBOL_GPL(queue_work_on);
1330
1331/**
1332 * queue_work - queue work on a workqueue
1333 * @wq: workqueue to use
1334 * @work: work to queue
1335 *
1336 * Returns %false if @work was already on a queue, %true otherwise.
1337 *
1338 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1339 * it can be processed by another CPU.
1340 */
1341bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1342{
1343        return queue_work_on(WORK_CPU_UNBOUND, wq, work);
1344}
1345EXPORT_SYMBOL_GPL(queue_work);
1346
1347void delayed_work_timer_fn(unsigned long __data)
1348{
1349        struct delayed_work *dwork = (struct delayed_work *)__data;
1350        struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1351
1352        /* should have been called from irqsafe timer with irq already off */
1353        __queue_work(dwork->cpu, cwq->wq, &dwork->work);
1354}
1355EXPORT_SYMBOL(delayed_work_timer_fn);
1356
1357static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1358                                struct delayed_work *dwork, unsigned long delay)
1359{
1360        struct timer_list *timer = &dwork->timer;
1361        struct work_struct *work = &dwork->work;
1362        unsigned int lcpu;
1363
1364        WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1365                     timer->data != (unsigned long)dwork);
1366        WARN_ON_ONCE(timer_pending(timer));
1367        WARN_ON_ONCE(!list_empty(&work->entry));
1368
1369        /*
1370         * If @delay is 0, queue @dwork->work immediately.  This is for
1371         * both optimization and correctness.  The earliest @timer can
1372         * expire is on the closest next tick and delayed_work users depend
1373         * on that there's no such delay when @delay is 0.
1374         */
1375        if (!delay) {
1376                __queue_work(cpu, wq, &dwork->work);
1377                return;
1378        }
1379
1380        timer_stats_timer_set_start_info(&dwork->timer);
1381
1382        /*
1383         * This stores cwq for the moment, for the timer_fn.  Note that the
1384         * work's gcwq is preserved to allow reentrance detection for
1385         * delayed works.
1386         */
1387        if (!(wq->flags & WQ_UNBOUND)) {
1388                struct global_cwq *gcwq = get_work_gcwq(work);
1389
1390                /*
1391                 * If we cannot get the last gcwq from @work directly,
1392                 * select the last CPU such that it avoids unnecessarily
1393                 * triggering non-reentrancy check in __queue_work().
1394                 */
1395                lcpu = cpu;
1396                if (gcwq)
1397                        lcpu = gcwq->cpu;
1398                if (lcpu == WORK_CPU_UNBOUND)
1399                        lcpu = raw_smp_processor_id();
1400        } else {
1401                lcpu = WORK_CPU_UNBOUND;
1402        }
1403
1404        set_work_cwq(work, get_cwq(lcpu, wq), 0);
1405
1406        dwork->cpu = cpu;
1407        timer->expires = jiffies + delay;
1408
1409        if (unlikely(cpu != WORK_CPU_UNBOUND))
1410                add_timer_on(timer, cpu);
1411        else
1412                add_timer(timer);
1413}
1414
1415/**
1416 * queue_delayed_work_on - queue work on specific CPU after delay
1417 * @cpu: CPU number to execute work on
1418 * @wq: workqueue to use
1419 * @dwork: work to queue
1420 * @delay: number of jiffies to wait before queueing
1421 *
1422 * Returns %false if @work was already on a queue, %true otherwise.  If
1423 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1424 * execution.
1425 */
1426bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1427                           struct delayed_work *dwork, unsigned long delay)
1428{
1429        struct work_struct *work = &dwork->work;
1430        bool ret = false;
1431        unsigned long flags;
1432
1433        /* read the comment in __queue_work() */
1434        local_irq_save(flags);
1435
1436        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1437                __queue_delayed_work(cpu, wq, dwork, delay);
1438                ret = true;
1439        }
1440
1441        local_irq_restore(flags);
1442        return ret;
1443}
1444EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1445
1446/**
1447 * queue_delayed_work - queue work on a workqueue after delay
1448 * @wq: workqueue to use
1449 * @dwork: delayable work to queue
1450 * @delay: number of jiffies to wait before queueing
1451 *
1452 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
1453 */
1454bool queue_delayed_work(struct workqueue_struct *wq,
1455                        struct delayed_work *dwork, unsigned long delay)
1456{
1457        return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1458}
1459EXPORT_SYMBOL_GPL(queue_delayed_work);
1460
1461/**
1462 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1463 * @cpu: CPU number to execute work on
1464 * @wq: workqueue to use
1465 * @dwork: work to queue
1466 * @delay: number of jiffies to wait before queueing
1467 *
1468 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1469 * modify @dwork's timer so that it expires after @delay.  If @delay is
1470 * zero, @work is guaranteed to be scheduled immediately regardless of its
1471 * current state.
1472 *
1473 * Returns %false if @dwork was idle and queued, %true if @dwork was
1474 * pending and its timer was modified.
1475 *
1476 * This function is safe to call from any context including IRQ handler.
1477 * See try_to_grab_pending() for details.
1478 */
1479bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1480                         struct delayed_work *dwork, unsigned long delay)
1481{
1482        unsigned long flags;
1483        int ret;
1484
1485        do {
1486                ret = try_to_grab_pending(&dwork->work, true, &flags);
1487        } while (unlikely(ret == -EAGAIN));
1488
1489        if (likely(ret >= 0)) {
1490                __queue_delayed_work(cpu, wq, dwork, delay);
1491                local_irq_restore(flags);
1492        }
1493
1494        /* -ENOENT from try_to_grab_pending() becomes %true */
1495        return ret;
1496}
1497EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1498
1499/**
1500 * mod_delayed_work - modify delay of or queue a delayed work
1501 * @wq: workqueue to use
1502 * @dwork: work to queue
1503 * @delay: number of jiffies to wait before queueing
1504 *
1505 * mod_delayed_work_on() on local CPU.
1506 */
1507bool mod_delayed_work(struct workqueue_struct *wq, struct delayed_work *dwork,
1508                      unsigned long delay)
1509{
1510        return mod_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
1511}
1512EXPORT_SYMBOL_GPL(mod_delayed_work);
1513
1514/**
1515 * worker_enter_idle - enter idle state
1516 * @worker: worker which is entering idle state
1517 *
1518 * @worker is entering idle state.  Update stats and idle timer if
1519 * necessary.
1520 *
1521 * LOCKING:
1522 * spin_lock_irq(gcwq->lock).
1523 */
1524static void worker_enter_idle(struct worker *worker)
1525{
1526        struct worker_pool *pool = worker->pool;
1527        struct global_cwq *gcwq = pool->gcwq;
1528
1529        BUG_ON(worker->flags & WORKER_IDLE);
1530        BUG_ON(!list_empty(&worker->entry) &&
1531               (worker->hentry.next || worker->hentry.pprev));
1532
1533        /* can't use worker_set_flags(), also called from start_worker() */
1534        worker->flags |= WORKER_IDLE;
1535        pool->nr_idle++;
1536        worker->last_active = jiffies;
1537
1538        /* idle_list is LIFO */
1539        list_add(&worker->entry, &pool->idle_list);
1540
1541        if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1542                mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1543
1544        /*
1545         * Sanity check nr_running.  Because gcwq_unbind_fn() releases
1546         * gcwq->lock between setting %WORKER_UNBOUND and zapping
1547         * nr_running, the warning may trigger spuriously.  Check iff
1548         * unbind is not in progress.
1549         */
1550        WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
1551                     pool->nr_workers == pool->nr_idle &&
1552                     atomic_read(get_pool_nr_running(pool)));
1553}
1554
1555/**
1556 * worker_leave_idle - leave idle state
1557 * @worker: worker which is leaving idle state
1558 *
1559 * @worker is leaving idle state.  Update stats.
1560 *
1561 * LOCKING:
1562 * spin_lock_irq(gcwq->lock).
1563 */
1564static void worker_leave_idle(struct worker *worker)
1565{
1566        struct worker_pool *pool = worker->pool;
1567
1568        BUG_ON(!(worker->flags & WORKER_IDLE));
1569        worker_clr_flags(worker, WORKER_IDLE);
1570        pool->nr_idle--;
1571        list_del_init(&worker->entry);
1572}
1573
1574/**
1575 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1576 * @worker: self
1577 *
1578 * Works which are scheduled while the cpu is online must at least be
1579 * scheduled to a worker which is bound to the cpu so that if they are
1580 * flushed from cpu callbacks while cpu is going down, they are
1581 * guaranteed to execute on the cpu.
1582 *
1583 * This function is to be used by rogue workers and rescuers to bind
1584 * themselves to the target cpu and may race with cpu going down or
1585 * coming online.  kthread_bind() can't be used because it may put the
1586 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1587 * verbatim as it's best effort and blocking and gcwq may be
1588 * [dis]associated in the meantime.
1589 *
1590 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1591 * binding against %GCWQ_DISASSOCIATED which is set during
1592 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1593 * enters idle state or fetches works without dropping lock, it can
1594 * guarantee the scheduling requirement described in the first paragraph.
1595 *
1596 * CONTEXT:
1597 * Might sleep.  Called without any lock but returns with gcwq->lock
1598 * held.
1599 *
1600 * RETURNS:
1601 * %true if the associated gcwq is online (@worker is successfully
1602 * bound), %false if offline.
1603 */
1604static bool worker_maybe_bind_and_lock(struct worker *worker)
1605__acquires(&gcwq->lock)
1606{
1607        struct global_cwq *gcwq = worker->pool->gcwq;
1608        struct task_struct *task = worker->task;
1609
1610        while (true) {
1611                /*
1612                 * The following call may fail, succeed or succeed
1613                 * without actually migrating the task to the cpu if
1614                 * it races with cpu hotunplug operation.  Verify
1615                 * against GCWQ_DISASSOCIATED.
1616                 */
1617                if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1618                        set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1619
1620                spin_lock_irq(&gcwq->lock);
1621                if (gcwq->flags & GCWQ_DISASSOCIATED)
1622                        return false;
1623                if (task_cpu(task) == gcwq->cpu &&
1624                    cpumask_equal(&current->cpus_allowed,
1625                                  get_cpu_mask(gcwq->cpu)))
1626                        return true;
1627                spin_unlock_irq(&gcwq->lock);
1628
1629                /*
1630                 * We've raced with CPU hot[un]plug.  Give it a breather
1631                 * and retry migration.  cond_resched() is required here;
1632                 * otherwise, we might deadlock against cpu_stop trying to
1633                 * bring down the CPU on non-preemptive kernel.
1634                 */
1635                cpu_relax();
1636                cond_resched();
1637        }
1638}
1639
1640/*
1641 * Rebind an idle @worker to its CPU.  worker_thread() will test
1642 * list_empty(@worker->entry) before leaving idle and call this function.
1643 */
1644static void idle_worker_rebind(struct worker *worker)
1645{
1646        struct global_cwq *gcwq = worker->pool->gcwq;
1647
1648        /* CPU may go down again inbetween, clear UNBOUND only on success */
1649        if (worker_maybe_bind_and_lock(worker))
1650                worker_clr_flags(worker, WORKER_UNBOUND);
1651
1652        /* rebind complete, become available again */
1653        list_add(&worker->entry, &worker->pool->idle_list);
1654        spin_unlock_irq(&gcwq->lock);
1655}
1656
1657/*
1658 * Function for @worker->rebind.work used to rebind unbound busy workers to
1659 * the associated cpu which is coming back online.  This is scheduled by
1660 * cpu up but can race with other cpu hotplug operations and may be
1661 * executed twice without intervening cpu down.
1662 */
1663static void busy_worker_rebind_fn(struct work_struct *work)
1664{
1665        struct worker *worker = container_of(work, struct worker, rebind_work);
1666        struct global_cwq *gcwq = worker->pool->gcwq;
1667
1668        if (worker_maybe_bind_and_lock(worker))
1669                worker_clr_flags(worker, WORKER_UNBOUND);
1670
1671        spin_unlock_irq(&gcwq->lock);
1672}
1673
1674/**
1675 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1676 * @gcwq: gcwq of interest
1677 *
1678 * @gcwq->cpu is coming online.  Rebind all workers to the CPU.  Rebinding
1679 * is different for idle and busy ones.
1680 *
1681 * Idle ones will be removed from the idle_list and woken up.  They will
1682 * add themselves back after completing rebind.  This ensures that the
1683 * idle_list doesn't contain any unbound workers when re-bound busy workers
1684 * try to perform local wake-ups for concurrency management.
1685 *
1686 * Busy workers can rebind after they finish their current work items.
1687 * Queueing the rebind work item at the head of the scheduled list is
1688 * enough.  Note that nr_running will be properly bumped as busy workers
1689 * rebind.
1690 *
1691 * On return, all non-manager workers are scheduled for rebind - see
1692 * manage_workers() for the manager special case.  Any idle worker
1693 * including the manager will not appear on @idle_list until rebind is
1694 * complete, making local wake-ups safe.
1695 */
1696static void rebind_workers(struct global_cwq *gcwq)
1697{
1698        struct worker_pool *pool;
1699        struct worker *worker, *n;
1700        struct hlist_node *pos;
1701        int i;
1702
1703        lockdep_assert_held(&gcwq->lock);
1704
1705        for_each_worker_pool(pool, gcwq)
1706                lockdep_assert_held(&pool->assoc_mutex);
1707
1708        /* dequeue and kick idle ones */
1709        for_each_worker_pool(pool, gcwq) {
1710                list_for_each_entry_safe(worker, n, &pool->idle_list, entry) {
1711                        /*
1712                         * idle workers should be off @pool->idle_list
1713                         * until rebind is complete to avoid receiving
1714                         * premature local wake-ups.
1715                         */
1716                        list_del_init(&worker->entry);
1717
1718                        /*
1719                         * worker_thread() will see the above dequeuing
1720                         * and call idle_worker_rebind().
1721                         */
1722                        wake_up_process(worker->task);
1723                }
1724        }
1725
1726        /* rebind busy workers */
1727        for_each_busy_worker(worker, i, pos, gcwq) {
1728                struct work_struct *rebind_work = &worker->rebind_work;
1729                struct workqueue_struct *wq;
1730
1731                if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1732                                     work_data_bits(rebind_work)))
1733                        continue;
1734
1735                debug_work_activate(rebind_work);
1736
1737                /*
1738                 * wq doesn't really matter but let's keep @worker->pool
1739                 * and @cwq->pool consistent for sanity.
1740                 */
1741                if (worker_pool_pri(worker->pool))
1742                        wq = system_highpri_wq;
1743                else
1744                        wq = system_wq;
1745
1746                insert_work(get_cwq(gcwq->cpu, wq), rebind_work,
1747                        worker->scheduled.next,
1748                        work_color_to_flags(WORK_NO_COLOR));
1749        }
1750}
1751
1752static struct worker *alloc_worker(void)
1753{
1754        struct worker *worker;
1755
1756        worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1757        if (worker) {
1758                INIT_LIST_HEAD(&worker->entry);
1759                INIT_LIST_HEAD(&worker->scheduled);
1760                INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
1761                /* on creation a worker is in !idle && prep state */
1762                worker->flags = WORKER_PREP;
1763        }
1764        return worker;
1765}
1766
1767/**
1768 * create_worker - create a new workqueue worker
1769 * @pool: pool the new worker will belong to
1770 *
1771 * Create a new worker which is bound to @pool.  The returned worker
1772 * can be started by calling start_worker() or destroyed using
1773 * destroy_worker().
1774 *
1775 * CONTEXT:
1776 * Might sleep.  Does GFP_KERNEL allocations.
1777 *
1778 * RETURNS:
1779 * Pointer to the newly created worker.
1780 */
1781static struct worker *create_worker(struct worker_pool *pool)
1782{
1783        struct global_cwq *gcwq = pool->gcwq;
1784        const char *pri = worker_pool_pri(pool) ? "H" : "";
1785        struct worker *worker = NULL;
1786        int id = -1;
1787
1788        spin_lock_irq(&gcwq->lock);
1789        while (ida_get_new(&pool->worker_ida, &id)) {
1790                spin_unlock_irq(&gcwq->lock);
1791                if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
1792                        goto fail;
1793                spin_lock_irq(&gcwq->lock);
1794        }
1795        spin_unlock_irq(&gcwq->lock);
1796
1797        worker = alloc_worker();
1798        if (!worker)
1799                goto fail;
1800
1801        worker->pool = pool;
1802        worker->id = id;
1803
1804        if (gcwq->cpu != WORK_CPU_UNBOUND)
1805                worker->task = kthread_create_on_node(worker_thread,
1806                                        worker, cpu_to_node(gcwq->cpu),
1807                                        "kworker/%u:%d%s", gcwq->cpu, id, pri);
1808        else
1809                worker->task = kthread_create(worker_thread, worker,
1810                                              "kworker/u:%d%s", id, pri);
1811        if (IS_ERR(worker->task))
1812                goto fail;
1813
1814        if (worker_pool_pri(pool))
1815                set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1816
1817        /*
1818         * Determine CPU binding of the new worker depending on
1819         * %GCWQ_DISASSOCIATED.  The caller is responsible for ensuring the
1820         * flag remains stable across this function.  See the comments
1821         * above the flag definition for details.
1822         *
1823         * As an unbound worker may later become a regular one if CPU comes
1824         * online, make sure every worker has %PF_THREAD_BOUND set.
1825         */
1826        if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
1827                kthread_bind(worker->task, gcwq->cpu);
1828        } else {
1829                worker->task->flags |= PF_THREAD_BOUND;
1830                worker->flags |= WORKER_UNBOUND;
1831        }
1832
1833        return worker;
1834fail:
1835        if (id >= 0) {
1836                spin_lock_irq(&gcwq->lock);
1837                ida_remove(&pool->worker_ida, id);
1838                spin_unlock_irq(&gcwq->lock);
1839        }
1840        kfree(worker);
1841        return NULL;
1842}
1843
1844/**
1845 * start_worker - start a newly created worker
1846 * @worker: worker to start
1847 *
1848 * Make the gcwq aware of @worker and start it.
1849 *
1850 * CONTEXT:
1851 * spin_lock_irq(gcwq->lock).
1852 */
1853static void start_worker(struct worker *worker)
1854{
1855        worker->flags |= WORKER_STARTED;
1856        worker->pool->nr_workers++;
1857        worker_enter_idle(worker);
1858        wake_up_process(worker->task);
1859}
1860
1861/**
1862 * destroy_worker - destroy a workqueue worker
1863 * @worker: worker to be destroyed
1864 *
1865 * Destroy @worker and adjust @gcwq stats accordingly.
1866 *
1867 * CONTEXT:
1868 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1869 */
1870static void destroy_worker(struct worker *worker)
1871{
1872        struct worker_pool *pool = worker->pool;
1873        struct global_cwq *gcwq = pool->gcwq;
1874        int id = worker->id;
1875
1876        /* sanity check frenzy */
1877        BUG_ON(worker->current_work);
1878        BUG_ON(!list_empty(&worker->scheduled));
1879
1880        if (worker->flags & WORKER_STARTED)
1881                pool->nr_workers--;
1882        if (worker->flags & WORKER_IDLE)
1883                pool->nr_idle--;
1884
1885        list_del_init(&worker->entry);
1886        worker->flags |= WORKER_DIE;
1887
1888        spin_unlock_irq(&gcwq->lock);
1889
1890        kthread_stop(worker->task);
1891        kfree(worker);
1892
1893        spin_lock_irq(&gcwq->lock);
1894        ida_remove(&pool->worker_ida, id);
1895}
1896
1897static void idle_worker_timeout(unsigned long __pool)
1898{
1899        struct worker_pool *pool = (void *)__pool;
1900        struct global_cwq *gcwq = pool->gcwq;
1901
1902        spin_lock_irq(&gcwq->lock);
1903
1904        if (too_many_workers(pool)) {
1905                struct worker *worker;
1906                unsigned long expires;
1907
1908                /* idle_list is kept in LIFO order, check the last one */
1909                worker = list_entry(pool->idle_list.prev, struct worker, entry);
1910                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1911
1912                if (time_before(jiffies, expires))
1913                        mod_timer(&pool->idle_timer, expires);
1914                else {
1915                        /* it's been idle for too long, wake up manager */
1916                        pool->flags |= POOL_MANAGE_WORKERS;
1917                        wake_up_worker(pool);
1918                }
1919        }
1920
1921        spin_unlock_irq(&gcwq->lock);
1922}
1923
1924static bool send_mayday(struct work_struct *work)
1925{
1926        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1927        struct workqueue_struct *wq = cwq->wq;
1928        unsigned int cpu;
1929
1930        if (!(wq->flags & WQ_RESCUER))
1931                return false;
1932
1933        /* mayday mayday mayday */
1934        cpu = cwq->pool->gcwq->cpu;
1935        /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1936        if (cpu == WORK_CPU_UNBOUND)
1937                cpu = 0;
1938        if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
1939                wake_up_process(wq->rescuer->task);
1940        return true;
1941}
1942
1943static void gcwq_mayday_timeout(unsigned long __pool)
1944{
1945        struct worker_pool *pool = (void *)__pool;
1946        struct global_cwq *gcwq = pool->gcwq;
1947        struct work_struct *work;
1948
1949        spin_lock_irq(&gcwq->lock);
1950
1951        if (need_to_create_worker(pool)) {
1952                /*
1953                 * We've been trying to create a new worker but
1954                 * haven't been successful.  We might be hitting an
1955                 * allocation deadlock.  Send distress signals to
1956                 * rescuers.
1957                 */
1958                list_for_each_entry(work, &pool->worklist, entry)
1959                        send_mayday(work);
1960        }
1961
1962        spin_unlock_irq(&gcwq->lock);
1963
1964        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1965}
1966
1967/**
1968 * maybe_create_worker - create a new worker if necessary
1969 * @pool: pool to create a new worker for
1970 *
1971 * Create a new worker for @pool if necessary.  @pool is guaranteed to
1972 * have at least one idle worker on return from this function.  If
1973 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1974 * sent to all rescuers with works scheduled on @pool to resolve
1975 * possible allocation deadlock.
1976 *
1977 * On return, need_to_create_worker() is guaranteed to be false and
1978 * may_start_working() true.
1979 *
1980 * LOCKING:
1981 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1982 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1983 * manager.
1984 *
1985 * RETURNS:
1986 * false if no action was taken and gcwq->lock stayed locked, true
1987 * otherwise.
1988 */
1989static bool maybe_create_worker(struct worker_pool *pool)
1990__releases(&gcwq->lock)
1991__acquires(&gcwq->lock)
1992{
1993        struct global_cwq *gcwq = pool->gcwq;
1994
1995        if (!need_to_create_worker(pool))
1996                return false;
1997restart:
1998        spin_unlock_irq(&gcwq->lock);
1999
2000        /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2001        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2002
2003        while (true) {
2004                struct worker *worker;
2005
2006                worker = create_worker(pool);
2007                if (worker) {
2008                        del_timer_sync(&pool->mayday_timer);
2009                        spin_lock_irq(&gcwq->lock);
2010                        start_worker(worker);
2011                        BUG_ON(need_to_create_worker(pool));
2012                        return true;
2013                }
2014
2015                if (!need_to_create_worker(pool))
2016                        break;
2017
2018                __set_current_state(TASK_INTERRUPTIBLE);
2019                schedule_timeout(CREATE_COOLDOWN);
2020
2021                if (!need_to_create_worker(pool))
2022                        break;
2023        }
2024
2025        del_timer_sync(&pool->mayday_timer);
2026        spin_lock_irq(&gcwq->lock);
2027        if (need_to_create_worker(pool))
2028                goto restart;
2029        return true;
2030}
2031
2032/**
2033 * maybe_destroy_worker - destroy workers which have been idle for a while
2034 * @pool: pool to destroy workers for
2035 *
2036 * Destroy @pool workers which have been idle for longer than
2037 * IDLE_WORKER_TIMEOUT.
2038 *
2039 * LOCKING:
2040 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2041 * multiple times.  Called only from manager.
2042 *
2043 * RETURNS:
2044 * false if no action was taken and gcwq->lock stayed locked, true
2045 * otherwise.
2046 */
2047static bool maybe_destroy_workers(struct worker_pool *pool)
2048{
2049        bool ret = false;
2050
2051        while (too_many_workers(pool)) {
2052                struct worker *worker;
2053                unsigned long expires;
2054
2055                worker = list_entry(pool->idle_list.prev, struct worker, entry);
2056                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2057
2058                if (time_before(jiffies, expires)) {
2059                        mod_timer(&pool->idle_timer, expires);
2060                        break;
2061                }
2062
2063                destroy_worker(worker);
2064                ret = true;
2065        }
2066
2067        return ret;
2068}
2069
2070/**
2071 * manage_workers - manage worker pool
2072 * @worker: self
2073 *
2074 * Assume the manager role and manage gcwq worker pool @worker belongs
2075 * to.  At any given time, there can be only zero or one manager per
2076 * gcwq.  The exclusion is handled automatically by this function.
2077 *
2078 * The caller can safely start processing works on false return.  On
2079 * true return, it's guaranteed that need_to_create_worker() is false
2080 * and may_start_working() is true.
2081 *
2082 * CONTEXT:
2083 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2084 * multiple times.  Does GFP_KERNEL allocations.
2085 *
2086 * RETURNS:
2087 * false if no action was taken and gcwq->lock stayed locked, true if
2088 * some action was taken.
2089 */
2090static bool manage_workers(struct worker *worker)
2091{
2092        struct worker_pool *pool = worker->pool;
2093        bool ret = false;
2094
2095        if (pool->flags & POOL_MANAGING_WORKERS)
2096                return ret;
2097
2098        pool->flags |= POOL_MANAGING_WORKERS;
2099
2100        /*
2101         * To simplify both worker management and CPU hotplug, hold off
2102         * management while hotplug is in progress.  CPU hotplug path can't
2103         * grab %POOL_MANAGING_WORKERS to achieve this because that can
2104         * lead to idle worker depletion (all become busy thinking someone
2105         * else is managing) which in turn can result in deadlock under
2106         * extreme circumstances.  Use @pool->assoc_mutex to synchronize
2107         * manager against CPU hotplug.
2108         *
2109         * assoc_mutex would always be free unless CPU hotplug is in
2110         * progress.  trylock first without dropping @gcwq->lock.
2111         */
2112        if (unlikely(!mutex_trylock(&pool->assoc_mutex))) {
2113                spin_unlock_irq(&pool->gcwq->lock);
2114                mutex_lock(&pool->assoc_mutex);
2115                /*
2116                 * CPU hotplug could have happened while we were waiting
2117                 * for assoc_mutex.  Hotplug itself can't handle us
2118                 * because manager isn't either on idle or busy list, and
2119                 * @gcwq's state and ours could have deviated.
2120                 *
2121                 * As hotplug is now excluded via assoc_mutex, we can
2122                 * simply try to bind.  It will succeed or fail depending
2123                 * on @gcwq's current state.  Try it and adjust
2124                 * %WORKER_UNBOUND accordingly.
2125                 */
2126                if (worker_maybe_bind_and_lock(worker))
2127                        worker->flags &= ~WORKER_UNBOUND;
2128                else
2129                        worker->flags |= WORKER_UNBOUND;
2130
2131                ret = true;
2132        }
2133
2134        pool->flags &= ~POOL_MANAGE_WORKERS;
2135
2136        /*
2137         * Destroy and then create so that may_start_working() is true
2138         * on return.
2139         */
2140        ret |= maybe_destroy_workers(pool);
2141        ret |= maybe_create_worker(pool);
2142
2143        pool->flags &= ~POOL_MANAGING_WORKERS;
2144        mutex_unlock(&pool->assoc_mutex);
2145        return ret;
2146}
2147
2148/**
2149 * process_one_work - process single work
2150 * @worker: self
2151 * @work: work to process
2152 *
2153 * Process @work.  This function contains all the logics necessary to
2154 * process a single work including synchronization against and
2155 * interaction with other workers on the same cpu, queueing and
2156 * flushing.  As long as context requirement is met, any worker can
2157 * call this function to process a work.
2158 *
2159 * CONTEXT:
2160 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
2161 */
2162static void process_one_work(struct worker *worker, struct work_struct *work)
2163__releases(&gcwq->lock)
2164__acquires(&gcwq->lock)
2165{
2166        struct cpu_workqueue_struct *cwq = get_work_cwq(work);
2167        struct worker_pool *pool = worker->pool;
2168        struct global_cwq *gcwq = pool->gcwq;
2169        struct hlist_head *bwh = busy_worker_head(gcwq, work);
2170        bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
2171        work_func_t f = work->func;
2172        int work_color;
2173        struct worker *collision;
2174#ifdef CONFIG_LOCKDEP
2175        /*
2176         * It is permissible to free the struct work_struct from
2177         * inside the function that is called from it, this we need to
2178         * take into account for lockdep too.  To avoid bogus "held
2179         * lock freed" warnings as well as problems when looking into
2180         * work->lockdep_map, make a copy and use that here.
2181         */
2182        struct lockdep_map lockdep_map;
2183
2184        lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2185#endif
2186        /*
2187         * Ensure we're on the correct CPU.  DISASSOCIATED test is
2188         * necessary to avoid spurious warnings from rescuers servicing the
2189         * unbound or a disassociated gcwq.
2190         */
2191        WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2192                     !(gcwq->flags & GCWQ_DISASSOCIATED) &&
2193                     raw_smp_processor_id() != gcwq->cpu);
2194
2195        /*
2196         * A single work shouldn't be executed concurrently by
2197         * multiple workers on a single cpu.  Check whether anyone is
2198         * already processing the work.  If so, defer the work to the
2199         * currently executing one.
2200         */
2201        collision = __find_worker_executing_work(gcwq, bwh, work);
2202        if (unlikely(collision)) {
2203                move_linked_works(work, &collision->scheduled, NULL);
2204                return;
2205        }
2206
2207        /* claim and dequeue */
2208        debug_work_deactivate(work);
2209        hlist_add_head(&worker->hentry, bwh);
2210        worker->current_work = work;
2211        worker->current_cwq = cwq;
2212        work_color = get_work_color(work);
2213
2214        list_del_init(&work->entry);
2215
2216        /*
2217         * CPU intensive works don't participate in concurrency
2218         * management.  They're the scheduler's responsibility.
2219         */
2220        if (unlikely(cpu_intensive))
2221                worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2222
2223        /*
2224         * Unbound gcwq isn't concurrency managed and work items should be
2225         * executed ASAP.  Wake up another worker if necessary.
2226         */
2227        if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2228                wake_up_worker(pool);
2229
2230        /*
2231         * Record the last CPU and clear PENDING which should be the last
2232         * update to @work.  Also, do this inside @gcwq->lock so that
2233         * PENDING and queued state changes happen together while IRQ is
2234         * disabled.
2235         */
2236        set_work_cpu_and_clear_pending(work, gcwq->cpu);
2237
2238        spin_unlock_irq(&gcwq->lock);
2239
2240        lock_map_acquire_read(&cwq->wq->lockdep_map);
2241        lock_map_acquire(&lockdep_map);
2242        trace_workqueue_execute_start(work);
2243        f(work);
2244        /*
2245         * While we must be careful to not use "work" after this, the trace
2246         * point will only record its address.
2247         */
2248        trace_workqueue_execute_end(work);
2249        lock_map_release(&lockdep_map);
2250        lock_map_release(&cwq->wq->lockdep_map);
2251
2252        if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2253                pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2254                       "     last function: %pf\n",
2255                       current->comm, preempt_count(), task_pid_nr(current), f);
2256                debug_show_held_locks(current);
2257                dump_stack();
2258        }
2259
2260        spin_lock_irq(&gcwq->lock);
2261
2262        /* clear cpu intensive status */
2263        if (unlikely(cpu_intensive))
2264                worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2265
2266        /* we're done with it, release */
2267        hlist_del_init(&worker->hentry);
2268        worker->current_work = NULL;
2269        worker->current_cwq = NULL;
2270        cwq_dec_nr_in_flight(cwq, work_color);
2271}
2272
2273/**
2274 * process_scheduled_works - process scheduled works
2275 * @worker: self
2276 *
2277 * Process all scheduled works.  Please note that the scheduled list
2278 * may change while processing a work, so this function repeatedly
2279 * fetches a work from the top and executes it.
2280 *
2281 * CONTEXT:
2282 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
2283 * multiple times.
2284 */
2285static void process_scheduled_works(struct worker *worker)
2286{
2287        while (!list_empty(&worker->scheduled)) {
2288                struct work_struct *work = list_first_entry(&worker->scheduled,
2289                                                struct work_struct, entry);
2290                process_one_work(worker, work);
2291        }
2292}
2293
2294/**
2295 * worker_thread - the worker thread function
2296 * @__worker: self
2297 *
2298 * The gcwq worker thread function.  There's a single dynamic pool of
2299 * these per each cpu.  These workers process all works regardless of
2300 * their specific target workqueue.  The only exception is works which
2301 * belong to workqueues with a rescuer which will be explained in
2302 * rescuer_thread().
2303 */
2304static int worker_thread(void *__worker)
2305{
2306        struct worker *worker = __worker;
2307        struct worker_pool *pool = worker->pool;
2308        struct global_cwq *gcwq = pool->gcwq;
2309
2310        /* tell the scheduler that this is a workqueue worker */
2311        worker->task->flags |= PF_WQ_WORKER;
2312woke_up:
2313        spin_lock_irq(&gcwq->lock);
2314
2315        /* we are off idle list if destruction or rebind is requested */
2316        if (unlikely(list_empty(&worker->entry))) {
2317                spin_unlock_irq(&gcwq->lock);
2318
2319                /* if DIE is set, destruction is requested */
2320                if (worker->flags & WORKER_DIE) {
2321                        worker->task->flags &= ~PF_WQ_WORKER;
2322                        return 0;
2323                }
2324
2325                /* otherwise, rebind */
2326                idle_worker_rebind(worker);
2327                goto woke_up;
2328        }
2329
2330        worker_leave_idle(worker);
2331recheck:
2332        /* no more worker necessary? */
2333        if (!need_more_worker(pool))
2334                goto sleep;
2335
2336        /* do we need to manage? */
2337        if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2338                goto recheck;
2339
2340        /*
2341         * ->scheduled list can only be filled while a worker is
2342         * preparing to process a work or actually processing it.
2343         * Make sure nobody diddled with it while I was sleeping.
2344         */
2345        BUG_ON(!list_empty(&worker->scheduled));
2346
2347        /*
2348         * When control reaches this point, we're guaranteed to have
2349         * at least one idle worker or that someone else has already
2350         * assumed the manager role.
2351         */
2352        worker_clr_flags(worker, WORKER_PREP);
2353
2354        do {
2355                struct work_struct *work =
2356                        list_first_entry(&pool->worklist,
2357                                         struct work_struct, entry);
2358
2359                if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2360                        /* optimization path, not strictly necessary */
2361                        process_one_work(worker, work);
2362                        if (unlikely(!list_empty(&worker->scheduled)))
2363                                process_scheduled_works(worker);
2364                } else {
2365                        move_linked_works(work, &worker->scheduled, NULL);
2366                        process_scheduled_works(worker);
2367                }
2368        } while (keep_working(pool));
2369
2370        worker_set_flags(worker, WORKER_PREP, false);
2371sleep:
2372        if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2373                goto recheck;
2374
2375        /*
2376         * gcwq->lock is held and there's no work to process and no
2377         * need to manage, sleep.  Workers are woken up only while
2378         * holding gcwq->lock or from local cpu, so setting the
2379         * current state before releasing gcwq->lock is enough to
2380         * prevent losing any event.
2381         */
2382        worker_enter_idle(worker);
2383        __set_current_state(TASK_INTERRUPTIBLE);
2384        spin_unlock_irq(&gcwq->lock);
2385        schedule();
2386        goto woke_up;
2387}
2388
2389/**
2390 * rescuer_thread - the rescuer thread function
2391 * @__wq: the associated workqueue
2392 *
2393 * Workqueue rescuer thread function.  There's one rescuer for each
2394 * workqueue which has WQ_RESCUER set.
2395 *
2396 * Regular work processing on a gcwq may block trying to create a new
2397 * worker which uses GFP_KERNEL allocation which has slight chance of
2398 * developing into deadlock if some works currently on the same queue
2399 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2400 * the problem rescuer solves.
2401 *
2402 * When such condition is possible, the gcwq summons rescuers of all
2403 * workqueues which have works queued on the gcwq and let them process
2404 * those works so that forward progress can be guaranteed.
2405 *
2406 * This should happen rarely.
2407 */
2408static int rescuer_thread(void *__wq)
2409{
2410        struct workqueue_struct *wq = __wq;
2411        struct worker *rescuer = wq->rescuer;
2412        struct list_head *scheduled = &rescuer->scheduled;
2413        bool is_unbound = wq->flags & WQ_UNBOUND;
2414        unsigned int cpu;
2415
2416        set_user_nice(current, RESCUER_NICE_LEVEL);
2417repeat:
2418        set_current_state(TASK_INTERRUPTIBLE);
2419
2420        if (kthread_should_stop()) {
2421                __set_current_state(TASK_RUNNING);
2422                return 0;
2423        }
2424
2425        /*
2426         * See whether any cpu is asking for help.  Unbounded
2427         * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2428         */
2429        for_each_mayday_cpu(cpu, wq->mayday_mask) {
2430                unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2431                struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2432                struct worker_pool *pool = cwq->pool;
2433                struct global_cwq *gcwq = pool->gcwq;
2434                struct work_struct *work, *n;
2435
2436                __set_current_state(TASK_RUNNING);
2437                mayday_clear_cpu(cpu, wq->mayday_mask);
2438
2439                /* migrate to the target cpu if possible */
2440                rescuer->pool = pool;
2441                worker_maybe_bind_and_lock(rescuer);
2442
2443                /*
2444                 * Slurp in all works issued via this workqueue and
2445                 * process'em.
2446                 */
2447                BUG_ON(!list_empty(&rescuer->scheduled));
2448                list_for_each_entry_safe(work, n, &pool->worklist, entry)
2449                        if (get_work_cwq(work) == cwq)
2450                                move_linked_works(work, scheduled, &n);
2451
2452                process_scheduled_works(rescuer);
2453
2454                /*
2455                 * Leave this gcwq.  If keep_working() is %true, notify a
2456                 * regular worker; otherwise, we end up with 0 concurrency
2457                 * and stalling the execution.
2458                 */
2459                if (keep_working(pool))
2460                        wake_up_worker(pool);
2461
2462                spin_unlock_irq(&gcwq->lock);
2463        }
2464
2465        schedule();
2466        goto repeat;
2467}
2468
2469struct wq_barrier {
2470        struct work_struct      work;
2471        struct completion       done;
2472};
2473
2474static void wq_barrier_func(struct work_struct *work)
2475{
2476        struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2477        complete(&barr->done);
2478}
2479
2480/**
2481 * insert_wq_barrier - insert a barrier work
2482 * @cwq: cwq to insert barrier into
2483 * @barr: wq_barrier to insert
2484 * @target: target work to attach @barr to
2485 * @worker: worker currently executing @target, NULL if @target is not executing
2486 *
2487 * @barr is linked to @target such that @barr is completed only after
2488 * @target finishes execution.  Please note that the ordering
2489 * guarantee is observed only with respect to @target and on the local
2490 * cpu.
2491 *
2492 * Currently, a queued barrier can't be canceled.  This is because
2493 * try_to_grab_pending() can't determine whether the work to be
2494 * grabbed is at the head of the queue and thus can't clear LINKED
2495 * flag of the previous work while there must be a valid next work
2496 * after a work with LINKED flag set.
2497 *
2498 * Note that when @worker is non-NULL, @target may be modified
2499 * underneath us, so we can't reliably determine cwq from @target.
2500 *
2501 * CONTEXT:
2502 * spin_lock_irq(gcwq->lock).
2503 */
2504static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2505                              struct wq_barrier *barr,
2506                              struct work_struct *target, struct worker *worker)
2507{
2508        struct list_head *head;
2509        unsigned int linked = 0;
2510
2511        /*
2512         * debugobject calls are safe here even with gcwq->lock locked
2513         * as we know for sure that this will not trigger any of the
2514         * checks and call back into the fixup functions where we
2515         * might deadlock.
2516         */
2517        INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2518        __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2519        init_completion(&barr->done);
2520
2521        /*
2522         * If @target is currently being executed, schedule the
2523         * barrier to the worker; otherwise, put it after @target.
2524         */
2525        if (worker)
2526                head = worker->scheduled.next;
2527        else {
2528                unsigned long *bits = work_data_bits(target);
2529
2530                head = target->entry.next;
2531                /* there can already be other linked works, inherit and set */
2532                linked = *bits & WORK_STRUCT_LINKED;
2533                __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2534        }
2535
2536        debug_work_activate(&barr->work);
2537        insert_work(cwq, &barr->work, head,
2538                    work_color_to_flags(WORK_NO_COLOR) | linked);
2539}
2540
2541/**
2542 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2543 * @wq: workqueue being flushed
2544 * @flush_color: new flush color, < 0 for no-op
2545 * @work_color: new work color, < 0 for no-op
2546 *
2547 * Prepare cwqs for workqueue flushing.
2548 *
2549 * If @flush_color is non-negative, flush_color on all cwqs should be
2550 * -1.  If no cwq has in-flight commands at the specified color, all
2551 * cwq->flush_color's stay at -1 and %false is returned.  If any cwq
2552 * has in flight commands, its cwq->flush_color is set to
2553 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2554 * wakeup logic is armed and %true is returned.
2555 *
2556 * The caller should have initialized @wq->first_flusher prior to
2557 * calling this function with non-negative @flush_color.  If
2558 * @flush_color is negative, no flush color update is done and %false
2559 * is returned.
2560 *
2561 * If @work_color is non-negative, all cwqs should have the same
2562 * work_color which is previous to @work_color and all will be
2563 * advanced to @work_color.
2564 *
2565 * CONTEXT:
2566 * mutex_lock(wq->flush_mutex).
2567 *
2568 * RETURNS:
2569 * %true if @flush_color >= 0 and there's something to flush.  %false
2570 * otherwise.
2571 */
2572static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2573                                      int flush_color, int work_color)
2574{
2575        bool wait = false;
2576        unsigned int cpu;
2577
2578        if (flush_color >= 0) {
2579                BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2580                atomic_set(&wq->nr_cwqs_to_flush, 1);
2581        }
2582
2583        for_each_cwq_cpu(cpu, wq) {
2584                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2585                struct global_cwq *gcwq = cwq->pool->gcwq;
2586
2587                spin_lock_irq(&gcwq->lock);
2588
2589                if (flush_color >= 0) {
2590                        BUG_ON(cwq->flush_color != -1);
2591
2592                        if (cwq->nr_in_flight[flush_color]) {
2593                                cwq->flush_color = flush_color;
2594                                atomic_inc(&wq->nr_cwqs_to_flush);
2595                                wait = true;
2596                        }
2597                }
2598
2599                if (work_color >= 0) {
2600                        BUG_ON(work_color != work_next_color(cwq->work_color));
2601                        cwq->work_color = work_color;
2602                }
2603
2604                spin_unlock_irq(&gcwq->lock);
2605        }
2606
2607        if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2608                complete(&wq->first_flusher->done);
2609
2610        return wait;
2611}
2612
2613/**
2614 * flush_workqueue - ensure that any scheduled work has run to completion.
2615 * @wq: workqueue to flush
2616 *
2617 * Forces execution of the workqueue and blocks until its completion.
2618 * This is typically used in driver shutdown handlers.
2619 *
2620 * We sleep until all works which were queued on entry have been handled,
2621 * but we are not livelocked by new incoming ones.
2622 */
2623void flush_workqueue(struct workqueue_struct *wq)
2624{
2625        struct wq_flusher this_flusher = {
2626                .list = LIST_HEAD_INIT(this_flusher.list),
2627                .flush_color = -1,
2628                .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2629        };
2630        int next_color;
2631
2632        lock_map_acquire(&wq->lockdep_map);
2633        lock_map_release(&wq->lockdep_map);
2634
2635        mutex_lock(&wq->flush_mutex);
2636
2637        /*
2638         * Start-to-wait phase
2639         */
2640        next_color = work_next_color(wq->work_color);
2641
2642        if (next_color != wq->flush_color) {
2643                /*
2644                 * Color space is not full.  The current work_color
2645                 * becomes our flush_color and work_color is advanced
2646                 * by one.
2647                 */
2648                BUG_ON(!list_empty(&wq->flusher_overflow));
2649                this_flusher.flush_color = wq->work_color;
2650                wq->work_color = next_color;
2651
2652                if (!wq->first_flusher) {
2653                        /* no flush in progress, become the first flusher */
2654                        BUG_ON(wq->flush_color != this_flusher.flush_color);
2655
2656                        wq->first_flusher = &this_flusher;
2657
2658                        if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2659                                                       wq->work_color)) {
2660                                /* nothing to flush, done */
2661                                wq->flush_color = next_color;
2662                                wq->first_flusher = NULL;
2663                                goto out_unlock;
2664                        }
2665                } else {
2666                        /* wait in queue */
2667                        BUG_ON(wq->flush_color == this_flusher.flush_color);
2668                        list_add_tail(&this_flusher.list, &wq->flusher_queue);
2669                        flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2670                }
2671        } else {
2672                /*
2673                 * Oops, color space is full, wait on overflow queue.
2674                 * The next flush completion will assign us
2675                 * flush_color and transfer to flusher_queue.
2676                 */
2677                list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2678        }
2679
2680        mutex_unlock(&wq->flush_mutex);
2681
2682        wait_for_completion(&this_flusher.done);
2683
2684        /*
2685         * Wake-up-and-cascade phase
2686         *
2687         * First flushers are responsible for cascading flushes and
2688         * handling overflow.  Non-first flushers can simply return.
2689         */
2690        if (wq->first_flusher != &this_flusher)
2691                return;
2692
2693        mutex_lock(&wq->flush_mutex);
2694
2695        /* we might have raced, check again with mutex held */
2696        if (wq->first_flusher != &this_flusher)
2697                goto out_unlock;
2698
2699        wq->first_flusher = NULL;
2700
2701        BUG_ON(!list_empty(&this_flusher.list));
2702        BUG_ON(wq->flush_color != this_flusher.flush_color);
2703
2704        while (true) {
2705                struct wq_flusher *next, *tmp;
2706
2707                /* complete all the flushers sharing the current flush color */
2708                list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2709                        if (next->flush_color != wq->flush_color)
2710                                break;
2711                        list_del_init(&next->list);
2712                        complete(&next->done);
2713                }
2714
2715                BUG_ON(!list_empty(&wq->flusher_overflow) &&
2716                       wq->flush_color != work_next_color(wq->work_color));
2717
2718                /* this flush_color is finished, advance by one */
2719                wq->flush_color = work_next_color(wq->flush_color);
2720
2721                /* one color has been freed, handle overflow queue */
2722                if (!list_empty(&wq->flusher_overflow)) {
2723                        /*
2724                         * Assign the same color to all overflowed
2725                         * flushers, advance work_color and append to
2726                         * flusher_queue.  This is the start-to-wait
2727                         * phase for these overflowed flushers.
2728                         */
2729                        list_for_each_entry(tmp, &wq->flusher_overflow, list)
2730                                tmp->flush_color = wq->work_color;
2731
2732                        wq->work_color = work_next_color(wq->work_color);
2733
2734                        list_splice_tail_init(&wq->flusher_overflow,
2735                                              &wq->flusher_queue);
2736                        flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2737                }
2738
2739                if (list_empty(&wq->flusher_queue)) {
2740                        BUG_ON(wq->flush_color != wq->work_color);
2741                        break;
2742                }
2743
2744                /*
2745                 * Need to flush more colors.  Make the next flusher
2746                 * the new first flusher and arm cwqs.
2747                 */
2748                BUG_ON(wq->flush_color == wq->work_color);
2749                BUG_ON(wq->flush_color != next->flush_color);
2750
2751                list_del_init(&next->list);
2752                wq->first_flusher = next;
2753
2754                if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2755                        break;
2756
2757                /*
2758                 * Meh... this color is already done, clear first
2759                 * flusher and repeat cascading.
2760                 */
2761                wq->first_flusher = NULL;
2762        }
2763
2764out_unlock:
2765        mutex_unlock(&wq->flush_mutex);
2766}
2767EXPORT_SYMBOL_GPL(flush_workqueue);
2768
2769/**
2770 * drain_workqueue - drain a workqueue
2771 * @wq: workqueue to drain
2772 *
2773 * Wait until the workqueue becomes empty.  While draining is in progress,
2774 * only chain queueing is allowed.  IOW, only currently pending or running
2775 * work items on @wq can queue further work items on it.  @wq is flushed
2776 * repeatedly until it becomes empty.  The number of flushing is detemined
2777 * by the depth of chaining and should be relatively short.  Whine if it
2778 * takes too long.
2779 */
2780void drain_workqueue(struct workqueue_struct *wq)
2781{
2782        unsigned int flush_cnt = 0;
2783        unsigned int cpu;
2784
2785        /*
2786         * __queue_work() needs to test whether there are drainers, is much
2787         * hotter than drain_workqueue() and already looks at @wq->flags.
2788         * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2789         */
2790        spin_lock(&workqueue_lock);
2791        if (!wq->nr_drainers++)
2792                wq->flags |= WQ_DRAINING;
2793        spin_unlock(&workqueue_lock);
2794reflush:
2795        flush_workqueue(wq);
2796
2797        for_each_cwq_cpu(cpu, wq) {
2798                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2799                bool drained;
2800
2801                spin_lock_irq(&cwq->pool->gcwq->lock);
2802                drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2803                spin_unlock_irq(&cwq->pool->gcwq->lock);
2804
2805                if (drained)
2806                        continue;
2807
2808                if (++flush_cnt == 10 ||
2809                    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2810                        pr_warn("workqueue %s: flush on destruction isn't complete after %u tries\n",
2811                                wq->name, flush_cnt);
2812                goto reflush;
2813        }
2814
2815        spin_lock(&workqueue_lock);
2816        if (!--wq->nr_drainers)
2817                wq->flags &= ~WQ_DRAINING;
2818        spin_unlock(&workqueue_lock);
2819}
2820EXPORT_SYMBOL_GPL(drain_workqueue);
2821
2822static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2823{
2824        struct worker *worker = NULL;
2825        struct global_cwq *gcwq;
2826        struct cpu_workqueue_struct *cwq;
2827
2828        might_sleep();
2829        gcwq = get_work_gcwq(work);
2830        if (!gcwq)
2831                return false;
2832
2833        spin_lock_irq(&gcwq->lock);
2834        if (!list_empty(&work->entry)) {
2835                /*
2836                 * See the comment near try_to_grab_pending()->smp_rmb().
2837                 * If it was re-queued to a different gcwq under us, we
2838                 * are not going to wait.
2839                 */
2840                smp_rmb();
2841                cwq = get_work_cwq(work);
2842                if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
2843                        goto already_gone;
2844        } else {
2845                worker = find_worker_executing_work(gcwq, work);
2846                if (!worker)
2847                        goto already_gone;
2848                cwq = worker->current_cwq;
2849        }
2850
2851        insert_wq_barrier(cwq, barr, work, worker);
2852        spin_unlock_irq(&gcwq->lock);
2853
2854        /*
2855         * If @max_active is 1 or rescuer is in use, flushing another work
2856         * item on the same workqueue may lead to deadlock.  Make sure the
2857         * flusher is not running on the same workqueue by verifying write
2858         * access.
2859         */
2860        if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2861                lock_map_acquire(&cwq->wq->lockdep_map);
2862        else
2863                lock_map_acquire_read(&cwq->wq->lockdep_map);
2864        lock_map_release(&cwq->wq->lockdep_map);
2865
2866        return true;
2867already_gone:
2868        spin_unlock_irq(&gcwq->lock);
2869        return false;
2870}
2871
2872/**
2873 * flush_work - wait for a work to finish executing the last queueing instance
2874 * @work: the work to flush
2875 *
2876 * Wait until @work has finished execution.  @work is guaranteed to be idle
2877 * on return if it hasn't been requeued since flush started.
2878 *
2879 * RETURNS:
2880 * %true if flush_work() waited for the work to finish execution,
2881 * %false if it was already idle.
2882 */
2883bool flush_work(struct work_struct *work)
2884{
2885        struct wq_barrier barr;
2886
2887        lock_map_acquire(&work->lockdep_map);
2888        lock_map_release(&work->lockdep_map);
2889
2890        if (start_flush_work(work, &barr)) {
2891                wait_for_completion(&barr.done);
2892                destroy_work_on_stack(&barr.work);
2893                return true;
2894        } else {
2895                return false;
2896        }
2897}
2898EXPORT_SYMBOL_GPL(flush_work);
2899
2900static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2901{
2902        unsigned long flags;
2903        int ret;
2904
2905        do {
2906                ret = try_to_grab_pending(work, is_dwork, &flags);
2907                /*
2908                 * If someone else is canceling, wait for the same event it
2909                 * would be waiting for before retrying.
2910                 */
2911                if (unlikely(ret == -ENOENT))
2912                        flush_work(work);
2913        } while (unlikely(ret < 0));
2914
2915        /* tell other tasks trying to grab @work to back off */
2916        mark_work_canceling(work);
2917        local_irq_restore(flags);
2918
2919        flush_work(work);
2920        clear_work_data(work);
2921        return ret;
2922}
2923
2924/**
2925 * cancel_work_sync - cancel a work and wait for it to finish
2926 * @work: the work to cancel
2927 *
2928 * Cancel @work and wait for its execution to finish.  This function
2929 * can be used even if the work re-queues itself or migrates to
2930 * another workqueue.  On return from this function, @work is
2931 * guaranteed to be not pending or executing on any CPU.
2932 *
2933 * cancel_work_sync(&delayed_work->work) must not be used for
2934 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2935 *
2936 * The caller must ensure that the workqueue on which @work was last
2937 * queued can't be destroyed before this function returns.
2938 *
2939 * RETURNS:
2940 * %true if @work was pending, %false otherwise.
2941 */
2942bool cancel_work_sync(struct work_struct *work)
2943{
2944        return __cancel_work_timer(work, false);
2945}
2946EXPORT_SYMBOL_GPL(cancel_work_sync);
2947
2948/**
2949 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2950 * @dwork: the delayed work to flush
2951 *
2952 * Delayed timer is cancelled and the pending work is queued for
2953 * immediate execution.  Like flush_work(), this function only
2954 * considers the last queueing instance of @dwork.
2955 *
2956 * RETURNS:
2957 * %true if flush_work() waited for the work to finish execution,
2958 * %false if it was already idle.
2959 */
2960bool flush_delayed_work(struct delayed_work *dwork)
2961{
2962        local_irq_disable();
2963        if (del_timer_sync(&dwork->timer))
2964                __queue_work(dwork->cpu,
2965                             get_work_cwq(&dwork->work)->wq, &dwork->work);
2966        local_irq_enable();
2967        return flush_work(&dwork->work);
2968}
2969EXPORT_SYMBOL(flush_delayed_work);
2970
2971/**
2972 * cancel_delayed_work - cancel a delayed work
2973 * @dwork: delayed_work to cancel
2974 *
2975 * Kill off a pending delayed_work.  Returns %true if @dwork was pending
2976 * and canceled; %false if wasn't pending.  Note that the work callback
2977 * function may still be running on return, unless it returns %true and the
2978 * work doesn't re-arm itself.  Explicitly flush or use
2979 * cancel_delayed_work_sync() to wait on it.
2980 *
2981 * This function is safe to call from any context including IRQ handler.
2982 */
2983bool cancel_delayed_work(struct delayed_work *dwork)
2984{
2985        unsigned long flags;
2986        int ret;
2987
2988        do {
2989                ret = try_to_grab_pending(&dwork->work, true, &flags);
2990        } while (unlikely(ret == -EAGAIN));
2991
2992        if (unlikely(ret < 0))
2993                return false;
2994
2995        set_work_cpu_and_clear_pending(&dwork->work, work_cpu(&dwork->work));
2996        local_irq_restore(flags);
2997        return ret;
2998}
2999EXPORT_SYMBOL(cancel_delayed_work);
3000
3001/**
3002 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3003 * @dwork: the delayed work cancel
3004 *
3005 * This is cancel_work_sync() for delayed works.
3006 *
3007 * RETURNS:
3008 * %true if @dwork was pending, %false otherwise.
3009 */
3010bool cancel_delayed_work_sync(struct delayed_work *dwork)
3011{
3012        return __cancel_work_timer(&dwork->work, true);
3013}
3014EXPORT_SYMBOL(cancel_delayed_work_sync);
3015
3016/**
3017 * schedule_work_on - put work task on a specific cpu
3018 * @cpu: cpu to put the work task on
3019 * @work: job to be done
3020 *
3021 * This puts a job on a specific cpu
3022 */
3023bool schedule_work_on(int cpu, struct work_struct *work)
3024{
3025        return queue_work_on(cpu, system_wq, work);
3026}
3027EXPORT_SYMBOL(schedule_work_on);
3028
3029/**
3030 * schedule_work - put work task in global workqueue
3031 * @work: job to be done
3032 *
3033 * Returns %false if @work was already on the kernel-global workqueue and
3034 * %true otherwise.
3035 *
3036 * This puts a job in the kernel-global workqueue if it was not already
3037 * queued and leaves it in the same position on the kernel-global
3038 * workqueue otherwise.
3039 */
3040bool schedule_work(struct work_struct *work)
3041{
3042        return queue_work(system_wq, work);
3043}
3044EXPORT_SYMBOL(schedule_work);
3045
3046/**
3047 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
3048 * @cpu: cpu to use
3049 * @dwork: job to be done
3050 * @delay: number of jiffies to wait
3051 *
3052 * After waiting for a given time this puts a job in the kernel-global
3053 * workqueue on the specified CPU.
3054 */
3055bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
3056                              unsigned long delay)
3057{
3058        return queue_delayed_work_on(cpu, system_wq, dwork, delay);
3059}
3060EXPORT_SYMBOL(schedule_delayed_work_on);
3061
3062/**
3063 * schedule_delayed_work - put work task in global workqueue after delay
3064 * @dwork: job to be done
3065 * @delay: number of jiffies to wait or 0 for immediate execution
3066 *
3067 * After waiting for a given time this puts a job in the kernel-global
3068 * workqueue.
3069 */
3070bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
3071{
3072        return queue_delayed_work(system_wq, dwork, delay);
3073}
3074EXPORT_SYMBOL(schedule_delayed_work);
3075
3076/**
3077 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3078 * @func: the function to call
3079 *
3080 * schedule_on_each_cpu() executes @func on each online CPU using the
3081 * system workqueue and blocks until all CPUs have completed.
3082 * schedule_on_each_cpu() is very slow.
3083 *
3084 * RETURNS:
3085 * 0 on success, -errno on failure.
3086 */
3087int schedule_on_each_cpu(work_func_t func)
3088{
3089        int cpu;
3090        struct work_struct __percpu *works;
3091
3092        works = alloc_percpu(struct work_struct);
3093        if (!works)
3094                return -ENOMEM;
3095
3096        get_online_cpus();
3097
3098        for_each_online_cpu(cpu) {
3099                struct work_struct *work = per_cpu_ptr(works, cpu);
3100
3101                INIT_WORK(work, func);
3102                schedule_work_on(cpu, work);
3103        }
3104
3105        for_each_online_cpu(cpu)
3106                flush_work(per_cpu_ptr(works, cpu));
3107
3108        put_online_cpus();
3109        free_percpu(works);
3110        return 0;
3111}
3112
3113/**
3114 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3115 *
3116 * Forces execution of the kernel-global workqueue and blocks until its
3117 * completion.
3118 *
3119 * Think twice before calling this function!  It's very easy to get into
3120 * trouble if you don't take great care.  Either of the following situations
3121 * will lead to deadlock:
3122 *
3123 *      One of the work items currently on the workqueue needs to acquire
3124 *      a lock held by your code or its caller.
3125 *
3126 *      Your code is running in the context of a work routine.
3127 *
3128 * They will be detected by lockdep when they occur, but the first might not
3129 * occur very often.  It depends on what work items are on the workqueue and
3130 * what locks they need, which you have no control over.
3131 *
3132 * In most situations flushing the entire workqueue is overkill; you merely
3133 * need to know that a particular work item isn't queued and isn't running.
3134 * In such cases you should use cancel_delayed_work_sync() or
3135 * cancel_work_sync() instead.
3136 */
3137void flush_scheduled_work(void)
3138{
3139        flush_workqueue(system_wq);
3140}
3141EXPORT_SYMBOL(flush_scheduled_work);
3142
3143/**
3144 * execute_in_process_context - reliably execute the routine with user context
3145 * @fn:         the function to execute
3146 * @ew:         guaranteed storage for the execute work structure (must
3147 *              be available when the work executes)
3148 *
3149 * Executes the function immediately if process context is available,
3150 * otherwise schedules the function for delayed execution.
3151 *
3152 * Returns:     0 - function was executed
3153 *              1 - function was scheduled for execution
3154 */
3155int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3156{
3157        if (!in_interrupt()) {
3158                fn(&ew->work);
3159                return 0;
3160        }
3161
3162        INIT_WORK(&ew->work, fn);
3163        schedule_work(&ew->work);
3164
3165        return 1;
3166}
3167EXPORT_SYMBOL_GPL(execute_in_process_context);
3168
3169int keventd_up(void)
3170{
3171        return system_wq != NULL;
3172}
3173
3174static int alloc_cwqs(struct workqueue_struct *wq)
3175{
3176        /*
3177         * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3178         * Make sure that the alignment isn't lower than that of
3179         * unsigned long long.
3180         */
3181        const size_t size = sizeof(struct cpu_workqueue_struct);
3182        const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3183                                   __alignof__(unsigned long long));
3184
3185        if (!(wq->flags & WQ_UNBOUND))
3186                wq->cpu_wq.pcpu = __alloc_percpu(size, align);
3187        else {
3188                void *ptr;
3189
3190                /*
3191                 * Allocate enough room to align cwq and put an extra
3192                 * pointer at the end pointing back to the originally
3193                 * allocated pointer which will be used for free.
3194                 */
3195                ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3196                if (ptr) {
3197                        wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3198                        *(void **)(wq->cpu_wq.single + 1) = ptr;
3199                }
3200        }
3201
3202        /* just in case, make sure it's actually aligned */
3203        BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3204        return wq->cpu_wq.v ? 0 : -ENOMEM;
3205}
3206
3207static void free_cwqs(struct workqueue_struct *wq)
3208{
3209        if (!(wq->flags & WQ_UNBOUND))
3210                free_percpu(wq->cpu_wq.pcpu);
3211        else if (wq->cpu_wq.single) {
3212                /* the pointer to free is stored right after the cwq */
3213                kfree(*(void **)(wq->cpu_wq.single + 1));
3214        }
3215}
3216
3217static int wq_clamp_max_active(int max_active, unsigned int flags,
3218                               const char *name)
3219{
3220        int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3221
3222        if (max_active < 1 || max_active > lim)
3223                pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3224                        max_active, name, 1, lim);
3225
3226        return clamp_val(max_active, 1, lim);
3227}
3228
3229struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3230                                               unsigned int flags,
3231                                               int max_active,
3232                                               struct lock_class_key *key,
3233                                               const char *lock_name, ...)
3234{
3235        va_list args, args1;
3236        struct workqueue_struct *wq;
3237        unsigned int cpu;
3238        size_t namelen;
3239
3240        /* determine namelen, allocate wq and format name */
3241        va_start(args, lock_name);
3242        va_copy(args1, args);
3243        namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3244
3245        wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3246        if (!wq)
3247                goto err;
3248
3249        vsnprintf(wq->name, namelen, fmt, args1);
3250        va_end(args);
3251        va_end(args1);
3252
3253        /*
3254         * Workqueues which may be used during memory reclaim should
3255         * have a rescuer to guarantee forward progress.
3256         */
3257        if (flags & WQ_MEM_RECLAIM)
3258                flags |= WQ_RESCUER;
3259
3260        max_active = max_active ?: WQ_DFL_ACTIVE;
3261        max_active = wq_clamp_max_active(max_active, flags, wq->name);
3262
3263        /* init wq */
3264        wq->flags = flags;
3265        wq->saved_max_active = max_active;
3266        mutex_init(&wq->flush_mutex);
3267        atomic_set(&wq->nr_cwqs_to_flush, 0);
3268        INIT_LIST_HEAD(&wq->flusher_queue);
3269        INIT_LIST_HEAD(&wq->flusher_overflow);
3270
3271        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3272        INIT_LIST_HEAD(&wq->list);
3273
3274        if (alloc_cwqs(wq) < 0)
3275                goto err;
3276
3277        for_each_cwq_cpu(cpu, wq) {
3278                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3279                struct global_cwq *gcwq = get_gcwq(cpu);
3280                int pool_idx = (bool)(flags & WQ_HIGHPRI);
3281
3282                BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3283                cwq->pool = &gcwq->pools[pool_idx];
3284                cwq->wq = wq;
3285                cwq->flush_color = -1;
3286                cwq->max_active = max_active;
3287                INIT_LIST_HEAD(&cwq->delayed_works);
3288        }
3289
3290        if (flags & WQ_RESCUER) {
3291                struct worker *rescuer;
3292
3293                if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3294                        goto err;
3295
3296                wq->rescuer = rescuer = alloc_worker();
3297                if (!rescuer)
3298                        goto err;
3299
3300                rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3301                                               wq->name);
3302                if (IS_ERR(rescuer->task))
3303                        goto err;
3304
3305                rescuer->task->flags |= PF_THREAD_BOUND;
3306                wake_up_process(rescuer->task);
3307        }
3308
3309        /*
3310         * workqueue_lock protects global freeze state and workqueues
3311         * list.  Grab it, set max_active accordingly and add the new
3312         * workqueue to workqueues list.
3313         */
3314        spin_lock(&workqueue_lock);
3315
3316        if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3317                for_each_cwq_cpu(cpu, wq)
3318                        get_cwq(cpu, wq)->max_active = 0;
3319
3320        list_add(&wq->list, &workqueues);
3321
3322        spin_unlock(&workqueue_lock);
3323
3324        return wq;
3325err:
3326        if (wq) {
3327                free_cwqs(wq);
3328                free_mayday_mask(wq->mayday_mask);
3329                kfree(wq->rescuer);
3330                kfree(wq);
3331        }
3332        return NULL;
3333}
3334EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3335
3336/**
3337 * destroy_workqueue - safely terminate a workqueue
3338 * @wq: target workqueue
3339 *
3340 * Safely destroy a workqueue. All work currently pending will be done first.
3341 */
3342void destroy_workqueue(struct workqueue_struct *wq)
3343{
3344        unsigned int cpu;
3345
3346        /* drain it before proceeding with destruction */
3347        drain_workqueue(wq);
3348
3349        /*
3350         * wq list is used to freeze wq, remove from list after
3351         * flushing is complete in case freeze races us.
3352         */
3353        spin_lock(&workqueue_lock);
3354        list_del(&wq->list);
3355        spin_unlock(&workqueue_lock);
3356
3357        /* sanity check */
3358        for_each_cwq_cpu(cpu, wq) {
3359                struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3360                int i;
3361
3362                for (i = 0; i < WORK_NR_COLORS; i++)
3363                        BUG_ON(cwq->nr_in_flight[i]);
3364                BUG_ON(cwq->nr_active);
3365                BUG_ON(!list_empty(&cwq->delayed_works));
3366        }
3367
3368        if (wq->flags & WQ_RESCUER) {
3369                kthread_stop(wq->rescuer->task);
3370                free_mayday_mask(wq->mayday_mask);
3371                kfree(wq->rescuer);
3372        }
3373
3374        free_cwqs(wq);
3375        kfree(wq);
3376}
3377EXPORT_SYMBOL_GPL(destroy_workqueue);
3378
3379/**
3380 * cwq_set_max_active - adjust max_active of a cwq
3381 * @cwq: target cpu_workqueue_struct
3382 * @max_active: new max_active value.
3383 *
3384 * Set @cwq->max_active to @max_active and activate delayed works if
3385 * increased.
3386 *
3387 * CONTEXT:
3388 * spin_lock_irq(gcwq->lock).
3389 */
3390static void cwq_set_max_active(struct cpu_workqueue_struct *cwq, int max_active)
3391{
3392        cwq->max_active = max_active;
3393
3394        while (!list_empty(&cwq->delayed_works) &&
3395               cwq->nr_active < cwq->max_active)
3396                cwq_activate_first_delayed(cwq);
3397}
3398
3399/**
3400 * workqueue_set_max_active - adjust max_active of a workqueue
3401 * @wq: target workqueue
3402 * @max_active: new max_active value.
3403 *
3404 * Set max_active of @wq to @max_active.
3405 *
3406 * CONTEXT:
3407 * Don't call from IRQ context.
3408 */
3409void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3410{
3411        unsigned int cpu;
3412
3413        max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3414
3415        spin_lock(&workqueue_lock);
3416
3417        wq->saved_max_active = max_active;
3418
3419        for_each_cwq_cpu(cpu, wq) {
3420                struct global_cwq *gcwq = get_gcwq(cpu);
3421
3422                spin_lock_irq(&gcwq->lock);
3423
3424                if (!(wq->flags & WQ_FREEZABLE) ||
3425                    !(gcwq->flags & GCWQ_FREEZING))
3426                        cwq_set_max_active(get_cwq(gcwq->cpu, wq), max_active);
3427
3428                spin_unlock_irq(&gcwq->lock);
3429        }
3430
3431        spin_unlock(&workqueue_lock);
3432}
3433EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3434
3435/**
3436 * workqueue_congested - test whether a workqueue is congested
3437 * @cpu: CPU in question
3438 * @wq: target workqueue
3439 *
3440 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
3441 * no synchronization around this function and the test result is
3442 * unreliable and only useful as advisory hints or for debugging.
3443 *
3444 * RETURNS:
3445 * %true if congested, %false otherwise.
3446 */
3447bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3448{
3449        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3450
3451        return !list_empty(&cwq->delayed_works);
3452}
3453EXPORT_SYMBOL_GPL(workqueue_congested);
3454
3455/**
3456 * work_cpu - return the last known associated cpu for @work
3457 * @work: the work of interest
3458 *
3459 * RETURNS:
3460 * CPU number if @work was ever queued.  WORK_CPU_NONE otherwise.
3461 */
3462unsigned int work_cpu(struct work_struct *work)
3463{
3464        struct global_cwq *gcwq = get_work_gcwq(work);
3465
3466        return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3467}
3468EXPORT_SYMBOL_GPL(work_cpu);
3469
3470/**
3471 * work_busy - test whether a work is currently pending or running
3472 * @work: the work to be tested
3473 *
3474 * Test whether @work is currently pending or running.  There is no
3475 * synchronization around this function and the test result is
3476 * unreliable and only useful as advisory hints or for debugging.
3477 * Especially for reentrant wqs, the pending state might hide the
3478 * running state.
3479 *
3480 * RETURNS:
3481 * OR'd bitmask of WORK_BUSY_* bits.
3482 */
3483unsigned int work_busy(struct work_struct *work)
3484{
3485        struct global_cwq *gcwq = get_work_gcwq(work);
3486        unsigned long flags;
3487        unsigned int ret = 0;
3488
3489        if (!gcwq)
3490                return 0;
3491
3492        spin_lock_irqsave(&gcwq->lock, flags);
3493
3494        if (work_pending(work))
3495                ret |= WORK_BUSY_PENDING;
3496        if (find_worker_executing_work(gcwq, work))
3497                ret |= WORK_BUSY_RUNNING;
3498
3499        spin_unlock_irqrestore(&gcwq->lock, flags);
3500
3501        return ret;
3502}
3503EXPORT_SYMBOL_GPL(work_busy);
3504
3505/*
3506 * CPU hotplug.
3507 *
3508 * There are two challenges in supporting CPU hotplug.  Firstly, there
3509 * are a lot of assumptions on strong associations among work, cwq and
3510 * gcwq which make migrating pending and scheduled works very
3511 * difficult to implement without impacting hot paths.  Secondly,
3512 * gcwqs serve mix of short, long and very long running works making
3513 * blocked draining impractical.
3514 *
3515 * This is solved by allowing a gcwq to be disassociated from the CPU
3516 * running as an unbound one and allowing it to be reattached later if the
3517 * cpu comes back online.
3518 */
3519
3520/* claim manager positions of all pools */
3521static void gcwq_claim_assoc_and_lock(struct global_cwq *gcwq)
3522{
3523        struct worker_pool *pool;
3524
3525        for_each_worker_pool(pool, gcwq)
3526                mutex_lock_nested(&pool->assoc_mutex, pool - gcwq->pools);
3527        spin_lock_irq(&gcwq->lock);
3528}
3529
3530/* release manager positions */
3531static void gcwq_release_assoc_and_unlock(struct global_cwq *gcwq)
3532{
3533        struct worker_pool *pool;
3534
3535        spin_unlock_irq(&gcwq->lock);
3536        for_each_worker_pool(pool, gcwq)
3537                mutex_unlock(&pool->assoc_mutex);
3538}
3539
3540static void gcwq_unbind_fn(struct work_struct *work)
3541{
3542        struct global_cwq *gcwq = get_gcwq(smp_processor_id());
3543        struct worker_pool *pool;
3544        struct worker *worker;
3545        struct hlist_node *pos;
3546        int i;
3547
3548        BUG_ON(gcwq->cpu != smp_processor_id());
3549
3550        gcwq_claim_assoc_and_lock(gcwq);
3551
3552        /*
3553         * We've claimed all manager positions.  Make all workers unbound
3554         * and set DISASSOCIATED.  Before this, all workers except for the
3555         * ones which are still executing works from before the last CPU
3556         * down must be on the cpu.  After this, they may become diasporas.
3557         */
3558        for_each_worker_pool(pool, gcwq)
3559                list_for_each_entry(worker, &pool->idle_list, entry)
3560                        worker->flags |= WORKER_UNBOUND;
3561
3562        for_each_busy_worker(worker, i, pos, gcwq)
3563                worker->flags |= WORKER_UNBOUND;
3564
3565        gcwq->flags |= GCWQ_DISASSOCIATED;
3566
3567        gcwq_release_assoc_and_unlock(gcwq);
3568
3569        /*
3570         * Call schedule() so that we cross rq->lock and thus can guarantee
3571         * sched callbacks see the %WORKER_UNBOUND flag.  This is necessary
3572         * as scheduler callbacks may be invoked from other cpus.
3573         */
3574        schedule();
3575
3576        /*
3577         * Sched callbacks are disabled now.  Zap nr_running.  After this,
3578         * nr_running stays zero and need_more_worker() and keep_working()
3579         * are always true as long as the worklist is not empty.  @gcwq now
3580         * behaves as unbound (in terms of concurrency management) gcwq
3581         * which is served by workers tied to the CPU.
3582         *
3583         * On return from this function, the current worker would trigger
3584         * unbound chain execution of pending work items if other workers
3585         * didn't already.
3586         */
3587        for_each_worker_pool(pool, gcwq)
3588                atomic_set(get_pool_nr_running(pool), 0);
3589}
3590
3591/*
3592 * Workqueues should be brought up before normal priority CPU notifiers.
3593 * This will be registered high priority CPU notifier.
3594 */
3595static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3596                                               unsigned long action,
3597                                               void *hcpu)
3598{
3599        unsigned int cpu = (unsigned long)hcpu;
3600        struct global_cwq *gcwq = get_gcwq(cpu);
3601        struct worker_pool *pool;
3602
3603        switch (action & ~CPU_TASKS_FROZEN) {
3604        case CPU_UP_PREPARE:
3605                for_each_worker_pool(pool, gcwq) {
3606                        struct worker *worker;
3607
3608                        if (pool->nr_workers)
3609                                continue;
3610
3611                        worker = create_worker(pool);
3612                        if (!worker)
3613                                return NOTIFY_BAD;
3614
3615                        spin_lock_irq(&gcwq->lock);
3616                        start_worker(worker);
3617                        spin_unlock_irq(&gcwq->lock);
3618                }
3619                break;
3620
3621        case CPU_DOWN_FAILED:
3622        case CPU_ONLINE:
3623                gcwq_claim_assoc_and_lock(gcwq);
3624                gcwq->flags &= ~GCWQ_DISASSOCIATED;
3625                rebind_workers(gcwq);
3626                gcwq_release_assoc_and_unlock(gcwq);
3627                break;
3628        }
3629        return NOTIFY_OK;
3630}
3631
3632/*
3633 * Workqueues should be brought down after normal priority CPU notifiers.
3634 * This will be registered as low priority CPU notifier.
3635 */
3636static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3637                                                 unsigned long action,
3638                                                 void *hcpu)
3639{
3640        unsigned int cpu = (unsigned long)hcpu;
3641        struct work_struct unbind_work;
3642
3643        switch (action & ~CPU_TASKS_FROZEN) {
3644        case CPU_DOWN_PREPARE:
3645                /* unbinding should happen on the local CPU */
3646                INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3647                queue_work_on(cpu, system_highpri_wq, &unbind_work);
3648                flush_work(&unbind_work);
3649                break;
3650        }
3651        return NOTIFY_OK;
3652}
3653
3654#ifdef CONFIG_SMP
3655
3656struct work_for_cpu {
3657        struct work_struct work;
3658        long (*fn)(void *);
3659        void *arg;
3660        long ret;
3661};
3662
3663static void work_for_cpu_fn(struct work_struct *work)
3664{
3665        struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3666
3667        wfc->ret = wfc->fn(wfc->arg);
3668}
3669
3670/**
3671 * work_on_cpu - run a function in user context on a particular cpu
3672 * @cpu: the cpu to run on
3673 * @fn: the function to run
3674 * @arg: the function arg
3675 *
3676 * This will return the value @fn returns.
3677 * It is up to the caller to ensure that the cpu doesn't go offline.
3678 * The caller must not hold any locks which would prevent @fn from completing.
3679 */
3680long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3681{
3682        struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3683
3684        INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3685        schedule_work_on(cpu, &wfc.work);
3686        flush_work(&wfc.work);
3687        return wfc.ret;
3688}
3689EXPORT_SYMBOL_GPL(work_on_cpu);
3690#endif /* CONFIG_SMP */
3691
3692#ifdef CONFIG_FREEZER
3693
3694/**
3695 * freeze_workqueues_begin - begin freezing workqueues
3696 *
3697 * Start freezing workqueues.  After this function returns, all freezable
3698 * workqueues will queue new works to their frozen_works list instead of
3699 * gcwq->worklist.
3700 *
3701 * CONTEXT:
3702 * Grabs and releases workqueue_lock and gcwq->lock's.
3703 */
3704void freeze_workqueues_begin(void)
3705{
3706        unsigned int cpu;
3707
3708        spin_lock(&workqueue_lock);
3709
3710        BUG_ON(workqueue_freezing);
3711        workqueue_freezing = true;
3712
3713        for_each_gcwq_cpu(cpu) {
3714                struct global_cwq *gcwq = get_gcwq(cpu);
3715                struct workqueue_struct *wq;
3716
3717                spin_lock_irq(&gcwq->lock);
3718
3719                BUG_ON(gcwq->flags & GCWQ_FREEZING);
3720                gcwq->flags |= GCWQ_FREEZING;
3721
3722                list_for_each_entry(wq, &workqueues, list) {
3723                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3724
3725                        if (cwq && wq->flags & WQ_FREEZABLE)
3726                                cwq->max_active = 0;
3727                }
3728
3729                spin_unlock_irq(&gcwq->lock);
3730        }
3731
3732        spin_unlock(&workqueue_lock);
3733}
3734
3735/**
3736 * freeze_workqueues_busy - are freezable workqueues still busy?
3737 *
3738 * Check whether freezing is complete.  This function must be called
3739 * between freeze_workqueues_begin() and thaw_workqueues().
3740 *
3741 * CONTEXT:
3742 * Grabs and releases workqueue_lock.
3743 *
3744 * RETURNS:
3745 * %true if some freezable workqueues are still busy.  %false if freezing
3746 * is complete.
3747 */
3748bool freeze_workqueues_busy(void)
3749{
3750        unsigned int cpu;
3751        bool busy = false;
3752
3753        spin_lock(&workqueue_lock);
3754
3755        BUG_ON(!workqueue_freezing);
3756
3757        for_each_gcwq_cpu(cpu) {
3758                struct workqueue_struct *wq;
3759                /*
3760                 * nr_active is monotonically decreasing.  It's safe
3761                 * to peek without lock.
3762                 */
3763                list_for_each_entry(wq, &workqueues, list) {
3764                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3765
3766                        if (!cwq || !(wq->flags & WQ_FREEZABLE))
3767                                continue;
3768
3769                        BUG_ON(cwq->nr_active < 0);
3770                        if (cwq->nr_active) {
3771                                busy = true;
3772                                goto out_unlock;
3773                        }
3774                }
3775        }
3776out_unlock:
3777        spin_unlock(&workqueue_lock);
3778        return busy;
3779}
3780
3781/**
3782 * thaw_workqueues - thaw workqueues
3783 *
3784 * Thaw workqueues.  Normal queueing is restored and all collected
3785 * frozen works are transferred to their respective gcwq worklists.
3786 *
3787 * CONTEXT:
3788 * Grabs and releases workqueue_lock and gcwq->lock's.
3789 */
3790void thaw_workqueues(void)
3791{
3792        unsigned int cpu;
3793
3794        spin_lock(&workqueue_lock);
3795
3796        if (!workqueue_freezing)
3797                goto out_unlock;
3798
3799        for_each_gcwq_cpu(cpu) {
3800                struct global_cwq *gcwq = get_gcwq(cpu);
3801                struct worker_pool *pool;
3802                struct workqueue_struct *wq;
3803
3804                spin_lock_irq(&gcwq->lock);
3805
3806                BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3807                gcwq->flags &= ~GCWQ_FREEZING;
3808
3809                list_for_each_entry(wq, &workqueues, list) {
3810                        struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3811
3812                        if (!cwq || !(wq->flags & WQ_FREEZABLE))
3813                                continue;
3814
3815                        /* restore max_active and repopulate worklist */
3816                        cwq_set_max_active(cwq, wq->saved_max_active);
3817                }
3818
3819                for_each_worker_pool(pool, gcwq)
3820                        wake_up_worker(pool);
3821
3822                spin_unlock_irq(&gcwq->lock);
3823        }
3824
3825        workqueue_freezing = false;
3826out_unlock:
3827        spin_unlock(&workqueue_lock);
3828}
3829#endif /* CONFIG_FREEZER */
3830
3831static int __init init_workqueues(void)
3832{
3833        unsigned int cpu;
3834        int i;
3835
3836        /* make sure we have enough bits for OFFQ CPU number */
3837        BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3838                     WORK_CPU_LAST);
3839
3840        cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3841        hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3842
3843        /* initialize gcwqs */
3844        for_each_gcwq_cpu(cpu) {
3845                struct global_cwq *gcwq = get_gcwq(cpu);
3846                struct worker_pool *pool;
3847
3848                spin_lock_init(&gcwq->lock);
3849                gcwq->cpu = cpu;
3850                gcwq->flags |= GCWQ_DISASSOCIATED;
3851
3852                for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3853                        INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3854
3855                for_each_worker_pool(pool, gcwq) {
3856                        pool->gcwq = gcwq;
3857                        INIT_LIST_HEAD(&pool->worklist);
3858                        INIT_LIST_HEAD(&pool->idle_list);
3859
3860                        init_timer_deferrable(&pool->idle_timer);
3861                        pool->idle_timer.function = idle_worker_timeout;
3862                        pool->idle_timer.data = (unsigned long)pool;
3863
3864                        setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3865                                    (unsigned long)pool);
3866
3867                        mutex_init(&pool->assoc_mutex);
3868                        ida_init(&pool->worker_ida);
3869                }
3870        }
3871
3872        /* create the initial worker */
3873        for_each_online_gcwq_cpu(cpu) {
3874                struct global_cwq *gcwq = get_gcwq(cpu);
3875                struct worker_pool *pool;
3876
3877                if (cpu != WORK_CPU_UNBOUND)
3878                        gcwq->flags &= ~GCWQ_DISASSOCIATED;
3879
3880                for_each_worker_pool(pool, gcwq) {
3881                        struct worker *worker;
3882
3883                        worker = create_worker(pool);
3884                        BUG_ON(!worker);
3885                        spin_lock_irq(&gcwq->lock);
3886                        start_worker(worker);
3887                        spin_unlock_irq(&gcwq->lock);
3888                }
3889        }
3890
3891        system_wq = alloc_workqueue("events", 0, 0);
3892        system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
3893        system_long_wq = alloc_workqueue("events_long", 0, 0);
3894        system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3895                                            WQ_UNBOUND_MAX_ACTIVE);
3896        system_freezable_wq = alloc_workqueue("events_freezable",
3897                                              WQ_FREEZABLE, 0);
3898        BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
3899               !system_unbound_wq || !system_freezable_wq);
3900        return 0;
3901}
3902early_initcall(init_workqueues);
3903
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.