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#include <linux/jhash.h>
  45#include <linux/hashtable.h>
  46#include <linux/rculist.h>
  47#include <linux/nodemask.h>
  48#include <linux/moduleparam.h>
  49#include <linux/uaccess.h>
  50
  51#include "workqueue_internal.h"
  52
  53enum {
  54        /*
  55         * worker_pool flags
  56         *
  57         * A bound pool is either associated or disassociated with its CPU.
  58         * While associated (!DISASSOCIATED), all workers are bound to the
  59         * CPU and none has %WORKER_UNBOUND set and concurrency management
  60         * is in effect.
  61         *
  62         * While DISASSOCIATED, the cpu may be offline and all workers have
  63         * %WORKER_UNBOUND set and concurrency management disabled, and may
  64         * be executing on any CPU.  The pool behaves as an unbound one.
  65         *
  66         * Note that DISASSOCIATED should be flipped only while holding
  67         * manager_mutex to avoid changing binding state while
  68         * create_worker() is in progress.
  69         */
  70        POOL_MANAGE_WORKERS     = 1 << 0,       /* need to manage workers */
  71        POOL_DISASSOCIATED      = 1 << 2,       /* cpu can't serve workers */
  72        POOL_FREEZING           = 1 << 3,       /* freeze in progress */
  73
  74        /* worker flags */
  75        WORKER_STARTED          = 1 << 0,       /* started */
  76        WORKER_DIE              = 1 << 1,       /* die die die */
  77        WORKER_IDLE             = 1 << 2,       /* is idle */
  78        WORKER_PREP             = 1 << 3,       /* preparing to run works */
  79        WORKER_CPU_INTENSIVE    = 1 << 6,       /* cpu intensive */
  80        WORKER_UNBOUND          = 1 << 7,       /* worker is unbound */
  81        WORKER_REBOUND          = 1 << 8,       /* worker was rebound */
  82
  83        WORKER_NOT_RUNNING      = WORKER_PREP | WORKER_CPU_INTENSIVE |
  84                                  WORKER_UNBOUND | WORKER_REBOUND,
  85
  86        NR_STD_WORKER_POOLS     = 2,            /* # standard pools per cpu */
  87
  88        UNBOUND_POOL_HASH_ORDER = 6,            /* hashed by pool->attrs */
  89        BUSY_WORKER_HASH_ORDER  = 6,            /* 64 pointers */
  90
  91        MAX_IDLE_WORKERS_RATIO  = 4,            /* 1/4 of busy can be idle */
  92        IDLE_WORKER_TIMEOUT     = 300 * HZ,     /* keep idle ones for 5 mins */
  93
  94        MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
  95                                                /* call for help after 10ms
  96                                                   (min two ticks) */
  97        MAYDAY_INTERVAL         = HZ / 10,      /* and then every 100ms */
  98        CREATE_COOLDOWN         = HZ,           /* time to breath after fail */
  99
 100        /*
 101         * Rescue workers are used only on emergencies and shared by
 102         * all cpus.  Give -20.
 103         */
 104        RESCUER_NICE_LEVEL      = -20,
 105        HIGHPRI_NICE_LEVEL      = -20,
 106
 107        WQ_NAME_LEN             = 24,
 108};
 109
 110/*
 111 * Structure fields follow one of the following exclusion rules.
 112 *
 113 * I: Modifiable by initialization/destruction paths and read-only for
 114 *    everyone else.
 115 *
 116 * P: Preemption protected.  Disabling preemption is enough and should
 117 *    only be modified and accessed from the local cpu.
 118 *
 119 * L: pool->lock protected.  Access with pool->lock held.
 120 *
 121 * X: During normal operation, modification requires pool->lock and should
 122 *    be done only from local cpu.  Either disabling preemption on local
 123 *    cpu or grabbing pool->lock is enough for read access.  If
 124 *    POOL_DISASSOCIATED is set, it's identical to L.
 125 *
 126 * MG: pool->manager_mutex and pool->lock protected.  Writes require both
 127 *     locks.  Reads can happen under either lock.
 128 *
 129 * PL: wq_pool_mutex protected.
 130 *
 131 * PR: wq_pool_mutex protected for writes.  Sched-RCU protected for reads.
 132 *
 133 * WQ: wq->mutex protected.
 134 *
 135 * WR: wq->mutex protected for writes.  Sched-RCU protected for reads.
 136 *
 137 * MD: wq_mayday_lock protected.
 138 */
 139
 140/* struct worker is defined in workqueue_internal.h */
 141
 142struct worker_pool {
 143        spinlock_t              lock;           /* the pool lock */
 144        int                     cpu;            /* I: the associated cpu */
 145        int                     node;           /* I: the associated node ID */
 146        int                     id;             /* I: pool ID */
 147        unsigned int            flags;          /* X: flags */
 148
 149        struct list_head        worklist;       /* L: list of pending works */
 150        int                     nr_workers;     /* L: total number of workers */
 151
 152        /* nr_idle includes the ones off idle_list for rebinding */
 153        int                     nr_idle;        /* L: currently idle ones */
 154
 155        struct list_head        idle_list;      /* X: list of idle workers */
 156        struct timer_list       idle_timer;     /* L: worker idle timeout */
 157        struct timer_list       mayday_timer;   /* L: SOS timer for workers */
 158
 159        /* a workers is either on busy_hash or idle_list, or the manager */
 160        DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
 161                                                /* L: hash of busy workers */
 162
 163        /* see manage_workers() for details on the two manager mutexes */
 164        struct mutex            manager_arb;    /* manager arbitration */
 165        struct mutex            manager_mutex;  /* manager exclusion */
 166        struct idr              worker_idr;     /* MG: worker IDs and iteration */
 167
 168        struct workqueue_attrs  *attrs;         /* I: worker attributes */
 169        struct hlist_node       hash_node;      /* PL: unbound_pool_hash node */
 170        int                     refcnt;         /* PL: refcnt for unbound pools */
 171
 172        /*
 173         * The current concurrency level.  As it's likely to be accessed
 174         * from other CPUs during try_to_wake_up(), put it in a separate
 175         * cacheline.
 176         */
 177        atomic_t                nr_running ____cacheline_aligned_in_smp;
 178
 179        /*
 180         * Destruction of pool is sched-RCU protected to allow dereferences
 181         * from get_work_pool().
 182         */
 183        struct rcu_head         rcu;
 184} ____cacheline_aligned_in_smp;
 185
 186/*
 187 * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
 188 * of work_struct->data are used for flags and the remaining high bits
 189 * point to the pwq; thus, pwqs need to be aligned at two's power of the
 190 * number of flag bits.
 191 */
 192struct pool_workqueue {
 193        struct worker_pool      *pool;          /* I: the associated pool */
 194        struct workqueue_struct *wq;            /* I: the owning workqueue */
 195        int                     work_color;     /* L: current color */
 196        int                     flush_color;    /* L: flushing color */
 197        int                     refcnt;         /* L: reference count */
 198        int                     nr_in_flight[WORK_NR_COLORS];
 199                                                /* L: nr of in_flight works */
 200        int                     nr_active;      /* L: nr of active works */
 201        int                     max_active;     /* L: max active works */
 202        struct list_head        delayed_works;  /* L: delayed works */
 203        struct list_head        pwqs_node;      /* WR: node on wq->pwqs */
 204        struct list_head        mayday_node;    /* MD: node on wq->maydays */
 205
 206        /*
 207         * Release of unbound pwq is punted to system_wq.  See put_pwq()
 208         * and pwq_unbound_release_workfn() for details.  pool_workqueue
 209         * itself is also sched-RCU protected so that the first pwq can be
 210         * determined without grabbing wq->mutex.
 211         */
 212        struct work_struct      unbound_release_work;
 213        struct rcu_head         rcu;
 214} __aligned(1 << WORK_STRUCT_FLAG_BITS);
 215
 216/*
 217 * Structure used to wait for workqueue flush.
 218 */
 219struct wq_flusher {
 220        struct list_head        list;           /* WQ: list of flushers */
 221        int                     flush_color;    /* WQ: flush color waiting for */
 222        struct completion       done;           /* flush completion */
 223};
 224
 225struct wq_device;
 226
 227/*
 228 * The externally visible workqueue.  It relays the issued work items to
 229 * the appropriate worker_pool through its pool_workqueues.
 230 */
 231struct workqueue_struct {
 232        struct list_head        pwqs;           /* WR: all pwqs of this wq */
 233        struct list_head        list;           /* PL: list of all workqueues */
 234
 235        struct mutex            mutex;          /* protects this wq */
 236        int                     work_color;     /* WQ: current work color */
 237        int                     flush_color;    /* WQ: current flush color */
 238        atomic_t                nr_pwqs_to_flush; /* flush in progress */
 239        struct wq_flusher       *first_flusher; /* WQ: first flusher */
 240        struct list_head        flusher_queue;  /* WQ: flush waiters */
 241        struct list_head        flusher_overflow; /* WQ: flush overflow list */
 242
 243        struct list_head        maydays;        /* MD: pwqs requesting rescue */
 244        struct worker           *rescuer;       /* I: rescue worker */
 245
 246        int                     nr_drainers;    /* WQ: drain in progress */
 247        int                     saved_max_active; /* WQ: saved pwq max_active */
 248
 249        struct workqueue_attrs  *unbound_attrs; /* WQ: only for unbound wqs */
 250        struct pool_workqueue   *dfl_pwq;       /* WQ: only for unbound wqs */
 251
 252#ifdef CONFIG_SYSFS
 253        struct wq_device        *wq_dev;        /* I: for sysfs interface */
 254#endif
 255#ifdef CONFIG_LOCKDEP
 256        struct lockdep_map      lockdep_map;
 257#endif
 258        char                    name[WQ_NAME_LEN]; /* I: workqueue name */
 259
 260        /* hot fields used during command issue, aligned to cacheline */
 261        unsigned int            flags ____cacheline_aligned; /* WQ: WQ_* flags */
 262        struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
 263        struct pool_workqueue __rcu *numa_pwq_tbl[]; /* FR: unbound pwqs indexed by node */
 264};
 265
 266static struct kmem_cache *pwq_cache;
 267
 268static int wq_numa_tbl_len;             /* highest possible NUMA node id + 1 */
 269static cpumask_var_t *wq_numa_possible_cpumask;
 270                                        /* possible CPUs of each node */
 271
 272static bool wq_disable_numa;
 273module_param_named(disable_numa, wq_disable_numa, bool, 0444);
 274
 275static bool wq_numa_enabled;            /* unbound NUMA affinity enabled */
 276
 277/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
 278static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
 279
 280static DEFINE_MUTEX(wq_pool_mutex);     /* protects pools and workqueues list */
 281static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
 282
 283static LIST_HEAD(workqueues);           /* PL: list of all workqueues */
 284static bool workqueue_freezing;         /* PL: have wqs started freezing? */
 285
 286/* the per-cpu worker pools */
 287static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
 288                                     cpu_worker_pools);
 289
 290static DEFINE_IDR(worker_pool_idr);     /* PR: idr of all pools */
 291
 292/* PL: hash of all unbound pools keyed by pool->attrs */
 293static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
 294
 295/* I: attributes used when instantiating standard unbound pools on demand */
 296static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
 297
 298struct workqueue_struct *system_wq __read_mostly;
 299EXPORT_SYMBOL(system_wq);
 300struct workqueue_struct *system_highpri_wq __read_mostly;
 301EXPORT_SYMBOL_GPL(system_highpri_wq);
 302struct workqueue_struct *system_long_wq __read_mostly;
 303EXPORT_SYMBOL_GPL(system_long_wq);
 304struct workqueue_struct *system_unbound_wq __read_mostly;
 305EXPORT_SYMBOL_GPL(system_unbound_wq);
 306struct workqueue_struct *system_freezable_wq __read_mostly;
 307EXPORT_SYMBOL_GPL(system_freezable_wq);
 308
 309static int worker_thread(void *__worker);
 310static void copy_workqueue_attrs(struct workqueue_attrs *to,
 311                                 const struct workqueue_attrs *from);
 312
 313#define CREATE_TRACE_POINTS
 314#include <trace/events/workqueue.h>
 315
 316#define assert_rcu_or_pool_mutex()                                      \
 317        rcu_lockdep_assert(rcu_read_lock_sched_held() ||                \
 318                           lockdep_is_held(&wq_pool_mutex),             \
 319                           "sched RCU or wq_pool_mutex should be held")
 320
 321#define assert_rcu_or_wq_mutex(wq)                                      \
 322        rcu_lockdep_assert(rcu_read_lock_sched_held() ||                \
 323                           lockdep_is_held(&wq->mutex),                 \
 324                           "sched RCU or wq->mutex should be held")
 325
 326#ifdef CONFIG_LOCKDEP
 327#define assert_manager_or_pool_lock(pool)                               \
 328        WARN_ONCE(debug_locks &&                                        \
 329                  !lockdep_is_held(&(pool)->manager_mutex) &&           \
 330                  !lockdep_is_held(&(pool)->lock),                      \
 331                  "pool->manager_mutex or ->lock should be held")
 332#else
 333#define assert_manager_or_pool_lock(pool)       do { } while (0)
 334#endif
 335
 336#define for_each_cpu_worker_pool(pool, cpu)                             \
 337        for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];               \
 338             (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
 339             (pool)++)
 340
 341/**
 342 * for_each_pool - iterate through all worker_pools in the system
 343 * @pool: iteration cursor
 344 * @pi: integer used for iteration
 345 *
 346 * This must be called either with wq_pool_mutex held or sched RCU read
 347 * locked.  If the pool needs to be used beyond the locking in effect, the
 348 * caller is responsible for guaranteeing that the pool stays online.
 349 *
 350 * The if/else clause exists only for the lockdep assertion and can be
 351 * ignored.
 352 */
 353#define for_each_pool(pool, pi)                                         \
 354        idr_for_each_entry(&worker_pool_idr, pool, pi)                  \
 355                if (({ assert_rcu_or_pool_mutex(); false; })) { }       \
 356                else
 357
 358/**
 359 * for_each_pool_worker - iterate through all workers of a worker_pool
 360 * @worker: iteration cursor
 361 * @wi: integer used for iteration
 362 * @pool: worker_pool to iterate workers of
 363 *
 364 * This must be called with either @pool->manager_mutex or ->lock held.
 365 *
 366 * The if/else clause exists only for the lockdep assertion and can be
 367 * ignored.
 368 */
 369#define for_each_pool_worker(worker, wi, pool)                          \
 370        idr_for_each_entry(&(pool)->worker_idr, (worker), (wi))         \
 371                if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
 372                else
 373
 374/**
 375 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
 376 * @pwq: iteration cursor
 377 * @wq: the target workqueue
 378 *
 379 * This must be called either with wq->mutex held or sched RCU read locked.
 380 * If the pwq needs to be used beyond the locking in effect, the caller is
 381 * responsible for guaranteeing that the pwq stays online.
 382 *
 383 * The if/else clause exists only for the lockdep assertion and can be
 384 * ignored.
 385 */
 386#define for_each_pwq(pwq, wq)                                           \
 387        list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)          \
 388                if (({ assert_rcu_or_wq_mutex(wq); false; })) { }       \
 389                else
 390
 391#ifdef CONFIG_DEBUG_OBJECTS_WORK
 392
 393static struct debug_obj_descr work_debug_descr;
 394
 395static void *work_debug_hint(void *addr)
 396{
 397        return ((struct work_struct *) addr)->func;
 398}
 399
 400/*
 401 * fixup_init is called when:
 402 * - an active object is initialized
 403 */
 404static int work_fixup_init(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_init(work, &work_debug_descr);
 412                return 1;
 413        default:
 414                return 0;
 415        }
 416}
 417
 418/*
 419 * fixup_activate is called when:
 420 * - an active object is activated
 421 * - an unknown object is activated (might be a statically initialized object)
 422 */
 423static int work_fixup_activate(void *addr, enum debug_obj_state state)
 424{
 425        struct work_struct *work = addr;
 426
 427        switch (state) {
 428
 429        case ODEBUG_STATE_NOTAVAILABLE:
 430                /*
 431                 * This is not really a fixup. The work struct was
 432                 * statically initialized. We just make sure that it
 433                 * is tracked in the object tracker.
 434                 */
 435                if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
 436                        debug_object_init(work, &work_debug_descr);
 437                        debug_object_activate(work, &work_debug_descr);
 438                        return 0;
 439                }
 440                WARN_ON_ONCE(1);
 441                return 0;
 442
 443        case ODEBUG_STATE_ACTIVE:
 444                WARN_ON(1);
 445
 446        default:
 447                return 0;
 448        }
 449}
 450
 451/*
 452 * fixup_free is called when:
 453 * - an active object is freed
 454 */
 455static int work_fixup_free(void *addr, enum debug_obj_state state)
 456{
 457        struct work_struct *work = addr;
 458
 459        switch (state) {
 460        case ODEBUG_STATE_ACTIVE:
 461                cancel_work_sync(work);
 462                debug_object_free(work, &work_debug_descr);
 463                return 1;
 464        default:
 465                return 0;
 466        }
 467}
 468
 469static struct debug_obj_descr work_debug_descr = {
 470        .name           = "work_struct",
 471        .debug_hint     = work_debug_hint,
 472        .fixup_init     = work_fixup_init,
 473        .fixup_activate = work_fixup_activate,
 474        .fixup_free     = work_fixup_free,
 475};
 476
 477static inline void debug_work_activate(struct work_struct *work)
 478{
 479        debug_object_activate(work, &work_debug_descr);
 480}
 481
 482static inline void debug_work_deactivate(struct work_struct *work)
 483{
 484        debug_object_deactivate(work, &work_debug_descr);
 485}
 486
 487void __init_work(struct work_struct *work, int onstack)
 488{
 489        if (onstack)
 490                debug_object_init_on_stack(work, &work_debug_descr);
 491        else
 492                debug_object_init(work, &work_debug_descr);
 493}
 494EXPORT_SYMBOL_GPL(__init_work);
 495
 496void destroy_work_on_stack(struct work_struct *work)
 497{
 498        debug_object_free(work, &work_debug_descr);
 499}
 500EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 501
 502#else
 503static inline void debug_work_activate(struct work_struct *work) { }
 504static inline void debug_work_deactivate(struct work_struct *work) { }
 505#endif
 506
 507/* allocate ID and assign it to @pool */
 508static int worker_pool_assign_id(struct worker_pool *pool)
 509{
 510        int ret;
 511
 512        lockdep_assert_held(&wq_pool_mutex);
 513
 514        ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
 515        if (ret >= 0) {
 516                pool->id = ret;
 517                return 0;
 518        }
 519        return ret;
 520}
 521
 522/**
 523 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
 524 * @wq: the target workqueue
 525 * @node: the node ID
 526 *
 527 * This must be called either with pwq_lock held or sched RCU read locked.
 528 * If the pwq needs to be used beyond the locking in effect, the caller is
 529 * responsible for guaranteeing that the pwq stays online.
 530 */
 531static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
 532                                                  int node)
 533{
 534        assert_rcu_or_wq_mutex(wq);
 535        return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
 536}
 537
 538static unsigned int work_color_to_flags(int color)
 539{
 540        return color << WORK_STRUCT_COLOR_SHIFT;
 541}
 542
 543static int get_work_color(struct work_struct *work)
 544{
 545        return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 546                ((1 << WORK_STRUCT_COLOR_BITS) - 1);
 547}
 548
 549static int work_next_color(int color)
 550{
 551        return (color + 1) % WORK_NR_COLORS;
 552}
 553
 554/*
 555 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
 556 * contain the pointer to the queued pwq.  Once execution starts, the flag
 557 * is cleared and the high bits contain OFFQ flags and pool ID.
 558 *
 559 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
 560 * and clear_work_data() can be used to set the pwq, pool or clear
 561 * work->data.  These functions should only be called while the work is
 562 * owned - ie. while the PENDING bit is set.
 563 *
 564 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
 565 * corresponding to a work.  Pool is available once the work has been
 566 * queued anywhere after initialization until it is sync canceled.  pwq is
 567 * available only while the work item is queued.
 568 *
 569 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
 570 * canceled.  While being canceled, a work item may have its PENDING set
 571 * but stay off timer and worklist for arbitrarily long and nobody should
 572 * try to steal the PENDING bit.
 573 */
 574static inline void set_work_data(struct work_struct *work, unsigned long data,
 575                                 unsigned long flags)
 576{
 577        WARN_ON_ONCE(!work_pending(work));
 578        atomic_long_set(&work->data, data | flags | work_static(work));
 579}
 580
 581static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
 582                         unsigned long extra_flags)
 583{
 584        set_work_data(work, (unsigned long)pwq,
 585                      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
 586}
 587
 588static void set_work_pool_and_keep_pending(struct work_struct *work,
 589                                           int pool_id)
 590{
 591        set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
 592                      WORK_STRUCT_PENDING);
 593}
 594
 595static void set_work_pool_and_clear_pending(struct work_struct *work,
 596                                            int pool_id)
 597{
 598        /*
 599         * The following wmb is paired with the implied mb in
 600         * test_and_set_bit(PENDING) and ensures all updates to @work made
 601         * here are visible to and precede any updates by the next PENDING
 602         * owner.
 603         */
 604        smp_wmb();
 605        set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
 606}
 607
 608static void clear_work_data(struct work_struct *work)
 609{
 610        smp_wmb();      /* see set_work_pool_and_clear_pending() */
 611        set_work_data(work, WORK_STRUCT_NO_POOL, 0);
 612}
 613
 614static struct pool_workqueue *get_work_pwq(struct work_struct *work)
 615{
 616        unsigned long data = atomic_long_read(&work->data);
 617
 618        if (data & WORK_STRUCT_PWQ)
 619                return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 620        else
 621                return NULL;
 622}
 623
 624/**
 625 * get_work_pool - return the worker_pool a given work was associated with
 626 * @work: the work item of interest
 627 *
 628 * Return the worker_pool @work was last associated with.  %NULL if none.
 629 *
 630 * Pools are created and destroyed under wq_pool_mutex, and allows read
 631 * access under sched-RCU read lock.  As such, this function should be
 632 * called under wq_pool_mutex or with preemption disabled.
 633 *
 634 * All fields of the returned pool are accessible as long as the above
 635 * mentioned locking is in effect.  If the returned pool needs to be used
 636 * beyond the critical section, the caller is responsible for ensuring the
 637 * returned pool is and stays online.
 638 */
 639static struct worker_pool *get_work_pool(struct work_struct *work)
 640{
 641        unsigned long data = atomic_long_read(&work->data);
 642        int pool_id;
 643
 644        assert_rcu_or_pool_mutex();
 645
 646        if (data & WORK_STRUCT_PWQ)
 647                return ((struct pool_workqueue *)
 648                        (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
 649
 650        pool_id = data >> WORK_OFFQ_POOL_SHIFT;
 651        if (pool_id == WORK_OFFQ_POOL_NONE)
 652                return NULL;
 653
 654        return idr_find(&worker_pool_idr, pool_id);
 655}
 656
 657/**
 658 * get_work_pool_id - return the worker pool ID a given work is associated with
 659 * @work: the work item of interest
 660 *
 661 * Return the worker_pool ID @work was last associated with.
 662 * %WORK_OFFQ_POOL_NONE if none.
 663 */
 664static int get_work_pool_id(struct work_struct *work)
 665{
 666        unsigned long data = atomic_long_read(&work->data);
 667
 668        if (data & WORK_STRUCT_PWQ)
 669                return ((struct pool_workqueue *)
 670                        (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
 671
 672        return data >> WORK_OFFQ_POOL_SHIFT;
 673}
 674
 675static void mark_work_canceling(struct work_struct *work)
 676{
 677        unsigned long pool_id = get_work_pool_id(work);
 678
 679        pool_id <<= WORK_OFFQ_POOL_SHIFT;
 680        set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
 681}
 682
 683static bool work_is_canceling(struct work_struct *work)
 684{
 685        unsigned long data = atomic_long_read(&work->data);
 686
 687        return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
 688}
 689
 690/*
 691 * Policy functions.  These define the policies on how the global worker
 692 * pools are managed.  Unless noted otherwise, these functions assume that
 693 * they're being called with pool->lock held.
 694 */
 695
 696static bool __need_more_worker(struct worker_pool *pool)
 697{
 698        return !atomic_read(&pool->nr_running);
 699}
 700
 701/*
 702 * Need to wake up a worker?  Called from anything but currently
 703 * running workers.
 704 *
 705 * Note that, because unbound workers never contribute to nr_running, this
 706 * function will always return %true for unbound pools as long as the
 707 * worklist isn't empty.
 708 */
 709static bool need_more_worker(struct worker_pool *pool)
 710{
 711        return !list_empty(&pool->worklist) && __need_more_worker(pool);
 712}
 713
 714/* Can I start working?  Called from busy but !running workers. */
 715static bool may_start_working(struct worker_pool *pool)
 716{
 717        return pool->nr_idle;
 718}
 719
 720/* Do I need to keep working?  Called from currently running workers. */
 721static bool keep_working(struct worker_pool *pool)
 722{
 723        return !list_empty(&pool->worklist) &&
 724                atomic_read(&pool->nr_running) <= 1;
 725}
 726
 727/* Do we need a new worker?  Called from manager. */
 728static bool need_to_create_worker(struct worker_pool *pool)
 729{
 730        return need_more_worker(pool) && !may_start_working(pool);
 731}
 732
 733/* Do I need to be the manager? */
 734static bool need_to_manage_workers(struct worker_pool *pool)
 735{
 736        return need_to_create_worker(pool) ||
 737                (pool->flags & POOL_MANAGE_WORKERS);
 738}
 739
 740/* Do we have too many workers and should some go away? */
 741static bool too_many_workers(struct worker_pool *pool)
 742{
 743        bool managing = mutex_is_locked(&pool->manager_arb);
 744        int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
 745        int nr_busy = pool->nr_workers - nr_idle;
 746
 747        /*
 748         * nr_idle and idle_list may disagree if idle rebinding is in
 749         * progress.  Never return %true if idle_list is empty.
 750         */
 751        if (list_empty(&pool->idle_list))
 752                return false;
 753
 754        return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 755}
 756
 757/*
 758 * Wake up functions.
 759 */
 760
 761/* Return the first worker.  Safe with preemption disabled */
 762static struct worker *first_worker(struct worker_pool *pool)
 763{
 764        if (unlikely(list_empty(&pool->idle_list)))
 765                return NULL;
 766
 767        return list_first_entry(&pool->idle_list, struct worker, entry);
 768}
 769
 770/**
 771 * wake_up_worker - wake up an idle worker
 772 * @pool: worker pool to wake worker from
 773 *
 774 * Wake up the first idle worker of @pool.
 775 *
 776 * CONTEXT:
 777 * spin_lock_irq(pool->lock).
 778 */
 779static void wake_up_worker(struct worker_pool *pool)
 780{
 781        struct worker *worker = first_worker(pool);
 782
 783        if (likely(worker))
 784                wake_up_process(worker->task);
 785}
 786
 787/**
 788 * wq_worker_waking_up - a worker is waking up
 789 * @task: task waking up
 790 * @cpu: CPU @task is waking up to
 791 *
 792 * This function is called during try_to_wake_up() when a worker is
 793 * being awoken.
 794 *
 795 * CONTEXT:
 796 * spin_lock_irq(rq->lock)
 797 */
 798void wq_worker_waking_up(struct task_struct *task, int cpu)
 799{
 800        struct worker *worker = kthread_data(task);
 801
 802        if (!(worker->flags & WORKER_NOT_RUNNING)) {
 803                WARN_ON_ONCE(worker->pool->cpu != cpu);
 804                atomic_inc(&worker->pool->nr_running);
 805        }
 806}
 807
 808/**
 809 * wq_worker_sleeping - a worker is going to sleep
 810 * @task: task going to sleep
 811 * @cpu: CPU in question, must be the current CPU number
 812 *
 813 * This function is called during schedule() when a busy worker is
 814 * going to sleep.  Worker on the same cpu can be woken up by
 815 * returning pointer to its task.
 816 *
 817 * CONTEXT:
 818 * spin_lock_irq(rq->lock)
 819 *
 820 * RETURNS:
 821 * Worker task on @cpu to wake up, %NULL if none.
 822 */
 823struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
 824{
 825        struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 826        struct worker_pool *pool;
 827
 828        /*
 829         * Rescuers, which may not have all the fields set up like normal
 830         * workers, also reach here, let's not access anything before
 831         * checking NOT_RUNNING.
 832         */
 833        if (worker->flags & WORKER_NOT_RUNNING)
 834                return NULL;
 835
 836        pool = worker->pool;
 837
 838        /* this can only happen on the local cpu */
 839        if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
 840                return NULL;
 841
 842        /*
 843         * The counterpart of the following dec_and_test, implied mb,
 844         * worklist not empty test sequence is in insert_work().
 845         * Please read comment there.
 846         *
 847         * NOT_RUNNING is clear.  This means that we're bound to and
 848         * running on the local cpu w/ rq lock held and preemption
 849         * disabled, which in turn means that none else could be
 850         * manipulating idle_list, so dereferencing idle_list without pool
 851         * lock is safe.
 852         */
 853        if (atomic_dec_and_test(&pool->nr_running) &&
 854            !list_empty(&pool->worklist))
 855                to_wakeup = first_worker(pool);
 856        return to_wakeup ? to_wakeup->task : NULL;
 857}
 858
 859/**
 860 * worker_set_flags - set worker flags and adjust nr_running accordingly
 861 * @worker: self
 862 * @flags: flags to set
 863 * @wakeup: wakeup an idle worker if necessary
 864 *
 865 * Set @flags in @worker->flags and adjust nr_running accordingly.  If
 866 * nr_running becomes zero and @wakeup is %true, an idle worker is
 867 * woken up.
 868 *
 869 * CONTEXT:
 870 * spin_lock_irq(pool->lock)
 871 */
 872static inline void worker_set_flags(struct worker *worker, unsigned int flags,
 873                                    bool wakeup)
 874{
 875        struct worker_pool *pool = worker->pool;
 876
 877        WARN_ON_ONCE(worker->task != current);
 878
 879        /*
 880         * If transitioning into NOT_RUNNING, adjust nr_running and
 881         * wake up an idle worker as necessary if requested by
 882         * @wakeup.
 883         */
 884        if ((flags & WORKER_NOT_RUNNING) &&
 885            !(worker->flags & WORKER_NOT_RUNNING)) {
 886                if (wakeup) {
 887                        if (atomic_dec_and_test(&pool->nr_running) &&
 888                            !list_empty(&pool->worklist))
 889                                wake_up_worker(pool);
 890                } else
 891                        atomic_dec(&pool->nr_running);
 892        }
 893
 894        worker->flags |= flags;
 895}
 896
 897/**
 898 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 899 * @worker: self
 900 * @flags: flags to clear
 901 *
 902 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 903 *
 904 * CONTEXT:
 905 * spin_lock_irq(pool->lock)
 906 */
 907static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 908{
 909        struct worker_pool *pool = worker->pool;
 910        unsigned int oflags = worker->flags;
 911
 912        WARN_ON_ONCE(worker->task != current);
 913
 914        worker->flags &= ~flags;
 915
 916        /*
 917         * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 918         * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 919         * of multiple flags, not a single flag.
 920         */
 921        if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 922                if (!(worker->flags & WORKER_NOT_RUNNING))
 923                        atomic_inc(&pool->nr_running);
 924}
 925
 926/**
 927 * find_worker_executing_work - find worker which is executing a work
 928 * @pool: pool of interest
 929 * @work: work to find worker for
 930 *
 931 * Find a worker which is executing @work on @pool by searching
 932 * @pool->busy_hash which is keyed by the address of @work.  For a worker
 933 * to match, its current execution should match the address of @work and
 934 * its work function.  This is to avoid unwanted dependency between
 935 * unrelated work executions through a work item being recycled while still
 936 * being executed.
 937 *
 938 * This is a bit tricky.  A work item may be freed once its execution
 939 * starts and nothing prevents the freed area from being recycled for
 940 * another work item.  If the same work item address ends up being reused
 941 * before the original execution finishes, workqueue will identify the
 942 * recycled work item as currently executing and make it wait until the
 943 * current execution finishes, introducing an unwanted dependency.
 944 *
 945 * This function checks the work item address and work function to avoid
 946 * false positives.  Note that this isn't complete as one may construct a
 947 * work function which can introduce dependency onto itself through a
 948 * recycled work item.  Well, if somebody wants to shoot oneself in the
 949 * foot that badly, there's only so much we can do, and if such deadlock
 950 * actually occurs, it should be easy to locate the culprit work function.
 951 *
 952 * CONTEXT:
 953 * spin_lock_irq(pool->lock).
 954 *
 955 * RETURNS:
 956 * Pointer to worker which is executing @work if found, NULL
 957 * otherwise.
 958 */
 959static struct worker *find_worker_executing_work(struct worker_pool *pool,
 960                                                 struct work_struct *work)
 961{
 962        struct worker *worker;
 963
 964        hash_for_each_possible(pool->busy_hash, worker, hentry,
 965                               (unsigned long)work)
 966                if (worker->current_work == work &&
 967                    worker->current_func == work->func)
 968                        return worker;
 969
 970        return NULL;
 971}
 972
 973/**
 974 * move_linked_works - move linked works to a list
 975 * @work: start of series of works to be scheduled
 976 * @head: target list to append @work to
 977 * @nextp: out paramter for nested worklist walking
 978 *
 979 * Schedule linked works starting from @work to @head.  Work series to
 980 * be scheduled starts at @work and includes any consecutive work with
 981 * WORK_STRUCT_LINKED set in its predecessor.
 982 *
 983 * If @nextp is not NULL, it's updated to point to the next work of
 984 * the last scheduled work.  This allows move_linked_works() to be
 985 * nested inside outer list_for_each_entry_safe().
 986 *
 987 * CONTEXT:
 988 * spin_lock_irq(pool->lock).
 989 */
 990static void move_linked_works(struct work_struct *work, struct list_head *head,
 991                              struct work_struct **nextp)
 992{
 993        struct work_struct *n;
 994
 995        /*
 996         * Linked worklist will always end before the end of the list,
 997         * use NULL for list head.
 998         */
 999        list_for_each_entry_safe_from(work, n, NULL, entry) {
1000                list_move_tail(&work->entry, head);
1001                if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1002                        break;
1003        }
1004
1005        /*
1006         * If we're already inside safe list traversal and have moved
1007         * multiple works to the scheduled queue, the next position
1008         * needs to be updated.
1009         */
1010        if (nextp)
1011                *nextp = n;
1012}
1013
1014/**
1015 * get_pwq - get an extra reference on the specified pool_workqueue
1016 * @pwq: pool_workqueue to get
1017 *
1018 * Obtain an extra reference on @pwq.  The caller should guarantee that
1019 * @pwq has positive refcnt and be holding the matching pool->lock.
1020 */
1021static void get_pwq(struct pool_workqueue *pwq)
1022{
1023        lockdep_assert_held(&pwq->pool->lock);
1024        WARN_ON_ONCE(pwq->refcnt <= 0);
1025        pwq->refcnt++;
1026}
1027
1028/**
1029 * put_pwq - put a pool_workqueue reference
1030 * @pwq: pool_workqueue to put
1031 *
1032 * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1033 * destruction.  The caller should be holding the matching pool->lock.
1034 */
1035static void put_pwq(struct pool_workqueue *pwq)
1036{
1037        lockdep_assert_held(&pwq->pool->lock);
1038        if (likely(--pwq->refcnt))
1039                return;
1040        if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1041                return;
1042        /*
1043         * @pwq can't be released under pool->lock, bounce to
1044         * pwq_unbound_release_workfn().  This never recurses on the same
1045         * pool->lock as this path is taken only for unbound workqueues and
1046         * the release work item is scheduled on a per-cpu workqueue.  To
1047         * avoid lockdep warning, unbound pool->locks are given lockdep
1048         * subclass of 1 in get_unbound_pool().
1049         */
1050        schedule_work(&pwq->unbound_release_work);
1051}
1052
1053/**
1054 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1055 * @pwq: pool_workqueue to put (can be %NULL)
1056 *
1057 * put_pwq() with locking.  This function also allows %NULL @pwq.
1058 */
1059static void put_pwq_unlocked(struct pool_workqueue *pwq)
1060{
1061        if (pwq) {
1062                /*
1063                 * As both pwqs and pools are sched-RCU protected, the
1064                 * following lock operations are safe.
1065                 */
1066                spin_lock_irq(&pwq->pool->lock);
1067                put_pwq(pwq);
1068                spin_unlock_irq(&pwq->pool->lock);
1069        }
1070}
1071
1072static void pwq_activate_delayed_work(struct work_struct *work)
1073{
1074        struct pool_workqueue *pwq = get_work_pwq(work);
1075
1076        trace_workqueue_activate_work(work);
1077        move_linked_works(work, &pwq->pool->worklist, NULL);
1078        __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1079        pwq->nr_active++;
1080}
1081
1082static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1083{
1084        struct work_struct *work = list_first_entry(&pwq->delayed_works,
1085                                                    struct work_struct, entry);
1086
1087        pwq_activate_delayed_work(work);
1088}
1089
1090/**
1091 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1092 * @pwq: pwq of interest
1093 * @color: color of work which left the queue
1094 *
1095 * A work either has completed or is removed from pending queue,
1096 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1097 *
1098 * CONTEXT:
1099 * spin_lock_irq(pool->lock).
1100 */
1101static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1102{
1103        /* uncolored work items don't participate in flushing or nr_active */
1104        if (color == WORK_NO_COLOR)
1105                goto out_put;
1106
1107        pwq->nr_in_flight[color]--;
1108
1109        pwq->nr_active--;
1110        if (!list_empty(&pwq->delayed_works)) {
1111                /* one down, submit a delayed one */
1112                if (pwq->nr_active < pwq->max_active)
1113                        pwq_activate_first_delayed(pwq);
1114        }
1115
1116        /* is flush in progress and are we at the flushing tip? */
1117        if (likely(pwq->flush_color != color))
1118                goto out_put;
1119
1120        /* are there still in-flight works? */
1121        if (pwq->nr_in_flight[color])
1122                goto out_put;
1123
1124        /* this pwq is done, clear flush_color */
1125        pwq->flush_color = -1;
1126
1127        /*
1128         * If this was the last pwq, wake up the first flusher.  It
1129         * will handle the rest.
1130         */
1131        if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1132                complete(&pwq->wq->first_flusher->done);
1133out_put:
1134        put_pwq(pwq);
1135}
1136
1137/**
1138 * try_to_grab_pending - steal work item from worklist and disable irq
1139 * @work: work item to steal
1140 * @is_dwork: @work is a delayed_work
1141 * @flags: place to store irq state
1142 *
1143 * Try to grab PENDING bit of @work.  This function can handle @work in any
1144 * stable state - idle, on timer or on worklist.  Return values are
1145 *
1146 *  1           if @work was pending and we successfully stole PENDING
1147 *  0           if @work was idle and we claimed PENDING
1148 *  -EAGAIN     if PENDING couldn't be grabbed at the moment, safe to busy-retry
1149 *  -ENOENT     if someone else is canceling @work, this state may persist
1150 *              for arbitrarily long
1151 *
1152 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1153 * interrupted while holding PENDING and @work off queue, irq must be
1154 * disabled on entry.  This, combined with delayed_work->timer being
1155 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1156 *
1157 * On successful return, >= 0, irq is disabled and the caller is
1158 * responsible for releasing it using local_irq_restore(*@flags).
1159 *
1160 * This function is safe to call from any context including IRQ handler.
1161 */
1162static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1163                               unsigned long *flags)
1164{
1165        struct worker_pool *pool;
1166        struct pool_workqueue *pwq;
1167
1168        local_irq_save(*flags);
1169
1170        /* try to steal the timer if it exists */
1171        if (is_dwork) {
1172                struct delayed_work *dwork = to_delayed_work(work);
1173
1174                /*
1175                 * dwork->timer is irqsafe.  If del_timer() fails, it's
1176                 * guaranteed that the timer is not queued anywhere and not
1177                 * running on the local CPU.
1178                 */
1179                if (likely(del_timer(&dwork->timer)))
1180                        return 1;
1181        }
1182
1183        /* try to claim PENDING the normal way */
1184        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1185                return 0;
1186
1187        /*
1188         * The queueing is in progress, or it is already queued. Try to
1189         * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1190         */
1191        pool = get_work_pool(work);
1192        if (!pool)
1193                goto fail;
1194
1195        spin_lock(&pool->lock);
1196        /*
1197         * work->data is guaranteed to point to pwq only while the work
1198         * item is queued on pwq->wq, and both updating work->data to point
1199         * to pwq on queueing and to pool on dequeueing are done under
1200         * pwq->pool->lock.  This in turn guarantees that, if work->data
1201         * points to pwq which is associated with a locked pool, the work
1202         * item is currently queued on that pool.
1203         */
1204        pwq = get_work_pwq(work);
1205        if (pwq && pwq->pool == pool) {
1206                debug_work_deactivate(work);
1207
1208                /*
1209                 * A delayed work item cannot be grabbed directly because
1210                 * it might have linked NO_COLOR work items which, if left
1211                 * on the delayed_list, will confuse pwq->nr_active
1212                 * management later on and cause stall.  Make sure the work
1213                 * item is activated before grabbing.
1214                 */
1215                if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1216                        pwq_activate_delayed_work(work);
1217
1218                list_del_init(&work->entry);
1219                pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1220
1221                /* work->data points to pwq iff queued, point to pool */
1222                set_work_pool_and_keep_pending(work, pool->id);
1223
1224                spin_unlock(&pool->lock);
1225                return 1;
1226        }
1227        spin_unlock(&pool->lock);
1228fail:
1229        local_irq_restore(*flags);
1230        if (work_is_canceling(work))
1231                return -ENOENT;
1232        cpu_relax();
1233        return -EAGAIN;
1234}
1235
1236/**
1237 * insert_work - insert a work into a pool
1238 * @pwq: pwq @work belongs to
1239 * @work: work to insert
1240 * @head: insertion point
1241 * @extra_flags: extra WORK_STRUCT_* flags to set
1242 *
1243 * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1244 * work_struct flags.
1245 *
1246 * CONTEXT:
1247 * spin_lock_irq(pool->lock).
1248 */
1249static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1250                        struct list_head *head, unsigned int extra_flags)
1251{
1252        struct worker_pool *pool = pwq->pool;
1253
1254        /* we own @work, set data and link */
1255        set_work_pwq(work, pwq, extra_flags);
1256        list_add_tail(&work->entry, head);
1257        get_pwq(pwq);
1258
1259        /*
1260         * Ensure either wq_worker_sleeping() sees the above
1261         * list_add_tail() or we see zero nr_running to avoid workers lying
1262         * around lazily while there are works to be processed.
1263         */
1264        smp_mb();
1265
1266        if (__need_more_worker(pool))
1267                wake_up_worker(pool);
1268}
1269
1270/*
1271 * Test whether @work is being queued from another work executing on the
1272 * same workqueue.
1273 */
1274static bool is_chained_work(struct workqueue_struct *wq)
1275{
1276        struct worker *worker;
1277
1278        worker = current_wq_worker();
1279        /*
1280         * Return %true iff I'm a worker execuing a work item on @wq.  If
1281         * I'm @worker, it's safe to dereference it without locking.
1282         */
1283        return worker && worker->current_pwq->wq == wq;
1284}
1285
1286static void __queue_work(int cpu, struct workqueue_struct *wq,
1287                         struct work_struct *work)
1288{
1289        struct pool_workqueue *pwq;
1290        struct worker_pool *last_pool;
1291        struct list_head *worklist;
1292        unsigned int work_flags;
1293        unsigned int req_cpu = cpu;
1294
1295        /*
1296         * While a work item is PENDING && off queue, a task trying to
1297         * steal the PENDING will busy-loop waiting for it to either get
1298         * queued or lose PENDING.  Grabbing PENDING and queueing should
1299         * happen with IRQ disabled.
1300         */
1301        WARN_ON_ONCE(!irqs_disabled());
1302
1303        debug_work_activate(work);
1304
1305        /* if dying, only works from the same workqueue are allowed */
1306        if (unlikely(wq->flags & __WQ_DRAINING) &&
1307            WARN_ON_ONCE(!is_chained_work(wq)))
1308                return;
1309retry:
1310        if (req_cpu == WORK_CPU_UNBOUND)
1311                cpu = raw_smp_processor_id();
1312
1313        /* pwq which will be used unless @work is executing elsewhere */
1314        if (!(wq->flags & WQ_UNBOUND))
1315                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1316        else
1317                pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1318
1319        /*
1320         * If @work was previously on a different pool, it might still be
1321         * running there, in which case the work needs to be queued on that
1322         * pool to guarantee non-reentrancy.
1323         */
1324        last_pool = get_work_pool(work);
1325        if (last_pool && last_pool != pwq->pool) {
1326                struct worker *worker;
1327
1328                spin_lock(&last_pool->lock);
1329
1330                worker = find_worker_executing_work(last_pool, work);
1331
1332                if (worker && worker->current_pwq->wq == wq) {
1333                        pwq = worker->current_pwq;
1334                } else {
1335                        /* meh... not running there, queue here */
1336                        spin_unlock(&last_pool->lock);
1337                        spin_lock(&pwq->pool->lock);
1338                }
1339        } else {
1340                spin_lock(&pwq->pool->lock);
1341        }
1342
1343        /*
1344         * pwq is determined and locked.  For unbound pools, we could have
1345         * raced with pwq release and it could already be dead.  If its
1346         * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1347         * without another pwq replacing it in the numa_pwq_tbl or while
1348         * work items are executing on it, so the retrying is guaranteed to
1349         * make forward-progress.
1350         */
1351        if (unlikely(!pwq->refcnt)) {
1352                if (wq->flags & WQ_UNBOUND) {
1353                        spin_unlock(&pwq->pool->lock);
1354                        cpu_relax();
1355                        goto retry;
1356                }
1357                /* oops */
1358                WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1359                          wq->name, cpu);
1360        }
1361
1362        /* pwq determined, queue */
1363        trace_workqueue_queue_work(req_cpu, pwq, work);
1364
1365        if (WARN_ON(!list_empty(&work->entry))) {
1366                spin_unlock(&pwq->pool->lock);
1367                return;
1368        }
1369
1370        pwq->nr_in_flight[pwq->work_color]++;
1371        work_flags = work_color_to_flags(pwq->work_color);
1372
1373        if (likely(pwq->nr_active < pwq->max_active)) {
1374                trace_workqueue_activate_work(work);
1375                pwq->nr_active++;
1376                worklist = &pwq->pool->worklist;
1377        } else {
1378                work_flags |= WORK_STRUCT_DELAYED;
1379                worklist = &pwq->delayed_works;
1380        }
1381
1382        insert_work(pwq, work, worklist, work_flags);
1383
1384        spin_unlock(&pwq->pool->lock);
1385}
1386
1387/**
1388 * queue_work_on - queue work on specific cpu
1389 * @cpu: CPU number to execute work on
1390 * @wq: workqueue to use
1391 * @work: work to queue
1392 *
1393 * Returns %false if @work was already on a queue, %true otherwise.
1394 *
1395 * We queue the work to a specific CPU, the caller must ensure it
1396 * can't go away.
1397 */
1398bool queue_work_on(int cpu, struct workqueue_struct *wq,
1399                   struct work_struct *work)
1400{
1401        bool ret = false;
1402        unsigned long flags;
1403
1404        local_irq_save(flags);
1405
1406        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1407                __queue_work(cpu, wq, work);
1408                ret = true;
1409        }
1410
1411        local_irq_restore(flags);
1412        return ret;
1413}
1414EXPORT_SYMBOL(queue_work_on);
1415
1416void delayed_work_timer_fn(unsigned long __data)
1417{
1418        struct delayed_work *dwork = (struct delayed_work *)__data;
1419
1420        /* should have been called from irqsafe timer with irq already off */
1421        __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1422}
1423EXPORT_SYMBOL(delayed_work_timer_fn);
1424
1425static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1426                                struct delayed_work *dwork, unsigned long delay)
1427{
1428        struct timer_list *timer = &dwork->timer;
1429        struct work_struct *work = &dwork->work;
1430
1431        WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1432                     timer->data != (unsigned long)dwork);
1433        WARN_ON_ONCE(timer_pending(timer));
1434        WARN_ON_ONCE(!list_empty(&work->entry));
1435
1436        /*
1437         * If @delay is 0, queue @dwork->work immediately.  This is for
1438         * both optimization and correctness.  The earliest @timer can
1439         * expire is on the closest next tick and delayed_work users depend
1440         * on that there's no such delay when @delay is 0.
1441         */
1442        if (!delay) {
1443                __queue_work(cpu, wq, &dwork->work);
1444                return;
1445        }
1446
1447        timer_stats_timer_set_start_info(&dwork->timer);
1448
1449        dwork->wq = wq;
1450        dwork->cpu = cpu;
1451        timer->expires = jiffies + delay;
1452
1453        if (unlikely(cpu != WORK_CPU_UNBOUND))
1454                add_timer_on(timer, cpu);
1455        else
1456                add_timer(timer);
1457}
1458
1459/**
1460 * queue_delayed_work_on - queue work on specific CPU after delay
1461 * @cpu: CPU number to execute work on
1462 * @wq: workqueue to use
1463 * @dwork: work to queue
1464 * @delay: number of jiffies to wait before queueing
1465 *
1466 * Returns %false if @work was already on a queue, %true otherwise.  If
1467 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1468 * execution.
1469 */
1470bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1471                           struct delayed_work *dwork, unsigned long delay)
1472{
1473        struct work_struct *work = &dwork->work;
1474        bool ret = false;
1475        unsigned long flags;
1476
1477        /* read the comment in __queue_work() */
1478        local_irq_save(flags);
1479
1480        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1481                __queue_delayed_work(cpu, wq, dwork, delay);
1482                ret = true;
1483        }
1484
1485        local_irq_restore(flags);
1486        return ret;
1487}
1488EXPORT_SYMBOL(queue_delayed_work_on);
1489
1490/**
1491 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @dwork: work to queue
1495 * @delay: number of jiffies to wait before queueing
1496 *
1497 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1498 * modify @dwork's timer so that it expires after @delay.  If @delay is
1499 * zero, @work is guaranteed to be scheduled immediately regardless of its
1500 * current state.
1501 *
1502 * Returns %false if @dwork was idle and queued, %true if @dwork was
1503 * pending and its timer was modified.
1504 *
1505 * This function is safe to call from any context including IRQ handler.
1506 * See try_to_grab_pending() for details.
1507 */
1508bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1509                         struct delayed_work *dwork, unsigned long delay)
1510{
1511        unsigned long flags;
1512        int ret;
1513
1514        do {
1515                ret = try_to_grab_pending(&dwork->work, true, &flags);
1516        } while (unlikely(ret == -EAGAIN));
1517
1518        if (likely(ret >= 0)) {
1519                __queue_delayed_work(cpu, wq, dwork, delay);
1520                local_irq_restore(flags);
1521        }
1522
1523        /* -ENOENT from try_to_grab_pending() becomes %true */
1524        return ret;
1525}
1526EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1527
1528/**
1529 * worker_enter_idle - enter idle state
1530 * @worker: worker which is entering idle state
1531 *
1532 * @worker is entering idle state.  Update stats and idle timer if
1533 * necessary.
1534 *
1535 * LOCKING:
1536 * spin_lock_irq(pool->lock).
1537 */
1538static void worker_enter_idle(struct worker *worker)
1539{
1540        struct worker_pool *pool = worker->pool;
1541
1542        if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1543            WARN_ON_ONCE(!list_empty(&worker->entry) &&
1544                         (worker->hentry.next || worker->hentry.pprev)))
1545                return;
1546
1547        /* can't use worker_set_flags(), also called from start_worker() */
1548        worker->flags |= WORKER_IDLE;
1549        pool->nr_idle++;
1550        worker->last_active = jiffies;
1551
1552        /* idle_list is LIFO */
1553        list_add(&worker->entry, &pool->idle_list);
1554
1555        if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1556                mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1557
1558        /*
1559         * Sanity check nr_running.  Because wq_unbind_fn() releases
1560         * pool->lock between setting %WORKER_UNBOUND and zapping
1561         * nr_running, the warning may trigger spuriously.  Check iff
1562         * unbind is not in progress.
1563         */
1564        WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1565                     pool->nr_workers == pool->nr_idle &&
1566                     atomic_read(&pool->nr_running));
1567}
1568
1569/**
1570 * worker_leave_idle - leave idle state
1571 * @worker: worker which is leaving idle state
1572 *
1573 * @worker is leaving idle state.  Update stats.
1574 *
1575 * LOCKING:
1576 * spin_lock_irq(pool->lock).
1577 */
1578static void worker_leave_idle(struct worker *worker)
1579{
1580        struct worker_pool *pool = worker->pool;
1581
1582        if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1583                return;
1584        worker_clr_flags(worker, WORKER_IDLE);
1585        pool->nr_idle--;
1586        list_del_init(&worker->entry);
1587}
1588
1589/**
1590 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1591 * @pool: target worker_pool
1592 *
1593 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1594 *
1595 * Works which are scheduled while the cpu is online must at least be
1596 * scheduled to a worker which is bound to the cpu so that if they are
1597 * flushed from cpu callbacks while cpu is going down, they are
1598 * guaranteed to execute on the cpu.
1599 *
1600 * This function is to be used by unbound workers and rescuers to bind
1601 * themselves to the target cpu and may race with cpu going down or
1602 * coming online.  kthread_bind() can't be used because it may put the
1603 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1604 * verbatim as it's best effort and blocking and pool may be
1605 * [dis]associated in the meantime.
1606 *
1607 * This function tries set_cpus_allowed() and locks pool and verifies the
1608 * binding against %POOL_DISASSOCIATED which is set during
1609 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1610 * enters idle state or fetches works without dropping lock, it can
1611 * guarantee the scheduling requirement described in the first paragraph.
1612 *
1613 * CONTEXT:
1614 * Might sleep.  Called without any lock but returns with pool->lock
1615 * held.
1616 *
1617 * RETURNS:
1618 * %true if the associated pool is online (@worker is successfully
1619 * bound), %false if offline.
1620 */
1621static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1622__acquires(&pool->lock)
1623{
1624        while (true) {
1625                /*
1626                 * The following call may fail, succeed or succeed
1627                 * without actually migrating the task to the cpu if
1628                 * it races with cpu hotunplug operation.  Verify
1629                 * against POOL_DISASSOCIATED.
1630                 */
1631                if (!(pool->flags & POOL_DISASSOCIATED))
1632                        set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1633
1634                spin_lock_irq(&pool->lock);
1635                if (pool->flags & POOL_DISASSOCIATED)
1636                        return false;
1637                if (task_cpu(current) == pool->cpu &&
1638                    cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
1639                        return true;
1640                spin_unlock_irq(&pool->lock);
1641
1642                /*
1643                 * We've raced with CPU hot[un]plug.  Give it a breather
1644                 * and retry migration.  cond_resched() is required here;
1645                 * otherwise, we might deadlock against cpu_stop trying to
1646                 * bring down the CPU on non-preemptive kernel.
1647                 */
1648                cpu_relax();
1649                cond_resched();
1650        }
1651}
1652
1653static struct worker *alloc_worker(void)
1654{
1655        struct worker *worker;
1656
1657        worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1658        if (worker) {
1659                INIT_LIST_HEAD(&worker->entry);
1660                INIT_LIST_HEAD(&worker->scheduled);
1661                /* on creation a worker is in !idle && prep state */
1662                worker->flags = WORKER_PREP;
1663        }
1664        return worker;
1665}
1666
1667/**
1668 * create_worker - create a new workqueue worker
1669 * @pool: pool the new worker will belong to
1670 *
1671 * Create a new worker which is bound to @pool.  The returned worker
1672 * can be started by calling start_worker() or destroyed using
1673 * destroy_worker().
1674 *
1675 * CONTEXT:
1676 * Might sleep.  Does GFP_KERNEL allocations.
1677 *
1678 * RETURNS:
1679 * Pointer to the newly created worker.
1680 */
1681static struct worker *create_worker(struct worker_pool *pool)
1682{
1683        struct worker *worker = NULL;
1684        int id = -1;
1685        char id_buf[16];
1686
1687        lockdep_assert_held(&pool->manager_mutex);
1688
1689        /*
1690         * ID is needed to determine kthread name.  Allocate ID first
1691         * without installing the pointer.
1692         */
1693        idr_preload(GFP_KERNEL);
1694        spin_lock_irq(&pool->lock);
1695
1696        id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1697
1698        spin_unlock_irq(&pool->lock);
1699        idr_preload_end();
1700        if (id < 0)
1701                goto fail;
1702
1703        worker = alloc_worker();
1704        if (!worker)
1705                goto fail;
1706
1707        worker->pool = pool;
1708        worker->id = id;
1709
1710        if (pool->cpu >= 0)
1711                snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1712                         pool->attrs->nice < 0  ? "H" : "");
1713        else
1714                snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1715
1716        worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1717                                              "kworker/%s", id_buf);
1718        if (IS_ERR(worker->task))
1719                goto fail;
1720
1721        /*
1722         * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723         * online CPUs.  It'll be re-applied when any of the CPUs come up.
1724         */
1725        set_user_nice(worker->task, pool->attrs->nice);
1726        set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1727
1728        /* prevent userland from meddling with cpumask of workqueue workers */
1729        worker->task->flags |= PF_NO_SETAFFINITY;
1730
1731        /*
1732         * The caller is responsible for ensuring %POOL_DISASSOCIATED
1733         * remains stable across this function.  See the comments above the
1734         * flag definition for details.
1735         */
1736        if (pool->flags & POOL_DISASSOCIATED)
1737                worker->flags |= WORKER_UNBOUND;
1738
1739        /* successful, commit the pointer to idr */
1740        spin_lock_irq(&pool->lock);
1741        idr_replace(&pool->worker_idr, worker, worker->id);
1742        spin_unlock_irq(&pool->lock);
1743
1744        return worker;
1745
1746fail:
1747        if (id >= 0) {
1748                spin_lock_irq(&pool->lock);
1749                idr_remove(&pool->worker_idr, id);
1750                spin_unlock_irq(&pool->lock);
1751        }
1752        kfree(worker);
1753        return NULL;
1754}
1755
1756/**
1757 * start_worker - start a newly created worker
1758 * @worker: worker to start
1759 *
1760 * Make the pool aware of @worker and start it.
1761 *
1762 * CONTEXT:
1763 * spin_lock_irq(pool->lock).
1764 */
1765static void start_worker(struct worker *worker)
1766{
1767        worker->flags |= WORKER_STARTED;
1768        worker->pool->nr_workers++;
1769        worker_enter_idle(worker);
1770        wake_up_process(worker->task);
1771}
1772
1773/**
1774 * create_and_start_worker - create and start a worker for a pool
1775 * @pool: the target pool
1776 *
1777 * Grab the managership of @pool and create and start a new worker for it.
1778 */
1779static int create_and_start_worker(struct worker_pool *pool)
1780{
1781        struct worker *worker;
1782
1783        mutex_lock(&pool->manager_mutex);
1784
1785        worker = create_worker(pool);
1786        if (worker) {
1787                spin_lock_irq(&pool->lock);
1788                start_worker(worker);
1789                spin_unlock_irq(&pool->lock);
1790        }
1791
1792        mutex_unlock(&pool->manager_mutex);
1793
1794        return worker ? 0 : -ENOMEM;
1795}
1796
1797/**
1798 * destroy_worker - destroy a workqueue worker
1799 * @worker: worker to be destroyed
1800 *
1801 * Destroy @worker and adjust @pool stats accordingly.
1802 *
1803 * CONTEXT:
1804 * spin_lock_irq(pool->lock) which is released and regrabbed.
1805 */
1806static void destroy_worker(struct worker *worker)
1807{
1808        struct worker_pool *pool = worker->pool;
1809
1810        lockdep_assert_held(&pool->manager_mutex);
1811        lockdep_assert_held(&pool->lock);
1812
1813        /* sanity check frenzy */
1814        if (WARN_ON(worker->current_work) ||
1815            WARN_ON(!list_empty(&worker->scheduled)))
1816                return;
1817
1818        if (worker->flags & WORKER_STARTED)
1819                pool->nr_workers--;
1820        if (worker->flags & WORKER_IDLE)
1821                pool->nr_idle--;
1822
1823        list_del_init(&worker->entry);
1824        worker->flags |= WORKER_DIE;
1825
1826        idr_remove(&pool->worker_idr, worker->id);
1827
1828        spin_unlock_irq(&pool->lock);
1829
1830        kthread_stop(worker->task);
1831        kfree(worker);
1832
1833        spin_lock_irq(&pool->lock);
1834}
1835
1836static void idle_worker_timeout(unsigned long __pool)
1837{
1838        struct worker_pool *pool = (void *)__pool;
1839
1840        spin_lock_irq(&pool->lock);
1841
1842        if (too_many_workers(pool)) {
1843                struct worker *worker;
1844                unsigned long expires;
1845
1846                /* idle_list is kept in LIFO order, check the last one */
1847                worker = list_entry(pool->idle_list.prev, struct worker, entry);
1848                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1849
1850                if (time_before(jiffies, expires))
1851                        mod_timer(&pool->idle_timer, expires);
1852                else {
1853                        /* it's been idle for too long, wake up manager */
1854                        pool->flags |= POOL_MANAGE_WORKERS;
1855                        wake_up_worker(pool);
1856                }
1857        }
1858
1859        spin_unlock_irq(&pool->lock);
1860}
1861
1862static void send_mayday(struct work_struct *work)
1863{
1864        struct pool_workqueue *pwq = get_work_pwq(work);
1865        struct workqueue_struct *wq = pwq->wq;
1866
1867        lockdep_assert_held(&wq_mayday_lock);
1868
1869        if (!wq->rescuer)
1870                return;
1871
1872        /* mayday mayday mayday */
1873        if (list_empty(&pwq->mayday_node)) {
1874                list_add_tail(&pwq->mayday_node, &wq->maydays);
1875                wake_up_process(wq->rescuer->task);
1876        }
1877}
1878
1879static void pool_mayday_timeout(unsigned long __pool)
1880{
1881        struct worker_pool *pool = (void *)__pool;
1882        struct work_struct *work;
1883
1884        spin_lock_irq(&wq_mayday_lock);         /* for wq->maydays */
1885        spin_lock(&pool->lock);
1886
1887        if (need_to_create_worker(pool)) {
1888                /*
1889                 * We've been trying to create a new worker but
1890                 * haven't been successful.  We might be hitting an
1891                 * allocation deadlock.  Send distress signals to
1892                 * rescuers.
1893                 */
1894                list_for_each_entry(work, &pool->worklist, entry)
1895                        send_mayday(work);
1896        }
1897
1898        spin_unlock(&pool->lock);
1899        spin_unlock_irq(&wq_mayday_lock);
1900
1901        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1902}
1903
1904/**
1905 * maybe_create_worker - create a new worker if necessary
1906 * @pool: pool to create a new worker for
1907 *
1908 * Create a new worker for @pool if necessary.  @pool is guaranteed to
1909 * have at least one idle worker on return from this function.  If
1910 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1911 * sent to all rescuers with works scheduled on @pool to resolve
1912 * possible allocation deadlock.
1913 *
1914 * On return, need_to_create_worker() is guaranteed to be %false and
1915 * may_start_working() %true.
1916 *
1917 * LOCKING:
1918 * spin_lock_irq(pool->lock) which may be released and regrabbed
1919 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1920 * manager.
1921 *
1922 * RETURNS:
1923 * %false if no action was taken and pool->lock stayed locked, %true
1924 * otherwise.
1925 */
1926static bool maybe_create_worker(struct worker_pool *pool)
1927__releases(&pool->lock)
1928__acquires(&pool->lock)
1929{
1930        if (!need_to_create_worker(pool))
1931                return false;
1932restart:
1933        spin_unlock_irq(&pool->lock);
1934
1935        /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1936        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1937
1938        while (true) {
1939                struct worker *worker;
1940
1941                worker = create_worker(pool);
1942                if (worker) {
1943                        del_timer_sync(&pool->mayday_timer);
1944                        spin_lock_irq(&pool->lock);
1945                        start_worker(worker);
1946                        if (WARN_ON_ONCE(need_to_create_worker(pool)))
1947                                goto restart;
1948                        return true;
1949                }
1950
1951                if (!need_to_create_worker(pool))
1952                        break;
1953
1954                __set_current_state(TASK_INTERRUPTIBLE);
1955                schedule_timeout(CREATE_COOLDOWN);
1956
1957                if (!need_to_create_worker(pool))
1958                        break;
1959        }
1960
1961        del_timer_sync(&pool->mayday_timer);
1962        spin_lock_irq(&pool->lock);
1963        if (need_to_create_worker(pool))
1964                goto restart;
1965        return true;
1966}
1967
1968/**
1969 * maybe_destroy_worker - destroy workers which have been idle for a while
1970 * @pool: pool to destroy workers for
1971 *
1972 * Destroy @pool workers which have been idle for longer than
1973 * IDLE_WORKER_TIMEOUT.
1974 *
1975 * LOCKING:
1976 * spin_lock_irq(pool->lock) which may be released and regrabbed
1977 * multiple times.  Called only from manager.
1978 *
1979 * RETURNS:
1980 * %false if no action was taken and pool->lock stayed locked, %true
1981 * otherwise.
1982 */
1983static bool maybe_destroy_workers(struct worker_pool *pool)
1984{
1985        bool ret = false;
1986
1987        while (too_many_workers(pool)) {
1988                struct worker *worker;
1989                unsigned long expires;
1990
1991                worker = list_entry(pool->idle_list.prev, struct worker, entry);
1992                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1993
1994                if (time_before(jiffies, expires)) {
1995                        mod_timer(&pool->idle_timer, expires);
1996                        break;
1997                }
1998
1999                destroy_worker(worker);
2000                ret = true;
2001        }
2002
2003        return ret;
2004}
2005
2006/**
2007 * manage_workers - manage worker pool
2008 * @worker: self
2009 *
2010 * Assume the manager role and manage the worker pool @worker belongs
2011 * to.  At any given time, there can be only zero or one manager per
2012 * pool.  The exclusion is handled automatically by this function.
2013 *
2014 * The caller can safely start processing works on false return.  On
2015 * true return, it's guaranteed that need_to_create_worker() is false
2016 * and may_start_working() is true.
2017 *
2018 * CONTEXT:
2019 * spin_lock_irq(pool->lock) which may be released and regrabbed
2020 * multiple times.  Does GFP_KERNEL allocations.
2021 *
2022 * RETURNS:
2023 * spin_lock_irq(pool->lock) which may be released and regrabbed
2024 * multiple times.  Does GFP_KERNEL allocations.
2025 */
2026static bool manage_workers(struct worker *worker)
2027{
2028        struct worker_pool *pool = worker->pool;
2029        bool ret = false;
2030
2031        /*
2032         * Managership is governed by two mutexes - manager_arb and
2033         * manager_mutex.  manager_arb handles arbitration of manager role.
2034         * Anyone who successfully grabs manager_arb wins the arbitration
2035         * and becomes the manager.  mutex_trylock() on pool->manager_arb
2036         * failure while holding pool->lock reliably indicates that someone
2037         * else is managing the pool and the worker which failed trylock
2038         * can proceed to executing work items.  This means that anyone
2039         * grabbing manager_arb is responsible for actually performing
2040         * manager duties.  If manager_arb is grabbed and released without
2041         * actual management, the pool may stall indefinitely.
2042         *
2043         * manager_mutex is used for exclusion of actual management
2044         * operations.  The holder of manager_mutex can be sure that none
2045         * of management operations, including creation and destruction of
2046         * workers, won't take place until the mutex is released.  Because
2047         * manager_mutex doesn't interfere with manager role arbitration,
2048         * it is guaranteed that the pool's management, while may be
2049         * delayed, won't be disturbed by someone else grabbing
2050         * manager_mutex.
2051         */
2052        if (!mutex_trylock(&pool->manager_arb))
2053                return ret;
2054
2055        /*
2056         * With manager arbitration won, manager_mutex would be free in
2057         * most cases.  trylock first without dropping @pool->lock.
2058         */
2059        if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2060                spin_unlock_irq(&pool->lock);
2061                mutex_lock(&pool->manager_mutex);
2062                spin_lock_irq(&pool->lock);
2063                ret = true;
2064        }
2065
2066        pool->flags &= ~POOL_MANAGE_WORKERS;
2067
2068        /*
2069         * Destroy and then create so that may_start_working() is true
2070         * on return.
2071         */
2072        ret |= maybe_destroy_workers(pool);
2073        ret |= maybe_create_worker(pool);
2074
2075        mutex_unlock(&pool->manager_mutex);
2076        mutex_unlock(&pool->manager_arb);
2077        return ret;
2078}
2079
2080/**
2081 * process_one_work - process single work
2082 * @worker: self
2083 * @work: work to process
2084 *
2085 * Process @work.  This function contains all the logics necessary to
2086 * process a single work including synchronization against and
2087 * interaction with other workers on the same cpu, queueing and
2088 * flushing.  As long as context requirement is met, any worker can
2089 * call this function to process a work.
2090 *
2091 * CONTEXT:
2092 * spin_lock_irq(pool->lock) which is released and regrabbed.
2093 */
2094static void process_one_work(struct worker *worker, struct work_struct *work)
2095__releases(&pool->lock)
2096__acquires(&pool->lock)
2097{
2098        struct pool_workqueue *pwq = get_work_pwq(work);
2099        struct worker_pool *pool = worker->pool;
2100        bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2101        int work_color;
2102        struct worker *collision;
2103#ifdef CONFIG_LOCKDEP
2104        /*
2105         * It is permissible to free the struct work_struct from
2106         * inside the function that is called from it, this we need to
2107         * take into account for lockdep too.  To avoid bogus "held
2108         * lock freed" warnings as well as problems when looking into
2109         * work->lockdep_map, make a copy and use that here.
2110         */
2111        struct lockdep_map lockdep_map;
2112
2113        lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2114#endif
2115        /*
2116         * Ensure we're on the correct CPU.  DISASSOCIATED test is
2117         * necessary to avoid spurious warnings from rescuers servicing the
2118         * unbound or a disassociated pool.
2119         */
2120        WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2121                     !(pool->flags & POOL_DISASSOCIATED) &&
2122                     raw_smp_processor_id() != pool->cpu);
2123
2124        /*
2125         * A single work shouldn't be executed concurrently by
2126         * multiple workers on a single cpu.  Check whether anyone is
2127         * already processing the work.  If so, defer the work to the
2128         * currently executing one.
2129         */
2130        collision = find_worker_executing_work(pool, work);
2131        if (unlikely(collision)) {
2132                move_linked_works(work, &collision->scheduled, NULL);
2133                return;
2134        }
2135
2136        /* claim and dequeue */
2137        debug_work_deactivate(work);
2138        hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2139        worker->current_work = work;
2140        worker->current_func = work->func;
2141        worker->current_pwq = pwq;
2142        work_color = get_work_color(work);
2143
2144        list_del_init(&work->entry);
2145
2146        /*
2147         * CPU intensive works don't participate in concurrency
2148         * management.  They're the scheduler's responsibility.
2149         */
2150        if (unlikely(cpu_intensive))
2151                worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2152
2153        /*
2154         * Unbound pool isn't concurrency managed and work items should be
2155         * executed ASAP.  Wake up another worker if necessary.
2156         */
2157        if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2158                wake_up_worker(pool);
2159
2160        /*
2161         * Record the last pool and clear PENDING which should be the last
2162         * update to @work.  Also, do this inside @pool->lock so that
2163         * PENDING and queued state changes happen together while IRQ is
2164         * disabled.
2165         */
2166        set_work_pool_and_clear_pending(work, pool->id);
2167
2168        spin_unlock_irq(&pool->lock);
2169
2170        lock_map_acquire_read(&pwq->wq->lockdep_map);
2171        lock_map_acquire(&lockdep_map);
2172        trace_workqueue_execute_start(work);
2173        worker->current_func(work);
2174        /*
2175         * While we must be careful to not use "work" after this, the trace
2176         * point will only record its address.
2177         */
2178        trace_workqueue_execute_end(work);
2179        lock_map_release(&lockdep_map);
2180        lock_map_release(&pwq->wq->lockdep_map);
2181
2182        if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2183                pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2184                       "     last function: %pf\n",
2185                       current->comm, preempt_count(), task_pid_nr(current),
2186                       worker->current_func);
2187                debug_show_held_locks(current);
2188                dump_stack();
2189        }
2190
2191        spin_lock_irq(&pool->lock);
2192
2193        /* clear cpu intensive status */
2194        if (unlikely(cpu_intensive))
2195                worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2196
2197        /* we're done with it, release */
2198        hash_del(&worker->hentry);
2199        worker->current_work = NULL;
2200        worker->current_func = NULL;
2201        worker->current_pwq = NULL;
2202        worker->desc_valid = false;
2203        pwq_dec_nr_in_flight(pwq, work_color);
2204}
2205
2206/**
2207 * process_scheduled_works - process scheduled works
2208 * @worker: self
2209 *
2210 * Process all scheduled works.  Please note that the scheduled list
2211 * may change while processing a work, so this function repeatedly
2212 * fetches a work from the top and executes it.
2213 *
2214 * CONTEXT:
2215 * spin_lock_irq(pool->lock) which may be released and regrabbed
2216 * multiple times.
2217 */
2218static void process_scheduled_works(struct worker *worker)
2219{
2220        while (!list_empty(&worker->scheduled)) {
2221                struct work_struct *work = list_first_entry(&worker->scheduled,
2222                                                struct work_struct, entry);
2223                process_one_work(worker, work);
2224        }
2225}
2226
2227/**
2228 * worker_thread - the worker thread function
2229 * @__worker: self
2230 *
2231 * The worker thread function.  All workers belong to a worker_pool -
2232 * either a per-cpu one or dynamic unbound one.  These workers process all
2233 * work items regardless of their specific target workqueue.  The only
2234 * exception is work items which belong to workqueues with a rescuer which
2235 * will be explained in rescuer_thread().
2236 */
2237static int worker_thread(void *__worker)
2238{
2239        struct worker *worker = __worker;
2240        struct worker_pool *pool = worker->pool;
2241
2242        /* tell the scheduler that this is a workqueue worker */
2243        worker->task->flags |= PF_WQ_WORKER;
2244woke_up:
2245        spin_lock_irq(&pool->lock);
2246
2247        /* am I supposed to die? */
2248        if (unlikely(worker->flags & WORKER_DIE)) {
2249                spin_unlock_irq(&pool->lock);
2250                WARN_ON_ONCE(!list_empty(&worker->entry));
2251                worker->task->flags &= ~PF_WQ_WORKER;
2252                return 0;
2253        }
2254
2255        worker_leave_idle(worker);
2256recheck:
2257        /* no more worker necessary? */
2258        if (!need_more_worker(pool))
2259                goto sleep;
2260
2261        /* do we need to manage? */
2262        if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2263                goto recheck;
2264
2265        /*
2266         * ->scheduled list can only be filled while a worker is
2267         * preparing to process a work or actually processing it.
2268         * Make sure nobody diddled with it while I was sleeping.
2269         */
2270        WARN_ON_ONCE(!list_empty(&worker->scheduled));
2271
2272        /*
2273         * Finish PREP stage.  We're guaranteed to have at least one idle
2274         * worker or that someone else has already assumed the manager
2275         * role.  This is where @worker starts participating in concurrency
2276         * management if applicable and concurrency management is restored
2277         * after being rebound.  See rebind_workers() for details.
2278         */
2279        worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2280
2281        do {
2282                struct work_struct *work =
2283                        list_first_entry(&pool->worklist,
2284                                         struct work_struct, entry);
2285
2286                if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2287                        /* optimization path, not strictly necessary */
2288                        process_one_work(worker, work);
2289                        if (unlikely(!list_empty(&worker->scheduled)))
2290                                process_scheduled_works(worker);
2291                } else {
2292                        move_linked_works(work, &worker->scheduled, NULL);
2293                        process_scheduled_works(worker);
2294                }
2295        } while (keep_working(pool));
2296
2297        worker_set_flags(worker, WORKER_PREP, false);
2298sleep:
2299        if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2300                goto recheck;
2301
2302        /*
2303         * pool->lock is held and there's no work to process and no need to
2304         * manage, sleep.  Workers are woken up only while holding
2305         * pool->lock or from local cpu, so setting the current state
2306         * before releasing pool->lock is enough to prevent losing any
2307         * event.
2308         */
2309        worker_enter_idle(worker);
2310        __set_current_state(TASK_INTERRUPTIBLE);
2311        spin_unlock_irq(&pool->lock);
2312        schedule();
2313        goto woke_up;
2314}
2315
2316/**
2317 * rescuer_thread - the rescuer thread function
2318 * @__rescuer: self
2319 *
2320 * Workqueue rescuer thread function.  There's one rescuer for each
2321 * workqueue which has WQ_MEM_RECLAIM set.
2322 *
2323 * Regular work processing on a pool may block trying to create a new
2324 * worker which uses GFP_KERNEL allocation which has slight chance of
2325 * developing into deadlock if some works currently on the same queue
2326 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2327 * the problem rescuer solves.
2328 *
2329 * When such condition is possible, the pool summons rescuers of all
2330 * workqueues which have works queued on the pool and let them process
2331 * those works so that forward progress can be guaranteed.
2332 *
2333 * This should happen rarely.
2334 */
2335static int rescuer_thread(void *__rescuer)
2336{
2337        struct worker *rescuer = __rescuer;
2338        struct workqueue_struct *wq = rescuer->rescue_wq;
2339        struct list_head *scheduled = &rescuer->scheduled;
2340
2341        set_user_nice(current, RESCUER_NICE_LEVEL);
2342
2343        /*
2344         * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2345         * doesn't participate in concurrency management.
2346         */
2347        rescuer->task->flags |= PF_WQ_WORKER;
2348repeat:
2349        set_current_state(TASK_INTERRUPTIBLE);
2350
2351        if (kthread_should_stop()) {
2352                __set_current_state(TASK_RUNNING);
2353                rescuer->task->flags &= ~PF_WQ_WORKER;
2354                return 0;
2355        }
2356
2357        /* see whether any pwq is asking for help */
2358        spin_lock_irq(&wq_mayday_lock);
2359
2360        while (!list_empty(&wq->maydays)) {
2361                struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2362                                        struct pool_workqueue, mayday_node);
2363                struct worker_pool *pool = pwq->pool;
2364                struct work_struct *work, *n;
2365
2366                __set_current_state(TASK_RUNNING);
2367                list_del_init(&pwq->mayday_node);
2368
2369                spin_unlock_irq(&wq_mayday_lock);
2370
2371                /* migrate to the target cpu if possible */
2372                worker_maybe_bind_and_lock(pool);
2373                rescuer->pool = pool;
2374
2375                /*
2376                 * Slurp in all works issued via this workqueue and
2377                 * process'em.
2378                 */
2379                WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2380                list_for_each_entry_safe(work, n, &pool->worklist, entry)
2381                        if (get_work_pwq(work) == pwq)
2382                                move_linked_works(work, scheduled, &n);
2383
2384                process_scheduled_works(rescuer);
2385
2386                /*
2387                 * Leave this pool.  If keep_working() is %true, notify a
2388                 * regular worker; otherwise, we end up with 0 concurrency
2389                 * and stalling the execution.
2390                 */
2391                if (keep_working(pool))
2392                        wake_up_worker(pool);
2393
2394                rescuer->pool = NULL;
2395                spin_unlock(&pool->lock);
2396                spin_lock(&wq_mayday_lock);
2397        }
2398
2399        spin_unlock_irq(&wq_mayday_lock);
2400
2401        /* rescuers should never participate in concurrency management */
2402        WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2403        schedule();
2404        goto repeat;
2405}
2406
2407struct wq_barrier {
2408        struct work_struct      work;
2409        struct completion       done;
2410};
2411
2412static void wq_barrier_func(struct work_struct *work)
2413{
2414        struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2415        complete(&barr->done);
2416}
2417
2418/**
2419 * insert_wq_barrier - insert a barrier work
2420 * @pwq: pwq to insert barrier into
2421 * @barr: wq_barrier to insert
2422 * @target: target work to attach @barr to
2423 * @worker: worker currently executing @target, NULL if @target is not executing
2424 *
2425 * @barr is linked to @target such that @barr is completed only after
2426 * @target finishes execution.  Please note that the ordering
2427 * guarantee is observed only with respect to @target and on the local
2428 * cpu.
2429 *
2430 * Currently, a queued barrier can't be canceled.  This is because
2431 * try_to_grab_pending() can't determine whether the work to be
2432 * grabbed is at the head of the queue and thus can't clear LINKED
2433 * flag of the previous work while there must be a valid next work
2434 * after a work with LINKED flag set.
2435 *
2436 * Note that when @worker is non-NULL, @target may be modified
2437 * underneath us, so we can't reliably determine pwq from @target.
2438 *
2439 * CONTEXT:
2440 * spin_lock_irq(pool->lock).
2441 */
2442static void insert_wq_barrier(struct pool_workqueue *pwq,
2443                              struct wq_barrier *barr,
2444                              struct work_struct *target, struct worker *worker)
2445{
2446        struct list_head *head;
2447        unsigned int linked = 0;
2448
2449        /*
2450         * debugobject calls are safe here even with pool->lock locked
2451         * as we know for sure that this will not trigger any of the
2452         * checks and call back into the fixup functions where we
2453         * might deadlock.
2454         */
2455        INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2456        __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2457        init_completion(&barr->done);
2458
2459        /*
2460         * If @target is currently being executed, schedule the
2461         * barrier to the worker; otherwise, put it after @target.
2462         */
2463        if (worker)
2464                head = worker->scheduled.next;
2465        else {
2466                unsigned long *bits = work_data_bits(target);
2467
2468                head = target->entry.next;
2469                /* there can already be other linked works, inherit and set */
2470                linked = *bits & WORK_STRUCT_LINKED;
2471                __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2472        }
2473
2474        debug_work_activate(&barr->work);
2475        insert_work(pwq, &barr->work, head,
2476                    work_color_to_flags(WORK_NO_COLOR) | linked);
2477}
2478
2479/**
2480 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2481 * @wq: workqueue being flushed
2482 * @flush_color: new flush color, < 0 for no-op
2483 * @work_color: new work color, < 0 for no-op
2484 *
2485 * Prepare pwqs for workqueue flushing.
2486 *
2487 * If @flush_color is non-negative, flush_color on all pwqs should be
2488 * -1.  If no pwq has in-flight commands at the specified color, all
2489 * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2490 * has in flight commands, its pwq->flush_color is set to
2491 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2492 * wakeup logic is armed and %true is returned.
2493 *
2494 * The caller should have initialized @wq->first_flusher prior to
2495 * calling this function with non-negative @flush_color.  If
2496 * @flush_color is negative, no flush color update is done and %false
2497 * is returned.
2498 *
2499 * If @work_color is non-negative, all pwqs should have the same
2500 * work_color which is previous to @work_color and all will be
2501 * advanced to @work_color.
2502 *
2503 * CONTEXT:
2504 * mutex_lock(wq->mutex).
2505 *
2506 * RETURNS:
2507 * %true if @flush_color >= 0 and there's something to flush.  %false
2508 * otherwise.
2509 */
2510static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2511                                      int flush_color, int work_color)
2512{
2513        bool wait = false;
2514        struct pool_workqueue *pwq;
2515
2516        if (flush_color >= 0) {
2517                WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2518                atomic_set(&wq->nr_pwqs_to_flush, 1);
2519        }
2520
2521        for_each_pwq(pwq, wq) {
2522                struct worker_pool *pool = pwq->pool;
2523
2524                spin_lock_irq(&pool->lock);
2525
2526                if (flush_color >= 0) {
2527                        WARN_ON_ONCE(pwq->flush_color != -1);
2528
2529                        if (pwq->nr_in_flight[flush_color]) {
2530                                pwq->flush_color = flush_color;
2531                                atomic_inc(&wq->nr_pwqs_to_flush);
2532                                wait = true;
2533                        }
2534                }
2535
2536                if (work_color >= 0) {
2537                        WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2538                        pwq->work_color = work_color;
2539                }
2540
2541                spin_unlock_irq(&pool->lock);
2542        }
2543
2544        if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2545                complete(&wq->first_flusher->done);
2546
2547        return wait;
2548}
2549
2550/**
2551 * flush_workqueue - ensure that any scheduled work has run to completion.
2552 * @wq: workqueue to flush
2553 *
2554 * This function sleeps until all work items which were queued on entry
2555 * have finished execution, but it is not livelocked by new incoming ones.
2556 */
2557void flush_workqueue(struct workqueue_struct *wq)
2558{
2559        struct wq_flusher this_flusher = {
2560                .list = LIST_HEAD_INIT(this_flusher.list),
2561                .flush_color = -1,
2562                .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2563        };
2564        int next_color;
2565
2566        lock_map_acquire(&wq->lockdep_map);
2567        lock_map_release(&wq->lockdep_map);
2568
2569        mutex_lock(&wq->mutex);
2570
2571        /*
2572         * Start-to-wait phase
2573         */
2574        next_color = work_next_color(wq->work_color);
2575
2576        if (next_color != wq->flush_color) {
2577                /*
2578                 * Color space is not full.  The current work_color
2579                 * becomes our flush_color and work_color is advanced
2580                 * by one.
2581                 */
2582                WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2583                this_flusher.flush_color = wq->work_color;
2584                wq->work_color = next_color;
2585
2586                if (!wq->first_flusher) {
2587                        /* no flush in progress, become the first flusher */
2588                        WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2589
2590                        wq->first_flusher = &this_flusher;
2591
2592                        if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2593                                                       wq->work_color)) {
2594                                /* nothing to flush, done */
2595                                wq->flush_color = next_color;
2596                                wq->first_flusher = NULL;
2597                                goto out_unlock;
2598                        }
2599                } else {
2600                        /* wait in queue */
2601                        WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2602                        list_add_tail(&this_flusher.list, &wq->flusher_queue);
2603                        flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2604                }
2605        } else {
2606                /*
2607                 * Oops, color space is full, wait on overflow queue.
2608                 * The next flush completion will assign us
2609                 * flush_color and transfer to flusher_queue.
2610                 */
2611                list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2612        }
2613
2614        mutex_unlock(&wq->mutex);
2615
2616        wait_for_completion(&this_flusher.done);
2617
2618        /*
2619         * Wake-up-and-cascade phase
2620         *
2621         * First flushers are responsible for cascading flushes and
2622         * handling overflow.  Non-first flushers can simply return.
2623         */
2624        if (wq->first_flusher != &this_flusher)
2625                return;
2626
2627        mutex_lock(&wq->mutex);
2628
2629        /* we might have raced, check again with mutex held */
2630        if (wq->first_flusher != &this_flusher)
2631                goto out_unlock;
2632
2633        wq->first_flusher = NULL;
2634
2635        WARN_ON_ONCE(!list_empty(&this_flusher.list));
2636        WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2637
2638        while (true) {
2639                struct wq_flusher *next, *tmp;
2640
2641                /* complete all the flushers sharing the current flush color */
2642                list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2643                        if (next->flush_color != wq->flush_color)
2644                                break;
2645                        list_del_init(&next->list);
2646                        complete(&next->done);
2647                }
2648
2649                WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2650                             wq->flush_color != work_next_color(wq->work_color));
2651
2652                /* this flush_color is finished, advance by one */
2653                wq->flush_color = work_next_color(wq->flush_color);
2654
2655                /* one color has been freed, handle overflow queue */
2656                if (!list_empty(&wq->flusher_overflow)) {
2657                        /*
2658                         * Assign the same color to all overflowed
2659                         * flushers, advance work_color and append to
2660                         * flusher_queue.  This is the start-to-wait
2661                         * phase for these overflowed flushers.
2662                         */
2663                        list_for_each_entry(tmp, &wq->flusher_overflow, list)
2664                                tmp->flush_color = wq->work_color;
2665
2666                        wq->work_color = work_next_color(wq->work_color);
2667
2668                        list_splice_tail_init(&wq->flusher_overflow,
2669                                              &wq->flusher_queue);
2670                        flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2671                }
2672
2673                if (list_empty(&wq->flusher_queue)) {
2674                        WARN_ON_ONCE(wq->flush_color != wq->work_color);
2675                        break;
2676                }
2677
2678                /*
2679                 * Need to flush more colors.  Make the next flusher
2680                 * the new first flusher and arm pwqs.
2681                 */
2682                WARN_ON_ONCE(wq->flush_color == wq->work_color);
2683                WARN_ON_ONCE(wq->flush_color != next->flush_color);
2684
2685                list_del_init(&next->list);
2686                wq->first_flusher = next;
2687
2688                if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2689                        break;
2690
2691                /*
2692                 * Meh... this color is already done, clear first
2693                 * flusher and repeat cascading.
2694                 */
2695                wq->first_flusher = NULL;
2696        }
2697
2698out_unlock:
2699        mutex_unlock(&wq->mutex);
2700}
2701EXPORT_SYMBOL_GPL(flush_workqueue);
2702
2703/**
2704 * drain_workqueue - drain a workqueue
2705 * @wq: workqueue to drain
2706 *
2707 * Wait until the workqueue becomes empty.  While draining is in progress,
2708 * only chain queueing is allowed.  IOW, only currently pending or running
2709 * work items on @wq can queue further work items on it.  @wq is flushed
2710 * repeatedly until it becomes empty.  The number of flushing is detemined
2711 * by the depth of chaining and should be relatively short.  Whine if it
2712 * takes too long.
2713 */
2714void drain_workqueue(struct workqueue_struct *wq)
2715{
2716        unsigned int flush_cnt = 0;
2717        struct pool_workqueue *pwq;
2718
2719        /*
2720         * __queue_work() needs to test whether there are drainers, is much
2721         * hotter than drain_workqueue() and already looks at @wq->flags.
2722         * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2723         */
2724        mutex_lock(&wq->mutex);
2725        if (!wq->nr_drainers++)
2726                wq->flags |= __WQ_DRAINING;
2727        mutex_unlock(&wq->mutex);
2728reflush:
2729        flush_workqueue(wq);
2730
2731        mutex_lock(&wq->mutex);
2732
2733        for_each_pwq(pwq, wq) {
2734                bool drained;
2735
2736                spin_lock_irq(&pwq->pool->lock);
2737                drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2738                spin_unlock_irq(&pwq->pool->lock);
2739
2740                if (drained)
2741                        continue;
2742
2743                if (++flush_cnt == 10 ||
2744                    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2745                        pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2746                                wq->name, flush_cnt);
2747
2748                mutex_unlock(&wq->mutex);
2749                goto reflush;
2750        }
2751
2752        if (!--wq->nr_drainers)
2753                wq->flags &= ~__WQ_DRAINING;
2754        mutex_unlock(&wq->mutex);
2755}
2756EXPORT_SYMBOL_GPL(drain_workqueue);
2757
2758static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2759{
2760        struct worker *worker = NULL;
2761        struct worker_pool *pool;
2762        struct pool_workqueue *pwq;
2763
2764        might_sleep();
2765
2766        local_irq_disable();
2767        pool = get_work_pool(work);
2768        if (!pool) {
2769                local_irq_enable();
2770                return false;
2771        }
2772
2773        spin_lock(&pool->lock);
2774        /* see the comment in try_to_grab_pending() with the same code */
2775        pwq = get_work_pwq(work);
2776        if (pwq) {
2777                if (unlikely(pwq->pool != pool))
2778                        goto already_gone;
2779        } else {
2780                worker = find_worker_executing_work(pool, work);
2781                if (!worker)
2782                        goto already_gone;
2783                pwq = worker->current_pwq;
2784        }
2785
2786        insert_wq_barrier(pwq, barr, work, worker);
2787        spin_unlock_irq(&pool->lock);
2788
2789        /*
2790         * If @max_active is 1 or rescuer is in use, flushing another work
2791         * item on the same workqueue may lead to deadlock.  Make sure the
2792         * flusher is not running on the same workqueue by verifying write
2793         * access.
2794         */
2795        if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2796                lock_map_acquire(&pwq->wq->lockdep_map);
2797        else
2798                lock_map_acquire_read(&pwq->wq->lockdep_map);
2799        lock_map_release(&pwq->wq->lockdep_map);
2800
2801        return true;
2802already_gone:
2803        spin_unlock_irq(&pool->lock);
2804        return false;
2805}
2806
2807/**
2808 * flush_work - wait for a work to finish executing the last queueing instance
2809 * @work: the work to flush
2810 *
2811 * Wait until @work has finished execution.  @work is guaranteed to be idle
2812 * on return if it hasn't been requeued since flush started.
2813 *
2814 * RETURNS:
2815 * %true if flush_work() waited for the work to finish execution,
2816 * %false if it was already idle.
2817 */
2818bool flush_work(struct work_struct *work)
2819{
2820        struct wq_barrier barr;
2821
2822        lock_map_acquire(&work->lockdep_map);
2823        lock_map_release(&work->lockdep_map);
2824
2825        if (start_flush_work(work, &barr)) {
2826                wait_for_completion(&barr.done);
2827                destroy_work_on_stack(&barr.work);
2828                return true;
2829        } else {
2830                return false;
2831        }
2832}
2833EXPORT_SYMBOL_GPL(flush_work);
2834
2835static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2836{
2837        unsigned long flags;
2838        int ret;
2839
2840        do {
2841                ret = try_to_grab_pending(work, is_dwork, &flags);
2842                /*
2843                 * If someone else is canceling, wait for the same event it
2844                 * would be waiting for before retrying.
2845                 */
2846                if (unlikely(ret == -ENOENT))
2847                        flush_work(work);
2848        } while (unlikely(ret < 0));
2849
2850        /* tell other tasks trying to grab @work to back off */
2851        mark_work_canceling(work);
2852        local_irq_restore(flags);
2853
2854        flush_work(work);
2855        clear_work_data(work);
2856        return ret;
2857}
2858
2859/**
2860 * cancel_work_sync - cancel a work and wait for it to finish
2861 * @work: the work to cancel
2862 *
2863 * Cancel @work and wait for its execution to finish.  This function
2864 * can be used even if the work re-queues itself or migrates to
2865 * another workqueue.  On return from this function, @work is
2866 * guaranteed to be not pending or executing on any CPU.
2867 *
2868 * cancel_work_sync(&delayed_work->work) must not be used for
2869 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2870 *
2871 * The caller must ensure that the workqueue on which @work was last
2872 * queued can't be destroyed before this function returns.
2873 *
2874 * RETURNS:
2875 * %true if @work was pending, %false otherwise.
2876 */
2877bool cancel_work_sync(struct work_struct *work)
2878{
2879        return __cancel_work_timer(work, false);
2880}
2881EXPORT_SYMBOL_GPL(cancel_work_sync);
2882
2883/**
2884 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2885 * @dwork: the delayed work to flush
2886 *
2887 * Delayed timer is cancelled and the pending work is queued for
2888 * immediate execution.  Like flush_work(), this function only
2889 * considers the last queueing instance of @dwork.
2890 *
2891 * RETURNS:
2892 * %true if flush_work() waited for the work to finish execution,
2893 * %false if it was already idle.
2894 */
2895bool flush_delayed_work(struct delayed_work *dwork)
2896{
2897        local_irq_disable();
2898        if (del_timer_sync(&dwork->timer))
2899                __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2900        local_irq_enable();
2901        return flush_work(&dwork->work);
2902}
2903EXPORT_SYMBOL(flush_delayed_work);
2904
2905/**
2906 * cancel_delayed_work - cancel a delayed work
2907 * @dwork: delayed_work to cancel
2908 *
2909 * Kill off a pending delayed_work.  Returns %true if @dwork was pending
2910 * and canceled; %false if wasn't pending.  Note that the work callback
2911 * function may still be running on return, unless it returns %true and the
2912 * work doesn't re-arm itself.  Explicitly flush or use
2913 * cancel_delayed_work_sync() to wait on it.
2914 *
2915 * This function is safe to call from any context including IRQ handler.
2916 */
2917bool cancel_delayed_work(struct delayed_work *dwork)
2918{
2919        unsigned long flags;
2920        int ret;
2921
2922        do {
2923                ret = try_to_grab_pending(&dwork->work, true, &flags);
2924        } while (unlikely(ret == -EAGAIN));
2925
2926        if (unlikely(ret < 0))
2927                return false;
2928
2929        set_work_pool_and_clear_pending(&dwork->work,
2930                                        get_work_pool_id(&dwork->work));
2931        local_irq_restore(flags);
2932        return ret;
2933}
2934EXPORT_SYMBOL(cancel_delayed_work);
2935
2936/**
2937 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2938 * @dwork: the delayed work cancel
2939 *
2940 * This is cancel_work_sync() for delayed works.
2941 *
2942 * RETURNS:
2943 * %true if @dwork was pending, %false otherwise.
2944 */
2945bool cancel_delayed_work_sync(struct delayed_work *dwork)
2946{
2947        return __cancel_work_timer(&dwork->work, true);
2948}
2949EXPORT_SYMBOL(cancel_delayed_work_sync);
2950
2951/**
2952 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2953 * @func: the function to call
2954 *
2955 * schedule_on_each_cpu() executes @func on each online CPU using the
2956 * system workqueue and blocks until all CPUs have completed.
2957 * schedule_on_each_cpu() is very slow.
2958 *
2959 * RETURNS:
2960 * 0 on success, -errno on failure.
2961 */
2962int schedule_on_each_cpu(work_func_t func)
2963{
2964        int cpu;
2965        struct work_struct __percpu *works;
2966
2967        works = alloc_percpu(struct work_struct);
2968        if (!works)
2969                return -ENOMEM;
2970
2971        get_online_cpus();
2972
2973        for_each_online_cpu(cpu) {
2974                struct work_struct *work = per_cpu_ptr(works, cpu);
2975
2976                INIT_WORK(work, func);
2977                schedule_work_on(cpu, work);
2978        }
2979
2980        for_each_online_cpu(cpu)
2981                flush_work(per_cpu_ptr(works, cpu));
2982
2983        put_online_cpus();
2984        free_percpu(works);
2985        return 0;
2986}
2987
2988/**
2989 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2990 *
2991 * Forces execution of the kernel-global workqueue and blocks until its
2992 * completion.
2993 *
2994 * Think twice before calling this function!  It's very easy to get into
2995 * trouble if you don't take great care.  Either of the following situations
2996 * will lead to deadlock:
2997 *
2998 *      One of the work items currently on the workqueue needs to acquire
2999 *      a lock held by your code or its caller.
3000 *
3001 *      Your code is running in the context of a work routine.
3002 *
3003 * They will be detected by lockdep when they occur, but the first might not
3004 * occur very often.  It depends on what work items are on the workqueue and
3005 * what locks they need, which you have no control over.
3006 *
3007 * In most situations flushing the entire workqueue is overkill; you merely
3008 * need to know that a particular work item isn't queued and isn't running.
3009 * In such cases you should use cancel_delayed_work_sync() or
3010 * cancel_work_sync() instead.
3011 */
3012void flush_scheduled_work(void)
3013{
3014        flush_workqueue(system_wq);
3015}
3016EXPORT_SYMBOL(flush_scheduled_work);
3017
3018/**
3019 * execute_in_process_context - reliably execute the routine with user context
3020 * @fn:         the function to execute
3021 * @ew:         guaranteed storage for the execute work structure (must
3022 *              be available when the work executes)
3023 *
3024 * Executes the function immediately if process context is available,
3025 * otherwise schedules the function for delayed execution.
3026 *
3027 * Returns:     0 - function was executed
3028 *              1 - function was scheduled for execution
3029 */
3030int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3031{
3032        if (!in_interrupt()) {
3033                fn(&ew->work);
3034                return 0;
3035        }
3036
3037        INIT_WORK(&ew->work, fn);
3038        schedule_work(&ew->work);
3039
3040        return 1;
3041}
3042EXPORT_SYMBOL_GPL(execute_in_process_context);
3043
3044#ifdef CONFIG_SYSFS
3045/*
3046 * Workqueues with WQ_SYSFS flag set is visible to userland via
3047 * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
3048 * following attributes.
3049 *
3050 *  per_cpu     RO bool : whether the workqueue is per-cpu or unbound
3051 *  max_active  RW int  : maximum number of in-flight work items
3052 *
3053 * Unbound workqueues have the following extra attributes.
3054 *
3055 *  id          RO int  : the associated pool ID
3056 *  nice        RW int  : nice value of the workers
3057 *  cpumask     RW mask : bitmask of allowed CPUs for the workers
3058 */
3059struct wq_device {
3060        struct workqueue_struct         *wq;
3061        struct device                   dev;
3062};
3063
3064static struct workqueue_struct *dev_to_wq(struct device *dev)
3065{
3066        struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3067
3068        return wq_dev->wq;
3069}
3070
3071static ssize_t wq_per_cpu_show(struct device *dev,
3072                               struct device_attribute *attr, char *buf)
3073{
3074        struct workqueue_struct *wq = dev_to_wq(dev);
3075
3076        return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3077}
3078
3079static ssize_t wq_max_active_show(struct device *dev,
3080                                  struct device_attribute *attr, char *buf)
3081{
3082        struct workqueue_struct *wq = dev_to_wq(dev);
3083
3084        return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3085}
3086
3087static ssize_t wq_max_active_store(struct device *dev,
3088                                   struct device_attribute *attr,
3089                                   const char *buf, size_t count)
3090{
3091        struct workqueue_struct *wq = dev_to_wq(dev);
3092        int val;
3093
3094        if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3095                return -EINVAL;
3096
3097        workqueue_set_max_active(wq, val);
3098        return count;
3099}
3100
3101static struct device_attribute wq_sysfs_attrs[] = {
3102        __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3103        __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3104        __ATTR_NULL,
3105};
3106
3107static ssize_t wq_pool_ids_show(struct device *dev,
3108                                struct device_attribute *attr, char *buf)
3109{
3110        struct workqueue_struct *wq = dev_to_wq(dev);
3111        const char *delim = "";
3112        int node, written = 0;
3113
3114        rcu_read_lock_sched();
3115        for_each_node(node) {
3116                written += scnprintf(buf + written, PAGE_SIZE - written,
3117                                     "%s%d:%d", delim, node,
3118                                     unbound_pwq_by_node(wq, node)->pool->id);
3119                delim = " ";
3120        }
3121        written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3122        rcu_read_unlock_sched();
3123
3124        return written;
3125}
3126
3127static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3128                            char *buf)
3129{
3130        struct workqueue_struct *wq = dev_to_wq(dev);
3131        int written;
3132
3133        mutex_lock(&wq->mutex);
3134        written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3135        mutex_unlock(&wq->mutex);
3136
3137        return written;
3138}
3139
3140/* prepare workqueue_attrs for sysfs store operations */
3141static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3142{
3143        struct workqueue_attrs *attrs;
3144
3145        attrs = alloc_workqueue_attrs(GFP_KERNEL);
3146        if (!attrs)
3147                return NULL;
3148
3149        mutex_lock(&wq->mutex);
3150        copy_workqueue_attrs(attrs, wq->unbound_attrs);
3151        mutex_unlock(&wq->mutex);
3152        return attrs;
3153}
3154
3155static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3156                             const char *buf, size_t count)
3157{
3158        struct workqueue_struct *wq = dev_to_wq(dev);
3159        struct workqueue_attrs *attrs;
3160        int ret;
3161
3162        attrs = wq_sysfs_prep_attrs(wq);
3163        if (!attrs)
3164                return -ENOMEM;
3165
3166        if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3167            attrs->nice >= -20 && attrs->nice <= 19)
3168                ret = apply_workqueue_attrs(wq, attrs);
3169        else
3170                ret = -EINVAL;
3171
3172        free_workqueue_attrs(attrs);
3173        return ret ?: count;
3174}
3175
3176static ssize_t wq_cpumask_show(struct device *dev,
3177                               struct device_attribute *attr, char *buf)
3178{
3179        struct workqueue_struct *wq = dev_to_wq(dev);
3180        int written;
3181
3182        mutex_lock(&wq->mutex);
3183        written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3184        mutex_unlock(&wq->mutex);
3185
3186        written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3187        return written;
3188}
3189
3190static ssize_t wq_cpumask_store(struct device *dev,
3191                                struct device_attribute *attr,
3192                                const char *buf, size_t count)
3193{
3194        struct workqueue_struct *wq = dev_to_wq(dev);
3195        struct workqueue_attrs *attrs;
3196        int ret;
3197
3198        attrs = wq_sysfs_prep_attrs(wq);
3199        if (!attrs)
3200                return -ENOMEM;
3201
3202        ret = cpumask_parse(buf, attrs->cpumask);
3203        if (!ret)
3204                ret = apply_workqueue_attrs(wq, attrs);
3205
3206        free_workqueue_attrs(attrs);
3207        return ret ?: count;
3208}
3209
3210static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3211                            char *buf)
3212{
3213        struct workqueue_struct *wq = dev_to_wq(dev);
3214        int written;
3215
3216        mutex_lock(&wq->mutex);
3217        written = scnprintf(buf, PAGE_SIZE, "%d\n",
3218                            !wq->unbound_attrs->no_numa);
3219        mutex_unlock(&wq->mutex);
3220
3221        return written;
3222}
3223
3224static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3225                             const char *buf, size_t count)
3226{
3227        struct workqueue_struct *wq = dev_to_wq(dev);
3228        struct workqueue_attrs *attrs;
3229        int v, ret;
3230
3231        attrs = wq_sysfs_prep_attrs(wq);
3232        if (!attrs)
3233                return -ENOMEM;
3234
3235        ret = -EINVAL;
3236        if (sscanf(buf, "%d", &v) == 1) {
3237                attrs->no_numa = !v;
3238                ret = apply_workqueue_attrs(wq, attrs);
3239        }
3240
3241        free_workqueue_attrs(attrs);
3242        return ret ?: count;
3243}
3244
3245static struct device_attribute wq_sysfs_unbound_attrs[] = {
3246        __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3247        __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3248        __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3249        __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3250        __ATTR_NULL,
3251};
3252
3253static struct bus_type wq_subsys = {
3254        .name                           = "workqueue",
3255        .dev_attrs                      = wq_sysfs_attrs,
3256};
3257
3258static int __init wq_sysfs_init(void)
3259{
3260        return subsys_virtual_register(&wq_subsys, NULL);
3261}
3262core_initcall(wq_sysfs_init);
3263
3264static void wq_device_release(struct device *dev)
3265{
3266        struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3267
3268        kfree(wq_dev);
3269}
3270
3271/**
3272 * workqueue_sysfs_register - make a workqueue visible in sysfs
3273 * @wq: the workqueue to register
3274 *
3275 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3276 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3277 * which is the preferred method.
3278 *
3279 * Workqueue user should use this function directly iff it wants to apply
3280 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3281 * apply_workqueue_attrs() may race against userland updating the
3282 * attributes.
3283 *
3284 * Returns 0 on success, -errno on failure.
3285 */
3286int workqueue_sysfs_register(struct workqueue_struct *wq)
3287{
3288        struct wq_device *wq_dev;
3289        int ret;
3290
3291        /*
3292         * Adjusting max_active or creating new pwqs by applyting
3293         * attributes breaks ordering guarantee.  Disallow exposing ordered
3294         * workqueues.
3295         */
3296        if (WARN_ON(wq->flags & __WQ_ORDERED))
3297                return -EINVAL;
3298
3299        wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3300        if (!wq_dev)
3301                return -ENOMEM;
3302
3303        wq_dev->wq = wq;
3304        wq_dev->dev.bus = &wq_subsys;
3305        wq_dev->dev.init_name = wq->name;
3306        wq_dev->dev.release = wq_device_release;
3307
3308        /*
3309         * unbound_attrs are created separately.  Suppress uevent until
3310         * everything is ready.
3311         */
3312        dev_set_uevent_suppress(&wq_dev->dev, true);
3313
3314        ret = device_register(&wq_dev->dev);
3315        if (ret) {
3316                kfree(wq_dev);
3317                wq->wq_dev = NULL;
3318                return ret;
3319        }
3320
3321        if (wq->flags & WQ_UNBOUND) {
3322                struct device_attribute *attr;
3323
3324                for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3325                        ret = device_create_file(&wq_dev->dev, attr);
3326                        if (ret) {
3327                                device_unregister(&wq_dev->dev);
3328                                wq->wq_dev = NULL;
3329                                return ret;
3330                        }
3331                }
3332        }
3333
3334        kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3335        return 0;
3336}
3337
3338/**
3339 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3340 * @wq: the workqueue to unregister
3341 *
3342 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3343 */
3344static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3345{
3346        struct wq_device *wq_dev = wq->wq_dev;
3347
3348        if (!wq->wq_dev)
3349                return;
3350
3351        wq->wq_dev = NULL;
3352        device_unregister(&wq_dev->dev);
3353}
3354#else   /* CONFIG_SYSFS */
3355static void workqueue_sysfs_unregister(struct workqueue_struct *wq)     { }
3356#endif  /* CONFIG_SYSFS */
3357
3358/**
3359 * free_workqueue_attrs - free a workqueue_attrs
3360 * @attrs: workqueue_attrs to free
3361 *
3362 * Undo alloc_workqueue_attrs().
3363 */
3364void free_workqueue_attrs(struct workqueue_attrs *attrs)
3365{
3366        if (attrs) {
3367                free_cpumask_var(attrs->cpumask);
3368                kfree(attrs);
3369        }
3370}
3371
3372/**
3373 * alloc_workqueue_attrs - allocate a workqueue_attrs
3374 * @gfp_mask: allocation mask to use
3375 *
3376 * Allocate a new workqueue_attrs, initialize with default settings and
3377 * return it.  Returns NULL on failure.
3378 */
3379struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3380{
3381        struct workqueue_attrs *attrs;
3382
3383        attrs = kzalloc(sizeof(*attrs), gfp_mask);
3384        if (!attrs)
3385                goto fail;
3386        if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3387                goto fail;
3388
3389        cpumask_copy(attrs->cpumask, cpu_possible_mask);
3390        return attrs;
3391fail:
3392        free_workqueue_attrs(attrs);
3393        return NULL;
3394}
3395
3396static void copy_workqueue_attrs(struct workqueue_attrs *to,
3397                                 const struct workqueue_attrs *from)
3398{
3399        to->nice = from->nice;
3400        cpumask_copy(to->cpumask, from->cpumask);
3401}
3402
3403/* hash value of the content of @attr */
3404static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3405{
3406        u32 hash = 0;
3407
3408        hash = jhash_1word(attrs->nice, hash);
3409        hash = jhash(cpumask_bits(attrs->cpumask),
3410                     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3411        return hash;
3412}
3413
3414/* content equality test */
3415static bool wqattrs_equal(const struct workqueue_attrs *a,
3416                          const struct workqueue_attrs *b)
3417{
3418        if (a->nice != b->nice)
3419                return false;
3420        if (!cpumask_equal(a->cpumask, b->cpumask))
3421                return false;
3422        return true;
3423}
3424
3425/**
3426 * init_worker_pool - initialize a newly zalloc'd worker_pool
3427 * @pool: worker_pool to initialize
3428 *
3429 * Initiailize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3430 * Returns 0 on success, -errno on failure.  Even on failure, all fields
3431 * inside @pool proper are initialized and put_unbound_pool() can be called
3432 * on @pool safely to release it.
3433 */
3434static int init_worker_pool(struct worker_pool *pool)
3435{
3436        spin_lock_init(&pool->lock);
3437        pool->id = -1;
3438        pool->cpu = -1;
3439        pool->node = NUMA_NO_NODE;
3440        pool->flags |= POOL_DISASSOCIATED;
3441        INIT_LIST_HEAD(&pool->worklist);
3442        INIT_LIST_HEAD(&pool->idle_list);
3443        hash_init(pool->busy_hash);
3444
3445        init_timer_deferrable(&pool->idle_timer);
3446        pool->idle_timer.function = idle_worker_timeout;
3447        pool->idle_timer.data = (unsigned long)pool;
3448
3449        setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3450                    (unsigned long)pool);
3451
3452        mutex_init(&pool->manager_arb);
3453        mutex_init(&pool->manager_mutex);
3454        idr_init(&pool->worker_idr);
3455
3456        INIT_HLIST_NODE(&pool->hash_node);
3457        pool->refcnt = 1;
3458
3459        /* shouldn't fail above this point */
3460        pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3461        if (!pool->attrs)
3462                return -ENOMEM;
3463        return 0;
3464}
3465
3466static void rcu_free_pool(struct rcu_head *rcu)
3467{
3468        struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3469
3470        idr_destroy(&pool->worker_idr);
3471        free_workqueue_attrs(pool->attrs);
3472        kfree(pool);
3473}
3474
3475/**
3476 * put_unbound_pool - put a worker_pool
3477 * @pool: worker_pool to put
3478 *
3479 * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3480 * safe manner.  get_unbound_pool() calls this function on its failure path
3481 * and this function should be able to release pools which went through,
3482 * successfully or not, init_worker_pool().
3483 *
3484 * Should be called with wq_pool_mutex held.
3485 */
3486static void put_unbound_pool(struct worker_pool *pool)
3487{
3488        struct worker *worker;
3489
3490        lockdep_assert_held(&wq_pool_mutex);
3491
3492        if (--pool->refcnt)
3493                return;
3494
3495        /* sanity checks */
3496        if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3497            WARN_ON(!list_empty(&pool->worklist)))
3498                return;
3499
3500        /* release id and unhash */
3501        if (pool->id >= 0)
3502                idr_remove(&worker_pool_idr, pool->id);
3503        hash_del(&pool->hash_node);
3504
3505        /*
3506         * Become the manager and destroy all workers.  Grabbing
3507         * manager_arb prevents @pool's workers from blocking on
3508         * manager_mutex.
3509         */
3510        mutex_lock(&pool->manager_arb);
3511        mutex_lock(&pool->manager_mutex);
3512        spin_lock_irq(&pool->lock);
3513
3514        while ((worker = first_worker(pool)))
3515                destroy_worker(worker);
3516        WARN_ON(pool->nr_workers || pool->nr_idle);
3517
3518        spin_unlock_irq(&pool->lock);
3519        mutex_unlock(&pool->manager_mutex);
3520        mutex_unlock(&pool->manager_arb);
3521
3522        /* shut down the timers */
3523        del_timer_sync(&pool->idle_timer);
3524        del_timer_sync(&pool->mayday_timer);
3525
3526        /* sched-RCU protected to allow dereferences from get_work_pool() */
3527        call_rcu_sched(&pool->rcu, rcu_free_pool);
3528}
3529
3530/**
3531 * get_unbound_pool - get a worker_pool with the specified attributes
3532 * @attrs: the attributes of the worker_pool to get
3533 *
3534 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3535 * reference count and return it.  If there already is a matching
3536 * worker_pool, it will be used; otherwise, this function attempts to
3537 * create a new one.  On failure, returns NULL.
3538 *
3539 * Should be called with wq_pool_mutex held.
3540 */
3541static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3542{
3543        u32 hash = wqattrs_hash(attrs);
3544        struct worker_pool *pool;
3545        int node;
3546
3547        lockdep_assert_held(&wq_pool_mutex);
3548
3549        /* do we already have a matching pool? */
3550        hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3551                if (wqattrs_equal(pool->attrs, attrs)) {
3552                        pool->refcnt++;
3553                        goto out_unlock;
3554                }
3555        }
3556
3557        /* nope, create a new one */
3558        pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3559        if (!pool || init_worker_pool(pool) < 0)
3560                goto fail;
3561
3562        if (workqueue_freezing)
3563                pool->flags |= POOL_FREEZING;
3564
3565        lockdep_set_subclass(&pool->lock, 1);   /* see put_pwq() */
3566        copy_workqueue_attrs(pool->attrs, attrs);
3567
3568        /* if cpumask is contained inside a NUMA node, we belong to that node */
3569        if (wq_numa_enabled) {
3570                for_each_node(node) {
3571                        if (cpumask_subset(pool->attrs->cpumask,
3572                                           wq_numa_possible_cpumask[node])) {
3573                                pool->node = node;
3574                                break;
3575                        }
3576                }
3577        }
3578
3579        if (worker_pool_assign_id(pool) < 0)
3580                goto fail;
3581
3582        /* create and start the initial worker */
3583        if (create_and_start_worker(pool) < 0)
3584                goto fail;
3585
3586        /* install */
3587        hash_add(unbound_pool_hash, &pool->hash_node, hash);
3588out_unlock:
3589        return pool;
3590fail:
3591        if (pool)
3592                put_unbound_pool(pool);
3593        return NULL;
3594}
3595
3596static void rcu_free_pwq(struct rcu_head *rcu)
3597{
3598        kmem_cache_free(pwq_cache,
3599                        container_of(rcu, struct pool_workqueue, rcu));
3600}
3601
3602/*
3603 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3604 * and needs to be destroyed.
3605 */
3606static void pwq_unbound_release_workfn(struct work_struct *work)
3607{
3608        struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3609                                                  unbound_release_work);
3610        struct workqueue_struct *wq = pwq->wq;
3611        struct worker_pool *pool = pwq->pool;
3612        bool is_last;
3613
3614        if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3615                return;
3616
3617        /*
3618         * Unlink @pwq.  Synchronization against wq->mutex isn't strictly
3619         * necessary on release but do it anyway.  It's easier to verify
3620         * and consistent with the linking path.
3621         */
3622        mutex_lock(&wq->mutex);
3623        list_del_rcu(&pwq->pwqs_node);
3624        is_last = list_empty(&wq->pwqs);
3625        mutex_unlock(&wq->mutex);
3626
3627        mutex_lock(&wq_pool_mutex);
3628        put_unbound_pool(pool);
3629        mutex_unlock(&wq_pool_mutex);
3630
3631        call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3632
3633        /*
3634         * If we're the last pwq going away, @wq is already dead and no one
3635         * is gonna access it anymore.  Free it.
3636         */
3637        if (is_last) {
3638                free_workqueue_attrs(wq->unbound_attrs);
3639                kfree(wq);
3640        }
3641}
3642
3643/**
3644 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3645 * @pwq: target pool_workqueue
3646 *
3647 * If @pwq isn't freezing, set @pwq->max_active to the associated
3648 * workqueue's saved_max_active and activate delayed work items
3649 * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3650 */
3651static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3652{
3653        struct workqueue_struct *wq = pwq->wq;
3654        bool freezable = wq->flags & WQ_FREEZABLE;
3655
3656        /* for @wq->saved_max_active */
3657        lockdep_assert_held(&wq->mutex);
3658
3659        /* fast exit for non-freezable wqs */
3660        if (!freezable && pwq->max_active == wq->saved_max_active)
3661                return;
3662
3663        spin_lock_irq(&pwq->pool->lock);
3664
3665        if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3666                pwq->max_active = wq->saved_max_active;
3667
3668                while (!list_empty(&pwq->delayed_works) &&
3669                       pwq->nr_active < pwq->max_active)
3670                        pwq_activate_first_delayed(pwq);
3671
3672                /*
3673                 * Need to kick a worker after thawed or an unbound wq's
3674                 * max_active is bumped.  It's a slow path.  Do it always.
3675                 */
3676                wake_up_worker(pwq->pool);
3677        } else {
3678                pwq->max_active = 0;
3679        }
3680
3681        spin_unlock_irq(&pwq->pool->lock);
3682}
3683
3684/* initialize newly alloced @pwq which is associated with @wq and @pool */
3685static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3686                     struct worker_pool *pool)
3687{
3688        BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3689
3690        memset(pwq, 0, sizeof(*pwq));
3691
3692        pwq->pool = pool;
3693        pwq->wq = wq;
3694        pwq->flush_color = -1;
3695        pwq->refcnt = 1;
3696        INIT_LIST_HEAD(&pwq->delayed_works);
3697        INIT_LIST_HEAD(&pwq->pwqs_node);
3698        INIT_LIST_HEAD(&pwq->mayday_node);
3699        INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3700}
3701
3702/* sync @pwq with the current state of its associated wq and link it */
3703static void link_pwq(struct pool_workqueue *pwq)
3704{
3705        struct workqueue_struct *wq = pwq->wq;
3706
3707        lockdep_assert_held(&wq->mutex);
3708
3709        /* may be called multiple times, ignore if already linked */
3710        if (!list_empty(&pwq->pwqs_node))
3711                return;
3712
3713        /*
3714         * Set the matching work_color.  This is synchronized with
3715         * wq->mutex to avoid confusing flush_workqueue().
3716         */
3717        pwq->work_color = wq->work_color;
3718
3719        /* sync max_active to the current setting */
3720        pwq_adjust_max_active(pwq);
3721
3722        /* link in @pwq */
3723        list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3724}
3725
3726/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3727static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3728                                        const struct workqueue_attrs *attrs)
3729{
3730        struct worker_pool *pool;
3731        struct pool_workqueue *pwq;
3732
3733        lockdep_assert_held(&wq_pool_mutex);
3734
3735        pool = get_unbound_pool(attrs);
3736        if (!pool)
3737                return NULL;
3738
3739        pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3740        if (!pwq) {
3741                put_unbound_pool(pool);
3742                return NULL;
3743        }
3744
3745        init_pwq(pwq, wq, pool);
3746        return pwq;
3747}
3748
3749/* undo alloc_unbound_pwq(), used only in the error path */
3750static void free_unbound_pwq(struct pool_workqueue *pwq)
3751{
3752        lockdep_assert_held(&wq_pool_mutex);
3753
3754        if (pwq) {
3755                put_unbound_pool(pwq->pool);
3756                kmem_cache_free(pwq_cache, pwq);
3757        }
3758}
3759
3760/**
3761 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3762 * @attrs: the wq_attrs of interest
3763 * @node: the target NUMA node
3764 * @cpu_going_down: if >= 0, the CPU to consider as offline
3765 * @cpumask: outarg, the resulting cpumask
3766 *
3767 * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3768 * @cpu_going_down is >= 0, that cpu is considered offline during
3769 * calculation.  The result is stored in @cpumask.  This function returns
3770 * %true if the resulting @cpumask is different from @attrs->cpumask,
3771 * %false if equal.
3772 *
3773 * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3774 * enabled and @node has online CPUs requested by @attrs, the returned
3775 * cpumask is the intersection of the possible CPUs of @node and
3776 * @attrs->cpumask.
3777 *
3778 * The caller is responsible for ensuring that the cpumask of @node stays
3779 * stable.
3780 */
3781static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3782                                 int cpu_going_down, cpumask_t *cpumask)
3783{
3784        if (!wq_numa_enabled || attrs->no_numa)
3785                goto use_dfl;
3786
3787        /* does @node have any online CPUs @attrs wants? */
3788        cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3789        if (cpu_going_down >= 0)
3790                cpumask_clear_cpu(cpu_going_down, cpumask);
3791
3792        if (cpumask_empty(cpumask))
3793                goto use_dfl;
3794
3795        /* yeap, return possible CPUs in @node that @attrs wants */
3796        cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3797        return !cpumask_equal(cpumask, attrs->cpumask);
3798
3799use_dfl:
3800        cpumask_copy(cpumask, attrs->cpumask);
3801        return false;
3802}
3803
3804/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3805static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3806                                                   int node,
3807                                                   struct pool_workqueue *pwq)
3808{
3809        struct pool_workqueue *old_pwq;
3810
3811        lockdep_assert_held(&wq->mutex);
3812
3813        /* link_pwq() can handle duplicate calls */
3814        link_pwq(pwq);
3815
3816        old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3817        rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3818        return old_pwq;
3819}
3820
3821/**
3822 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3823 * @wq: the target workqueue
3824 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3825 *
3826 * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3827 * machines, this function maps a separate pwq to each NUMA node with
3828 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3829 * NUMA node it was issued on.  Older pwqs are released as in-flight work
3830 * items finish.  Note that a work item which repeatedly requeues itself
3831 * back-to-back will stay on its current pwq.
3832 *
3833 * Performs GFP_KERNEL allocations.  Returns 0 on success and -errno on
3834 * failure.
3835 */
3836int apply_workqueue_attrs(struct workqueue_struct *wq,
3837                          const struct workqueue_attrs *attrs)
3838{
3839        struct workqueue_attrs *new_attrs, *tmp_attrs;
3840        struct pool_workqueue **pwq_tbl, *dfl_pwq;
3841        int node, ret;
3842
3843        /* only unbound workqueues can change attributes */
3844        if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3845                return -EINVAL;
3846
3847        /* creating multiple pwqs breaks ordering guarantee */
3848        if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3849                return -EINVAL;
3850
3851        pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3852        new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3853        tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3854        if (!pwq_tbl || !new_attrs || !tmp_attrs)
3855                goto enomem;
3856
3857        /* make a copy of @attrs and sanitize it */
3858        copy_workqueue_attrs(new_attrs, attrs);
3859        cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3860
3861        /*
3862         * We may create multiple pwqs with differing cpumasks.  Make a
3863         * copy of @new_attrs which will be modified and used to obtain
3864         * pools.
3865         */
3866        copy_workqueue_attrs(tmp_attrs, new_attrs);
3867
3868        /*
3869         * CPUs should stay stable across pwq creations and installations.
3870         * Pin CPUs, determine the target cpumask for each node and create
3871         * pwqs accordingly.
3872         */
3873        get_online_cpus();
3874
3875        mutex_lock(&wq_pool_mutex);
3876
3877        /*
3878         * If something goes wrong during CPU up/down, we'll fall back to
3879         * the default pwq covering whole @attrs->cpumask.  Always create
3880         * it even if we don't use it immediately.
3881         */
3882        dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3883        if (!dfl_pwq)
3884                goto enomem_pwq;
3885
3886        for_each_node(node) {
3887                if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3888                        pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3889                        if (!pwq_tbl[node])
3890                                goto enomem_pwq;
3891                } else {
3892                        dfl_pwq->refcnt++;
3893                        pwq_tbl[node] = dfl_pwq;
3894                }
3895        }
3896
3897        mutex_unlock(&wq_pool_mutex);
3898
3899        /* all pwqs have been created successfully, let's install'em */
3900        mutex_lock(&wq->mutex);
3901
3902        copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3903
3904        /* save the previous pwq and install the new one */
3905        for_each_node(node)
3906                pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3907
3908        /* @dfl_pwq might not have been used, ensure it's linked */
3909        link_pwq(dfl_pwq);
3910        swap(wq->dfl_pwq, dfl_pwq);
3911
3912        mutex_unlock(&wq->mutex);
3913
3914        /* put the old pwqs */
3915        for_each_node(node)
3916                put_pwq_unlocked(pwq_tbl[node]);
3917        put_pwq_unlocked(dfl_pwq);
3918
3919        put_online_cpus();
3920        ret = 0;
3921        /* fall through */
3922out_free:
3923        free_workqueue_attrs(tmp_attrs);
3924        free_workqueue_attrs(new_attrs);
3925        kfree(pwq_tbl);
3926        return ret;
3927
3928enomem_pwq:
3929        free_unbound_pwq(dfl_pwq);
3930        for_each_node(node)
3931                if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3932                        free_unbound_pwq(pwq_tbl[node]);
3933        mutex_unlock(&wq_pool_mutex);
3934        put_online_cpus();
3935enomem:
3936        ret = -ENOMEM;
3937        goto out_free;
3938}
3939
3940/**
3941 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3942 * @wq: the target workqueue
3943 * @cpu: the CPU coming up or going down
3944 * @online: whether @cpu is coming up or going down
3945 *
3946 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3947 * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3948 * @wq accordingly.
3949 *
3950 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3951 * falls back to @wq->dfl_pwq which may not be optimal but is always
3952 * correct.
3953 *
3954 * Note that when the last allowed CPU of a NUMA node goes offline for a
3955 * workqueue with a cpumask spanning multiple nodes, the workers which were
3956 * already executing the work items for the workqueue will lose their CPU
3957 * affinity and may execute on any CPU.  This is similar to how per-cpu
3958 * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3959 * affinity, it's the user's responsibility to flush the work item from
3960 * CPU_DOWN_PREPARE.
3961 */
3962static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3963                                   bool online)
3964{
3965        int node = cpu_to_node(cpu);
3966        int cpu_off = online ? -1 : cpu;
3967        struct pool_workqueue *old_pwq = NULL, *pwq;
3968        struct workqueue_attrs *target_attrs;
3969        cpumask_t *cpumask;
3970
3971        lockdep_assert_held(&wq_pool_mutex);
3972
3973        if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
3974                return;
3975
3976        /*
3977         * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3978         * Let's use a preallocated one.  The following buf is protected by
3979         * CPU hotplug exclusion.
3980         */
3981        target_attrs = wq_update_unbound_numa_attrs_buf;
3982        cpumask = target_attrs->cpumask;
3983
3984        mutex_lock(&wq->mutex);
3985        if (wq->unbound_attrs->no_numa)
3986                goto out_unlock;
3987
3988        copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3989        pwq = unbound_pwq_by_node(wq, node);
3990
3991        /*
3992         * Let's determine what needs to be done.  If the target cpumask is
3993         * different from wq's, we need to compare it to @pwq's and create
3994         * a new one if they don't match.  If the target cpumask equals
3995         * wq's, the default pwq should be used.  If @pwq is already the
3996         * default one, nothing to do; otherwise, install the default one.
3997         */
3998        if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
3999                if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4000                        goto out_unlock;
4001        } else {
4002                if (pwq == wq->dfl_pwq)
4003                        goto out_unlock;
4004                else
4005                        goto use_dfl_pwq;
4006        }
4007
4008        mutex_unlock(&wq->mutex);
4009
4010        /* create a new pwq */
4011        pwq = alloc_unbound_pwq(wq, target_attrs);
4012        if (!pwq) {
4013                pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4014                           wq->name);
4015                goto out_unlock;
4016        }
4017
4018        /*
4019         * Install the new pwq.  As this function is called only from CPU
4020         * hotplug callbacks and applying a new attrs is wrapped with
4021         * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4022         * inbetween.
4023         */
4024        mutex_lock(&wq->mutex);
4025        old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4026        goto out_unlock;
4027
4028use_dfl_pwq:
4029        spin_lock_irq(&wq->dfl_pwq->pool->lock);
4030        get_pwq(wq->dfl_pwq);
4031        spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4032        old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4033out_unlock:
4034        mutex_unlock(&wq->mutex);
4035        put_pwq_unlocked(old_pwq);
4036}
4037
4038static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4039{
4040        bool highpri = wq->flags & WQ_HIGHPRI;
4041        int cpu;
4042
4043        if (!(wq->flags & WQ_UNBOUND)) {
4044                wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4045                if (!wq->cpu_pwqs)
4046                        return -ENOMEM;
4047
4048                for_each_possible_cpu(cpu) {
4049                        struct pool_workqueue *pwq =
4050                                per_cpu_ptr(wq->cpu_pwqs, cpu);
4051                        struct worker_pool *cpu_pools =
4052                                per_cpu(cpu_worker_pools, cpu);
4053
4054                        init_pwq(pwq, wq, &cpu_pools[highpri]);
4055
4056                        mutex_lock(&wq->mutex);
4057                        link_pwq(pwq);
4058                        mutex_unlock(&wq->mutex);
4059                }
4060                return 0;
4061        } else {
4062                return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4063        }
4064}
4065
4066static int wq_clamp_max_active(int max_active, unsigned int flags,
4067                               const char *name)
4068{
4069        int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4070
4071        if (max_active < 1 || max_active > lim)
4072                pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4073                        max_active, name, 1, lim);
4074
4075        return clamp_val(max_active, 1, lim);
4076}
4077
4078struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4079                                               unsigned int flags,
4080                                               int max_active,
4081                                               struct lock_class_key *key,
4082                                               const char *lock_name, ...)
4083{
4084        size_t tbl_size = 0;
4085        va_list args;
4086        struct workqueue_struct *wq;
4087        struct pool_workqueue *pwq;
4088
4089        /* allocate wq and format name */
4090        if (flags & WQ_UNBOUND)
4091                tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4092
4093        wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4094        if (!wq)
4095                return NULL;
4096
4097        if (flags & WQ_UNBOUND) {
4098                wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4099                if (!wq->unbound_attrs)
4100                        goto err_free_wq;
4101        }
4102
4103        va_start(args, lock_name);
4104        vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4105        va_end(args);
4106
4107        max_active = max_active ?: WQ_DFL_ACTIVE;
4108        max_active = wq_clamp_max_active(max_active, flags, wq->name);
4109
4110        /* init wq */
4111        wq->flags = flags;
4112        wq->saved_max_active = max_active;
4113        mutex_init(&wq->mutex);
4114        atomic_set(&wq->nr_pwqs_to_flush, 0);
4115        INIT_LIST_HEAD(&wq->pwqs);
4116        INIT_LIST_HEAD(&wq->flusher_queue);
4117        INIT_LIST_HEAD(&wq->flusher_overflow);
4118        INIT_LIST_HEAD(&wq->maydays);
4119
4120        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4121        INIT_LIST_HEAD(&wq->list);
4122
4123        if (alloc_and_link_pwqs(wq) < 0)
4124                goto err_free_wq;
4125
4126        /*
4127         * Workqueues which may be used during memory reclaim should
4128         * have a rescuer to guarantee forward progress.
4129         */
4130        if (flags & WQ_MEM_RECLAIM) {
4131                struct worker *rescuer;
4132
4133                rescuer = alloc_worker();
4134                if (!rescuer)
4135                        goto err_destroy;
4136
4137                rescuer->rescue_wq = wq;
4138                rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4139                                               wq->name);
4140                if (IS_ERR(rescuer->task)) {
4141                        kfree(rescuer);
4142                        goto err_destroy;
4143                }
4144
4145                wq->rescuer = rescuer;
4146                rescuer->task->flags |= PF_NO_SETAFFINITY;
4147                wake_up_process(rescuer->task);
4148        }
4149
4150        if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4151                goto err_destroy;
4152
4153        /*
4154         * wq_pool_mutex protects global freeze state and workqueues list.
4155         * Grab it, adjust max_active and add the new @wq to workqueues
4156         * list.
4157         */
4158        mutex_lock(&wq_pool_mutex);
4159
4160        mutex_lock(&wq->mutex);
4161        for_each_pwq(pwq, wq)
4162                pwq_adjust_max_active(pwq);
4163        mutex_unlock(&wq->mutex);
4164
4165        list_add(&wq->list, &workqueues);
4166
4167        mutex_unlock(&wq_pool_mutex);
4168
4169        return wq;
4170
4171err_free_wq:
4172        free_workqueue_attrs(wq->unbound_attrs);
4173        kfree(wq);
4174        return NULL;
4175err_destroy:
4176        destroy_workqueue(wq);
4177        return NULL;
4178}
4179EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4180
4181/**
4182 * destroy_workqueue - safely terminate a workqueue
4183 * @wq: target workqueue
4184 *
4185 * Safely destroy a workqueue. All work currently pending will be done first.
4186 */
4187void destroy_workqueue(struct workqueue_struct *wq)
4188{
4189        struct pool_workqueue *pwq;
4190        int node;
4191
4192        /* drain it before proceeding with destruction */
4193        drain_workqueue(wq);
4194
4195        /* sanity checks */
4196        mutex_lock(&wq->mutex);
4197        for_each_pwq(pwq, wq) {
4198                int i;
4199
4200                for (i = 0; i < WORK_NR_COLORS; i++) {
4201                        if (WARN_ON(pwq->nr_in_flight[i])) {
4202                                mutex_unlock(&wq->mutex);
4203                                return;
4204                        }
4205                }
4206
4207                if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4208                    WARN_ON(pwq->nr_active) ||
4209                    WARN_ON(!list_empty(&pwq->delayed_works))) {
4210                        mutex_unlock(&wq->mutex);
4211                        return;
4212                }
4213        }
4214        mutex_unlock(&wq->mutex);
4215
4216        /*
4217         * wq list is used to freeze wq, remove from list after
4218         * flushing is complete in case freeze races us.
4219         */
4220        mutex_lock(&wq_pool_mutex);
4221        list_del_init(&wq->list);
4222        mutex_unlock(&wq_pool_mutex);
4223
4224        workqueue_sysfs_unregister(wq);
4225
4226        if (wq->rescuer) {
4227                kthread_stop(wq->rescuer->task);
4228                kfree(wq->rescuer);
4229                wq->rescuer = NULL;
4230        }
4231
4232        if (!(wq->flags & WQ_UNBOUND)) {
4233                /*
4234                 * The base ref is never dropped on per-cpu pwqs.  Directly
4235                 * free the pwqs and wq.
4236                 */
4237                free_percpu(wq->cpu_pwqs);
4238                kfree(wq);
4239        } else {
4240                /*
4241                 * We're the sole accessor of @wq at this point.  Directly
4242                 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4243                 * @wq will be freed when the last pwq is released.
4244                 */
4245                for_each_node(node) {
4246                        pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4247                        RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4248                        put_pwq_unlocked(pwq);
4249                }
4250
4251                /*
4252                 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4253                 * put.  Don't access it afterwards.
4254                 */
4255                pwq = wq->dfl_pwq;
4256                wq->dfl_pwq = NULL;
4257                put_pwq_unlocked(pwq);
4258        }
4259}
4260EXPORT_SYMBOL_GPL(destroy_workqueue);
4261
4262/**
4263 * workqueue_set_max_active - adjust max_active of a workqueue
4264 * @wq: target workqueue
4265 * @max_active: new max_active value.
4266 *
4267 * Set max_active of @wq to @max_active.
4268 *
4269 * CONTEXT:
4270 * Don't call from IRQ context.
4271 */
4272void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4273{
4274        struct pool_workqueue *pwq;
4275
4276        /* disallow meddling with max_active for ordered workqueues */
4277        if (WARN_ON(wq->flags & __WQ_ORDERED))
4278                return;
4279
4280        max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4281
4282        mutex_lock(&wq->mutex);
4283
4284        wq->saved_max_active = max_active;
4285
4286        for_each_pwq(pwq, wq)
4287                pwq_adjust_max_active(pwq);
4288
4289        mutex_unlock(&wq->mutex);
4290}
4291EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4292
4293/**
4294 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4295 *
4296 * Determine whether %current is a workqueue rescuer.  Can be used from
4297 * work functions to determine whether it's being run off the rescuer task.
4298 */
4299bool current_is_workqueue_rescuer(void)
4300{
4301        struct worker *worker = current_wq_worker();
4302
4303        return worker && worker->rescue_wq;
4304}
4305
4306/**
4307 * workqueue_congested - test whether a workqueue is congested
4308 * @cpu: CPU in question
4309 * @wq: target workqueue
4310 *
4311 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4312 * no synchronization around this function and the test result is
4313 * unreliable and only useful as advisory hints or for debugging.
4314 *
4315 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4316 * Note that both per-cpu and unbound workqueues may be associated with
4317 * multiple pool_workqueues which have separate congested states.  A
4318 * workqueue being congested on one CPU doesn't mean the workqueue is also
4319 * contested on other CPUs / NUMA nodes.
4320 *
4321 * RETURNS:
4322 * %true if congested, %false otherwise.
4323 */
4324bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4325{
4326        struct pool_workqueue *pwq;
4327        bool ret;
4328
4329        rcu_read_lock_sched();
4330
4331        if (cpu == WORK_CPU_UNBOUND)
4332                cpu = smp_processor_id();
4333
4334        if (!(wq->flags & WQ_UNBOUND))
4335                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4336        else
4337                pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4338
4339        ret = !list_empty(&pwq->delayed_works);
4340        rcu_read_unlock_sched();
4341
4342        return ret;
4343}
4344EXPORT_SYMBOL_GPL(workqueue_congested);
4345
4346/**
4347 * work_busy - test whether a work is currently pending or running
4348 * @work: the work to be tested
4349 *
4350 * Test whether @work is currently pending or running.  There is no
4351 * synchronization around this function and the test result is
4352 * unreliable and only useful as advisory hints or for debugging.
4353 *
4354 * RETURNS:
4355 * OR'd bitmask of WORK_BUSY_* bits.
4356 */
4357unsigned int work_busy(struct work_struct *work)
4358{
4359        struct worker_pool *pool;
4360        unsigned long flags;
4361        unsigned int ret = 0;
4362
4363        if (work_pending(work))
4364                ret |= WORK_BUSY_PENDING;
4365
4366        local_irq_save(flags);
4367        pool = get_work_pool(work);
4368        if (pool) {
4369                spin_lock(&pool->lock);
4370                if (find_worker_executing_work(pool, work))
4371                        ret |= WORK_BUSY_RUNNING;
4372                spin_unlock(&pool->lock);
4373        }
4374        local_irq_restore(flags);
4375
4376        return ret;
4377}
4378EXPORT_SYMBOL_GPL(work_busy);
4379
4380/**
4381 * set_worker_desc - set description for the current work item
4382 * @fmt: printf-style format string
4383 * @...: arguments for the format string
4384 *
4385 * This function can be called by a running work function to describe what
4386 * the work item is about.  If the worker task gets dumped, this
4387 * information will be printed out together to help debugging.  The
4388 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4389 */
4390void set_worker_desc(const char *fmt, ...)
4391{
4392        struct worker *worker = current_wq_worker();
4393        va_list args;
4394
4395        if (worker) {
4396                va_start(args, fmt);
4397                vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4398                va_end(args);
4399                worker->desc_valid = true;
4400        }
4401}
4402
4403/**
4404 * print_worker_info - print out worker information and description
4405 * @log_lvl: the log level to use when printing
4406 * @task: target task
4407 *
4408 * If @task is a worker and currently executing a work item, print out the
4409 * name of the workqueue being serviced and worker description set with
4410 * set_worker_desc() by the currently executing work item.
4411 *
4412 * This function can be safely called on any task as long as the
4413 * task_struct itself is accessible.  While safe, this function isn't
4414 * synchronized and may print out mixups or garbages of limited length.
4415 */
4416void print_worker_info(const char *log_lvl, struct task_struct *task)
4417{
4418        work_func_t *fn = NULL;
4419        char name[WQ_NAME_LEN] = { };
4420        char desc[WORKER_DESC_LEN] = { };
4421        struct pool_workqueue *pwq = NULL;
4422        struct workqueue_struct *wq = NULL;
4423        bool desc_valid = false;
4424        struct worker *worker;
4425
4426        if (!(task->flags & PF_WQ_WORKER))
4427                return;
4428
4429        /*
4430         * This function is called without any synchronization and @task
4431         * could be in any state.  Be careful with dereferences.
4432         */
4433        worker = probe_kthread_data(task);
4434
4435        /*
4436         * Carefully copy the associated workqueue's workfn and name.  Keep
4437         * the original last '\0' in case the original contains garbage.
4438         */
4439        probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4440        probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4441        probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4442        probe_kernel_read(name, wq->name, sizeof(name) - 1);
4443
4444        /* copy worker description */
4445        probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4446        if (desc_valid)
4447                probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4448
4449        if (fn || name[0] || desc[0]) {
4450                printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4451                if (desc[0])
4452                        pr_cont(" (%s)", desc);
4453                pr_cont("\n");
4454        }
4455}
4456
4457/*
4458 * CPU hotplug.
4459 *
4460 * There are two challenges in supporting CPU hotplug.  Firstly, there
4461 * are a lot of assumptions on strong associations among work, pwq and
4462 * pool which make migrating pending and scheduled works very
4463 * difficult to implement without impacting hot paths.  Secondly,
4464 * worker pools serve mix of short, long and very long running works making
4465 * blocked draining impractical.
4466 *
4467 * This is solved by allowing the pools to be disassociated from the CPU
4468 * running as an unbound one and allowing it to be reattached later if the
4469 * cpu comes back online.
4470 */
4471
4472static void wq_unbind_fn(struct work_struct *work)
4473{
4474        int cpu = smp_processor_id();
4475        struct worker_pool *pool;
4476        struct worker *worker;
4477        int wi;
4478
4479        for_each_cpu_worker_pool(pool, cpu) {
4480                WARN_ON_ONCE(cpu != smp_processor_id());
4481
4482                mutex_lock(&pool->manager_mutex);
4483                spin_lock_irq(&pool->lock);
4484
4485                /*
4486                 * We've blocked all manager operations.  Make all workers
4487                 * unbound and set DISASSOCIATED.  Before this, all workers
4488                 * except for the ones which are still executing works from
4489                 * before the last CPU down must be on the cpu.  After
4490                 * this, they may become diasporas.
4491                 */
4492                for_each_pool_worker(worker, wi, pool)
4493                        worker->flags |= WORKER_UNBOUND;
4494
4495                pool->flags |= POOL_DISASSOCIATED;
4496
4497                spin_unlock_irq(&pool->lock);
4498                mutex_unlock(&pool->manager_mutex);
4499
4500                /*
4501                 * Call schedule() so that we cross rq->lock and thus can
4502                 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4503                 * This is necessary as scheduler callbacks may be invoked
4504                 * from other cpus.
4505                 */
4506                schedule();
4507
4508                /*
4509                 * Sched callbacks are disabled now.  Zap nr_running.
4510                 * After this, nr_running stays zero and need_more_worker()
4511                 * and keep_working() are always true as long as the
4512                 * worklist is not empty.  This pool now behaves as an
4513                 * unbound (in terms of concurrency management) pool which
4514                 * are served by workers tied to the pool.
4515                 */
4516                atomic_set(&pool->nr_running, 0);
4517
4518                /*
4519                 * With concurrency management just turned off, a busy
4520                 * worker blocking could lead to lengthy stalls.  Kick off
4521                 * unbound chain execution of currently pending work items.
4522                 */
4523                spin_lock_irq(&pool->lock);
4524                wake_up_worker(pool);
4525                spin_unlock_irq(&pool->lock);
4526        }
4527}
4528
4529/**
4530 * rebind_workers - rebind all workers of a pool to the associated CPU
4531 * @pool: pool of interest
4532 *
4533 * @pool->cpu is coming online.  Rebind all workers to the CPU.
4534 */
4535static void rebind_workers(struct worker_pool *pool)
4536{
4537        struct worker *worker;
4538        int wi;
4539
4540        lockdep_assert_held(&pool->manager_mutex);
4541
4542        /*
4543         * Restore CPU affinity of all workers.  As all idle workers should
4544         * be on the run-queue of the associated CPU before any local
4545         * wake-ups for concurrency management happen, restore CPU affinty
4546         * of all workers first and then clear UNBOUND.  As we're called
4547         * from CPU_ONLINE, the following shouldn't fail.
4548         */
4549        for_each_pool_worker(worker, wi, pool)
4550                WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4551                                                  pool->attrs->cpumask) < 0);
4552
4553        spin_lock_irq(&pool->lock);
4554
4555        for_each_pool_worker(worker, wi, pool) {
4556                unsigned int worker_flags = worker->flags;
4557
4558                /*
4559                 * A bound idle worker should actually be on the runqueue
4560                 * of the associated CPU for local wake-ups targeting it to
4561                 * work.  Kick all idle workers so that they migrate to the
4562                 * associated CPU.  Doing this in the same loop as
4563                 * replacing UNBOUND with REBOUND is safe as no worker will
4564                 * be bound before @pool->lock is released.
4565                 */
4566                if (worker_flags & WORKER_IDLE)
4567                        wake_up_process(worker->task);
4568
4569                /*
4570                 * We want to clear UNBOUND but can't directly call
4571                 * worker_clr_flags() or adjust nr_running.  Atomically
4572                 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4573                 * @worker will clear REBOUND using worker_clr_flags() when
4574                 * it initiates the next execution cycle thus restoring
4575                 * concurrency management.  Note that when or whether
4576                 * @worker clears REBOUND doesn't affect correctness.
4577                 *
4578                 * ACCESS_ONCE() is necessary because @worker->flags may be
4579                 * tested without holding any lock in
4580                 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4581                 * fail incorrectly leading to premature concurrency
4582                 * management operations.
4583                 */
4584                WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4585                worker_flags |= WORKER_REBOUND;
4586                worker_flags &= ~WORKER_UNBOUND;
4587                ACCESS_ONCE(worker->flags) = worker_flags;
4588        }
4589
4590        spin_unlock_irq(&pool->lock);
4591}
4592
4593/**
4594 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4595 * @pool: unbound pool of interest
4596 * @cpu: the CPU which is coming up
4597 *
4598 * An unbound pool may end up with a cpumask which doesn't have any online
4599 * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4600 * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4601 * online CPU before, cpus_allowed of all its workers should be restored.
4602 */
4603static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4604{
4605        static cpumask_t cpumask;
4606        struct worker *worker;
4607        int wi;
4608
4609        lockdep_assert_held(&pool->manager_mutex);
4610
4611        /* is @cpu allowed for @pool? */
4612        if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4613                return;
4614
4615        /* is @cpu the only online CPU? */
4616        cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4617        if (cpumask_weight(&cpumask) != 1)
4618                return;
4619
4620        /* as we're called from CPU_ONLINE, the following shouldn't fail */
4621        for_each_pool_worker(worker, wi, pool)
4622                WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4623                                                  pool->attrs->cpumask) < 0);
4624}
4625
4626/*
4627 * Workqueues should be brought up before normal priority CPU notifiers.
4628 * This will be registered high priority CPU notifier.
4629 */
4630static int __cpuinit workqueue_cpu_up_callback(struct notifier_block *nfb,
4631                                               unsigned long action,
4632                                               void *hcpu)
4633{
4634        int cpu = (unsigned long)hcpu;
4635        struct worker_pool *pool;
4636        struct workqueue_struct *wq;
4637        int pi;
4638
4639        switch (action & ~CPU_TASKS_FROZEN) {
4640        case CPU_UP_PREPARE:
4641                for_each_cpu_worker_pool(pool, cpu) {
4642                        if (pool->nr_workers)
4643                                continue;
4644                        if (create_and_start_worker(pool) < 0)
4645                                return NOTIFY_BAD;
4646                }
4647                break;
4648
4649        case CPU_DOWN_FAILED:
4650        case CPU_ONLINE:
4651                mutex_lock(&wq_pool_mutex);
4652
4653                for_each_pool(pool, pi) {
4654                        mutex_lock(&pool->manager_mutex);
4655
4656                        if (pool->cpu == cpu) {
4657                                spin_lock_irq(&pool->lock);
4658                                pool->flags &= ~POOL_DISASSOCIATED;
4659                                spin_unlock_irq(&pool->lock);
4660
4661                                rebind_workers(pool);
4662                        } else if (pool->cpu < 0) {
4663                                restore_unbound_workers_cpumask(pool, cpu);
4664                        }
4665
4666                        mutex_unlock(&pool->manager_mutex);
4667                }
4668
4669                /* update NUMA affinity of unbound workqueues */
4670                list_for_each_entry(wq, &workqueues, list)
4671                        wq_update_unbound_numa(wq, cpu, true);
4672
4673                mutex_unlock(&wq_pool_mutex);
4674                break;
4675        }
4676        return NOTIFY_OK;
4677}
4678
4679/*
4680 * Workqueues should be brought down after normal priority CPU notifiers.
4681 * This will be registered as low priority CPU notifier.
4682 */
4683static int __cpuinit workqueue_cpu_down_callback(struct notifier_block *nfb,
4684                                                 unsigned long action,
4685                                                 void *hcpu)
4686{
4687        int cpu = (unsigned long)hcpu;
4688        struct work_struct unbind_work;
4689        struct workqueue_struct *wq;
4690
4691        switch (action & ~CPU_TASKS_FROZEN) {
4692        case CPU_DOWN_PREPARE:
4693                /* unbinding per-cpu workers should happen on the local CPU */
4694                INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4695                queue_work_on(cpu, system_highpri_wq, &unbind_work);
4696
4697                /* update NUMA affinity of unbound workqueues */
4698                mutex_lock(&wq_pool_mutex);
4699                list_for_each_entry(wq, &workqueues, list)
4700                        wq_update_unbound_numa(wq, cpu, false);
4701                mutex_unlock(&wq_pool_mutex);
4702
4703                /* wait for per-cpu unbinding to finish */
4704                flush_work(&unbind_work);
4705                break;
4706        }
4707        return NOTIFY_OK;
4708}
4709
4710#ifdef CONFIG_SMP
4711
4712struct work_for_cpu {
4713        struct work_struct work;
4714        long (*fn)(void *);
4715        void *arg;
4716        long ret;
4717};
4718
4719static void work_for_cpu_fn(struct work_struct *work)
4720{
4721        struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4722
4723        wfc->ret = wfc->fn(wfc->arg);
4724}
4725
4726/**
4727 * work_on_cpu - run a function in user context on a particular cpu
4728 * @cpu: the cpu to run on
4729 * @fn: the function to run
4730 * @arg: the function arg
4731 *
4732 * This will return the value @fn returns.
4733 * It is up to the caller to ensure that the cpu doesn't go offline.
4734 * The caller must not hold any locks which would prevent @fn from completing.
4735 */
4736long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4737{
4738        struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4739
4740        INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4741        schedule_work_on(cpu, &wfc.work);
4742        flush_work(&wfc.work);
4743        return wfc.ret;
4744}
4745EXPORT_SYMBOL_GPL(work_on_cpu);
4746#endif /* CONFIG_SMP */
4747
4748#ifdef CONFIG_FREEZER
4749
4750/**
4751 * freeze_workqueues_begin - begin freezing workqueues
4752 *
4753 * Start freezing workqueues.  After this function returns, all freezable
4754 * workqueues will queue new works to their delayed_works list instead of
4755 * pool->worklist.
4756 *
4757 * CONTEXT:
4758 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4759 */
4760void freeze_workqueues_begin(void)
4761{
4762        struct worker_pool *pool;
4763        struct workqueue_struct *wq;
4764        struct pool_workqueue *pwq;
4765        int pi;
4766
4767        mutex_lock(&wq_pool_mutex);
4768
4769        WARN_ON_ONCE(workqueue_freezing);
4770        workqueue_freezing = true;
4771
4772        /* set FREEZING */
4773        for_each_pool(pool, pi) {
4774                spin_lock_irq(&pool->lock);
4775                WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4776                pool->flags |= POOL_FREEZING;
4777                spin_unlock_irq(&pool->lock);
4778        }
4779
4780        list_for_each_entry(wq, &workqueues, list) {
4781                mutex_lock(&wq->mutex);
4782                for_each_pwq(pwq, wq)
4783                        pwq_adjust_max_active(pwq);
4784                mutex_unlock(&wq->mutex);
4785        }
4786
4787        mutex_unlock(&wq_pool_mutex);
4788}
4789
4790/**
4791 * freeze_workqueues_busy - are freezable workqueues still busy?
4792 *
4793 * Check whether freezing is complete.  This function must be called
4794 * between freeze_workqueues_begin() and thaw_workqueues().
4795 *
4796 * CONTEXT:
4797 * Grabs and releases wq_pool_mutex.
4798 *
4799 * RETURNS:
4800 * %true if some freezable workqueues are still busy.  %false if freezing
4801 * is complete.
4802 */
4803bool freeze_workqueues_busy(void)
4804{
4805        bool busy = false;
4806        struct workqueue_struct *wq;
4807        struct pool_workqueue *pwq;
4808
4809        mutex_lock(&wq_pool_mutex);
4810
4811        WARN_ON_ONCE(!workqueue_freezing);
4812
4813        list_for_each_entry(wq, &workqueues, list) {
4814                if (!(wq->flags & WQ_FREEZABLE))
4815                        continue;
4816                /*
4817                 * nr_active is monotonically decreasing.  It's safe
4818                 * to peek without lock.
4819                 */
4820                rcu_read_lock_sched();
4821                for_each_pwq(pwq, wq) {
4822                        WARN_ON_ONCE(pwq->nr_active < 0);
4823                        if (pwq->nr_active) {
4824                                busy = true;
4825                                rcu_read_unlock_sched();
4826                                goto out_unlock;
4827                        }
4828                }
4829                rcu_read_unlock_sched();
4830        }
4831out_unlock:
4832        mutex_unlock(&wq_pool_mutex);
4833        return busy;
4834}
4835
4836/**
4837 * thaw_workqueues - thaw workqueues
4838 *
4839 * Thaw workqueues.  Normal queueing is restored and all collected
4840 * frozen works are transferred to their respective pool worklists.
4841 *
4842 * CONTEXT:
4843 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4844 */
4845void thaw_workqueues(void)
4846{
4847        struct workqueue_struct *wq;
4848        struct pool_workqueue *pwq;
4849        struct worker_pool *pool;
4850        int pi;
4851
4852        mutex_lock(&wq_pool_mutex);
4853
4854        if (!workqueue_freezing)
4855                goto out_unlock;
4856
4857        /* clear FREEZING */
4858        for_each_pool(pool, pi) {
4859                spin_lock_irq(&pool->lock);
4860                WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4861                pool->flags &= ~POOL_FREEZING;
4862                spin_unlock_irq(&pool->lock);
4863        }
4864
4865        /* restore max_active and repopulate worklist */
4866        list_for_each_entry(wq, &workqueues, list) {
4867                mutex_lock(&wq->mutex);
4868                for_each_pwq(pwq, wq)
4869                        pwq_adjust_max_active(pwq);
4870                mutex_unlock(&wq->mutex);
4871        }
4872
4873        workqueue_freezing = false;
4874out_unlock:
4875        mutex_unlock(&wq_pool_mutex);
4876}
4877#endif /* CONFIG_FREEZER */
4878
4879static void __init wq_numa_init(void)
4880{
4881        cpumask_var_t *tbl;
4882        int node, cpu;
4883
4884        /* determine NUMA pwq table len - highest node id + 1 */
4885        for_each_node(node)
4886                wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4887
4888        if (num_possible_nodes() <= 1)
4889                return;
4890
4891        if (wq_disable_numa) {
4892                pr_info("workqueue: NUMA affinity support disabled\n");
4893                return;
4894        }
4895
4896        wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4897        BUG_ON(!wq_update_unbound_numa_attrs_buf);
4898
4899        /*
4900         * We want masks of possible CPUs of each node which isn't readily
4901         * available.  Build one from cpu_to_node() which should have been
4902         * fully initialized by now.
4903         */
4904        tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4905        BUG_ON(!tbl);
4906
4907        for_each_node(node)
4908                BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4909                                node_online(node) ? node : NUMA_NO_NODE));
4910
4911        for_each_possible_cpu(cpu) {
4912                node = cpu_to_node(cpu);
4913                if (WARN_ON(node == NUMA_NO_NODE)) {
4914                        pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4915                        /* happens iff arch is bonkers, let's just proceed */
4916                        return;
4917                }
4918                cpumask_set_cpu(cpu, tbl[node]);
4919        }
4920
4921        wq_numa_possible_cpumask = tbl;
4922        wq_numa_enabled = true;
4923}
4924
4925static int __init init_workqueues(void)
4926{
4927        int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4928        int i, cpu;
4929
4930        /* make sure we have enough bits for OFFQ pool ID */
4931        BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4932                     WORK_CPU_END * NR_STD_WORKER_POOLS);
4933
4934        WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4935
4936        pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4937
4938        cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4939        hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4940
4941        wq_numa_init();
4942
4943        /* initialize CPU pools */
4944        for_each_possible_cpu(cpu) {
4945                struct worker_pool *pool;
4946
4947                i = 0;
4948                for_each_cpu_worker_pool(pool, cpu) {
4949                        BUG_ON(init_worker_pool(pool));
4950                        pool->cpu = cpu;
4951                        cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
4952                        pool->attrs->nice = std_nice[i++];
4953                        pool->node = cpu_to_node(cpu);
4954
4955                        /* alloc pool ID */
4956                        mutex_lock(&wq_pool_mutex);
4957                        BUG_ON(worker_pool_assign_id(pool));
4958                        mutex_unlock(&wq_pool_mutex);
4959                }
4960        }
4961
4962        /* create the initial worker */
4963        for_each_online_cpu(cpu) {
4964                struct worker_pool *pool;
4965
4966                for_each_cpu_worker_pool(pool, cpu) {
4967                        pool->flags &= ~POOL_DISASSOCIATED;
4968                        BUG_ON(create_and_start_worker(pool) < 0);
4969                }
4970        }
4971
4972        /* create default unbound wq attrs */
4973        for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
4974                struct workqueue_attrs *attrs;
4975
4976                BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
4977                attrs->nice = std_nice[i];
4978                unbound_std_wq_attrs[i] = attrs;
4979        }
4980
4981        system_wq = alloc_workqueue("events", 0, 0);
4982        system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
4983        system_long_wq = alloc_workqueue("events_long", 0, 0);
4984        system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
4985                                            WQ_UNBOUND_MAX_ACTIVE);
4986        system_freezable_wq = alloc_workqueue("events_freezable",
4987                                              WQ_FREEZABLE, 0);
4988        BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
4989               !system_unbound_wq || !system_freezable_wq);
4990        return 0;
4991}
4992early_initcall(init_workqueues);
4993
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.