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
 275/* see the comment above the definition of WQ_POWER_EFFICIENT */
 276#ifdef CONFIG_WQ_POWER_EFFICIENT_DEFAULT
 277static bool wq_power_efficient = true;
 278#else
 279static bool wq_power_efficient;
 280#endif
 281
 282module_param_named(power_efficient, wq_power_efficient, bool, 0444);
 283
 284static bool wq_numa_enabled;            /* unbound NUMA affinity enabled */
 285
 286/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
 287static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
 288
 289static DEFINE_MUTEX(wq_pool_mutex);     /* protects pools and workqueues list */
 290static DEFINE_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */
 291
 292static LIST_HEAD(workqueues);           /* PL: list of all workqueues */
 293static bool workqueue_freezing;         /* PL: have wqs started freezing? */
 294
 295/* the per-cpu worker pools */
 296static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS],
 297                                     cpu_worker_pools);
 298
 299static DEFINE_IDR(worker_pool_idr);     /* PR: idr of all pools */
 300
 301/* PL: hash of all unbound pools keyed by pool->attrs */
 302static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
 303
 304/* I: attributes used when instantiating standard unbound pools on demand */
 305static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
 306
 307struct workqueue_struct *system_wq __read_mostly;
 308EXPORT_SYMBOL(system_wq);
 309struct workqueue_struct *system_highpri_wq __read_mostly;
 310EXPORT_SYMBOL_GPL(system_highpri_wq);
 311struct workqueue_struct *system_long_wq __read_mostly;
 312EXPORT_SYMBOL_GPL(system_long_wq);
 313struct workqueue_struct *system_unbound_wq __read_mostly;
 314EXPORT_SYMBOL_GPL(system_unbound_wq);
 315struct workqueue_struct *system_freezable_wq __read_mostly;
 316EXPORT_SYMBOL_GPL(system_freezable_wq);
 317struct workqueue_struct *system_power_efficient_wq __read_mostly;
 318EXPORT_SYMBOL_GPL(system_power_efficient_wq);
 319struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
 320EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
 321
 322static int worker_thread(void *__worker);
 323static void copy_workqueue_attrs(struct workqueue_attrs *to,
 324                                 const struct workqueue_attrs *from);
 325
 326#define CREATE_TRACE_POINTS
 327#include <trace/events/workqueue.h>
 328
 329#define assert_rcu_or_pool_mutex()                                      \
 330        rcu_lockdep_assert(rcu_read_lock_sched_held() ||                \
 331                           lockdep_is_held(&wq_pool_mutex),             \
 332                           "sched RCU or wq_pool_mutex should be held")
 333
 334#define assert_rcu_or_wq_mutex(wq)                                      \
 335        rcu_lockdep_assert(rcu_read_lock_sched_held() ||                \
 336                           lockdep_is_held(&wq->mutex),                 \
 337                           "sched RCU or wq->mutex should be held")
 338
 339#ifdef CONFIG_LOCKDEP
 340#define assert_manager_or_pool_lock(pool)                               \
 341        WARN_ONCE(debug_locks &&                                        \
 342                  !lockdep_is_held(&(pool)->manager_mutex) &&           \
 343                  !lockdep_is_held(&(pool)->lock),                      \
 344                  "pool->manager_mutex or ->lock should be held")
 345#else
 346#define assert_manager_or_pool_lock(pool)       do { } while (0)
 347#endif
 348
 349#define for_each_cpu_worker_pool(pool, cpu)                             \
 350        for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];               \
 351             (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
 352             (pool)++)
 353
 354/**
 355 * for_each_pool - iterate through all worker_pools in the system
 356 * @pool: iteration cursor
 357 * @pi: integer used for iteration
 358 *
 359 * This must be called either with wq_pool_mutex held or sched RCU read
 360 * locked.  If the pool needs to be used beyond the locking in effect, the
 361 * caller is responsible for guaranteeing that the pool stays online.
 362 *
 363 * The if/else clause exists only for the lockdep assertion and can be
 364 * ignored.
 365 */
 366#define for_each_pool(pool, pi)                                         \
 367        idr_for_each_entry(&worker_pool_idr, pool, pi)                  \
 368                if (({ assert_rcu_or_pool_mutex(); false; })) { }       \
 369                else
 370
 371/**
 372 * for_each_pool_worker - iterate through all workers of a worker_pool
 373 * @worker: iteration cursor
 374 * @wi: integer used for iteration
 375 * @pool: worker_pool to iterate workers of
 376 *
 377 * This must be called with either @pool->manager_mutex or ->lock held.
 378 *
 379 * The if/else clause exists only for the lockdep assertion and can be
 380 * ignored.
 381 */
 382#define for_each_pool_worker(worker, wi, pool)                          \
 383        idr_for_each_entry(&(pool)->worker_idr, (worker), (wi))         \
 384                if (({ assert_manager_or_pool_lock((pool)); false; })) { } \
 385                else
 386
 387/**
 388 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
 389 * @pwq: iteration cursor
 390 * @wq: the target workqueue
 391 *
 392 * This must be called either with wq->mutex held or sched RCU read locked.
 393 * If the pwq needs to be used beyond the locking in effect, the caller is
 394 * responsible for guaranteeing that the pwq stays online.
 395 *
 396 * The if/else clause exists only for the lockdep assertion and can be
 397 * ignored.
 398 */
 399#define for_each_pwq(pwq, wq)                                           \
 400        list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)          \
 401                if (({ assert_rcu_or_wq_mutex(wq); false; })) { }       \
 402                else
 403
 404#ifdef CONFIG_DEBUG_OBJECTS_WORK
 405
 406static struct debug_obj_descr work_debug_descr;
 407
 408static void *work_debug_hint(void *addr)
 409{
 410        return ((struct work_struct *) addr)->func;
 411}
 412
 413/*
 414 * fixup_init is called when:
 415 * - an active object is initialized
 416 */
 417static int work_fixup_init(void *addr, enum debug_obj_state state)
 418{
 419        struct work_struct *work = addr;
 420
 421        switch (state) {
 422        case ODEBUG_STATE_ACTIVE:
 423                cancel_work_sync(work);
 424                debug_object_init(work, &work_debug_descr);
 425                return 1;
 426        default:
 427                return 0;
 428        }
 429}
 430
 431/*
 432 * fixup_activate is called when:
 433 * - an active object is activated
 434 * - an unknown object is activated (might be a statically initialized object)
 435 */
 436static int work_fixup_activate(void *addr, enum debug_obj_state state)
 437{
 438        struct work_struct *work = addr;
 439
 440        switch (state) {
 441
 442        case ODEBUG_STATE_NOTAVAILABLE:
 443                /*
 444                 * This is not really a fixup. The work struct was
 445                 * statically initialized. We just make sure that it
 446                 * is tracked in the object tracker.
 447                 */
 448                if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
 449                        debug_object_init(work, &work_debug_descr);
 450                        debug_object_activate(work, &work_debug_descr);
 451                        return 0;
 452                }
 453                WARN_ON_ONCE(1);
 454                return 0;
 455
 456        case ODEBUG_STATE_ACTIVE:
 457                WARN_ON(1);
 458
 459        default:
 460                return 0;
 461        }
 462}
 463
 464/*
 465 * fixup_free is called when:
 466 * - an active object is freed
 467 */
 468static int work_fixup_free(void *addr, enum debug_obj_state state)
 469{
 470        struct work_struct *work = addr;
 471
 472        switch (state) {
 473        case ODEBUG_STATE_ACTIVE:
 474                cancel_work_sync(work);
 475                debug_object_free(work, &work_debug_descr);
 476                return 1;
 477        default:
 478                return 0;
 479        }
 480}
 481
 482static struct debug_obj_descr work_debug_descr = {
 483        .name           = "work_struct",
 484        .debug_hint     = work_debug_hint,
 485        .fixup_init     = work_fixup_init,
 486        .fixup_activate = work_fixup_activate,
 487        .fixup_free     = work_fixup_free,
 488};
 489
 490static inline void debug_work_activate(struct work_struct *work)
 491{
 492        debug_object_activate(work, &work_debug_descr);
 493}
 494
 495static inline void debug_work_deactivate(struct work_struct *work)
 496{
 497        debug_object_deactivate(work, &work_debug_descr);
 498}
 499
 500void __init_work(struct work_struct *work, int onstack)
 501{
 502        if (onstack)
 503                debug_object_init_on_stack(work, &work_debug_descr);
 504        else
 505                debug_object_init(work, &work_debug_descr);
 506}
 507EXPORT_SYMBOL_GPL(__init_work);
 508
 509void destroy_work_on_stack(struct work_struct *work)
 510{
 511        debug_object_free(work, &work_debug_descr);
 512}
 513EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 514
 515#else
 516static inline void debug_work_activate(struct work_struct *work) { }
 517static inline void debug_work_deactivate(struct work_struct *work) { }
 518#endif
 519
 520/* allocate ID and assign it to @pool */
 521static int worker_pool_assign_id(struct worker_pool *pool)
 522{
 523        int ret;
 524
 525        lockdep_assert_held(&wq_pool_mutex);
 526
 527        ret = idr_alloc(&worker_pool_idr, pool, 0, 0, GFP_KERNEL);
 528        if (ret >= 0) {
 529                pool->id = ret;
 530                return 0;
 531        }
 532        return ret;
 533}
 534
 535/**
 536 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
 537 * @wq: the target workqueue
 538 * @node: the node ID
 539 *
 540 * This must be called either with pwq_lock held or sched RCU read locked.
 541 * If the pwq needs to be used beyond the locking in effect, the caller is
 542 * responsible for guaranteeing that the pwq stays online.
 543 */
 544static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
 545                                                  int node)
 546{
 547        assert_rcu_or_wq_mutex(wq);
 548        return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
 549}
 550
 551static unsigned int work_color_to_flags(int color)
 552{
 553        return color << WORK_STRUCT_COLOR_SHIFT;
 554}
 555
 556static int get_work_color(struct work_struct *work)
 557{
 558        return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 559                ((1 << WORK_STRUCT_COLOR_BITS) - 1);
 560}
 561
 562static int work_next_color(int color)
 563{
 564        return (color + 1) % WORK_NR_COLORS;
 565}
 566
 567/*
 568 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
 569 * contain the pointer to the queued pwq.  Once execution starts, the flag
 570 * is cleared and the high bits contain OFFQ flags and pool ID.
 571 *
 572 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
 573 * and clear_work_data() can be used to set the pwq, pool or clear
 574 * work->data.  These functions should only be called while the work is
 575 * owned - ie. while the PENDING bit is set.
 576 *
 577 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
 578 * corresponding to a work.  Pool is available once the work has been
 579 * queued anywhere after initialization until it is sync canceled.  pwq is
 580 * available only while the work item is queued.
 581 *
 582 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
 583 * canceled.  While being canceled, a work item may have its PENDING set
 584 * but stay off timer and worklist for arbitrarily long and nobody should
 585 * try to steal the PENDING bit.
 586 */
 587static inline void set_work_data(struct work_struct *work, unsigned long data,
 588                                 unsigned long flags)
 589{
 590        WARN_ON_ONCE(!work_pending(work));
 591        atomic_long_set(&work->data, data | flags | work_static(work));
 592}
 593
 594static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
 595                         unsigned long extra_flags)
 596{
 597        set_work_data(work, (unsigned long)pwq,
 598                      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
 599}
 600
 601static void set_work_pool_and_keep_pending(struct work_struct *work,
 602                                           int pool_id)
 603{
 604        set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
 605                      WORK_STRUCT_PENDING);
 606}
 607
 608static void set_work_pool_and_clear_pending(struct work_struct *work,
 609                                            int pool_id)
 610{
 611        /*
 612         * The following wmb is paired with the implied mb in
 613         * test_and_set_bit(PENDING) and ensures all updates to @work made
 614         * here are visible to and precede any updates by the next PENDING
 615         * owner.
 616         */
 617        smp_wmb();
 618        set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
 619}
 620
 621static void clear_work_data(struct work_struct *work)
 622{
 623        smp_wmb();      /* see set_work_pool_and_clear_pending() */
 624        set_work_data(work, WORK_STRUCT_NO_POOL, 0);
 625}
 626
 627static struct pool_workqueue *get_work_pwq(struct work_struct *work)
 628{
 629        unsigned long data = atomic_long_read(&work->data);
 630
 631        if (data & WORK_STRUCT_PWQ)
 632                return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 633        else
 634                return NULL;
 635}
 636
 637/**
 638 * get_work_pool - return the worker_pool a given work was associated with
 639 * @work: the work item of interest
 640 *
 641 * Return the worker_pool @work was last associated with.  %NULL if none.
 642 *
 643 * Pools are created and destroyed under wq_pool_mutex, and allows read
 644 * access under sched-RCU read lock.  As such, this function should be
 645 * called under wq_pool_mutex or with preemption disabled.
 646 *
 647 * All fields of the returned pool are accessible as long as the above
 648 * mentioned locking is in effect.  If the returned pool needs to be used
 649 * beyond the critical section, the caller is responsible for ensuring the
 650 * returned pool is and stays online.
 651 */
 652static struct worker_pool *get_work_pool(struct work_struct *work)
 653{
 654        unsigned long data = atomic_long_read(&work->data);
 655        int pool_id;
 656
 657        assert_rcu_or_pool_mutex();
 658
 659        if (data & WORK_STRUCT_PWQ)
 660                return ((struct pool_workqueue *)
 661                        (data & WORK_STRUCT_WQ_DATA_MASK))->pool;
 662
 663        pool_id = data >> WORK_OFFQ_POOL_SHIFT;
 664        if (pool_id == WORK_OFFQ_POOL_NONE)
 665                return NULL;
 666
 667        return idr_find(&worker_pool_idr, pool_id);
 668}
 669
 670/**
 671 * get_work_pool_id - return the worker pool ID a given work is associated with
 672 * @work: the work item of interest
 673 *
 674 * Return the worker_pool ID @work was last associated with.
 675 * %WORK_OFFQ_POOL_NONE if none.
 676 */
 677static int get_work_pool_id(struct work_struct *work)
 678{
 679        unsigned long data = atomic_long_read(&work->data);
 680
 681        if (data & WORK_STRUCT_PWQ)
 682                return ((struct pool_workqueue *)
 683                        (data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
 684
 685        return data >> WORK_OFFQ_POOL_SHIFT;
 686}
 687
 688static void mark_work_canceling(struct work_struct *work)
 689{
 690        unsigned long pool_id = get_work_pool_id(work);
 691
 692        pool_id <<= WORK_OFFQ_POOL_SHIFT;
 693        set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
 694}
 695
 696static bool work_is_canceling(struct work_struct *work)
 697{
 698        unsigned long data = atomic_long_read(&work->data);
 699
 700        return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
 701}
 702
 703/*
 704 * Policy functions.  These define the policies on how the global worker
 705 * pools are managed.  Unless noted otherwise, these functions assume that
 706 * they're being called with pool->lock held.
 707 */
 708
 709static bool __need_more_worker(struct worker_pool *pool)
 710{
 711        return !atomic_read(&pool->nr_running);
 712}
 713
 714/*
 715 * Need to wake up a worker?  Called from anything but currently
 716 * running workers.
 717 *
 718 * Note that, because unbound workers never contribute to nr_running, this
 719 * function will always return %true for unbound pools as long as the
 720 * worklist isn't empty.
 721 */
 722static bool need_more_worker(struct worker_pool *pool)
 723{
 724        return !list_empty(&pool->worklist) && __need_more_worker(pool);
 725}
 726
 727/* Can I start working?  Called from busy but !running workers. */
 728static bool may_start_working(struct worker_pool *pool)
 729{
 730        return pool->nr_idle;
 731}
 732
 733/* Do I need to keep working?  Called from currently running workers. */
 734static bool keep_working(struct worker_pool *pool)
 735{
 736        return !list_empty(&pool->worklist) &&
 737                atomic_read(&pool->nr_running) <= 1;
 738}
 739
 740/* Do we need a new worker?  Called from manager. */
 741static bool need_to_create_worker(struct worker_pool *pool)
 742{
 743        return need_more_worker(pool) && !may_start_working(pool);
 744}
 745
 746/* Do I need to be the manager? */
 747static bool need_to_manage_workers(struct worker_pool *pool)
 748{
 749        return need_to_create_worker(pool) ||
 750                (pool->flags & POOL_MANAGE_WORKERS);
 751}
 752
 753/* Do we have too many workers and should some go away? */
 754static bool too_many_workers(struct worker_pool *pool)
 755{
 756        bool managing = mutex_is_locked(&pool->manager_arb);
 757        int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
 758        int nr_busy = pool->nr_workers - nr_idle;
 759
 760        /*
 761         * nr_idle and idle_list may disagree if idle rebinding is in
 762         * progress.  Never return %true if idle_list is empty.
 763         */
 764        if (list_empty(&pool->idle_list))
 765                return false;
 766
 767        return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 768}
 769
 770/*
 771 * Wake up functions.
 772 */
 773
 774/* Return the first worker.  Safe with preemption disabled */
 775static struct worker *first_worker(struct worker_pool *pool)
 776{
 777        if (unlikely(list_empty(&pool->idle_list)))
 778                return NULL;
 779
 780        return list_first_entry(&pool->idle_list, struct worker, entry);
 781}
 782
 783/**
 784 * wake_up_worker - wake up an idle worker
 785 * @pool: worker pool to wake worker from
 786 *
 787 * Wake up the first idle worker of @pool.
 788 *
 789 * CONTEXT:
 790 * spin_lock_irq(pool->lock).
 791 */
 792static void wake_up_worker(struct worker_pool *pool)
 793{
 794        struct worker *worker = first_worker(pool);
 795
 796        if (likely(worker))
 797                wake_up_process(worker->task);
 798}
 799
 800/**
 801 * wq_worker_waking_up - a worker is waking up
 802 * @task: task waking up
 803 * @cpu: CPU @task is waking up to
 804 *
 805 * This function is called during try_to_wake_up() when a worker is
 806 * being awoken.
 807 *
 808 * CONTEXT:
 809 * spin_lock_irq(rq->lock)
 810 */
 811void wq_worker_waking_up(struct task_struct *task, int cpu)
 812{
 813        struct worker *worker = kthread_data(task);
 814
 815        if (!(worker->flags & WORKER_NOT_RUNNING)) {
 816                WARN_ON_ONCE(worker->pool->cpu != cpu);
 817                atomic_inc(&worker->pool->nr_running);
 818        }
 819}
 820
 821/**
 822 * wq_worker_sleeping - a worker is going to sleep
 823 * @task: task going to sleep
 824 * @cpu: CPU in question, must be the current CPU number
 825 *
 826 * This function is called during schedule() when a busy worker is
 827 * going to sleep.  Worker on the same cpu can be woken up by
 828 * returning pointer to its task.
 829 *
 830 * CONTEXT:
 831 * spin_lock_irq(rq->lock)
 832 *
 833 * RETURNS:
 834 * Worker task on @cpu to wake up, %NULL if none.
 835 */
 836struct task_struct *wq_worker_sleeping(struct task_struct *task, int cpu)
 837{
 838        struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 839        struct worker_pool *pool;
 840
 841        /*
 842         * Rescuers, which may not have all the fields set up like normal
 843         * workers, also reach here, let's not access anything before
 844         * checking NOT_RUNNING.
 845         */
 846        if (worker->flags & WORKER_NOT_RUNNING)
 847                return NULL;
 848
 849        pool = worker->pool;
 850
 851        /* this can only happen on the local cpu */
 852        if (WARN_ON_ONCE(cpu != raw_smp_processor_id()))
 853                return NULL;
 854
 855        /*
 856         * The counterpart of the following dec_and_test, implied mb,
 857         * worklist not empty test sequence is in insert_work().
 858         * Please read comment there.
 859         *
 860         * NOT_RUNNING is clear.  This means that we're bound to and
 861         * running on the local cpu w/ rq lock held and preemption
 862         * disabled, which in turn means that none else could be
 863         * manipulating idle_list, so dereferencing idle_list without pool
 864         * lock is safe.
 865         */
 866        if (atomic_dec_and_test(&pool->nr_running) &&
 867            !list_empty(&pool->worklist))
 868                to_wakeup = first_worker(pool);
 869        return to_wakeup ? to_wakeup->task : NULL;
 870}
 871
 872/**
 873 * worker_set_flags - set worker flags and adjust nr_running accordingly
 874 * @worker: self
 875 * @flags: flags to set
 876 * @wakeup: wakeup an idle worker if necessary
 877 *
 878 * Set @flags in @worker->flags and adjust nr_running accordingly.  If
 879 * nr_running becomes zero and @wakeup is %true, an idle worker is
 880 * woken up.
 881 *
 882 * CONTEXT:
 883 * spin_lock_irq(pool->lock)
 884 */
 885static inline void worker_set_flags(struct worker *worker, unsigned int flags,
 886                                    bool wakeup)
 887{
 888        struct worker_pool *pool = worker->pool;
 889
 890        WARN_ON_ONCE(worker->task != current);
 891
 892        /*
 893         * If transitioning into NOT_RUNNING, adjust nr_running and
 894         * wake up an idle worker as necessary if requested by
 895         * @wakeup.
 896         */
 897        if ((flags & WORKER_NOT_RUNNING) &&
 898            !(worker->flags & WORKER_NOT_RUNNING)) {
 899                if (wakeup) {
 900                        if (atomic_dec_and_test(&pool->nr_running) &&
 901                            !list_empty(&pool->worklist))
 902                                wake_up_worker(pool);
 903                } else
 904                        atomic_dec(&pool->nr_running);
 905        }
 906
 907        worker->flags |= flags;
 908}
 909
 910/**
 911 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 912 * @worker: self
 913 * @flags: flags to clear
 914 *
 915 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 916 *
 917 * CONTEXT:
 918 * spin_lock_irq(pool->lock)
 919 */
 920static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 921{
 922        struct worker_pool *pool = worker->pool;
 923        unsigned int oflags = worker->flags;
 924
 925        WARN_ON_ONCE(worker->task != current);
 926
 927        worker->flags &= ~flags;
 928
 929        /*
 930         * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 931         * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 932         * of multiple flags, not a single flag.
 933         */
 934        if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 935                if (!(worker->flags & WORKER_NOT_RUNNING))
 936                        atomic_inc(&pool->nr_running);
 937}
 938
 939/**
 940 * find_worker_executing_work - find worker which is executing a work
 941 * @pool: pool of interest
 942 * @work: work to find worker for
 943 *
 944 * Find a worker which is executing @work on @pool by searching
 945 * @pool->busy_hash which is keyed by the address of @work.  For a worker
 946 * to match, its current execution should match the address of @work and
 947 * its work function.  This is to avoid unwanted dependency between
 948 * unrelated work executions through a work item being recycled while still
 949 * being executed.
 950 *
 951 * This is a bit tricky.  A work item may be freed once its execution
 952 * starts and nothing prevents the freed area from being recycled for
 953 * another work item.  If the same work item address ends up being reused
 954 * before the original execution finishes, workqueue will identify the
 955 * recycled work item as currently executing and make it wait until the
 956 * current execution finishes, introducing an unwanted dependency.
 957 *
 958 * This function checks the work item address and work function to avoid
 959 * false positives.  Note that this isn't complete as one may construct a
 960 * work function which can introduce dependency onto itself through a
 961 * recycled work item.  Well, if somebody wants to shoot oneself in the
 962 * foot that badly, there's only so much we can do, and if such deadlock
 963 * actually occurs, it should be easy to locate the culprit work function.
 964 *
 965 * CONTEXT:
 966 * spin_lock_irq(pool->lock).
 967 *
 968 * RETURNS:
 969 * Pointer to worker which is executing @work if found, NULL
 970 * otherwise.
 971 */
 972static struct worker *find_worker_executing_work(struct worker_pool *pool,
 973                                                 struct work_struct *work)
 974{
 975        struct worker *worker;
 976
 977        hash_for_each_possible(pool->busy_hash, worker, hentry,
 978                               (unsigned long)work)
 979                if (worker->current_work == work &&
 980                    worker->current_func == work->func)
 981                        return worker;
 982
 983        return NULL;
 984}
 985
 986/**
 987 * move_linked_works - move linked works to a list
 988 * @work: start of series of works to be scheduled
 989 * @head: target list to append @work to
 990 * @nextp: out paramter for nested worklist walking
 991 *
 992 * Schedule linked works starting from @work to @head.  Work series to
 993 * be scheduled starts at @work and includes any consecutive work with
 994 * WORK_STRUCT_LINKED set in its predecessor.
 995 *
 996 * If @nextp is not NULL, it's updated to point to the next work of
 997 * the last scheduled work.  This allows move_linked_works() to be
 998 * nested inside outer list_for_each_entry_safe().
 999 *
1000 * CONTEXT:
1001 * spin_lock_irq(pool->lock).
1002 */
1003static void move_linked_works(struct work_struct *work, struct list_head *head,
1004                              struct work_struct **nextp)
1005{
1006        struct work_struct *n;
1007
1008        /*
1009         * Linked worklist will always end before the end of the list,
1010         * use NULL for list head.
1011         */
1012        list_for_each_entry_safe_from(work, n, NULL, entry) {
1013                list_move_tail(&work->entry, head);
1014                if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1015                        break;
1016        }
1017
1018        /*
1019         * If we're already inside safe list traversal and have moved
1020         * multiple works to the scheduled queue, the next position
1021         * needs to be updated.
1022         */
1023        if (nextp)
1024                *nextp = n;
1025}
1026
1027/**
1028 * get_pwq - get an extra reference on the specified pool_workqueue
1029 * @pwq: pool_workqueue to get
1030 *
1031 * Obtain an extra reference on @pwq.  The caller should guarantee that
1032 * @pwq has positive refcnt and be holding the matching pool->lock.
1033 */
1034static void get_pwq(struct pool_workqueue *pwq)
1035{
1036        lockdep_assert_held(&pwq->pool->lock);
1037        WARN_ON_ONCE(pwq->refcnt <= 0);
1038        pwq->refcnt++;
1039}
1040
1041/**
1042 * put_pwq - put a pool_workqueue reference
1043 * @pwq: pool_workqueue to put
1044 *
1045 * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1046 * destruction.  The caller should be holding the matching pool->lock.
1047 */
1048static void put_pwq(struct pool_workqueue *pwq)
1049{
1050        lockdep_assert_held(&pwq->pool->lock);
1051        if (likely(--pwq->refcnt))
1052                return;
1053        if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1054                return;
1055        /*
1056         * @pwq can't be released under pool->lock, bounce to
1057         * pwq_unbound_release_workfn().  This never recurses on the same
1058         * pool->lock as this path is taken only for unbound workqueues and
1059         * the release work item is scheduled on a per-cpu workqueue.  To
1060         * avoid lockdep warning, unbound pool->locks are given lockdep
1061         * subclass of 1 in get_unbound_pool().
1062         */
1063        schedule_work(&pwq->unbound_release_work);
1064}
1065
1066/**
1067 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1068 * @pwq: pool_workqueue to put (can be %NULL)
1069 *
1070 * put_pwq() with locking.  This function also allows %NULL @pwq.
1071 */
1072static void put_pwq_unlocked(struct pool_workqueue *pwq)
1073{
1074        if (pwq) {
1075                /*
1076                 * As both pwqs and pools are sched-RCU protected, the
1077                 * following lock operations are safe.
1078                 */
1079                spin_lock_irq(&pwq->pool->lock);
1080                put_pwq(pwq);
1081                spin_unlock_irq(&pwq->pool->lock);
1082        }
1083}
1084
1085static void pwq_activate_delayed_work(struct work_struct *work)
1086{
1087        struct pool_workqueue *pwq = get_work_pwq(work);
1088
1089        trace_workqueue_activate_work(work);
1090        move_linked_works(work, &pwq->pool->worklist, NULL);
1091        __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1092        pwq->nr_active++;
1093}
1094
1095static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1096{
1097        struct work_struct *work = list_first_entry(&pwq->delayed_works,
1098                                                    struct work_struct, entry);
1099
1100        pwq_activate_delayed_work(work);
1101}
1102
1103/**
1104 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1105 * @pwq: pwq of interest
1106 * @color: color of work which left the queue
1107 *
1108 * A work either has completed or is removed from pending queue,
1109 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1110 *
1111 * CONTEXT:
1112 * spin_lock_irq(pool->lock).
1113 */
1114static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1115{
1116        /* uncolored work items don't participate in flushing or nr_active */
1117        if (color == WORK_NO_COLOR)
1118                goto out_put;
1119
1120        pwq->nr_in_flight[color]--;
1121
1122        pwq->nr_active--;
1123        if (!list_empty(&pwq->delayed_works)) {
1124                /* one down, submit a delayed one */
1125                if (pwq->nr_active < pwq->max_active)
1126                        pwq_activate_first_delayed(pwq);
1127        }
1128
1129        /* is flush in progress and are we at the flushing tip? */
1130        if (likely(pwq->flush_color != color))
1131                goto out_put;
1132
1133        /* are there still in-flight works? */
1134        if (pwq->nr_in_flight[color])
1135                goto out_put;
1136
1137        /* this pwq is done, clear flush_color */
1138        pwq->flush_color = -1;
1139
1140        /*
1141         * If this was the last pwq, wake up the first flusher.  It
1142         * will handle the rest.
1143         */
1144        if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1145                complete(&pwq->wq->first_flusher->done);
1146out_put:
1147        put_pwq(pwq);
1148}
1149
1150/**
1151 * try_to_grab_pending - steal work item from worklist and disable irq
1152 * @work: work item to steal
1153 * @is_dwork: @work is a delayed_work
1154 * @flags: place to store irq state
1155 *
1156 * Try to grab PENDING bit of @work.  This function can handle @work in any
1157 * stable state - idle, on timer or on worklist.  Return values are
1158 *
1159 *  1           if @work was pending and we successfully stole PENDING
1160 *  0           if @work was idle and we claimed PENDING
1161 *  -EAGAIN     if PENDING couldn't be grabbed at the moment, safe to busy-retry
1162 *  -ENOENT     if someone else is canceling @work, this state may persist
1163 *              for arbitrarily long
1164 *
1165 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1166 * interrupted while holding PENDING and @work off queue, irq must be
1167 * disabled on entry.  This, combined with delayed_work->timer being
1168 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1169 *
1170 * On successful return, >= 0, irq is disabled and the caller is
1171 * responsible for releasing it using local_irq_restore(*@flags).
1172 *
1173 * This function is safe to call from any context including IRQ handler.
1174 */
1175static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1176                               unsigned long *flags)
1177{
1178        struct worker_pool *pool;
1179        struct pool_workqueue *pwq;
1180
1181        local_irq_save(*flags);
1182
1183        /* try to steal the timer if it exists */
1184        if (is_dwork) {
1185                struct delayed_work *dwork = to_delayed_work(work);
1186
1187                /*
1188                 * dwork->timer is irqsafe.  If del_timer() fails, it's
1189                 * guaranteed that the timer is not queued anywhere and not
1190                 * running on the local CPU.
1191                 */
1192                if (likely(del_timer(&dwork->timer)))
1193                        return 1;
1194        }
1195
1196        /* try to claim PENDING the normal way */
1197        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1198                return 0;
1199
1200        /*
1201         * The queueing is in progress, or it is already queued. Try to
1202         * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1203         */
1204        pool = get_work_pool(work);
1205        if (!pool)
1206                goto fail;
1207
1208        spin_lock(&pool->lock);
1209        /*
1210         * work->data is guaranteed to point to pwq only while the work
1211         * item is queued on pwq->wq, and both updating work->data to point
1212         * to pwq on queueing and to pool on dequeueing are done under
1213         * pwq->pool->lock.  This in turn guarantees that, if work->data
1214         * points to pwq which is associated with a locked pool, the work
1215         * item is currently queued on that pool.
1216         */
1217        pwq = get_work_pwq(work);
1218        if (pwq && pwq->pool == pool) {
1219                debug_work_deactivate(work);
1220
1221                /*
1222                 * A delayed work item cannot be grabbed directly because
1223                 * it might have linked NO_COLOR work items which, if left
1224                 * on the delayed_list, will confuse pwq->nr_active
1225                 * management later on and cause stall.  Make sure the work
1226                 * item is activated before grabbing.
1227                 */
1228                if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1229                        pwq_activate_delayed_work(work);
1230
1231                list_del_init(&work->entry);
1232                pwq_dec_nr_in_flight(get_work_pwq(work), get_work_color(work));
1233
1234                /* work->data points to pwq iff queued, point to pool */
1235                set_work_pool_and_keep_pending(work, pool->id);
1236
1237                spin_unlock(&pool->lock);
1238                return 1;
1239        }
1240        spin_unlock(&pool->lock);
1241fail:
1242        local_irq_restore(*flags);
1243        if (work_is_canceling(work))
1244                return -ENOENT;
1245        cpu_relax();
1246        return -EAGAIN;
1247}
1248
1249/**
1250 * insert_work - insert a work into a pool
1251 * @pwq: pwq @work belongs to
1252 * @work: work to insert
1253 * @head: insertion point
1254 * @extra_flags: extra WORK_STRUCT_* flags to set
1255 *
1256 * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1257 * work_struct flags.
1258 *
1259 * CONTEXT:
1260 * spin_lock_irq(pool->lock).
1261 */
1262static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1263                        struct list_head *head, unsigned int extra_flags)
1264{
1265        struct worker_pool *pool = pwq->pool;
1266
1267        /* we own @work, set data and link */
1268        set_work_pwq(work, pwq, extra_flags);
1269        list_add_tail(&work->entry, head);
1270        get_pwq(pwq);
1271
1272        /*
1273         * Ensure either wq_worker_sleeping() sees the above
1274         * list_add_tail() or we see zero nr_running to avoid workers lying
1275         * around lazily while there are works to be processed.
1276         */
1277        smp_mb();
1278
1279        if (__need_more_worker(pool))
1280                wake_up_worker(pool);
1281}
1282
1283/*
1284 * Test whether @work is being queued from another work executing on the
1285 * same workqueue.
1286 */
1287static bool is_chained_work(struct workqueue_struct *wq)
1288{
1289        struct worker *worker;
1290
1291        worker = current_wq_worker();
1292        /*
1293         * Return %true iff I'm a worker execuing a work item on @wq.  If
1294         * I'm @worker, it's safe to dereference it without locking.
1295         */
1296        return worker && worker->current_pwq->wq == wq;
1297}
1298
1299static void __queue_work(int cpu, struct workqueue_struct *wq,
1300                         struct work_struct *work)
1301{
1302        struct pool_workqueue *pwq;
1303        struct worker_pool *last_pool;
1304        struct list_head *worklist;
1305        unsigned int work_flags;
1306        unsigned int req_cpu = cpu;
1307
1308        /*
1309         * While a work item is PENDING && off queue, a task trying to
1310         * steal the PENDING will busy-loop waiting for it to either get
1311         * queued or lose PENDING.  Grabbing PENDING and queueing should
1312         * happen with IRQ disabled.
1313         */
1314        WARN_ON_ONCE(!irqs_disabled());
1315
1316        debug_work_activate(work);
1317
1318        /* if dying, only works from the same workqueue are allowed */
1319        if (unlikely(wq->flags & __WQ_DRAINING) &&
1320            WARN_ON_ONCE(!is_chained_work(wq)))
1321                return;
1322retry:
1323        if (req_cpu == WORK_CPU_UNBOUND)
1324                cpu = raw_smp_processor_id();
1325
1326        /* pwq which will be used unless @work is executing elsewhere */
1327        if (!(wq->flags & WQ_UNBOUND))
1328                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1329        else
1330                pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1331
1332        /*
1333         * If @work was previously on a different pool, it might still be
1334         * running there, in which case the work needs to be queued on that
1335         * pool to guarantee non-reentrancy.
1336         */
1337        last_pool = get_work_pool(work);
1338        if (last_pool && last_pool != pwq->pool) {
1339                struct worker *worker;
1340
1341                spin_lock(&last_pool->lock);
1342
1343                worker = find_worker_executing_work(last_pool, work);
1344
1345                if (worker && worker->current_pwq->wq == wq) {
1346                        pwq = worker->current_pwq;
1347                } else {
1348                        /* meh... not running there, queue here */
1349                        spin_unlock(&last_pool->lock);
1350                        spin_lock(&pwq->pool->lock);
1351                }
1352        } else {
1353                spin_lock(&pwq->pool->lock);
1354        }
1355
1356        /*
1357         * pwq is determined and locked.  For unbound pools, we could have
1358         * raced with pwq release and it could already be dead.  If its
1359         * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1360         * without another pwq replacing it in the numa_pwq_tbl or while
1361         * work items are executing on it, so the retrying is guaranteed to
1362         * make forward-progress.
1363         */
1364        if (unlikely(!pwq->refcnt)) {
1365                if (wq->flags & WQ_UNBOUND) {
1366                        spin_unlock(&pwq->pool->lock);
1367                        cpu_relax();
1368                        goto retry;
1369                }
1370                /* oops */
1371                WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1372                          wq->name, cpu);
1373        }
1374
1375        /* pwq determined, queue */
1376        trace_workqueue_queue_work(req_cpu, pwq, work);
1377
1378        if (WARN_ON(!list_empty(&work->entry))) {
1379                spin_unlock(&pwq->pool->lock);
1380                return;
1381        }
1382
1383        pwq->nr_in_flight[pwq->work_color]++;
1384        work_flags = work_color_to_flags(pwq->work_color);
1385
1386        if (likely(pwq->nr_active < pwq->max_active)) {
1387                trace_workqueue_activate_work(work);
1388                pwq->nr_active++;
1389                worklist = &pwq->pool->worklist;
1390        } else {
1391                work_flags |= WORK_STRUCT_DELAYED;
1392                worklist = &pwq->delayed_works;
1393        }
1394
1395        insert_work(pwq, work, worklist, work_flags);
1396
1397        spin_unlock(&pwq->pool->lock);
1398}
1399
1400/**
1401 * queue_work_on - queue work on specific cpu
1402 * @cpu: CPU number to execute work on
1403 * @wq: workqueue to use
1404 * @work: work to queue
1405 *
1406 * Returns %false if @work was already on a queue, %true otherwise.
1407 *
1408 * We queue the work to a specific CPU, the caller must ensure it
1409 * can't go away.
1410 */
1411bool queue_work_on(int cpu, struct workqueue_struct *wq,
1412                   struct work_struct *work)
1413{
1414        bool ret = false;
1415        unsigned long flags;
1416
1417        local_irq_save(flags);
1418
1419        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1420                __queue_work(cpu, wq, work);
1421                ret = true;
1422        }
1423
1424        local_irq_restore(flags);
1425        return ret;
1426}
1427EXPORT_SYMBOL(queue_work_on);
1428
1429void delayed_work_timer_fn(unsigned long __data)
1430{
1431        struct delayed_work *dwork = (struct delayed_work *)__data;
1432
1433        /* should have been called from irqsafe timer with irq already off */
1434        __queue_work(dwork->cpu, dwork->wq, &dwork->work);
1435}
1436EXPORT_SYMBOL(delayed_work_timer_fn);
1437
1438static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1439                                struct delayed_work *dwork, unsigned long delay)
1440{
1441        struct timer_list *timer = &dwork->timer;
1442        struct work_struct *work = &dwork->work;
1443
1444        WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1445                     timer->data != (unsigned long)dwork);
1446        WARN_ON_ONCE(timer_pending(timer));
1447        WARN_ON_ONCE(!list_empty(&work->entry));
1448
1449        /*
1450         * If @delay is 0, queue @dwork->work immediately.  This is for
1451         * both optimization and correctness.  The earliest @timer can
1452         * expire is on the closest next tick and delayed_work users depend
1453         * on that there's no such delay when @delay is 0.
1454         */
1455        if (!delay) {
1456                __queue_work(cpu, wq, &dwork->work);
1457                return;
1458        }
1459
1460        timer_stats_timer_set_start_info(&dwork->timer);
1461
1462        dwork->wq = wq;
1463        dwork->cpu = cpu;
1464        timer->expires = jiffies + delay;
1465
1466        if (unlikely(cpu != WORK_CPU_UNBOUND))
1467                add_timer_on(timer, cpu);
1468        else
1469                add_timer(timer);
1470}
1471
1472/**
1473 * queue_delayed_work_on - queue work on specific CPU after delay
1474 * @cpu: CPU number to execute work on
1475 * @wq: workqueue to use
1476 * @dwork: work to queue
1477 * @delay: number of jiffies to wait before queueing
1478 *
1479 * Returns %false if @work was already on a queue, %true otherwise.  If
1480 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1481 * execution.
1482 */
1483bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1484                           struct delayed_work *dwork, unsigned long delay)
1485{
1486        struct work_struct *work = &dwork->work;
1487        bool ret = false;
1488        unsigned long flags;
1489
1490        /* read the comment in __queue_work() */
1491        local_irq_save(flags);
1492
1493        if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1494                __queue_delayed_work(cpu, wq, dwork, delay);
1495                ret = true;
1496        }
1497
1498        local_irq_restore(flags);
1499        return ret;
1500}
1501EXPORT_SYMBOL(queue_delayed_work_on);
1502
1503/**
1504 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1505 * @cpu: CPU number to execute work on
1506 * @wq: workqueue to use
1507 * @dwork: work to queue
1508 * @delay: number of jiffies to wait before queueing
1509 *
1510 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1511 * modify @dwork's timer so that it expires after @delay.  If @delay is
1512 * zero, @work is guaranteed to be scheduled immediately regardless of its
1513 * current state.
1514 *
1515 * Returns %false if @dwork was idle and queued, %true if @dwork was
1516 * pending and its timer was modified.
1517 *
1518 * This function is safe to call from any context including IRQ handler.
1519 * See try_to_grab_pending() for details.
1520 */
1521bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1522                         struct delayed_work *dwork, unsigned long delay)
1523{
1524        unsigned long flags;
1525        int ret;
1526
1527        do {
1528                ret = try_to_grab_pending(&dwork->work, true, &flags);
1529        } while (unlikely(ret == -EAGAIN));
1530
1531        if (likely(ret >= 0)) {
1532                __queue_delayed_work(cpu, wq, dwork, delay);
1533                local_irq_restore(flags);
1534        }
1535
1536        /* -ENOENT from try_to_grab_pending() becomes %true */
1537        return ret;
1538}
1539EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1540
1541/**
1542 * worker_enter_idle - enter idle state
1543 * @worker: worker which is entering idle state
1544 *
1545 * @worker is entering idle state.  Update stats and idle timer if
1546 * necessary.
1547 *
1548 * LOCKING:
1549 * spin_lock_irq(pool->lock).
1550 */
1551static void worker_enter_idle(struct worker *worker)
1552{
1553        struct worker_pool *pool = worker->pool;
1554
1555        if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1556            WARN_ON_ONCE(!list_empty(&worker->entry) &&
1557                         (worker->hentry.next || worker->hentry.pprev)))
1558                return;
1559
1560        /* can't use worker_set_flags(), also called from start_worker() */
1561        worker->flags |= WORKER_IDLE;
1562        pool->nr_idle++;
1563        worker->last_active = jiffies;
1564
1565        /* idle_list is LIFO */
1566        list_add(&worker->entry, &pool->idle_list);
1567
1568        if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1569                mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1570
1571        /*
1572         * Sanity check nr_running.  Because wq_unbind_fn() releases
1573         * pool->lock between setting %WORKER_UNBOUND and zapping
1574         * nr_running, the warning may trigger spuriously.  Check iff
1575         * unbind is not in progress.
1576         */
1577        WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1578                     pool->nr_workers == pool->nr_idle &&
1579                     atomic_read(&pool->nr_running));
1580}
1581
1582/**
1583 * worker_leave_idle - leave idle state
1584 * @worker: worker which is leaving idle state
1585 *
1586 * @worker is leaving idle state.  Update stats.
1587 *
1588 * LOCKING:
1589 * spin_lock_irq(pool->lock).
1590 */
1591static void worker_leave_idle(struct worker *worker)
1592{
1593        struct worker_pool *pool = worker->pool;
1594
1595        if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1596                return;
1597        worker_clr_flags(worker, WORKER_IDLE);
1598        pool->nr_idle--;
1599        list_del_init(&worker->entry);
1600}
1601
1602/**
1603 * worker_maybe_bind_and_lock - try to bind %current to worker_pool and lock it
1604 * @pool: target worker_pool
1605 *
1606 * Bind %current to the cpu of @pool if it is associated and lock @pool.
1607 *
1608 * Works which are scheduled while the cpu is online must at least be
1609 * scheduled to a worker which is bound to the cpu so that if they are
1610 * flushed from cpu callbacks while cpu is going down, they are
1611 * guaranteed to execute on the cpu.
1612 *
1613 * This function is to be used by unbound workers and rescuers to bind
1614 * themselves to the target cpu and may race with cpu going down or
1615 * coming online.  kthread_bind() can't be used because it may put the
1616 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1617 * verbatim as it's best effort and blocking and pool may be
1618 * [dis]associated in the meantime.
1619 *
1620 * This function tries set_cpus_allowed() and locks pool and verifies the
1621 * binding against %POOL_DISASSOCIATED which is set during
1622 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1623 * enters idle state or fetches works without dropping lock, it can
1624 * guarantee the scheduling requirement described in the first paragraph.
1625 *
1626 * CONTEXT:
1627 * Might sleep.  Called without any lock but returns with pool->lock
1628 * held.
1629 *
1630 * RETURNS:
1631 * %true if the associated pool is online (@worker is successfully
1632 * bound), %false if offline.
1633 */
1634static bool worker_maybe_bind_and_lock(struct worker_pool *pool)
1635__acquires(&pool->lock)
1636{
1637        while (true) {
1638                /*
1639                 * The following call may fail, succeed or succeed
1640                 * without actually migrating the task to the cpu if
1641                 * it races with cpu hotunplug operation.  Verify
1642                 * against POOL_DISASSOCIATED.
1643                 */
1644                if (!(pool->flags & POOL_DISASSOCIATED))
1645                        set_cpus_allowed_ptr(current, pool->attrs->cpumask);
1646
1647                spin_lock_irq(&pool->lock);
1648                if (pool->flags & POOL_DISASSOCIATED)
1649                        return false;
1650                if (task_cpu(current) == pool->cpu &&
1651                    cpumask_equal(&current->cpus_allowed, pool->attrs->cpumask))
1652                        return true;
1653                spin_unlock_irq(&pool->lock);
1654
1655                /*
1656                 * We've raced with CPU hot[un]plug.  Give it a breather
1657                 * and retry migration.  cond_resched() is required here;
1658                 * otherwise, we might deadlock against cpu_stop trying to
1659                 * bring down the CPU on non-preemptive kernel.
1660                 */
1661                cpu_relax();
1662                cond_resched();
1663        }
1664}
1665
1666static struct worker *alloc_worker(void)
1667{
1668        struct worker *worker;
1669
1670        worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1671        if (worker) {
1672                INIT_LIST_HEAD(&worker->entry);
1673                INIT_LIST_HEAD(&worker->scheduled);
1674                /* on creation a worker is in !idle && prep state */
1675                worker->flags = WORKER_PREP;
1676        }
1677        return worker;
1678}
1679
1680/**
1681 * create_worker - create a new workqueue worker
1682 * @pool: pool the new worker will belong to
1683 *
1684 * Create a new worker which is bound to @pool.  The returned worker
1685 * can be started by calling start_worker() or destroyed using
1686 * destroy_worker().
1687 *
1688 * CONTEXT:
1689 * Might sleep.  Does GFP_KERNEL allocations.
1690 *
1691 * RETURNS:
1692 * Pointer to the newly created worker.
1693 */
1694static struct worker *create_worker(struct worker_pool *pool)
1695{
1696        struct worker *worker = NULL;
1697        int id = -1;
1698        char id_buf[16];
1699
1700        lockdep_assert_held(&pool->manager_mutex);
1701
1702        /*
1703         * ID is needed to determine kthread name.  Allocate ID first
1704         * without installing the pointer.
1705         */
1706        idr_preload(GFP_KERNEL);
1707        spin_lock_irq(&pool->lock);
1708
1709        id = idr_alloc(&pool->worker_idr, NULL, 0, 0, GFP_NOWAIT);
1710
1711        spin_unlock_irq(&pool->lock);
1712        idr_preload_end();
1713        if (id < 0)
1714                goto fail;
1715
1716        worker = alloc_worker();
1717        if (!worker)
1718                goto fail;
1719
1720        worker->pool = pool;
1721        worker->id = id;
1722
1723        if (pool->cpu >= 0)
1724                snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1725                         pool->attrs->nice < 0  ? "H" : "");
1726        else
1727                snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1728
1729        worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1730                                              "kworker/%s", id_buf);
1731        if (IS_ERR(worker->task))
1732                goto fail;
1733
1734        /*
1735         * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1736         * online CPUs.  It'll be re-applied when any of the CPUs come up.
1737         */
1738        set_user_nice(worker->task, pool->attrs->nice);
1739        set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1740
1741        /* prevent userland from meddling with cpumask of workqueue workers */
1742        worker->task->flags |= PF_NO_SETAFFINITY;
1743
1744        /*
1745         * The caller is responsible for ensuring %POOL_DISASSOCIATED
1746         * remains stable across this function.  See the comments above the
1747         * flag definition for details.
1748         */
1749        if (pool->flags & POOL_DISASSOCIATED)
1750                worker->flags |= WORKER_UNBOUND;
1751
1752        /* successful, commit the pointer to idr */
1753        spin_lock_irq(&pool->lock);
1754        idr_replace(&pool->worker_idr, worker, worker->id);
1755        spin_unlock_irq(&pool->lock);
1756
1757        return worker;
1758
1759fail:
1760        if (id >= 0) {
1761                spin_lock_irq(&pool->lock);
1762                idr_remove(&pool->worker_idr, id);
1763                spin_unlock_irq(&pool->lock);
1764        }
1765        kfree(worker);
1766        return NULL;
1767}
1768
1769/**
1770 * start_worker - start a newly created worker
1771 * @worker: worker to start
1772 *
1773 * Make the pool aware of @worker and start it.
1774 *
1775 * CONTEXT:
1776 * spin_lock_irq(pool->lock).
1777 */
1778static void start_worker(struct worker *worker)
1779{
1780        worker->flags |= WORKER_STARTED;
1781        worker->pool->nr_workers++;
1782        worker_enter_idle(worker);
1783        wake_up_process(worker->task);
1784}
1785
1786/**
1787 * create_and_start_worker - create and start a worker for a pool
1788 * @pool: the target pool
1789 *
1790 * Grab the managership of @pool and create and start a new worker for it.
1791 */
1792static int create_and_start_worker(struct worker_pool *pool)
1793{
1794        struct worker *worker;
1795
1796        mutex_lock(&pool->manager_mutex);
1797
1798        worker = create_worker(pool);
1799        if (worker) {
1800                spin_lock_irq(&pool->lock);
1801                start_worker(worker);
1802                spin_unlock_irq(&pool->lock);
1803        }
1804
1805        mutex_unlock(&pool->manager_mutex);
1806
1807        return worker ? 0 : -ENOMEM;
1808}
1809
1810/**
1811 * destroy_worker - destroy a workqueue worker
1812 * @worker: worker to be destroyed
1813 *
1814 * Destroy @worker and adjust @pool stats accordingly.
1815 *
1816 * CONTEXT:
1817 * spin_lock_irq(pool->lock) which is released and regrabbed.
1818 */
1819static void destroy_worker(struct worker *worker)
1820{
1821        struct worker_pool *pool = worker->pool;
1822
1823        lockdep_assert_held(&pool->manager_mutex);
1824        lockdep_assert_held(&pool->lock);
1825
1826        /* sanity check frenzy */
1827        if (WARN_ON(worker->current_work) ||
1828            WARN_ON(!list_empty(&worker->scheduled)))
1829                return;
1830
1831        if (worker->flags & WORKER_STARTED)
1832                pool->nr_workers--;
1833        if (worker->flags & WORKER_IDLE)
1834                pool->nr_idle--;
1835
1836        list_del_init(&worker->entry);
1837        worker->flags |= WORKER_DIE;
1838
1839        idr_remove(&pool->worker_idr, worker->id);
1840
1841        spin_unlock_irq(&pool->lock);
1842
1843        kthread_stop(worker->task);
1844        kfree(worker);
1845
1846        spin_lock_irq(&pool->lock);
1847}
1848
1849static void idle_worker_timeout(unsigned long __pool)
1850{
1851        struct worker_pool *pool = (void *)__pool;
1852
1853        spin_lock_irq(&pool->lock);
1854
1855        if (too_many_workers(pool)) {
1856                struct worker *worker;
1857                unsigned long expires;
1858
1859                /* idle_list is kept in LIFO order, check the last one */
1860                worker = list_entry(pool->idle_list.prev, struct worker, entry);
1861                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1862
1863                if (time_before(jiffies, expires))
1864                        mod_timer(&pool->idle_timer, expires);
1865                else {
1866                        /* it's been idle for too long, wake up manager */
1867                        pool->flags |= POOL_MANAGE_WORKERS;
1868                        wake_up_worker(pool);
1869                }
1870        }
1871
1872        spin_unlock_irq(&pool->lock);
1873}
1874
1875static void send_mayday(struct work_struct *work)
1876{
1877        struct pool_workqueue *pwq = get_work_pwq(work);
1878        struct workqueue_struct *wq = pwq->wq;
1879
1880        lockdep_assert_held(&wq_mayday_lock);
1881
1882        if (!wq->rescuer)
1883                return;
1884
1885        /* mayday mayday mayday */
1886        if (list_empty(&pwq->mayday_node)) {
1887                list_add_tail(&pwq->mayday_node, &wq->maydays);
1888                wake_up_process(wq->rescuer->task);
1889        }
1890}
1891
1892static void pool_mayday_timeout(unsigned long __pool)
1893{
1894        struct worker_pool *pool = (void *)__pool;
1895        struct work_struct *work;
1896
1897        spin_lock_irq(&wq_mayday_lock);         /* for wq->maydays */
1898        spin_lock(&pool->lock);
1899
1900        if (need_to_create_worker(pool)) {
1901                /*
1902                 * We've been trying to create a new worker but
1903                 * haven't been successful.  We might be hitting an
1904                 * allocation deadlock.  Send distress signals to
1905                 * rescuers.
1906                 */
1907                list_for_each_entry(work, &pool->worklist, entry)
1908                        send_mayday(work);
1909        }
1910
1911        spin_unlock(&pool->lock);
1912        spin_unlock_irq(&wq_mayday_lock);
1913
1914        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1915}
1916
1917/**
1918 * maybe_create_worker - create a new worker if necessary
1919 * @pool: pool to create a new worker for
1920 *
1921 * Create a new worker for @pool if necessary.  @pool is guaranteed to
1922 * have at least one idle worker on return from this function.  If
1923 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1924 * sent to all rescuers with works scheduled on @pool to resolve
1925 * possible allocation deadlock.
1926 *
1927 * On return, need_to_create_worker() is guaranteed to be %false and
1928 * may_start_working() %true.
1929 *
1930 * LOCKING:
1931 * spin_lock_irq(pool->lock) which may be released and regrabbed
1932 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1933 * manager.
1934 *
1935 * RETURNS:
1936 * %false if no action was taken and pool->lock stayed locked, %true
1937 * otherwise.
1938 */
1939static bool maybe_create_worker(struct worker_pool *pool)
1940__releases(&pool->lock)
1941__acquires(&pool->lock)
1942{
1943        if (!need_to_create_worker(pool))
1944                return false;
1945restart:
1946        spin_unlock_irq(&pool->lock);
1947
1948        /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1949        mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1950
1951        while (true) {
1952                struct worker *worker;
1953
1954                worker = create_worker(pool);
1955                if (worker) {
1956                        del_timer_sync(&pool->mayday_timer);
1957                        spin_lock_irq(&pool->lock);
1958                        start_worker(worker);
1959                        if (WARN_ON_ONCE(need_to_create_worker(pool)))
1960                                goto restart;
1961                        return true;
1962                }
1963
1964                if (!need_to_create_worker(pool))
1965                        break;
1966
1967                __set_current_state(TASK_INTERRUPTIBLE);
1968                schedule_timeout(CREATE_COOLDOWN);
1969
1970                if (!need_to_create_worker(pool))
1971                        break;
1972        }
1973
1974        del_timer_sync(&pool->mayday_timer);
1975        spin_lock_irq(&pool->lock);
1976        if (need_to_create_worker(pool))
1977                goto restart;
1978        return true;
1979}
1980
1981/**
1982 * maybe_destroy_worker - destroy workers which have been idle for a while
1983 * @pool: pool to destroy workers for
1984 *
1985 * Destroy @pool workers which have been idle for longer than
1986 * IDLE_WORKER_TIMEOUT.
1987 *
1988 * LOCKING:
1989 * spin_lock_irq(pool->lock) which may be released and regrabbed
1990 * multiple times.  Called only from manager.
1991 *
1992 * RETURNS:
1993 * %false if no action was taken and pool->lock stayed locked, %true
1994 * otherwise.
1995 */
1996static bool maybe_destroy_workers(struct worker_pool *pool)
1997{
1998        bool ret = false;
1999
2000        while (too_many_workers(pool)) {
2001                struct worker *worker;
2002                unsigned long expires;
2003
2004                worker = list_entry(pool->idle_list.prev, struct worker, entry);
2005                expires = worker->last_active + IDLE_WORKER_TIMEOUT;
2006
2007                if (time_before(jiffies, expires)) {
2008                        mod_timer(&pool->idle_timer, expires);
2009                        break;
2010                }
2011
2012                destroy_worker(worker);
2013                ret = true;
2014        }
2015
2016        return ret;
2017}
2018
2019/**
2020 * manage_workers - manage worker pool
2021 * @worker: self
2022 *
2023 * Assume the manager role and manage the worker pool @worker belongs
2024 * to.  At any given time, there can be only zero or one manager per
2025 * pool.  The exclusion is handled automatically by this function.
2026 *
2027 * The caller can safely start processing works on false return.  On
2028 * true return, it's guaranteed that need_to_create_worker() is false
2029 * and may_start_working() is true.
2030 *
2031 * CONTEXT:
2032 * spin_lock_irq(pool->lock) which may be released and regrabbed
2033 * multiple times.  Does GFP_KERNEL allocations.
2034 *
2035 * RETURNS:
2036 * spin_lock_irq(pool->lock) which may be released and regrabbed
2037 * multiple times.  Does GFP_KERNEL allocations.
2038 */
2039static bool manage_workers(struct worker *worker)
2040{
2041        struct worker_pool *pool = worker->pool;
2042        bool ret = false;
2043
2044        /*
2045         * Managership is governed by two mutexes - manager_arb and
2046         * manager_mutex.  manager_arb handles arbitration of manager role.
2047         * Anyone who successfully grabs manager_arb wins the arbitration
2048         * and becomes the manager.  mutex_trylock() on pool->manager_arb
2049         * failure while holding pool->lock reliably indicates that someone
2050         * else is managing the pool and the worker which failed trylock
2051         * can proceed to executing work items.  This means that anyone
2052         * grabbing manager_arb is responsible for actually performing
2053         * manager duties.  If manager_arb is grabbed and released without
2054         * actual management, the pool may stall indefinitely.
2055         *
2056         * manager_mutex is used for exclusion of actual management
2057         * operations.  The holder of manager_mutex can be sure that none
2058         * of management operations, including creation and destruction of
2059         * workers, won't take place until the mutex is released.  Because
2060         * manager_mutex doesn't interfere with manager role arbitration,
2061         * it is guaranteed that the pool's management, while may be
2062         * delayed, won't be disturbed by someone else grabbing
2063         * manager_mutex.
2064         */
2065        if (!mutex_trylock(&pool->manager_arb))
2066                return ret;
2067
2068        /*
2069         * With manager arbitration won, manager_mutex would be free in
2070         * most cases.  trylock first without dropping @pool->lock.
2071         */
2072        if (unlikely(!mutex_trylock(&pool->manager_mutex))) {
2073                spin_unlock_irq(&pool->lock);
2074                mutex_lock(&pool->manager_mutex);
2075                spin_lock_irq(&pool->lock);
2076                ret = true;
2077        }
2078
2079        pool->flags &= ~POOL_MANAGE_WORKERS;
2080
2081        /*
2082         * Destroy and then create so that may_start_working() is true
2083         * on return.
2084         */
2085        ret |= maybe_destroy_workers(pool);
2086        ret |= maybe_create_worker(pool);
2087
2088        mutex_unlock(&pool->manager_mutex);
2089        mutex_unlock(&pool->manager_arb);
2090        return ret;
2091}
2092
2093/**
2094 * process_one_work - process single work
2095 * @worker: self
2096 * @work: work to process
2097 *
2098 * Process @work.  This function contains all the logics necessary to
2099 * process a single work including synchronization against and
2100 * interaction with other workers on the same cpu, queueing and
2101 * flushing.  As long as context requirement is met, any worker can
2102 * call this function to process a work.
2103 *
2104 * CONTEXT:
2105 * spin_lock_irq(pool->lock) which is released and regrabbed.
2106 */
2107static void process_one_work(struct worker *worker, struct work_struct *work)
2108__releases(&pool->lock)
2109__acquires(&pool->lock)
2110{
2111        struct pool_workqueue *pwq = get_work_pwq(work);
2112        struct worker_pool *pool = worker->pool;
2113        bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2114        int work_color;
2115        struct worker *collision;
2116#ifdef CONFIG_LOCKDEP
2117        /*
2118         * It is permissible to free the struct work_struct from
2119         * inside the function that is called from it, this we need to
2120         * take into account for lockdep too.  To avoid bogus "held
2121         * lock freed" warnings as well as problems when looking into
2122         * work->lockdep_map, make a copy and use that here.
2123         */
2124        struct lockdep_map lockdep_map;
2125
2126        lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2127#endif
2128        /*
2129         * Ensure we're on the correct CPU.  DISASSOCIATED test is
2130         * necessary to avoid spurious warnings from rescuers servicing the
2131         * unbound or a disassociated pool.
2132         */
2133        WARN_ON_ONCE(!(worker->flags & WORKER_UNBOUND) &&
2134                     !(pool->flags & POOL_DISASSOCIATED) &&
2135                     raw_smp_processor_id() != pool->cpu);
2136
2137        /*
2138         * A single work shouldn't be executed concurrently by
2139         * multiple workers on a single cpu.  Check whether anyone is
2140         * already processing the work.  If so, defer the work to the
2141         * currently executing one.
2142         */
2143        collision = find_worker_executing_work(pool, work);
2144        if (unlikely(collision)) {
2145                move_linked_works(work, &collision->scheduled, NULL);
2146                return;
2147        }
2148
2149        /* claim and dequeue */
2150        debug_work_deactivate(work);
2151        hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2152        worker->current_work = work;
2153        worker->current_func = work->func;
2154        worker->current_pwq = pwq;
2155        work_color = get_work_color(work);
2156
2157        list_del_init(&work->entry);
2158
2159        /*
2160         * CPU intensive works don't participate in concurrency
2161         * management.  They're the scheduler's responsibility.
2162         */
2163        if (unlikely(cpu_intensive))
2164                worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2165
2166        /*
2167         * Unbound pool isn't concurrency managed and work items should be
2168         * executed ASAP.  Wake up another worker if necessary.
2169         */
2170        if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2171                wake_up_worker(pool);
2172
2173        /*
2174         * Record the last pool and clear PENDING which should be the last
2175         * update to @work.  Also, do this inside @pool->lock so that
2176         * PENDING and queued state changes happen together while IRQ is
2177         * disabled.
2178         */
2179        set_work_pool_and_clear_pending(work, pool->id);
2180
2181        spin_unlock_irq(&pool->lock);
2182
2183        lock_map_acquire_read(&pwq->wq->lockdep_map);
2184        lock_map_acquire(&lockdep_map);
2185        trace_workqueue_execute_start(work);
2186        worker->current_func(work);
2187        /*
2188         * While we must be careful to not use "work" after this, the trace
2189         * point will only record its address.
2190         */
2191        trace_workqueue_execute_end(work);
2192        lock_map_release(&lockdep_map);
2193        lock_map_release(&pwq->wq->lockdep_map);
2194
2195        if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2196                pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2197                       "     last function: %pf\n",
2198                       current->comm, preempt_count(), task_pid_nr(current),
2199                       worker->current_func);
2200                debug_show_held_locks(current);
2201                dump_stack();
2202        }
2203
2204        /*
2205         * The following prevents a kworker from hogging CPU on !PREEMPT
2206         * kernels, where a requeueing work item waiting for something to
2207         * happen could deadlock with stop_machine as such work item could
2208         * indefinitely requeue itself while all other CPUs are trapped in
2209         * stop_machine.
2210         */
2211        cond_resched();
2212
2213        spin_lock_irq(&pool->lock);
2214
2215        /* clear cpu intensive status */
2216        if (unlikely(cpu_intensive))
2217                worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2218
2219        /* we're done with it, release */
2220        hash_del(&worker->hentry);
2221        worker->current_work = NULL;
2222        worker->current_func = NULL;
2223        worker->current_pwq = NULL;
2224        worker->desc_valid = false;
2225        pwq_dec_nr_in_flight(pwq, work_color);
2226}
2227
2228/**
2229 * process_scheduled_works - process scheduled works
2230 * @worker: self
2231 *
2232 * Process all scheduled works.  Please note that the scheduled list
2233 * may change while processing a work, so this function repeatedly
2234 * fetches a work from the top and executes it.
2235 *
2236 * CONTEXT:
2237 * spin_lock_irq(pool->lock) which may be released and regrabbed
2238 * multiple times.
2239 */
2240static void process_scheduled_works(struct worker *worker)
2241{
2242        while (!list_empty(&worker->scheduled)) {
2243                struct work_struct *work = list_first_entry(&worker->scheduled,
2244                                                struct work_struct, entry);
2245                process_one_work(worker, work);
2246        }
2247}
2248
2249/**
2250 * worker_thread - the worker thread function
2251 * @__worker: self
2252 *
2253 * The worker thread function.  All workers belong to a worker_pool -
2254 * either a per-cpu one or dynamic unbound one.  These workers process all
2255 * work items regardless of their specific target workqueue.  The only
2256 * exception is work items which belong to workqueues with a rescuer which
2257 * will be explained in rescuer_thread().
2258 */
2259static int worker_thread(void *__worker)
2260{
2261        struct worker *worker = __worker;
2262        struct worker_pool *pool = worker->pool;
2263
2264        /* tell the scheduler that this is a workqueue worker */
2265        worker->task->flags |= PF_WQ_WORKER;
2266woke_up:
2267        spin_lock_irq(&pool->lock);
2268
2269        /* am I supposed to die? */
2270        if (unlikely(worker->flags & WORKER_DIE)) {
2271                spin_unlock_irq(&pool->lock);
2272                WARN_ON_ONCE(!list_empty(&worker->entry));
2273                worker->task->flags &= ~PF_WQ_WORKER;
2274                return 0;
2275        }
2276
2277        worker_leave_idle(worker);
2278recheck:
2279        /* no more worker necessary? */
2280        if (!need_more_worker(pool))
2281                goto sleep;
2282
2283        /* do we need to manage? */
2284        if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2285                goto recheck;
2286
2287        /*
2288         * ->scheduled list can only be filled while a worker is
2289         * preparing to process a work or actually processing it.
2290         * Make sure nobody diddled with it while I was sleeping.
2291         */
2292        WARN_ON_ONCE(!list_empty(&worker->scheduled));
2293
2294        /*
2295         * Finish PREP stage.  We're guaranteed to have at least one idle
2296         * worker or that someone else has already assumed the manager
2297         * role.  This is where @worker starts participating in concurrency
2298         * management if applicable and concurrency management is restored
2299         * after being rebound.  See rebind_workers() for details.
2300         */
2301        worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2302
2303        do {
2304                struct work_struct *work =
2305                        list_first_entry(&pool->worklist,
2306                                         struct work_struct, entry);
2307
2308                if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2309                        /* optimization path, not strictly necessary */
2310                        process_one_work(worker, work);
2311                        if (unlikely(!list_empty(&worker->scheduled)))
2312                                process_scheduled_works(worker);
2313                } else {
2314                        move_linked_works(work, &worker->scheduled, NULL);
2315                        process_scheduled_works(worker);
2316                }
2317        } while (keep_working(pool));
2318
2319        worker_set_flags(worker, WORKER_PREP, false);
2320sleep:
2321        if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
2322                goto recheck;
2323
2324        /*
2325         * pool->lock is held and there's no work to process and no need to
2326         * manage, sleep.  Workers are woken up only while holding
2327         * pool->lock or from local cpu, so setting the current state
2328         * before releasing pool->lock is enough to prevent losing any
2329         * event.
2330         */
2331        worker_enter_idle(worker);
2332        __set_current_state(TASK_INTERRUPTIBLE);
2333        spin_unlock_irq(&pool->lock);
2334        schedule();
2335        goto woke_up;
2336}
2337
2338/**
2339 * rescuer_thread - the rescuer thread function
2340 * @__rescuer: self
2341 *
2342 * Workqueue rescuer thread function.  There's one rescuer for each
2343 * workqueue which has WQ_MEM_RECLAIM set.
2344 *
2345 * Regular work processing on a pool may block trying to create a new
2346 * worker which uses GFP_KERNEL allocation which has slight chance of
2347 * developing into deadlock if some works currently on the same queue
2348 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2349 * the problem rescuer solves.
2350 *
2351 * When such condition is possible, the pool summons rescuers of all
2352 * workqueues which have works queued on the pool and let them process
2353 * those works so that forward progress can be guaranteed.
2354 *
2355 * This should happen rarely.
2356 */
2357static int rescuer_thread(void *__rescuer)
2358{
2359        struct worker *rescuer = __rescuer;
2360        struct workqueue_struct *wq = rescuer->rescue_wq;
2361        struct list_head *scheduled = &rescuer->scheduled;
2362
2363        set_user_nice(current, RESCUER_NICE_LEVEL);
2364
2365        /*
2366         * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2367         * doesn't participate in concurrency management.
2368         */
2369        rescuer->task->flags |= PF_WQ_WORKER;
2370repeat:
2371        set_current_state(TASK_INTERRUPTIBLE);
2372
2373        if (kthread_should_stop()) {
2374                __set_current_state(TASK_RUNNING);
2375                rescuer->task->flags &= ~PF_WQ_WORKER;
2376                return 0;
2377        }
2378
2379        /* see whether any pwq is asking for help */
2380        spin_lock_irq(&wq_mayday_lock);
2381
2382        while (!list_empty(&wq->maydays)) {
2383                struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2384                                        struct pool_workqueue, mayday_node);
2385                struct worker_pool *pool = pwq->pool;
2386                struct work_struct *work, *n;
2387
2388                __set_current_state(TASK_RUNNING);
2389                list_del_init(&pwq->mayday_node);
2390
2391                spin_unlock_irq(&wq_mayday_lock);
2392
2393                /* migrate to the target cpu if possible */
2394                worker_maybe_bind_and_lock(pool);
2395                rescuer->pool = pool;
2396
2397                /*
2398                 * Slurp in all works issued via this workqueue and
2399                 * process'em.
2400                 */
2401                WARN_ON_ONCE(!list_empty(&rescuer->scheduled));
2402                list_for_each_entry_safe(work, n, &pool->worklist, entry)
2403                        if (get_work_pwq(work) == pwq)
2404                                move_linked_works(work, scheduled, &n);
2405
2406                process_scheduled_works(rescuer);
2407
2408                /*
2409                 * Leave this pool.  If keep_working() is %true, notify a
2410                 * regular worker; otherwise, we end up with 0 concurrency
2411                 * and stalling the execution.
2412                 */
2413                if (keep_working(pool))
2414                        wake_up_worker(pool);
2415
2416                rescuer->pool = NULL;
2417                spin_unlock(&pool->lock);
2418                spin_lock(&wq_mayday_lock);
2419        }
2420
2421        spin_unlock_irq(&wq_mayday_lock);
2422
2423        /* rescuers should never participate in concurrency management */
2424        WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2425        schedule();
2426        goto repeat;
2427}
2428
2429struct wq_barrier {
2430        struct work_struct      work;
2431        struct completion       done;
2432};
2433
2434static void wq_barrier_func(struct work_struct *work)
2435{
2436        struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2437        complete(&barr->done);
2438}
2439
2440/**
2441 * insert_wq_barrier - insert a barrier work
2442 * @pwq: pwq to insert barrier into
2443 * @barr: wq_barrier to insert
2444 * @target: target work to attach @barr to
2445 * @worker: worker currently executing @target, NULL if @target is not executing
2446 *
2447 * @barr is linked to @target such that @barr is completed only after
2448 * @target finishes execution.  Please note that the ordering
2449 * guarantee is observed only with respect to @target and on the local
2450 * cpu.
2451 *
2452 * Currently, a queued barrier can't be canceled.  This is because
2453 * try_to_grab_pending() can't determine whether the work to be
2454 * grabbed is at the head of the queue and thus can't clear LINKED
2455 * flag of the previous work while there must be a valid next work
2456 * after a work with LINKED flag set.
2457 *
2458 * Note that when @worker is non-NULL, @target may be modified
2459 * underneath us, so we can't reliably determine pwq from @target.
2460 *
2461 * CONTEXT:
2462 * spin_lock_irq(pool->lock).
2463 */
2464static void insert_wq_barrier(struct pool_workqueue *pwq,
2465                              struct wq_barrier *barr,
2466                              struct work_struct *target, struct worker *worker)
2467{
2468        struct list_head *head;
2469        unsigned int linked = 0;
2470
2471        /*
2472         * debugobject calls are safe here even with pool->lock locked
2473         * as we know for sure that this will not trigger any of the
2474         * checks and call back into the fixup functions where we
2475         * might deadlock.
2476         */
2477        INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2478        __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2479        init_completion(&barr->done);
2480
2481        /*
2482         * If @target is currently being executed, schedule the
2483         * barrier to the worker; otherwise, put it after @target.
2484         */
2485        if (worker)
2486                head = worker->scheduled.next;
2487        else {
2488                unsigned long *bits = work_data_bits(target);
2489
2490                head = target->entry.next;
2491                /* there can already be other linked works, inherit and set */
2492                linked = *bits & WORK_STRUCT_LINKED;
2493                __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2494        }
2495
2496        debug_work_activate(&barr->work);
2497        insert_work(pwq, &barr->work, head,
2498                    work_color_to_flags(WORK_NO_COLOR) | linked);
2499}
2500
2501/**
2502 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2503 * @wq: workqueue being flushed
2504 * @flush_color: new flush color, < 0 for no-op
2505 * @work_color: new work color, < 0 for no-op
2506 *
2507 * Prepare pwqs for workqueue flushing.
2508 *
2509 * If @flush_color is non-negative, flush_color on all pwqs should be
2510 * -1.  If no pwq has in-flight commands at the specified color, all
2511 * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2512 * has in flight commands, its pwq->flush_color is set to
2513 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2514 * wakeup logic is armed and %true is returned.
2515 *
2516 * The caller should have initialized @wq->first_flusher prior to
2517 * calling this function with non-negative @flush_color.  If
2518 * @flush_color is negative, no flush color update is done and %false
2519 * is returned.
2520 *
2521 * If @work_color is non-negative, all pwqs should have the same
2522 * work_color which is previous to @work_color and all will be
2523 * advanced to @work_color.
2524 *
2525 * CONTEXT:
2526 * mutex_lock(wq->mutex).
2527 *
2528 * RETURNS:
2529 * %true if @flush_color >= 0 and there's something to flush.  %false
2530 * otherwise.
2531 */
2532static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2533                                      int flush_color, int work_color)
2534{
2535        bool wait = false;
2536        struct pool_workqueue *pwq;
2537
2538        if (flush_color >= 0) {
2539                WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2540                atomic_set(&wq->nr_pwqs_to_flush, 1);
2541        }
2542
2543        for_each_pwq(pwq, wq) {
2544                struct worker_pool *pool = pwq->pool;
2545
2546                spin_lock_irq(&pool->lock);
2547
2548                if (flush_color >= 0) {
2549                        WARN_ON_ONCE(pwq->flush_color != -1);
2550
2551                        if (pwq->nr_in_flight[flush_color]) {
2552                                pwq->flush_color = flush_color;
2553                                atomic_inc(&wq->nr_pwqs_to_flush);
2554                                wait = true;
2555                        }
2556                }
2557
2558                if (work_color >= 0) {
2559                        WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2560                        pwq->work_color = work_color;
2561                }
2562
2563                spin_unlock_irq(&pool->lock);
2564        }
2565
2566        if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2567                complete(&wq->first_flusher->done);
2568
2569        return wait;
2570}
2571
2572/**
2573 * flush_workqueue - ensure that any scheduled work has run to completion.
2574 * @wq: workqueue to flush
2575 *
2576 * This function sleeps until all work items which were queued on entry
2577 * have finished execution, but it is not livelocked by new incoming ones.
2578 */
2579void flush_workqueue(struct workqueue_struct *wq)
2580{
2581        struct wq_flusher this_flusher = {
2582                .list = LIST_HEAD_INIT(this_flusher.list),
2583                .flush_color = -1,
2584                .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2585        };
2586        int next_color;
2587
2588        lock_map_acquire(&wq->lockdep_map);
2589        lock_map_release(&wq->lockdep_map);
2590
2591        mutex_lock(&wq->mutex);
2592
2593        /*
2594         * Start-to-wait phase
2595         */
2596        next_color = work_next_color(wq->work_color);
2597
2598        if (next_color != wq->flush_color) {
2599                /*
2600                 * Color space is not full.  The current work_color
2601                 * becomes our flush_color and work_color is advanced
2602                 * by one.
2603                 */
2604                WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2605                this_flusher.flush_color = wq->work_color;
2606                wq->work_color = next_color;
2607
2608                if (!wq->first_flusher) {
2609                        /* no flush in progress, become the first flusher */
2610                        WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2611
2612                        wq->first_flusher = &this_flusher;
2613
2614                        if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2615                                                       wq->work_color)) {
2616                                /* nothing to flush, done */
2617                                wq->flush_color = next_color;
2618                                wq->first_flusher = NULL;
2619                                goto out_unlock;
2620                        }
2621                } else {
2622                        /* wait in queue */
2623                        WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2624                        list_add_tail(&this_flusher.list, &wq->flusher_queue);
2625                        flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2626                }
2627        } else {
2628                /*
2629                 * Oops, color space is full, wait on overflow queue.
2630                 * The next flush completion will assign us
2631                 * flush_color and transfer to flusher_queue.
2632                 */
2633                list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2634        }
2635
2636        mutex_unlock(&wq->mutex);
2637
2638        wait_for_completion(&this_flusher.done);
2639
2640        /*
2641         * Wake-up-and-cascade phase
2642         *
2643         * First flushers are responsible for cascading flushes and
2644         * handling overflow.  Non-first flushers can simply return.
2645         */
2646        if (wq->first_flusher != &this_flusher)
2647                return;
2648
2649        mutex_lock(&wq->mutex);
2650
2651        /* we might have raced, check again with mutex held */
2652        if (wq->first_flusher != &this_flusher)
2653                goto out_unlock;
2654
2655        wq->first_flusher = NULL;
2656
2657        WARN_ON_ONCE(!list_empty(&this_flusher.list));
2658        WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2659
2660        while (true) {
2661                struct wq_flusher *next, *tmp;
2662
2663                /* complete all the flushers sharing the current flush color */
2664                list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2665                        if (next->flush_color != wq->flush_color)
2666                                break;
2667                        list_del_init(&next->list);
2668                        complete(&next->done);
2669                }
2670
2671                WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2672                             wq->flush_color != work_next_color(wq->work_color));
2673
2674                /* this flush_color is finished, advance by one */
2675                wq->flush_color = work_next_color(wq->flush_color);
2676
2677                /* one color has been freed, handle overflow queue */
2678                if (!list_empty(&wq->flusher_overflow)) {
2679                        /*
2680                         * Assign the same color to all overflowed
2681                         * flushers, advance work_color and append to
2682                         * flusher_queue.  This is the start-to-wait
2683                         * phase for these overflowed flushers.
2684                         */
2685                        list_for_each_entry(tmp, &wq->flusher_overflow, list)
2686                                tmp->flush_color = wq->work_color;
2687
2688                        wq->work_color = work_next_color(wq->work_color);
2689
2690                        list_splice_tail_init(&wq->flusher_overflow,
2691                                              &wq->flusher_queue);
2692                        flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2693                }
2694
2695                if (list_empty(&wq->flusher_queue)) {
2696                        WARN_ON_ONCE(wq->flush_color != wq->work_color);
2697                        break;
2698                }
2699
2700                /*
2701                 * Need to flush more colors.  Make the next flusher
2702                 * the new first flusher and arm pwqs.
2703                 */
2704                WARN_ON_ONCE(wq->flush_color == wq->work_color);
2705                WARN_ON_ONCE(wq->flush_color != next->flush_color);
2706
2707                list_del_init(&next->list);
2708                wq->first_flusher = next;
2709
2710                if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2711                        break;
2712
2713                /*
2714                 * Meh... this color is already done, clear first
2715                 * flusher and repeat cascading.
2716                 */
2717                wq->first_flusher = NULL;
2718        }
2719
2720out_unlock:
2721        mutex_unlock(&wq->mutex);
2722}
2723EXPORT_SYMBOL_GPL(flush_workqueue);
2724
2725/**
2726 * drain_workqueue - drain a workqueue
2727 * @wq: workqueue to drain
2728 *
2729 * Wait until the workqueue becomes empty.  While draining is in progress,
2730 * only chain queueing is allowed.  IOW, only currently pending or running
2731 * work items on @wq can queue further work items on it.  @wq is flushed
2732 * repeatedly until it becomes empty.  The number of flushing is detemined
2733 * by the depth of chaining and should be relatively short.  Whine if it
2734 * takes too long.
2735 */
2736void drain_workqueue(struct workqueue_struct *wq)
2737{
2738        unsigned int flush_cnt = 0;
2739        struct pool_workqueue *pwq;
2740
2741        /*
2742         * __queue_work() needs to test whether there are drainers, is much
2743         * hotter than drain_workqueue() and already looks at @wq->flags.
2744         * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2745         */
2746        mutex_lock(&wq->mutex);
2747        if (!wq->nr_drainers++)
2748                wq->flags |= __WQ_DRAINING;
2749        mutex_unlock(&wq->mutex);
2750reflush:
2751        flush_workqueue(wq);
2752
2753        mutex_lock(&wq->mutex);
2754
2755        for_each_pwq(pwq, wq) {
2756                bool drained;
2757
2758                spin_lock_irq(&pwq->pool->lock);
2759                drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2760                spin_unlock_irq(&pwq->pool->lock);
2761
2762                if (drained)
2763                        continue;
2764
2765                if (++flush_cnt == 10 ||
2766                    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2767                        pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2768                                wq->name, flush_cnt);
2769
2770                mutex_unlock(&wq->mutex);
2771                goto reflush;
2772        }
2773
2774        if (!--wq->nr_drainers)
2775                wq->flags &= ~__WQ_DRAINING;
2776        mutex_unlock(&wq->mutex);
2777}
2778EXPORT_SYMBOL_GPL(drain_workqueue);
2779
2780static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
2781{
2782        struct worker *worker = NULL;
2783        struct worker_pool *pool;
2784        struct pool_workqueue *pwq;
2785
2786        might_sleep();
2787
2788        local_irq_disable();
2789        pool = get_work_pool(work);
2790        if (!pool) {
2791                local_irq_enable();
2792                return false;
2793        }
2794
2795        spin_lock(&pool->lock);
2796        /* see the comment in try_to_grab_pending() with the same code */
2797        pwq = get_work_pwq(work);
2798        if (pwq) {
2799                if (unlikely(pwq->pool != pool))
2800                        goto already_gone;
2801        } else {
2802                worker = find_worker_executing_work(pool, work);
2803                if (!worker)
2804                        goto already_gone;
2805                pwq = worker->current_pwq;
2806        }
2807
2808        insert_wq_barrier(pwq, barr, work, worker);
2809        spin_unlock_irq(&pool->lock);
2810
2811        /*
2812         * If @max_active is 1 or rescuer is in use, flushing another work
2813         * item on the same workqueue may lead to deadlock.  Make sure the
2814         * flusher is not running on the same workqueue by verifying write
2815         * access.
2816         */
2817        if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2818                lock_map_acquire(&pwq->wq->lockdep_map);
2819        else
2820                lock_map_acquire_read(&pwq->wq->lockdep_map);
2821        lock_map_release(&pwq->wq->lockdep_map);
2822
2823        return true;
2824already_gone:
2825        spin_unlock_irq(&pool->lock);
2826        return false;
2827}
2828
2829static bool __flush_work(struct work_struct *work)
2830{
2831        struct wq_barrier barr;
2832
2833        if (start_flush_work(work, &barr)) {
2834                wait_for_completion(&barr.done);
2835                destroy_work_on_stack(&barr.work);
2836                return true;
2837        } else {
2838                return false;
2839        }
2840}
2841
2842/**
2843 * flush_work - wait for a work to finish executing the last queueing instance
2844 * @work: the work to flush
2845 *
2846 * Wait until @work has finished execution.  @work is guaranteed to be idle
2847 * on return if it hasn't been requeued since flush started.
2848 *
2849 * RETURNS:
2850 * %true if flush_work() waited for the work to finish execution,
2851 * %false if it was already idle.
2852 */
2853bool flush_work(struct work_struct *work)
2854{
2855        lock_map_acquire(&work->lockdep_map);
2856        lock_map_release(&work->lockdep_map);
2857
2858        return __flush_work(work);
2859}
2860EXPORT_SYMBOL_GPL(flush_work);
2861
2862static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
2863{
2864        unsigned long flags;
2865        int ret;
2866
2867        do {
2868                ret = try_to_grab_pending(work, is_dwork, &flags);
2869                /*
2870                 * If someone else is canceling, wait for the same event it
2871                 * would be waiting for before retrying.
2872                 */
2873                if (unlikely(ret == -ENOENT))
2874                        flush_work(work);
2875        } while (unlikely(ret < 0));
2876
2877        /* tell other tasks trying to grab @work to back off */
2878        mark_work_canceling(work);
2879        local_irq_restore(flags);
2880
2881        flush_work(work);
2882        clear_work_data(work);
2883        return ret;
2884}
2885
2886/**
2887 * cancel_work_sync - cancel a work and wait for it to finish
2888 * @work: the work to cancel
2889 *
2890 * Cancel @work and wait for its execution to finish.  This function
2891 * can be used even if the work re-queues itself or migrates to
2892 * another workqueue.  On return from this function, @work is
2893 * guaranteed to be not pending or executing on any CPU.
2894 *
2895 * cancel_work_sync(&delayed_work->work) must not be used for
2896 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2897 *
2898 * The caller must ensure that the workqueue on which @work was last
2899 * queued can't be destroyed before this function returns.
2900 *
2901 * RETURNS:
2902 * %true if @work was pending, %false otherwise.
2903 */
2904bool cancel_work_sync(struct work_struct *work)
2905{
2906        return __cancel_work_timer(work, false);
2907}
2908EXPORT_SYMBOL_GPL(cancel_work_sync);
2909
2910/**
2911 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2912 * @dwork: the delayed work to flush
2913 *
2914 * Delayed timer is cancelled and the pending work is queued for
2915 * immediate execution.  Like flush_work(), this function only
2916 * considers the last queueing instance of @dwork.
2917 *
2918 * RETURNS:
2919 * %true if flush_work() waited for the work to finish execution,
2920 * %false if it was already idle.
2921 */
2922bool flush_delayed_work(struct delayed_work *dwork)
2923{
2924        local_irq_disable();
2925        if (del_timer_sync(&dwork->timer))
2926                __queue_work(dwork->cpu, dwork->wq, &dwork->work);
2927        local_irq_enable();
2928        return flush_work(&dwork->work);
2929}
2930EXPORT_SYMBOL(flush_delayed_work);
2931
2932/**
2933 * cancel_delayed_work - cancel a delayed work
2934 * @dwork: delayed_work to cancel
2935 *
2936 * Kill off a pending delayed_work.  Returns %true if @dwork was pending
2937 * and canceled; %false if wasn't pending.  Note that the work callback
2938 * function may still be running on return, unless it returns %true and the
2939 * work doesn't re-arm itself.  Explicitly flush or use
2940 * cancel_delayed_work_sync() to wait on it.
2941 *
2942 * This function is safe to call from any context including IRQ handler.
2943 */
2944bool cancel_delayed_work(struct delayed_work *dwork)
2945{
2946        unsigned long flags;
2947        int ret;
2948
2949        do {
2950                ret = try_to_grab_pending(&dwork->work, true, &flags);
2951        } while (unlikely(ret == -EAGAIN));
2952
2953        if (unlikely(ret < 0))
2954                return false;
2955
2956        set_work_pool_and_clear_pending(&dwork->work,
2957                                        get_work_pool_id(&dwork->work));
2958        local_irq_restore(flags);
2959        return ret;
2960}
2961EXPORT_SYMBOL(cancel_delayed_work);
2962
2963/**
2964 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2965 * @dwork: the delayed work cancel
2966 *
2967 * This is cancel_work_sync() for delayed works.
2968 *
2969 * RETURNS:
2970 * %true if @dwork was pending, %false otherwise.
2971 */
2972bool cancel_delayed_work_sync(struct delayed_work *dwork)
2973{
2974        return __cancel_work_timer(&dwork->work, true);
2975}
2976EXPORT_SYMBOL(cancel_delayed_work_sync);
2977
2978/**
2979 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2980 * @func: the function to call
2981 *
2982 * schedule_on_each_cpu() executes @func on each online CPU using the
2983 * system workqueue and blocks until all CPUs have completed.
2984 * schedule_on_each_cpu() is very slow.
2985 *
2986 * RETURNS:
2987 * 0 on success, -errno on failure.
2988 */
2989int schedule_on_each_cpu(work_func_t func)
2990{
2991        int cpu;
2992        struct work_struct __percpu *works;
2993
2994        works = alloc_percpu(struct work_struct);
2995        if (!works)
2996                return -ENOMEM;
2997
2998        get_online_cpus();
2999
3000        for_each_online_cpu(cpu) {
3001                struct work_struct *work = per_cpu_ptr(works, cpu);
3002
3003                INIT_WORK(work, func);
3004                schedule_work_on(cpu, work);
3005        }
3006
3007        for_each_online_cpu(cpu)
3008                flush_work(per_cpu_ptr(works, cpu));
3009
3010        put_online_cpus();
3011        free_percpu(works);
3012        return 0;
3013}
3014
3015/**
3016 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3017 *
3018 * Forces execution of the kernel-global workqueue and blocks until its
3019 * completion.
3020 *
3021 * Think twice before calling this function!  It's very easy to get into
3022 * trouble if you don't take great care.  Either of the following situations
3023 * will lead to deadlock:
3024 *
3025 *      One of the work items currently on the workqueue needs to acquire
3026 *      a lock held by your code or its caller.
3027 *
3028 *      Your code is running in the context of a work routine.
3029 *
3030 * They will be detected by lockdep when they occur, but the first might not
3031 * occur very often.  It depends on what work items are on the workqueue and
3032 * what locks they need, which you have no control over.
3033 *
3034 * In most situations flushing the entire workqueue is overkill; you merely
3035 * need to know that a particular work item isn't queued and isn't running.
3036 * In such cases you should use cancel_delayed_work_sync() or
3037 * cancel_work_sync() instead.
3038 */
3039void flush_scheduled_work(void)
3040{
3041        flush_workqueue(system_wq);
3042}
3043EXPORT_SYMBOL(flush_scheduled_work);
3044
3045/**
3046 * execute_in_process_context - reliably execute the routine with user context
3047 * @fn:         the function to execute
3048 * @ew:         guaranteed storage for the execute work structure (must
3049 *              be available when the work executes)
3050 *
3051 * Executes the function immediately if process context is available,
3052 * otherwise schedules the function for delayed execution.
3053 *
3054 * Returns:     0 - function was executed
3055 *              1 - function was scheduled for execution
3056 */
3057int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3058{
3059        if (!in_interrupt()) {
3060                fn(&ew->work);
3061                return 0;
3062        }
3063
3064        INIT_WORK(&ew->work, fn);
3065        schedule_work(&ew->work);
3066
3067        return 1;
3068}
3069EXPORT_SYMBOL_GPL(execute_in_process_context);
3070
3071#ifdef CONFIG_SYSFS
3072/*
3073 * Workqueues with WQ_SYSFS flag set is visible to userland via
3074 * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
3075 * following attributes.
3076 *
3077 *  per_cpu     RO bool : whether the workqueue is per-cpu or unbound
3078 *  max_active  RW int  : maximum number of in-flight work items
3079 *
3080 * Unbound workqueues have the following extra attributes.
3081 *
3082 *  id          RO int  : the associated pool ID
3083 *  nice        RW int  : nice value of the workers
3084 *  cpumask     RW mask : bitmask of allowed CPUs for the workers
3085 */
3086struct wq_device {
3087        struct workqueue_struct         *wq;
3088        struct device                   dev;
3089};
3090
3091static struct workqueue_struct *dev_to_wq(struct device *dev)
3092{
3093        struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3094
3095        return wq_dev->wq;
3096}
3097
3098static ssize_t wq_per_cpu_show(struct device *dev,
3099                               struct device_attribute *attr, char *buf)
3100{
3101        struct workqueue_struct *wq = dev_to_wq(dev);
3102
3103        return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
3104}
3105
3106static ssize_t wq_max_active_show(struct device *dev,
3107                                  struct device_attribute *attr, char *buf)
3108{
3109        struct workqueue_struct *wq = dev_to_wq(dev);
3110
3111        return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
3112}
3113
3114static ssize_t wq_max_active_store(struct device *dev,
3115                                   struct device_attribute *attr,
3116                                   const char *buf, size_t count)
3117{
3118        struct workqueue_struct *wq = dev_to_wq(dev);
3119        int val;
3120
3121        if (sscanf(buf, "%d", &val) != 1 || val <= 0)
3122                return -EINVAL;
3123
3124        workqueue_set_max_active(wq, val);
3125        return count;
3126}
3127
3128static struct device_attribute wq_sysfs_attrs[] = {
3129        __ATTR(per_cpu, 0444, wq_per_cpu_show, NULL),
3130        __ATTR(max_active, 0644, wq_max_active_show, wq_max_active_store),
3131        __ATTR_NULL,
3132};
3133
3134static ssize_t wq_pool_ids_show(struct device *dev,
3135                                struct device_attribute *attr, char *buf)
3136{
3137        struct workqueue_struct *wq = dev_to_wq(dev);
3138        const char *delim = "";
3139        int node, written = 0;
3140
3141        rcu_read_lock_sched();
3142        for_each_node(node) {
3143                written += scnprintf(buf + written, PAGE_SIZE - written,
3144                                     "%s%d:%d", delim, node,
3145                                     unbound_pwq_by_node(wq, node)->pool->id);
3146                delim = " ";
3147        }
3148        written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3149        rcu_read_unlock_sched();
3150
3151        return written;
3152}
3153
3154static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
3155                            char *buf)
3156{
3157        struct workqueue_struct *wq = dev_to_wq(dev);
3158        int written;
3159
3160        mutex_lock(&wq->mutex);
3161        written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
3162        mutex_unlock(&wq->mutex);
3163
3164        return written;
3165}
3166
3167/* prepare workqueue_attrs for sysfs store operations */
3168static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
3169{
3170        struct workqueue_attrs *attrs;
3171
3172        attrs = alloc_workqueue_attrs(GFP_KERNEL);
3173        if (!attrs)
3174                return NULL;
3175
3176        mutex_lock(&wq->mutex);
3177        copy_workqueue_attrs(attrs, wq->unbound_attrs);
3178        mutex_unlock(&wq->mutex);
3179        return attrs;
3180}
3181
3182static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
3183                             const char *buf, size_t count)
3184{
3185        struct workqueue_struct *wq = dev_to_wq(dev);
3186        struct workqueue_attrs *attrs;
3187        int ret;
3188
3189        attrs = wq_sysfs_prep_attrs(wq);
3190        if (!attrs)
3191                return -ENOMEM;
3192
3193        if (sscanf(buf, "%d", &attrs->nice) == 1 &&
3194            attrs->nice >= -20 && attrs->nice <= 19)
3195                ret = apply_workqueue_attrs(wq, attrs);
3196        else
3197                ret = -EINVAL;
3198
3199        free_workqueue_attrs(attrs);
3200        return ret ?: count;
3201}
3202
3203static ssize_t wq_cpumask_show(struct device *dev,
3204                               struct device_attribute *attr, char *buf)
3205{
3206        struct workqueue_struct *wq = dev_to_wq(dev);
3207        int written;
3208
3209        mutex_lock(&wq->mutex);
3210        written = cpumask_scnprintf(buf, PAGE_SIZE, wq->unbound_attrs->cpumask);
3211        mutex_unlock(&wq->mutex);
3212
3213        written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
3214        return written;
3215}
3216
3217static ssize_t wq_cpumask_store(struct device *dev,
3218                                struct device_attribute *attr,
3219                                const char *buf, size_t count)
3220{
3221        struct workqueue_struct *wq = dev_to_wq(dev);
3222        struct workqueue_attrs *attrs;
3223        int ret;
3224
3225        attrs = wq_sysfs_prep_attrs(wq);
3226        if (!attrs)
3227                return -ENOMEM;
3228
3229        ret = cpumask_parse(buf, attrs->cpumask);
3230        if (!ret)
3231                ret = apply_workqueue_attrs(wq, attrs);
3232
3233        free_workqueue_attrs(attrs);
3234        return ret ?: count;
3235}
3236
3237static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
3238                            char *buf)
3239{
3240        struct workqueue_struct *wq = dev_to_wq(dev);
3241        int written;
3242
3243        mutex_lock(&wq->mutex);
3244        written = scnprintf(buf, PAGE_SIZE, "%d\n",
3245                            !wq->unbound_attrs->no_numa);
3246        mutex_unlock(&wq->mutex);
3247
3248        return written;
3249}
3250
3251static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
3252                             const char *buf, size_t count)
3253{
3254        struct workqueue_struct *wq = dev_to_wq(dev);
3255        struct workqueue_attrs *attrs;
3256        int v, ret;
3257
3258        attrs = wq_sysfs_prep_attrs(wq);
3259        if (!attrs)
3260                return -ENOMEM;
3261
3262        ret = -EINVAL;
3263        if (sscanf(buf, "%d", &v) == 1) {
3264                attrs->no_numa = !v;
3265                ret = apply_workqueue_attrs(wq, attrs);
3266        }
3267
3268        free_workqueue_attrs(attrs);
3269        return ret ?: count;
3270}
3271
3272static struct device_attribute wq_sysfs_unbound_attrs[] = {
3273        __ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
3274        __ATTR(nice, 0644, wq_nice_show, wq_nice_store),
3275        __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
3276        __ATTR(numa, 0644, wq_numa_show, wq_numa_store),
3277        __ATTR_NULL,
3278};
3279
3280static struct bus_type wq_subsys = {
3281        .name                           = "workqueue",
3282        .dev_attrs                      = wq_sysfs_attrs,
3283};
3284
3285static int __init wq_sysfs_init(void)
3286{
3287        return subsys_virtual_register(&wq_subsys, NULL);
3288}
3289core_initcall(wq_sysfs_init);
3290
3291static void wq_device_release(struct device *dev)
3292{
3293        struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
3294
3295        kfree(wq_dev);
3296}
3297
3298/**
3299 * workqueue_sysfs_register - make a workqueue visible in sysfs
3300 * @wq: the workqueue to register
3301 *
3302 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
3303 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
3304 * which is the preferred method.
3305 *
3306 * Workqueue user should use this function directly iff it wants to apply
3307 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
3308 * apply_workqueue_attrs() may race against userland updating the
3309 * attributes.
3310 *
3311 * Returns 0 on success, -errno on failure.
3312 */
3313int workqueue_sysfs_register(struct workqueue_struct *wq)
3314{
3315        struct wq_device *wq_dev;
3316        int ret;
3317
3318        /*
3319         * Adjusting max_active or creating new pwqs by applyting
3320         * attributes breaks ordering guarantee.  Disallow exposing ordered
3321         * workqueues.
3322         */
3323        if (WARN_ON(wq->flags & __WQ_ORDERED))
3324                return -EINVAL;
3325
3326        wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
3327        if (!wq_dev)
3328                return -ENOMEM;
3329
3330        wq_dev->wq = wq;
3331        wq_dev->dev.bus = &wq_subsys;
3332        wq_dev->dev.init_name = wq->name;
3333        wq_dev->dev.release = wq_device_release;
3334
3335        /*
3336         * unbound_attrs are created separately.  Suppress uevent until
3337         * everything is ready.
3338         */
3339        dev_set_uevent_suppress(&wq_dev->dev, true);
3340
3341        ret = device_register(&wq_dev->dev);
3342        if (ret) {
3343                kfree(wq_dev);
3344                wq->wq_dev = NULL;
3345                return ret;
3346        }
3347
3348        if (wq->flags & WQ_UNBOUND) {
3349                struct device_attribute *attr;
3350
3351                for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
3352                        ret = device_create_file(&wq_dev->dev, attr);
3353                        if (ret) {
3354                                device_unregister(&wq_dev->dev);
3355                                wq->wq_dev = NULL;
3356                                return ret;
3357                        }
3358                }
3359        }
3360
3361        kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
3362        return 0;
3363}
3364
3365/**
3366 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
3367 * @wq: the workqueue to unregister
3368 *
3369 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
3370 */
3371static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
3372{
3373        struct wq_device *wq_dev = wq->wq_dev;
3374
3375        if (!wq->wq_dev)
3376                return;
3377
3378        wq->wq_dev = NULL;
3379        device_unregister(&wq_dev->dev);
3380}
3381#else   /* CONFIG_SYSFS */
3382static void workqueue_sysfs_unregister(struct workqueue_struct *wq)     { }
3383#endif  /* CONFIG_SYSFS */
3384
3385/**
3386 * free_workqueue_attrs - free a workqueue_attrs
3387 * @attrs: workqueue_attrs to free
3388 *
3389 * Undo alloc_workqueue_attrs().
3390 */
3391void free_workqueue_attrs(struct workqueue_attrs *attrs)
3392{
3393        if (attrs) {
3394                free_cpumask_var(attrs->cpumask);
3395                kfree(attrs);
3396        }
3397}
3398
3399/**
3400 * alloc_workqueue_attrs - allocate a workqueue_attrs
3401 * @gfp_mask: allocation mask to use
3402 *
3403 * Allocate a new workqueue_attrs, initialize with default settings and
3404 * return it.  Returns NULL on failure.
3405 */
3406struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3407{
3408        struct workqueue_attrs *attrs;
3409
3410        attrs = kzalloc(sizeof(*attrs), gfp_mask);
3411        if (!attrs)
3412                goto fail;
3413        if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3414                goto fail;
3415
3416        cpumask_copy(attrs->cpumask, cpu_possible_mask);
3417        return attrs;
3418fail:
3419        free_workqueue_attrs(attrs);
3420        return NULL;
3421}
3422
3423static void copy_workqueue_attrs(struct workqueue_attrs *to,
3424                                 const struct workqueue_attrs *from)
3425{
3426        to->nice = from->nice;
3427        cpumask_copy(to->cpumask, from->cpumask);
3428        /*
3429         * Unlike hash and equality test, this function doesn't ignore
3430         * ->no_numa as it is used for both pool and wq attrs.  Instead,
3431         * get_unbound_pool() explicitly clears ->no_numa after copying.
3432         */
3433        to->no_numa = from->no_numa;
3434}
3435
3436/* hash value of the content of @attr */
3437static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3438{
3439        u32 hash = 0;
3440
3441        hash = jhash_1word(attrs->nice, hash);
3442        hash = jhash(cpumask_bits(attrs->cpumask),
3443                     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3444        return hash;
3445}
3446
3447/* content equality test */
3448static bool wqattrs_equal(const struct workqueue_attrs *a,
3449                          const struct workqueue_attrs *b)
3450{
3451        if (a->nice != b->nice)
3452                return false;
3453        if (!cpumask_equal(a->cpumask, b->cpumask))
3454                return false;
3455        return true;
3456}
3457
3458/**
3459 * init_worker_pool - initialize a newly zalloc'd worker_pool
3460 * @pool: worker_pool to initialize
3461 *
3462 * Initiailize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3463 * Returns 0 on success, -errno on failure.  Even on failure, all fields
3464 * inside @pool proper are initialized and put_unbound_pool() can be called
3465 * on @pool safely to release it.
3466 */
3467static int init_worker_pool(struct worker_pool *pool)
3468{
3469        spin_lock_init(&pool->lock);
3470        pool->id = -1;
3471        pool->cpu = -1;
3472        pool->node = NUMA_NO_NODE;
3473        pool->flags |= POOL_DISASSOCIATED;
3474        INIT_LIST_HEAD(&pool->worklist);
3475        INIT_LIST_HEAD(&pool->idle_list);
3476        hash_init(pool->busy_hash);
3477
3478        init_timer_deferrable(&pool->idle_timer);
3479        pool->idle_timer.function = idle_worker_timeout;
3480        pool->idle_timer.data = (unsigned long)pool;
3481
3482        setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3483                    (unsigned long)pool);
3484
3485        mutex_init(&pool->manager_arb);
3486        mutex_init(&pool->manager_mutex);
3487        idr_init(&pool->worker_idr);
3488
3489        INIT_HLIST_NODE(&pool->hash_node);
3490        pool->refcnt = 1;
3491
3492        /* shouldn't fail above this point */
3493        pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3494        if (!pool->attrs)
3495                return -ENOMEM;
3496        return 0;
3497}
3498
3499static void rcu_free_pool(struct rcu_head *rcu)
3500{
3501        struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3502
3503        idr_destroy(&pool->worker_idr);
3504        free_workqueue_attrs(pool->attrs);
3505        kfree(pool);
3506}
3507
3508/**
3509 * put_unbound_pool - put a worker_pool
3510 * @pool: worker_pool to put
3511 *
3512 * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3513 * safe manner.  get_unbound_pool() calls this function on its failure path
3514 * and this function should be able to release pools which went through,
3515 * successfully or not, init_worker_pool().
3516 *
3517 * Should be called with wq_pool_mutex held.
3518 */
3519static void put_unbound_pool(struct worker_pool *pool)
3520{
3521        struct worker *worker;
3522
3523        lockdep_assert_held(&wq_pool_mutex);
3524
3525        if (--pool->refcnt)
3526                return;
3527
3528        /* sanity checks */
3529        if (WARN_ON(!(pool->flags & POOL_DISASSOCIATED)) ||
3530            WARN_ON(!list_empty(&pool->worklist)))
3531                return;
3532
3533        /* release id and unhash */
3534        if (pool->id >= 0)
3535                idr_remove(&worker_pool_idr, pool->id);
3536        hash_del(&pool->hash_node);
3537
3538        /*
3539         * Become the manager and destroy all workers.  Grabbing
3540         * manager_arb prevents @pool's workers from blocking on
3541         * manager_mutex.
3542         */
3543        mutex_lock(&pool->manager_arb);
3544        mutex_lock(&pool->manager_mutex);
3545        spin_lock_irq(&pool->lock);
3546
3547        while ((worker = first_worker(pool)))
3548                destroy_worker(worker);
3549        WARN_ON(pool->nr_workers || pool->nr_idle);
3550
3551        spin_unlock_irq(&pool->lock);
3552        mutex_unlock(&pool->manager_mutex);
3553        mutex_unlock(&pool->manager_arb);
3554
3555        /* shut down the timers */
3556        del_timer_sync(&pool->idle_timer);
3557        del_timer_sync(&pool->mayday_timer);
3558
3559        /* sched-RCU protected to allow dereferences from get_work_pool() */
3560        call_rcu_sched(&pool->rcu, rcu_free_pool);
3561}
3562
3563/**
3564 * get_unbound_pool - get a worker_pool with the specified attributes
3565 * @attrs: the attributes of the worker_pool to get
3566 *
3567 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3568 * reference count and return it.  If there already is a matching
3569 * worker_pool, it will be used; otherwise, this function attempts to
3570 * create a new one.  On failure, returns NULL.
3571 *
3572 * Should be called with wq_pool_mutex held.
3573 */
3574static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3575{
3576        u32 hash = wqattrs_hash(attrs);
3577        struct worker_pool *pool;
3578        int node;
3579
3580        lockdep_assert_held(&wq_pool_mutex);
3581
3582        /* do we already have a matching pool? */
3583        hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3584                if (wqattrs_equal(pool->attrs, attrs)) {
3585                        pool->refcnt++;
3586                        goto out_unlock;
3587                }
3588        }
3589
3590        /* nope, create a new one */
3591        pool = kzalloc(sizeof(*pool), GFP_KERNEL);
3592        if (!pool || init_worker_pool(pool) < 0)
3593                goto fail;
3594
3595        if (workqueue_freezing)
3596                pool->flags |= POOL_FREEZING;
3597
3598        lockdep_set_subclass(&pool->lock, 1);   /* see put_pwq() */
3599        copy_workqueue_attrs(pool->attrs, attrs);
3600
3601        /*
3602         * no_numa isn't a worker_pool attribute, always clear it.  See
3603         * 'struct workqueue_attrs' comments for detail.
3604         */
3605        pool->attrs->no_numa = false;
3606
3607        /* if cpumask is contained inside a NUMA node, we belong to that node */
3608        if (wq_numa_enabled) {
3609                for_each_node(node) {
3610                        if (cpumask_subset(pool->attrs->cpumask,
3611                                           wq_numa_possible_cpumask[node])) {
3612                                pool->node = node;
3613                                break;
3614                        }
3615                }
3616        }
3617
3618        if (worker_pool_assign_id(pool) < 0)
3619                goto fail;
3620
3621        /* create and start the initial worker */
3622        if (create_and_start_worker(pool) < 0)
3623                goto fail;
3624
3625        /* install */
3626        hash_add(unbound_pool_hash, &pool->hash_node, hash);
3627out_unlock:
3628        return pool;
3629fail:
3630        if (pool)
3631                put_unbound_pool(pool);
3632        return NULL;
3633}
3634
3635static void rcu_free_pwq(struct rcu_head *rcu)
3636{
3637        kmem_cache_free(pwq_cache,
3638                        container_of(rcu, struct pool_workqueue, rcu));
3639}
3640
3641/*
3642 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3643 * and needs to be destroyed.
3644 */
3645static void pwq_unbound_release_workfn(struct work_struct *work)
3646{
3647        struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3648                                                  unbound_release_work);
3649        struct workqueue_struct *wq = pwq->wq;
3650        struct worker_pool *pool = pwq->pool;
3651        bool is_last;
3652
3653        if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3654                return;
3655
3656        /*
3657         * Unlink @pwq.  Synchronization against wq->mutex isn't strictly
3658         * necessary on release but do it anyway.  It's easier to verify
3659         * and consistent with the linking path.
3660         */
3661        mutex_lock(&wq->mutex);
3662        list_del_rcu(&pwq->pwqs_node);
3663        is_last = list_empty(&wq->pwqs);
3664        mutex_unlock(&wq->mutex);
3665
3666        mutex_lock(&wq_pool_mutex);
3667        put_unbound_pool(pool);
3668        mutex_unlock(&wq_pool_mutex);
3669
3670        call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3671
3672        /*
3673         * If we're the last pwq going away, @wq is already dead and no one
3674         * is gonna access it anymore.  Free it.
3675         */
3676        if (is_last) {
3677                free_workqueue_attrs(wq->unbound_attrs);
3678                kfree(wq);
3679        }
3680}
3681
3682/**
3683 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3684 * @pwq: target pool_workqueue
3685 *
3686 * If @pwq isn't freezing, set @pwq->max_active to the associated
3687 * workqueue's saved_max_active and activate delayed work items
3688 * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3689 */
3690static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3691{
3692        struct workqueue_struct *wq = pwq->wq;
3693        bool freezable = wq->flags & WQ_FREEZABLE;
3694
3695        /* for @wq->saved_max_active */
3696        lockdep_assert_held(&wq->mutex);
3697
3698        /* fast exit for non-freezable wqs */
3699        if (!freezable && pwq->max_active == wq->saved_max_active)
3700                return;
3701
3702        spin_lock_irq(&pwq->pool->lock);
3703
3704        if (!freezable || !(pwq->pool->flags & POOL_FREEZING)) {
3705                pwq->max_active = wq->saved_max_active;
3706
3707                while (!list_empty(&pwq->delayed_works) &&
3708                       pwq->nr_active < pwq->max_active)
3709                        pwq_activate_first_delayed(pwq);
3710
3711                /*
3712                 * Need to kick a worker after thawed or an unbound wq's
3713                 * max_active is bumped.  It's a slow path.  Do it always.
3714                 */
3715                wake_up_worker(pwq->pool);
3716        } else {
3717                pwq->max_active = 0;
3718        }
3719
3720        spin_unlock_irq(&pwq->pool->lock);
3721}
3722
3723/* initialize newly alloced @pwq which is associated with @wq and @pool */
3724static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3725                     struct worker_pool *pool)
3726{
3727        BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3728
3729        memset(pwq, 0, sizeof(*pwq));
3730
3731        pwq->pool = pool;
3732        pwq->wq = wq;
3733        pwq->flush_color = -1;
3734        pwq->refcnt = 1;
3735        INIT_LIST_HEAD(&pwq->delayed_works);
3736        INIT_LIST_HEAD(&pwq->pwqs_node);
3737        INIT_LIST_HEAD(&pwq->mayday_node);
3738        INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3739}
3740
3741/* sync @pwq with the current state of its associated wq and link it */
3742static void link_pwq(struct pool_workqueue *pwq)
3743{
3744        struct workqueue_struct *wq = pwq->wq;
3745
3746        lockdep_assert_held(&wq->mutex);
3747
3748        /* may be called multiple times, ignore if already linked */
3749        if (!list_empty(&pwq->pwqs_node))
3750                return;
3751
3752        /*
3753         * Set the matching work_color.  This is synchronized with
3754         * wq->mutex to avoid confusing flush_workqueue().
3755         */
3756        pwq->work_color = wq->work_color;
3757
3758        /* sync max_active to the current setting */
3759        pwq_adjust_max_active(pwq);
3760
3761        /* link in @pwq */
3762        list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3763}
3764
3765/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3766static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3767                                        const struct workqueue_attrs *attrs)
3768{
3769        struct worker_pool *pool;
3770        struct pool_workqueue *pwq;
3771
3772        lockdep_assert_held(&wq_pool_mutex);
3773
3774        pool = get_unbound_pool(attrs);
3775        if (!pool)
3776                return NULL;
3777
3778        pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3779        if (!pwq) {
3780                put_unbound_pool(pool);
3781                return NULL;
3782        }
3783
3784        init_pwq(pwq, wq, pool);
3785        return pwq;
3786}
3787
3788/* undo alloc_unbound_pwq(), used only in the error path */
3789static void free_unbound_pwq(struct pool_workqueue *pwq)
3790{
3791        lockdep_assert_held(&wq_pool_mutex);
3792
3793        if (pwq) {
3794                put_unbound_pool(pwq->pool);
3795                kmem_cache_free(pwq_cache, pwq);
3796        }
3797}
3798
3799/**
3800 * wq_calc_node_mask - calculate a wq_attrs' cpumask for the specified node
3801 * @attrs: the wq_attrs of interest
3802 * @node: the target NUMA node
3803 * @cpu_going_down: if >= 0, the CPU to consider as offline
3804 * @cpumask: outarg, the resulting cpumask
3805 *
3806 * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3807 * @cpu_going_down is >= 0, that cpu is considered offline during
3808 * calculation.  The result is stored in @cpumask.  This function returns
3809 * %true if the resulting @cpumask is different from @attrs->cpumask,
3810 * %false if equal.
3811 *
3812 * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3813 * enabled and @node has online CPUs requested by @attrs, the returned
3814 * cpumask is the intersection of the possible CPUs of @node and
3815 * @attrs->cpumask.
3816 *
3817 * The caller is responsible for ensuring that the cpumask of @node stays
3818 * stable.
3819 */
3820static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3821                                 int cpu_going_down, cpumask_t *cpumask)
3822{
3823        if (!wq_numa_enabled || attrs->no_numa)
3824                goto use_dfl;
3825
3826        /* does @node have any online CPUs @attrs wants? */
3827        cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3828        if (cpu_going_down >= 0)
3829                cpumask_clear_cpu(cpu_going_down, cpumask);
3830
3831        if (cpumask_empty(cpumask))
3832                goto use_dfl;
3833
3834        /* yeap, return possible CPUs in @node that @attrs wants */
3835        cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3836        return !cpumask_equal(cpumask, attrs->cpumask);
3837
3838use_dfl:
3839        cpumask_copy(cpumask, attrs->cpumask);
3840        return false;
3841}
3842
3843/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3844static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3845                                                   int node,
3846                                                   struct pool_workqueue *pwq)
3847{
3848        struct pool_workqueue *old_pwq;
3849
3850        lockdep_assert_held(&wq->mutex);
3851
3852        /* link_pwq() can handle duplicate calls */
3853        link_pwq(pwq);
3854
3855        old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3856        rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3857        return old_pwq;
3858}
3859
3860/**
3861 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3862 * @wq: the target workqueue
3863 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3864 *
3865 * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3866 * machines, this function maps a separate pwq to each NUMA node with
3867 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3868 * NUMA node it was issued on.  Older pwqs are released as in-flight work
3869 * items finish.  Note that a work item which repeatedly requeues itself
3870 * back-to-back will stay on its current pwq.
3871 *
3872 * Performs GFP_KERNEL allocations.  Returns 0 on success and -errno on
3873 * failure.
3874 */
3875int apply_workqueue_attrs(struct workqueue_struct *wq,
3876                          const struct workqueue_attrs *attrs)
3877{
3878        struct workqueue_attrs *new_attrs, *tmp_attrs;
3879        struct pool_workqueue **pwq_tbl, *dfl_pwq;
3880        int node, ret;
3881
3882        /* only unbound workqueues can change attributes */
3883        if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3884                return -EINVAL;
3885
3886        /* creating multiple pwqs breaks ordering guarantee */
3887        if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3888                return -EINVAL;
3889
3890        pwq_tbl = kzalloc(wq_numa_tbl_len * sizeof(pwq_tbl[0]), GFP_KERNEL);
3891        new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3892        tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3893        if (!pwq_tbl || !new_attrs || !tmp_attrs)
3894                goto enomem;
3895
3896        /* make a copy of @attrs and sanitize it */
3897        copy_workqueue_attrs(new_attrs, attrs);
3898        cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3899
3900        /*
3901         * We may create multiple pwqs with differing cpumasks.  Make a
3902         * copy of @new_attrs which will be modified and used to obtain
3903         * pools.
3904         */
3905        copy_workqueue_attrs(tmp_attrs, new_attrs);
3906
3907        /*
3908         * CPUs should stay stable across pwq creations and installations.
3909         * Pin CPUs, determine the target cpumask for each node and create
3910         * pwqs accordingly.
3911         */
3912        get_online_cpus();
3913
3914        mutex_lock(&wq_pool_mutex);
3915
3916        /*
3917         * If something goes wrong during CPU up/down, we'll fall back to
3918         * the default pwq covering whole @attrs->cpumask.  Always create
3919         * it even if we don't use it immediately.
3920         */
3921        dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3922        if (!dfl_pwq)
3923                goto enomem_pwq;
3924
3925        for_each_node(node) {
3926                if (wq_calc_node_cpumask(attrs, node, -1, tmp_attrs->cpumask)) {
3927                        pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3928                        if (!pwq_tbl[node])
3929                                goto enomem_pwq;
3930                } else {
3931                        dfl_pwq->refcnt++;
3932                        pwq_tbl[node] = dfl_pwq;
3933                }
3934        }
3935
3936        mutex_unlock(&wq_pool_mutex);
3937
3938        /* all pwqs have been created successfully, let's install'em */
3939        mutex_lock(&wq->mutex);
3940
3941        copy_workqueue_attrs(wq->unbound_attrs, new_attrs);
3942
3943        /* save the previous pwq and install the new one */
3944        for_each_node(node)
3945                pwq_tbl[node] = numa_pwq_tbl_install(wq, node, pwq_tbl[node]);
3946
3947        /* @dfl_pwq might not have been used, ensure it's linked */
3948        link_pwq(dfl_pwq);
3949        swap(wq->dfl_pwq, dfl_pwq);
3950
3951        mutex_unlock(&wq->mutex);
3952
3953        /* put the old pwqs */
3954        for_each_node(node)
3955                put_pwq_unlocked(pwq_tbl[node]);
3956        put_pwq_unlocked(dfl_pwq);
3957
3958        put_online_cpus();
3959        ret = 0;
3960        /* fall through */
3961out_free:
3962        free_workqueue_attrs(tmp_attrs);
3963        free_workqueue_attrs(new_attrs);
3964        kfree(pwq_tbl);
3965        return ret;
3966
3967enomem_pwq:
3968        free_unbound_pwq(dfl_pwq);
3969        for_each_node(node)
3970                if (pwq_tbl && pwq_tbl[node] != dfl_pwq)
3971                        free_unbound_pwq(pwq_tbl[node]);
3972        mutex_unlock(&wq_pool_mutex);
3973        put_online_cpus();
3974enomem:
3975        ret = -ENOMEM;
3976        goto out_free;
3977}
3978
3979/**
3980 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3981 * @wq: the target workqueue
3982 * @cpu: the CPU coming up or going down
3983 * @online: whether @cpu is coming up or going down
3984 *
3985 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3986 * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3987 * @wq accordingly.
3988 *
3989 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3990 * falls back to @wq->dfl_pwq which may not be optimal but is always
3991 * correct.
3992 *
3993 * Note that when the last allowed CPU of a NUMA node goes offline for a
3994 * workqueue with a cpumask spanning multiple nodes, the workers which were
3995 * already executing the work items for the workqueue will lose their CPU
3996 * affinity and may execute on any CPU.  This is similar to how per-cpu
3997 * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3998 * affinity, it's the user's responsibility to flush the work item from
3999 * CPU_DOWN_PREPARE.
4000 */
4001static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4002                                   bool online)
4003{
4004        int node = cpu_to_node(cpu);
4005        int cpu_off = online ? -1 : cpu;
4006        struct pool_workqueue *old_pwq = NULL, *pwq;
4007        struct workqueue_attrs *target_attrs;
4008        cpumask_t *cpumask;
4009
4010        lockdep_assert_held(&wq_pool_mutex);
4011
4012        if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND))
4013                return;
4014
4015        /*
4016         * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4017         * Let's use a preallocated one.  The following buf is protected by
4018         * CPU hotplug exclusion.
4019         */
4020        target_attrs = wq_update_unbound_numa_attrs_buf;
4021        cpumask = target_attrs->cpumask;
4022
4023        mutex_lock(&wq->mutex);
4024        if (wq->unbound_attrs->no_numa)
4025                goto out_unlock;
4026
4027        copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4028        pwq = unbound_pwq_by_node(wq, node);
4029
4030        /*
4031         * Let's determine what needs to be done.  If the target cpumask is
4032         * different from wq's, we need to compare it to @pwq's and create
4033         * a new one if they don't match.  If the target cpumask equals
4034         * wq's, the default pwq should be used.  If @pwq is already the
4035         * default one, nothing to do; otherwise, install the default one.
4036         */
4037        if (wq_calc_node_cpumask(wq->unbound_attrs, node, cpu_off, cpumask)) {
4038                if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4039                        goto out_unlock;
4040        } else {
4041                if (pwq == wq->dfl_pwq)
4042                        goto out_unlock;
4043                else
4044                        goto use_dfl_pwq;
4045        }
4046
4047        mutex_unlock(&wq->mutex);
4048
4049        /* create a new pwq */
4050        pwq = alloc_unbound_pwq(wq, target_attrs);
4051        if (!pwq) {
4052                pr_warning("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4053                           wq->name);
4054                goto out_unlock;
4055        }
4056
4057        /*
4058         * Install the new pwq.  As this function is called only from CPU
4059         * hotplug callbacks and applying a new attrs is wrapped with
4060         * get/put_online_cpus(), @wq->unbound_attrs couldn't have changed
4061         * inbetween.
4062         */
4063        mutex_lock(&wq->mutex);
4064        old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4065        goto out_unlock;
4066
4067use_dfl_pwq:
4068        spin_lock_irq(&wq->dfl_pwq->pool->lock);
4069        get_pwq(wq->dfl_pwq);
4070        spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4071        old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4072out_unlock:
4073        mutex_unlock(&wq->mutex);
4074        put_pwq_unlocked(old_pwq);
4075}
4076
4077static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4078{
4079        bool highpri = wq->flags & WQ_HIGHPRI;
4080        int cpu;
4081
4082        if (!(wq->flags & WQ_UNBOUND)) {
4083                wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4084                if (!wq->cpu_pwqs)
4085                        return -ENOMEM;
4086
4087                for_each_possible_cpu(cpu) {
4088                        struct pool_workqueue *pwq =
4089                                per_cpu_ptr(wq->cpu_pwqs, cpu);
4090                        struct worker_pool *cpu_pools =
4091                                per_cpu(cpu_worker_pools, cpu);
4092
4093                        init_pwq(pwq, wq, &cpu_pools[highpri]);
4094
4095                        mutex_lock(&wq->mutex);
4096                        link_pwq(pwq);
4097                        mutex_unlock(&wq->mutex);
4098                }
4099                return 0;
4100        } else {
4101                return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4102        }
4103}
4104
4105static int wq_clamp_max_active(int max_active, unsigned int flags,
4106                               const char *name)
4107{
4108        int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4109
4110        if (max_active < 1 || max_active > lim)
4111                pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4112                        max_active, name, 1, lim);
4113
4114        return clamp_val(max_active, 1, lim);
4115}
4116
4117struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
4118                                               unsigned int flags,
4119                                               int max_active,
4120                                               struct lock_class_key *key,
4121                                               const char *lock_name, ...)
4122{
4123        size_t tbl_size = 0;
4124        va_list args;
4125        struct workqueue_struct *wq;
4126        struct pool_workqueue *pwq;
4127
4128        /* see the comment above the definition of WQ_POWER_EFFICIENT */
4129        if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4130                flags |= WQ_UNBOUND;
4131
4132        /* allocate wq and format name */
4133        if (flags & WQ_UNBOUND)
4134                tbl_size = wq_numa_tbl_len * sizeof(wq->numa_pwq_tbl[0]);
4135
4136        wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4137        if (!wq)
4138                return NULL;
4139
4140        if (flags & WQ_UNBOUND) {
4141                wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
4142                if (!wq->unbound_attrs)
4143                        goto err_free_wq;
4144        }
4145
4146        va_start(args, lock_name);
4147        vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4148        va_end(args);
4149
4150        max_active = max_active ?: WQ_DFL_ACTIVE;
4151        max_active = wq_clamp_max_active(max_active, flags, wq->name);
4152
4153        /* init wq */
4154        wq->flags = flags;
4155        wq->saved_max_active = max_active;
4156        mutex_init(&wq->mutex);
4157        atomic_set(&wq->nr_pwqs_to_flush, 0);
4158        INIT_LIST_HEAD(&wq->pwqs);
4159        INIT_LIST_HEAD(&wq->flusher_queue);
4160        INIT_LIST_HEAD(&wq->flusher_overflow);
4161        INIT_LIST_HEAD(&wq->maydays);
4162
4163        lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
4164        INIT_LIST_HEAD(&wq->list);
4165
4166        if (alloc_and_link_pwqs(wq) < 0)
4167                goto err_free_wq;
4168
4169        /*
4170         * Workqueues which may be used during memory reclaim should
4171         * have a rescuer to guarantee forward progress.
4172         */
4173        if (flags & WQ_MEM_RECLAIM) {
4174                struct worker *rescuer;
4175
4176                rescuer = alloc_worker();
4177                if (!rescuer)
4178                        goto err_destroy;
4179
4180                rescuer->rescue_wq = wq;
4181                rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
4182                                               wq->name);
4183                if (IS_ERR(rescuer->task)) {
4184                        kfree(rescuer);
4185                        goto err_destroy;
4186                }
4187
4188                wq->rescuer = rescuer;
4189                rescuer->task->flags |= PF_NO_SETAFFINITY;
4190                wake_up_process(rescuer->task);
4191        }
4192
4193        if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4194                goto err_destroy;
4195
4196        /*
4197         * wq_pool_mutex protects global freeze state and workqueues list.
4198         * Grab it, adjust max_active and add the new @wq to workqueues
4199         * list.
4200         */
4201        mutex_lock(&wq_pool_mutex);
4202
4203        mutex_lock(&wq->mutex);
4204        for_each_pwq(pwq, wq)
4205                pwq_adjust_max_active(pwq);
4206        mutex_unlock(&wq->mutex);
4207
4208        list_add(&wq->list, &workqueues);
4209
4210        mutex_unlock(&wq_pool_mutex);
4211
4212        return wq;
4213
4214err_free_wq:
4215        free_workqueue_attrs(wq->unbound_attrs);
4216        kfree(wq);
4217        return NULL;
4218err_destroy:
4219        destroy_workqueue(wq);
4220        return NULL;
4221}
4222EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4223
4224/**
4225 * destroy_workqueue - safely terminate a workqueue
4226 * @wq: target workqueue
4227 *
4228 * Safely destroy a workqueue. All work currently pending will be done first.
4229 */
4230void destroy_workqueue(struct workqueue_struct *wq)
4231{
4232        struct pool_workqueue *pwq;
4233        int node;
4234
4235        /* drain it before proceeding with destruction */
4236        drain_workqueue(wq);
4237
4238        /* sanity checks */
4239        mutex_lock(&wq->mutex);
4240        for_each_pwq(pwq, wq) {
4241                int i;
4242
4243                for (i = 0; i < WORK_NR_COLORS; i++) {
4244                        if (WARN_ON(pwq->nr_in_flight[i])) {
4245                                mutex_unlock(&wq->mutex);
4246                                return;
4247                        }
4248                }
4249
4250                if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4251                    WARN_ON(pwq->nr_active) ||
4252                    WARN_ON(!list_empty(&pwq->delayed_works))) {
4253                        mutex_unlock(&wq->mutex);
4254                        return;
4255                }
4256        }
4257        mutex_unlock(&wq->mutex);
4258
4259        /*
4260         * wq list is used to freeze wq, remove from list after
4261         * flushing is complete in case freeze races us.
4262         */
4263        mutex_lock(&wq_pool_mutex);
4264        list_del_init(&wq->list);
4265        mutex_unlock(&wq_pool_mutex);
4266
4267        workqueue_sysfs_unregister(wq);
4268
4269        if (wq->rescuer) {
4270                kthread_stop(wq->rescuer->task);
4271                kfree(wq->rescuer);
4272                wq->rescuer = NULL;
4273        }
4274
4275        if (!(wq->flags & WQ_UNBOUND)) {
4276                /*
4277                 * The base ref is never dropped on per-cpu pwqs.  Directly
4278                 * free the pwqs and wq.
4279                 */
4280                free_percpu(wq->cpu_pwqs);
4281                kfree(wq);
4282        } else {
4283                /*
4284                 * We're the sole accessor of @wq at this point.  Directly
4285                 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4286                 * @wq will be freed when the last pwq is released.
4287                 */
4288                for_each_node(node) {
4289                        pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4290                        RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4291                        put_pwq_unlocked(pwq);
4292                }
4293
4294                /*
4295                 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4296                 * put.  Don't access it afterwards.
4297                 */
4298                pwq = wq->dfl_pwq;
4299                wq->dfl_pwq = NULL;
4300                put_pwq_unlocked(pwq);
4301        }
4302}
4303EXPORT_SYMBOL_GPL(destroy_workqueue);
4304
4305/**
4306 * workqueue_set_max_active - adjust max_active of a workqueue
4307 * @wq: target workqueue
4308 * @max_active: new max_active value.
4309 *
4310 * Set max_active of @wq to @max_active.
4311 *
4312 * CONTEXT:
4313 * Don't call from IRQ context.
4314 */
4315void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4316{
4317        struct pool_workqueue *pwq;
4318
4319        /* disallow meddling with max_active for ordered workqueues */
4320        if (WARN_ON(wq->flags & __WQ_ORDERED))
4321                return;
4322
4323        max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4324
4325        mutex_lock(&wq->mutex);
4326
4327        wq->saved_max_active = max_active;
4328
4329        for_each_pwq(pwq, wq)
4330                pwq_adjust_max_active(pwq);
4331
4332        mutex_unlock(&wq->mutex);
4333}
4334EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4335
4336/**
4337 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4338 *
4339 * Determine whether %current is a workqueue rescuer.  Can be used from
4340 * work functions to determine whether it's being run off the rescuer task.
4341 */
4342bool current_is_workqueue_rescuer(void)
4343{
4344        struct worker *worker = current_wq_worker();
4345
4346        return worker && worker->rescue_wq;
4347}
4348
4349/**
4350 * workqueue_congested - test whether a workqueue is congested
4351 * @cpu: CPU in question
4352 * @wq: target workqueue
4353 *
4354 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4355 * no synchronization around this function and the test result is
4356 * unreliable and only useful as advisory hints or for debugging.
4357 *
4358 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4359 * Note that both per-cpu and unbound workqueues may be associated with
4360 * multiple pool_workqueues which have separate congested states.  A
4361 * workqueue being congested on one CPU doesn't mean the workqueue is also
4362 * contested on other CPUs / NUMA nodes.
4363 *
4364 * RETURNS:
4365 * %true if congested, %false otherwise.
4366 */
4367bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4368{
4369        struct pool_workqueue *pwq;
4370        bool ret;
4371
4372        rcu_read_lock_sched();
4373
4374        if (cpu == WORK_CPU_UNBOUND)
4375                cpu = smp_processor_id();
4376
4377        if (!(wq->flags & WQ_UNBOUND))
4378                pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4379        else
4380                pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4381
4382        ret = !list_empty(&pwq->delayed_works);
4383        rcu_read_unlock_sched();
4384
4385        return ret;
4386}
4387EXPORT_SYMBOL_GPL(workqueue_congested);
4388
4389/**
4390 * work_busy - test whether a work is currently pending or running
4391 * @work: the work to be tested
4392 *
4393 * Test whether @work is currently pending or running.  There is no
4394 * synchronization around this function and the test result is
4395 * unreliable and only useful as advisory hints or for debugging.
4396 *
4397 * RETURNS:
4398 * OR'd bitmask of WORK_BUSY_* bits.
4399 */
4400unsigned int work_busy(struct work_struct *work)
4401{
4402        struct worker_pool *pool;
4403        unsigned long flags;
4404        unsigned int ret = 0;
4405
4406        if (work_pending(work))
4407                ret |= WORK_BUSY_PENDING;
4408
4409        local_irq_save(flags);
4410        pool = get_work_pool(work);
4411        if (pool) {
4412                spin_lock(&pool->lock);
4413                if (find_worker_executing_work(pool, work))
4414                        ret |= WORK_BUSY_RUNNING;
4415                spin_unlock(&pool->lock);
4416        }
4417        local_irq_restore(flags);
4418
4419        return ret;
4420}
4421EXPORT_SYMBOL_GPL(work_busy);
4422
4423/**
4424 * set_worker_desc - set description for the current work item
4425 * @fmt: printf-style format string
4426 * @...: arguments for the format string
4427 *
4428 * This function can be called by a running work function to describe what
4429 * the work item is about.  If the worker task gets dumped, this
4430 * information will be printed out together to help debugging.  The
4431 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4432 */
4433void set_worker_desc(const char *fmt, ...)
4434{
4435        struct worker *worker = current_wq_worker();
4436        va_list args;
4437
4438        if (worker) {
4439                va_start(args, fmt);
4440                vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4441                va_end(args);
4442                worker->desc_valid = true;
4443        }
4444}
4445
4446/**
4447 * print_worker_info - print out worker information and description
4448 * @log_lvl: the log level to use when printing
4449 * @task: target task
4450 *
4451 * If @task is a worker and currently executing a work item, print out the
4452 * name of the workqueue being serviced and worker description set with
4453 * set_worker_desc() by the currently executing work item.
4454 *
4455 * This function can be safely called on any task as long as the
4456 * task_struct itself is accessible.  While safe, this function isn't
4457 * synchronized and may print out mixups or garbages of limited length.
4458 */
4459void print_worker_info(const char *log_lvl, struct task_struct *task)
4460{
4461        work_func_t *fn = NULL;
4462        char name[WQ_NAME_LEN] = { };
4463        char desc[WORKER_DESC_LEN] = { };
4464        struct pool_workqueue *pwq = NULL;
4465        struct workqueue_struct *wq = NULL;
4466        bool desc_valid = false;
4467        struct worker *worker;
4468
4469        if (!(task->flags & PF_WQ_WORKER))
4470                return;
4471
4472        /*
4473         * This function is called without any synchronization and @task
4474         * could be in any state.  Be careful with dereferences.
4475         */
4476        worker = probe_kthread_data(task);
4477
4478        /*
4479         * Carefully copy the associated workqueue's workfn and name.  Keep
4480         * the original last '\0' in case the original contains garbage.
4481         */
4482        probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4483        probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4484        probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4485        probe_kernel_read(name, wq->name, sizeof(name) - 1);
4486
4487        /* copy worker description */
4488        probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4489        if (desc_valid)
4490                probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4491
4492        if (fn || name[0] || desc[0]) {
4493                printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4494                if (desc[0])
4495                        pr_cont(" (%s)", desc);
4496                pr_cont("\n");
4497        }
4498}
4499
4500/*
4501 * CPU hotplug.
4502 *
4503 * There are two challenges in supporting CPU hotplug.  Firstly, there
4504 * are a lot of assumptions on strong associations among work, pwq and
4505 * pool which make migrating pending and scheduled works very
4506 * difficult to implement without impacting hot paths.  Secondly,
4507 * worker pools serve mix of short, long and very long running works making
4508 * blocked draining impractical.
4509 *
4510 * This is solved by allowing the pools to be disassociated from the CPU
4511 * running as an unbound one and allowing it to be reattached later if the
4512 * cpu comes back online.
4513 */
4514
4515static void wq_unbind_fn(struct work_struct *work)
4516{
4517        int cpu = smp_processor_id();
4518        struct worker_pool *pool;
4519        struct worker *worker;
4520        int wi;
4521
4522        for_each_cpu_worker_pool(pool, cpu) {
4523                WARN_ON_ONCE(cpu != smp_processor_id());
4524
4525                mutex_lock(&pool->manager_mutex);
4526                spin_lock_irq(&pool->lock);
4527
4528                /*
4529                 * We've blocked all manager operations.  Make all workers
4530                 * unbound and set DISASSOCIATED.  Before this, all workers
4531                 * except for the ones which are still executing works from
4532                 * before the last CPU down must be on the cpu.  After
4533                 * this, they may become diasporas.
4534                 */
4535                for_each_pool_worker(worker, wi, pool)
4536                        worker->flags |= WORKER_UNBOUND;
4537
4538                pool->flags |= POOL_DISASSOCIATED;
4539
4540                spin_unlock_irq(&pool->lock);
4541                mutex_unlock(&pool->manager_mutex);
4542
4543                /*
4544                 * Call schedule() so that we cross rq->lock and thus can
4545                 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4546                 * This is necessary as scheduler callbacks may be invoked
4547                 * from other cpus.
4548                 */
4549                schedule();
4550
4551                /*
4552                 * Sched callbacks are disabled now.  Zap nr_running.
4553                 * After this, nr_running stays zero and need_more_worker()
4554                 * and keep_working() are always true as long as the
4555                 * worklist is not empty.  This pool now behaves as an
4556                 * unbound (in terms of concurrency management) pool which
4557                 * are served by workers tied to the pool.
4558                 */
4559                atomic_set(&pool->nr_running, 0);
4560
4561                /*
4562                 * With concurrency management just turned off, a busy
4563                 * worker blocking could lead to lengthy stalls.  Kick off
4564                 * unbound chain execution of currently pending work items.
4565                 */
4566                spin_lock_irq(&pool->lock);
4567                wake_up_worker(pool);
4568                spin_unlock_irq(&pool->lock);
4569        }
4570}
4571
4572/**
4573 * rebind_workers - rebind all workers of a pool to the associated CPU
4574 * @pool: pool of interest
4575 *
4576 * @pool->cpu is coming online.  Rebind all workers to the CPU.
4577 */
4578static void rebind_workers(struct worker_pool *pool)
4579{
4580        struct worker *worker;
4581        int wi;
4582
4583        lockdep_assert_held(&pool->manager_mutex);
4584
4585        /*
4586         * Restore CPU affinity of all workers.  As all idle workers should
4587         * be on the run-queue of the associated CPU before any local
4588         * wake-ups for concurrency management happen, restore CPU affinty
4589         * of all workers first and then clear UNBOUND.  As we're called
4590         * from CPU_ONLINE, the following shouldn't fail.
4591         */
4592        for_each_pool_worker(worker, wi, pool)
4593                WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4594                                                  pool->attrs->cpumask) < 0);
4595
4596        spin_lock_irq(&pool->lock);
4597
4598        for_each_pool_worker(worker, wi, pool) {
4599                unsigned int worker_flags = worker->flags;
4600
4601                /*
4602                 * A bound idle worker should actually be on the runqueue
4603                 * of the associated CPU for local wake-ups targeting it to
4604                 * work.  Kick all idle workers so that they migrate to the
4605                 * associated CPU.  Doing this in the same loop as
4606                 * replacing UNBOUND with REBOUND is safe as no worker will
4607                 * be bound before @pool->lock is released.
4608                 */
4609                if (worker_flags & WORKER_IDLE)
4610                        wake_up_process(worker->task);
4611
4612                /*
4613                 * We want to clear UNBOUND but can't directly call
4614                 * worker_clr_flags() or adjust nr_running.  Atomically
4615                 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4616                 * @worker will clear REBOUND using worker_clr_flags() when
4617                 * it initiates the next execution cycle thus restoring
4618                 * concurrency management.  Note that when or whether
4619                 * @worker clears REBOUND doesn't affect correctness.
4620                 *
4621                 * ACCESS_ONCE() is necessary because @worker->flags may be
4622                 * tested without holding any lock in
4623                 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4624                 * fail incorrectly leading to premature concurrency
4625                 * management operations.
4626                 */
4627                WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4628                worker_flags |= WORKER_REBOUND;
4629                worker_flags &= ~WORKER_UNBOUND;
4630                ACCESS_ONCE(worker->flags) = worker_flags;
4631        }
4632
4633        spin_unlock_irq(&pool->lock);
4634}
4635
4636/**
4637 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4638 * @pool: unbound pool of interest
4639 * @cpu: the CPU which is coming up
4640 *
4641 * An unbound pool may end up with a cpumask which doesn't have any online
4642 * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4643 * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4644 * online CPU before, cpus_allowed of all its workers should be restored.
4645 */
4646static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4647{
4648        static cpumask_t cpumask;
4649        struct worker *worker;
4650        int wi;
4651
4652        lockdep_assert_held(&pool->manager_mutex);
4653
4654        /* is @cpu allowed for @pool? */
4655        if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4656                return;
4657
4658        /* is @cpu the only online CPU? */
4659        cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4660        if (cpumask_weight(&cpumask) != 1)
4661                return;
4662
4663        /* as we're called from CPU_ONLINE, the following shouldn't fail */
4664        for_each_pool_worker(worker, wi, pool)
4665                WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4666                                                  pool->attrs->cpumask) < 0);
4667}
4668
4669/*
4670 * Workqueues should be brought up before normal priority CPU notifiers.
4671 * This will be registered high priority CPU notifier.
4672 */
4673static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4674                                               unsigned long action,
4675                                               void *hcpu)
4676{
4677        int cpu = (unsigned long)hcpu;
4678        struct worker_pool *pool;
4679        struct workqueue_struct *wq;
4680        int pi;
4681
4682        switch (action & ~CPU_TASKS_FROZEN) {
4683        case CPU_UP_PREPARE:
4684                for_each_cpu_worker_pool(pool, cpu) {
4685                        if (pool->nr_workers)
4686                                continue;
4687                        if (create_and_start_worker(pool) < 0)
4688                                return NOTIFY_BAD;
4689                }
4690                break;
4691
4692        case CPU_DOWN_FAILED:
4693        case CPU_ONLINE:
4694                mutex_lock(&wq_pool_mutex);
4695
4696                for_each_pool(pool, pi) {
4697                        mutex_lock(&pool->manager_mutex);
4698
4699                        if (pool->cpu == cpu) {
4700                                spin_lock_irq(&pool->lock);
4701                                pool->flags &= ~POOL_DISASSOCIATED;
4702                                spin_unlock_irq(&pool->lock);
4703
4704                                rebind_workers(pool);
4705                        } else if (pool->cpu < 0) {
4706                                restore_unbound_workers_cpumask(pool, cpu);
4707                        }
4708
4709                        mutex_unlock(&pool->manager_mutex);
4710                }
4711
4712                /* update NUMA affinity of unbound workqueues */
4713                list_for_each_entry(wq, &workqueues, list)
4714                        wq_update_unbound_numa(wq, cpu, true);
4715
4716                mutex_unlock(&wq_pool_mutex);
4717                break;
4718        }
4719        return NOTIFY_OK;
4720}
4721
4722/*
4723 * Workqueues should be brought down after normal priority CPU notifiers.
4724 * This will be registered as low priority CPU notifier.
4725 */
4726static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4727                                                 unsigned long action,
4728                                                 void *hcpu)
4729{
4730        int cpu = (unsigned long)hcpu;
4731        struct work_struct unbind_work;
4732        struct workqueue_struct *wq;
4733
4734        switch (action & ~CPU_TASKS_FROZEN) {
4735        case CPU_DOWN_PREPARE:
4736                /* unbinding per-cpu workers should happen on the local CPU */
4737                INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4738                queue_work_on(cpu, system_highpri_wq, &unbind_work);
4739
4740                /* update NUMA affinity of unbound workqueues */
4741                mutex_lock(&wq_pool_mutex);
4742                list_for_each_entry(wq, &workqueues, list)
4743                        wq_update_unbound_numa(wq, cpu, false);
4744                mutex_unlock(&wq_pool_mutex);
4745
4746                /* wait for per-cpu unbinding to finish */
4747                flush_work(&unbind_work);
4748                break;
4749        }
4750        return NOTIFY_OK;
4751}
4752
4753#ifdef CONFIG_SMP
4754
4755struct work_for_cpu {
4756        struct work_struct work;
4757        long (*fn)(void *);
4758        void *arg;
4759        long ret;
4760};
4761
4762static void work_for_cpu_fn(struct work_struct *work)
4763{
4764        struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4765
4766        wfc->ret = wfc->fn(wfc->arg);
4767}
4768
4769/**
4770 * work_on_cpu - run a function in user context on a particular cpu
4771 * @cpu: the cpu to run on
4772 * @fn: the function to run
4773 * @arg: the function arg
4774 *
4775 * This will return the value @fn returns.
4776 * It is up to the caller to ensure that the cpu doesn't go offline.
4777 * The caller must not hold any locks which would prevent @fn from completing.
4778 */
4779long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4780{
4781        struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4782
4783        INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4784        schedule_work_on(cpu, &wfc.work);
4785
4786        /*
4787         * The work item is on-stack and can't lead to deadlock through
4788         * flushing.  Use __flush_work() to avoid spurious lockdep warnings
4789         * when work_on_cpu()s are nested.
4790         */
4791        __flush_work(&wfc.work);
4792
4793        return wfc.ret;
4794}
4795EXPORT_SYMBOL_GPL(work_on_cpu);
4796#endif /* CONFIG_SMP */
4797
4798#ifdef CONFIG_FREEZER
4799
4800/**
4801 * freeze_workqueues_begin - begin freezing workqueues
4802 *
4803 * Start freezing workqueues.  After this function returns, all freezable
4804 * workqueues will queue new works to their delayed_works list instead of
4805 * pool->worklist.
4806 *
4807 * CONTEXT:
4808 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4809 */
4810void freeze_workqueues_begin(void)
4811{
4812        struct worker_pool *pool;
4813        struct workqueue_struct *wq;
4814        struct pool_workqueue *pwq;
4815        int pi;
4816
4817        mutex_lock(&wq_pool_mutex);
4818
4819        WARN_ON_ONCE(workqueue_freezing);
4820        workqueue_freezing = true;
4821
4822        /* set FREEZING */
4823        for_each_pool(pool, pi) {
4824                spin_lock_irq(&pool->lock);
4825                WARN_ON_ONCE(pool->flags & POOL_FREEZING);
4826                pool->flags |= POOL_FREEZING;
4827                spin_unlock_irq(&pool->lock);
4828        }
4829
4830        list_for_each_entry(wq, &workqueues, list) {
4831                mutex_lock(&wq->mutex);
4832                for_each_pwq(pwq, wq)
4833                        pwq_adjust_max_active(pwq);
4834                mutex_unlock(&wq->mutex);
4835        }
4836
4837        mutex_unlock(&wq_pool_mutex);
4838}
4839
4840/**
4841 * freeze_workqueues_busy - are freezable workqueues still busy?
4842 *
4843 * Check whether freezing is complete.  This function must be called
4844 * between freeze_workqueues_begin() and thaw_workqueues().
4845 *
4846 * CONTEXT:
4847 * Grabs and releases wq_pool_mutex.
4848 *
4849 * RETURNS:
4850 * %true if some freezable workqueues are still busy.  %false if freezing
4851 * is complete.
4852 */
4853bool freeze_workqueues_busy(void)
4854{
4855        bool busy = false;
4856        struct workqueue_struct *wq;
4857        struct pool_workqueue *pwq;
4858
4859        mutex_lock(&wq_pool_mutex);
4860
4861        WARN_ON_ONCE(!workqueue_freezing);
4862
4863        list_for_each_entry(wq, &workqueues, list) {
4864                if (!(wq->flags & WQ_FREEZABLE))
4865                        continue;
4866                /*
4867                 * nr_active is monotonically decreasing.  It's safe
4868                 * to peek without lock.
4869                 */
4870                rcu_read_lock_sched();
4871                for_each_pwq(pwq, wq) {
4872                        WARN_ON_ONCE(pwq->nr_active < 0);
4873                        if (pwq->nr_active) {
4874                                busy = true;
4875                                rcu_read_unlock_sched();
4876                                goto out_unlock;
4877                        }
4878                }
4879                rcu_read_unlock_sched();
4880        }
4881out_unlock:
4882        mutex_unlock(&wq_pool_mutex);
4883        return busy;
4884}
4885
4886/**
4887 * thaw_workqueues - thaw workqueues
4888 *
4889 * Thaw workqueues.  Normal queueing is restored and all collected
4890 * frozen works are transferred to their respective pool worklists.
4891 *
4892 * CONTEXT:
4893 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4894 */
4895void thaw_workqueues(void)
4896{
4897        struct workqueue_struct *wq;
4898        struct pool_workqueue *pwq;
4899        struct worker_pool *pool;
4900        int pi;
4901
4902        mutex_lock(&wq_pool_mutex);
4903
4904        if (!workqueue_freezing)
4905                goto out_unlock;
4906
4907        /* clear FREEZING */
4908        for_each_pool(pool, pi) {
4909                spin_lock_irq(&pool->lock);
4910                WARN_ON_ONCE(!(pool->flags & POOL_FREEZING));
4911                pool->flags &= ~POOL_FREEZING;
4912                spin_unlock_irq(&pool->lock);
4913        }
4914
4915        /* restore max_active and repopulate worklist */
4916        list_for_each_entry(wq, &workqueues, list) {
4917                mutex_lock(&wq->mutex);
4918                for_each_pwq(pwq, wq)
4919                        pwq_adjust_max_active(pwq);
4920                mutex_unlock(&wq->mutex);
4921        }
4922
4923        workqueue_freezing = false;
4924out_unlock:
4925        mutex_unlock(&wq_pool_mutex);
4926}
4927#endif /* CONFIG_FREEZER */
4928
4929static void __init wq_numa_init(void)
4930{
4931        cpumask_var_t *tbl;
4932        int node, cpu;
4933
4934        /* determine NUMA pwq table len - highest node id + 1 */
4935        for_each_node(node)
4936                wq_numa_tbl_len = max(wq_numa_tbl_len, node + 1);
4937
4938        if (num_possible_nodes() <= 1)
4939                return;
4940
4941        if (wq_disable_numa) {
4942                pr_info("workqueue: NUMA affinity support disabled\n");
4943                return;
4944        }
4945
4946        wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
4947        BUG_ON(!wq_update_unbound_numa_attrs_buf);
4948
4949        /*
4950         * We want masks of possible CPUs of each node which isn't readily
4951         * available.  Build one from cpu_to_node() which should have been
4952         * fully initialized by now.
4953         */
4954        tbl = kzalloc(wq_numa_tbl_len * sizeof(tbl[0]), GFP_KERNEL);
4955        BUG_ON(!tbl);
4956
4957        for_each_node(node)
4958                BUG_ON(!alloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
4959                                node_online(node) ? node : NUMA_NO_NODE));
4960
4961        for_each_possible_cpu(cpu) {
4962                node = cpu_to_node(cpu);
4963                if (WARN_ON(node == NUMA_NO_NODE)) {
4964                        pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
4965                        /* happens iff arch is bonkers, let's just proceed */
4966                        return;
4967                }
4968                cpumask_set_cpu(cpu, tbl[node]);
4969        }
4970
4971        wq_numa_possible_cpumask = tbl;
4972        wq_numa_enabled = true;
4973}
4974
4975static int __init init_workqueues(void)
4976{
4977        int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
4978        int i, cpu;
4979
4980        /* make sure we have enough bits for OFFQ pool ID */
4981        BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_POOL_SHIFT)) <
4982                     WORK_CPU_END * NR_STD_WORKER_POOLS);
4983
4984        WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
4985
4986        pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
4987
4988        cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
4989        hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
4990
4991        wq_numa_init();
4992
4993        /* initialize CPU pools */
4994        for_each_possible_cpu(cpu) {
4995                struct worker_pool *pool;
4996
4997                i = 0;
4998                for_each_cpu_worker_pool(pool, cpu) {
4999                        BUG_ON(init_worker_pool(pool));
5000                        pool->cpu = cpu;
5001                        cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5002                        pool->attrs->nice = std_nice[i++];
5003                        pool->node = cpu_to_node(cpu);
5004
5005                        /* alloc pool ID */
5006                        mutex_lock(&wq_pool_mutex);
5007                        BUG_ON(worker_pool_assign_id(pool));
5008                        mutex_unlock(&wq_pool_mutex);
5009                }
5010        }
5011
5012        /* create the initial worker */
5013        for_each_online_cpu(cpu) {
5014                struct worker_pool *pool;
5015
5016                for_each_cpu_worker_pool(pool, cpu) {
5017                        pool->flags &= ~POOL_DISASSOCIATED;
5018                        BUG_ON(create_and_start_worker(pool) < 0);
5019                }
5020        }
5021
5022        /* create default unbound wq attrs */
5023        for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5024                struct workqueue_attrs *attrs;
5025
5026                BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5027                attrs->nice = std_nice[i];
5028                unbound_std_wq_attrs[i] = attrs;
5029        }
5030
5031        system_wq = alloc_workqueue("events", 0, 0);
5032        system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5033        system_long_wq = alloc_workqueue("events_long", 0, 0);
5034        system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5035                                            WQ_UNBOUND_MAX_ACTIVE);
5036        system_freezable_wq = alloc_workqueue("events_freezable",
5037                                              WQ_FREEZABLE, 0);
5038        system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5039                                              WQ_POWER_EFFICIENT, 0);
5040        system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5041                                              WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5042                                              0);
5043        BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5044               !system_unbound_wq || !system_freezable_wq ||
5045               !system_power_efficient_wq ||
5046               !system_freezable_power_efficient_wq);
5047        return 0;
5048}
5049early_initcall(init_workqueues);
5050
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.