Linux Audio

Check our new training course

Loading...
v4.6
   1/*
   2 * kernel/workqueue.c - generic async execution with shared worker pool
   3 *
   4 * Copyright (C) 2002		Ingo Molnar
   5 *
   6 *   Derived from the taskqueue/keventd code by:
   7 *     David Woodhouse <dwmw2@infradead.org>
   8 *     Andrew Morton
   9 *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
  10 *     Theodore Ts'o <tytso@mit.edu>
  11 *
  12 * Made to use alloc_percpu by Christoph Lameter.
  13 *
  14 * Copyright (C) 2010		SUSE Linux Products GmbH
  15 * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
  16 *
  17 * This is the generic async execution mechanism.  Work items as are
  18 * executed in process context.  The worker pool is shared and
  19 * automatically managed.  There are two worker pools for each CPU (one for
  20 * normal work items and the other for high priority ones) and some extra
  21 * pools for workqueues which are not bound to any specific CPU - the
  22 * number of these backing pools is dynamic.
  23 *
  24 * Please read Documentation/workqueue.txt for details.
  25 */
  26
  27#include <linux/export.h>
  28#include <linux/kernel.h>
  29#include <linux/sched.h>
  30#include <linux/init.h>
  31#include <linux/signal.h>
  32#include <linux/completion.h>
  33#include <linux/workqueue.h>
  34#include <linux/slab.h>
  35#include <linux/cpu.h>
  36#include <linux/notifier.h>
  37#include <linux/kthread.h>
  38#include <linux/hardirq.h>
  39#include <linux/mempolicy.h>
  40#include <linux/freezer.h>
  41#include <linux/kallsyms.h>
  42#include <linux/debug_locks.h>
  43#include <linux/lockdep.h>
  44#include <linux/idr.h>
  45#include <linux/jhash.h>
  46#include <linux/hashtable.h>
  47#include <linux/rculist.h>
  48#include <linux/nodemask.h>
  49#include <linux/moduleparam.h>
  50#include <linux/uaccess.h>
  51
  52#include "workqueue_internal.h"
  53
  54enum {
  55	/*
  56	 * worker_pool flags
  57	 *
  58	 * A bound pool is either associated or disassociated with its CPU.
  59	 * While associated (!DISASSOCIATED), all workers are bound to the
  60	 * CPU and none has %WORKER_UNBOUND set and concurrency management
  61	 * is in effect.
  62	 *
  63	 * While DISASSOCIATED, the cpu may be offline and all workers have
  64	 * %WORKER_UNBOUND set and concurrency management disabled, and may
  65	 * be executing on any CPU.  The pool behaves as an unbound one.
  66	 *
  67	 * Note that DISASSOCIATED should be flipped only while holding
  68	 * attach_mutex to avoid changing binding state while
  69	 * worker_attach_to_pool() is in progress.
  70	 */
  71	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
  72
  73	/* worker flags */
 
  74	WORKER_DIE		= 1 << 1,	/* die die die */
  75	WORKER_IDLE		= 1 << 2,	/* is idle */
  76	WORKER_PREP		= 1 << 3,	/* preparing to run works */
 
 
  77	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
  78	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
  79	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
  80
  81	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
  82				  WORKER_UNBOUND | WORKER_REBOUND,
  83
  84	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
 
 
 
 
 
  85
  86	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
  87	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
 
 
  88
  89	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
  90	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
  91
  92	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
  93						/* call for help after 10ms
  94						   (min two ticks) */
  95	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
  96	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
 
  97
  98	/*
  99	 * Rescue workers are used only on emergencies and shared by
 100	 * all cpus.  Give MIN_NICE.
 101	 */
 102	RESCUER_NICE_LEVEL	= MIN_NICE,
 103	HIGHPRI_NICE_LEVEL	= MIN_NICE,
 104
 105	WQ_NAME_LEN		= 24,
 106};
 107
 108/*
 109 * Structure fields follow one of the following exclusion rules.
 110 *
 111 * I: Modifiable by initialization/destruction paths and read-only for
 112 *    everyone else.
 113 *
 114 * P: Preemption protected.  Disabling preemption is enough and should
 115 *    only be modified and accessed from the local cpu.
 116 *
 117 * L: pool->lock protected.  Access with pool->lock held.
 118 *
 119 * X: During normal operation, modification requires pool->lock and should
 120 *    be done only from local cpu.  Either disabling preemption on local
 121 *    cpu or grabbing pool->lock is enough for read access.  If
 122 *    POOL_DISASSOCIATED is set, it's identical to L.
 123 *
 124 * A: pool->attach_mutex protected.
 
 
 
 125 *
 126 * PL: wq_pool_mutex protected.
 127 *
 128 * PR: wq_pool_mutex protected for writes.  Sched-RCU protected for reads.
 129 *
 130 * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
 131 *
 132 * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
 133 *      sched-RCU for reads.
 134 *
 135 * WQ: wq->mutex protected.
 136 *
 137 * WR: wq->mutex protected for writes.  Sched-RCU protected for reads.
 138 *
 139 * MD: wq_mayday_lock protected.
 140 */
 141
 142/* struct worker is defined in workqueue_internal.h */
 143
 144struct worker_pool {
 145	spinlock_t		lock;		/* the pool lock */
 146	int			cpu;		/* I: the associated cpu */
 147	int			node;		/* I: the associated node ID */
 148	int			id;		/* I: pool ID */
 149	unsigned int		flags;		/* X: flags */
 
 
 
 
 150
 151	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
 
 
 
 
 
 
 
 
 
 
 152
 
 
 
 
 
 
 
 153	struct list_head	worklist;	/* L: list of pending works */
 154	int			nr_workers;	/* L: total number of workers */
 
 155
 156	/* nr_idle includes the ones off idle_list for rebinding */
 157	int			nr_idle;	/* L: currently idle ones */
 158
 
 159	struct list_head	idle_list;	/* X: list of idle workers */
 160	struct timer_list	idle_timer;	/* L: worker idle timeout */
 161	struct timer_list	mayday_timer;	/* L: SOS timer for workers */
 162
 163	/* a workers is either on busy_hash or idle_list, or the manager */
 164	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
 165						/* L: hash of busy workers */
 166
 167	/* see manage_workers() for details on the two manager mutexes */
 168	struct mutex		manager_arb;	/* manager arbitration */
 169	struct worker		*manager;	/* L: purely informational */
 170	struct mutex		attach_mutex;	/* attach/detach exclusion */
 171	struct list_head	workers;	/* A: attached workers */
 172	struct completion	*detach_completion; /* all workers detached */
 173
 174	struct ida		worker_ida;	/* worker IDs for task name */
 175
 176	struct workqueue_attrs	*attrs;		/* I: worker attributes */
 177	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
 178	int			refcnt;		/* PL: refcnt for unbound pools */
 179
 180	/*
 181	 * The current concurrency level.  As it's likely to be accessed
 182	 * from other CPUs during try_to_wake_up(), put it in a separate
 183	 * cacheline.
 184	 */
 185	atomic_t		nr_running ____cacheline_aligned_in_smp;
 186
 187	/*
 188	 * Destruction of pool is sched-RCU protected to allow dereferences
 189	 * from get_work_pool().
 190	 */
 191	struct rcu_head		rcu;
 192} ____cacheline_aligned_in_smp;
 193
 194/*
 195 * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
 196 * of work_struct->data are used for flags and the remaining high bits
 197 * point to the pwq; thus, pwqs need to be aligned at two's power of the
 198 * number of flag bits.
 199 */
 200struct pool_workqueue {
 201	struct worker_pool	*pool;		/* I: the associated pool */
 202	struct workqueue_struct *wq;		/* I: the owning workqueue */
 203	int			work_color;	/* L: current color */
 204	int			flush_color;	/* L: flushing color */
 205	int			refcnt;		/* L: reference count */
 206	int			nr_in_flight[WORK_NR_COLORS];
 207						/* L: nr of in_flight works */
 208	int			nr_active;	/* L: nr of active works */
 209	int			max_active;	/* L: max active works */
 210	struct list_head	delayed_works;	/* L: delayed works */
 211	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
 212	struct list_head	mayday_node;	/* MD: node on wq->maydays */
 213
 214	/*
 215	 * Release of unbound pwq is punted to system_wq.  See put_pwq()
 216	 * and pwq_unbound_release_workfn() for details.  pool_workqueue
 217	 * itself is also sched-RCU protected so that the first pwq can be
 218	 * determined without grabbing wq->mutex.
 219	 */
 220	struct work_struct	unbound_release_work;
 221	struct rcu_head		rcu;
 222} __aligned(1 << WORK_STRUCT_FLAG_BITS);
 223
 224/*
 225 * Structure used to wait for workqueue flush.
 226 */
 227struct wq_flusher {
 228	struct list_head	list;		/* WQ: list of flushers */
 229	int			flush_color;	/* WQ: flush color waiting for */
 230	struct completion	done;		/* flush completion */
 231};
 232
 233struct wq_device;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 234
 235/*
 236 * The externally visible workqueue.  It relays the issued work items to
 237 * the appropriate worker_pool through its pool_workqueues.
 238 */
 239struct workqueue_struct {
 240	struct list_head	pwqs;		/* WR: all pwqs of this wq */
 241	struct list_head	list;		/* PR: list of all workqueues */
 242
 243	struct mutex		mutex;		/* protects this wq */
 244	int			work_color;	/* WQ: current work color */
 245	int			flush_color;	/* WQ: current flush color */
 246	atomic_t		nr_pwqs_to_flush; /* flush in progress */
 247	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
 248	struct list_head	flusher_queue;	/* WQ: flush waiters */
 249	struct list_head	flusher_overflow; /* WQ: flush overflow list */
 
 
 
 
 
 250
 251	struct list_head	maydays;	/* MD: pwqs requesting rescue */
 252	struct worker		*rescuer;	/* I: rescue worker */
 253
 254	int			nr_drainers;	/* WQ: drain in progress */
 255	int			saved_max_active; /* WQ: saved pwq max_active */
 256
 257	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
 258	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
 259
 260#ifdef CONFIG_SYSFS
 261	struct wq_device	*wq_dev;	/* I: for sysfs interface */
 262#endif
 263#ifdef CONFIG_LOCKDEP
 264	struct lockdep_map	lockdep_map;
 265#endif
 266	char			name[WQ_NAME_LEN]; /* I: workqueue name */
 267
 268	/*
 269	 * Destruction of workqueue_struct is sched-RCU protected to allow
 270	 * walking the workqueues list without grabbing wq_pool_mutex.
 271	 * This is used to dump all workqueues from sysrq.
 272	 */
 273	struct rcu_head		rcu;
 274
 275	/* hot fields used during command issue, aligned to cacheline */
 276	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
 277	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
 278	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
 279};
 280
 281static struct kmem_cache *pwq_cache;
 282
 283static cpumask_var_t *wq_numa_possible_cpumask;
 284					/* possible CPUs of each node */
 285
 286static bool wq_disable_numa;
 287module_param_named(disable_numa, wq_disable_numa, bool, 0444);
 288
 289/* see the comment above the definition of WQ_POWER_EFFICIENT */
 290static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
 291module_param_named(power_efficient, wq_power_efficient, bool, 0444);
 292
 293static bool wq_numa_enabled;		/* unbound NUMA affinity enabled */
 294
 295/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
 296static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
 297
 298static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
 299static DEFINE_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
 300
 301static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
 302static bool workqueue_freezing;		/* PL: have wqs started freezing? */
 303
 304/* PL: allowable cpus for unbound wqs and work items */
 305static cpumask_var_t wq_unbound_cpumask;
 306
 307/* CPU where unbound work was last round robin scheduled from this CPU */
 308static DEFINE_PER_CPU(int, wq_rr_cpu_last);
 309
 310/*
 311 * Local execution of unbound work items is no longer guaranteed.  The
 312 * following always forces round-robin CPU selection on unbound work items
 313 * to uncover usages which depend on it.
 314 */
 315#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
 316static bool wq_debug_force_rr_cpu = true;
 317#else
 318static bool wq_debug_force_rr_cpu = false;
 319#endif
 320module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
 321
 322/* the per-cpu worker pools */
 323static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
 324
 325static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
 326
 327/* PL: hash of all unbound pools keyed by pool->attrs */
 328static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
 329
 330/* I: attributes used when instantiating standard unbound pools on demand */
 331static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
 332
 333/* I: attributes used when instantiating ordered pools on demand */
 334static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
 335
 336struct workqueue_struct *system_wq __read_mostly;
 337EXPORT_SYMBOL(system_wq);
 338struct workqueue_struct *system_highpri_wq __read_mostly;
 339EXPORT_SYMBOL_GPL(system_highpri_wq);
 340struct workqueue_struct *system_long_wq __read_mostly;
 341EXPORT_SYMBOL_GPL(system_long_wq);
 342struct workqueue_struct *system_unbound_wq __read_mostly;
 343EXPORT_SYMBOL_GPL(system_unbound_wq);
 344struct workqueue_struct *system_freezable_wq __read_mostly;
 
 
 
 
 
 345EXPORT_SYMBOL_GPL(system_freezable_wq);
 346struct workqueue_struct *system_power_efficient_wq __read_mostly;
 347EXPORT_SYMBOL_GPL(system_power_efficient_wq);
 348struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
 349EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
 350
 351static int worker_thread(void *__worker);
 352static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
 353
 354#define CREATE_TRACE_POINTS
 355#include <trace/events/workqueue.h>
 356
 357#define assert_rcu_or_pool_mutex()					\
 358	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
 359			 !lockdep_is_held(&wq_pool_mutex),		\
 360			 "sched RCU or wq_pool_mutex should be held")
 361
 362#define assert_rcu_or_wq_mutex(wq)					\
 363	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
 364			 !lockdep_is_held(&wq->mutex),			\
 365			 "sched RCU or wq->mutex should be held")
 366
 367#define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
 368	RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held() &&			\
 369			 !lockdep_is_held(&wq->mutex) &&		\
 370			 !lockdep_is_held(&wq_pool_mutex),		\
 371			 "sched RCU, wq->mutex or wq_pool_mutex should be held")
 372
 373#define for_each_cpu_worker_pool(pool, cpu)				\
 374	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
 375	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
 376	     (pool)++)
 377
 378/**
 379 * for_each_pool - iterate through all worker_pools in the system
 380 * @pool: iteration cursor
 381 * @pi: integer used for iteration
 382 *
 383 * This must be called either with wq_pool_mutex held or sched RCU read
 384 * locked.  If the pool needs to be used beyond the locking in effect, the
 385 * caller is responsible for guaranteeing that the pool stays online.
 386 *
 387 * The if/else clause exists only for the lockdep assertion and can be
 388 * ignored.
 389 */
 390#define for_each_pool(pool, pi)						\
 391	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
 392		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
 393		else
 394
 395/**
 396 * for_each_pool_worker - iterate through all workers of a worker_pool
 397 * @worker: iteration cursor
 398 * @pool: worker_pool to iterate workers of
 399 *
 400 * This must be called with @pool->attach_mutex.
 401 *
 402 * The if/else clause exists only for the lockdep assertion and can be
 403 * ignored.
 404 */
 405#define for_each_pool_worker(worker, pool)				\
 406	list_for_each_entry((worker), &(pool)->workers, node)		\
 407		if (({ lockdep_assert_held(&pool->attach_mutex); false; })) { } \
 408		else
 409
 410/**
 411 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
 412 * @pwq: iteration cursor
 413 * @wq: the target workqueue
 414 *
 415 * This must be called either with wq->mutex held or sched RCU read locked.
 416 * If the pwq needs to be used beyond the locking in effect, the caller is
 417 * responsible for guaranteeing that the pwq stays online.
 418 *
 419 * The if/else clause exists only for the lockdep assertion and can be
 420 * ignored.
 421 */
 422#define for_each_pwq(pwq, wq)						\
 423	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)		\
 424		if (({ assert_rcu_or_wq_mutex(wq); false; })) { }	\
 425		else
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 426
 427#ifdef CONFIG_DEBUG_OBJECTS_WORK
 428
 429static struct debug_obj_descr work_debug_descr;
 430
 431static void *work_debug_hint(void *addr)
 432{
 433	return ((struct work_struct *) addr)->func;
 434}
 435
 436/*
 437 * fixup_init is called when:
 438 * - an active object is initialized
 439 */
 440static int work_fixup_init(void *addr, enum debug_obj_state state)
 441{
 442	struct work_struct *work = addr;
 443
 444	switch (state) {
 445	case ODEBUG_STATE_ACTIVE:
 446		cancel_work_sync(work);
 447		debug_object_init(work, &work_debug_descr);
 448		return 1;
 449	default:
 450		return 0;
 451	}
 452}
 453
 454/*
 455 * fixup_activate is called when:
 456 * - an active object is activated
 457 * - an unknown object is activated (might be a statically initialized object)
 458 */
 459static int work_fixup_activate(void *addr, enum debug_obj_state state)
 460{
 461	struct work_struct *work = addr;
 462
 463	switch (state) {
 464
 465	case ODEBUG_STATE_NOTAVAILABLE:
 466		/*
 467		 * This is not really a fixup. The work struct was
 468		 * statically initialized. We just make sure that it
 469		 * is tracked in the object tracker.
 470		 */
 471		if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
 472			debug_object_init(work, &work_debug_descr);
 473			debug_object_activate(work, &work_debug_descr);
 474			return 0;
 475		}
 476		WARN_ON_ONCE(1);
 477		return 0;
 478
 479	case ODEBUG_STATE_ACTIVE:
 480		WARN_ON(1);
 481
 482	default:
 483		return 0;
 484	}
 485}
 486
 487/*
 488 * fixup_free is called when:
 489 * - an active object is freed
 490 */
 491static int work_fixup_free(void *addr, enum debug_obj_state state)
 492{
 493	struct work_struct *work = addr;
 494
 495	switch (state) {
 496	case ODEBUG_STATE_ACTIVE:
 497		cancel_work_sync(work);
 498		debug_object_free(work, &work_debug_descr);
 499		return 1;
 500	default:
 501		return 0;
 502	}
 503}
 504
 505static struct debug_obj_descr work_debug_descr = {
 506	.name		= "work_struct",
 507	.debug_hint	= work_debug_hint,
 508	.fixup_init	= work_fixup_init,
 509	.fixup_activate	= work_fixup_activate,
 510	.fixup_free	= work_fixup_free,
 511};
 512
 513static inline void debug_work_activate(struct work_struct *work)
 514{
 515	debug_object_activate(work, &work_debug_descr);
 516}
 517
 518static inline void debug_work_deactivate(struct work_struct *work)
 519{
 520	debug_object_deactivate(work, &work_debug_descr);
 521}
 522
 523void __init_work(struct work_struct *work, int onstack)
 524{
 525	if (onstack)
 526		debug_object_init_on_stack(work, &work_debug_descr);
 527	else
 528		debug_object_init(work, &work_debug_descr);
 529}
 530EXPORT_SYMBOL_GPL(__init_work);
 531
 532void destroy_work_on_stack(struct work_struct *work)
 533{
 534	debug_object_free(work, &work_debug_descr);
 535}
 536EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 537
 538void destroy_delayed_work_on_stack(struct delayed_work *work)
 539{
 540	destroy_timer_on_stack(&work->timer);
 541	debug_object_free(&work->work, &work_debug_descr);
 542}
 543EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
 544
 545#else
 546static inline void debug_work_activate(struct work_struct *work) { }
 547static inline void debug_work_deactivate(struct work_struct *work) { }
 548#endif
 549
 550/**
 551 * worker_pool_assign_id - allocate ID and assing it to @pool
 552 * @pool: the pool pointer of interest
 553 *
 554 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
 555 * successfully, -errno on failure.
 
 
 
 556 */
 557static int worker_pool_assign_id(struct worker_pool *pool)
 558{
 559	int ret;
 560
 561	lockdep_assert_held(&wq_pool_mutex);
 
 
 
 
 
 
 562
 563	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
 564			GFP_KERNEL);
 565	if (ret >= 0) {
 566		pool->id = ret;
 567		return 0;
 568	}
 569	return ret;
 570}
 571
 572/**
 573 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
 574 * @wq: the target workqueue
 575 * @node: the node ID
 576 *
 577 * This must be called with any of wq_pool_mutex, wq->mutex or sched RCU
 578 * read locked.
 579 * If the pwq needs to be used beyond the locking in effect, the caller is
 580 * responsible for guaranteeing that the pwq stays online.
 581 *
 582 * Return: The unbound pool_workqueue for @node.
 583 */
 584static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
 585						  int node)
 586{
 587	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
 
 
 
 
 588
 589	/*
 590	 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
 591	 * delayed item is pending.  The plan is to keep CPU -> NODE
 592	 * mapping valid and stable across CPU on/offlines.  Once that
 593	 * happens, this workaround can be removed.
 594	 */
 595	if (unlikely(node == NUMA_NO_NODE))
 596		return wq->dfl_pwq;
 597
 598	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
 
 
 
 
 
 
 
 
 599}
 600
 601static unsigned int work_color_to_flags(int color)
 602{
 603	return color << WORK_STRUCT_COLOR_SHIFT;
 604}
 605
 606static int get_work_color(struct work_struct *work)
 607{
 608	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 609		((1 << WORK_STRUCT_COLOR_BITS) - 1);
 610}
 611
 612static int work_next_color(int color)
 613{
 614	return (color + 1) % WORK_NR_COLORS;
 615}
 616
 617/*
 618 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
 619 * contain the pointer to the queued pwq.  Once execution starts, the flag
 620 * is cleared and the high bits contain OFFQ flags and pool ID.
 621 *
 622 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
 623 * and clear_work_data() can be used to set the pwq, pool or clear
 624 * work->data.  These functions should only be called while the work is
 625 * owned - ie. while the PENDING bit is set.
 626 *
 627 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
 628 * corresponding to a work.  Pool is available once the work has been
 629 * queued anywhere after initialization until it is sync canceled.  pwq is
 630 * available only while the work item is queued.
 631 *
 632 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
 633 * canceled.  While being canceled, a work item may have its PENDING set
 634 * but stay off timer and worklist for arbitrarily long and nobody should
 635 * try to steal the PENDING bit.
 636 */
 637static inline void set_work_data(struct work_struct *work, unsigned long data,
 638				 unsigned long flags)
 639{
 640	WARN_ON_ONCE(!work_pending(work));
 641	atomic_long_set(&work->data, data | flags | work_static(work));
 642}
 643
 644static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
 
 645			 unsigned long extra_flags)
 646{
 647	set_work_data(work, (unsigned long)pwq,
 648		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
 649}
 650
 651static void set_work_pool_and_keep_pending(struct work_struct *work,
 652					   int pool_id)
 653{
 654	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
 655		      WORK_STRUCT_PENDING);
 656}
 657
 658static void set_work_pool_and_clear_pending(struct work_struct *work,
 659					    int pool_id)
 660{
 661	/*
 662	 * The following wmb is paired with the implied mb in
 663	 * test_and_set_bit(PENDING) and ensures all updates to @work made
 664	 * here are visible to and precede any updates by the next PENDING
 665	 * owner.
 666	 */
 667	smp_wmb();
 668	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
 669	/*
 670	 * The following mb guarantees that previous clear of a PENDING bit
 671	 * will not be reordered with any speculative LOADS or STORES from
 672	 * work->current_func, which is executed afterwards.  This possible
 673	 * reordering can lead to a missed execution on attempt to qeueue
 674	 * the same @work.  E.g. consider this case:
 675	 *
 676	 *   CPU#0                         CPU#1
 677	 *   ----------------------------  --------------------------------
 678	 *
 679	 * 1  STORE event_indicated
 680	 * 2  queue_work_on() {
 681	 * 3    test_and_set_bit(PENDING)
 682	 * 4 }                             set_..._and_clear_pending() {
 683	 * 5                                 set_work_data() # clear bit
 684	 * 6                                 smp_mb()
 685	 * 7                               work->current_func() {
 686	 * 8				      LOAD event_indicated
 687	 *				   }
 688	 *
 689	 * Without an explicit full barrier speculative LOAD on line 8 can
 690	 * be executed before CPU#0 does STORE on line 1.  If that happens,
 691	 * CPU#0 observes the PENDING bit is still set and new execution of
 692	 * a @work is not queued in a hope, that CPU#1 will eventually
 693	 * finish the queued @work.  Meanwhile CPU#1 does not see
 694	 * event_indicated is set, because speculative LOAD was executed
 695	 * before actual STORE.
 696	 */
 697	smp_mb();
 698}
 699
 700static void clear_work_data(struct work_struct *work)
 701{
 702	smp_wmb();	/* see set_work_pool_and_clear_pending() */
 703	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
 704}
 705
 706static struct pool_workqueue *get_work_pwq(struct work_struct *work)
 707{
 708	unsigned long data = atomic_long_read(&work->data);
 709
 710	if (data & WORK_STRUCT_PWQ)
 711		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 712	else
 713		return NULL;
 714}
 715
 716/**
 717 * get_work_pool - return the worker_pool a given work was associated with
 718 * @work: the work item of interest
 719 *
 720 * Pools are created and destroyed under wq_pool_mutex, and allows read
 721 * access under sched-RCU read lock.  As such, this function should be
 722 * called under wq_pool_mutex or with preemption disabled.
 723 *
 724 * All fields of the returned pool are accessible as long as the above
 725 * mentioned locking is in effect.  If the returned pool needs to be used
 726 * beyond the critical section, the caller is responsible for ensuring the
 727 * returned pool is and stays online.
 728 *
 729 * Return: The worker_pool @work was last associated with.  %NULL if none.
 730 */
 731static struct worker_pool *get_work_pool(struct work_struct *work)
 732{
 733	unsigned long data = atomic_long_read(&work->data);
 734	int pool_id;
 735
 736	assert_rcu_or_pool_mutex();
 
 
 737
 738	if (data & WORK_STRUCT_PWQ)
 739		return ((struct pool_workqueue *)
 740			(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
 741
 742	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
 743	if (pool_id == WORK_OFFQ_POOL_NONE)
 744		return NULL;
 745
 746	return idr_find(&worker_pool_idr, pool_id);
 747}
 748
 749/**
 750 * get_work_pool_id - return the worker pool ID a given work is associated with
 751 * @work: the work item of interest
 752 *
 753 * Return: The worker_pool ID @work was last associated with.
 754 * %WORK_OFFQ_POOL_NONE if none.
 755 */
 756static int get_work_pool_id(struct work_struct *work)
 757{
 758	unsigned long data = atomic_long_read(&work->data);
 759
 760	if (data & WORK_STRUCT_PWQ)
 761		return ((struct pool_workqueue *)
 762			(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
 763
 764	return data >> WORK_OFFQ_POOL_SHIFT;
 765}
 766
 767static void mark_work_canceling(struct work_struct *work)
 768{
 769	unsigned long pool_id = get_work_pool_id(work);
 770
 771	pool_id <<= WORK_OFFQ_POOL_SHIFT;
 772	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
 773}
 774
 775static bool work_is_canceling(struct work_struct *work)
 776{
 777	unsigned long data = atomic_long_read(&work->data);
 778
 779	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
 780}
 781
 782/*
 783 * Policy functions.  These define the policies on how the global worker
 784 * pools are managed.  Unless noted otherwise, these functions assume that
 785 * they're being called with pool->lock held.
 786 */
 787
 788static bool __need_more_worker(struct worker_pool *pool)
 789{
 790	return !atomic_read(&pool->nr_running);
 
 791}
 792
 793/*
 794 * Need to wake up a worker?  Called from anything but currently
 795 * running workers.
 796 *
 797 * Note that, because unbound workers never contribute to nr_running, this
 798 * function will always return %true for unbound pools as long as the
 799 * worklist isn't empty.
 800 */
 801static bool need_more_worker(struct worker_pool *pool)
 802{
 803	return !list_empty(&pool->worklist) && __need_more_worker(pool);
 804}
 805
 806/* Can I start working?  Called from busy but !running workers. */
 807static bool may_start_working(struct worker_pool *pool)
 808{
 809	return pool->nr_idle;
 810}
 811
 812/* Do I need to keep working?  Called from currently running workers. */
 813static bool keep_working(struct worker_pool *pool)
 814{
 815	return !list_empty(&pool->worklist) &&
 816		atomic_read(&pool->nr_running) <= 1;
 
 
 
 817}
 818
 819/* Do we need a new worker?  Called from manager. */
 820static bool need_to_create_worker(struct worker_pool *pool)
 821{
 822	return need_more_worker(pool) && !may_start_working(pool);
 
 
 
 
 
 
 823}
 824
 825/* Do we have too many workers and should some go away? */
 826static bool too_many_workers(struct worker_pool *pool)
 827{
 828	bool managing = mutex_is_locked(&pool->manager_arb);
 829	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
 830	int nr_busy = pool->nr_workers - nr_idle;
 831
 832	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 833}
 834
 835/*
 836 * Wake up functions.
 837 */
 838
 839/* Return the first idle worker.  Safe with preemption disabled */
 840static struct worker *first_idle_worker(struct worker_pool *pool)
 841{
 842	if (unlikely(list_empty(&pool->idle_list)))
 843		return NULL;
 844
 845	return list_first_entry(&pool->idle_list, struct worker, entry);
 846}
 847
 848/**
 849 * wake_up_worker - wake up an idle worker
 850 * @pool: worker pool to wake worker from
 851 *
 852 * Wake up the first idle worker of @pool.
 853 *
 854 * CONTEXT:
 855 * spin_lock_irq(pool->lock).
 856 */
 857static void wake_up_worker(struct worker_pool *pool)
 858{
 859	struct worker *worker = first_idle_worker(pool);
 860
 861	if (likely(worker))
 862		wake_up_process(worker->task);
 863}
 864
 865/**
 866 * wq_worker_waking_up - a worker is waking up
 867 * @task: task waking up
 868 * @cpu: CPU @task is waking up to
 869 *
 870 * This function is called during try_to_wake_up() when a worker is
 871 * being awoken.
 872 *
 873 * CONTEXT:
 874 * spin_lock_irq(rq->lock)
 875 */
 876void wq_worker_waking_up(struct task_struct *task, int cpu)
 877{
 878	struct worker *worker = kthread_data(task);
 879
 880	if (!(worker->flags & WORKER_NOT_RUNNING)) {
 881		WARN_ON_ONCE(worker->pool->cpu != cpu);
 882		atomic_inc(&worker->pool->nr_running);
 883	}
 884}
 885
 886/**
 887 * wq_worker_sleeping - a worker is going to sleep
 888 * @task: task going to sleep
 
 889 *
 890 * This function is called during schedule() when a busy worker is
 891 * going to sleep.  Worker on the same cpu can be woken up by
 892 * returning pointer to its task.
 893 *
 894 * CONTEXT:
 895 * spin_lock_irq(rq->lock)
 896 *
 897 * Return:
 898 * Worker task on @cpu to wake up, %NULL if none.
 899 */
 900struct task_struct *wq_worker_sleeping(struct task_struct *task)
 
 901{
 902	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 903	struct worker_pool *pool;
 
 904
 905	/*
 906	 * Rescuers, which may not have all the fields set up like normal
 907	 * workers, also reach here, let's not access anything before
 908	 * checking NOT_RUNNING.
 909	 */
 910	if (worker->flags & WORKER_NOT_RUNNING)
 911		return NULL;
 912
 913	pool = worker->pool;
 914
 915	/* this can only happen on the local cpu */
 916	if (WARN_ON_ONCE(pool->cpu != raw_smp_processor_id()))
 917		return NULL;
 918
 919	/*
 920	 * The counterpart of the following dec_and_test, implied mb,
 921	 * worklist not empty test sequence is in insert_work().
 922	 * Please read comment there.
 923	 *
 924	 * NOT_RUNNING is clear.  This means that we're bound to and
 925	 * running on the local cpu w/ rq lock held and preemption
 926	 * disabled, which in turn means that none else could be
 927	 * manipulating idle_list, so dereferencing idle_list without pool
 928	 * lock is safe.
 929	 */
 930	if (atomic_dec_and_test(&pool->nr_running) &&
 931	    !list_empty(&pool->worklist))
 932		to_wakeup = first_idle_worker(pool);
 933	return to_wakeup ? to_wakeup->task : NULL;
 934}
 935
 936/**
 937 * worker_set_flags - set worker flags and adjust nr_running accordingly
 938 * @worker: self
 939 * @flags: flags to set
 
 940 *
 941 * Set @flags in @worker->flags and adjust nr_running accordingly.
 
 
 942 *
 943 * CONTEXT:
 944 * spin_lock_irq(pool->lock)
 945 */
 946static inline void worker_set_flags(struct worker *worker, unsigned int flags)
 
 947{
 948	struct worker_pool *pool = worker->pool;
 949
 950	WARN_ON_ONCE(worker->task != current);
 951
 952	/* If transitioning into NOT_RUNNING, adjust nr_running. */
 
 
 
 
 953	if ((flags & WORKER_NOT_RUNNING) &&
 954	    !(worker->flags & WORKER_NOT_RUNNING)) {
 955		atomic_dec(&pool->nr_running);
 
 
 
 
 
 
 
 956	}
 957
 958	worker->flags |= flags;
 959}
 960
 961/**
 962 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 963 * @worker: self
 964 * @flags: flags to clear
 965 *
 966 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 967 *
 968 * CONTEXT:
 969 * spin_lock_irq(pool->lock)
 970 */
 971static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 972{
 973	struct worker_pool *pool = worker->pool;
 974	unsigned int oflags = worker->flags;
 975
 976	WARN_ON_ONCE(worker->task != current);
 977
 978	worker->flags &= ~flags;
 979
 980	/*
 981	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 982	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 983	 * of multiple flags, not a single flag.
 984	 */
 985	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 986		if (!(worker->flags & WORKER_NOT_RUNNING))
 987			atomic_inc(&pool->nr_running);
 988}
 989
 990/**
 991 * find_worker_executing_work - find worker which is executing a work
 992 * @pool: pool of interest
 993 * @work: work to find worker for
 994 *
 995 * Find a worker which is executing @work on @pool by searching
 996 * @pool->busy_hash which is keyed by the address of @work.  For a worker
 997 * to match, its current execution should match the address of @work and
 998 * its work function.  This is to avoid unwanted dependency between
 999 * unrelated work executions through a work item being recycled while still
1000 * being executed.
1001 *
1002 * This is a bit tricky.  A work item may be freed once its execution
1003 * starts and nothing prevents the freed area from being recycled for
1004 * another work item.  If the same work item address ends up being reused
1005 * before the original execution finishes, workqueue will identify the
1006 * recycled work item as currently executing and make it wait until the
1007 * current execution finishes, introducing an unwanted dependency.
1008 *
1009 * This function checks the work item address and work function to avoid
1010 * false positives.  Note that this isn't complete as one may construct a
1011 * work function which can introduce dependency onto itself through a
1012 * recycled work item.  Well, if somebody wants to shoot oneself in the
1013 * foot that badly, there's only so much we can do, and if such deadlock
1014 * actually occurs, it should be easy to locate the culprit work function.
1015 *
1016 * CONTEXT:
1017 * spin_lock_irq(pool->lock).
1018 *
1019 * Return:
1020 * Pointer to worker which is executing @work if found, %NULL
1021 * otherwise.
1022 */
1023static struct worker *find_worker_executing_work(struct worker_pool *pool,
1024						 struct work_struct *work)
1025{
1026	struct worker *worker;
 
1027
1028	hash_for_each_possible(pool->busy_hash, worker, hentry,
1029			       (unsigned long)work)
1030		if (worker->current_work == work &&
1031		    worker->current_func == work->func)
1032			return worker;
1033
1034	return NULL;
1035}
1036
1037/**
1038 * move_linked_works - move linked works to a list
1039 * @work: start of series of works to be scheduled
1040 * @head: target list to append @work to
1041 * @nextp: out parameter for nested worklist walking
1042 *
1043 * Schedule linked works starting from @work to @head.  Work series to
1044 * be scheduled starts at @work and includes any consecutive work with
1045 * WORK_STRUCT_LINKED set in its predecessor.
1046 *
1047 * If @nextp is not NULL, it's updated to point to the next work of
1048 * the last scheduled work.  This allows move_linked_works() to be
1049 * nested inside outer list_for_each_entry_safe().
1050 *
1051 * CONTEXT:
1052 * spin_lock_irq(pool->lock).
1053 */
1054static void move_linked_works(struct work_struct *work, struct list_head *head,
1055			      struct work_struct **nextp)
1056{
1057	struct work_struct *n;
1058
1059	/*
1060	 * Linked worklist will always end before the end of the list,
1061	 * use NULL for list head.
1062	 */
1063	list_for_each_entry_safe_from(work, n, NULL, entry) {
1064		list_move_tail(&work->entry, head);
1065		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1066			break;
1067	}
1068
1069	/*
1070	 * If we're already inside safe list traversal and have moved
1071	 * multiple works to the scheduled queue, the next position
1072	 * needs to be updated.
1073	 */
1074	if (nextp)
1075		*nextp = n;
1076}
1077
1078/**
1079 * get_pwq - get an extra reference on the specified pool_workqueue
1080 * @pwq: pool_workqueue to get
1081 *
1082 * Obtain an extra reference on @pwq.  The caller should guarantee that
1083 * @pwq has positive refcnt and be holding the matching pool->lock.
1084 */
1085static void get_pwq(struct pool_workqueue *pwq)
1086{
1087	lockdep_assert_held(&pwq->pool->lock);
1088	WARN_ON_ONCE(pwq->refcnt <= 0);
1089	pwq->refcnt++;
1090}
1091
1092/**
1093 * put_pwq - put a pool_workqueue reference
1094 * @pwq: pool_workqueue to put
1095 *
1096 * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1097 * destruction.  The caller should be holding the matching pool->lock.
1098 */
1099static void put_pwq(struct pool_workqueue *pwq)
1100{
1101	lockdep_assert_held(&pwq->pool->lock);
1102	if (likely(--pwq->refcnt))
1103		return;
1104	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1105		return;
1106	/*
1107	 * @pwq can't be released under pool->lock, bounce to
1108	 * pwq_unbound_release_workfn().  This never recurses on the same
1109	 * pool->lock as this path is taken only for unbound workqueues and
1110	 * the release work item is scheduled on a per-cpu workqueue.  To
1111	 * avoid lockdep warning, unbound pool->locks are given lockdep
1112	 * subclass of 1 in get_unbound_pool().
1113	 */
1114	schedule_work(&pwq->unbound_release_work);
1115}
1116
1117/**
1118 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1119 * @pwq: pool_workqueue to put (can be %NULL)
1120 *
1121 * put_pwq() with locking.  This function also allows %NULL @pwq.
 
 
1122 */
1123static void put_pwq_unlocked(struct pool_workqueue *pwq)
 
 
1124{
1125	if (pwq) {
1126		/*
1127		 * As both pwqs and pools are sched-RCU protected, the
1128		 * following lock operations are safe.
1129		 */
1130		spin_lock_irq(&pwq->pool->lock);
1131		put_pwq(pwq);
1132		spin_unlock_irq(&pwq->pool->lock);
1133	}
1134}
1135
1136static void pwq_activate_delayed_work(struct work_struct *work)
1137{
1138	struct pool_workqueue *pwq = get_work_pwq(work);
1139
1140	trace_workqueue_activate_work(work);
1141	if (list_empty(&pwq->pool->worklist))
1142		pwq->pool->watchdog_ts = jiffies;
1143	move_linked_works(work, &pwq->pool->worklist, NULL);
1144	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1145	pwq->nr_active++;
1146}
1147
1148static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1149{
1150	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1151						    struct work_struct, entry);
1152
1153	pwq_activate_delayed_work(work);
 
 
 
1154}
1155
1156/**
1157 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1158 * @pwq: pwq of interest
1159 * @color: color of work which left the queue
1160 *
1161 * A work either has completed or is removed from pending queue,
1162 * decrement nr_in_flight of its pwq and handle workqueue flushing.
 
1163 *
1164 * CONTEXT:
1165 * spin_lock_irq(pool->lock).
 
 
 
 
1166 */
1167static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
 
1168{
1169	/* uncolored work items don't participate in flushing or nr_active */
1170	if (color == WORK_NO_COLOR)
1171		goto out_put;
1172
1173	pwq->nr_in_flight[color]--;
1174
1175	pwq->nr_active--;
1176	if (!list_empty(&pwq->delayed_works)) {
1177		/* one down, submit a delayed one */
1178		if (pwq->nr_active < pwq->max_active)
1179			pwq_activate_first_delayed(pwq);
1180	}
1181
1182	/* is flush in progress and are we at the flushing tip? */
1183	if (likely(pwq->flush_color != color))
1184		goto out_put;
1185
1186	/* are there still in-flight works? */
1187	if (pwq->nr_in_flight[color])
1188		goto out_put;
1189
1190	/* this pwq is done, clear flush_color */
1191	pwq->flush_color = -1;
1192
1193	/*
1194	 * If this was the last pwq, wake up the first flusher.  It
1195	 * will handle the rest.
1196	 */
1197	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1198		complete(&pwq->wq->first_flusher->done);
1199out_put:
1200	put_pwq(pwq);
1201}
1202
1203/**
1204 * try_to_grab_pending - steal work item from worklist and disable irq
1205 * @work: work item to steal
1206 * @is_dwork: @work is a delayed_work
1207 * @flags: place to store irq state
1208 *
1209 * Try to grab PENDING bit of @work.  This function can handle @work in any
1210 * stable state - idle, on timer or on worklist.
 
 
 
 
1211 *
1212 * Return:
1213 *  1		if @work was pending and we successfully stole PENDING
1214 *  0		if @work was idle and we claimed PENDING
1215 *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1216 *  -ENOENT	if someone else is canceling @work, this state may persist
1217 *		for arbitrarily long
1218 *
1219 * Note:
1220 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1221 * interrupted while holding PENDING and @work off queue, irq must be
1222 * disabled on entry.  This, combined with delayed_work->timer being
1223 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1224 *
1225 * On successful return, >= 0, irq is disabled and the caller is
1226 * responsible for releasing it using local_irq_restore(*@flags).
1227 *
1228 * This function is safe to call from any context including IRQ handler.
 
1229 */
1230static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1231			       unsigned long *flags)
1232{
1233	struct worker_pool *pool;
1234	struct pool_workqueue *pwq;
1235
1236	local_irq_save(*flags);
 
1237
1238	/* try to steal the timer if it exists */
1239	if (is_dwork) {
1240		struct delayed_work *dwork = to_delayed_work(work);
1241
1242		/*
1243		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1244		 * guaranteed that the timer is not queued anywhere and not
1245		 * running on the local CPU.
1246		 */
1247		if (likely(del_timer(&dwork->timer)))
1248			return 1;
1249	}
1250
1251	/* try to claim PENDING the normal way */
1252	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1253		return 0;
1254
1255	/*
1256	 * The queueing is in progress, or it is already queued. Try to
1257	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1258	 */
1259	pool = get_work_pool(work);
1260	if (!pool)
1261		goto fail;
1262
1263	spin_lock(&pool->lock);
1264	/*
1265	 * work->data is guaranteed to point to pwq only while the work
1266	 * item is queued on pwq->wq, and both updating work->data to point
1267	 * to pwq on queueing and to pool on dequeueing are done under
1268	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1269	 * points to pwq which is associated with a locked pool, the work
1270	 * item is currently queued on that pool.
1271	 */
1272	pwq = get_work_pwq(work);
1273	if (pwq && pwq->pool == pool) {
1274		debug_work_deactivate(work);
1275
1276		/*
1277		 * A delayed work item cannot be grabbed directly because
1278		 * it might have linked NO_COLOR work items which, if left
1279		 * on the delayed_list, will confuse pwq->nr_active
1280		 * management later on and cause stall.  Make sure the work
1281		 * item is activated before grabbing.
1282		 */
1283		if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1284			pwq_activate_delayed_work(work);
1285
1286		list_del_init(&work->entry);
1287		pwq_dec_nr_in_flight(pwq, get_work_color(work));
1288
1289		/* work->data points to pwq iff queued, point to pool */
1290		set_work_pool_and_keep_pending(work, pool->id);
1291
1292		spin_unlock(&pool->lock);
1293		return 1;
1294	}
1295	spin_unlock(&pool->lock);
1296fail:
1297	local_irq_restore(*flags);
1298	if (work_is_canceling(work))
1299		return -ENOENT;
1300	cpu_relax();
1301	return -EAGAIN;
1302}
1303
1304/**
1305 * insert_work - insert a work into a pool
1306 * @pwq: pwq @work belongs to
1307 * @work: work to insert
1308 * @head: insertion point
1309 * @extra_flags: extra WORK_STRUCT_* flags to set
1310 *
1311 * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1312 * work_struct flags.
1313 *
1314 * CONTEXT:
1315 * spin_lock_irq(pool->lock).
1316 */
1317static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1318			struct list_head *head, unsigned int extra_flags)
 
1319{
1320	struct worker_pool *pool = pwq->pool;
1321
1322	/* we own @work, set data and link */
1323	set_work_pwq(work, pwq, extra_flags);
 
 
 
 
 
 
 
1324	list_add_tail(&work->entry, head);
1325	get_pwq(pwq);
1326
1327	/*
1328	 * Ensure either wq_worker_sleeping() sees the above
1329	 * list_add_tail() or we see zero nr_running to avoid workers lying
1330	 * around lazily while there are works to be processed.
1331	 */
1332	smp_mb();
1333
1334	if (__need_more_worker(pool))
1335		wake_up_worker(pool);
1336}
1337
1338/*
1339 * Test whether @work is being queued from another work executing on the
1340 * same workqueue.
 
1341 */
1342static bool is_chained_work(struct workqueue_struct *wq)
1343{
1344	struct worker *worker;
1345
1346	worker = current_wq_worker();
1347	/*
1348	 * Return %true iff I'm a worker execuing a work item on @wq.  If
1349	 * I'm @worker, it's safe to dereference it without locking.
1350	 */
1351	return worker && worker->current_pwq->wq == wq;
1352}
1353
1354/*
1355 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1356 * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1357 * avoid perturbing sensitive tasks.
1358 */
1359static int wq_select_unbound_cpu(int cpu)
1360{
1361	static bool printed_dbg_warning;
1362	int new_cpu;
1363
1364	if (likely(!wq_debug_force_rr_cpu)) {
1365		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1366			return cpu;
1367	} else if (!printed_dbg_warning) {
1368		pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1369		printed_dbg_warning = true;
1370	}
1371
1372	if (cpumask_empty(wq_unbound_cpumask))
1373		return cpu;
 
 
 
1374
1375	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1376	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1377	if (unlikely(new_cpu >= nr_cpu_ids)) {
1378		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1379		if (unlikely(new_cpu >= nr_cpu_ids))
1380			return cpu;
 
 
 
 
 
 
1381	}
1382	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1383
1384	return new_cpu;
1385}
1386
1387static void __queue_work(int cpu, struct workqueue_struct *wq,
1388			 struct work_struct *work)
1389{
1390	struct pool_workqueue *pwq;
1391	struct worker_pool *last_pool;
1392	struct list_head *worklist;
1393	unsigned int work_flags;
1394	unsigned int req_cpu = cpu;
1395
1396	/*
1397	 * While a work item is PENDING && off queue, a task trying to
1398	 * steal the PENDING will busy-loop waiting for it to either get
1399	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1400	 * happen with IRQ disabled.
1401	 */
1402	WARN_ON_ONCE(!irqs_disabled());
1403
1404	debug_work_activate(work);
1405
1406	/* if draining, only works from the same workqueue are allowed */
1407	if (unlikely(wq->flags & __WQ_DRAINING) &&
1408	    WARN_ON_ONCE(!is_chained_work(wq)))
1409		return;
1410retry:
1411	if (req_cpu == WORK_CPU_UNBOUND)
1412		cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1413
1414	/* pwq which will be used unless @work is executing elsewhere */
1415	if (!(wq->flags & WQ_UNBOUND))
1416		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1417	else
1418		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1419
1420	/*
1421	 * If @work was previously on a different pool, it might still be
1422	 * running there, in which case the work needs to be queued on that
1423	 * pool to guarantee non-reentrancy.
1424	 */
1425	last_pool = get_work_pool(work);
1426	if (last_pool && last_pool != pwq->pool) {
1427		struct worker *worker;
1428
1429		spin_lock(&last_pool->lock);
 
 
 
 
 
 
 
 
 
1430
1431		worker = find_worker_executing_work(last_pool, work);
1432
1433		if (worker && worker->current_pwq->wq == wq) {
1434			pwq = worker->current_pwq;
1435		} else {
1436			/* meh... not running there, queue here */
1437			spin_unlock(&last_pool->lock);
1438			spin_lock(&pwq->pool->lock);
1439		}
1440	} else {
1441		spin_lock(&pwq->pool->lock);
1442	}
1443
1444	/*
1445	 * pwq is determined and locked.  For unbound pools, we could have
1446	 * raced with pwq release and it could already be dead.  If its
1447	 * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1448	 * without another pwq replacing it in the numa_pwq_tbl or while
1449	 * work items are executing on it, so the retrying is guaranteed to
1450	 * make forward-progress.
1451	 */
1452	if (unlikely(!pwq->refcnt)) {
1453		if (wq->flags & WQ_UNBOUND) {
1454			spin_unlock(&pwq->pool->lock);
1455			cpu_relax();
1456			goto retry;
1457		}
1458		/* oops */
1459		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1460			  wq->name, cpu);
1461	}
1462
1463	/* pwq determined, queue */
1464	trace_workqueue_queue_work(req_cpu, pwq, work);
 
1465
1466	if (WARN_ON(!list_empty(&work->entry))) {
1467		spin_unlock(&pwq->pool->lock);
1468		return;
1469	}
1470
1471	pwq->nr_in_flight[pwq->work_color]++;
1472	work_flags = work_color_to_flags(pwq->work_color);
1473
1474	if (likely(pwq->nr_active < pwq->max_active)) {
1475		trace_workqueue_activate_work(work);
1476		pwq->nr_active++;
1477		worklist = &pwq->pool->worklist;
1478		if (list_empty(worklist))
1479			pwq->pool->watchdog_ts = jiffies;
1480	} else {
1481		work_flags |= WORK_STRUCT_DELAYED;
1482		worklist = &pwq->delayed_works;
1483	}
1484
1485	insert_work(pwq, work, worklist, work_flags);
1486
1487	spin_unlock(&pwq->pool->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1488}
 
1489
1490/**
1491 * queue_work_on - queue work on specific cpu
1492 * @cpu: CPU number to execute work on
1493 * @wq: workqueue to use
1494 * @work: work to queue
1495 *
 
 
1496 * We queue the work to a specific CPU, the caller must ensure it
1497 * can't go away.
1498 *
1499 * Return: %false if @work was already on a queue, %true otherwise.
1500 */
1501bool queue_work_on(int cpu, struct workqueue_struct *wq,
1502		   struct work_struct *work)
1503{
1504	bool ret = false;
1505	unsigned long flags;
1506
1507	local_irq_save(flags);
1508
1509	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1510		__queue_work(cpu, wq, work);
1511		ret = true;
1512	}
1513
1514	local_irq_restore(flags);
1515	return ret;
1516}
1517EXPORT_SYMBOL(queue_work_on);
1518
1519void delayed_work_timer_fn(unsigned long __data)
1520{
1521	struct delayed_work *dwork = (struct delayed_work *)__data;
 
1522
1523	/* should have been called from irqsafe timer with irq already off */
1524	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1525}
1526EXPORT_SYMBOL(delayed_work_timer_fn);
1527
1528static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1529				struct delayed_work *dwork, unsigned long delay)
 
 
 
 
 
 
 
 
1530{
1531	struct timer_list *timer = &dwork->timer;
1532	struct work_struct *work = &dwork->work;
1533
1534	WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1535		     timer->data != (unsigned long)dwork);
1536	WARN_ON_ONCE(timer_pending(timer));
1537	WARN_ON_ONCE(!list_empty(&work->entry));
1538
1539	/*
1540	 * If @delay is 0, queue @dwork->work immediately.  This is for
1541	 * both optimization and correctness.  The earliest @timer can
1542	 * expire is on the closest next tick and delayed_work users depend
1543	 * on that there's no such delay when @delay is 0.
1544	 */
1545	if (!delay) {
1546		__queue_work(cpu, wq, &dwork->work);
1547		return;
1548	}
1549
1550	timer_stats_timer_set_start_info(&dwork->timer);
1551
1552	dwork->wq = wq;
1553	dwork->cpu = cpu;
1554	timer->expires = jiffies + delay;
1555
1556	if (unlikely(cpu != WORK_CPU_UNBOUND))
1557		add_timer_on(timer, cpu);
1558	else
1559		add_timer(timer);
1560}
 
1561
1562/**
1563 * queue_delayed_work_on - queue work on specific CPU after delay
1564 * @cpu: CPU number to execute work on
1565 * @wq: workqueue to use
1566 * @dwork: work to queue
1567 * @delay: number of jiffies to wait before queueing
1568 *
1569 * Return: %false if @work was already on a queue, %true otherwise.  If
1570 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1571 * execution.
1572 */
1573bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1574			   struct delayed_work *dwork, unsigned long delay)
1575{
 
 
1576	struct work_struct *work = &dwork->work;
1577	bool ret = false;
1578	unsigned long flags;
1579
1580	/* read the comment in __queue_work() */
1581	local_irq_save(flags);
1582
1583	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1584		__queue_delayed_work(cpu, wq, dwork, delay);
1585		ret = true;
1586	}
1587
1588	local_irq_restore(flags);
1589	return ret;
1590}
1591EXPORT_SYMBOL(queue_delayed_work_on);
1592
1593/**
1594 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1595 * @cpu: CPU number to execute work on
1596 * @wq: workqueue to use
1597 * @dwork: work to queue
1598 * @delay: number of jiffies to wait before queueing
1599 *
1600 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1601 * modify @dwork's timer so that it expires after @delay.  If @delay is
1602 * zero, @work is guaranteed to be scheduled immediately regardless of its
1603 * current state.
1604 *
1605 * Return: %false if @dwork was idle and queued, %true if @dwork was
1606 * pending and its timer was modified.
1607 *
1608 * This function is safe to call from any context including IRQ handler.
1609 * See try_to_grab_pending() for details.
1610 */
1611bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1612			 struct delayed_work *dwork, unsigned long delay)
1613{
1614	unsigned long flags;
1615	int ret;
1616
1617	do {
1618		ret = try_to_grab_pending(&dwork->work, true, &flags);
1619	} while (unlikely(ret == -EAGAIN));
 
 
 
 
1620
1621	if (likely(ret >= 0)) {
1622		__queue_delayed_work(cpu, wq, dwork, delay);
1623		local_irq_restore(flags);
1624	}
 
 
1625
1626	/* -ENOENT from try_to_grab_pending() becomes %true */
 
 
 
 
 
 
 
 
 
 
 
1627	return ret;
1628}
1629EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1630
1631/**
1632 * worker_enter_idle - enter idle state
1633 * @worker: worker which is entering idle state
1634 *
1635 * @worker is entering idle state.  Update stats and idle timer if
1636 * necessary.
1637 *
1638 * LOCKING:
1639 * spin_lock_irq(pool->lock).
1640 */
1641static void worker_enter_idle(struct worker *worker)
1642{
1643	struct worker_pool *pool = worker->pool;
1644
1645	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1646	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1647			 (worker->hentry.next || worker->hentry.pprev)))
1648		return;
1649
1650	/* can't use worker_set_flags(), also called from create_worker() */
1651	worker->flags |= WORKER_IDLE;
1652	pool->nr_idle++;
1653	worker->last_active = jiffies;
1654
1655	/* idle_list is LIFO */
1656	list_add(&worker->entry, &pool->idle_list);
1657
1658	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1659		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1660
1661	/*
1662	 * Sanity check nr_running.  Because wq_unbind_fn() releases
1663	 * pool->lock between setting %WORKER_UNBOUND and zapping
1664	 * nr_running, the warning may trigger spuriously.  Check iff
1665	 * unbind is not in progress.
1666	 */
1667	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1668		     pool->nr_workers == pool->nr_idle &&
1669		     atomic_read(&pool->nr_running));
 
 
 
 
 
 
1670}
1671
1672/**
1673 * worker_leave_idle - leave idle state
1674 * @worker: worker which is leaving idle state
1675 *
1676 * @worker is leaving idle state.  Update stats.
1677 *
1678 * LOCKING:
1679 * spin_lock_irq(pool->lock).
1680 */
1681static void worker_leave_idle(struct worker *worker)
1682{
1683	struct worker_pool *pool = worker->pool;
1684
1685	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1686		return;
1687	worker_clr_flags(worker, WORKER_IDLE);
1688	pool->nr_idle--;
1689	list_del_init(&worker->entry);
1690}
1691
1692static struct worker *alloc_worker(int node)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1693{
1694	struct worker *worker;
 
1695
1696	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1697	if (worker) {
1698		INIT_LIST_HEAD(&worker->entry);
1699		INIT_LIST_HEAD(&worker->scheduled);
1700		INIT_LIST_HEAD(&worker->node);
1701		/* on creation a worker is in !idle && prep state */
1702		worker->flags = WORKER_PREP;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1703	}
1704	return worker;
1705}
1706
1707/**
1708 * worker_attach_to_pool() - attach a worker to a pool
1709 * @worker: worker to be attached
1710 * @pool: the target pool
1711 *
1712 * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
1713 * cpu-binding of @worker are kept coordinated with the pool across
1714 * cpu-[un]hotplugs.
1715 */
1716static void worker_attach_to_pool(struct worker *worker,
1717				   struct worker_pool *pool)
1718{
1719	mutex_lock(&pool->attach_mutex);
 
1720
1721	/*
1722	 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1723	 * online CPUs.  It'll be re-applied when any of the CPUs come up.
1724	 */
1725	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1726
1727	/*
1728	 * The pool->attach_mutex ensures %POOL_DISASSOCIATED remains
1729	 * stable across this function.  See the comments above the
1730	 * flag definition for details.
1731	 */
1732	if (pool->flags & POOL_DISASSOCIATED)
1733		worker->flags |= WORKER_UNBOUND;
1734
1735	list_add_tail(&worker->node, &pool->workers);
1736
1737	mutex_unlock(&pool->attach_mutex);
1738}
1739
1740/**
1741 * worker_detach_from_pool() - detach a worker from its pool
1742 * @worker: worker which is attached to its pool
1743 * @pool: the pool @worker is attached to
1744 *
1745 * Undo the attaching which had been done in worker_attach_to_pool().  The
1746 * caller worker shouldn't access to the pool after detached except it has
1747 * other reference to the pool.
1748 */
1749static void worker_detach_from_pool(struct worker *worker,
1750				    struct worker_pool *pool)
1751{
1752	struct completion *detach_completion = NULL;
1753
1754	mutex_lock(&pool->attach_mutex);
1755	list_del(&worker->node);
1756	if (list_empty(&pool->workers))
1757		detach_completion = pool->detach_completion;
1758	mutex_unlock(&pool->attach_mutex);
1759
1760	/* clear leftover flags without pool->lock after it is detached */
1761	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1762
1763	if (detach_completion)
1764		complete(detach_completion);
 
 
 
 
 
 
 
1765}
1766
1767/**
1768 * create_worker - create a new workqueue worker
1769 * @pool: pool the new worker will belong to
 
1770 *
1771 * Create and start a new worker which is attached to @pool.
 
 
1772 *
1773 * CONTEXT:
1774 * Might sleep.  Does GFP_KERNEL allocations.
1775 *
1776 * Return:
1777 * Pointer to the newly created worker.
1778 */
1779static struct worker *create_worker(struct worker_pool *pool)
1780{
 
1781	struct worker *worker = NULL;
1782	int id = -1;
1783	char id_buf[16];
1784
1785	/* ID is needed to determine kthread name */
1786	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1787	if (id < 0)
1788		goto fail;
 
 
 
 
1789
1790	worker = alloc_worker(pool->node);
1791	if (!worker)
1792		goto fail;
1793
1794	worker->pool = pool;
1795	worker->id = id;
1796
1797	if (pool->cpu >= 0)
1798		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1799			 pool->attrs->nice < 0  ? "H" : "");
 
 
1800	else
1801		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1802
1803	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1804					      "kworker/%s", id_buf);
1805	if (IS_ERR(worker->task))
1806		goto fail;
1807
1808	set_user_nice(worker->task, pool->attrs->nice);
1809	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1810
1811	/* successful, attach the worker to the pool */
1812	worker_attach_to_pool(worker, pool);
1813
1814	/* start the newly created worker */
1815	spin_lock_irq(&pool->lock);
1816	worker->pool->nr_workers++;
1817	worker_enter_idle(worker);
1818	wake_up_process(worker->task);
1819	spin_unlock_irq(&pool->lock);
1820
1821	return worker;
1822
1823fail:
1824	if (id >= 0)
1825		ida_simple_remove(&pool->worker_ida, id);
 
 
 
1826	kfree(worker);
1827	return NULL;
1828}
1829
1830/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1831 * destroy_worker - destroy a workqueue worker
1832 * @worker: worker to be destroyed
1833 *
1834 * Destroy @worker and adjust @pool stats accordingly.  The worker should
1835 * be idle.
1836 *
1837 * CONTEXT:
1838 * spin_lock_irq(pool->lock).
1839 */
1840static void destroy_worker(struct worker *worker)
1841{
1842	struct worker_pool *pool = worker->pool;
1843
1844	lockdep_assert_held(&pool->lock);
1845
1846	/* sanity check frenzy */
1847	if (WARN_ON(worker->current_work) ||
1848	    WARN_ON(!list_empty(&worker->scheduled)) ||
1849	    WARN_ON(!(worker->flags & WORKER_IDLE)))
1850		return;
1851
1852	pool->nr_workers--;
1853	pool->nr_idle--;
 
 
1854
1855	list_del_init(&worker->entry);
1856	worker->flags |= WORKER_DIE;
1857	wake_up_process(worker->task);
 
 
 
 
 
 
 
1858}
1859
1860static void idle_worker_timeout(unsigned long __pool)
1861{
1862	struct worker_pool *pool = (void *)__pool;
1863
1864	spin_lock_irq(&pool->lock);
1865
1866	while (too_many_workers(pool)) {
1867		struct worker *worker;
1868		unsigned long expires;
1869
1870		/* idle_list is kept in LIFO order, check the last one */
1871		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1872		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1873
1874		if (time_before(jiffies, expires)) {
1875			mod_timer(&pool->idle_timer, expires);
1876			break;
 
 
 
1877		}
1878
1879		destroy_worker(worker);
1880	}
1881
1882	spin_unlock_irq(&pool->lock);
1883}
1884
1885static void send_mayday(struct work_struct *work)
1886{
1887	struct pool_workqueue *pwq = get_work_pwq(work);
1888	struct workqueue_struct *wq = pwq->wq;
1889
1890	lockdep_assert_held(&wq_mayday_lock);
1891
1892	if (!wq->rescuer)
1893		return;
1894
1895	/* mayday mayday mayday */
1896	if (list_empty(&pwq->mayday_node)) {
1897		/*
1898		 * If @pwq is for an unbound wq, its base ref may be put at
1899		 * any time due to an attribute change.  Pin @pwq until the
1900		 * rescuer is done with it.
1901		 */
1902		get_pwq(pwq);
1903		list_add_tail(&pwq->mayday_node, &wq->maydays);
1904		wake_up_process(wq->rescuer->task);
1905	}
1906}
1907
1908static void pool_mayday_timeout(unsigned long __pool)
1909{
1910	struct worker_pool *pool = (void *)__pool;
1911	struct work_struct *work;
1912
1913	spin_lock_irq(&pool->lock);
1914	spin_lock(&wq_mayday_lock);		/* for wq->maydays */
1915
1916	if (need_to_create_worker(pool)) {
1917		/*
1918		 * We've been trying to create a new worker but
1919		 * haven't been successful.  We might be hitting an
1920		 * allocation deadlock.  Send distress signals to
1921		 * rescuers.
1922		 */
1923		list_for_each_entry(work, &pool->worklist, entry)
1924			send_mayday(work);
1925	}
1926
1927	spin_unlock(&wq_mayday_lock);
1928	spin_unlock_irq(&pool->lock);
1929
1930	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1931}
1932
1933/**
1934 * maybe_create_worker - create a new worker if necessary
1935 * @pool: pool to create a new worker for
1936 *
1937 * Create a new worker for @pool if necessary.  @pool is guaranteed to
1938 * have at least one idle worker on return from this function.  If
1939 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1940 * sent to all rescuers with works scheduled on @pool to resolve
1941 * possible allocation deadlock.
1942 *
1943 * On return, need_to_create_worker() is guaranteed to be %false and
1944 * may_start_working() %true.
1945 *
1946 * LOCKING:
1947 * spin_lock_irq(pool->lock) which may be released and regrabbed
1948 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1949 * manager.
 
 
 
 
1950 */
1951static void maybe_create_worker(struct worker_pool *pool)
1952__releases(&pool->lock)
1953__acquires(&pool->lock)
1954{
 
 
1955restart:
1956	spin_unlock_irq(&pool->lock);
1957
1958	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1959	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1960
1961	while (true) {
1962		if (create_worker(pool) || !need_to_create_worker(pool))
 
 
 
 
 
 
 
 
 
 
 
1963			break;
1964
1965		schedule_timeout_interruptible(CREATE_COOLDOWN);
 
1966
1967		if (!need_to_create_worker(pool))
1968			break;
1969	}
1970
1971	del_timer_sync(&pool->mayday_timer);
1972	spin_lock_irq(&pool->lock);
1973	/*
1974	 * This is necessary even after a new worker was just successfully
1975	 * created as @pool->lock was dropped and the new worker might have
1976	 * already become busy.
1977	 */
1978	if (need_to_create_worker(pool))
1979		goto restart;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1980}
1981
1982/**
1983 * manage_workers - manage worker pool
1984 * @worker: self
1985 *
1986 * Assume the manager role and manage the worker pool @worker belongs
1987 * to.  At any given time, there can be only zero or one manager per
1988 * pool.  The exclusion is handled automatically by this function.
1989 *
1990 * The caller can safely start processing works on false return.  On
1991 * true return, it's guaranteed that need_to_create_worker() is false
1992 * and may_start_working() is true.
1993 *
1994 * CONTEXT:
1995 * spin_lock_irq(pool->lock) which may be released and regrabbed
1996 * multiple times.  Does GFP_KERNEL allocations.
1997 *
1998 * Return:
1999 * %false if the pool doesn't need management and the caller can safely
2000 * start processing works, %true if management function was performed and
2001 * the conditions that the caller verified before calling the function may
2002 * no longer be true.
2003 */
2004static bool manage_workers(struct worker *worker)
2005{
2006	struct worker_pool *pool = worker->pool;
 
 
 
 
 
 
 
2007
2008	/*
2009	 * Anyone who successfully grabs manager_arb wins the arbitration
2010	 * and becomes the manager.  mutex_trylock() on pool->manager_arb
2011	 * failure while holding pool->lock reliably indicates that someone
2012	 * else is managing the pool and the worker which failed trylock
2013	 * can proceed to executing work items.  This means that anyone
2014	 * grabbing manager_arb is responsible for actually performing
2015	 * manager duties.  If manager_arb is grabbed and released without
2016	 * actual management, the pool may stall indefinitely.
2017	 */
2018	if (!mutex_trylock(&pool->manager_arb))
2019		return false;
2020	pool->manager = worker;
2021
2022	maybe_create_worker(pool);
2023
2024	pool->manager = NULL;
2025	mutex_unlock(&pool->manager_arb);
2026	return true;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2027}
2028
2029/**
2030 * process_one_work - process single work
2031 * @worker: self
2032 * @work: work to process
2033 *
2034 * Process @work.  This function contains all the logics necessary to
2035 * process a single work including synchronization against and
2036 * interaction with other workers on the same cpu, queueing and
2037 * flushing.  As long as context requirement is met, any worker can
2038 * call this function to process a work.
2039 *
2040 * CONTEXT:
2041 * spin_lock_irq(pool->lock) which is released and regrabbed.
2042 */
2043static void process_one_work(struct worker *worker, struct work_struct *work)
2044__releases(&pool->lock)
2045__acquires(&pool->lock)
2046{
2047	struct pool_workqueue *pwq = get_work_pwq(work);
2048	struct worker_pool *pool = worker->pool;
2049	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
 
 
2050	int work_color;
2051	struct worker *collision;
2052#ifdef CONFIG_LOCKDEP
2053	/*
2054	 * It is permissible to free the struct work_struct from
2055	 * inside the function that is called from it, this we need to
2056	 * take into account for lockdep too.  To avoid bogus "held
2057	 * lock freed" warnings as well as problems when looking into
2058	 * work->lockdep_map, make a copy and use that here.
2059	 */
2060	struct lockdep_map lockdep_map;
2061
2062	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2063#endif
2064	/* ensure we're on the correct CPU */
2065	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2066		     raw_smp_processor_id() != pool->cpu);
2067
2068	/*
2069	 * A single work shouldn't be executed concurrently by
2070	 * multiple workers on a single cpu.  Check whether anyone is
2071	 * already processing the work.  If so, defer the work to the
2072	 * currently executing one.
2073	 */
2074	collision = find_worker_executing_work(pool, work);
2075	if (unlikely(collision)) {
2076		move_linked_works(work, &collision->scheduled, NULL);
2077		return;
2078	}
2079
2080	/* claim and dequeue */
2081	debug_work_deactivate(work);
2082	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2083	worker->current_work = work;
2084	worker->current_func = work->func;
2085	worker->current_pwq = pwq;
2086	work_color = get_work_color(work);
2087
 
 
2088	list_del_init(&work->entry);
2089
2090	/*
2091	 * CPU intensive works don't participate in concurrency management.
2092	 * They're the scheduler's responsibility.  This takes @worker out
2093	 * of concurrency management and the next code block will chain
2094	 * execution of the pending work items.
2095	 */
2096	if (unlikely(cpu_intensive))
2097		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
 
2098
2099	/*
2100	 * Wake up another worker if necessary.  The condition is always
2101	 * false for normal per-cpu workers since nr_running would always
2102	 * be >= 1 at this point.  This is used to chain execution of the
2103	 * pending work items for WORKER_NOT_RUNNING workers such as the
2104	 * UNBOUND and CPU_INTENSIVE ones.
2105	 */
2106	if (need_more_worker(pool))
2107		wake_up_worker(pool);
2108
2109	/*
2110	 * Record the last pool and clear PENDING which should be the last
2111	 * update to @work.  Also, do this inside @pool->lock so that
2112	 * PENDING and queued state changes happen together while IRQ is
2113	 * disabled.
2114	 */
2115	set_work_pool_and_clear_pending(work, pool->id);
 
2116
2117	spin_unlock_irq(&pool->lock);
2118
2119	lock_map_acquire_read(&pwq->wq->lockdep_map);
 
2120	lock_map_acquire(&lockdep_map);
2121	trace_workqueue_execute_start(work);
2122	worker->current_func(work);
2123	/*
2124	 * While we must be careful to not use "work" after this, the trace
2125	 * point will only record its address.
2126	 */
2127	trace_workqueue_execute_end(work);
2128	lock_map_release(&lockdep_map);
2129	lock_map_release(&pwq->wq->lockdep_map);
2130
2131	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2132		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2133		       "     last function: %pf\n",
2134		       current->comm, preempt_count(), task_pid_nr(current),
2135		       worker->current_func);
 
2136		debug_show_held_locks(current);
2137		dump_stack();
2138	}
2139
2140	/*
2141	 * The following prevents a kworker from hogging CPU on !PREEMPT
2142	 * kernels, where a requeueing work item waiting for something to
2143	 * happen could deadlock with stop_machine as such work item could
2144	 * indefinitely requeue itself while all other CPUs are trapped in
2145	 * stop_machine. At the same time, report a quiescent RCU state so
2146	 * the same condition doesn't freeze RCU.
2147	 */
2148	cond_resched_rcu_qs();
2149
2150	spin_lock_irq(&pool->lock);
2151
2152	/* clear cpu intensive status */
2153	if (unlikely(cpu_intensive))
2154		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2155
2156	/* we're done with it, release */
2157	hash_del(&worker->hentry);
2158	worker->current_work = NULL;
2159	worker->current_func = NULL;
2160	worker->current_pwq = NULL;
2161	worker->desc_valid = false;
2162	pwq_dec_nr_in_flight(pwq, work_color);
2163}
2164
2165/**
2166 * process_scheduled_works - process scheduled works
2167 * @worker: self
2168 *
2169 * Process all scheduled works.  Please note that the scheduled list
2170 * may change while processing a work, so this function repeatedly
2171 * fetches a work from the top and executes it.
2172 *
2173 * CONTEXT:
2174 * spin_lock_irq(pool->lock) which may be released and regrabbed
2175 * multiple times.
2176 */
2177static void process_scheduled_works(struct worker *worker)
2178{
2179	while (!list_empty(&worker->scheduled)) {
2180		struct work_struct *work = list_first_entry(&worker->scheduled,
2181						struct work_struct, entry);
2182		process_one_work(worker, work);
2183	}
2184}
2185
2186/**
2187 * worker_thread - the worker thread function
2188 * @__worker: self
2189 *
2190 * The worker thread function.  All workers belong to a worker_pool -
2191 * either a per-cpu one or dynamic unbound one.  These workers process all
2192 * work items regardless of their specific target workqueue.  The only
2193 * exception is work items which belong to workqueues with a rescuer which
2194 * will be explained in rescuer_thread().
2195 *
2196 * Return: 0
2197 */
2198static int worker_thread(void *__worker)
2199{
2200	struct worker *worker = __worker;
2201	struct worker_pool *pool = worker->pool;
2202
2203	/* tell the scheduler that this is a workqueue worker */
2204	worker->task->flags |= PF_WQ_WORKER;
2205woke_up:
2206	spin_lock_irq(&pool->lock);
2207
2208	/* am I supposed to die? */
2209	if (unlikely(worker->flags & WORKER_DIE)) {
2210		spin_unlock_irq(&pool->lock);
2211		WARN_ON_ONCE(!list_empty(&worker->entry));
2212		worker->task->flags &= ~PF_WQ_WORKER;
2213
2214		set_task_comm(worker->task, "kworker/dying");
2215		ida_simple_remove(&pool->worker_ida, worker->id);
2216		worker_detach_from_pool(worker, pool);
2217		kfree(worker);
2218		return 0;
2219	}
2220
2221	worker_leave_idle(worker);
2222recheck:
2223	/* no more worker necessary? */
2224	if (!need_more_worker(pool))
2225		goto sleep;
2226
2227	/* do we need to manage? */
2228	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2229		goto recheck;
2230
2231	/*
2232	 * ->scheduled list can only be filled while a worker is
2233	 * preparing to process a work or actually processing it.
2234	 * Make sure nobody diddled with it while I was sleeping.
2235	 */
2236	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2237
2238	/*
2239	 * Finish PREP stage.  We're guaranteed to have at least one idle
2240	 * worker or that someone else has already assumed the manager
2241	 * role.  This is where @worker starts participating in concurrency
2242	 * management if applicable and concurrency management is restored
2243	 * after being rebound.  See rebind_workers() for details.
2244	 */
2245	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2246
2247	do {
2248		struct work_struct *work =
2249			list_first_entry(&pool->worklist,
2250					 struct work_struct, entry);
2251
2252		pool->watchdog_ts = jiffies;
2253
2254		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2255			/* optimization path, not strictly necessary */
2256			process_one_work(worker, work);
2257			if (unlikely(!list_empty(&worker->scheduled)))
2258				process_scheduled_works(worker);
2259		} else {
2260			move_linked_works(work, &worker->scheduled, NULL);
2261			process_scheduled_works(worker);
2262		}
2263	} while (keep_working(pool));
2264
2265	worker_set_flags(worker, WORKER_PREP);
2266sleep:
 
 
 
2267	/*
2268	 * pool->lock is held and there's no work to process and no need to
2269	 * manage, sleep.  Workers are woken up only while holding
2270	 * pool->lock or from local cpu, so setting the current state
2271	 * before releasing pool->lock is enough to prevent losing any
2272	 * event.
2273	 */
2274	worker_enter_idle(worker);
2275	__set_current_state(TASK_INTERRUPTIBLE);
2276	spin_unlock_irq(&pool->lock);
2277	schedule();
2278	goto woke_up;
2279}
2280
2281/**
2282 * rescuer_thread - the rescuer thread function
2283 * @__rescuer: self
2284 *
2285 * Workqueue rescuer thread function.  There's one rescuer for each
2286 * workqueue which has WQ_MEM_RECLAIM set.
2287 *
2288 * Regular work processing on a pool may block trying to create a new
2289 * worker which uses GFP_KERNEL allocation which has slight chance of
2290 * developing into deadlock if some works currently on the same queue
2291 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2292 * the problem rescuer solves.
2293 *
2294 * When such condition is possible, the pool summons rescuers of all
2295 * workqueues which have works queued on the pool and let them process
2296 * those works so that forward progress can be guaranteed.
2297 *
2298 * This should happen rarely.
2299 *
2300 * Return: 0
2301 */
2302static int rescuer_thread(void *__rescuer)
2303{
2304	struct worker *rescuer = __rescuer;
2305	struct workqueue_struct *wq = rescuer->rescue_wq;
2306	struct list_head *scheduled = &rescuer->scheduled;
2307	bool should_stop;
 
2308
2309	set_user_nice(current, RESCUER_NICE_LEVEL);
2310
2311	/*
2312	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2313	 * doesn't participate in concurrency management.
2314	 */
2315	rescuer->task->flags |= PF_WQ_WORKER;
2316repeat:
2317	set_current_state(TASK_INTERRUPTIBLE);
2318
 
 
 
2319	/*
2320	 * By the time the rescuer is requested to stop, the workqueue
2321	 * shouldn't have any work pending, but @wq->maydays may still have
2322	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2323	 * all the work items before the rescuer got to them.  Go through
2324	 * @wq->maydays processing before acting on should_stop so that the
2325	 * list is always empty on exit.
2326	 */
2327	should_stop = kthread_should_stop();
2328
2329	/* see whether any pwq is asking for help */
2330	spin_lock_irq(&wq_mayday_lock);
2331
2332	while (!list_empty(&wq->maydays)) {
2333		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2334					struct pool_workqueue, mayday_node);
2335		struct worker_pool *pool = pwq->pool;
2336		struct work_struct *work, *n;
2337		bool first = true;
2338
2339		__set_current_state(TASK_RUNNING);
2340		list_del_init(&pwq->mayday_node);
2341
2342		spin_unlock_irq(&wq_mayday_lock);
2343
2344		worker_attach_to_pool(rescuer, pool);
2345
2346		spin_lock_irq(&pool->lock);
2347		rescuer->pool = pool;
2348
2349		/*
2350		 * Slurp in all works issued via this workqueue and
2351		 * process'em.
2352		 */
2353		WARN_ON_ONCE(!list_empty(scheduled));
2354		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2355			if (get_work_pwq(work) == pwq) {
2356				if (first)
2357					pool->watchdog_ts = jiffies;
2358				move_linked_works(work, scheduled, &n);
2359			}
2360			first = false;
2361		}
2362
2363		if (!list_empty(scheduled)) {
2364			process_scheduled_works(rescuer);
2365
2366			/*
2367			 * The above execution of rescued work items could
2368			 * have created more to rescue through
2369			 * pwq_activate_first_delayed() or chained
2370			 * queueing.  Let's put @pwq back on mayday list so
2371			 * that such back-to-back work items, which may be
2372			 * being used to relieve memory pressure, don't
2373			 * incur MAYDAY_INTERVAL delay inbetween.
2374			 */
2375			if (need_to_create_worker(pool)) {
2376				spin_lock(&wq_mayday_lock);
2377				get_pwq(pwq);
2378				list_move_tail(&pwq->mayday_node, &wq->maydays);
2379				spin_unlock(&wq_mayday_lock);
2380			}
2381		}
2382
2383		/*
2384		 * Put the reference grabbed by send_mayday().  @pool won't
2385		 * go away while we're still attached to it.
2386		 */
2387		put_pwq(pwq);
2388
2389		/*
2390		 * Leave this pool.  If need_more_worker() is %true, notify a
2391		 * regular worker; otherwise, we end up with 0 concurrency
2392		 * and stalling the execution.
2393		 */
2394		if (need_more_worker(pool))
2395			wake_up_worker(pool);
2396
2397		rescuer->pool = NULL;
2398		spin_unlock_irq(&pool->lock);
2399
2400		worker_detach_from_pool(rescuer, pool);
2401
2402		spin_lock_irq(&wq_mayday_lock);
2403	}
2404
2405	spin_unlock_irq(&wq_mayday_lock);
2406
2407	if (should_stop) {
2408		__set_current_state(TASK_RUNNING);
2409		rescuer->task->flags &= ~PF_WQ_WORKER;
2410		return 0;
2411	}
2412
2413	/* rescuers should never participate in concurrency management */
2414	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2415	schedule();
2416	goto repeat;
2417}
2418
2419/**
2420 * check_flush_dependency - check for flush dependency sanity
2421 * @target_wq: workqueue being flushed
2422 * @target_work: work item being flushed (NULL for workqueue flushes)
2423 *
2424 * %current is trying to flush the whole @target_wq or @target_work on it.
2425 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2426 * reclaiming memory or running on a workqueue which doesn't have
2427 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2428 * a deadlock.
2429 */
2430static void check_flush_dependency(struct workqueue_struct *target_wq,
2431				   struct work_struct *target_work)
2432{
2433	work_func_t target_func = target_work ? target_work->func : NULL;
2434	struct worker *worker;
2435
2436	if (target_wq->flags & WQ_MEM_RECLAIM)
2437		return;
2438
2439	worker = current_wq_worker();
2440
2441	WARN_ONCE(current->flags & PF_MEMALLOC,
2442		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%pf",
2443		  current->pid, current->comm, target_wq->name, target_func);
2444	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2445			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2446		  "workqueue: WQ_MEM_RECLAIM %s:%pf is flushing !WQ_MEM_RECLAIM %s:%pf",
2447		  worker->current_pwq->wq->name, worker->current_func,
2448		  target_wq->name, target_func);
2449}
2450
2451struct wq_barrier {
2452	struct work_struct	work;
2453	struct completion	done;
2454	struct task_struct	*task;	/* purely informational */
2455};
2456
2457static void wq_barrier_func(struct work_struct *work)
2458{
2459	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2460	complete(&barr->done);
2461}
2462
2463/**
2464 * insert_wq_barrier - insert a barrier work
2465 * @pwq: pwq to insert barrier into
2466 * @barr: wq_barrier to insert
2467 * @target: target work to attach @barr to
2468 * @worker: worker currently executing @target, NULL if @target is not executing
2469 *
2470 * @barr is linked to @target such that @barr is completed only after
2471 * @target finishes execution.  Please note that the ordering
2472 * guarantee is observed only with respect to @target and on the local
2473 * cpu.
2474 *
2475 * Currently, a queued barrier can't be canceled.  This is because
2476 * try_to_grab_pending() can't determine whether the work to be
2477 * grabbed is at the head of the queue and thus can't clear LINKED
2478 * flag of the previous work while there must be a valid next work
2479 * after a work with LINKED flag set.
2480 *
2481 * Note that when @worker is non-NULL, @target may be modified
2482 * underneath us, so we can't reliably determine pwq from @target.
2483 *
2484 * CONTEXT:
2485 * spin_lock_irq(pool->lock).
2486 */
2487static void insert_wq_barrier(struct pool_workqueue *pwq,
2488			      struct wq_barrier *barr,
2489			      struct work_struct *target, struct worker *worker)
2490{
2491	struct list_head *head;
2492	unsigned int linked = 0;
2493
2494	/*
2495	 * debugobject calls are safe here even with pool->lock locked
2496	 * as we know for sure that this will not trigger any of the
2497	 * checks and call back into the fixup functions where we
2498	 * might deadlock.
2499	 */
2500	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2501	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2502	init_completion(&barr->done);
2503	barr->task = current;
2504
2505	/*
2506	 * If @target is currently being executed, schedule the
2507	 * barrier to the worker; otherwise, put it after @target.
2508	 */
2509	if (worker)
2510		head = worker->scheduled.next;
2511	else {
2512		unsigned long *bits = work_data_bits(target);
2513
2514		head = target->entry.next;
2515		/* there can already be other linked works, inherit and set */
2516		linked = *bits & WORK_STRUCT_LINKED;
2517		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2518	}
2519
2520	debug_work_activate(&barr->work);
2521	insert_work(pwq, &barr->work, head,
2522		    work_color_to_flags(WORK_NO_COLOR) | linked);
2523}
2524
2525/**
2526 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2527 * @wq: workqueue being flushed
2528 * @flush_color: new flush color, < 0 for no-op
2529 * @work_color: new work color, < 0 for no-op
2530 *
2531 * Prepare pwqs for workqueue flushing.
2532 *
2533 * If @flush_color is non-negative, flush_color on all pwqs should be
2534 * -1.  If no pwq has in-flight commands at the specified color, all
2535 * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2536 * has in flight commands, its pwq->flush_color is set to
2537 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2538 * wakeup logic is armed and %true is returned.
2539 *
2540 * The caller should have initialized @wq->first_flusher prior to
2541 * calling this function with non-negative @flush_color.  If
2542 * @flush_color is negative, no flush color update is done and %false
2543 * is returned.
2544 *
2545 * If @work_color is non-negative, all pwqs should have the same
2546 * work_color which is previous to @work_color and all will be
2547 * advanced to @work_color.
2548 *
2549 * CONTEXT:
2550 * mutex_lock(wq->mutex).
2551 *
2552 * Return:
2553 * %true if @flush_color >= 0 and there's something to flush.  %false
2554 * otherwise.
2555 */
2556static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2557				      int flush_color, int work_color)
2558{
2559	bool wait = false;
2560	struct pool_workqueue *pwq;
2561
2562	if (flush_color >= 0) {
2563		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2564		atomic_set(&wq->nr_pwqs_to_flush, 1);
2565	}
2566
2567	for_each_pwq(pwq, wq) {
2568		struct worker_pool *pool = pwq->pool;
 
2569
2570		spin_lock_irq(&pool->lock);
2571
2572		if (flush_color >= 0) {
2573			WARN_ON_ONCE(pwq->flush_color != -1);
2574
2575			if (pwq->nr_in_flight[flush_color]) {
2576				pwq->flush_color = flush_color;
2577				atomic_inc(&wq->nr_pwqs_to_flush);
2578				wait = true;
2579			}
2580		}
2581
2582		if (work_color >= 0) {
2583			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2584			pwq->work_color = work_color;
2585		}
2586
2587		spin_unlock_irq(&pool->lock);
2588	}
2589
2590	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2591		complete(&wq->first_flusher->done);
2592
2593	return wait;
2594}
2595
2596/**
2597 * flush_workqueue - ensure that any scheduled work has run to completion.
2598 * @wq: workqueue to flush
2599 *
2600 * This function sleeps until all work items which were queued on entry
2601 * have finished execution, but it is not livelocked by new incoming ones.
 
 
 
2602 */
2603void flush_workqueue(struct workqueue_struct *wq)
2604{
2605	struct wq_flusher this_flusher = {
2606		.list = LIST_HEAD_INIT(this_flusher.list),
2607		.flush_color = -1,
2608		.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2609	};
2610	int next_color;
2611
2612	lock_map_acquire(&wq->lockdep_map);
2613	lock_map_release(&wq->lockdep_map);
2614
2615	mutex_lock(&wq->mutex);
2616
2617	/*
2618	 * Start-to-wait phase
2619	 */
2620	next_color = work_next_color(wq->work_color);
2621
2622	if (next_color != wq->flush_color) {
2623		/*
2624		 * Color space is not full.  The current work_color
2625		 * becomes our flush_color and work_color is advanced
2626		 * by one.
2627		 */
2628		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2629		this_flusher.flush_color = wq->work_color;
2630		wq->work_color = next_color;
2631
2632		if (!wq->first_flusher) {
2633			/* no flush in progress, become the first flusher */
2634			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2635
2636			wq->first_flusher = &this_flusher;
2637
2638			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2639						       wq->work_color)) {
2640				/* nothing to flush, done */
2641				wq->flush_color = next_color;
2642				wq->first_flusher = NULL;
2643				goto out_unlock;
2644			}
2645		} else {
2646			/* wait in queue */
2647			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2648			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2649			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2650		}
2651	} else {
2652		/*
2653		 * Oops, color space is full, wait on overflow queue.
2654		 * The next flush completion will assign us
2655		 * flush_color and transfer to flusher_queue.
2656		 */
2657		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2658	}
2659
2660	check_flush_dependency(wq, NULL);
2661
2662	mutex_unlock(&wq->mutex);
2663
2664	wait_for_completion(&this_flusher.done);
2665
2666	/*
2667	 * Wake-up-and-cascade phase
2668	 *
2669	 * First flushers are responsible for cascading flushes and
2670	 * handling overflow.  Non-first flushers can simply return.
2671	 */
2672	if (wq->first_flusher != &this_flusher)
2673		return;
2674
2675	mutex_lock(&wq->mutex);
2676
2677	/* we might have raced, check again with mutex held */
2678	if (wq->first_flusher != &this_flusher)
2679		goto out_unlock;
2680
2681	wq->first_flusher = NULL;
2682
2683	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2684	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2685
2686	while (true) {
2687		struct wq_flusher *next, *tmp;
2688
2689		/* complete all the flushers sharing the current flush color */
2690		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2691			if (next->flush_color != wq->flush_color)
2692				break;
2693			list_del_init(&next->list);
2694			complete(&next->done);
2695		}
2696
2697		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2698			     wq->flush_color != work_next_color(wq->work_color));
2699
2700		/* this flush_color is finished, advance by one */
2701		wq->flush_color = work_next_color(wq->flush_color);
2702
2703		/* one color has been freed, handle overflow queue */
2704		if (!list_empty(&wq->flusher_overflow)) {
2705			/*
2706			 * Assign the same color to all overflowed
2707			 * flushers, advance work_color and append to
2708			 * flusher_queue.  This is the start-to-wait
2709			 * phase for these overflowed flushers.
2710			 */
2711			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2712				tmp->flush_color = wq->work_color;
2713
2714			wq->work_color = work_next_color(wq->work_color);
2715
2716			list_splice_tail_init(&wq->flusher_overflow,
2717					      &wq->flusher_queue);
2718			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2719		}
2720
2721		if (list_empty(&wq->flusher_queue)) {
2722			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2723			break;
2724		}
2725
2726		/*
2727		 * Need to flush more colors.  Make the next flusher
2728		 * the new first flusher and arm pwqs.
2729		 */
2730		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2731		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2732
2733		list_del_init(&next->list);
2734		wq->first_flusher = next;
2735
2736		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2737			break;
2738
2739		/*
2740		 * Meh... this color is already done, clear first
2741		 * flusher and repeat cascading.
2742		 */
2743		wq->first_flusher = NULL;
2744	}
2745
2746out_unlock:
2747	mutex_unlock(&wq->mutex);
2748}
2749EXPORT_SYMBOL(flush_workqueue);
2750
2751/**
2752 * drain_workqueue - drain a workqueue
2753 * @wq: workqueue to drain
2754 *
2755 * Wait until the workqueue becomes empty.  While draining is in progress,
2756 * only chain queueing is allowed.  IOW, only currently pending or running
2757 * work items on @wq can queue further work items on it.  @wq is flushed
2758 * repeatedly until it becomes empty.  The number of flushing is determined
2759 * by the depth of chaining and should be relatively short.  Whine if it
2760 * takes too long.
2761 */
2762void drain_workqueue(struct workqueue_struct *wq)
2763{
2764	unsigned int flush_cnt = 0;
2765	struct pool_workqueue *pwq;
2766
2767	/*
2768	 * __queue_work() needs to test whether there are drainers, is much
2769	 * hotter than drain_workqueue() and already looks at @wq->flags.
2770	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2771	 */
2772	mutex_lock(&wq->mutex);
2773	if (!wq->nr_drainers++)
2774		wq->flags |= __WQ_DRAINING;
2775	mutex_unlock(&wq->mutex);
2776reflush:
2777	flush_workqueue(wq);
2778
2779	mutex_lock(&wq->mutex);
2780
2781	for_each_pwq(pwq, wq) {
2782		bool drained;
2783
2784		spin_lock_irq(&pwq->pool->lock);
2785		drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2786		spin_unlock_irq(&pwq->pool->lock);
2787
2788		if (drained)
2789			continue;
2790
2791		if (++flush_cnt == 10 ||
2792		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2793			pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2794				wq->name, flush_cnt);
2795
2796		mutex_unlock(&wq->mutex);
2797		goto reflush;
2798	}
2799
 
2800	if (!--wq->nr_drainers)
2801		wq->flags &= ~__WQ_DRAINING;
2802	mutex_unlock(&wq->mutex);
2803}
2804EXPORT_SYMBOL_GPL(drain_workqueue);
2805
2806static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr)
 
2807{
2808	struct worker *worker = NULL;
2809	struct worker_pool *pool;
2810	struct pool_workqueue *pwq;
2811
2812	might_sleep();
2813
2814	local_irq_disable();
2815	pool = get_work_pool(work);
2816	if (!pool) {
2817		local_irq_enable();
2818		return false;
2819	}
2820
2821	spin_lock(&pool->lock);
2822	/* see the comment in try_to_grab_pending() with the same code */
2823	pwq = get_work_pwq(work);
2824	if (pwq) {
2825		if (unlikely(pwq->pool != pool))
 
 
 
 
 
2826			goto already_gone;
2827	} else {
2828		worker = find_worker_executing_work(pool, work);
2829		if (!worker)
2830			goto already_gone;
2831		pwq = worker->current_pwq;
2832	}
2833
2834	check_flush_dependency(pwq->wq, work);
2835
2836	insert_wq_barrier(pwq, barr, work, worker);
2837	spin_unlock_irq(&pool->lock);
2838
2839	/*
2840	 * If @max_active is 1 or rescuer is in use, flushing another work
2841	 * item on the same workqueue may lead to deadlock.  Make sure the
2842	 * flusher is not running on the same workqueue by verifying write
2843	 * access.
2844	 */
2845	if (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)
2846		lock_map_acquire(&pwq->wq->lockdep_map);
2847	else
2848		lock_map_acquire_read(&pwq->wq->lockdep_map);
2849	lock_map_release(&pwq->wq->lockdep_map);
2850
2851	return true;
2852already_gone:
2853	spin_unlock_irq(&pool->lock);
2854	return false;
2855}
2856
2857/**
2858 * flush_work - wait for a work to finish executing the last queueing instance
2859 * @work: the work to flush
2860 *
2861 * Wait until @work has finished execution.  @work is guaranteed to be idle
2862 * on return if it hasn't been requeued since flush started.
 
 
 
 
 
 
 
2863 *
2864 * Return:
2865 * %true if flush_work() waited for the work to finish execution,
2866 * %false if it was already idle.
2867 */
2868bool flush_work(struct work_struct *work)
2869{
2870	struct wq_barrier barr;
2871
2872	lock_map_acquire(&work->lockdep_map);
2873	lock_map_release(&work->lockdep_map);
2874
2875	if (start_flush_work(work, &barr)) {
2876		wait_for_completion(&barr.done);
2877		destroy_work_on_stack(&barr.work);
2878		return true;
2879	} else {
2880		return false;
2881	}
2882}
2883EXPORT_SYMBOL_GPL(flush_work);
2884
2885struct cwt_wait {
2886	wait_queue_t		wait;
2887	struct work_struct	*work;
2888};
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2889
2890static int cwt_wakefn(wait_queue_t *wait, unsigned mode, int sync, void *key)
2891{
2892	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
 
2893
2894	if (cwait->work != key)
2895		return 0;
2896	return autoremove_wake_function(wait, mode, sync, key);
 
 
 
 
 
2897}
2898
2899static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2900{
2901	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
2902	unsigned long flags;
2903	int ret;
2904
2905	do {
2906		ret = try_to_grab_pending(work, is_dwork, &flags);
2907		/*
2908		 * If someone else is already canceling, wait for it to
2909		 * finish.  flush_work() doesn't work for PREEMPT_NONE
2910		 * because we may get scheduled between @work's completion
2911		 * and the other canceling task resuming and clearing
2912		 * CANCELING - flush_work() will return false immediately
2913		 * as @work is no longer busy, try_to_grab_pending() will
2914		 * return -ENOENT as @work is still being canceled and the
2915		 * other canceling task won't be able to clear CANCELING as
2916		 * we're hogging the CPU.
2917		 *
2918		 * Let's wait for completion using a waitqueue.  As this
2919		 * may lead to the thundering herd problem, use a custom
2920		 * wake function which matches @work along with exclusive
2921		 * wait and wakeup.
2922		 */
2923		if (unlikely(ret == -ENOENT)) {
2924			struct cwt_wait cwait;
2925
2926			init_wait(&cwait.wait);
2927			cwait.wait.func = cwt_wakefn;
2928			cwait.work = work;
2929
2930			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
2931						  TASK_UNINTERRUPTIBLE);
2932			if (work_is_canceling(work))
2933				schedule();
2934			finish_wait(&cancel_waitq, &cwait.wait);
2935		}
2936	} while (unlikely(ret < 0));
2937
2938	/* tell other tasks trying to grab @work to back off */
2939	mark_work_canceling(work);
2940	local_irq_restore(flags);
 
 
2941
2942	flush_work(work);
2943	clear_work_data(work);
 
 
 
 
 
 
 
 
 
 
 
 
 
2944
2945	/*
2946	 * Paired with prepare_to_wait() above so that either
2947	 * waitqueue_active() is visible here or !work_is_canceling() is
2948	 * visible there.
2949	 */
2950	smp_mb();
2951	if (waitqueue_active(&cancel_waitq))
2952		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2953
2954	return ret;
2955}
2956
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2957/**
2958 * cancel_work_sync - cancel a work and wait for it to finish
2959 * @work: the work to cancel
2960 *
2961 * Cancel @work and wait for its execution to finish.  This function
2962 * can be used even if the work re-queues itself or migrates to
2963 * another workqueue.  On return from this function, @work is
2964 * guaranteed to be not pending or executing on any CPU.
2965 *
2966 * cancel_work_sync(&delayed_work->work) must not be used for
2967 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2968 *
2969 * The caller must ensure that the workqueue on which @work was last
2970 * queued can't be destroyed before this function returns.
2971 *
2972 * Return:
2973 * %true if @work was pending, %false otherwise.
2974 */
2975bool cancel_work_sync(struct work_struct *work)
2976{
2977	return __cancel_work_timer(work, false);
2978}
2979EXPORT_SYMBOL_GPL(cancel_work_sync);
2980
2981/**
2982 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2983 * @dwork: the delayed work to flush
2984 *
2985 * Delayed timer is cancelled and the pending work is queued for
2986 * immediate execution.  Like flush_work(), this function only
2987 * considers the last queueing instance of @dwork.
2988 *
2989 * Return:
2990 * %true if flush_work() waited for the work to finish execution,
2991 * %false if it was already idle.
2992 */
2993bool flush_delayed_work(struct delayed_work *dwork)
2994{
2995	local_irq_disable();
2996	if (del_timer_sync(&dwork->timer))
2997		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
2998	local_irq_enable();
2999	return flush_work(&dwork->work);
3000}
3001EXPORT_SYMBOL(flush_delayed_work);
3002
3003/**
3004 * cancel_delayed_work - cancel a delayed work
3005 * @dwork: delayed_work to cancel
3006 *
3007 * Kill off a pending delayed_work.
3008 *
3009 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3010 * pending.
3011 *
3012 * Note:
3013 * The work callback function may still be running on return, unless
3014 * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3015 * use cancel_delayed_work_sync() to wait on it.
3016 *
3017 * This function is safe to call from any context including IRQ handler.
 
 
3018 */
3019bool cancel_delayed_work(struct delayed_work *dwork)
3020{
3021	unsigned long flags;
3022	int ret;
3023
3024	do {
3025		ret = try_to_grab_pending(&dwork->work, true, &flags);
3026	} while (unlikely(ret == -EAGAIN));
3027
3028	if (unlikely(ret < 0))
3029		return false;
3030
3031	set_work_pool_and_clear_pending(&dwork->work,
3032					get_work_pool_id(&dwork->work));
3033	local_irq_restore(flags);
3034	return ret;
3035}
3036EXPORT_SYMBOL(cancel_delayed_work);
3037
3038/**
3039 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3040 * @dwork: the delayed work cancel
3041 *
3042 * This is cancel_work_sync() for delayed works.
3043 *
3044 * Return:
3045 * %true if @dwork was pending, %false otherwise.
3046 */
3047bool cancel_delayed_work_sync(struct delayed_work *dwork)
3048{
3049	return __cancel_work_timer(&dwork->work, true);
3050}
3051EXPORT_SYMBOL(cancel_delayed_work_sync);
3052
3053/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3054 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3055 * @func: the function to call
3056 *
3057 * schedule_on_each_cpu() executes @func on each online CPU using the
3058 * system workqueue and blocks until all CPUs have completed.
3059 * schedule_on_each_cpu() is very slow.
3060 *
3061 * Return:
3062 * 0 on success, -errno on failure.
3063 */
3064int schedule_on_each_cpu(work_func_t func)
3065{
3066	int cpu;
3067	struct work_struct __percpu *works;
3068
3069	works = alloc_percpu(struct work_struct);
3070	if (!works)
3071		return -ENOMEM;
3072
3073	get_online_cpus();
3074
3075	for_each_online_cpu(cpu) {
3076		struct work_struct *work = per_cpu_ptr(works, cpu);
3077
3078		INIT_WORK(work, func);
3079		schedule_work_on(cpu, work);
3080	}
3081
3082	for_each_online_cpu(cpu)
3083		flush_work(per_cpu_ptr(works, cpu));
3084
3085	put_online_cpus();
3086	free_percpu(works);
3087	return 0;
3088}
3089
3090/**
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3091 * execute_in_process_context - reliably execute the routine with user context
3092 * @fn:		the function to execute
3093 * @ew:		guaranteed storage for the execute work structure (must
3094 *		be available when the work executes)
3095 *
3096 * Executes the function immediately if process context is available,
3097 * otherwise schedules the function for delayed execution.
3098 *
3099 * Return:	0 - function was executed
3100 *		1 - function was scheduled for execution
3101 */
3102int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3103{
3104	if (!in_interrupt()) {
3105		fn(&ew->work);
3106		return 0;
3107	}
3108
3109	INIT_WORK(&ew->work, fn);
3110	schedule_work(&ew->work);
3111
3112	return 1;
3113}
3114EXPORT_SYMBOL_GPL(execute_in_process_context);
3115
3116/**
3117 * free_workqueue_attrs - free a workqueue_attrs
3118 * @attrs: workqueue_attrs to free
3119 *
3120 * Undo alloc_workqueue_attrs().
3121 */
3122void free_workqueue_attrs(struct workqueue_attrs *attrs)
3123{
3124	if (attrs) {
3125		free_cpumask_var(attrs->cpumask);
3126		kfree(attrs);
3127	}
3128}
3129
3130/**
3131 * alloc_workqueue_attrs - allocate a workqueue_attrs
3132 * @gfp_mask: allocation mask to use
3133 *
3134 * Allocate a new workqueue_attrs, initialize with default settings and
3135 * return it.
3136 *
3137 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3138 */
3139struct workqueue_attrs *alloc_workqueue_attrs(gfp_t gfp_mask)
3140{
3141	struct workqueue_attrs *attrs;
3142
3143	attrs = kzalloc(sizeof(*attrs), gfp_mask);
3144	if (!attrs)
3145		goto fail;
3146	if (!alloc_cpumask_var(&attrs->cpumask, gfp_mask))
3147		goto fail;
3148
3149	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3150	return attrs;
3151fail:
3152	free_workqueue_attrs(attrs);
3153	return NULL;
3154}
3155
3156static void copy_workqueue_attrs(struct workqueue_attrs *to,
3157				 const struct workqueue_attrs *from)
3158{
3159	to->nice = from->nice;
3160	cpumask_copy(to->cpumask, from->cpumask);
3161	/*
3162	 * Unlike hash and equality test, this function doesn't ignore
3163	 * ->no_numa as it is used for both pool and wq attrs.  Instead,
3164	 * get_unbound_pool() explicitly clears ->no_numa after copying.
3165	 */
3166	to->no_numa = from->no_numa;
3167}
3168
3169/* hash value of the content of @attr */
3170static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3171{
3172	u32 hash = 0;
3173
3174	hash = jhash_1word(attrs->nice, hash);
3175	hash = jhash(cpumask_bits(attrs->cpumask),
3176		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3177	return hash;
3178}
3179
3180/* content equality test */
3181static bool wqattrs_equal(const struct workqueue_attrs *a,
3182			  const struct workqueue_attrs *b)
3183{
3184	if (a->nice != b->nice)
3185		return false;
3186	if (!cpumask_equal(a->cpumask, b->cpumask))
3187		return false;
3188	return true;
3189}
3190
3191/**
3192 * init_worker_pool - initialize a newly zalloc'd worker_pool
3193 * @pool: worker_pool to initialize
3194 *
3195 * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3196 *
3197 * Return: 0 on success, -errno on failure.  Even on failure, all fields
3198 * inside @pool proper are initialized and put_unbound_pool() can be called
3199 * on @pool safely to release it.
3200 */
3201static int init_worker_pool(struct worker_pool *pool)
3202{
3203	spin_lock_init(&pool->lock);
3204	pool->id = -1;
3205	pool->cpu = -1;
3206	pool->node = NUMA_NO_NODE;
3207	pool->flags |= POOL_DISASSOCIATED;
3208	pool->watchdog_ts = jiffies;
3209	INIT_LIST_HEAD(&pool->worklist);
3210	INIT_LIST_HEAD(&pool->idle_list);
3211	hash_init(pool->busy_hash);
3212
3213	init_timer_deferrable(&pool->idle_timer);
3214	pool->idle_timer.function = idle_worker_timeout;
3215	pool->idle_timer.data = (unsigned long)pool;
3216
3217	setup_timer(&pool->mayday_timer, pool_mayday_timeout,
3218		    (unsigned long)pool);
3219
3220	mutex_init(&pool->manager_arb);
3221	mutex_init(&pool->attach_mutex);
3222	INIT_LIST_HEAD(&pool->workers);
3223
3224	ida_init(&pool->worker_ida);
3225	INIT_HLIST_NODE(&pool->hash_node);
3226	pool->refcnt = 1;
3227
3228	/* shouldn't fail above this point */
3229	pool->attrs = alloc_workqueue_attrs(GFP_KERNEL);
3230	if (!pool->attrs)
3231		return -ENOMEM;
3232	return 0;
3233}
3234
3235static void rcu_free_wq(struct rcu_head *rcu)
3236{
3237	struct workqueue_struct *wq =
3238		container_of(rcu, struct workqueue_struct, rcu);
3239
3240	if (!(wq->flags & WQ_UNBOUND))
3241		free_percpu(wq->cpu_pwqs);
3242	else
3243		free_workqueue_attrs(wq->unbound_attrs);
3244
3245	kfree(wq->rescuer);
3246	kfree(wq);
3247}
3248
3249static void rcu_free_pool(struct rcu_head *rcu)
3250{
3251	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3252
3253	ida_destroy(&pool->worker_ida);
3254	free_workqueue_attrs(pool->attrs);
3255	kfree(pool);
3256}
3257
3258/**
3259 * put_unbound_pool - put a worker_pool
3260 * @pool: worker_pool to put
3261 *
3262 * Put @pool.  If its refcnt reaches zero, it gets destroyed in sched-RCU
3263 * safe manner.  get_unbound_pool() calls this function on its failure path
3264 * and this function should be able to release pools which went through,
3265 * successfully or not, init_worker_pool().
3266 *
3267 * Should be called with wq_pool_mutex held.
3268 */
3269static void put_unbound_pool(struct worker_pool *pool)
3270{
3271	DECLARE_COMPLETION_ONSTACK(detach_completion);
3272	struct worker *worker;
3273
3274	lockdep_assert_held(&wq_pool_mutex);
3275
3276	if (--pool->refcnt)
3277		return;
3278
3279	/* sanity checks */
3280	if (WARN_ON(!(pool->cpu < 0)) ||
3281	    WARN_ON(!list_empty(&pool->worklist)))
3282		return;
3283
3284	/* release id and unhash */
3285	if (pool->id >= 0)
3286		idr_remove(&worker_pool_idr, pool->id);
3287	hash_del(&pool->hash_node);
3288
3289	/*
3290	 * Become the manager and destroy all workers.  Grabbing
3291	 * manager_arb prevents @pool's workers from blocking on
3292	 * attach_mutex.
3293	 */
3294	mutex_lock(&pool->manager_arb);
3295
3296	spin_lock_irq(&pool->lock);
3297	while ((worker = first_idle_worker(pool)))
3298		destroy_worker(worker);
3299	WARN_ON(pool->nr_workers || pool->nr_idle);
3300	spin_unlock_irq(&pool->lock);
3301
3302	mutex_lock(&pool->attach_mutex);
3303	if (!list_empty(&pool->workers))
3304		pool->detach_completion = &detach_completion;
3305	mutex_unlock(&pool->attach_mutex);
3306
3307	if (pool->detach_completion)
3308		wait_for_completion(pool->detach_completion);
3309
3310	mutex_unlock(&pool->manager_arb);
3311
3312	/* shut down the timers */
3313	del_timer_sync(&pool->idle_timer);
3314	del_timer_sync(&pool->mayday_timer);
3315
3316	/* sched-RCU protected to allow dereferences from get_work_pool() */
3317	call_rcu_sched(&pool->rcu, rcu_free_pool);
3318}
3319
3320/**
3321 * get_unbound_pool - get a worker_pool with the specified attributes
3322 * @attrs: the attributes of the worker_pool to get
3323 *
3324 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3325 * reference count and return it.  If there already is a matching
3326 * worker_pool, it will be used; otherwise, this function attempts to
3327 * create a new one.
3328 *
3329 * Should be called with wq_pool_mutex held.
3330 *
3331 * Return: On success, a worker_pool with the same attributes as @attrs.
3332 * On failure, %NULL.
3333 */
3334static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3335{
3336	u32 hash = wqattrs_hash(attrs);
3337	struct worker_pool *pool;
3338	int node;
3339	int target_node = NUMA_NO_NODE;
3340
3341	lockdep_assert_held(&wq_pool_mutex);
3342
3343	/* do we already have a matching pool? */
3344	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3345		if (wqattrs_equal(pool->attrs, attrs)) {
3346			pool->refcnt++;
3347			return pool;
3348		}
3349	}
3350
3351	/* if cpumask is contained inside a NUMA node, we belong to that node */
3352	if (wq_numa_enabled) {
3353		for_each_node(node) {
3354			if (cpumask_subset(attrs->cpumask,
3355					   wq_numa_possible_cpumask[node])) {
3356				target_node = node;
3357				break;
3358			}
3359		}
3360	}
3361
3362	/* nope, create a new one */
3363	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3364	if (!pool || init_worker_pool(pool) < 0)
3365		goto fail;
3366
3367	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3368	copy_workqueue_attrs(pool->attrs, attrs);
3369	pool->node = target_node;
3370
3371	/*
3372	 * no_numa isn't a worker_pool attribute, always clear it.  See
3373	 * 'struct workqueue_attrs' comments for detail.
3374	 */
3375	pool->attrs->no_numa = false;
3376
3377	if (worker_pool_assign_id(pool) < 0)
3378		goto fail;
3379
3380	/* create and start the initial worker */
3381	if (!create_worker(pool))
3382		goto fail;
3383
3384	/* install */
3385	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3386
3387	return pool;
3388fail:
3389	if (pool)
3390		put_unbound_pool(pool);
3391	return NULL;
3392}
3393
3394static void rcu_free_pwq(struct rcu_head *rcu)
3395{
3396	kmem_cache_free(pwq_cache,
3397			container_of(rcu, struct pool_workqueue, rcu));
3398}
3399
3400/*
3401 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3402 * and needs to be destroyed.
3403 */
3404static void pwq_unbound_release_workfn(struct work_struct *work)
3405{
3406	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3407						  unbound_release_work);
3408	struct workqueue_struct *wq = pwq->wq;
3409	struct worker_pool *pool = pwq->pool;
3410	bool is_last;
3411
3412	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3413		return;
3414
3415	mutex_lock(&wq->mutex);
3416	list_del_rcu(&pwq->pwqs_node);
3417	is_last = list_empty(&wq->pwqs);
3418	mutex_unlock(&wq->mutex);
3419
3420	mutex_lock(&wq_pool_mutex);
3421	put_unbound_pool(pool);
3422	mutex_unlock(&wq_pool_mutex);
3423
3424	call_rcu_sched(&pwq->rcu, rcu_free_pwq);
3425
3426	/*
3427	 * If we're the last pwq going away, @wq is already dead and no one
3428	 * is gonna access it anymore.  Schedule RCU free.
3429	 */
3430	if (is_last)
3431		call_rcu_sched(&wq->rcu, rcu_free_wq);
3432}
3433
3434/**
3435 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3436 * @pwq: target pool_workqueue
3437 *
3438 * If @pwq isn't freezing, set @pwq->max_active to the associated
3439 * workqueue's saved_max_active and activate delayed work items
3440 * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3441 */
3442static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3443{
3444	struct workqueue_struct *wq = pwq->wq;
3445	bool freezable = wq->flags & WQ_FREEZABLE;
3446
3447	/* for @wq->saved_max_active */
3448	lockdep_assert_held(&wq->mutex);
3449
3450	/* fast exit for non-freezable wqs */
3451	if (!freezable && pwq->max_active == wq->saved_max_active)
3452		return;
3453
3454	spin_lock_irq(&pwq->pool->lock);
3455
3456	/*
3457	 * During [un]freezing, the caller is responsible for ensuring that
3458	 * this function is called at least once after @workqueue_freezing
3459	 * is updated and visible.
3460	 */
3461	if (!freezable || !workqueue_freezing) {
3462		pwq->max_active = wq->saved_max_active;
3463
3464		while (!list_empty(&pwq->delayed_works) &&
3465		       pwq->nr_active < pwq->max_active)
3466			pwq_activate_first_delayed(pwq);
3467
3468		/*
3469		 * Need to kick a worker after thawed or an unbound wq's
3470		 * max_active is bumped.  It's a slow path.  Do it always.
 
3471		 */
3472		wake_up_worker(pwq->pool);
3473	} else {
3474		pwq->max_active = 0;
3475	}
3476
3477	spin_unlock_irq(&pwq->pool->lock);
3478}
3479
3480/* initialize newly alloced @pwq which is associated with @wq and @pool */
3481static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3482		     struct worker_pool *pool)
3483{
3484	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3485
3486	memset(pwq, 0, sizeof(*pwq));
3487
3488	pwq->pool = pool;
3489	pwq->wq = wq;
3490	pwq->flush_color = -1;
3491	pwq->refcnt = 1;
3492	INIT_LIST_HEAD(&pwq->delayed_works);
3493	INIT_LIST_HEAD(&pwq->pwqs_node);
3494	INIT_LIST_HEAD(&pwq->mayday_node);
3495	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3496}
3497
3498/* sync @pwq with the current state of its associated wq and link it */
3499static void link_pwq(struct pool_workqueue *pwq)
3500{
3501	struct workqueue_struct *wq = pwq->wq;
3502
3503	lockdep_assert_held(&wq->mutex);
3504
3505	/* may be called multiple times, ignore if already linked */
3506	if (!list_empty(&pwq->pwqs_node))
3507		return;
3508
3509	/* set the matching work_color */
3510	pwq->work_color = wq->work_color;
3511
3512	/* sync max_active to the current setting */
3513	pwq_adjust_max_active(pwq);
3514
3515	/* link in @pwq */
3516	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3517}
3518
3519/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3520static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3521					const struct workqueue_attrs *attrs)
3522{
3523	struct worker_pool *pool;
3524	struct pool_workqueue *pwq;
3525
3526	lockdep_assert_held(&wq_pool_mutex);
3527
3528	pool = get_unbound_pool(attrs);
3529	if (!pool)
3530		return NULL;
3531
3532	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3533	if (!pwq) {
3534		put_unbound_pool(pool);
3535		return NULL;
3536	}
3537
3538	init_pwq(pwq, wq, pool);
3539	return pwq;
3540}
3541
3542/**
3543 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3544 * @attrs: the wq_attrs of the default pwq of the target workqueue
3545 * @node: the target NUMA node
3546 * @cpu_going_down: if >= 0, the CPU to consider as offline
3547 * @cpumask: outarg, the resulting cpumask
3548 *
3549 * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3550 * @cpu_going_down is >= 0, that cpu is considered offline during
3551 * calculation.  The result is stored in @cpumask.
3552 *
3553 * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3554 * enabled and @node has online CPUs requested by @attrs, the returned
3555 * cpumask is the intersection of the possible CPUs of @node and
3556 * @attrs->cpumask.
3557 *
3558 * The caller is responsible for ensuring that the cpumask of @node stays
3559 * stable.
3560 *
3561 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3562 * %false if equal.
3563 */
3564static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3565				 int cpu_going_down, cpumask_t *cpumask)
3566{
3567	if (!wq_numa_enabled || attrs->no_numa)
3568		goto use_dfl;
3569
3570	/* does @node have any online CPUs @attrs wants? */
3571	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3572	if (cpu_going_down >= 0)
3573		cpumask_clear_cpu(cpu_going_down, cpumask);
3574
3575	if (cpumask_empty(cpumask))
3576		goto use_dfl;
3577
3578	/* yeap, return possible CPUs in @node that @attrs wants */
3579	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3580	return !cpumask_equal(cpumask, attrs->cpumask);
3581
3582use_dfl:
3583	cpumask_copy(cpumask, attrs->cpumask);
3584	return false;
3585}
3586
3587/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3588static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3589						   int node,
3590						   struct pool_workqueue *pwq)
3591{
3592	struct pool_workqueue *old_pwq;
3593
3594	lockdep_assert_held(&wq_pool_mutex);
3595	lockdep_assert_held(&wq->mutex);
3596
3597	/* link_pwq() can handle duplicate calls */
3598	link_pwq(pwq);
3599
3600	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3601	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3602	return old_pwq;
3603}
3604
3605/* context to store the prepared attrs & pwqs before applying */
3606struct apply_wqattrs_ctx {
3607	struct workqueue_struct	*wq;		/* target workqueue */
3608	struct workqueue_attrs	*attrs;		/* attrs to apply */
3609	struct list_head	list;		/* queued for batching commit */
3610	struct pool_workqueue	*dfl_pwq;
3611	struct pool_workqueue	*pwq_tbl[];
3612};
3613
3614/* free the resources after success or abort */
3615static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3616{
3617	if (ctx) {
3618		int node;
3619
3620		for_each_node(node)
3621			put_pwq_unlocked(ctx->pwq_tbl[node]);
3622		put_pwq_unlocked(ctx->dfl_pwq);
3623
3624		free_workqueue_attrs(ctx->attrs);
3625
3626		kfree(ctx);
3627	}
3628}
3629
3630/* allocate the attrs and pwqs for later installation */
3631static struct apply_wqattrs_ctx *
3632apply_wqattrs_prepare(struct workqueue_struct *wq,
3633		      const struct workqueue_attrs *attrs)
3634{
3635	struct apply_wqattrs_ctx *ctx;
3636	struct workqueue_attrs *new_attrs, *tmp_attrs;
3637	int node;
3638
3639	lockdep_assert_held(&wq_pool_mutex);
3640
3641	ctx = kzalloc(sizeof(*ctx) + nr_node_ids * sizeof(ctx->pwq_tbl[0]),
3642		      GFP_KERNEL);
3643
3644	new_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3645	tmp_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3646	if (!ctx || !new_attrs || !tmp_attrs)
3647		goto out_free;
3648
3649	/*
3650	 * Calculate the attrs of the default pwq.
3651	 * If the user configured cpumask doesn't overlap with the
3652	 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3653	 */
3654	copy_workqueue_attrs(new_attrs, attrs);
3655	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3656	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3657		cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3658
3659	/*
3660	 * We may create multiple pwqs with differing cpumasks.  Make a
3661	 * copy of @new_attrs which will be modified and used to obtain
3662	 * pools.
3663	 */
3664	copy_workqueue_attrs(tmp_attrs, new_attrs);
3665
3666	/*
3667	 * If something goes wrong during CPU up/down, we'll fall back to
3668	 * the default pwq covering whole @attrs->cpumask.  Always create
3669	 * it even if we don't use it immediately.
3670	 */
3671	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3672	if (!ctx->dfl_pwq)
3673		goto out_free;
3674
3675	for_each_node(node) {
3676		if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3677			ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3678			if (!ctx->pwq_tbl[node])
3679				goto out_free;
3680		} else {
3681			ctx->dfl_pwq->refcnt++;
3682			ctx->pwq_tbl[node] = ctx->dfl_pwq;
3683		}
3684	}
3685
3686	/* save the user configured attrs and sanitize it. */
3687	copy_workqueue_attrs(new_attrs, attrs);
3688	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3689	ctx->attrs = new_attrs;
3690
3691	ctx->wq = wq;
3692	free_workqueue_attrs(tmp_attrs);
3693	return ctx;
3694
3695out_free:
3696	free_workqueue_attrs(tmp_attrs);
3697	free_workqueue_attrs(new_attrs);
3698	apply_wqattrs_cleanup(ctx);
3699	return NULL;
3700}
3701
3702/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3703static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3704{
3705	int node;
3706
3707	/* all pwqs have been created successfully, let's install'em */
3708	mutex_lock(&ctx->wq->mutex);
3709
3710	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3711
3712	/* save the previous pwq and install the new one */
3713	for_each_node(node)
3714		ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3715							  ctx->pwq_tbl[node]);
3716
3717	/* @dfl_pwq might not have been used, ensure it's linked */
3718	link_pwq(ctx->dfl_pwq);
3719	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3720
3721	mutex_unlock(&ctx->wq->mutex);
3722}
3723
3724static void apply_wqattrs_lock(void)
3725{
3726	/* CPUs should stay stable across pwq creations and installations */
3727	get_online_cpus();
3728	mutex_lock(&wq_pool_mutex);
3729}
3730
3731static void apply_wqattrs_unlock(void)
3732{
3733	mutex_unlock(&wq_pool_mutex);
3734	put_online_cpus();
3735}
3736
3737static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3738					const struct workqueue_attrs *attrs)
3739{
3740	struct apply_wqattrs_ctx *ctx;
3741
3742	/* only unbound workqueues can change attributes */
3743	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3744		return -EINVAL;
3745
3746	/* creating multiple pwqs breaks ordering guarantee */
3747	if (WARN_ON((wq->flags & __WQ_ORDERED) && !list_empty(&wq->pwqs)))
3748		return -EINVAL;
3749
3750	ctx = apply_wqattrs_prepare(wq, attrs);
3751	if (!ctx)
3752		return -ENOMEM;
3753
3754	/* the ctx has been prepared successfully, let's commit it */
3755	apply_wqattrs_commit(ctx);
3756	apply_wqattrs_cleanup(ctx);
3757
3758	return 0;
3759}
3760
3761/**
3762 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
3763 * @wq: the target workqueue
3764 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
3765 *
3766 * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
3767 * machines, this function maps a separate pwq to each NUMA node with
3768 * possibles CPUs in @attrs->cpumask so that work items are affine to the
3769 * NUMA node it was issued on.  Older pwqs are released as in-flight work
3770 * items finish.  Note that a work item which repeatedly requeues itself
3771 * back-to-back will stay on its current pwq.
3772 *
3773 * Performs GFP_KERNEL allocations.
3774 *
3775 * Return: 0 on success and -errno on failure.
3776 */
3777int apply_workqueue_attrs(struct workqueue_struct *wq,
3778			  const struct workqueue_attrs *attrs)
3779{
3780	int ret;
3781
3782	apply_wqattrs_lock();
3783	ret = apply_workqueue_attrs_locked(wq, attrs);
3784	apply_wqattrs_unlock();
3785
3786	return ret;
3787}
3788
3789/**
3790 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
3791 * @wq: the target workqueue
3792 * @cpu: the CPU coming up or going down
3793 * @online: whether @cpu is coming up or going down
3794 *
3795 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
3796 * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
3797 * @wq accordingly.
3798 *
3799 * If NUMA affinity can't be adjusted due to memory allocation failure, it
3800 * falls back to @wq->dfl_pwq which may not be optimal but is always
3801 * correct.
3802 *
3803 * Note that when the last allowed CPU of a NUMA node goes offline for a
3804 * workqueue with a cpumask spanning multiple nodes, the workers which were
3805 * already executing the work items for the workqueue will lose their CPU
3806 * affinity and may execute on any CPU.  This is similar to how per-cpu
3807 * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
3808 * affinity, it's the user's responsibility to flush the work item from
3809 * CPU_DOWN_PREPARE.
3810 */
3811static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
3812				   bool online)
3813{
3814	int node = cpu_to_node(cpu);
3815	int cpu_off = online ? -1 : cpu;
3816	struct pool_workqueue *old_pwq = NULL, *pwq;
3817	struct workqueue_attrs *target_attrs;
3818	cpumask_t *cpumask;
3819
3820	lockdep_assert_held(&wq_pool_mutex);
3821
3822	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
3823	    wq->unbound_attrs->no_numa)
3824		return;
3825
3826	/*
3827	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
3828	 * Let's use a preallocated one.  The following buf is protected by
3829	 * CPU hotplug exclusion.
3830	 */
3831	target_attrs = wq_update_unbound_numa_attrs_buf;
3832	cpumask = target_attrs->cpumask;
3833
3834	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
3835	pwq = unbound_pwq_by_node(wq, node);
3836
3837	/*
3838	 * Let's determine what needs to be done.  If the target cpumask is
3839	 * different from the default pwq's, we need to compare it to @pwq's
3840	 * and create a new one if they don't match.  If the target cpumask
3841	 * equals the default pwq's, the default pwq should be used.
3842	 */
3843	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
3844		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
3845			return;
3846	} else {
3847		goto use_dfl_pwq;
3848	}
3849
3850	/* create a new pwq */
3851	pwq = alloc_unbound_pwq(wq, target_attrs);
3852	if (!pwq) {
3853		pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
3854			wq->name);
3855		goto use_dfl_pwq;
3856	}
3857
3858	/* Install the new pwq. */
3859	mutex_lock(&wq->mutex);
3860	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
3861	goto out_unlock;
3862
3863use_dfl_pwq:
3864	mutex_lock(&wq->mutex);
3865	spin_lock_irq(&wq->dfl_pwq->pool->lock);
3866	get_pwq(wq->dfl_pwq);
3867	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
3868	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
3869out_unlock:
3870	mutex_unlock(&wq->mutex);
3871	put_pwq_unlocked(old_pwq);
3872}
3873
3874static int alloc_and_link_pwqs(struct workqueue_struct *wq)
3875{
3876	bool highpri = wq->flags & WQ_HIGHPRI;
3877	int cpu, ret;
3878
3879	if (!(wq->flags & WQ_UNBOUND)) {
3880		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
3881		if (!wq->cpu_pwqs)
3882			return -ENOMEM;
3883
3884		for_each_possible_cpu(cpu) {
3885			struct pool_workqueue *pwq =
3886				per_cpu_ptr(wq->cpu_pwqs, cpu);
3887			struct worker_pool *cpu_pools =
3888				per_cpu(cpu_worker_pools, cpu);
3889
3890			init_pwq(pwq, wq, &cpu_pools[highpri]);
3891
3892			mutex_lock(&wq->mutex);
3893			link_pwq(pwq);
3894			mutex_unlock(&wq->mutex);
3895		}
3896		return 0;
3897	} else if (wq->flags & __WQ_ORDERED) {
3898		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
3899		/* there should only be single pwq for ordering guarantee */
3900		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
3901			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
3902		     "ordering guarantee broken for workqueue %s\n", wq->name);
3903		return ret;
3904	} else {
3905		return apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
3906	}
3907}
3908
3909static int wq_clamp_max_active(int max_active, unsigned int flags,
3910			       const char *name)
3911{
3912	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3913
3914	if (max_active < 1 || max_active > lim)
3915		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
3916			max_active, name, 1, lim);
 
3917
3918	return clamp_val(max_active, 1, lim);
3919}
3920
3921struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
3922					       unsigned int flags,
3923					       int max_active,
3924					       struct lock_class_key *key,
3925					       const char *lock_name, ...)
3926{
3927	size_t tbl_size = 0;
3928	va_list args;
3929	struct workqueue_struct *wq;
3930	struct pool_workqueue *pwq;
3931
3932	/* see the comment above the definition of WQ_POWER_EFFICIENT */
3933	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
3934		flags |= WQ_UNBOUND;
3935
3936	/* allocate wq and format name */
3937	if (flags & WQ_UNBOUND)
3938		tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
 
3939
3940	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
3941	if (!wq)
3942		return NULL;
3943
3944	if (flags & WQ_UNBOUND) {
3945		wq->unbound_attrs = alloc_workqueue_attrs(GFP_KERNEL);
3946		if (!wq->unbound_attrs)
3947			goto err_free_wq;
3948	}
3949
3950	va_start(args, lock_name);
3951	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
3952	va_end(args);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3953
3954	max_active = max_active ?: WQ_DFL_ACTIVE;
3955	max_active = wq_clamp_max_active(max_active, flags, wq->name);
3956
3957	/* init wq */
3958	wq->flags = flags;
3959	wq->saved_max_active = max_active;
3960	mutex_init(&wq->mutex);
3961	atomic_set(&wq->nr_pwqs_to_flush, 0);
3962	INIT_LIST_HEAD(&wq->pwqs);
3963	INIT_LIST_HEAD(&wq->flusher_queue);
3964	INIT_LIST_HEAD(&wq->flusher_overflow);
3965	INIT_LIST_HEAD(&wq->maydays);
3966
3967	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3968	INIT_LIST_HEAD(&wq->list);
3969
3970	if (alloc_and_link_pwqs(wq) < 0)
3971		goto err_free_wq;
3972
3973	/*
3974	 * Workqueues which may be used during memory reclaim should
3975	 * have a rescuer to guarantee forward progress.
3976	 */
3977	if (flags & WQ_MEM_RECLAIM) {
 
 
 
 
 
 
 
 
3978		struct worker *rescuer;
3979
3980		rescuer = alloc_worker(NUMA_NO_NODE);
 
 
 
3981		if (!rescuer)
3982			goto err_destroy;
3983
3984		rescuer->rescue_wq = wq;
3985		rescuer->task = kthread_create(rescuer_thread, rescuer, "%s",
3986					       wq->name);
3987		if (IS_ERR(rescuer->task)) {
3988			kfree(rescuer);
3989			goto err_destroy;
3990		}
3991
3992		wq->rescuer = rescuer;
3993		kthread_bind_mask(rescuer->task, cpu_possible_mask);
3994		wake_up_process(rescuer->task);
3995	}
3996
3997	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
3998		goto err_destroy;
3999
4000	/*
4001	 * wq_pool_mutex protects global freeze state and workqueues list.
4002	 * Grab it, adjust max_active and add the new @wq to workqueues
4003	 * list.
4004	 */
4005	mutex_lock(&wq_pool_mutex);
4006
4007	mutex_lock(&wq->mutex);
4008	for_each_pwq(pwq, wq)
4009		pwq_adjust_max_active(pwq);
4010	mutex_unlock(&wq->mutex);
4011
4012	list_add_tail_rcu(&wq->list, &workqueues);
4013
4014	mutex_unlock(&wq_pool_mutex);
4015
4016	return wq;
4017
4018err_free_wq:
4019	free_workqueue_attrs(wq->unbound_attrs);
4020	kfree(wq);
4021	return NULL;
4022err_destroy:
4023	destroy_workqueue(wq);
4024	return NULL;
4025}
4026EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
4027
4028/**
4029 * destroy_workqueue - safely terminate a workqueue
4030 * @wq: target workqueue
4031 *
4032 * Safely destroy a workqueue. All work currently pending will be done first.
4033 */
4034void destroy_workqueue(struct workqueue_struct *wq)
4035{
4036	struct pool_workqueue *pwq;
4037	int node;
4038
4039	/* drain it before proceeding with destruction */
4040	drain_workqueue(wq);
4041
4042	/* sanity checks */
4043	mutex_lock(&wq->mutex);
4044	for_each_pwq(pwq, wq) {
4045		int i;
4046
4047		for (i = 0; i < WORK_NR_COLORS; i++) {
4048			if (WARN_ON(pwq->nr_in_flight[i])) {
4049				mutex_unlock(&wq->mutex);
4050				return;
4051			}
4052		}
4053
4054		if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4055		    WARN_ON(pwq->nr_active) ||
4056		    WARN_ON(!list_empty(&pwq->delayed_works))) {
4057			mutex_unlock(&wq->mutex);
4058			return;
4059		}
4060	}
4061	mutex_unlock(&wq->mutex);
4062
4063	/*
4064	 * wq list is used to freeze wq, remove from list after
4065	 * flushing is complete in case freeze races us.
4066	 */
4067	mutex_lock(&wq_pool_mutex);
4068	list_del_rcu(&wq->list);
4069	mutex_unlock(&wq_pool_mutex);
 
 
 
 
 
4070
4071	workqueue_sysfs_unregister(wq);
 
 
 
 
4072
4073	if (wq->rescuer)
4074		kthread_stop(wq->rescuer->task);
4075
4076	if (!(wq->flags & WQ_UNBOUND)) {
4077		/*
4078		 * The base ref is never dropped on per-cpu pwqs.  Directly
4079		 * schedule RCU free.
4080		 */
4081		call_rcu_sched(&wq->rcu, rcu_free_wq);
4082	} else {
4083		/*
4084		 * We're the sole accessor of @wq at this point.  Directly
4085		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4086		 * @wq will be freed when the last pwq is released.
4087		 */
4088		for_each_node(node) {
4089			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4090			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4091			put_pwq_unlocked(pwq);
4092		}
4093
4094		/*
4095		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4096		 * put.  Don't access it afterwards.
4097		 */
4098		pwq = wq->dfl_pwq;
4099		wq->dfl_pwq = NULL;
4100		put_pwq_unlocked(pwq);
4101	}
 
 
 
4102}
4103EXPORT_SYMBOL_GPL(destroy_workqueue);
4104
4105/**
4106 * workqueue_set_max_active - adjust max_active of a workqueue
4107 * @wq: target workqueue
4108 * @max_active: new max_active value.
4109 *
4110 * Set max_active of @wq to @max_active.
4111 *
4112 * CONTEXT:
4113 * Don't call from IRQ context.
4114 */
4115void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4116{
4117	struct pool_workqueue *pwq;
4118
4119	/* disallow meddling with max_active for ordered workqueues */
4120	if (WARN_ON(wq->flags & __WQ_ORDERED))
4121		return;
4122
4123	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4124
4125	mutex_lock(&wq->mutex);
4126
4127	wq->saved_max_active = max_active;
4128
4129	for_each_pwq(pwq, wq)
4130		pwq_adjust_max_active(pwq);
4131
4132	mutex_unlock(&wq->mutex);
4133}
4134EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4135
4136/**
4137 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4138 *
4139 * Determine whether %current is a workqueue rescuer.  Can be used from
4140 * work functions to determine whether it's being run off the rescuer task.
4141 *
4142 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4143 */
4144bool current_is_workqueue_rescuer(void)
4145{
4146	struct worker *worker = current_wq_worker();
4147
4148	return worker && worker->rescue_wq;
4149}
 
4150
4151/**
4152 * workqueue_congested - test whether a workqueue is congested
4153 * @cpu: CPU in question
4154 * @wq: target workqueue
4155 *
4156 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4157 * no synchronization around this function and the test result is
4158 * unreliable and only useful as advisory hints or for debugging.
4159 *
4160 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4161 * Note that both per-cpu and unbound workqueues may be associated with
4162 * multiple pool_workqueues which have separate congested states.  A
4163 * workqueue being congested on one CPU doesn't mean the workqueue is also
4164 * contested on other CPUs / NUMA nodes.
4165 *
4166 * Return:
4167 * %true if congested, %false otherwise.
4168 */
4169bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4170{
4171	struct pool_workqueue *pwq;
4172	bool ret;
4173
4174	rcu_read_lock_sched();
4175
4176	if (cpu == WORK_CPU_UNBOUND)
4177		cpu = smp_processor_id();
4178
4179	if (!(wq->flags & WQ_UNBOUND))
4180		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4181	else
4182		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4183
4184	ret = !list_empty(&pwq->delayed_works);
4185	rcu_read_unlock_sched();
 
 
 
 
 
 
 
 
4186
4187	return ret;
4188}
4189EXPORT_SYMBOL_GPL(workqueue_congested);
4190
4191/**
4192 * work_busy - test whether a work is currently pending or running
4193 * @work: the work to be tested
4194 *
4195 * Test whether @work is currently pending or running.  There is no
4196 * synchronization around this function and the test result is
4197 * unreliable and only useful as advisory hints or for debugging.
 
 
4198 *
4199 * Return:
4200 * OR'd bitmask of WORK_BUSY_* bits.
4201 */
4202unsigned int work_busy(struct work_struct *work)
4203{
4204	struct worker_pool *pool;
4205	unsigned long flags;
4206	unsigned int ret = 0;
4207
 
 
 
 
 
4208	if (work_pending(work))
4209		ret |= WORK_BUSY_PENDING;
 
 
4210
4211	local_irq_save(flags);
4212	pool = get_work_pool(work);
4213	if (pool) {
4214		spin_lock(&pool->lock);
4215		if (find_worker_executing_work(pool, work))
4216			ret |= WORK_BUSY_RUNNING;
4217		spin_unlock(&pool->lock);
4218	}
4219	local_irq_restore(flags);
4220
4221	return ret;
4222}
4223EXPORT_SYMBOL_GPL(work_busy);
4224
4225/**
4226 * set_worker_desc - set description for the current work item
4227 * @fmt: printf-style format string
4228 * @...: arguments for the format string
4229 *
4230 * This function can be called by a running work function to describe what
4231 * the work item is about.  If the worker task gets dumped, this
4232 * information will be printed out together to help debugging.  The
4233 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4234 */
4235void set_worker_desc(const char *fmt, ...)
4236{
4237	struct worker *worker = current_wq_worker();
4238	va_list args;
4239
4240	if (worker) {
4241		va_start(args, fmt);
4242		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4243		va_end(args);
4244		worker->desc_valid = true;
4245	}
4246}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4247
4248/**
4249 * print_worker_info - print out worker information and description
4250 * @log_lvl: the log level to use when printing
4251 * @task: target task
4252 *
4253 * If @task is a worker and currently executing a work item, print out the
4254 * name of the workqueue being serviced and worker description set with
4255 * set_worker_desc() by the currently executing work item.
4256 *
4257 * This function can be safely called on any task as long as the
4258 * task_struct itself is accessible.  While safe, this function isn't
4259 * synchronized and may print out mixups or garbages of limited length.
 
 
 
4260 */
4261void print_worker_info(const char *log_lvl, struct task_struct *task)
 
 
 
 
 
 
4262{
4263	work_func_t *fn = NULL;
4264	char name[WQ_NAME_LEN] = { };
4265	char desc[WORKER_DESC_LEN] = { };
4266	struct pool_workqueue *pwq = NULL;
4267	struct workqueue_struct *wq = NULL;
4268	bool desc_valid = false;
4269	struct worker *worker;
 
 
 
 
4270
4271	if (!(task->flags & PF_WQ_WORKER))
4272		return;
4273
 
4274	/*
4275	 * This function is called without any synchronization and @task
4276	 * could be in any state.  Be careful with dereferences.
 
4277	 */
4278	worker = probe_kthread_data(task);
 
 
4279
4280	/*
4281	 * Carefully copy the associated workqueue's workfn and name.  Keep
4282	 * the original last '\0' in case the original contains garbage.
4283	 */
4284	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4285	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4286	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4287	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4288
4289	/* copy worker description */
4290	probe_kernel_read(&desc_valid, &worker->desc_valid, sizeof(desc_valid));
4291	if (desc_valid)
4292		probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
4293
4294	if (fn || name[0] || desc[0]) {
4295		printk("%sWorkqueue: %s %pf", log_lvl, name, fn);
4296		if (desc[0])
4297			pr_cont(" (%s)", desc);
4298		pr_cont("\n");
4299	}
4300}
4301
4302static void pr_cont_pool_info(struct worker_pool *pool)
4303{
4304	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4305	if (pool->node != NUMA_NO_NODE)
4306		pr_cont(" node=%d", pool->node);
4307	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4308}
4309
4310static void pr_cont_work(bool comma, struct work_struct *work)
4311{
4312	if (work->func == wq_barrier_func) {
4313		struct wq_barrier *barr;
4314
4315		barr = container_of(work, struct wq_barrier, work);
4316
4317		pr_cont("%s BAR(%d)", comma ? "," : "",
4318			task_pid_nr(barr->task));
4319	} else {
4320		pr_cont("%s %pf", comma ? "," : "", work->func);
4321	}
4322}
4323
4324static void show_pwq(struct pool_workqueue *pwq)
4325{
4326	struct worker_pool *pool = pwq->pool;
4327	struct work_struct *work;
4328	struct worker *worker;
4329	bool has_in_flight = false, has_pending = false;
4330	int bkt;
 
 
4331
4332	pr_info("  pwq %d:", pool->id);
4333	pr_cont_pool_info(pool);
 
 
 
 
 
4334
4335	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4336		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
 
4337
4338	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4339		if (worker->current_pwq == pwq) {
4340			has_in_flight = true;
4341			break;
4342		}
4343	}
4344	if (has_in_flight) {
4345		bool comma = false;
4346
4347		pr_info("    in-flight:");
4348		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4349			if (worker->current_pwq != pwq)
4350				continue;
 
 
 
 
 
 
 
 
 
 
4351
4352			pr_cont("%s %d%s:%pf", comma ? "," : "",
4353				task_pid_nr(worker->task),
4354				worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4355				worker->current_func);
4356			list_for_each_entry(work, &worker->scheduled, entry)
4357				pr_cont_work(false, work);
4358			comma = true;
4359		}
4360		pr_cont("\n");
4361	}
4362
4363	list_for_each_entry(work, &pool->worklist, entry) {
4364		if (get_work_pwq(work) == pwq) {
4365			has_pending = true;
4366			break;
4367		}
4368	}
4369	if (has_pending) {
4370		bool comma = false;
4371
4372		pr_info("    pending:");
4373		list_for_each_entry(work, &pool->worklist, entry) {
4374			if (get_work_pwq(work) != pwq)
4375				continue;
4376
4377			pr_cont_work(comma, work);
4378			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
 
 
 
 
 
 
4379		}
4380		pr_cont("\n");
4381	}
4382
4383	if (!list_empty(&pwq->delayed_works)) {
4384		bool comma = false;
4385
4386		pr_info("    delayed:");
4387		list_for_each_entry(work, &pwq->delayed_works, entry) {
4388			pr_cont_work(comma, work);
4389			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4390		}
4391		pr_cont("\n");
4392	}
4393}
4394
4395/**
4396 * show_workqueue_state - dump workqueue state
4397 *
4398 * Called from a sysrq handler and prints out all busy workqueues and
4399 * pools.
4400 */
4401void show_workqueue_state(void)
4402{
4403	struct workqueue_struct *wq;
4404	struct worker_pool *pool;
4405	unsigned long flags;
4406	int pi;
4407
4408	rcu_read_lock_sched();
 
 
 
 
 
 
 
4409
4410	pr_info("Showing busy workqueues and worker pools:\n");
 
 
4411
4412	list_for_each_entry_rcu(wq, &workqueues, list) {
4413		struct pool_workqueue *pwq;
4414		bool idle = true;
 
 
 
 
 
 
4415
4416		for_each_pwq(pwq, wq) {
4417			if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4418				idle = false;
4419				break;
4420			}
4421		}
4422		if (idle)
4423			continue;
4424
4425		pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4426
4427		for_each_pwq(pwq, wq) {
4428			spin_lock_irqsave(&pwq->pool->lock, flags);
4429			if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4430				show_pwq(pwq);
4431			spin_unlock_irqrestore(&pwq->pool->lock, flags);
4432		}
4433	}
4434
4435	for_each_pool(pool, pi) {
4436		struct worker *worker;
4437		bool first = true;
4438
4439		spin_lock_irqsave(&pool->lock, flags);
4440		if (pool->nr_workers == pool->nr_idle)
4441			goto next_pool;
4442
4443		pr_info("pool %d:", pool->id);
4444		pr_cont_pool_info(pool);
4445		pr_cont(" hung=%us workers=%d",
4446			jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4447			pool->nr_workers);
4448		if (pool->manager)
4449			pr_cont(" manager: %d",
4450				task_pid_nr(pool->manager->task));
4451		list_for_each_entry(worker, &pool->idle_list, entry) {
4452			pr_cont(" %s%d", first ? "idle: " : "",
4453				task_pid_nr(worker->task));
4454			first = false;
4455		}
4456		pr_cont("\n");
4457	next_pool:
4458		spin_unlock_irqrestore(&pool->lock, flags);
4459	}
4460
4461	rcu_read_unlock_sched();
 
 
 
 
 
 
 
 
4462}
4463
4464/*
4465 * CPU hotplug.
 
 
4466 *
4467 * There are two challenges in supporting CPU hotplug.  Firstly, there
4468 * are a lot of assumptions on strong associations among work, pwq and
4469 * pool which make migrating pending and scheduled works very
4470 * difficult to implement without impacting hot paths.  Secondly,
4471 * worker pools serve mix of short, long and very long running works making
4472 * blocked draining impractical.
4473 *
4474 * This is solved by allowing the pools to be disassociated from the CPU
4475 * running as an unbound one and allowing it to be reattached later if the
4476 * cpu comes back online.
4477 */
4478
4479static void wq_unbind_fn(struct work_struct *work)
4480{
4481	int cpu = smp_processor_id();
4482	struct worker_pool *pool;
4483	struct worker *worker;
4484
4485	for_each_cpu_worker_pool(pool, cpu) {
4486		mutex_lock(&pool->attach_mutex);
4487		spin_lock_irq(&pool->lock);
4488
4489		/*
4490		 * We've blocked all attach/detach operations. Make all workers
4491		 * unbound and set DISASSOCIATED.  Before this, all workers
4492		 * except for the ones which are still executing works from
4493		 * before the last CPU down must be on the cpu.  After
4494		 * this, they may become diasporas.
4495		 */
4496		for_each_pool_worker(worker, pool)
4497			worker->flags |= WORKER_UNBOUND;
4498
4499		pool->flags |= POOL_DISASSOCIATED;
4500
4501		spin_unlock_irq(&pool->lock);
4502		mutex_unlock(&pool->attach_mutex);
4503
4504		/*
4505		 * Call schedule() so that we cross rq->lock and thus can
4506		 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4507		 * This is necessary as scheduler callbacks may be invoked
4508		 * from other cpus.
4509		 */
4510		schedule();
4511
4512		/*
4513		 * Sched callbacks are disabled now.  Zap nr_running.
4514		 * After this, nr_running stays zero and need_more_worker()
4515		 * and keep_working() are always true as long as the
4516		 * worklist is not empty.  This pool now behaves as an
4517		 * unbound (in terms of concurrency management) pool which
4518		 * are served by workers tied to the pool.
4519		 */
4520		atomic_set(&pool->nr_running, 0);
4521
4522		/*
4523		 * With concurrency management just turned off, a busy
4524		 * worker blocking could lead to lengthy stalls.  Kick off
4525		 * unbound chain execution of currently pending work items.
4526		 */
4527		spin_lock_irq(&pool->lock);
4528		wake_up_worker(pool);
4529		spin_unlock_irq(&pool->lock);
4530	}
4531}
4532
4533/**
4534 * rebind_workers - rebind all workers of a pool to the associated CPU
4535 * @pool: pool of interest
4536 *
4537 * @pool->cpu is coming online.  Rebind all workers to the CPU.
4538 */
4539static void rebind_workers(struct worker_pool *pool)
4540{
4541	struct worker *worker;
4542
4543	lockdep_assert_held(&pool->attach_mutex);
4544
4545	/*
4546	 * Restore CPU affinity of all workers.  As all idle workers should
4547	 * be on the run-queue of the associated CPU before any local
4548	 * wake-ups for concurrency management happen, restore CPU affinity
4549	 * of all workers first and then clear UNBOUND.  As we're called
4550	 * from CPU_ONLINE, the following shouldn't fail.
4551	 */
4552	for_each_pool_worker(worker, pool)
4553		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4554						  pool->attrs->cpumask) < 0);
4555
4556	spin_lock_irq(&pool->lock);
4557
4558	/*
4559	 * XXX: CPU hotplug notifiers are weird and can call DOWN_FAILED
4560	 * w/o preceding DOWN_PREPARE.  Work around it.  CPU hotplug is
4561	 * being reworked and this can go away in time.
4562	 */
4563	if (!(pool->flags & POOL_DISASSOCIATED)) {
4564		spin_unlock_irq(&pool->lock);
4565		return;
 
 
 
 
 
 
 
 
4566	}
4567
4568	pool->flags &= ~POOL_DISASSOCIATED;
4569
4570	for_each_pool_worker(worker, pool) {
4571		unsigned int worker_flags = worker->flags;
4572
4573		/*
4574		 * A bound idle worker should actually be on the runqueue
4575		 * of the associated CPU for local wake-ups targeting it to
4576		 * work.  Kick all idle workers so that they migrate to the
4577		 * associated CPU.  Doing this in the same loop as
4578		 * replacing UNBOUND with REBOUND is safe as no worker will
4579		 * be bound before @pool->lock is released.
4580		 */
4581		if (worker_flags & WORKER_IDLE)
4582			wake_up_process(worker->task);
 
 
 
4583
 
4584		/*
4585		 * We want to clear UNBOUND but can't directly call
4586		 * worker_clr_flags() or adjust nr_running.  Atomically
4587		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4588		 * @worker will clear REBOUND using worker_clr_flags() when
4589		 * it initiates the next execution cycle thus restoring
4590		 * concurrency management.  Note that when or whether
4591		 * @worker clears REBOUND doesn't affect correctness.
4592		 *
4593		 * ACCESS_ONCE() is necessary because @worker->flags may be
4594		 * tested without holding any lock in
4595		 * wq_worker_waking_up().  Without it, NOT_RUNNING test may
4596		 * fail incorrectly leading to premature concurrency
4597		 * management operations.
4598		 */
4599		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4600		worker_flags |= WORKER_REBOUND;
4601		worker_flags &= ~WORKER_UNBOUND;
4602		ACCESS_ONCE(worker->flags) = worker_flags;
4603	}
4604
4605	spin_unlock_irq(&pool->lock);
4606}
4607
4608/**
4609 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4610 * @pool: unbound pool of interest
4611 * @cpu: the CPU which is coming up
4612 *
4613 * An unbound pool may end up with a cpumask which doesn't have any online
4614 * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4615 * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4616 * online CPU before, cpus_allowed of all its workers should be restored.
4617 */
4618static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4619{
4620	static cpumask_t cpumask;
4621	struct worker *worker;
4622
4623	lockdep_assert_held(&pool->attach_mutex);
 
 
 
 
 
 
 
4624
4625	/* is @cpu allowed for @pool? */
4626	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4627		return;
 
 
 
 
 
 
 
 
 
 
4628
4629	/* is @cpu the only online CPU? */
4630	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
4631	if (cpumask_weight(&cpumask) != 1)
4632		return;
4633
4634	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4635	for_each_pool_worker(worker, pool)
4636		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4637						  pool->attrs->cpumask) < 0);
4638}
4639
4640/*
4641 * Workqueues should be brought up before normal priority CPU notifiers.
4642 * This will be registered high priority CPU notifier.
4643 */
4644static int workqueue_cpu_up_callback(struct notifier_block *nfb,
4645					       unsigned long action,
4646					       void *hcpu)
4647{
4648	int cpu = (unsigned long)hcpu;
4649	struct worker_pool *pool;
4650	struct workqueue_struct *wq;
4651	int pi;
4652
4653	switch (action & ~CPU_TASKS_FROZEN) {
4654	case CPU_UP_PREPARE:
4655		for_each_cpu_worker_pool(pool, cpu) {
4656			if (pool->nr_workers)
4657				continue;
4658			if (!create_worker(pool))
4659				return NOTIFY_BAD;
4660		}
4661		break;
4662
4663	case CPU_DOWN_FAILED:
4664	case CPU_ONLINE:
4665		mutex_lock(&wq_pool_mutex);
4666
4667		for_each_pool(pool, pi) {
4668			mutex_lock(&pool->attach_mutex);
4669
4670			if (pool->cpu == cpu)
4671				rebind_workers(pool);
4672			else if (pool->cpu < 0)
4673				restore_unbound_workers_cpumask(pool, cpu);
4674
4675			mutex_unlock(&pool->attach_mutex);
4676		}
4677
4678		/* update NUMA affinity of unbound workqueues */
4679		list_for_each_entry(wq, &workqueues, list)
4680			wq_update_unbound_numa(wq, cpu, true);
4681
4682		mutex_unlock(&wq_pool_mutex);
4683		break;
4684	}
4685	return NOTIFY_OK;
4686}
4687
4688/*
4689 * Workqueues should be brought down after normal priority CPU notifiers.
4690 * This will be registered as low priority CPU notifier.
4691 */
4692static int workqueue_cpu_down_callback(struct notifier_block *nfb,
4693						 unsigned long action,
4694						 void *hcpu)
4695{
4696	int cpu = (unsigned long)hcpu;
4697	struct work_struct unbind_work;
4698	struct workqueue_struct *wq;
4699
4700	switch (action & ~CPU_TASKS_FROZEN) {
4701	case CPU_DOWN_PREPARE:
4702		/* unbinding per-cpu workers should happen on the local CPU */
4703		INIT_WORK_ONSTACK(&unbind_work, wq_unbind_fn);
4704		queue_work_on(cpu, system_highpri_wq, &unbind_work);
4705
4706		/* update NUMA affinity of unbound workqueues */
4707		mutex_lock(&wq_pool_mutex);
4708		list_for_each_entry(wq, &workqueues, list)
4709			wq_update_unbound_numa(wq, cpu, false);
4710		mutex_unlock(&wq_pool_mutex);
4711
4712		/* wait for per-cpu unbinding to finish */
4713		flush_work(&unbind_work);
4714		destroy_work_on_stack(&unbind_work);
4715		break;
4716	}
4717	return NOTIFY_OK;
4718}
4719
4720#ifdef CONFIG_SMP
4721
4722struct work_for_cpu {
4723	struct work_struct work;
4724	long (*fn)(void *);
4725	void *arg;
4726	long ret;
4727};
4728
4729static void work_for_cpu_fn(struct work_struct *work)
4730{
4731	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
4732
4733	wfc->ret = wfc->fn(wfc->arg);
4734}
4735
4736/**
4737 * work_on_cpu - run a function in thread context on a particular cpu
4738 * @cpu: the cpu to run on
4739 * @fn: the function to run
4740 * @arg: the function arg
4741 *
 
4742 * It is up to the caller to ensure that the cpu doesn't go offline.
4743 * The caller must not hold any locks which would prevent @fn from completing.
4744 *
4745 * Return: The value @fn returns.
4746 */
4747long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
4748{
4749	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
4750
4751	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
4752	schedule_work_on(cpu, &wfc.work);
4753	flush_work(&wfc.work);
4754	destroy_work_on_stack(&wfc.work);
4755	return wfc.ret;
4756}
4757EXPORT_SYMBOL_GPL(work_on_cpu);
4758#endif /* CONFIG_SMP */
4759
4760#ifdef CONFIG_FREEZER
4761
4762/**
4763 * freeze_workqueues_begin - begin freezing workqueues
4764 *
4765 * Start freezing workqueues.  After this function returns, all freezable
4766 * workqueues will queue new works to their delayed_works list instead of
4767 * pool->worklist.
4768 *
4769 * CONTEXT:
4770 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4771 */
4772void freeze_workqueues_begin(void)
4773{
4774	struct workqueue_struct *wq;
4775	struct pool_workqueue *pwq;
4776
4777	mutex_lock(&wq_pool_mutex);
4778
4779	WARN_ON_ONCE(workqueue_freezing);
4780	workqueue_freezing = true;
4781
4782	list_for_each_entry(wq, &workqueues, list) {
4783		mutex_lock(&wq->mutex);
4784		for_each_pwq(pwq, wq)
4785			pwq_adjust_max_active(pwq);
4786		mutex_unlock(&wq->mutex);
 
 
 
 
 
 
 
 
 
 
 
 
4787	}
4788
4789	mutex_unlock(&wq_pool_mutex);
4790}
4791
4792/**
4793 * freeze_workqueues_busy - are freezable workqueues still busy?
4794 *
4795 * Check whether freezing is complete.  This function must be called
4796 * between freeze_workqueues_begin() and thaw_workqueues().
4797 *
4798 * CONTEXT:
4799 * Grabs and releases wq_pool_mutex.
4800 *
4801 * Return:
4802 * %true if some freezable workqueues are still busy.  %false if freezing
4803 * is complete.
4804 */
4805bool freeze_workqueues_busy(void)
4806{
 
4807	bool busy = false;
4808	struct workqueue_struct *wq;
4809	struct pool_workqueue *pwq;
4810
4811	mutex_lock(&wq_pool_mutex);
4812
4813	WARN_ON_ONCE(!workqueue_freezing);
4814
4815	list_for_each_entry(wq, &workqueues, list) {
4816		if (!(wq->flags & WQ_FREEZABLE))
4817			continue;
4818		/*
4819		 * nr_active is monotonically decreasing.  It's safe
4820		 * to peek without lock.
4821		 */
4822		rcu_read_lock_sched();
4823		for_each_pwq(pwq, wq) {
4824			WARN_ON_ONCE(pwq->nr_active < 0);
4825			if (pwq->nr_active) {
 
 
 
 
4826				busy = true;
4827				rcu_read_unlock_sched();
4828				goto out_unlock;
4829			}
4830		}
4831		rcu_read_unlock_sched();
4832	}
4833out_unlock:
4834	mutex_unlock(&wq_pool_mutex);
4835	return busy;
4836}
4837
4838/**
4839 * thaw_workqueues - thaw workqueues
4840 *
4841 * Thaw workqueues.  Normal queueing is restored and all collected
4842 * frozen works are transferred to their respective pool worklists.
4843 *
4844 * CONTEXT:
4845 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
4846 */
4847void thaw_workqueues(void)
4848{
4849	struct workqueue_struct *wq;
4850	struct pool_workqueue *pwq;
4851
4852	mutex_lock(&wq_pool_mutex);
4853
4854	if (!workqueue_freezing)
4855		goto out_unlock;
4856
4857	workqueue_freezing = false;
 
 
4858
4859	/* restore max_active and repopulate worklist */
4860	list_for_each_entry(wq, &workqueues, list) {
4861		mutex_lock(&wq->mutex);
4862		for_each_pwq(pwq, wq)
4863			pwq_adjust_max_active(pwq);
4864		mutex_unlock(&wq->mutex);
4865	}
4866
4867out_unlock:
4868	mutex_unlock(&wq_pool_mutex);
4869}
4870#endif /* CONFIG_FREEZER */
4871
4872static int workqueue_apply_unbound_cpumask(void)
4873{
4874	LIST_HEAD(ctxs);
4875	int ret = 0;
4876	struct workqueue_struct *wq;
4877	struct apply_wqattrs_ctx *ctx, *n;
4878
4879	lockdep_assert_held(&wq_pool_mutex);
 
4880
4881	list_for_each_entry(wq, &workqueues, list) {
4882		if (!(wq->flags & WQ_UNBOUND))
4883			continue;
4884		/* creating multiple pwqs breaks ordering guarantee */
4885		if (wq->flags & __WQ_ORDERED)
4886			continue;
4887
4888		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
4889		if (!ctx) {
4890			ret = -ENOMEM;
4891			break;
4892		}
4893
4894		list_add_tail(&ctx->list, &ctxs);
4895	}
4896
4897	list_for_each_entry_safe(ctx, n, &ctxs, list) {
4898		if (!ret)
4899			apply_wqattrs_commit(ctx);
4900		apply_wqattrs_cleanup(ctx);
4901	}
4902
4903	return ret;
4904}
4905
4906/**
4907 *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
4908 *  @cpumask: the cpumask to set
4909 *
4910 *  The low-level workqueues cpumask is a global cpumask that limits
4911 *  the affinity of all unbound workqueues.  This function check the @cpumask
4912 *  and apply it to all unbound workqueues and updates all pwqs of them.
4913 *
4914 *  Retun:	0	- Success
4915 *  		-EINVAL	- Invalid @cpumask
4916 *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
4917 */
4918int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
4919{
4920	int ret = -EINVAL;
4921	cpumask_var_t saved_cpumask;
4922
4923	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
4924		return -ENOMEM;
4925
4926	cpumask_and(cpumask, cpumask, cpu_possible_mask);
4927	if (!cpumask_empty(cpumask)) {
4928		apply_wqattrs_lock();
4929
4930		/* save the old wq_unbound_cpumask. */
4931		cpumask_copy(saved_cpumask, wq_unbound_cpumask);
4932
4933		/* update wq_unbound_cpumask at first and apply it to wqs. */
4934		cpumask_copy(wq_unbound_cpumask, cpumask);
4935		ret = workqueue_apply_unbound_cpumask();
4936
4937		/* restore the wq_unbound_cpumask when failed. */
4938		if (ret < 0)
4939			cpumask_copy(wq_unbound_cpumask, saved_cpumask);
4940
4941		apply_wqattrs_unlock();
4942	}
4943
4944	free_cpumask_var(saved_cpumask);
4945	return ret;
4946}
4947
4948#ifdef CONFIG_SYSFS
4949/*
4950 * Workqueues with WQ_SYSFS flag set is visible to userland via
4951 * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
4952 * following attributes.
4953 *
4954 *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
4955 *  max_active	RW int	: maximum number of in-flight work items
4956 *
4957 * Unbound workqueues have the following extra attributes.
4958 *
4959 *  id		RO int	: the associated pool ID
4960 *  nice	RW int	: nice value of the workers
4961 *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
4962 */
4963struct wq_device {
4964	struct workqueue_struct		*wq;
4965	struct device			dev;
4966};
4967
4968static struct workqueue_struct *dev_to_wq(struct device *dev)
4969{
4970	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
4971
4972	return wq_dev->wq;
4973}
4974
4975static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
4976			    char *buf)
4977{
4978	struct workqueue_struct *wq = dev_to_wq(dev);
4979
4980	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
4981}
4982static DEVICE_ATTR_RO(per_cpu);
4983
4984static ssize_t max_active_show(struct device *dev,
4985			       struct device_attribute *attr, char *buf)
4986{
4987	struct workqueue_struct *wq = dev_to_wq(dev);
4988
4989	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
4990}
4991
4992static ssize_t max_active_store(struct device *dev,
4993				struct device_attribute *attr, const char *buf,
4994				size_t count)
4995{
4996	struct workqueue_struct *wq = dev_to_wq(dev);
4997	int val;
4998
4999	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5000		return -EINVAL;
5001
5002	workqueue_set_max_active(wq, val);
5003	return count;
5004}
5005static DEVICE_ATTR_RW(max_active);
5006
5007static struct attribute *wq_sysfs_attrs[] = {
5008	&dev_attr_per_cpu.attr,
5009	&dev_attr_max_active.attr,
5010	NULL,
5011};
5012ATTRIBUTE_GROUPS(wq_sysfs);
5013
5014static ssize_t wq_pool_ids_show(struct device *dev,
5015				struct device_attribute *attr, char *buf)
5016{
5017	struct workqueue_struct *wq = dev_to_wq(dev);
5018	const char *delim = "";
5019	int node, written = 0;
5020
5021	rcu_read_lock_sched();
5022	for_each_node(node) {
5023		written += scnprintf(buf + written, PAGE_SIZE - written,
5024				     "%s%d:%d", delim, node,
5025				     unbound_pwq_by_node(wq, node)->pool->id);
5026		delim = " ";
5027	}
5028	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5029	rcu_read_unlock_sched();
5030
5031	return written;
5032}
5033
5034static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5035			    char *buf)
5036{
5037	struct workqueue_struct *wq = dev_to_wq(dev);
5038	int written;
5039
5040	mutex_lock(&wq->mutex);
5041	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5042	mutex_unlock(&wq->mutex);
5043
5044	return written;
5045}
5046
5047/* prepare workqueue_attrs for sysfs store operations */
5048static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5049{
5050	struct workqueue_attrs *attrs;
5051
5052	lockdep_assert_held(&wq_pool_mutex);
5053
5054	attrs = alloc_workqueue_attrs(GFP_KERNEL);
5055	if (!attrs)
5056		return NULL;
5057
5058	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5059	return attrs;
5060}
5061
5062static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5063			     const char *buf, size_t count)
5064{
5065	struct workqueue_struct *wq = dev_to_wq(dev);
5066	struct workqueue_attrs *attrs;
5067	int ret = -ENOMEM;
5068
5069	apply_wqattrs_lock();
5070
5071	attrs = wq_sysfs_prep_attrs(wq);
5072	if (!attrs)
5073		goto out_unlock;
5074
5075	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5076	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5077		ret = apply_workqueue_attrs_locked(wq, attrs);
5078	else
5079		ret = -EINVAL;
5080
5081out_unlock:
5082	apply_wqattrs_unlock();
5083	free_workqueue_attrs(attrs);
5084	return ret ?: count;
5085}
5086
5087static ssize_t wq_cpumask_show(struct device *dev,
5088			       struct device_attribute *attr, char *buf)
5089{
5090	struct workqueue_struct *wq = dev_to_wq(dev);
5091	int written;
5092
5093	mutex_lock(&wq->mutex);
5094	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5095			    cpumask_pr_args(wq->unbound_attrs->cpumask));
5096	mutex_unlock(&wq->mutex);
5097	return written;
5098}
5099
5100static ssize_t wq_cpumask_store(struct device *dev,
5101				struct device_attribute *attr,
5102				const char *buf, size_t count)
5103{
5104	struct workqueue_struct *wq = dev_to_wq(dev);
5105	struct workqueue_attrs *attrs;
5106	int ret = -ENOMEM;
5107
5108	apply_wqattrs_lock();
5109
5110	attrs = wq_sysfs_prep_attrs(wq);
5111	if (!attrs)
5112		goto out_unlock;
5113
5114	ret = cpumask_parse(buf, attrs->cpumask);
5115	if (!ret)
5116		ret = apply_workqueue_attrs_locked(wq, attrs);
5117
5118out_unlock:
5119	apply_wqattrs_unlock();
5120	free_workqueue_attrs(attrs);
5121	return ret ?: count;
5122}
5123
5124static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5125			    char *buf)
5126{
5127	struct workqueue_struct *wq = dev_to_wq(dev);
5128	int written;
5129
5130	mutex_lock(&wq->mutex);
5131	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5132			    !wq->unbound_attrs->no_numa);
5133	mutex_unlock(&wq->mutex);
5134
5135	return written;
5136}
5137
5138static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5139			     const char *buf, size_t count)
5140{
5141	struct workqueue_struct *wq = dev_to_wq(dev);
5142	struct workqueue_attrs *attrs;
5143	int v, ret = -ENOMEM;
5144
5145	apply_wqattrs_lock();
5146
5147	attrs = wq_sysfs_prep_attrs(wq);
5148	if (!attrs)
5149		goto out_unlock;
5150
5151	ret = -EINVAL;
5152	if (sscanf(buf, "%d", &v) == 1) {
5153		attrs->no_numa = !v;
5154		ret = apply_workqueue_attrs_locked(wq, attrs);
5155	}
5156
 
5157out_unlock:
5158	apply_wqattrs_unlock();
5159	free_workqueue_attrs(attrs);
5160	return ret ?: count;
5161}
5162
5163static struct device_attribute wq_sysfs_unbound_attrs[] = {
5164	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5165	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5166	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5167	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5168	__ATTR_NULL,
5169};
5170
5171static struct bus_type wq_subsys = {
5172	.name				= "workqueue",
5173	.dev_groups			= wq_sysfs_groups,
5174};
5175
5176static ssize_t wq_unbound_cpumask_show(struct device *dev,
5177		struct device_attribute *attr, char *buf)
5178{
5179	int written;
5180
5181	mutex_lock(&wq_pool_mutex);
5182	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5183			    cpumask_pr_args(wq_unbound_cpumask));
5184	mutex_unlock(&wq_pool_mutex);
5185
5186	return written;
5187}
5188
5189static ssize_t wq_unbound_cpumask_store(struct device *dev,
5190		struct device_attribute *attr, const char *buf, size_t count)
5191{
5192	cpumask_var_t cpumask;
5193	int ret;
5194
5195	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5196		return -ENOMEM;
5197
5198	ret = cpumask_parse(buf, cpumask);
5199	if (!ret)
5200		ret = workqueue_set_unbound_cpumask(cpumask);
5201
5202	free_cpumask_var(cpumask);
5203	return ret ? ret : count;
5204}
5205
5206static struct device_attribute wq_sysfs_cpumask_attr =
5207	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5208	       wq_unbound_cpumask_store);
5209
5210static int __init wq_sysfs_init(void)
5211{
5212	int err;
5213
5214	err = subsys_virtual_register(&wq_subsys, NULL);
5215	if (err)
5216		return err;
5217
5218	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5219}
5220core_initcall(wq_sysfs_init);
5221
5222static void wq_device_release(struct device *dev)
5223{
5224	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5225
5226	kfree(wq_dev);
5227}
5228
5229/**
5230 * workqueue_sysfs_register - make a workqueue visible in sysfs
5231 * @wq: the workqueue to register
5232 *
5233 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5234 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5235 * which is the preferred method.
5236 *
5237 * Workqueue user should use this function directly iff it wants to apply
5238 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5239 * apply_workqueue_attrs() may race against userland updating the
5240 * attributes.
5241 *
5242 * Return: 0 on success, -errno on failure.
5243 */
5244int workqueue_sysfs_register(struct workqueue_struct *wq)
5245{
5246	struct wq_device *wq_dev;
5247	int ret;
5248
5249	/*
5250	 * Adjusting max_active or creating new pwqs by applying
5251	 * attributes breaks ordering guarantee.  Disallow exposing ordered
5252	 * workqueues.
5253	 */
5254	if (WARN_ON(wq->flags & __WQ_ORDERED))
5255		return -EINVAL;
5256
5257	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5258	if (!wq_dev)
5259		return -ENOMEM;
5260
5261	wq_dev->wq = wq;
5262	wq_dev->dev.bus = &wq_subsys;
5263	wq_dev->dev.release = wq_device_release;
5264	dev_set_name(&wq_dev->dev, "%s", wq->name);
5265
5266	/*
5267	 * unbound_attrs are created separately.  Suppress uevent until
5268	 * everything is ready.
5269	 */
5270	dev_set_uevent_suppress(&wq_dev->dev, true);
5271
5272	ret = device_register(&wq_dev->dev);
5273	if (ret) {
5274		kfree(wq_dev);
5275		wq->wq_dev = NULL;
5276		return ret;
5277	}
5278
5279	if (wq->flags & WQ_UNBOUND) {
5280		struct device_attribute *attr;
5281
5282		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5283			ret = device_create_file(&wq_dev->dev, attr);
5284			if (ret) {
5285				device_unregister(&wq_dev->dev);
5286				wq->wq_dev = NULL;
5287				return ret;
5288			}
5289		}
5290	}
5291
5292	dev_set_uevent_suppress(&wq_dev->dev, false);
5293	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5294	return 0;
5295}
5296
5297/**
5298 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5299 * @wq: the workqueue to unregister
5300 *
5301 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5302 */
5303static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5304{
5305	struct wq_device *wq_dev = wq->wq_dev;
5306
5307	if (!wq->wq_dev)
5308		return;
5309
5310	wq->wq_dev = NULL;
5311	device_unregister(&wq_dev->dev);
5312}
5313#else	/* CONFIG_SYSFS */
5314static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
5315#endif	/* CONFIG_SYSFS */
5316
5317/*
5318 * Workqueue watchdog.
5319 *
5320 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5321 * flush dependency, a concurrency managed work item which stays RUNNING
5322 * indefinitely.  Workqueue stalls can be very difficult to debug as the
5323 * usual warning mechanisms don't trigger and internal workqueue state is
5324 * largely opaque.
5325 *
5326 * Workqueue watchdog monitors all worker pools periodically and dumps
5327 * state if some pools failed to make forward progress for a while where
5328 * forward progress is defined as the first item on ->worklist changing.
5329 *
5330 * This mechanism is controlled through the kernel parameter
5331 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5332 * corresponding sysfs parameter file.
5333 */
5334#ifdef CONFIG_WQ_WATCHDOG
5335
5336static void wq_watchdog_timer_fn(unsigned long data);
5337
5338static unsigned long wq_watchdog_thresh = 30;
5339static struct timer_list wq_watchdog_timer =
5340	TIMER_DEFERRED_INITIALIZER(wq_watchdog_timer_fn, 0, 0);
5341
5342static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5343static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5344
5345static void wq_watchdog_reset_touched(void)
5346{
5347	int cpu;
5348
5349	wq_watchdog_touched = jiffies;
5350	for_each_possible_cpu(cpu)
5351		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5352}
5353
5354static void wq_watchdog_timer_fn(unsigned long data)
5355{
5356	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5357	bool lockup_detected = false;
5358	struct worker_pool *pool;
5359	int pi;
5360
5361	if (!thresh)
5362		return;
5363
5364	rcu_read_lock();
5365
5366	for_each_pool(pool, pi) {
5367		unsigned long pool_ts, touched, ts;
5368
5369		if (list_empty(&pool->worklist))
5370			continue;
5371
5372		/* get the latest of pool and touched timestamps */
5373		pool_ts = READ_ONCE(pool->watchdog_ts);
5374		touched = READ_ONCE(wq_watchdog_touched);
5375
5376		if (time_after(pool_ts, touched))
5377			ts = pool_ts;
5378		else
5379			ts = touched;
5380
5381		if (pool->cpu >= 0) {
5382			unsigned long cpu_touched =
5383				READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5384						  pool->cpu));
5385			if (time_after(cpu_touched, ts))
5386				ts = cpu_touched;
5387		}
5388
5389		/* did we stall? */
5390		if (time_after(jiffies, ts + thresh)) {
5391			lockup_detected = true;
5392			pr_emerg("BUG: workqueue lockup - pool");
5393			pr_cont_pool_info(pool);
5394			pr_cont(" stuck for %us!\n",
5395				jiffies_to_msecs(jiffies - pool_ts) / 1000);
5396		}
5397	}
5398
5399	rcu_read_unlock();
5400
5401	if (lockup_detected)
5402		show_workqueue_state();
5403
5404	wq_watchdog_reset_touched();
5405	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5406}
5407
5408void wq_watchdog_touch(int cpu)
5409{
5410	if (cpu >= 0)
5411		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5412	else
5413		wq_watchdog_touched = jiffies;
5414}
5415
5416static void wq_watchdog_set_thresh(unsigned long thresh)
5417{
5418	wq_watchdog_thresh = 0;
5419	del_timer_sync(&wq_watchdog_timer);
5420
5421	if (thresh) {
5422		wq_watchdog_thresh = thresh;
5423		wq_watchdog_reset_touched();
5424		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5425	}
5426}
5427
5428static int wq_watchdog_param_set_thresh(const char *val,
5429					const struct kernel_param *kp)
5430{
5431	unsigned long thresh;
5432	int ret;
5433
5434	ret = kstrtoul(val, 0, &thresh);
5435	if (ret)
5436		return ret;
5437
5438	if (system_wq)
5439		wq_watchdog_set_thresh(thresh);
5440	else
5441		wq_watchdog_thresh = thresh;
5442
5443	return 0;
5444}
5445
5446static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5447	.set	= wq_watchdog_param_set_thresh,
5448	.get	= param_get_ulong,
5449};
5450
5451module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5452		0644);
5453
5454static void wq_watchdog_init(void)
5455{
5456	wq_watchdog_set_thresh(wq_watchdog_thresh);
5457}
5458
5459#else	/* CONFIG_WQ_WATCHDOG */
5460
5461static inline void wq_watchdog_init(void) { }
5462
5463#endif	/* CONFIG_WQ_WATCHDOG */
5464
5465static void __init wq_numa_init(void)
5466{
5467	cpumask_var_t *tbl;
5468	int node, cpu;
5469
5470	if (num_possible_nodes() <= 1)
5471		return;
5472
5473	if (wq_disable_numa) {
5474		pr_info("workqueue: NUMA affinity support disabled\n");
5475		return;
5476	}
5477
5478	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs(GFP_KERNEL);
5479	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5480
5481	/*
5482	 * We want masks of possible CPUs of each node which isn't readily
5483	 * available.  Build one from cpu_to_node() which should have been
5484	 * fully initialized by now.
5485	 */
5486	tbl = kzalloc(nr_node_ids * sizeof(tbl[0]), GFP_KERNEL);
5487	BUG_ON(!tbl);
5488
5489	for_each_node(node)
5490		BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5491				node_online(node) ? node : NUMA_NO_NODE));
5492
5493	for_each_possible_cpu(cpu) {
5494		node = cpu_to_node(cpu);
5495		if (WARN_ON(node == NUMA_NO_NODE)) {
5496			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5497			/* happens iff arch is bonkers, let's just proceed */
5498			return;
5499		}
5500		cpumask_set_cpu(cpu, tbl[node]);
5501	}
5502
5503	wq_numa_possible_cpumask = tbl;
5504	wq_numa_enabled = true;
5505}
 
5506
5507static int __init init_workqueues(void)
5508{
5509	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5510	int i, cpu;
5511
5512	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
 
5513
5514	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5515	cpumask_copy(wq_unbound_cpumask, cpu_possible_mask);
 
5516
5517	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
 
 
 
5518
5519	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
5520	hotcpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
 
5521
5522	wq_numa_init();
 
 
5523
5524	/* initialize CPU pools */
5525	for_each_possible_cpu(cpu) {
5526		struct worker_pool *pool;
5527
5528		i = 0;
5529		for_each_cpu_worker_pool(pool, cpu) {
5530			BUG_ON(init_worker_pool(pool));
5531			pool->cpu = cpu;
5532			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5533			pool->attrs->nice = std_nice[i++];
5534			pool->node = cpu_to_node(cpu);
5535
5536			/* alloc pool ID */
5537			mutex_lock(&wq_pool_mutex);
5538			BUG_ON(worker_pool_assign_id(pool));
5539			mutex_unlock(&wq_pool_mutex);
5540		}
5541	}
5542
5543	/* create the initial worker */
5544	for_each_online_cpu(cpu) {
5545		struct worker_pool *pool;
5546
5547		for_each_cpu_worker_pool(pool, cpu) {
5548			pool->flags &= ~POOL_DISASSOCIATED;
5549			BUG_ON(!create_worker(pool));
5550		}
5551	}
5552
5553	/* create default unbound and ordered wq attrs */
5554	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5555		struct workqueue_attrs *attrs;
5556
5557		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5558		attrs->nice = std_nice[i];
5559		unbound_std_wq_attrs[i] = attrs;
5560
5561		/*
5562		 * An ordered wq should have only one pwq as ordering is
5563		 * guaranteed by max_active which is enforced by pwqs.
5564		 * Turn off NUMA so that dfl_pwq is used for all nodes.
5565		 */
5566		BUG_ON(!(attrs = alloc_workqueue_attrs(GFP_KERNEL)));
5567		attrs->nice = std_nice[i];
5568		attrs->no_numa = true;
5569		ordered_wq_attrs[i] = attrs;
5570	}
5571
5572	system_wq = alloc_workqueue("events", 0, 0);
5573	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5574	system_long_wq = alloc_workqueue("events_long", 0, 0);
 
5575	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5576					    WQ_UNBOUND_MAX_ACTIVE);
5577	system_freezable_wq = alloc_workqueue("events_freezable",
5578					      WQ_FREEZABLE, 0);
5579	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5580					      WQ_POWER_EFFICIENT, 0);
5581	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5582					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5583					      0);
5584	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5585	       !system_unbound_wq || !system_freezable_wq ||
5586	       !system_power_efficient_wq ||
5587	       !system_freezable_power_efficient_wq);
5588
5589	wq_watchdog_init();
5590
5591	return 0;
5592}
5593early_initcall(init_workqueues);
v3.5.6
   1/*
   2 * kernel/workqueue.c - generic async execution with shared worker pool
   3 *
   4 * Copyright (C) 2002		Ingo Molnar
   5 *
   6 *   Derived from the taskqueue/keventd code by:
   7 *     David Woodhouse <dwmw2@infradead.org>
   8 *     Andrew Morton
   9 *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
  10 *     Theodore Ts'o <tytso@mit.edu>
  11 *
  12 * Made to use alloc_percpu by Christoph Lameter.
  13 *
  14 * Copyright (C) 2010		SUSE Linux Products GmbH
  15 * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
  16 *
  17 * This is the generic async execution mechanism.  Work items as are
  18 * executed in process context.  The worker pool is shared and
  19 * automatically managed.  There is one worker pool for each CPU and
  20 * one extra for works which are better served by workers which are
  21 * not bound to any specific CPU.
 
  22 *
  23 * Please read Documentation/workqueue.txt for details.
  24 */
  25
  26#include <linux/export.h>
  27#include <linux/kernel.h>
  28#include <linux/sched.h>
  29#include <linux/init.h>
  30#include <linux/signal.h>
  31#include <linux/completion.h>
  32#include <linux/workqueue.h>
  33#include <linux/slab.h>
  34#include <linux/cpu.h>
  35#include <linux/notifier.h>
  36#include <linux/kthread.h>
  37#include <linux/hardirq.h>
  38#include <linux/mempolicy.h>
  39#include <linux/freezer.h>
  40#include <linux/kallsyms.h>
  41#include <linux/debug_locks.h>
  42#include <linux/lockdep.h>
  43#include <linux/idr.h>
 
 
 
 
 
 
  44
  45#include "workqueue_sched.h"
  46
  47enum {
  48	/* global_cwq flags */
  49	GCWQ_MANAGE_WORKERS	= 1 << 0,	/* need to manage workers */
  50	GCWQ_MANAGING_WORKERS	= 1 << 1,	/* managing workers */
  51	GCWQ_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
  52	GCWQ_FREEZING		= 1 << 3,	/* freeze in progress */
  53	GCWQ_HIGHPRI_PENDING	= 1 << 4,	/* highpri works on queue */
 
 
 
 
 
 
 
 
 
 
 
  54
  55	/* worker flags */
  56	WORKER_STARTED		= 1 << 0,	/* started */
  57	WORKER_DIE		= 1 << 1,	/* die die die */
  58	WORKER_IDLE		= 1 << 2,	/* is idle */
  59	WORKER_PREP		= 1 << 3,	/* preparing to run works */
  60	WORKER_ROGUE		= 1 << 4,	/* not bound to any cpu */
  61	WORKER_REBIND		= 1 << 5,	/* mom is home, come back */
  62	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
  63	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
 
  64
  65	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
  66				  WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
  67
  68	/* gcwq->trustee_state */
  69	TRUSTEE_START		= 0,		/* start */
  70	TRUSTEE_IN_CHARGE	= 1,		/* trustee in charge of gcwq */
  71	TRUSTEE_BUTCHER		= 2,		/* butcher workers */
  72	TRUSTEE_RELEASE		= 3,		/* release workers */
  73	TRUSTEE_DONE		= 4,		/* trustee is done */
  74
 
  75	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
  76	BUSY_WORKER_HASH_SIZE	= 1 << BUSY_WORKER_HASH_ORDER,
  77	BUSY_WORKER_HASH_MASK	= BUSY_WORKER_HASH_SIZE - 1,
  78
  79	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
  80	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
  81
  82	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
  83						/* call for help after 10ms
  84						   (min two ticks) */
  85	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
  86	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
  87	TRUSTEE_COOLDOWN	= HZ / 10,	/* for trustee draining */
  88
  89	/*
  90	 * Rescue workers are used only on emergencies and shared by
  91	 * all cpus.  Give -20.
  92	 */
  93	RESCUER_NICE_LEVEL	= -20,
 
 
 
  94};
  95
  96/*
  97 * Structure fields follow one of the following exclusion rules.
  98 *
  99 * I: Modifiable by initialization/destruction paths and read-only for
 100 *    everyone else.
 101 *
 102 * P: Preemption protected.  Disabling preemption is enough and should
 103 *    only be modified and accessed from the local cpu.
 104 *
 105 * L: gcwq->lock protected.  Access with gcwq->lock held.
 
 
 
 
 
 106 *
 107 * X: During normal operation, modification requires gcwq->lock and
 108 *    should be done only from local cpu.  Either disabling preemption
 109 *    on local cpu or grabbing gcwq->lock is enough for read access.
 110 *    If GCWQ_DISASSOCIATED is set, it's identical to L.
 111 *
 112 * F: wq->flush_mutex protected.
 113 *
 114 * W: workqueue_lock protected.
 
 
 
 
 
 
 
 
 
 
 
 115 */
 116
 117struct global_cwq;
 118
 119/*
 120 * The poor guys doing the actual heavy lifting.  All on-duty workers
 121 * are either serving the manager role, on idle list or on busy hash.
 122 */
 123struct worker {
 124	/* on idle list while idle, on busy hash table while busy */
 125	union {
 126		struct list_head	entry;	/* L: while idle */
 127		struct hlist_node	hentry;	/* L: while busy */
 128	};
 129
 130	struct work_struct	*current_work;	/* L: work being processed */
 131	struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
 132	struct list_head	scheduled;	/* L: scheduled works */
 133	struct task_struct	*task;		/* I: worker task */
 134	struct global_cwq	*gcwq;		/* I: the associated gcwq */
 135	/* 64 bytes boundary on 64bit, 32 on 32bit */
 136	unsigned long		last_active;	/* L: last active timestamp */
 137	unsigned int		flags;		/* X: flags */
 138	int			id;		/* I: worker id */
 139	struct work_struct	rebind_work;	/* L: rebind worker to cpu */
 140};
 141
 142/*
 143 * Global per-cpu workqueue.  There's one and only one for each cpu
 144 * and all works are queued and processed here regardless of their
 145 * target workqueues.
 146 */
 147struct global_cwq {
 148	spinlock_t		lock;		/* the gcwq lock */
 149	struct list_head	worklist;	/* L: list of pending works */
 150	unsigned int		cpu;		/* I: the associated cpu */
 151	unsigned int		flags;		/* L: GCWQ_* flags */
 152
 153	int			nr_workers;	/* L: total number of workers */
 154	int			nr_idle;	/* L: currently idle ones */
 155
 156	/* workers are chained either in the idle_list or busy_hash */
 157	struct list_head	idle_list;	/* X: list of idle workers */
 158	struct hlist_head	busy_hash[BUSY_WORKER_HASH_SIZE];
 
 
 
 
 159						/* L: hash of busy workers */
 160
 161	struct timer_list	idle_timer;	/* L: worker idle timeout */
 162	struct timer_list	mayday_timer;	/* L: SOS timer for dworkers */
 
 
 
 
 
 
 
 
 
 
 163
 164	struct ida		worker_ida;	/* L: for worker IDs */
 
 
 
 
 
 165
 166	struct task_struct	*trustee;	/* L: for gcwq shutdown */
 167	unsigned int		trustee_state;	/* L: trustee state */
 168	wait_queue_head_t	trustee_wait;	/* trustee wait */
 169	struct worker		*first_idle;	/* L: first idle worker */
 
 170} ____cacheline_aligned_in_smp;
 171
 172/*
 173 * The per-CPU workqueue.  The lower WORK_STRUCT_FLAG_BITS of
 174 * work_struct->data are used for flags and thus cwqs need to be
 175 * aligned at two's power of the number of flag bits.
 
 176 */
 177struct cpu_workqueue_struct {
 178	struct global_cwq	*gcwq;		/* I: the associated gcwq */
 179	struct workqueue_struct *wq;		/* I: the owning workqueue */
 180	int			work_color;	/* L: current color */
 181	int			flush_color;	/* L: flushing color */
 
 182	int			nr_in_flight[WORK_NR_COLORS];
 183						/* L: nr of in_flight works */
 184	int			nr_active;	/* L: nr of active works */
 185	int			max_active;	/* L: max active works */
 186	struct list_head	delayed_works;	/* L: delayed works */
 187};
 
 
 
 
 
 
 
 
 
 
 
 188
 189/*
 190 * Structure used to wait for workqueue flush.
 191 */
 192struct wq_flusher {
 193	struct list_head	list;		/* F: list of flushers */
 194	int			flush_color;	/* F: flush color waiting for */
 195	struct completion	done;		/* flush completion */
 196};
 197
 198/*
 199 * All cpumasks are assumed to be always set on UP and thus can't be
 200 * used to determine whether there's something to be done.
 201 */
 202#ifdef CONFIG_SMP
 203typedef cpumask_var_t mayday_mask_t;
 204#define mayday_test_and_set_cpu(cpu, mask)	\
 205	cpumask_test_and_set_cpu((cpu), (mask))
 206#define mayday_clear_cpu(cpu, mask)		cpumask_clear_cpu((cpu), (mask))
 207#define for_each_mayday_cpu(cpu, mask)		for_each_cpu((cpu), (mask))
 208#define alloc_mayday_mask(maskp, gfp)		zalloc_cpumask_var((maskp), (gfp))
 209#define free_mayday_mask(mask)			free_cpumask_var((mask))
 210#else
 211typedef unsigned long mayday_mask_t;
 212#define mayday_test_and_set_cpu(cpu, mask)	test_and_set_bit(0, &(mask))
 213#define mayday_clear_cpu(cpu, mask)		clear_bit(0, &(mask))
 214#define for_each_mayday_cpu(cpu, mask)		if ((cpu) = 0, (mask))
 215#define alloc_mayday_mask(maskp, gfp)		true
 216#define free_mayday_mask(mask)			do { } while (0)
 217#endif
 218
 219/*
 220 * The externally visible workqueue abstraction is an array of
 221 * per-CPU workqueues:
 222 */
 223struct workqueue_struct {
 224	unsigned int		flags;		/* W: WQ_* flags */
 225	union {
 226		struct cpu_workqueue_struct __percpu	*pcpu;
 227		struct cpu_workqueue_struct		*single;
 228		unsigned long				v;
 229	} cpu_wq;				/* I: cwq's */
 230	struct list_head	list;		/* W: list of all workqueues */
 231
 232	struct mutex		flush_mutex;	/* protects wq flushing */
 233	int			work_color;	/* F: current work color */
 234	int			flush_color;	/* F: current flush color */
 235	atomic_t		nr_cwqs_to_flush; /* flush in progress */
 236	struct wq_flusher	*first_flusher;	/* F: first flusher */
 237	struct list_head	flusher_queue;	/* F: flush waiters */
 238	struct list_head	flusher_overflow; /* F: flush overflow list */
 239
 240	mayday_mask_t		mayday_mask;	/* cpus requesting rescue */
 241	struct worker		*rescuer;	/* I: rescue worker */
 242
 243	int			nr_drainers;	/* W: drain in progress */
 244	int			saved_max_active; /* W: saved cwq max_active */
 
 
 
 
 
 
 
 245#ifdef CONFIG_LOCKDEP
 246	struct lockdep_map	lockdep_map;
 247#endif
 248	char			name[];		/* I: workqueue name */
 
 
 
 
 
 
 
 
 
 
 
 
 249};
 250
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 251struct workqueue_struct *system_wq __read_mostly;
 
 
 
 252struct workqueue_struct *system_long_wq __read_mostly;
 253struct workqueue_struct *system_nrt_wq __read_mostly;
 254struct workqueue_struct *system_unbound_wq __read_mostly;
 
 255struct workqueue_struct *system_freezable_wq __read_mostly;
 256struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
 257EXPORT_SYMBOL_GPL(system_wq);
 258EXPORT_SYMBOL_GPL(system_long_wq);
 259EXPORT_SYMBOL_GPL(system_nrt_wq);
 260EXPORT_SYMBOL_GPL(system_unbound_wq);
 261EXPORT_SYMBOL_GPL(system_freezable_wq);
 262EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
 
 
 
 
 
 
 263
 264#define CREATE_TRACE_POINTS
 265#include <trace/events/workqueue.h>
 266
 267#define for_each_busy_worker(worker, i, pos, gcwq)			\
 268	for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)			\
 269		hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 270
 271static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
 272				  unsigned int sw)
 273{
 274	if (cpu < nr_cpu_ids) {
 275		if (sw & 1) {
 276			cpu = cpumask_next(cpu, mask);
 277			if (cpu < nr_cpu_ids)
 278				return cpu;
 279		}
 280		if (sw & 2)
 281			return WORK_CPU_UNBOUND;
 282	}
 283	return WORK_CPU_NONE;
 284}
 285
 286static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
 287				struct workqueue_struct *wq)
 288{
 289	return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
 290}
 291
 292/*
 293 * CPU iterators
 294 *
 295 * An extra gcwq is defined for an invalid cpu number
 296 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
 297 * specific CPU.  The following iterators are similar to
 298 * for_each_*_cpu() iterators but also considers the unbound gcwq.
 299 *
 300 * for_each_gcwq_cpu()		: possible CPUs + WORK_CPU_UNBOUND
 301 * for_each_online_gcwq_cpu()	: online CPUs + WORK_CPU_UNBOUND
 302 * for_each_cwq_cpu()		: possible CPUs for bound workqueues,
 303 *				  WORK_CPU_UNBOUND for unbound workqueues
 304 */
 305#define for_each_gcwq_cpu(cpu)						\
 306	for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3);		\
 307	     (cpu) < WORK_CPU_NONE;					\
 308	     (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
 309
 310#define for_each_online_gcwq_cpu(cpu)					\
 311	for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3);		\
 312	     (cpu) < WORK_CPU_NONE;					\
 313	     (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
 314
 315#define for_each_cwq_cpu(cpu, wq)					\
 316	for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq));	\
 317	     (cpu) < WORK_CPU_NONE;					\
 318	     (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
 319
 320#ifdef CONFIG_DEBUG_OBJECTS_WORK
 321
 322static struct debug_obj_descr work_debug_descr;
 323
 324static void *work_debug_hint(void *addr)
 325{
 326	return ((struct work_struct *) addr)->func;
 327}
 328
 329/*
 330 * fixup_init is called when:
 331 * - an active object is initialized
 332 */
 333static int work_fixup_init(void *addr, enum debug_obj_state state)
 334{
 335	struct work_struct *work = addr;
 336
 337	switch (state) {
 338	case ODEBUG_STATE_ACTIVE:
 339		cancel_work_sync(work);
 340		debug_object_init(work, &work_debug_descr);
 341		return 1;
 342	default:
 343		return 0;
 344	}
 345}
 346
 347/*
 348 * fixup_activate is called when:
 349 * - an active object is activated
 350 * - an unknown object is activated (might be a statically initialized object)
 351 */
 352static int work_fixup_activate(void *addr, enum debug_obj_state state)
 353{
 354	struct work_struct *work = addr;
 355
 356	switch (state) {
 357
 358	case ODEBUG_STATE_NOTAVAILABLE:
 359		/*
 360		 * This is not really a fixup. The work struct was
 361		 * statically initialized. We just make sure that it
 362		 * is tracked in the object tracker.
 363		 */
 364		if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
 365			debug_object_init(work, &work_debug_descr);
 366			debug_object_activate(work, &work_debug_descr);
 367			return 0;
 368		}
 369		WARN_ON_ONCE(1);
 370		return 0;
 371
 372	case ODEBUG_STATE_ACTIVE:
 373		WARN_ON(1);
 374
 375	default:
 376		return 0;
 377	}
 378}
 379
 380/*
 381 * fixup_free is called when:
 382 * - an active object is freed
 383 */
 384static int work_fixup_free(void *addr, enum debug_obj_state state)
 385{
 386	struct work_struct *work = addr;
 387
 388	switch (state) {
 389	case ODEBUG_STATE_ACTIVE:
 390		cancel_work_sync(work);
 391		debug_object_free(work, &work_debug_descr);
 392		return 1;
 393	default:
 394		return 0;
 395	}
 396}
 397
 398static struct debug_obj_descr work_debug_descr = {
 399	.name		= "work_struct",
 400	.debug_hint	= work_debug_hint,
 401	.fixup_init	= work_fixup_init,
 402	.fixup_activate	= work_fixup_activate,
 403	.fixup_free	= work_fixup_free,
 404};
 405
 406static inline void debug_work_activate(struct work_struct *work)
 407{
 408	debug_object_activate(work, &work_debug_descr);
 409}
 410
 411static inline void debug_work_deactivate(struct work_struct *work)
 412{
 413	debug_object_deactivate(work, &work_debug_descr);
 414}
 415
 416void __init_work(struct work_struct *work, int onstack)
 417{
 418	if (onstack)
 419		debug_object_init_on_stack(work, &work_debug_descr);
 420	else
 421		debug_object_init(work, &work_debug_descr);
 422}
 423EXPORT_SYMBOL_GPL(__init_work);
 424
 425void destroy_work_on_stack(struct work_struct *work)
 426{
 427	debug_object_free(work, &work_debug_descr);
 428}
 429EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 430
 
 
 
 
 
 
 
 431#else
 432static inline void debug_work_activate(struct work_struct *work) { }
 433static inline void debug_work_deactivate(struct work_struct *work) { }
 434#endif
 435
 436/* Serializes the accesses to the list of workqueues. */
 437static DEFINE_SPINLOCK(workqueue_lock);
 438static LIST_HEAD(workqueues);
 439static bool workqueue_freezing;		/* W: have wqs started freezing? */
 440
 441/*
 442 * The almighty global cpu workqueues.  nr_running is the only field
 443 * which is expected to be used frequently by other cpus via
 444 * try_to_wake_up().  Put it in a separate cacheline.
 445 */
 446static DEFINE_PER_CPU(struct global_cwq, global_cwq);
 447static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
 
 448
 449/*
 450 * Global cpu workqueue and nr_running counter for unbound gcwq.  The
 451 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
 452 * workers have WORKER_UNBOUND set.
 453 */
 454static struct global_cwq unbound_global_cwq;
 455static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0);	/* always 0 */
 456
 457static int worker_thread(void *__worker);
 
 
 
 
 
 
 
 458
 459static struct global_cwq *get_gcwq(unsigned int cpu)
 
 
 
 
 
 
 
 
 
 
 
 
 
 460{
 461	if (cpu != WORK_CPU_UNBOUND)
 462		return &per_cpu(global_cwq, cpu);
 463	else
 464		return &unbound_global_cwq;
 465}
 466
 467static atomic_t *get_gcwq_nr_running(unsigned int cpu)
 468{
 469	if (cpu != WORK_CPU_UNBOUND)
 470		return &per_cpu(gcwq_nr_running, cpu);
 471	else
 472		return &unbound_gcwq_nr_running;
 473}
 
 474
 475static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
 476					    struct workqueue_struct *wq)
 477{
 478	if (!(wq->flags & WQ_UNBOUND)) {
 479		if (likely(cpu < nr_cpu_ids))
 480			return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
 481	} else if (likely(cpu == WORK_CPU_UNBOUND))
 482		return wq->cpu_wq.single;
 483	return NULL;
 484}
 485
 486static unsigned int work_color_to_flags(int color)
 487{
 488	return color << WORK_STRUCT_COLOR_SHIFT;
 489}
 490
 491static int get_work_color(struct work_struct *work)
 492{
 493	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 494		((1 << WORK_STRUCT_COLOR_BITS) - 1);
 495}
 496
 497static int work_next_color(int color)
 498{
 499	return (color + 1) % WORK_NR_COLORS;
 500}
 501
 502/*
 503 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
 504 * work is on queue.  Once execution starts, WORK_STRUCT_CWQ is
 505 * cleared and the work data contains the cpu number it was last on.
 506 *
 507 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
 508 * cwq, cpu or clear work->data.  These functions should only be
 509 * called while the work is owned - ie. while the PENDING bit is set.
 510 *
 511 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
 512 * corresponding to a work.  gcwq is available once the work has been
 513 * queued anywhere after initialization.  cwq is available only from
 514 * queueing until execution starts.
 
 
 
 
 
 
 515 */
 516static inline void set_work_data(struct work_struct *work, unsigned long data,
 517				 unsigned long flags)
 518{
 519	BUG_ON(!work_pending(work));
 520	atomic_long_set(&work->data, data | flags | work_static(work));
 521}
 522
 523static void set_work_cwq(struct work_struct *work,
 524			 struct cpu_workqueue_struct *cwq,
 525			 unsigned long extra_flags)
 526{
 527	set_work_data(work, (unsigned long)cwq,
 528		      WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
 
 
 
 
 
 
 
 529}
 530
 531static void set_work_cpu(struct work_struct *work, unsigned int cpu)
 
 532{
 533	set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 534}
 535
 536static void clear_work_data(struct work_struct *work)
 537{
 538	set_work_data(work, WORK_STRUCT_NO_CPU, 0);
 
 539}
 540
 541static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
 542{
 543	unsigned long data = atomic_long_read(&work->data);
 544
 545	if (data & WORK_STRUCT_CWQ)
 546		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 547	else
 548		return NULL;
 549}
 550
 551static struct global_cwq *get_work_gcwq(struct work_struct *work)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 552{
 553	unsigned long data = atomic_long_read(&work->data);
 554	unsigned int cpu;
 555
 556	if (data & WORK_STRUCT_CWQ)
 557		return ((struct cpu_workqueue_struct *)
 558			(data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
 559
 560	cpu = data >> WORK_STRUCT_FLAG_BITS;
 561	if (cpu == WORK_CPU_NONE)
 
 
 
 
 562		return NULL;
 563
 564	BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
 565	return get_gcwq(cpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 566}
 567
 568/*
 569 * Policy functions.  These define the policies on how the global
 570 * worker pool is managed.  Unless noted otherwise, these functions
 571 * assume that they're being called with gcwq->lock held.
 572 */
 573
 574static bool __need_more_worker(struct global_cwq *gcwq)
 575{
 576	return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
 577		gcwq->flags & GCWQ_HIGHPRI_PENDING;
 578}
 579
 580/*
 581 * Need to wake up a worker?  Called from anything but currently
 582 * running workers.
 
 
 
 
 583 */
 584static bool need_more_worker(struct global_cwq *gcwq)
 585{
 586	return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
 587}
 588
 589/* Can I start working?  Called from busy but !running workers. */
 590static bool may_start_working(struct global_cwq *gcwq)
 591{
 592	return gcwq->nr_idle;
 593}
 594
 595/* Do I need to keep working?  Called from currently running workers. */
 596static bool keep_working(struct global_cwq *gcwq)
 597{
 598	atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
 599
 600	return !list_empty(&gcwq->worklist) &&
 601		(atomic_read(nr_running) <= 1 ||
 602		 gcwq->flags & GCWQ_HIGHPRI_PENDING);
 603}
 604
 605/* Do we need a new worker?  Called from manager. */
 606static bool need_to_create_worker(struct global_cwq *gcwq)
 607{
 608	return need_more_worker(gcwq) && !may_start_working(gcwq);
 609}
 610
 611/* Do I need to be the manager? */
 612static bool need_to_manage_workers(struct global_cwq *gcwq)
 613{
 614	return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
 615}
 616
 617/* Do we have too many workers and should some go away? */
 618static bool too_many_workers(struct global_cwq *gcwq)
 619{
 620	bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
 621	int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
 622	int nr_busy = gcwq->nr_workers - nr_idle;
 623
 624	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 625}
 626
 627/*
 628 * Wake up functions.
 629 */
 630
 631/* Return the first worker.  Safe with preemption disabled */
 632static struct worker *first_worker(struct global_cwq *gcwq)
 633{
 634	if (unlikely(list_empty(&gcwq->idle_list)))
 635		return NULL;
 636
 637	return list_first_entry(&gcwq->idle_list, struct worker, entry);
 638}
 639
 640/**
 641 * wake_up_worker - wake up an idle worker
 642 * @gcwq: gcwq to wake worker for
 643 *
 644 * Wake up the first idle worker of @gcwq.
 645 *
 646 * CONTEXT:
 647 * spin_lock_irq(gcwq->lock).
 648 */
 649static void wake_up_worker(struct global_cwq *gcwq)
 650{
 651	struct worker *worker = first_worker(gcwq);
 652
 653	if (likely(worker))
 654		wake_up_process(worker->task);
 655}
 656
 657/**
 658 * wq_worker_waking_up - a worker is waking up
 659 * @task: task waking up
 660 * @cpu: CPU @task is waking up to
 661 *
 662 * This function is called during try_to_wake_up() when a worker is
 663 * being awoken.
 664 *
 665 * CONTEXT:
 666 * spin_lock_irq(rq->lock)
 667 */
 668void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
 669{
 670	struct worker *worker = kthread_data(task);
 671
 672	if (!(worker->flags & WORKER_NOT_RUNNING))
 673		atomic_inc(get_gcwq_nr_running(cpu));
 
 
 674}
 675
 676/**
 677 * wq_worker_sleeping - a worker is going to sleep
 678 * @task: task going to sleep
 679 * @cpu: CPU in question, must be the current CPU number
 680 *
 681 * This function is called during schedule() when a busy worker is
 682 * going to sleep.  Worker on the same cpu can be woken up by
 683 * returning pointer to its task.
 684 *
 685 * CONTEXT:
 686 * spin_lock_irq(rq->lock)
 687 *
 688 * RETURNS:
 689 * Worker task on @cpu to wake up, %NULL if none.
 690 */
 691struct task_struct *wq_worker_sleeping(struct task_struct *task,
 692				       unsigned int cpu)
 693{
 694	struct worker *worker = kthread_data(task), *to_wakeup = NULL;
 695	struct global_cwq *gcwq = get_gcwq(cpu);
 696	atomic_t *nr_running = get_gcwq_nr_running(cpu);
 697
 
 
 
 
 
 698	if (worker->flags & WORKER_NOT_RUNNING)
 699		return NULL;
 700
 
 
 701	/* this can only happen on the local cpu */
 702	BUG_ON(cpu != raw_smp_processor_id());
 
 703
 704	/*
 705	 * The counterpart of the following dec_and_test, implied mb,
 706	 * worklist not empty test sequence is in insert_work().
 707	 * Please read comment there.
 708	 *
 709	 * NOT_RUNNING is clear.  This means that trustee is not in
 710	 * charge and we're running on the local cpu w/ rq lock held
 711	 * and preemption disabled, which in turn means that none else
 712	 * could be manipulating idle_list, so dereferencing idle_list
 713	 * without gcwq lock is safe.
 714	 */
 715	if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
 716		to_wakeup = first_worker(gcwq);
 
 717	return to_wakeup ? to_wakeup->task : NULL;
 718}
 719
 720/**
 721 * worker_set_flags - set worker flags and adjust nr_running accordingly
 722 * @worker: self
 723 * @flags: flags to set
 724 * @wakeup: wakeup an idle worker if necessary
 725 *
 726 * Set @flags in @worker->flags and adjust nr_running accordingly.  If
 727 * nr_running becomes zero and @wakeup is %true, an idle worker is
 728 * woken up.
 729 *
 730 * CONTEXT:
 731 * spin_lock_irq(gcwq->lock)
 732 */
 733static inline void worker_set_flags(struct worker *worker, unsigned int flags,
 734				    bool wakeup)
 735{
 736	struct global_cwq *gcwq = worker->gcwq;
 737
 738	WARN_ON_ONCE(worker->task != current);
 739
 740	/*
 741	 * If transitioning into NOT_RUNNING, adjust nr_running and
 742	 * wake up an idle worker as necessary if requested by
 743	 * @wakeup.
 744	 */
 745	if ((flags & WORKER_NOT_RUNNING) &&
 746	    !(worker->flags & WORKER_NOT_RUNNING)) {
 747		atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
 748
 749		if (wakeup) {
 750			if (atomic_dec_and_test(nr_running) &&
 751			    !list_empty(&gcwq->worklist))
 752				wake_up_worker(gcwq);
 753		} else
 754			atomic_dec(nr_running);
 755	}
 756
 757	worker->flags |= flags;
 758}
 759
 760/**
 761 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 762 * @worker: self
 763 * @flags: flags to clear
 764 *
 765 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 766 *
 767 * CONTEXT:
 768 * spin_lock_irq(gcwq->lock)
 769 */
 770static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 771{
 772	struct global_cwq *gcwq = worker->gcwq;
 773	unsigned int oflags = worker->flags;
 774
 775	WARN_ON_ONCE(worker->task != current);
 776
 777	worker->flags &= ~flags;
 778
 779	/*
 780	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 781	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 782	 * of multiple flags, not a single flag.
 783	 */
 784	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 785		if (!(worker->flags & WORKER_NOT_RUNNING))
 786			atomic_inc(get_gcwq_nr_running(gcwq->cpu));
 787}
 788
 789/**
 790 * busy_worker_head - return the busy hash head for a work
 791 * @gcwq: gcwq of interest
 792 * @work: work to be hashed
 793 *
 794 * Return hash head of @gcwq for @work.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 795 *
 796 * CONTEXT:
 797 * spin_lock_irq(gcwq->lock).
 798 *
 799 * RETURNS:
 800 * Pointer to the hash head.
 
 801 */
 802static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
 803					   struct work_struct *work)
 804{
 805	const int base_shift = ilog2(sizeof(struct work_struct));
 806	unsigned long v = (unsigned long)work;
 807
 808	/* simple shift and fold hash, do we need something better? */
 809	v >>= base_shift;
 810	v += v >> BUSY_WORKER_HASH_ORDER;
 811	v &= BUSY_WORKER_HASH_MASK;
 
 812
 813	return &gcwq->busy_hash[v];
 814}
 815
 816/**
 817 * __find_worker_executing_work - find worker which is executing a work
 818 * @gcwq: gcwq of interest
 819 * @bwh: hash head as returned by busy_worker_head()
 820 * @work: work to find worker for
 
 
 
 
 821 *
 822 * Find a worker which is executing @work on @gcwq.  @bwh should be
 823 * the hash head obtained by calling busy_worker_head() with the same
 824 * work.
 825 *
 826 * CONTEXT:
 827 * spin_lock_irq(gcwq->lock).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 828 *
 829 * RETURNS:
 830 * Pointer to worker which is executing @work if found, NULL
 831 * otherwise.
 832 */
 833static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
 834						   struct hlist_head *bwh,
 835						   struct work_struct *work)
 836{
 837	struct worker *worker;
 838	struct hlist_node *tmp;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 839
 840	hlist_for_each_entry(worker, tmp, bwh, hentry)
 841		if (worker->current_work == work)
 842			return worker;
 843	return NULL;
 844}
 845
 846/**
 847 * find_worker_executing_work - find worker which is executing a work
 848 * @gcwq: gcwq of interest
 849 * @work: work to find worker for
 850 *
 851 * Find a worker which is executing @work on @gcwq.  This function is
 852 * identical to __find_worker_executing_work() except that this
 853 * function calculates @bwh itself.
 854 *
 855 * CONTEXT:
 856 * spin_lock_irq(gcwq->lock).
 857 *
 858 * RETURNS:
 859 * Pointer to worker which is executing @work if found, NULL
 860 * otherwise.
 861 */
 862static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
 863						 struct work_struct *work)
 864{
 865	return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
 866					    work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 867}
 868
 869/**
 870 * gcwq_determine_ins_pos - find insertion position
 871 * @gcwq: gcwq of interest
 872 * @cwq: cwq a work is being queued for
 
 873 *
 874 * A work for @cwq is about to be queued on @gcwq, determine insertion
 875 * position for the work.  If @cwq is for HIGHPRI wq, the work is
 876 * queued at the head of the queue but in FIFO order with respect to
 877 * other HIGHPRI works; otherwise, at the end of the queue.  This
 878 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
 879 * there are HIGHPRI works pending.
 880 *
 881 * CONTEXT:
 882 * spin_lock_irq(gcwq->lock).
 
 
 
 
 
 
 
 
 
 
 
 
 
 883 *
 884 * RETURNS:
 885 * Pointer to inserstion position.
 886 */
 887static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
 888					       struct cpu_workqueue_struct *cwq)
 889{
 890	struct work_struct *twork;
 
 891
 892	if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
 893		return &gcwq->worklist;
 894
 895	list_for_each_entry(twork, &gcwq->worklist, entry) {
 896		struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
 
 897
 898		if (!(tcwq->wq->flags & WQ_HIGHPRI))
 899			break;
 
 
 
 
 
 900	}
 901
 902	gcwq->flags |= GCWQ_HIGHPRI_PENDING;
 903	return &twork->entry;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 904}
 905
 906/**
 907 * insert_work - insert a work into gcwq
 908 * @cwq: cwq @work belongs to
 909 * @work: work to insert
 910 * @head: insertion point
 911 * @extra_flags: extra WORK_STRUCT_* flags to set
 912 *
 913 * Insert @work which belongs to @cwq into @gcwq after @head.
 914 * @extra_flags is or'd to work_struct flags.
 915 *
 916 * CONTEXT:
 917 * spin_lock_irq(gcwq->lock).
 918 */
 919static void insert_work(struct cpu_workqueue_struct *cwq,
 920			struct work_struct *work, struct list_head *head,
 921			unsigned int extra_flags)
 922{
 923	struct global_cwq *gcwq = cwq->gcwq;
 924
 925	/* we own @work, set data and link */
 926	set_work_cwq(work, cwq, extra_flags);
 927
 928	/*
 929	 * Ensure that we get the right work->data if we see the
 930	 * result of list_add() below, see try_to_grab_pending().
 931	 */
 932	smp_wmb();
 933
 934	list_add_tail(&work->entry, head);
 
 935
 936	/*
 937	 * Ensure either worker_sched_deactivated() sees the above
 938	 * list_add_tail() or we see zero nr_running to avoid workers
 939	 * lying around lazily while there are works to be processed.
 940	 */
 941	smp_mb();
 942
 943	if (__need_more_worker(gcwq))
 944		wake_up_worker(gcwq);
 945}
 946
 947/*
 948 * Test whether @work is being queued from another work executing on the
 949 * same workqueue.  This is rather expensive and should only be used from
 950 * cold paths.
 951 */
 952static bool is_chained_work(struct workqueue_struct *wq)
 953{
 954	unsigned long flags;
 955	unsigned int cpu;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 956
 957	for_each_gcwq_cpu(cpu) {
 958		struct global_cwq *gcwq = get_gcwq(cpu);
 959		struct worker *worker;
 960		struct hlist_node *pos;
 961		int i;
 962
 963		spin_lock_irqsave(&gcwq->lock, flags);
 964		for_each_busy_worker(worker, i, pos, gcwq) {
 965			if (worker->task != current)
 966				continue;
 967			spin_unlock_irqrestore(&gcwq->lock, flags);
 968			/*
 969			 * I'm @worker, no locking necessary.  See if @work
 970			 * is headed to the same workqueue.
 971			 */
 972			return worker->current_cwq->wq == wq;
 973		}
 974		spin_unlock_irqrestore(&gcwq->lock, flags);
 975	}
 976	return false;
 
 
 977}
 978
 979static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
 980			 struct work_struct *work)
 981{
 982	struct global_cwq *gcwq;
 983	struct cpu_workqueue_struct *cwq;
 984	struct list_head *worklist;
 985	unsigned int work_flags;
 986	unsigned long flags;
 
 
 
 
 
 
 
 
 987
 988	debug_work_activate(work);
 989
 990	/* if dying, only works from the same workqueue are allowed */
 991	if (unlikely(wq->flags & WQ_DRAINING) &&
 992	    WARN_ON_ONCE(!is_chained_work(wq)))
 993		return;
 
 
 
 994
 995	/* determine gcwq to use */
 996	if (!(wq->flags & WQ_UNBOUND)) {
 997		struct global_cwq *last_gcwq;
 
 
 998
 999		if (unlikely(cpu == WORK_CPU_UNBOUND))
1000			cpu = raw_smp_processor_id();
 
 
 
 
 
 
1001
1002		/*
1003		 * It's multi cpu.  If @wq is non-reentrant and @work
1004		 * was previously on a different cpu, it might still
1005		 * be running there, in which case the work needs to
1006		 * be queued on that cpu to guarantee non-reentrance.
1007		 */
1008		gcwq = get_gcwq(cpu);
1009		if (wq->flags & WQ_NON_REENTRANT &&
1010		    (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1011			struct worker *worker;
1012
1013			spin_lock_irqsave(&last_gcwq->lock, flags);
1014
1015			worker = find_worker_executing_work(last_gcwq, work);
 
 
 
 
 
 
 
 
 
1016
1017			if (worker && worker->current_cwq->wq == wq)
1018				gcwq = last_gcwq;
1019			else {
1020				/* meh... not running there, queue here */
1021				spin_unlock_irqrestore(&last_gcwq->lock, flags);
1022				spin_lock_irqsave(&gcwq->lock, flags);
1023			}
1024		} else
1025			spin_lock_irqsave(&gcwq->lock, flags);
1026	} else {
1027		gcwq = get_gcwq(WORK_CPU_UNBOUND);
1028		spin_lock_irqsave(&gcwq->lock, flags);
 
 
 
 
 
1029	}
1030
1031	/* gcwq determined, get cwq and queue */
1032	cwq = get_cwq(gcwq->cpu, wq);
1033	trace_workqueue_queue_work(cpu, cwq, work);
1034
1035	if (WARN_ON(!list_empty(&work->entry))) {
1036		spin_unlock_irqrestore(&gcwq->lock, flags);
1037		return;
1038	}
1039
1040	cwq->nr_in_flight[cwq->work_color]++;
1041	work_flags = work_color_to_flags(cwq->work_color);
1042
1043	if (likely(cwq->nr_active < cwq->max_active)) {
1044		trace_workqueue_activate_work(work);
1045		cwq->nr_active++;
1046		worklist = gcwq_determine_ins_pos(gcwq, cwq);
 
 
1047	} else {
1048		work_flags |= WORK_STRUCT_DELAYED;
1049		worklist = &cwq->delayed_works;
1050	}
1051
1052	insert_work(cwq, work, worklist, work_flags);
1053
1054	spin_unlock_irqrestore(&gcwq->lock, flags);
1055}
1056
1057/**
1058 * queue_work - queue work on a workqueue
1059 * @wq: workqueue to use
1060 * @work: work to queue
1061 *
1062 * Returns 0 if @work was already on a queue, non-zero otherwise.
1063 *
1064 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1065 * it can be processed by another CPU.
1066 */
1067int queue_work(struct workqueue_struct *wq, struct work_struct *work)
1068{
1069	int ret;
1070
1071	ret = queue_work_on(get_cpu(), wq, work);
1072	put_cpu();
1073
1074	return ret;
1075}
1076EXPORT_SYMBOL_GPL(queue_work);
1077
1078/**
1079 * queue_work_on - queue work on specific cpu
1080 * @cpu: CPU number to execute work on
1081 * @wq: workqueue to use
1082 * @work: work to queue
1083 *
1084 * Returns 0 if @work was already on a queue, non-zero otherwise.
1085 *
1086 * We queue the work to a specific CPU, the caller must ensure it
1087 * can't go away.
 
 
1088 */
1089int
1090queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1091{
1092	int ret = 0;
 
 
 
1093
1094	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1095		__queue_work(cpu, wq, work);
1096		ret = 1;
1097	}
 
 
1098	return ret;
1099}
1100EXPORT_SYMBOL_GPL(queue_work_on);
1101
1102static void delayed_work_timer_fn(unsigned long __data)
1103{
1104	struct delayed_work *dwork = (struct delayed_work *)__data;
1105	struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1106
1107	__queue_work(smp_processor_id(), cwq->wq, &dwork->work);
 
1108}
 
1109
1110/**
1111 * queue_delayed_work - queue work on a workqueue after delay
1112 * @wq: workqueue to use
1113 * @dwork: delayable work to queue
1114 * @delay: number of jiffies to wait before queueing
1115 *
1116 * Returns 0 if @work was already on a queue, non-zero otherwise.
1117 */
1118int queue_delayed_work(struct workqueue_struct *wq,
1119			struct delayed_work *dwork, unsigned long delay)
1120{
1121	if (delay == 0)
1122		return queue_work(wq, &dwork->work);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1123
1124	return queue_delayed_work_on(-1, wq, dwork, delay);
 
 
 
 
 
 
 
1125}
1126EXPORT_SYMBOL_GPL(queue_delayed_work);
1127
1128/**
1129 * queue_delayed_work_on - queue work on specific CPU after delay
1130 * @cpu: CPU number to execute work on
1131 * @wq: workqueue to use
1132 * @dwork: work to queue
1133 * @delay: number of jiffies to wait before queueing
1134 *
1135 * Returns 0 if @work was already on a queue, non-zero otherwise.
 
 
1136 */
1137int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1138			struct delayed_work *dwork, unsigned long delay)
1139{
1140	int ret = 0;
1141	struct timer_list *timer = &dwork->timer;
1142	struct work_struct *work = &dwork->work;
 
 
 
 
 
1143
1144	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1145		unsigned int lcpu;
 
 
1146
1147		BUG_ON(timer_pending(timer));
1148		BUG_ON(!list_empty(&work->entry));
 
 
1149
1150		timer_stats_timer_set_start_info(&dwork->timer);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1151
1152		/*
1153		 * This stores cwq for the moment, for the timer_fn.
1154		 * Note that the work's gcwq is preserved to allow
1155		 * reentrance detection for delayed works.
1156		 */
1157		if (!(wq->flags & WQ_UNBOUND)) {
1158			struct global_cwq *gcwq = get_work_gcwq(work);
1159
1160			if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1161				lcpu = gcwq->cpu;
1162			else
1163				lcpu = raw_smp_processor_id();
1164		} else
1165			lcpu = WORK_CPU_UNBOUND;
1166
1167		set_work_cwq(work, get_cwq(lcpu, wq), 0);
1168
1169		timer->expires = jiffies + delay;
1170		timer->data = (unsigned long)dwork;
1171		timer->function = delayed_work_timer_fn;
1172
1173		if (unlikely(cpu >= 0))
1174			add_timer_on(timer, cpu);
1175		else
1176			add_timer(timer);
1177		ret = 1;
1178	}
1179	return ret;
1180}
1181EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1182
1183/**
1184 * worker_enter_idle - enter idle state
1185 * @worker: worker which is entering idle state
1186 *
1187 * @worker is entering idle state.  Update stats and idle timer if
1188 * necessary.
1189 *
1190 * LOCKING:
1191 * spin_lock_irq(gcwq->lock).
1192 */
1193static void worker_enter_idle(struct worker *worker)
1194{
1195	struct global_cwq *gcwq = worker->gcwq;
1196
1197	BUG_ON(worker->flags & WORKER_IDLE);
1198	BUG_ON(!list_empty(&worker->entry) &&
1199	       (worker->hentry.next || worker->hentry.pprev));
 
1200
1201	/* can't use worker_set_flags(), also called from start_worker() */
1202	worker->flags |= WORKER_IDLE;
1203	gcwq->nr_idle++;
1204	worker->last_active = jiffies;
1205
1206	/* idle_list is LIFO */
1207	list_add(&worker->entry, &gcwq->idle_list);
 
 
 
1208
1209	if (likely(!(worker->flags & WORKER_ROGUE))) {
1210		if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1211			mod_timer(&gcwq->idle_timer,
1212				  jiffies + IDLE_WORKER_TIMEOUT);
1213	} else
1214		wake_up_all(&gcwq->trustee_wait);
1215
1216	/*
1217	 * Sanity check nr_running.  Because trustee releases gcwq->lock
1218	 * between setting %WORKER_ROGUE and zapping nr_running, the
1219	 * warning may trigger spuriously.  Check iff trustee is idle.
1220	 */
1221	WARN_ON_ONCE(gcwq->trustee_state == TRUSTEE_DONE &&
1222		     gcwq->nr_workers == gcwq->nr_idle &&
1223		     atomic_read(get_gcwq_nr_running(gcwq->cpu)));
1224}
1225
1226/**
1227 * worker_leave_idle - leave idle state
1228 * @worker: worker which is leaving idle state
1229 *
1230 * @worker is leaving idle state.  Update stats.
1231 *
1232 * LOCKING:
1233 * spin_lock_irq(gcwq->lock).
1234 */
1235static void worker_leave_idle(struct worker *worker)
1236{
1237	struct global_cwq *gcwq = worker->gcwq;
1238
1239	BUG_ON(!(worker->flags & WORKER_IDLE));
 
1240	worker_clr_flags(worker, WORKER_IDLE);
1241	gcwq->nr_idle--;
1242	list_del_init(&worker->entry);
1243}
1244
1245/**
1246 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1247 * @worker: self
1248 *
1249 * Works which are scheduled while the cpu is online must at least be
1250 * scheduled to a worker which is bound to the cpu so that if they are
1251 * flushed from cpu callbacks while cpu is going down, they are
1252 * guaranteed to execute on the cpu.
1253 *
1254 * This function is to be used by rogue workers and rescuers to bind
1255 * themselves to the target cpu and may race with cpu going down or
1256 * coming online.  kthread_bind() can't be used because it may put the
1257 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1258 * verbatim as it's best effort and blocking and gcwq may be
1259 * [dis]associated in the meantime.
1260 *
1261 * This function tries set_cpus_allowed() and locks gcwq and verifies
1262 * the binding against GCWQ_DISASSOCIATED which is set during
1263 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1264 * idle state or fetches works without dropping lock, it can guarantee
1265 * the scheduling requirement described in the first paragraph.
1266 *
1267 * CONTEXT:
1268 * Might sleep.  Called without any lock but returns with gcwq->lock
1269 * held.
1270 *
1271 * RETURNS:
1272 * %true if the associated gcwq is online (@worker is successfully
1273 * bound), %false if offline.
1274 */
1275static bool worker_maybe_bind_and_lock(struct worker *worker)
1276__acquires(&gcwq->lock)
1277{
1278	struct global_cwq *gcwq = worker->gcwq;
1279	struct task_struct *task = worker->task;
1280
1281	while (true) {
1282		/*
1283		 * The following call may fail, succeed or succeed
1284		 * without actually migrating the task to the cpu if
1285		 * it races with cpu hotunplug operation.  Verify
1286		 * against GCWQ_DISASSOCIATED.
1287		 */
1288		if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1289			set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
1290
1291		spin_lock_irq(&gcwq->lock);
1292		if (gcwq->flags & GCWQ_DISASSOCIATED)
1293			return false;
1294		if (task_cpu(task) == gcwq->cpu &&
1295		    cpumask_equal(&current->cpus_allowed,
1296				  get_cpu_mask(gcwq->cpu)))
1297			return true;
1298		spin_unlock_irq(&gcwq->lock);
1299
1300		/*
1301		 * We've raced with CPU hot[un]plug.  Give it a breather
1302		 * and retry migration.  cond_resched() is required here;
1303		 * otherwise, we might deadlock against cpu_stop trying to
1304		 * bring down the CPU on non-preemptive kernel.
1305		 */
1306		cpu_relax();
1307		cond_resched();
1308	}
 
1309}
1310
1311/*
1312 * Function for worker->rebind_work used to rebind rogue busy workers
1313 * to the associated cpu which is coming back online.  This is
1314 * scheduled by cpu up but can race with other cpu hotplug operations
1315 * and may be executed twice without intervening cpu down.
 
 
 
1316 */
1317static void worker_rebind_fn(struct work_struct *work)
 
1318{
1319	struct worker *worker = container_of(work, struct worker, rebind_work);
1320	struct global_cwq *gcwq = worker->gcwq;
1321
1322	if (worker_maybe_bind_and_lock(worker))
1323		worker_clr_flags(worker, WORKER_REBIND);
 
 
 
1324
1325	spin_unlock_irq(&gcwq->lock);
 
 
 
 
 
 
 
 
 
 
1326}
1327
1328static struct worker *alloc_worker(void)
 
 
 
 
 
 
 
 
 
 
1329{
1330	struct worker *worker;
 
 
 
 
 
 
 
 
 
1331
1332	worker = kzalloc(sizeof(*worker), GFP_KERNEL);
1333	if (worker) {
1334		INIT_LIST_HEAD(&worker->entry);
1335		INIT_LIST_HEAD(&worker->scheduled);
1336		INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1337		/* on creation a worker is in !idle && prep state */
1338		worker->flags = WORKER_PREP;
1339	}
1340	return worker;
1341}
1342
1343/**
1344 * create_worker - create a new workqueue worker
1345 * @gcwq: gcwq the new worker will belong to
1346 * @bind: whether to set affinity to @cpu or not
1347 *
1348 * Create a new worker which is bound to @gcwq.  The returned worker
1349 * can be started by calling start_worker() or destroyed using
1350 * destroy_worker().
1351 *
1352 * CONTEXT:
1353 * Might sleep.  Does GFP_KERNEL allocations.
1354 *
1355 * RETURNS:
1356 * Pointer to the newly created worker.
1357 */
1358static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
1359{
1360	bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
1361	struct worker *worker = NULL;
1362	int id = -1;
 
1363
1364	spin_lock_irq(&gcwq->lock);
1365	while (ida_get_new(&gcwq->worker_ida, &id)) {
1366		spin_unlock_irq(&gcwq->lock);
1367		if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
1368			goto fail;
1369		spin_lock_irq(&gcwq->lock);
1370	}
1371	spin_unlock_irq(&gcwq->lock);
1372
1373	worker = alloc_worker();
1374	if (!worker)
1375		goto fail;
1376
1377	worker->gcwq = gcwq;
1378	worker->id = id;
1379
1380	if (!on_unbound_cpu)
1381		worker->task = kthread_create_on_node(worker_thread,
1382						      worker,
1383						      cpu_to_node(gcwq->cpu),
1384						      "kworker/%u:%d", gcwq->cpu, id);
1385	else
1386		worker->task = kthread_create(worker_thread, worker,
1387					      "kworker/u:%d", id);
 
 
1388	if (IS_ERR(worker->task))
1389		goto fail;
1390
1391	/*
1392	 * A rogue worker will become a regular one if CPU comes
1393	 * online later on.  Make sure every worker has
1394	 * PF_THREAD_BOUND set.
1395	 */
1396	if (bind && !on_unbound_cpu)
1397		kthread_bind(worker->task, gcwq->cpu);
1398	else {
1399		worker->task->flags |= PF_THREAD_BOUND;
1400		if (on_unbound_cpu)
1401			worker->flags |= WORKER_UNBOUND;
1402	}
1403
1404	return worker;
 
1405fail:
1406	if (id >= 0) {
1407		spin_lock_irq(&gcwq->lock);
1408		ida_remove(&gcwq->worker_ida, id);
1409		spin_unlock_irq(&gcwq->lock);
1410	}
1411	kfree(worker);
1412	return NULL;
1413}
1414
1415/**
1416 * start_worker - start a newly created worker
1417 * @worker: worker to start
1418 *
1419 * Make the gcwq aware of @worker and start it.
1420 *
1421 * CONTEXT:
1422 * spin_lock_irq(gcwq->lock).
1423 */
1424static void start_worker(struct worker *worker)
1425{
1426	worker->flags |= WORKER_STARTED;
1427	worker->gcwq->nr_workers++;
1428	worker_enter_idle(worker);
1429	wake_up_process(worker->task);
1430}
1431
1432/**
1433 * destroy_worker - destroy a workqueue worker
1434 * @worker: worker to be destroyed
1435 *
1436 * Destroy @worker and adjust @gcwq stats accordingly.
 
1437 *
1438 * CONTEXT:
1439 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1440 */
1441static void destroy_worker(struct worker *worker)
1442{
1443	struct global_cwq *gcwq = worker->gcwq;
1444	int id = worker->id;
 
1445
1446	/* sanity check frenzy */
1447	BUG_ON(worker->current_work);
1448	BUG_ON(!list_empty(&worker->scheduled));
 
 
1449
1450	if (worker->flags & WORKER_STARTED)
1451		gcwq->nr_workers--;
1452	if (worker->flags & WORKER_IDLE)
1453		gcwq->nr_idle--;
1454
1455	list_del_init(&worker->entry);
1456	worker->flags |= WORKER_DIE;
1457
1458	spin_unlock_irq(&gcwq->lock);
1459
1460	kthread_stop(worker->task);
1461	kfree(worker);
1462
1463	spin_lock_irq(&gcwq->lock);
1464	ida_remove(&gcwq->worker_ida, id);
1465}
1466
1467static void idle_worker_timeout(unsigned long __gcwq)
1468{
1469	struct global_cwq *gcwq = (void *)__gcwq;
1470
1471	spin_lock_irq(&gcwq->lock);
1472
1473	if (too_many_workers(gcwq)) {
1474		struct worker *worker;
1475		unsigned long expires;
1476
1477		/* idle_list is kept in LIFO order, check the last one */
1478		worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1479		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1480
1481		if (time_before(jiffies, expires))
1482			mod_timer(&gcwq->idle_timer, expires);
1483		else {
1484			/* it's been idle for too long, wake up manager */
1485			gcwq->flags |= GCWQ_MANAGE_WORKERS;
1486			wake_up_worker(gcwq);
1487		}
 
 
1488	}
1489
1490	spin_unlock_irq(&gcwq->lock);
1491}
1492
1493static bool send_mayday(struct work_struct *work)
1494{
1495	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1496	struct workqueue_struct *wq = cwq->wq;
1497	unsigned int cpu;
 
1498
1499	if (!(wq->flags & WQ_RESCUER))
1500		return false;
1501
1502	/* mayday mayday mayday */
1503	cpu = cwq->gcwq->cpu;
1504	/* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1505	if (cpu == WORK_CPU_UNBOUND)
1506		cpu = 0;
1507	if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
 
 
 
1508		wake_up_process(wq->rescuer->task);
1509	return true;
1510}
1511
1512static void gcwq_mayday_timeout(unsigned long __gcwq)
1513{
1514	struct global_cwq *gcwq = (void *)__gcwq;
1515	struct work_struct *work;
1516
1517	spin_lock_irq(&gcwq->lock);
 
1518
1519	if (need_to_create_worker(gcwq)) {
1520		/*
1521		 * We've been trying to create a new worker but
1522		 * haven't been successful.  We might be hitting an
1523		 * allocation deadlock.  Send distress signals to
1524		 * rescuers.
1525		 */
1526		list_for_each_entry(work, &gcwq->worklist, entry)
1527			send_mayday(work);
1528	}
1529
1530	spin_unlock_irq(&gcwq->lock);
 
1531
1532	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1533}
1534
1535/**
1536 * maybe_create_worker - create a new worker if necessary
1537 * @gcwq: gcwq to create a new worker for
1538 *
1539 * Create a new worker for @gcwq if necessary.  @gcwq is guaranteed to
1540 * have at least one idle worker on return from this function.  If
1541 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1542 * sent to all rescuers with works scheduled on @gcwq to resolve
1543 * possible allocation deadlock.
1544 *
1545 * On return, need_to_create_worker() is guaranteed to be false and
1546 * may_start_working() true.
1547 *
1548 * LOCKING:
1549 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1550 * multiple times.  Does GFP_KERNEL allocations.  Called only from
1551 * manager.
1552 *
1553 * RETURNS:
1554 * false if no action was taken and gcwq->lock stayed locked, true
1555 * otherwise.
1556 */
1557static bool maybe_create_worker(struct global_cwq *gcwq)
1558__releases(&gcwq->lock)
1559__acquires(&gcwq->lock)
1560{
1561	if (!need_to_create_worker(gcwq))
1562		return false;
1563restart:
1564	spin_unlock_irq(&gcwq->lock);
1565
1566	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1567	mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1568
1569	while (true) {
1570		struct worker *worker;
1571
1572		worker = create_worker(gcwq, true);
1573		if (worker) {
1574			del_timer_sync(&gcwq->mayday_timer);
1575			spin_lock_irq(&gcwq->lock);
1576			start_worker(worker);
1577			BUG_ON(need_to_create_worker(gcwq));
1578			return true;
1579		}
1580
1581		if (!need_to_create_worker(gcwq))
1582			break;
1583
1584		__set_current_state(TASK_INTERRUPTIBLE);
1585		schedule_timeout(CREATE_COOLDOWN);
1586
1587		if (!need_to_create_worker(gcwq))
1588			break;
1589	}
1590
1591	del_timer_sync(&gcwq->mayday_timer);
1592	spin_lock_irq(&gcwq->lock);
1593	if (need_to_create_worker(gcwq))
 
 
 
 
 
1594		goto restart;
1595	return true;
1596}
1597
1598/**
1599 * maybe_destroy_worker - destroy workers which have been idle for a while
1600 * @gcwq: gcwq to destroy workers for
1601 *
1602 * Destroy @gcwq workers which have been idle for longer than
1603 * IDLE_WORKER_TIMEOUT.
1604 *
1605 * LOCKING:
1606 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1607 * multiple times.  Called only from manager.
1608 *
1609 * RETURNS:
1610 * false if no action was taken and gcwq->lock stayed locked, true
1611 * otherwise.
1612 */
1613static bool maybe_destroy_workers(struct global_cwq *gcwq)
1614{
1615	bool ret = false;
1616
1617	while (too_many_workers(gcwq)) {
1618		struct worker *worker;
1619		unsigned long expires;
1620
1621		worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1622		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1623
1624		if (time_before(jiffies, expires)) {
1625			mod_timer(&gcwq->idle_timer, expires);
1626			break;
1627		}
1628
1629		destroy_worker(worker);
1630		ret = true;
1631	}
1632
1633	return ret;
1634}
1635
1636/**
1637 * manage_workers - manage worker pool
1638 * @worker: self
1639 *
1640 * Assume the manager role and manage gcwq worker pool @worker belongs
1641 * to.  At any given time, there can be only zero or one manager per
1642 * gcwq.  The exclusion is handled automatically by this function.
1643 *
1644 * The caller can safely start processing works on false return.  On
1645 * true return, it's guaranteed that need_to_create_worker() is false
1646 * and may_start_working() is true.
1647 *
1648 * CONTEXT:
1649 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1650 * multiple times.  Does GFP_KERNEL allocations.
1651 *
1652 * RETURNS:
1653 * false if no action was taken and gcwq->lock stayed locked, true if
1654 * some action was taken.
 
 
1655 */
1656static bool manage_workers(struct worker *worker)
1657{
1658	struct global_cwq *gcwq = worker->gcwq;
1659	bool ret = false;
1660
1661	if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1662		return ret;
1663
1664	gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1665	gcwq->flags |= GCWQ_MANAGING_WORKERS;
1666
1667	/*
1668	 * Destroy and then create so that may_start_working() is true
1669	 * on return.
 
 
 
 
 
 
1670	 */
1671	ret |= maybe_destroy_workers(gcwq);
1672	ret |= maybe_create_worker(gcwq);
 
1673
1674	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1675
1676	/*
1677	 * The trustee might be waiting to take over the manager
1678	 * position, tell it we're done.
1679	 */
1680	if (unlikely(gcwq->trustee))
1681		wake_up_all(&gcwq->trustee_wait);
1682
1683	return ret;
1684}
1685
1686/**
1687 * move_linked_works - move linked works to a list
1688 * @work: start of series of works to be scheduled
1689 * @head: target list to append @work to
1690 * @nextp: out paramter for nested worklist walking
1691 *
1692 * Schedule linked works starting from @work to @head.  Work series to
1693 * be scheduled starts at @work and includes any consecutive work with
1694 * WORK_STRUCT_LINKED set in its predecessor.
1695 *
1696 * If @nextp is not NULL, it's updated to point to the next work of
1697 * the last scheduled work.  This allows move_linked_works() to be
1698 * nested inside outer list_for_each_entry_safe().
1699 *
1700 * CONTEXT:
1701 * spin_lock_irq(gcwq->lock).
1702 */
1703static void move_linked_works(struct work_struct *work, struct list_head *head,
1704			      struct work_struct **nextp)
1705{
1706	struct work_struct *n;
1707
1708	/*
1709	 * Linked worklist will always end before the end of the list,
1710	 * use NULL for list head.
1711	 */
1712	list_for_each_entry_safe_from(work, n, NULL, entry) {
1713		list_move_tail(&work->entry, head);
1714		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1715			break;
1716	}
1717
1718	/*
1719	 * If we're already inside safe list traversal and have moved
1720	 * multiple works to the scheduled queue, the next position
1721	 * needs to be updated.
1722	 */
1723	if (nextp)
1724		*nextp = n;
1725}
1726
1727static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1728{
1729	struct work_struct *work = list_first_entry(&cwq->delayed_works,
1730						    struct work_struct, entry);
1731	struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
1732
1733	trace_workqueue_activate_work(work);
1734	move_linked_works(work, pos, NULL);
1735	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1736	cwq->nr_active++;
1737}
1738
1739/**
1740 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1741 * @cwq: cwq of interest
1742 * @color: color of work which left the queue
1743 * @delayed: for a delayed work
1744 *
1745 * A work either has completed or is removed from pending queue,
1746 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1747 *
1748 * CONTEXT:
1749 * spin_lock_irq(gcwq->lock).
1750 */
1751static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1752				 bool delayed)
1753{
1754	/* ignore uncolored works */
1755	if (color == WORK_NO_COLOR)
1756		return;
1757
1758	cwq->nr_in_flight[color]--;
1759
1760	if (!delayed) {
1761		cwq->nr_active--;
1762		if (!list_empty(&cwq->delayed_works)) {
1763			/* one down, submit a delayed one */
1764			if (cwq->nr_active < cwq->max_active)
1765				cwq_activate_first_delayed(cwq);
1766		}
1767	}
1768
1769	/* is flush in progress and are we at the flushing tip? */
1770	if (likely(cwq->flush_color != color))
1771		return;
1772
1773	/* are there still in-flight works? */
1774	if (cwq->nr_in_flight[color])
1775		return;
1776
1777	/* this cwq is done, clear flush_color */
1778	cwq->flush_color = -1;
1779
1780	/*
1781	 * If this was the last cwq, wake up the first flusher.  It
1782	 * will handle the rest.
1783	 */
1784	if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1785		complete(&cwq->wq->first_flusher->done);
1786}
1787
1788/**
1789 * process_one_work - process single work
1790 * @worker: self
1791 * @work: work to process
1792 *
1793 * Process @work.  This function contains all the logics necessary to
1794 * process a single work including synchronization against and
1795 * interaction with other workers on the same cpu, queueing and
1796 * flushing.  As long as context requirement is met, any worker can
1797 * call this function to process a work.
1798 *
1799 * CONTEXT:
1800 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
1801 */
1802static void process_one_work(struct worker *worker, struct work_struct *work)
1803__releases(&gcwq->lock)
1804__acquires(&gcwq->lock)
1805{
1806	struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1807	struct global_cwq *gcwq = cwq->gcwq;
1808	struct hlist_head *bwh = busy_worker_head(gcwq, work);
1809	bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
1810	work_func_t f = work->func;
1811	int work_color;
1812	struct worker *collision;
1813#ifdef CONFIG_LOCKDEP
1814	/*
1815	 * It is permissible to free the struct work_struct from
1816	 * inside the function that is called from it, this we need to
1817	 * take into account for lockdep too.  To avoid bogus "held
1818	 * lock freed" warnings as well as problems when looking into
1819	 * work->lockdep_map, make a copy and use that here.
1820	 */
1821	struct lockdep_map lockdep_map;
1822
1823	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
1824#endif
 
 
 
 
1825	/*
1826	 * A single work shouldn't be executed concurrently by
1827	 * multiple workers on a single cpu.  Check whether anyone is
1828	 * already processing the work.  If so, defer the work to the
1829	 * currently executing one.
1830	 */
1831	collision = __find_worker_executing_work(gcwq, bwh, work);
1832	if (unlikely(collision)) {
1833		move_linked_works(work, &collision->scheduled, NULL);
1834		return;
1835	}
1836
1837	/* claim and process */
1838	debug_work_deactivate(work);
1839	hlist_add_head(&worker->hentry, bwh);
1840	worker->current_work = work;
1841	worker->current_cwq = cwq;
 
1842	work_color = get_work_color(work);
1843
1844	/* record the current cpu number in the work data and dequeue */
1845	set_work_cpu(work, gcwq->cpu);
1846	list_del_init(&work->entry);
1847
1848	/*
1849	 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1850	 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
 
 
1851	 */
1852	if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1853		struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1854						struct work_struct, entry);
1855
1856		if (!list_empty(&gcwq->worklist) &&
1857		    get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1858			wake_up_worker(gcwq);
1859		else
1860			gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1861	}
 
 
 
1862
1863	/*
1864	 * CPU intensive works don't participate in concurrency
1865	 * management.  They're the scheduler's responsibility.
 
 
1866	 */
1867	if (unlikely(cpu_intensive))
1868		worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1869
1870	spin_unlock_irq(&gcwq->lock);
1871
1872	work_clear_pending(work);
1873	lock_map_acquire_read(&cwq->wq->lockdep_map);
1874	lock_map_acquire(&lockdep_map);
1875	trace_workqueue_execute_start(work);
1876	f(work);
1877	/*
1878	 * While we must be careful to not use "work" after this, the trace
1879	 * point will only record its address.
1880	 */
1881	trace_workqueue_execute_end(work);
1882	lock_map_release(&lockdep_map);
1883	lock_map_release(&cwq->wq->lockdep_map);
1884
1885	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1886		printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1887		       "%s/0x%08x/%d\n",
1888		       current->comm, preempt_count(), task_pid_nr(current));
1889		printk(KERN_ERR "    last function: ");
1890		print_symbol("%s\n", (unsigned long)f);
1891		debug_show_held_locks(current);
1892		dump_stack();
1893	}
1894
1895	spin_lock_irq(&gcwq->lock);
 
 
 
 
 
 
 
 
 
 
1896
1897	/* clear cpu intensive status */
1898	if (unlikely(cpu_intensive))
1899		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1900
1901	/* we're done with it, release */
1902	hlist_del_init(&worker->hentry);
1903	worker->current_work = NULL;
1904	worker->current_cwq = NULL;
1905	cwq_dec_nr_in_flight(cwq, work_color, false);
 
 
1906}
1907
1908/**
1909 * process_scheduled_works - process scheduled works
1910 * @worker: self
1911 *
1912 * Process all scheduled works.  Please note that the scheduled list
1913 * may change while processing a work, so this function repeatedly
1914 * fetches a work from the top and executes it.
1915 *
1916 * CONTEXT:
1917 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1918 * multiple times.
1919 */
1920static void process_scheduled_works(struct worker *worker)
1921{
1922	while (!list_empty(&worker->scheduled)) {
1923		struct work_struct *work = list_first_entry(&worker->scheduled,
1924						struct work_struct, entry);
1925		process_one_work(worker, work);
1926	}
1927}
1928
1929/**
1930 * worker_thread - the worker thread function
1931 * @__worker: self
1932 *
1933 * The gcwq worker thread function.  There's a single dynamic pool of
1934 * these per each cpu.  These workers process all works regardless of
1935 * their specific target workqueue.  The only exception is works which
1936 * belong to workqueues with a rescuer which will be explained in
1937 * rescuer_thread().
 
 
1938 */
1939static int worker_thread(void *__worker)
1940{
1941	struct worker *worker = __worker;
1942	struct global_cwq *gcwq = worker->gcwq;
1943
1944	/* tell the scheduler that this is a workqueue worker */
1945	worker->task->flags |= PF_WQ_WORKER;
1946woke_up:
1947	spin_lock_irq(&gcwq->lock);
1948
1949	/* DIE can be set only while we're idle, checking here is enough */
1950	if (worker->flags & WORKER_DIE) {
1951		spin_unlock_irq(&gcwq->lock);
 
1952		worker->task->flags &= ~PF_WQ_WORKER;
 
 
 
 
 
1953		return 0;
1954	}
1955
1956	worker_leave_idle(worker);
1957recheck:
1958	/* no more worker necessary? */
1959	if (!need_more_worker(gcwq))
1960		goto sleep;
1961
1962	/* do we need to manage? */
1963	if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1964		goto recheck;
1965
1966	/*
1967	 * ->scheduled list can only be filled while a worker is
1968	 * preparing to process a work or actually processing it.
1969	 * Make sure nobody diddled with it while I was sleeping.
1970	 */
1971	BUG_ON(!list_empty(&worker->scheduled));
1972
1973	/*
1974	 * When control reaches this point, we're guaranteed to have
1975	 * at least one idle worker or that someone else has already
1976	 * assumed the manager role.
 
 
1977	 */
1978	worker_clr_flags(worker, WORKER_PREP);
1979
1980	do {
1981		struct work_struct *work =
1982			list_first_entry(&gcwq->worklist,
1983					 struct work_struct, entry);
1984
 
 
1985		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1986			/* optimization path, not strictly necessary */
1987			process_one_work(worker, work);
1988			if (unlikely(!list_empty(&worker->scheduled)))
1989				process_scheduled_works(worker);
1990		} else {
1991			move_linked_works(work, &worker->scheduled, NULL);
1992			process_scheduled_works(worker);
1993		}
1994	} while (keep_working(gcwq));
1995
1996	worker_set_flags(worker, WORKER_PREP, false);
1997sleep:
1998	if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1999		goto recheck;
2000
2001	/*
2002	 * gcwq->lock is held and there's no work to process and no
2003	 * need to manage, sleep.  Workers are woken up only while
2004	 * holding gcwq->lock or from local cpu, so setting the
2005	 * current state before releasing gcwq->lock is enough to
2006	 * prevent losing any event.
2007	 */
2008	worker_enter_idle(worker);
2009	__set_current_state(TASK_INTERRUPTIBLE);
2010	spin_unlock_irq(&gcwq->lock);
2011	schedule();
2012	goto woke_up;
2013}
2014
2015/**
2016 * rescuer_thread - the rescuer thread function
2017 * @__wq: the associated workqueue
2018 *
2019 * Workqueue rescuer thread function.  There's one rescuer for each
2020 * workqueue which has WQ_RESCUER set.
2021 *
2022 * Regular work processing on a gcwq may block trying to create a new
2023 * worker which uses GFP_KERNEL allocation which has slight chance of
2024 * developing into deadlock if some works currently on the same queue
2025 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2026 * the problem rescuer solves.
2027 *
2028 * When such condition is possible, the gcwq summons rescuers of all
2029 * workqueues which have works queued on the gcwq and let them process
2030 * those works so that forward progress can be guaranteed.
2031 *
2032 * This should happen rarely.
 
 
2033 */
2034static int rescuer_thread(void *__wq)
2035{
2036	struct workqueue_struct *wq = __wq;
2037	struct worker *rescuer = wq->rescuer;
2038	struct list_head *scheduled = &rescuer->scheduled;
2039	bool is_unbound = wq->flags & WQ_UNBOUND;
2040	unsigned int cpu;
2041
2042	set_user_nice(current, RESCUER_NICE_LEVEL);
 
 
 
 
 
 
2043repeat:
2044	set_current_state(TASK_INTERRUPTIBLE);
2045
2046	if (kthread_should_stop())
2047		return 0;
2048
2049	/*
2050	 * See whether any cpu is asking for help.  Unbounded
2051	 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2052	 */
2053	for_each_mayday_cpu(cpu, wq->mayday_mask) {
2054		unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2055		struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
2056		struct global_cwq *gcwq = cwq->gcwq;
 
 
 
 
 
 
 
 
 
2057		struct work_struct *work, *n;
 
2058
2059		__set_current_state(TASK_RUNNING);
2060		mayday_clear_cpu(cpu, wq->mayday_mask);
2061
2062		/* migrate to the target cpu if possible */
2063		rescuer->gcwq = gcwq;
2064		worker_maybe_bind_and_lock(rescuer);
 
 
 
2065
2066		/*
2067		 * Slurp in all works issued via this workqueue and
2068		 * process'em.
2069		 */
2070		BUG_ON(!list_empty(&rescuer->scheduled));
2071		list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2072			if (get_work_cwq(work) == cwq)
 
 
2073				move_linked_works(work, scheduled, &n);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2074
2075		process_scheduled_works(rescuer);
 
 
 
 
2076
2077		/*
2078		 * Leave this gcwq.  If keep_working() is %true, notify a
2079		 * regular worker; otherwise, we end up with 0 concurrency
2080		 * and stalling the execution.
2081		 */
2082		if (keep_working(gcwq))
2083			wake_up_worker(gcwq);
2084
2085		spin_unlock_irq(&gcwq->lock);
 
 
 
 
 
2086	}
2087
 
 
 
 
 
 
 
 
 
 
2088	schedule();
2089	goto repeat;
2090}
2091
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2092struct wq_barrier {
2093	struct work_struct	work;
2094	struct completion	done;
 
2095};
2096
2097static void wq_barrier_func(struct work_struct *work)
2098{
2099	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2100	complete(&barr->done);
2101}
2102
2103/**
2104 * insert_wq_barrier - insert a barrier work
2105 * @cwq: cwq to insert barrier into
2106 * @barr: wq_barrier to insert
2107 * @target: target work to attach @barr to
2108 * @worker: worker currently executing @target, NULL if @target is not executing
2109 *
2110 * @barr is linked to @target such that @barr is completed only after
2111 * @target finishes execution.  Please note that the ordering
2112 * guarantee is observed only with respect to @target and on the local
2113 * cpu.
2114 *
2115 * Currently, a queued barrier can't be canceled.  This is because
2116 * try_to_grab_pending() can't determine whether the work to be
2117 * grabbed is at the head of the queue and thus can't clear LINKED
2118 * flag of the previous work while there must be a valid next work
2119 * after a work with LINKED flag set.
2120 *
2121 * Note that when @worker is non-NULL, @target may be modified
2122 * underneath us, so we can't reliably determine cwq from @target.
2123 *
2124 * CONTEXT:
2125 * spin_lock_irq(gcwq->lock).
2126 */
2127static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
2128			      struct wq_barrier *barr,
2129			      struct work_struct *target, struct worker *worker)
2130{
2131	struct list_head *head;
2132	unsigned int linked = 0;
2133
2134	/*
2135	 * debugobject calls are safe here even with gcwq->lock locked
2136	 * as we know for sure that this will not trigger any of the
2137	 * checks and call back into the fixup functions where we
2138	 * might deadlock.
2139	 */
2140	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2141	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2142	init_completion(&barr->done);
 
2143
2144	/*
2145	 * If @target is currently being executed, schedule the
2146	 * barrier to the worker; otherwise, put it after @target.
2147	 */
2148	if (worker)
2149		head = worker->scheduled.next;
2150	else {
2151		unsigned long *bits = work_data_bits(target);
2152
2153		head = target->entry.next;
2154		/* there can already be other linked works, inherit and set */
2155		linked = *bits & WORK_STRUCT_LINKED;
2156		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2157	}
2158
2159	debug_work_activate(&barr->work);
2160	insert_work(cwq, &barr->work, head,
2161		    work_color_to_flags(WORK_NO_COLOR) | linked);
2162}
2163
2164/**
2165 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2166 * @wq: workqueue being flushed
2167 * @flush_color: new flush color, < 0 for no-op
2168 * @work_color: new work color, < 0 for no-op
2169 *
2170 * Prepare cwqs for workqueue flushing.
2171 *
2172 * If @flush_color is non-negative, flush_color on all cwqs should be
2173 * -1.  If no cwq has in-flight commands at the specified color, all
2174 * cwq->flush_color's stay at -1 and %false is returned.  If any cwq
2175 * has in flight commands, its cwq->flush_color is set to
2176 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2177 * wakeup logic is armed and %true is returned.
2178 *
2179 * The caller should have initialized @wq->first_flusher prior to
2180 * calling this function with non-negative @flush_color.  If
2181 * @flush_color is negative, no flush color update is done and %false
2182 * is returned.
2183 *
2184 * If @work_color is non-negative, all cwqs should have the same
2185 * work_color which is previous to @work_color and all will be
2186 * advanced to @work_color.
2187 *
2188 * CONTEXT:
2189 * mutex_lock(wq->flush_mutex).
2190 *
2191 * RETURNS:
2192 * %true if @flush_color >= 0 and there's something to flush.  %false
2193 * otherwise.
2194 */
2195static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2196				      int flush_color, int work_color)
2197{
2198	bool wait = false;
2199	unsigned int cpu;
2200
2201	if (flush_color >= 0) {
2202		BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2203		atomic_set(&wq->nr_cwqs_to_flush, 1);
2204	}
2205
2206	for_each_cwq_cpu(cpu, wq) {
2207		struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
2208		struct global_cwq *gcwq = cwq->gcwq;
2209
2210		spin_lock_irq(&gcwq->lock);
2211
2212		if (flush_color >= 0) {
2213			BUG_ON(cwq->flush_color != -1);
2214
2215			if (cwq->nr_in_flight[flush_color]) {
2216				cwq->flush_color = flush_color;
2217				atomic_inc(&wq->nr_cwqs_to_flush);
2218				wait = true;
2219			}
2220		}
2221
2222		if (work_color >= 0) {
2223			BUG_ON(work_color != work_next_color(cwq->work_color));
2224			cwq->work_color = work_color;
2225		}
2226
2227		spin_unlock_irq(&gcwq->lock);
2228	}
2229
2230	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2231		complete(&wq->first_flusher->done);
2232
2233	return wait;
2234}
2235
2236/**
2237 * flush_workqueue - ensure that any scheduled work has run to completion.
2238 * @wq: workqueue to flush
2239 *
2240 * Forces execution of the workqueue and blocks until its completion.
2241 * This is typically used in driver shutdown handlers.
2242 *
2243 * We sleep until all works which were queued on entry have been handled,
2244 * but we are not livelocked by new incoming ones.
2245 */
2246void flush_workqueue(struct workqueue_struct *wq)
2247{
2248	struct wq_flusher this_flusher = {
2249		.list = LIST_HEAD_INIT(this_flusher.list),
2250		.flush_color = -1,
2251		.done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2252	};
2253	int next_color;
2254
2255	lock_map_acquire(&wq->lockdep_map);
2256	lock_map_release(&wq->lockdep_map);
2257
2258	mutex_lock(&wq->flush_mutex);
2259
2260	/*
2261	 * Start-to-wait phase
2262	 */
2263	next_color = work_next_color(wq->work_color);
2264
2265	if (next_color != wq->flush_color) {
2266		/*
2267		 * Color space is not full.  The current work_color
2268		 * becomes our flush_color and work_color is advanced
2269		 * by one.
2270		 */
2271		BUG_ON(!list_empty(&wq->flusher_overflow));
2272		this_flusher.flush_color = wq->work_color;
2273		wq->work_color = next_color;
2274
2275		if (!wq->first_flusher) {
2276			/* no flush in progress, become the first flusher */
2277			BUG_ON(wq->flush_color != this_flusher.flush_color);
2278
2279			wq->first_flusher = &this_flusher;
2280
2281			if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2282						       wq->work_color)) {
2283				/* nothing to flush, done */
2284				wq->flush_color = next_color;
2285				wq->first_flusher = NULL;
2286				goto out_unlock;
2287			}
2288		} else {
2289			/* wait in queue */
2290			BUG_ON(wq->flush_color == this_flusher.flush_color);
2291			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2292			flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2293		}
2294	} else {
2295		/*
2296		 * Oops, color space is full, wait on overflow queue.
2297		 * The next flush completion will assign us
2298		 * flush_color and transfer to flusher_queue.
2299		 */
2300		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2301	}
2302
2303	mutex_unlock(&wq->flush_mutex);
 
 
2304
2305	wait_for_completion(&this_flusher.done);
2306
2307	/*
2308	 * Wake-up-and-cascade phase
2309	 *
2310	 * First flushers are responsible for cascading flushes and
2311	 * handling overflow.  Non-first flushers can simply return.
2312	 */
2313	if (wq->first_flusher != &this_flusher)
2314		return;
2315
2316	mutex_lock(&wq->flush_mutex);
2317
2318	/* we might have raced, check again with mutex held */
2319	if (wq->first_flusher != &this_flusher)
2320		goto out_unlock;
2321
2322	wq->first_flusher = NULL;
2323
2324	BUG_ON(!list_empty(&this_flusher.list));
2325	BUG_ON(wq->flush_color != this_flusher.flush_color);
2326
2327	while (true) {
2328		struct wq_flusher *next, *tmp;
2329
2330		/* complete all the flushers sharing the current flush color */
2331		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2332			if (next->flush_color != wq->flush_color)
2333				break;
2334			list_del_init(&next->list);
2335			complete(&next->done);
2336		}
2337
2338		BUG_ON(!list_empty(&wq->flusher_overflow) &&
2339		       wq->flush_color != work_next_color(wq->work_color));
2340
2341		/* this flush_color is finished, advance by one */
2342		wq->flush_color = work_next_color(wq->flush_color);
2343
2344		/* one color has been freed, handle overflow queue */
2345		if (!list_empty(&wq->flusher_overflow)) {
2346			/*
2347			 * Assign the same color to all overflowed
2348			 * flushers, advance work_color and append to
2349			 * flusher_queue.  This is the start-to-wait
2350			 * phase for these overflowed flushers.
2351			 */
2352			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2353				tmp->flush_color = wq->work_color;
2354
2355			wq->work_color = work_next_color(wq->work_color);
2356
2357			list_splice_tail_init(&wq->flusher_overflow,
2358					      &wq->flusher_queue);
2359			flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2360		}
2361
2362		if (list_empty(&wq->flusher_queue)) {
2363			BUG_ON(wq->flush_color != wq->work_color);
2364			break;
2365		}
2366
2367		/*
2368		 * Need to flush more colors.  Make the next flusher
2369		 * the new first flusher and arm cwqs.
2370		 */
2371		BUG_ON(wq->flush_color == wq->work_color);
2372		BUG_ON(wq->flush_color != next->flush_color);
2373
2374		list_del_init(&next->list);
2375		wq->first_flusher = next;
2376
2377		if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2378			break;
2379
2380		/*
2381		 * Meh... this color is already done, clear first
2382		 * flusher and repeat cascading.
2383		 */
2384		wq->first_flusher = NULL;
2385	}
2386
2387out_unlock:
2388	mutex_unlock(&wq->flush_mutex);
2389}
2390EXPORT_SYMBOL_GPL(flush_workqueue);
2391
2392/**
2393 * drain_workqueue - drain a workqueue
2394 * @wq: workqueue to drain
2395 *
2396 * Wait until the workqueue becomes empty.  While draining is in progress,
2397 * only chain queueing is allowed.  IOW, only currently pending or running
2398 * work items on @wq can queue further work items on it.  @wq is flushed
2399 * repeatedly until it becomes empty.  The number of flushing is detemined
2400 * by the depth of chaining and should be relatively short.  Whine if it
2401 * takes too long.
2402 */
2403void drain_workqueue(struct workqueue_struct *wq)
2404{
2405	unsigned int flush_cnt = 0;
2406	unsigned int cpu;
2407
2408	/*
2409	 * __queue_work() needs to test whether there are drainers, is much
2410	 * hotter than drain_workqueue() and already looks at @wq->flags.
2411	 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2412	 */
2413	spin_lock(&workqueue_lock);
2414	if (!wq->nr_drainers++)
2415		wq->flags |= WQ_DRAINING;
2416	spin_unlock(&workqueue_lock);
2417reflush:
2418	flush_workqueue(wq);
2419
2420	for_each_cwq_cpu(cpu, wq) {
2421		struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
2422		bool drained;
2423
2424		spin_lock_irq(&cwq->gcwq->lock);
2425		drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
2426		spin_unlock_irq(&cwq->gcwq->lock);
2427
2428		if (drained)
2429			continue;
2430
2431		if (++flush_cnt == 10 ||
2432		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2433			pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2434				   wq->name, flush_cnt);
 
 
2435		goto reflush;
2436	}
2437
2438	spin_lock(&workqueue_lock);
2439	if (!--wq->nr_drainers)
2440		wq->flags &= ~WQ_DRAINING;
2441	spin_unlock(&workqueue_lock);
2442}
2443EXPORT_SYMBOL_GPL(drain_workqueue);
2444
2445static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2446			     bool wait_executing)
2447{
2448	struct worker *worker = NULL;
2449	struct global_cwq *gcwq;
2450	struct cpu_workqueue_struct *cwq;
2451
2452	might_sleep();
2453	gcwq = get_work_gcwq(work);
2454	if (!gcwq)
 
 
 
2455		return false;
 
2456
2457	spin_lock_irq(&gcwq->lock);
2458	if (!list_empty(&work->entry)) {
2459		/*
2460		 * See the comment near try_to_grab_pending()->smp_rmb().
2461		 * If it was re-queued to a different gcwq under us, we
2462		 * are not going to wait.
2463		 */
2464		smp_rmb();
2465		cwq = get_work_cwq(work);
2466		if (unlikely(!cwq || gcwq != cwq->gcwq))
2467			goto already_gone;
2468	} else if (wait_executing) {
2469		worker = find_worker_executing_work(gcwq, work);
2470		if (!worker)
2471			goto already_gone;
2472		cwq = worker->current_cwq;
2473	} else
2474		goto already_gone;
 
2475
2476	insert_wq_barrier(cwq, barr, work, worker);
2477	spin_unlock_irq(&gcwq->lock);
2478
2479	/*
2480	 * If @max_active is 1 or rescuer is in use, flushing another work
2481	 * item on the same workqueue may lead to deadlock.  Make sure the
2482	 * flusher is not running on the same workqueue by verifying write
2483	 * access.
2484	 */
2485	if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2486		lock_map_acquire(&cwq->wq->lockdep_map);
2487	else
2488		lock_map_acquire_read(&cwq->wq->lockdep_map);
2489	lock_map_release(&cwq->wq->lockdep_map);
2490
2491	return true;
2492already_gone:
2493	spin_unlock_irq(&gcwq->lock);
2494	return false;
2495}
2496
2497/**
2498 * flush_work - wait for a work to finish executing the last queueing instance
2499 * @work: the work to flush
2500 *
2501 * Wait until @work has finished execution.  This function considers
2502 * only the last queueing instance of @work.  If @work has been
2503 * enqueued across different CPUs on a non-reentrant workqueue or on
2504 * multiple workqueues, @work might still be executing on return on
2505 * some of the CPUs from earlier queueing.
2506 *
2507 * If @work was queued only on a non-reentrant, ordered or unbound
2508 * workqueue, @work is guaranteed to be idle on return if it hasn't
2509 * been requeued since flush started.
2510 *
2511 * RETURNS:
2512 * %true if flush_work() waited for the work to finish execution,
2513 * %false if it was already idle.
2514 */
2515bool flush_work(struct work_struct *work)
2516{
2517	struct wq_barrier barr;
2518
2519	lock_map_acquire(&work->lockdep_map);
2520	lock_map_release(&work->lockdep_map);
2521
2522	if (start_flush_work(work, &barr, true)) {
2523		wait_for_completion(&barr.done);
2524		destroy_work_on_stack(&barr.work);
2525		return true;
2526	} else
2527		return false;
 
2528}
2529EXPORT_SYMBOL_GPL(flush_work);
2530
2531static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2532{
2533	struct wq_barrier barr;
2534	struct worker *worker;
2535
2536	spin_lock_irq(&gcwq->lock);
2537
2538	worker = find_worker_executing_work(gcwq, work);
2539	if (unlikely(worker))
2540		insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2541
2542	spin_unlock_irq(&gcwq->lock);
2543
2544	if (unlikely(worker)) {
2545		wait_for_completion(&barr.done);
2546		destroy_work_on_stack(&barr.work);
2547		return true;
2548	} else
2549		return false;
2550}
2551
2552static bool wait_on_work(struct work_struct *work)
2553{
2554	bool ret = false;
2555	int cpu;
2556
2557	might_sleep();
2558
2559	lock_map_acquire(&work->lockdep_map);
2560	lock_map_release(&work->lockdep_map);
2561
2562	for_each_gcwq_cpu(cpu)
2563		ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2564	return ret;
2565}
2566
2567/**
2568 * flush_work_sync - wait until a work has finished execution
2569 * @work: the work to flush
2570 *
2571 * Wait until @work has finished execution.  On return, it's
2572 * guaranteed that all queueing instances of @work which happened
2573 * before this function is called are finished.  In other words, if
2574 * @work hasn't been requeued since this function was called, @work is
2575 * guaranteed to be idle on return.
2576 *
2577 * RETURNS:
2578 * %true if flush_work_sync() waited for the work to finish execution,
2579 * %false if it was already idle.
2580 */
2581bool flush_work_sync(struct work_struct *work)
2582{
2583	struct wq_barrier barr;
2584	bool pending, waited;
 
2585
2586	/* we'll wait for executions separately, queue barr only if pending */
2587	pending = start_flush_work(work, &barr, false);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2588
2589	/* wait for executions to finish */
2590	waited = wait_on_work(work);
 
 
 
 
 
 
 
 
 
2591
2592	/* wait for the pending one */
2593	if (pending) {
2594		wait_for_completion(&barr.done);
2595		destroy_work_on_stack(&barr.work);
2596	}
2597
2598	return pending || waited;
2599}
2600EXPORT_SYMBOL_GPL(flush_work_sync);
2601
2602/*
2603 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
2604 * so this work can't be re-armed in any way.
2605 */
2606static int try_to_grab_pending(struct work_struct *work)
2607{
2608	struct global_cwq *gcwq;
2609	int ret = -1;
2610
2611	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
2612		return 0;
2613
2614	/*
2615	 * The queueing is in progress, or it is already queued. Try to
2616	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
 
2617	 */
2618	gcwq = get_work_gcwq(work);
2619	if (!gcwq)
2620		return ret;
2621
2622	spin_lock_irq(&gcwq->lock);
2623	if (!list_empty(&work->entry)) {
2624		/*
2625		 * This work is queued, but perhaps we locked the wrong gcwq.
2626		 * In that case we must see the new value after rmb(), see
2627		 * insert_work()->wmb().
2628		 */
2629		smp_rmb();
2630		if (gcwq == get_work_gcwq(work)) {
2631			debug_work_deactivate(work);
2632			list_del_init(&work->entry);
2633			cwq_dec_nr_in_flight(get_work_cwq(work),
2634				get_work_color(work),
2635				*work_data_bits(work) & WORK_STRUCT_DELAYED);
2636			ret = 1;
2637		}
2638	}
2639	spin_unlock_irq(&gcwq->lock);
2640
2641	return ret;
2642}
2643
2644static bool __cancel_work_timer(struct work_struct *work,
2645				struct timer_list* timer)
2646{
2647	int ret;
2648
2649	do {
2650		ret = (timer && likely(del_timer(timer)));
2651		if (!ret)
2652			ret = try_to_grab_pending(work);
2653		wait_on_work(work);
2654	} while (unlikely(ret < 0));
2655
2656	clear_work_data(work);
2657	return ret;
2658}
2659
2660/**
2661 * cancel_work_sync - cancel a work and wait for it to finish
2662 * @work: the work to cancel
2663 *
2664 * Cancel @work and wait for its execution to finish.  This function
2665 * can be used even if the work re-queues itself or migrates to
2666 * another workqueue.  On return from this function, @work is
2667 * guaranteed to be not pending or executing on any CPU.
2668 *
2669 * cancel_work_sync(&delayed_work->work) must not be used for
2670 * delayed_work's.  Use cancel_delayed_work_sync() instead.
2671 *
2672 * The caller must ensure that the workqueue on which @work was last
2673 * queued can't be destroyed before this function returns.
2674 *
2675 * RETURNS:
2676 * %true if @work was pending, %false otherwise.
2677 */
2678bool cancel_work_sync(struct work_struct *work)
2679{
2680	return __cancel_work_timer(work, NULL);
2681}
2682EXPORT_SYMBOL_GPL(cancel_work_sync);
2683
2684/**
2685 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2686 * @dwork: the delayed work to flush
2687 *
2688 * Delayed timer is cancelled and the pending work is queued for
2689 * immediate execution.  Like flush_work(), this function only
2690 * considers the last queueing instance of @dwork.
2691 *
2692 * RETURNS:
2693 * %true if flush_work() waited for the work to finish execution,
2694 * %false if it was already idle.
2695 */
2696bool flush_delayed_work(struct delayed_work *dwork)
2697{
 
2698	if (del_timer_sync(&dwork->timer))
2699		__queue_work(raw_smp_processor_id(),
2700			     get_work_cwq(&dwork->work)->wq, &dwork->work);
2701	return flush_work(&dwork->work);
2702}
2703EXPORT_SYMBOL(flush_delayed_work);
2704
2705/**
2706 * flush_delayed_work_sync - wait for a dwork to finish
2707 * @dwork: the delayed work to flush
 
 
 
 
 
2708 *
2709 * Delayed timer is cancelled and the pending work is queued for
2710 * execution immediately.  Other than timer handling, its behavior
2711 * is identical to flush_work_sync().
 
2712 *
2713 * RETURNS:
2714 * %true if flush_work_sync() waited for the work to finish execution,
2715 * %false if it was already idle.
2716 */
2717bool flush_delayed_work_sync(struct delayed_work *dwork)
2718{
2719	if (del_timer_sync(&dwork->timer))
2720		__queue_work(raw_smp_processor_id(),
2721			     get_work_cwq(&dwork->work)->wq, &dwork->work);
2722	return flush_work_sync(&dwork->work);
 
 
 
 
 
 
 
 
 
 
2723}
2724EXPORT_SYMBOL(flush_delayed_work_sync);
2725
2726/**
2727 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2728 * @dwork: the delayed work cancel
2729 *
2730 * This is cancel_work_sync() for delayed works.
2731 *
2732 * RETURNS:
2733 * %true if @dwork was pending, %false otherwise.
2734 */
2735bool cancel_delayed_work_sync(struct delayed_work *dwork)
2736{
2737	return __cancel_work_timer(&dwork->work, &dwork->timer);
2738}
2739EXPORT_SYMBOL(cancel_delayed_work_sync);
2740
2741/**
2742 * schedule_work - put work task in global workqueue
2743 * @work: job to be done
2744 *
2745 * Returns zero if @work was already on the kernel-global workqueue and
2746 * non-zero otherwise.
2747 *
2748 * This puts a job in the kernel-global workqueue if it was not already
2749 * queued and leaves it in the same position on the kernel-global
2750 * workqueue otherwise.
2751 */
2752int schedule_work(struct work_struct *work)
2753{
2754	return queue_work(system_wq, work);
2755}
2756EXPORT_SYMBOL(schedule_work);
2757
2758/*
2759 * schedule_work_on - put work task on a specific cpu
2760 * @cpu: cpu to put the work task on
2761 * @work: job to be done
2762 *
2763 * This puts a job on a specific cpu
2764 */
2765int schedule_work_on(int cpu, struct work_struct *work)
2766{
2767	return queue_work_on(cpu, system_wq, work);
2768}
2769EXPORT_SYMBOL(schedule_work_on);
2770
2771/**
2772 * schedule_delayed_work - put work task in global workqueue after delay
2773 * @dwork: job to be done
2774 * @delay: number of jiffies to wait or 0 for immediate execution
2775 *
2776 * After waiting for a given time this puts a job in the kernel-global
2777 * workqueue.
2778 */
2779int schedule_delayed_work(struct delayed_work *dwork,
2780					unsigned long delay)
2781{
2782	return queue_delayed_work(system_wq, dwork, delay);
2783}
2784EXPORT_SYMBOL(schedule_delayed_work);
2785
2786/**
2787 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2788 * @cpu: cpu to use
2789 * @dwork: job to be done
2790 * @delay: number of jiffies to wait
2791 *
2792 * After waiting for a given time this puts a job in the kernel-global
2793 * workqueue on the specified CPU.
2794 */
2795int schedule_delayed_work_on(int cpu,
2796			struct delayed_work *dwork, unsigned long delay)
2797{
2798	return queue_delayed_work_on(cpu, system_wq, dwork, delay);
2799}
2800EXPORT_SYMBOL(schedule_delayed_work_on);
2801
2802/**
2803 * schedule_on_each_cpu - execute a function synchronously on each online CPU
2804 * @func: the function to call
2805 *
2806 * schedule_on_each_cpu() executes @func on each online CPU using the
2807 * system workqueue and blocks until all CPUs have completed.
2808 * schedule_on_each_cpu() is very slow.
2809 *
2810 * RETURNS:
2811 * 0 on success, -errno on failure.
2812 */
2813int schedule_on_each_cpu(work_func_t func)
2814{
2815	int cpu;
2816	struct work_struct __percpu *works;
2817
2818	works = alloc_percpu(struct work_struct);
2819	if (!works)
2820		return -ENOMEM;
2821
2822	get_online_cpus();
2823
2824	for_each_online_cpu(cpu) {
2825		struct work_struct *work = per_cpu_ptr(works, cpu);
2826
2827		INIT_WORK(work, func);
2828		schedule_work_on(cpu, work);
2829	}
2830
2831	for_each_online_cpu(cpu)
2832		flush_work(per_cpu_ptr(works, cpu));
2833
2834	put_online_cpus();
2835	free_percpu(works);
2836	return 0;
2837}
2838
2839/**
2840 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2841 *
2842 * Forces execution of the kernel-global workqueue and blocks until its
2843 * completion.
2844 *
2845 * Think twice before calling this function!  It's very easy to get into
2846 * trouble if you don't take great care.  Either of the following situations
2847 * will lead to deadlock:
2848 *
2849 *	One of the work items currently on the workqueue needs to acquire
2850 *	a lock held by your code or its caller.
2851 *
2852 *	Your code is running in the context of a work routine.
2853 *
2854 * They will be detected by lockdep when they occur, but the first might not
2855 * occur very often.  It depends on what work items are on the workqueue and
2856 * what locks they need, which you have no control over.
2857 *
2858 * In most situations flushing the entire workqueue is overkill; you merely
2859 * need to know that a particular work item isn't queued and isn't running.
2860 * In such cases you should use cancel_delayed_work_sync() or
2861 * cancel_work_sync() instead.
2862 */
2863void flush_scheduled_work(void)
2864{
2865	flush_workqueue(system_wq);
2866}
2867EXPORT_SYMBOL(flush_scheduled_work);
2868
2869/**
2870 * execute_in_process_context - reliably execute the routine with user context
2871 * @fn:		the function to execute
2872 * @ew:		guaranteed storage for the execute work structure (must
2873 *		be available when the work executes)
2874 *
2875 * Executes the function immediately if process context is available,
2876 * otherwise schedules the function for delayed execution.
2877 *
2878 * Returns:	0 - function was executed
2879 *		1 - function was scheduled for execution
2880 */
2881int execute_in_process_context(work_func_t fn, struct execute_work *ew)
2882{
2883	if (!in_interrupt()) {
2884		fn(&ew->work);
2885		return 0;
2886	}
2887
2888	INIT_WORK(&ew->work, fn);
2889	schedule_work(&ew->work);
2890
2891	return 1;
2892}
2893EXPORT_SYMBOL_GPL(execute_in_process_context);
2894
2895int keventd_up(void)
 
 
 
 
 
 
2896{
2897	return system_wq != NULL;
 
 
 
2898}
2899
2900static int alloc_cwqs(struct workqueue_struct *wq)
 
 
 
 
 
 
 
 
 
2901{
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2902	/*
2903	 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2904	 * Make sure that the alignment isn't lower than that of
2905	 * unsigned long long.
2906	 */
2907	const size_t size = sizeof(struct cpu_workqueue_struct);
2908	const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2909				   __alignof__(unsigned long long));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2910
2911	if (!(wq->flags & WQ_UNBOUND))
2912		wq->cpu_wq.pcpu = __alloc_percpu(size, align);
2913	else {
2914		void *ptr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2915
2916		/*
2917		 * Allocate enough room to align cwq and put an extra
2918		 * pointer at the end pointing back to the originally
2919		 * allocated pointer which will be used for free.
2920		 */
2921		ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2922		if (ptr) {
2923			wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2924			*(void **)(wq->cpu_wq.single + 1) = ptr;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2925		}
2926	}
2927
2928	/* just in case, make sure it's actually aligned */
2929	BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2930	return wq->cpu_wq.v ? 0 : -ENOMEM;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2931}
2932
2933static void free_cwqs(struct workqueue_struct *wq)
2934{
2935	if (!(wq->flags & WQ_UNBOUND))
2936		free_percpu(wq->cpu_wq.pcpu);
2937	else if (wq->cpu_wq.single) {
2938		/* the pointer to free is stored right after the cwq */
2939		kfree(*(void **)(wq->cpu_wq.single + 1));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2940	}
2941}
2942
2943static int wq_clamp_max_active(int max_active, unsigned int flags,
2944			       const char *name)
2945{
2946	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2947
2948	if (max_active < 1 || max_active > lim)
2949		printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2950		       "is out of range, clamping between %d and %d\n",
2951		       max_active, name, 1, lim);
2952
2953	return clamp_val(max_active, 1, lim);
2954}
2955
2956struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
2957					       unsigned int flags,
2958					       int max_active,
2959					       struct lock_class_key *key,
2960					       const char *lock_name, ...)
2961{
2962	va_list args, args1;
 
2963	struct workqueue_struct *wq;
2964	unsigned int cpu;
2965	size_t namelen;
 
 
 
2966
2967	/* determine namelen, allocate wq and format name */
2968	va_start(args, lock_name);
2969	va_copy(args1, args);
2970	namelen = vsnprintf(NULL, 0, fmt, args) + 1;
2971
2972	wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
2973	if (!wq)
2974		goto err;
2975
2976	vsnprintf(wq->name, namelen, fmt, args1);
 
 
 
 
 
 
 
2977	va_end(args);
2978	va_end(args1);
2979
2980	/*
2981	 * Workqueues which may be used during memory reclaim should
2982	 * have a rescuer to guarantee forward progress.
2983	 */
2984	if (flags & WQ_MEM_RECLAIM)
2985		flags |= WQ_RESCUER;
2986
2987	/*
2988	 * Unbound workqueues aren't concurrency managed and should be
2989	 * dispatched to workers immediately.
2990	 */
2991	if (flags & WQ_UNBOUND)
2992		flags |= WQ_HIGHPRI;
2993
2994	max_active = max_active ?: WQ_DFL_ACTIVE;
2995	max_active = wq_clamp_max_active(max_active, flags, wq->name);
2996
2997	/* init wq */
2998	wq->flags = flags;
2999	wq->saved_max_active = max_active;
3000	mutex_init(&wq->flush_mutex);
3001	atomic_set(&wq->nr_cwqs_to_flush, 0);
 
3002	INIT_LIST_HEAD(&wq->flusher_queue);
3003	INIT_LIST_HEAD(&wq->flusher_overflow);
 
3004
3005	lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
3006	INIT_LIST_HEAD(&wq->list);
3007
3008	if (alloc_cwqs(wq) < 0)
3009		goto err;
3010
3011	for_each_cwq_cpu(cpu, wq) {
3012		struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3013		struct global_cwq *gcwq = get_gcwq(cpu);
3014
3015		BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3016		cwq->gcwq = gcwq;
3017		cwq->wq = wq;
3018		cwq->flush_color = -1;
3019		cwq->max_active = max_active;
3020		INIT_LIST_HEAD(&cwq->delayed_works);
3021	}
3022
3023	if (flags & WQ_RESCUER) {
3024		struct worker *rescuer;
3025
3026		if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
3027			goto err;
3028
3029		wq->rescuer = rescuer = alloc_worker();
3030		if (!rescuer)
3031			goto err;
3032
3033		rescuer->task = kthread_create(rescuer_thread, wq, "%s",
 
3034					       wq->name);
3035		if (IS_ERR(rescuer->task))
3036			goto err;
 
 
3037
3038		rescuer->task->flags |= PF_THREAD_BOUND;
 
3039		wake_up_process(rescuer->task);
3040	}
3041
 
 
 
3042	/*
3043	 * workqueue_lock protects global freeze state and workqueues
3044	 * list.  Grab it, set max_active accordingly and add the new
3045	 * workqueue to workqueues list.
3046	 */
3047	spin_lock(&workqueue_lock);
3048
3049	if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
3050		for_each_cwq_cpu(cpu, wq)
3051			get_cwq(cpu, wq)->max_active = 0;
 
3052
3053	list_add(&wq->list, &workqueues);
3054
3055	spin_unlock(&workqueue_lock);
3056
3057	return wq;
3058err:
3059	if (wq) {
3060		free_cwqs(wq);
3061		free_mayday_mask(wq->mayday_mask);
3062		kfree(wq->rescuer);
3063		kfree(wq);
3064	}
3065	return NULL;
3066}
3067EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
3068
3069/**
3070 * destroy_workqueue - safely terminate a workqueue
3071 * @wq: target workqueue
3072 *
3073 * Safely destroy a workqueue. All work currently pending will be done first.
3074 */
3075void destroy_workqueue(struct workqueue_struct *wq)
3076{
3077	unsigned int cpu;
 
3078
3079	/* drain it before proceeding with destruction */
3080	drain_workqueue(wq);
3081
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3082	/*
3083	 * wq list is used to freeze wq, remove from list after
3084	 * flushing is complete in case freeze races us.
3085	 */
3086	spin_lock(&workqueue_lock);
3087	list_del(&wq->list);
3088	spin_unlock(&workqueue_lock);
3089
3090	/* sanity check */
3091	for_each_cwq_cpu(cpu, wq) {
3092		struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3093		int i;
3094
3095		for (i = 0; i < WORK_NR_COLORS; i++)
3096			BUG_ON(cwq->nr_in_flight[i]);
3097		BUG_ON(cwq->nr_active);
3098		BUG_ON(!list_empty(&cwq->delayed_works));
3099	}
3100
3101	if (wq->flags & WQ_RESCUER) {
3102		kthread_stop(wq->rescuer->task);
3103		free_mayday_mask(wq->mayday_mask);
3104		kfree(wq->rescuer);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3105	}
3106
3107	free_cwqs(wq);
3108	kfree(wq);
3109}
3110EXPORT_SYMBOL_GPL(destroy_workqueue);
3111
3112/**
3113 * workqueue_set_max_active - adjust max_active of a workqueue
3114 * @wq: target workqueue
3115 * @max_active: new max_active value.
3116 *
3117 * Set max_active of @wq to @max_active.
3118 *
3119 * CONTEXT:
3120 * Don't call from IRQ context.
3121 */
3122void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3123{
3124	unsigned int cpu;
 
 
 
 
3125
3126	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
3127
3128	spin_lock(&workqueue_lock);
3129
3130	wq->saved_max_active = max_active;
3131
3132	for_each_cwq_cpu(cpu, wq) {
3133		struct global_cwq *gcwq = get_gcwq(cpu);
3134
3135		spin_lock_irq(&gcwq->lock);
 
 
3136
3137		if (!(wq->flags & WQ_FREEZABLE) ||
3138		    !(gcwq->flags & GCWQ_FREEZING))
3139			get_cwq(gcwq->cpu, wq)->max_active = max_active;
3140
3141		spin_unlock_irq(&gcwq->lock);
3142	}
 
 
 
 
 
3143
3144	spin_unlock(&workqueue_lock);
3145}
3146EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3147
3148/**
3149 * workqueue_congested - test whether a workqueue is congested
3150 * @cpu: CPU in question
3151 * @wq: target workqueue
3152 *
3153 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
3154 * no synchronization around this function and the test result is
3155 * unreliable and only useful as advisory hints or for debugging.
3156 *
3157 * RETURNS:
 
 
 
 
 
 
3158 * %true if congested, %false otherwise.
3159 */
3160bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3161{
3162	struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
 
 
 
 
 
3163
3164	return !list_empty(&cwq->delayed_works);
3165}
3166EXPORT_SYMBOL_GPL(workqueue_congested);
 
3167
3168/**
3169 * work_cpu - return the last known associated cpu for @work
3170 * @work: the work of interest
3171 *
3172 * RETURNS:
3173 * CPU number if @work was ever queued.  WORK_CPU_NONE otherwise.
3174 */
3175unsigned int work_cpu(struct work_struct *work)
3176{
3177	struct global_cwq *gcwq = get_work_gcwq(work);
3178
3179	return gcwq ? gcwq->cpu : WORK_CPU_NONE;
3180}
3181EXPORT_SYMBOL_GPL(work_cpu);
3182
3183/**
3184 * work_busy - test whether a work is currently pending or running
3185 * @work: the work to be tested
3186 *
3187 * Test whether @work is currently pending or running.  There is no
3188 * synchronization around this function and the test result is
3189 * unreliable and only useful as advisory hints or for debugging.
3190 * Especially for reentrant wqs, the pending state might hide the
3191 * running state.
3192 *
3193 * RETURNS:
3194 * OR'd bitmask of WORK_BUSY_* bits.
3195 */
3196unsigned int work_busy(struct work_struct *work)
3197{
3198	struct global_cwq *gcwq = get_work_gcwq(work);
3199	unsigned long flags;
3200	unsigned int ret = 0;
3201
3202	if (!gcwq)
3203		return false;
3204
3205	spin_lock_irqsave(&gcwq->lock, flags);
3206
3207	if (work_pending(work))
3208		ret |= WORK_BUSY_PENDING;
3209	if (find_worker_executing_work(gcwq, work))
3210		ret |= WORK_BUSY_RUNNING;
3211
3212	spin_unlock_irqrestore(&gcwq->lock, flags);
 
 
 
 
 
 
 
 
3213
3214	return ret;
3215}
3216EXPORT_SYMBOL_GPL(work_busy);
3217
3218/*
3219 * CPU hotplug.
3220 *
3221 * There are two challenges in supporting CPU hotplug.  Firstly, there
3222 * are a lot of assumptions on strong associations among work, cwq and
3223 * gcwq which make migrating pending and scheduled works very
3224 * difficult to implement without impacting hot paths.  Secondly,
3225 * gcwqs serve mix of short, long and very long running works making
3226 * blocked draining impractical.
3227 *
3228 * This is solved by allowing a gcwq to be detached from CPU, running
3229 * it with unbound (rogue) workers and allowing it to be reattached
3230 * later if the cpu comes back online.  A separate thread is created
3231 * to govern a gcwq in such state and is called the trustee of the
3232 * gcwq.
3233 *
3234 * Trustee states and their descriptions.
3235 *
3236 * START	Command state used on startup.  On CPU_DOWN_PREPARE, a
3237 *		new trustee is started with this state.
3238 *
3239 * IN_CHARGE	Once started, trustee will enter this state after
3240 *		assuming the manager role and making all existing
3241 *		workers rogue.  DOWN_PREPARE waits for trustee to
3242 *		enter this state.  After reaching IN_CHARGE, trustee
3243 *		tries to execute the pending worklist until it's empty
3244 *		and the state is set to BUTCHER, or the state is set
3245 *		to RELEASE.
3246 *
3247 * BUTCHER	Command state which is set by the cpu callback after
3248 *		the cpu has went down.  Once this state is set trustee
3249 *		knows that there will be no new works on the worklist
3250 *		and once the worklist is empty it can proceed to
3251 *		killing idle workers.
3252 *
3253 * RELEASE	Command state which is set by the cpu callback if the
3254 *		cpu down has been canceled or it has come online
3255 *		again.  After recognizing this state, trustee stops
3256 *		trying to drain or butcher and clears ROGUE, rebinds
3257 *		all remaining workers back to the cpu and releases
3258 *		manager role.
3259 *
3260 * DONE		Trustee will enter this state after BUTCHER or RELEASE
3261 *		is complete.
3262 *
3263 *          trustee                 CPU                draining
3264 *         took over                down               complete
3265 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3266 *                        |                     |                  ^
3267 *                        | CPU is back online  v   return workers |
3268 *                         ----------------> RELEASE --------------
3269 */
 
 
 
 
3270
3271/**
3272 * trustee_wait_event_timeout - timed event wait for trustee
3273 * @cond: condition to wait for
3274 * @timeout: timeout in jiffies
3275 *
3276 * wait_event_timeout() for trustee to use.  Handles locking and
3277 * checks for RELEASE request.
3278 *
3279 * CONTEXT:
3280 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3281 * multiple times.  To be used by trustee.
3282 *
3283 * RETURNS:
3284 * Positive indicating left time if @cond is satisfied, 0 if timed
3285 * out, -1 if canceled.
3286 */
3287#define trustee_wait_event_timeout(cond, timeout) ({			\
3288	long __ret = (timeout);						\
3289	while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) &&	\
3290	       __ret) {							\
3291		spin_unlock_irq(&gcwq->lock);				\
3292		__wait_event_timeout(gcwq->trustee_wait, (cond) ||	\
3293			(gcwq->trustee_state == TRUSTEE_RELEASE),	\
3294			__ret);						\
3295		spin_lock_irq(&gcwq->lock);				\
3296	}								\
3297	gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret);		\
3298})
3299
3300/**
3301 * trustee_wait_event - event wait for trustee
3302 * @cond: condition to wait for
 
3303 *
3304 * wait_event() for trustee to use.  Automatically handles locking and
3305 * checks for CANCEL request.
 
3306 *
3307 * CONTEXT:
3308 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3309 * multiple times.  To be used by trustee.
3310 *
3311 * RETURNS:
3312 * 0 if @cond is satisfied, -1 if canceled.
3313 */
3314#define trustee_wait_event(cond) ({					\
3315	long __ret1;							\
3316	__ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3317	__ret1 < 0 ? -1 : 0;						\
3318})
3319
3320static int __cpuinit trustee_thread(void *__gcwq)
3321{
3322	struct global_cwq *gcwq = __gcwq;
 
 
 
 
 
3323	struct worker *worker;
3324	struct work_struct *work;
3325	struct hlist_node *pos;
3326	long rc;
3327	int i;
3328
3329	BUG_ON(gcwq->cpu != smp_processor_id());
 
3330
3331	spin_lock_irq(&gcwq->lock);
3332	/*
3333	 * Claim the manager position and make all workers rogue.
3334	 * Trustee must be bound to the target cpu and can't be
3335	 * cancelled.
3336	 */
3337	BUG_ON(gcwq->cpu != smp_processor_id());
3338	rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3339	BUG_ON(rc < 0);
3340
3341	gcwq->flags |= GCWQ_MANAGING_WORKERS;
 
 
 
 
 
 
 
3342
3343	list_for_each_entry(worker, &gcwq->idle_list, entry)
3344		worker->flags |= WORKER_ROGUE;
 
 
3345
3346	for_each_busy_worker(worker, i, pos, gcwq)
3347		worker->flags |= WORKER_ROGUE;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3348
3349	/*
3350	 * Call schedule() so that we cross rq->lock and thus can
3351	 * guarantee sched callbacks see the rogue flag.  This is
3352	 * necessary as scheduler callbacks may be invoked from other
3353	 * cpus.
3354	 */
3355	spin_unlock_irq(&gcwq->lock);
3356	schedule();
3357	spin_lock_irq(&gcwq->lock);
3358
3359	/*
3360	 * Sched callbacks are disabled now.  Zap nr_running.  After
3361	 * this, nr_running stays zero and need_more_worker() and
3362	 * keep_working() are always true as long as the worklist is
3363	 * not empty.
3364	 */
3365	atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
3366
3367	spin_unlock_irq(&gcwq->lock);
3368	del_timer_sync(&gcwq->idle_timer);
3369	spin_lock_irq(&gcwq->lock);
3370
3371	/*
3372	 * We're now in charge.  Notify and proceed to drain.  We need
3373	 * to keep the gcwq running during the whole CPU down
3374	 * procedure as other cpu hotunplug callbacks may need to
3375	 * flush currently running tasks.
3376	 */
3377	gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3378	wake_up_all(&gcwq->trustee_wait);
3379
3380	/*
3381	 * The original cpu is in the process of dying and may go away
3382	 * anytime now.  When that happens, we and all workers would
3383	 * be migrated to other cpus.  Try draining any left work.  We
3384	 * want to get it over with ASAP - spam rescuers, wake up as
3385	 * many idlers as necessary and create new ones till the
3386	 * worklist is empty.  Note that if the gcwq is frozen, there
3387	 * may be frozen works in freezable cwqs.  Don't declare
3388	 * completion while frozen.
3389	 */
3390	while (gcwq->nr_workers != gcwq->nr_idle ||
3391	       gcwq->flags & GCWQ_FREEZING ||
3392	       gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
3393		int nr_works = 0;
3394
3395		list_for_each_entry(work, &gcwq->worklist, entry) {
3396			send_mayday(work);
3397			nr_works++;
 
 
 
 
3398		}
 
 
3399
3400		list_for_each_entry(worker, &gcwq->idle_list, entry) {
3401			if (!nr_works--)
3402				break;
3403			wake_up_process(worker->task);
3404		}
 
 
 
 
 
 
 
 
3405
3406		if (need_to_create_worker(gcwq)) {
3407			spin_unlock_irq(&gcwq->lock);
3408			worker = create_worker(gcwq, false);
3409			spin_lock_irq(&gcwq->lock);
3410			if (worker) {
3411				worker->flags |= WORKER_ROGUE;
3412				start_worker(worker);
3413			}
3414		}
 
 
3415
3416		/* give a breather */
3417		if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3418			break;
 
 
 
 
 
 
3419	}
 
3420
3421	/*
3422	 * Either all works have been scheduled and cpu is down, or
3423	 * cpu down has already been canceled.  Wait for and butcher
3424	 * all workers till we're canceled.
3425	 */
3426	do {
3427		rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3428		while (!list_empty(&gcwq->idle_list))
3429			destroy_worker(list_first_entry(&gcwq->idle_list,
3430							struct worker, entry));
3431	} while (gcwq->nr_workers && rc >= 0);
 
3432
3433	/*
3434	 * At this point, either draining has completed and no worker
3435	 * is left, or cpu down has been canceled or the cpu is being
3436	 * brought back up.  There shouldn't be any idle one left.
3437	 * Tell the remaining busy ones to rebind once it finishes the
3438	 * currently scheduled works by scheduling the rebind_work.
3439	 */
3440	WARN_ON(!list_empty(&gcwq->idle_list));
3441
3442	for_each_busy_worker(worker, i, pos, gcwq) {
3443		struct work_struct *rebind_work = &worker->rebind_work;
3444		unsigned long worker_flags = worker->flags;
3445
3446		/*
3447		 * Rebind_work may race with future cpu hotplug
3448		 * operations.  Use a separate flag to mark that
3449		 * rebinding is scheduled.  The morphing should
3450		 * be atomic.
3451		 */
3452		worker_flags |= WORKER_REBIND;
3453		worker_flags &= ~WORKER_ROGUE;
3454		ACCESS_ONCE(worker->flags) = worker_flags;
3455
3456		/* queue rebind_work, wq doesn't matter, use the default one */
3457		if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3458				     work_data_bits(rebind_work)))
 
 
 
 
3459			continue;
3460
3461		debug_work_activate(rebind_work);
3462		insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
3463			    worker->scheduled.next,
3464			    work_color_to_flags(WORK_NO_COLOR));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3465	}
3466
3467	/* relinquish manager role */
3468	gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3469
3470	/* notify completion */
3471	gcwq->trustee = NULL;
3472	gcwq->trustee_state = TRUSTEE_DONE;
3473	wake_up_all(&gcwq->trustee_wait);
3474	spin_unlock_irq(&gcwq->lock);
3475	return 0;
3476}
3477
3478/**
3479 * wait_trustee_state - wait for trustee to enter the specified state
3480 * @gcwq: gcwq the trustee of interest belongs to
3481 * @state: target state to wait for
3482 *
3483 * Wait for the trustee to reach @state.  DONE is already matched.
 
 
 
 
 
3484 *
3485 * CONTEXT:
3486 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3487 * multiple times.  To be used by cpu_callback.
3488 */
3489static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
3490__releases(&gcwq->lock)
3491__acquires(&gcwq->lock)
3492{
3493	if (!(gcwq->trustee_state == state ||
3494	      gcwq->trustee_state == TRUSTEE_DONE)) {
3495		spin_unlock_irq(&gcwq->lock);
3496		__wait_event(gcwq->trustee_wait,
3497			     gcwq->trustee_state == state ||
3498			     gcwq->trustee_state == TRUSTEE_DONE);
3499		spin_lock_irq(&gcwq->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3500	}
3501}
3502
3503static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3504						unsigned long action,
3505						void *hcpu)
3506{
3507	unsigned int cpu = (unsigned long)hcpu;
3508	struct global_cwq *gcwq = get_gcwq(cpu);
3509	struct task_struct *new_trustee = NULL;
3510	struct worker *uninitialized_var(new_worker);
3511	unsigned long flags;
 
 
 
 
 
 
 
 
 
 
 
 
 
3512
3513	action &= ~CPU_TASKS_FROZEN;
3514
3515	switch (action) {
3516	case CPU_DOWN_PREPARE:
3517		new_trustee = kthread_create(trustee_thread, gcwq,
3518					     "workqueue_trustee/%d\n", cpu);
3519		if (IS_ERR(new_trustee))
3520			return notifier_from_errno(PTR_ERR(new_trustee));
3521		kthread_bind(new_trustee, cpu);
3522		/* fall through */
3523	case CPU_UP_PREPARE:
3524		BUG_ON(gcwq->first_idle);
3525		new_worker = create_worker(gcwq, false);
3526		if (!new_worker) {
3527			if (new_trustee)
3528				kthread_stop(new_trustee);
3529			return NOTIFY_BAD;
3530		}
3531	}
3532
3533	/* some are called w/ irq disabled, don't disturb irq status */
3534	spin_lock_irqsave(&gcwq->lock, flags);
 
 
3535
3536	switch (action) {
3537	case CPU_DOWN_PREPARE:
3538		/* initialize trustee and tell it to acquire the gcwq */
3539		BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3540		gcwq->trustee = new_trustee;
3541		gcwq->trustee_state = TRUSTEE_START;
3542		wake_up_process(gcwq->trustee);
3543		wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
3544		/* fall through */
3545	case CPU_UP_PREPARE:
3546		BUG_ON(gcwq->first_idle);
3547		gcwq->first_idle = new_worker;
3548		break;
3549
3550	case CPU_DYING:
3551		/*
3552		 * Before this, the trustee and all workers except for
3553		 * the ones which are still executing works from
3554		 * before the last CPU down must be on the cpu.  After
3555		 * this, they'll all be diasporas.
 
 
 
 
 
 
 
 
 
3556		 */
3557		gcwq->flags |= GCWQ_DISASSOCIATED;
3558		break;
 
 
 
 
 
 
3559
3560	case CPU_POST_DEAD:
3561		gcwq->trustee_state = TRUSTEE_BUTCHER;
3562		/* fall through */
3563	case CPU_UP_CANCELED:
3564		destroy_worker(gcwq->first_idle);
3565		gcwq->first_idle = NULL;
3566		break;
 
 
 
 
 
 
 
3567
3568	case CPU_DOWN_FAILED:
3569	case CPU_ONLINE:
3570		gcwq->flags &= ~GCWQ_DISASSOCIATED;
3571		if (gcwq->trustee_state != TRUSTEE_DONE) {
3572			gcwq->trustee_state = TRUSTEE_RELEASE;
3573			wake_up_process(gcwq->trustee);
3574			wait_trustee_state(gcwq, TRUSTEE_DONE);
3575		}
3576
3577		/*
3578		 * Trustee is done and there might be no worker left.
3579		 * Put the first_idle in and request a real manager to
3580		 * take a look.
3581		 */
3582		spin_unlock_irq(&gcwq->lock);
3583		kthread_bind(gcwq->first_idle->task, cpu);
3584		spin_lock_irq(&gcwq->lock);
3585		gcwq->flags |= GCWQ_MANAGE_WORKERS;
3586		start_worker(gcwq->first_idle);
3587		gcwq->first_idle = NULL;
3588		break;
3589	}
3590
3591	spin_unlock_irqrestore(&gcwq->lock, flags);
 
 
 
3592
3593	return notifier_from_errno(0);
 
 
 
3594}
3595
3596/*
3597 * Workqueues should be brought up before normal priority CPU notifiers.
3598 * This will be registered high priority CPU notifier.
3599 */
3600static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3601					       unsigned long action,
3602					       void *hcpu)
3603{
 
 
 
 
 
3604	switch (action & ~CPU_TASKS_FROZEN) {
3605	case CPU_UP_PREPARE:
3606	case CPU_UP_CANCELED:
 
 
 
 
 
 
 
3607	case CPU_DOWN_FAILED:
3608	case CPU_ONLINE:
3609		return workqueue_cpu_callback(nfb, action, hcpu);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3610	}
3611	return NOTIFY_OK;
3612}
3613
3614/*
3615 * Workqueues should be brought down after normal priority CPU notifiers.
3616 * This will be registered as low priority CPU notifier.
3617 */
3618static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3619						 unsigned long action,
3620						 void *hcpu)
3621{
 
 
 
 
3622	switch (action & ~CPU_TASKS_FROZEN) {
3623	case CPU_DOWN_PREPARE:
3624	case CPU_DYING:
3625	case CPU_POST_DEAD:
3626		return workqueue_cpu_callback(nfb, action, hcpu);
 
 
 
 
 
 
 
 
 
 
 
3627	}
3628	return NOTIFY_OK;
3629}
3630
3631#ifdef CONFIG_SMP
3632
3633struct work_for_cpu {
3634	struct work_struct work;
3635	long (*fn)(void *);
3636	void *arg;
3637	long ret;
3638};
3639
3640static void work_for_cpu_fn(struct work_struct *work)
3641{
3642	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
3643
3644	wfc->ret = wfc->fn(wfc->arg);
3645}
3646
3647/**
3648 * work_on_cpu - run a function in user context on a particular cpu
3649 * @cpu: the cpu to run on
3650 * @fn: the function to run
3651 * @arg: the function arg
3652 *
3653 * This will return the value @fn returns.
3654 * It is up to the caller to ensure that the cpu doesn't go offline.
3655 * The caller must not hold any locks which would prevent @fn from completing.
 
 
3656 */
3657long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3658{
3659	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
3660
3661	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
3662	schedule_work_on(cpu, &wfc.work);
3663	flush_work(&wfc.work);
 
3664	return wfc.ret;
3665}
3666EXPORT_SYMBOL_GPL(work_on_cpu);
3667#endif /* CONFIG_SMP */
3668
3669#ifdef CONFIG_FREEZER
3670
3671/**
3672 * freeze_workqueues_begin - begin freezing workqueues
3673 *
3674 * Start freezing workqueues.  After this function returns, all freezable
3675 * workqueues will queue new works to their frozen_works list instead of
3676 * gcwq->worklist.
3677 *
3678 * CONTEXT:
3679 * Grabs and releases workqueue_lock and gcwq->lock's.
3680 */
3681void freeze_workqueues_begin(void)
3682{
3683	unsigned int cpu;
 
3684
3685	spin_lock(&workqueue_lock);
3686
3687	BUG_ON(workqueue_freezing);
3688	workqueue_freezing = true;
3689
3690	for_each_gcwq_cpu(cpu) {
3691		struct global_cwq *gcwq = get_gcwq(cpu);
3692		struct workqueue_struct *wq;
3693
3694		spin_lock_irq(&gcwq->lock);
3695
3696		BUG_ON(gcwq->flags & GCWQ_FREEZING);
3697		gcwq->flags |= GCWQ_FREEZING;
3698
3699		list_for_each_entry(wq, &workqueues, list) {
3700			struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3701
3702			if (cwq && wq->flags & WQ_FREEZABLE)
3703				cwq->max_active = 0;
3704		}
3705
3706		spin_unlock_irq(&gcwq->lock);
3707	}
3708
3709	spin_unlock(&workqueue_lock);
3710}
3711
3712/**
3713 * freeze_workqueues_busy - are freezable workqueues still busy?
3714 *
3715 * Check whether freezing is complete.  This function must be called
3716 * between freeze_workqueues_begin() and thaw_workqueues().
3717 *
3718 * CONTEXT:
3719 * Grabs and releases workqueue_lock.
3720 *
3721 * RETURNS:
3722 * %true if some freezable workqueues are still busy.  %false if freezing
3723 * is complete.
3724 */
3725bool freeze_workqueues_busy(void)
3726{
3727	unsigned int cpu;
3728	bool busy = false;
 
 
3729
3730	spin_lock(&workqueue_lock);
3731
3732	BUG_ON(!workqueue_freezing);
3733
3734	for_each_gcwq_cpu(cpu) {
3735		struct workqueue_struct *wq;
 
3736		/*
3737		 * nr_active is monotonically decreasing.  It's safe
3738		 * to peek without lock.
3739		 */
3740		list_for_each_entry(wq, &workqueues, list) {
3741			struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3742
3743			if (!cwq || !(wq->flags & WQ_FREEZABLE))
3744				continue;
3745
3746			BUG_ON(cwq->nr_active < 0);
3747			if (cwq->nr_active) {
3748				busy = true;
 
3749				goto out_unlock;
3750			}
3751		}
 
3752	}
3753out_unlock:
3754	spin_unlock(&workqueue_lock);
3755	return busy;
3756}
3757
3758/**
3759 * thaw_workqueues - thaw workqueues
3760 *
3761 * Thaw workqueues.  Normal queueing is restored and all collected
3762 * frozen works are transferred to their respective gcwq worklists.
3763 *
3764 * CONTEXT:
3765 * Grabs and releases workqueue_lock and gcwq->lock's.
3766 */
3767void thaw_workqueues(void)
3768{
3769	unsigned int cpu;
 
3770
3771	spin_lock(&workqueue_lock);
3772
3773	if (!workqueue_freezing)
3774		goto out_unlock;
3775
3776	for_each_gcwq_cpu(cpu) {
3777		struct global_cwq *gcwq = get_gcwq(cpu);
3778		struct workqueue_struct *wq;
3779
3780		spin_lock_irq(&gcwq->lock);
 
 
 
 
 
 
3781
3782		BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3783		gcwq->flags &= ~GCWQ_FREEZING;
 
 
3784
3785		list_for_each_entry(wq, &workqueues, list) {
3786			struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
 
 
 
 
3787
3788			if (!cwq || !(wq->flags & WQ_FREEZABLE))
3789				continue;
3790
3791			/* restore max_active and repopulate worklist */
3792			cwq->max_active = wq->saved_max_active;
 
 
 
 
3793
3794			while (!list_empty(&cwq->delayed_works) &&
3795			       cwq->nr_active < cwq->max_active)
3796				cwq_activate_first_delayed(cwq);
 
3797		}
3798
3799		wake_up_worker(gcwq);
 
3800
3801		spin_unlock_irq(&gcwq->lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3802	}
3803
3804	workqueue_freezing = false;
3805out_unlock:
3806	spin_unlock(&workqueue_lock);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3807}
3808#endif /* CONFIG_FREEZER */
3809
3810static int __init init_workqueues(void)
3811{
3812	unsigned int cpu;
3813	int i;
3814
3815	cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3816	cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
3817
3818	/* initialize gcwqs */
3819	for_each_gcwq_cpu(cpu) {
3820		struct global_cwq *gcwq = get_gcwq(cpu);
3821
3822		spin_lock_init(&gcwq->lock);
3823		INIT_LIST_HEAD(&gcwq->worklist);
3824		gcwq->cpu = cpu;
3825		gcwq->flags |= GCWQ_DISASSOCIATED;
3826
3827		INIT_LIST_HEAD(&gcwq->idle_list);
3828		for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3829			INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3830
3831		init_timer_deferrable(&gcwq->idle_timer);
3832		gcwq->idle_timer.function = idle_worker_timeout;
3833		gcwq->idle_timer.data = (unsigned long)gcwq;
3834
3835		setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3836			    (unsigned long)gcwq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3837
3838		ida_init(&gcwq->worker_ida);
 
 
3839
3840		gcwq->trustee_state = TRUSTEE_DONE;
3841		init_waitqueue_head(&gcwq->trustee_wait);
 
 
3842	}
3843
3844	/* create the initial worker */
3845	for_each_online_gcwq_cpu(cpu) {
3846		struct global_cwq *gcwq = get_gcwq(cpu);
3847		struct worker *worker;
 
 
 
3848
3849		if (cpu != WORK_CPU_UNBOUND)
3850			gcwq->flags &= ~GCWQ_DISASSOCIATED;
3851		worker = create_worker(gcwq, true);
3852		BUG_ON(!worker);
3853		spin_lock_irq(&gcwq->lock);
3854		start_worker(worker);
3855		spin_unlock_irq(&gcwq->lock);
 
 
3856	}
3857
3858	system_wq = alloc_workqueue("events", 0, 0);
 
3859	system_long_wq = alloc_workqueue("events_long", 0, 0);
3860	system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
3861	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3862					    WQ_UNBOUND_MAX_ACTIVE);
3863	system_freezable_wq = alloc_workqueue("events_freezable",
3864					      WQ_FREEZABLE, 0);
3865	system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3866			WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
3867	BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
 
 
 
3868	       !system_unbound_wq || !system_freezable_wq ||
3869		!system_nrt_freezable_wq);
 
 
 
 
3870	return 0;
3871}
3872early_initcall(init_workqueues);