Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Apr 14-17, 2025
Register
Loading...
v5.4
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * kernel/workqueue.c - generic async execution with shared worker pool
   4 *
   5 * Copyright (C) 2002		Ingo Molnar
   6 *
   7 *   Derived from the taskqueue/keventd code by:
   8 *     David Woodhouse <dwmw2@infradead.org>
   9 *     Andrew Morton
  10 *     Kai Petzke <wpp@marie.physik.tu-berlin.de>
  11 *     Theodore Ts'o <tytso@mit.edu>
  12 *
  13 * Made to use alloc_percpu by Christoph Lameter.
  14 *
  15 * Copyright (C) 2010		SUSE Linux Products GmbH
  16 * Copyright (C) 2010		Tejun Heo <tj@kernel.org>
  17 *
  18 * This is the generic async execution mechanism.  Work items as are
  19 * executed in process context.  The worker pool is shared and
  20 * automatically managed.  There are two worker pools for each CPU (one for
  21 * normal work items and the other for high priority ones) and some extra
  22 * pools for workqueues which are not bound to any specific CPU - the
  23 * number of these backing pools is dynamic.
  24 *
  25 * Please read Documentation/core-api/workqueue.rst for details.
  26 */
  27
  28#include <linux/export.h>
  29#include <linux/kernel.h>
  30#include <linux/sched.h>
  31#include <linux/init.h>
  32#include <linux/signal.h>
  33#include <linux/completion.h>
  34#include <linux/workqueue.h>
  35#include <linux/slab.h>
  36#include <linux/cpu.h>
  37#include <linux/notifier.h>
  38#include <linux/kthread.h>
  39#include <linux/hardirq.h>
  40#include <linux/mempolicy.h>
  41#include <linux/freezer.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#include <linux/sched/isolation.h>
  52#include <linux/nmi.h>
  53
  54#include "workqueue_internal.h"
  55
  56enum {
  57	/*
  58	 * worker_pool flags
  59	 *
  60	 * A bound pool is either associated or disassociated with its CPU.
  61	 * While associated (!DISASSOCIATED), all workers are bound to the
  62	 * CPU and none has %WORKER_UNBOUND set and concurrency management
  63	 * is in effect.
  64	 *
  65	 * While DISASSOCIATED, the cpu may be offline and all workers have
  66	 * %WORKER_UNBOUND set and concurrency management disabled, and may
  67	 * be executing on any CPU.  The pool behaves as an unbound one.
  68	 *
  69	 * Note that DISASSOCIATED should be flipped only while holding
  70	 * wq_pool_attach_mutex to avoid changing binding state while
  71	 * worker_attach_to_pool() is in progress.
  72	 */
  73	POOL_MANAGER_ACTIVE	= 1 << 0,	/* being managed */
  74	POOL_DISASSOCIATED	= 1 << 2,	/* cpu can't serve workers */
  75
  76	/* worker flags */
  77	WORKER_DIE		= 1 << 1,	/* die die die */
  78	WORKER_IDLE		= 1 << 2,	/* is idle */
  79	WORKER_PREP		= 1 << 3,	/* preparing to run works */
  80	WORKER_CPU_INTENSIVE	= 1 << 6,	/* cpu intensive */
  81	WORKER_UNBOUND		= 1 << 7,	/* worker is unbound */
  82	WORKER_REBOUND		= 1 << 8,	/* worker was rebound */
  83
  84	WORKER_NOT_RUNNING	= WORKER_PREP | WORKER_CPU_INTENSIVE |
  85				  WORKER_UNBOUND | WORKER_REBOUND,
  86
  87	NR_STD_WORKER_POOLS	= 2,		/* # standard pools per cpu */
  88
  89	UNBOUND_POOL_HASH_ORDER	= 6,		/* hashed by pool->attrs */
  90	BUSY_WORKER_HASH_ORDER	= 6,		/* 64 pointers */
  91
  92	MAX_IDLE_WORKERS_RATIO	= 4,		/* 1/4 of busy can be idle */
  93	IDLE_WORKER_TIMEOUT	= 300 * HZ,	/* keep idle ones for 5 mins */
  94
  95	MAYDAY_INITIAL_TIMEOUT  = HZ / 100 >= 2 ? HZ / 100 : 2,
  96						/* call for help after 10ms
  97						   (min two ticks) */
  98	MAYDAY_INTERVAL		= HZ / 10,	/* and then every 100ms */
  99	CREATE_COOLDOWN		= HZ,		/* time to breath after fail */
 100
 101	/*
 102	 * Rescue workers are used only on emergencies and shared by
 103	 * all cpus.  Give MIN_NICE.
 104	 */
 105	RESCUER_NICE_LEVEL	= MIN_NICE,
 106	HIGHPRI_NICE_LEVEL	= MIN_NICE,
 107
 108	WQ_NAME_LEN		= 24,
 109};
 110
 111/*
 112 * Structure fields follow one of the following exclusion rules.
 113 *
 114 * I: Modifiable by initialization/destruction paths and read-only for
 115 *    everyone else.
 116 *
 117 * P: Preemption protected.  Disabling preemption is enough and should
 118 *    only be modified and accessed from the local cpu.
 119 *
 120 * L: pool->lock protected.  Access with pool->lock held.
 121 *
 122 * X: During normal operation, modification requires pool->lock and should
 123 *    be done only from local cpu.  Either disabling preemption on local
 124 *    cpu or grabbing pool->lock is enough for read access.  If
 125 *    POOL_DISASSOCIATED is set, it's identical to L.
 126 *
 127 * A: wq_pool_attach_mutex protected.
 128 *
 129 * PL: wq_pool_mutex protected.
 130 *
 131 * PR: wq_pool_mutex protected for writes.  RCU protected for reads.
 132 *
 133 * PW: wq_pool_mutex and wq->mutex protected for writes.  Either for reads.
 134 *
 135 * PWR: wq_pool_mutex and wq->mutex protected for writes.  Either or
 136 *      RCU for reads.
 137 *
 138 * WQ: wq->mutex protected.
 139 *
 140 * WR: wq->mutex protected for writes.  RCU protected for reads.
 141 *
 142 * MD: wq_mayday_lock protected.
 143 */
 144
 145/* struct worker is defined in workqueue_internal.h */
 146
 147struct worker_pool {
 148	spinlock_t		lock;		/* the pool lock */
 149	int			cpu;		/* I: the associated cpu */
 150	int			node;		/* I: the associated node ID */
 151	int			id;		/* I: pool ID */
 152	unsigned int		flags;		/* X: flags */
 153
 154	unsigned long		watchdog_ts;	/* L: watchdog timestamp */
 155
 156	struct list_head	worklist;	/* L: list of pending works */
 157
 158	int			nr_workers;	/* L: total number of workers */
 159	int			nr_idle;	/* L: currently idle workers */
 
 
 160
 161	struct list_head	idle_list;	/* X: list of idle workers */
 162	struct timer_list	idle_timer;	/* L: worker idle timeout */
 163	struct timer_list	mayday_timer;	/* L: SOS timer for workers */
 164
 165	/* a workers is either on busy_hash or idle_list, or the manager */
 166	DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER);
 167						/* L: hash of busy workers */
 168
 
 
 169	struct worker		*manager;	/* L: purely informational */
 
 170	struct list_head	workers;	/* A: attached workers */
 171	struct completion	*detach_completion; /* all workers detached */
 172
 173	struct ida		worker_ida;	/* worker IDs for task name */
 174
 175	struct workqueue_attrs	*attrs;		/* I: worker attributes */
 176	struct hlist_node	hash_node;	/* PL: unbound_pool_hash node */
 177	int			refcnt;		/* PL: refcnt for unbound pools */
 178
 179	/*
 180	 * The current concurrency level.  As it's likely to be accessed
 181	 * from other CPUs during try_to_wake_up(), put it in a separate
 182	 * cacheline.
 183	 */
 184	atomic_t		nr_running ____cacheline_aligned_in_smp;
 185
 186	/*
 187	 * Destruction of pool is RCU protected to allow dereferences
 188	 * from get_work_pool().
 189	 */
 190	struct rcu_head		rcu;
 191} ____cacheline_aligned_in_smp;
 192
 193/*
 194 * The per-pool workqueue.  While queued, the lower WORK_STRUCT_FLAG_BITS
 195 * of work_struct->data are used for flags and the remaining high bits
 196 * point to the pwq; thus, pwqs need to be aligned at two's power of the
 197 * number of flag bits.
 198 */
 199struct pool_workqueue {
 200	struct worker_pool	*pool;		/* I: the associated pool */
 201	struct workqueue_struct *wq;		/* I: the owning workqueue */
 202	int			work_color;	/* L: current color */
 203	int			flush_color;	/* L: flushing color */
 204	int			refcnt;		/* L: reference count */
 205	int			nr_in_flight[WORK_NR_COLORS];
 206						/* L: nr of in_flight works */
 207	int			nr_active;	/* L: nr of active works */
 208	int			max_active;	/* L: max active works */
 209	struct list_head	delayed_works;	/* L: delayed works */
 210	struct list_head	pwqs_node;	/* WR: node on wq->pwqs */
 211	struct list_head	mayday_node;	/* MD: node on wq->maydays */
 212
 213	/*
 214	 * Release of unbound pwq is punted to system_wq.  See put_pwq()
 215	 * and pwq_unbound_release_workfn() for details.  pool_workqueue
 216	 * itself is also RCU protected so that the first pwq can be
 217	 * determined without grabbing wq->mutex.
 218	 */
 219	struct work_struct	unbound_release_work;
 220	struct rcu_head		rcu;
 221} __aligned(1 << WORK_STRUCT_FLAG_BITS);
 222
 223/*
 224 * Structure used to wait for workqueue flush.
 225 */
 226struct wq_flusher {
 227	struct list_head	list;		/* WQ: list of flushers */
 228	int			flush_color;	/* WQ: flush color waiting for */
 229	struct completion	done;		/* flush completion */
 230};
 231
 232struct wq_device;
 233
 234/*
 235 * The externally visible workqueue.  It relays the issued work items to
 236 * the appropriate worker_pool through its pool_workqueues.
 237 */
 238struct workqueue_struct {
 239	struct list_head	pwqs;		/* WR: all pwqs of this wq */
 240	struct list_head	list;		/* PR: list of all workqueues */
 241
 242	struct mutex		mutex;		/* protects this wq */
 243	int			work_color;	/* WQ: current work color */
 244	int			flush_color;	/* WQ: current flush color */
 245	atomic_t		nr_pwqs_to_flush; /* flush in progress */
 246	struct wq_flusher	*first_flusher;	/* WQ: first flusher */
 247	struct list_head	flusher_queue;	/* WQ: flush waiters */
 248	struct list_head	flusher_overflow; /* WQ: flush overflow list */
 249
 250	struct list_head	maydays;	/* MD: pwqs requesting rescue */
 251	struct worker		*rescuer;	/* I: rescue worker */
 252
 253	int			nr_drainers;	/* WQ: drain in progress */
 254	int			saved_max_active; /* WQ: saved pwq max_active */
 255
 256	struct workqueue_attrs	*unbound_attrs;	/* PW: only for unbound wqs */
 257	struct pool_workqueue	*dfl_pwq;	/* PW: only for unbound wqs */
 258
 259#ifdef CONFIG_SYSFS
 260	struct wq_device	*wq_dev;	/* I: for sysfs interface */
 261#endif
 262#ifdef CONFIG_LOCKDEP
 263	char			*lock_name;
 264	struct lock_class_key	key;
 265	struct lockdep_map	lockdep_map;
 266#endif
 267	char			name[WQ_NAME_LEN]; /* I: workqueue name */
 268
 269	/*
 270	 * Destruction of workqueue_struct is RCU protected to allow walking
 271	 * the workqueues list without grabbing wq_pool_mutex.
 272	 * This is used to dump all workqueues from sysrq.
 273	 */
 274	struct rcu_head		rcu;
 275
 276	/* hot fields used during command issue, aligned to cacheline */
 277	unsigned int		flags ____cacheline_aligned; /* WQ: WQ_* flags */
 278	struct pool_workqueue __percpu *cpu_pwqs; /* I: per-cpu pwqs */
 279	struct pool_workqueue __rcu *numa_pwq_tbl[]; /* PWR: unbound pwqs indexed by node */
 280};
 281
 282static struct kmem_cache *pwq_cache;
 283
 284static cpumask_var_t *wq_numa_possible_cpumask;
 285					/* possible CPUs of each node */
 286
 287static bool wq_disable_numa;
 288module_param_named(disable_numa, wq_disable_numa, bool, 0444);
 289
 290/* see the comment above the definition of WQ_POWER_EFFICIENT */
 291static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT);
 292module_param_named(power_efficient, wq_power_efficient, bool, 0444);
 293
 294static bool wq_online;			/* can kworkers be created yet? */
 295
 296static bool wq_numa_enabled;		/* unbound NUMA affinity enabled */
 297
 298/* buf for wq_update_unbound_numa_attrs(), protected by CPU hotplug exclusion */
 299static struct workqueue_attrs *wq_update_unbound_numa_attrs_buf;
 300
 301static DEFINE_MUTEX(wq_pool_mutex);	/* protects pools and workqueues list */
 302static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */
 303static DEFINE_SPINLOCK(wq_mayday_lock);	/* protects wq->maydays list */
 304static DECLARE_WAIT_QUEUE_HEAD(wq_manager_wait); /* wait for manager to go away */
 305
 306static LIST_HEAD(workqueues);		/* PR: list of all workqueues */
 307static bool workqueue_freezing;		/* PL: have wqs started freezing? */
 308
 309/* PL: allowable cpus for unbound wqs and work items */
 310static cpumask_var_t wq_unbound_cpumask;
 311
 312/* CPU where unbound work was last round robin scheduled from this CPU */
 313static DEFINE_PER_CPU(int, wq_rr_cpu_last);
 314
 315/*
 316 * Local execution of unbound work items is no longer guaranteed.  The
 317 * following always forces round-robin CPU selection on unbound work items
 318 * to uncover usages which depend on it.
 319 */
 320#ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU
 321static bool wq_debug_force_rr_cpu = true;
 322#else
 323static bool wq_debug_force_rr_cpu = false;
 324#endif
 325module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644);
 326
 327/* the per-cpu worker pools */
 328static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools);
 329
 330static DEFINE_IDR(worker_pool_idr);	/* PR: idr of all pools */
 331
 332/* PL: hash of all unbound pools keyed by pool->attrs */
 333static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER);
 334
 335/* I: attributes used when instantiating standard unbound pools on demand */
 336static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS];
 337
 338/* I: attributes used when instantiating ordered pools on demand */
 339static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS];
 340
 341struct workqueue_struct *system_wq __read_mostly;
 342EXPORT_SYMBOL(system_wq);
 343struct workqueue_struct *system_highpri_wq __read_mostly;
 344EXPORT_SYMBOL_GPL(system_highpri_wq);
 345struct workqueue_struct *system_long_wq __read_mostly;
 346EXPORT_SYMBOL_GPL(system_long_wq);
 347struct workqueue_struct *system_unbound_wq __read_mostly;
 348EXPORT_SYMBOL_GPL(system_unbound_wq);
 349struct workqueue_struct *system_freezable_wq __read_mostly;
 350EXPORT_SYMBOL_GPL(system_freezable_wq);
 351struct workqueue_struct *system_power_efficient_wq __read_mostly;
 352EXPORT_SYMBOL_GPL(system_power_efficient_wq);
 353struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly;
 354EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq);
 355
 356static int worker_thread(void *__worker);
 357static void workqueue_sysfs_unregister(struct workqueue_struct *wq);
 358
 359#define CREATE_TRACE_POINTS
 360#include <trace/events/workqueue.h>
 361
 362#define assert_rcu_or_pool_mutex()					\
 363	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
 364			 !lockdep_is_held(&wq_pool_mutex),		\
 365			 "RCU or wq_pool_mutex should be held")
 366
 367#define assert_rcu_or_wq_mutex(wq)					\
 368	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
 369			 !lockdep_is_held(&wq->mutex),			\
 370			 "RCU or wq->mutex should be held")
 371
 372#define assert_rcu_or_wq_mutex_or_pool_mutex(wq)			\
 373	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&			\
 374			 !lockdep_is_held(&wq->mutex) &&		\
 375			 !lockdep_is_held(&wq_pool_mutex),		\
 376			 "RCU, wq->mutex or wq_pool_mutex should be held")
 377
 378#define for_each_cpu_worker_pool(pool, cpu)				\
 379	for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0];		\
 380	     (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \
 381	     (pool)++)
 382
 383/**
 384 * for_each_pool - iterate through all worker_pools in the system
 385 * @pool: iteration cursor
 386 * @pi: integer used for iteration
 387 *
 388 * This must be called either with wq_pool_mutex held or RCU read
 389 * locked.  If the pool needs to be used beyond the locking in effect, the
 390 * caller is responsible for guaranteeing that the pool stays online.
 391 *
 392 * The if/else clause exists only for the lockdep assertion and can be
 393 * ignored.
 394 */
 395#define for_each_pool(pool, pi)						\
 396	idr_for_each_entry(&worker_pool_idr, pool, pi)			\
 397		if (({ assert_rcu_or_pool_mutex(); false; })) { }	\
 398		else
 399
 400/**
 401 * for_each_pool_worker - iterate through all workers of a worker_pool
 402 * @worker: iteration cursor
 403 * @pool: worker_pool to iterate workers of
 404 *
 405 * This must be called with wq_pool_attach_mutex.
 406 *
 407 * The if/else clause exists only for the lockdep assertion and can be
 408 * ignored.
 409 */
 410#define for_each_pool_worker(worker, pool)				\
 411	list_for_each_entry((worker), &(pool)->workers, node)		\
 412		if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \
 413		else
 414
 415/**
 416 * for_each_pwq - iterate through all pool_workqueues of the specified workqueue
 417 * @pwq: iteration cursor
 418 * @wq: the target workqueue
 419 *
 420 * This must be called either with wq->mutex held or RCU read locked.
 421 * If the pwq needs to be used beyond the locking in effect, the caller is
 422 * responsible for guaranteeing that the pwq stays online.
 423 *
 424 * The if/else clause exists only for the lockdep assertion and can be
 425 * ignored.
 426 */
 427#define for_each_pwq(pwq, wq)						\
 428	list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node)		\
 429		if (({ assert_rcu_or_wq_mutex(wq); false; })) { }	\
 430		else
 431
 432#ifdef CONFIG_DEBUG_OBJECTS_WORK
 433
 434static struct debug_obj_descr work_debug_descr;
 435
 436static void *work_debug_hint(void *addr)
 437{
 438	return ((struct work_struct *) addr)->func;
 439}
 440
 441static bool work_is_static_object(void *addr)
 
 
 
 
 442{
 443	struct work_struct *work = addr;
 444
 445	return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work));
 
 
 
 
 
 
 
 446}
 447
 448/*
 449 * fixup_init is called when:
 450 * - an active object is initialized
 
 451 */
 452static bool work_fixup_init(void *addr, enum debug_obj_state state)
 453{
 454	struct work_struct *work = addr;
 455
 456	switch (state) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 457	case ODEBUG_STATE_ACTIVE:
 458		cancel_work_sync(work);
 459		debug_object_init(work, &work_debug_descr);
 460		return true;
 461	default:
 462		return false;
 463	}
 464}
 465
 466/*
 467 * fixup_free is called when:
 468 * - an active object is freed
 469 */
 470static bool work_fixup_free(void *addr, enum debug_obj_state state)
 471{
 472	struct work_struct *work = addr;
 473
 474	switch (state) {
 475	case ODEBUG_STATE_ACTIVE:
 476		cancel_work_sync(work);
 477		debug_object_free(work, &work_debug_descr);
 478		return true;
 479	default:
 480		return false;
 481	}
 482}
 483
 484static struct debug_obj_descr work_debug_descr = {
 485	.name		= "work_struct",
 486	.debug_hint	= work_debug_hint,
 487	.is_static_object = work_is_static_object,
 488	.fixup_init	= work_fixup_init,
 
 489	.fixup_free	= work_fixup_free,
 490};
 491
 492static inline void debug_work_activate(struct work_struct *work)
 493{
 494	debug_object_activate(work, &work_debug_descr);
 495}
 496
 497static inline void debug_work_deactivate(struct work_struct *work)
 498{
 499	debug_object_deactivate(work, &work_debug_descr);
 500}
 501
 502void __init_work(struct work_struct *work, int onstack)
 503{
 504	if (onstack)
 505		debug_object_init_on_stack(work, &work_debug_descr);
 506	else
 507		debug_object_init(work, &work_debug_descr);
 508}
 509EXPORT_SYMBOL_GPL(__init_work);
 510
 511void destroy_work_on_stack(struct work_struct *work)
 512{
 513	debug_object_free(work, &work_debug_descr);
 514}
 515EXPORT_SYMBOL_GPL(destroy_work_on_stack);
 516
 517void destroy_delayed_work_on_stack(struct delayed_work *work)
 518{
 519	destroy_timer_on_stack(&work->timer);
 520	debug_object_free(&work->work, &work_debug_descr);
 521}
 522EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack);
 523
 524#else
 525static inline void debug_work_activate(struct work_struct *work) { }
 526static inline void debug_work_deactivate(struct work_struct *work) { }
 527#endif
 528
 529/**
 530 * worker_pool_assign_id - allocate ID and assing it to @pool
 531 * @pool: the pool pointer of interest
 532 *
 533 * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned
 534 * successfully, -errno on failure.
 535 */
 536static int worker_pool_assign_id(struct worker_pool *pool)
 537{
 538	int ret;
 539
 540	lockdep_assert_held(&wq_pool_mutex);
 541
 542	ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE,
 543			GFP_KERNEL);
 544	if (ret >= 0) {
 545		pool->id = ret;
 546		return 0;
 547	}
 548	return ret;
 549}
 550
 551/**
 552 * unbound_pwq_by_node - return the unbound pool_workqueue for the given node
 553 * @wq: the target workqueue
 554 * @node: the node ID
 555 *
 556 * This must be called with any of wq_pool_mutex, wq->mutex or RCU
 557 * read locked.
 558 * If the pwq needs to be used beyond the locking in effect, the caller is
 559 * responsible for guaranteeing that the pwq stays online.
 560 *
 561 * Return: The unbound pool_workqueue for @node.
 562 */
 563static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
 564						  int node)
 565{
 566	assert_rcu_or_wq_mutex_or_pool_mutex(wq);
 567
 568	/*
 569	 * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
 570	 * delayed item is pending.  The plan is to keep CPU -> NODE
 571	 * mapping valid and stable across CPU on/offlines.  Once that
 572	 * happens, this workaround can be removed.
 573	 */
 574	if (unlikely(node == NUMA_NO_NODE))
 575		return wq->dfl_pwq;
 576
 577	return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
 578}
 579
 580static unsigned int work_color_to_flags(int color)
 581{
 582	return color << WORK_STRUCT_COLOR_SHIFT;
 583}
 584
 585static int get_work_color(struct work_struct *work)
 586{
 587	return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
 588		((1 << WORK_STRUCT_COLOR_BITS) - 1);
 589}
 590
 591static int work_next_color(int color)
 592{
 593	return (color + 1) % WORK_NR_COLORS;
 594}
 595
 596/*
 597 * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data
 598 * contain the pointer to the queued pwq.  Once execution starts, the flag
 599 * is cleared and the high bits contain OFFQ flags and pool ID.
 600 *
 601 * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling()
 602 * and clear_work_data() can be used to set the pwq, pool or clear
 603 * work->data.  These functions should only be called while the work is
 604 * owned - ie. while the PENDING bit is set.
 605 *
 606 * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq
 607 * corresponding to a work.  Pool is available once the work has been
 608 * queued anywhere after initialization until it is sync canceled.  pwq is
 609 * available only while the work item is queued.
 610 *
 611 * %WORK_OFFQ_CANCELING is used to mark a work item which is being
 612 * canceled.  While being canceled, a work item may have its PENDING set
 613 * but stay off timer and worklist for arbitrarily long and nobody should
 614 * try to steal the PENDING bit.
 615 */
 616static inline void set_work_data(struct work_struct *work, unsigned long data,
 617				 unsigned long flags)
 618{
 619	WARN_ON_ONCE(!work_pending(work));
 620	atomic_long_set(&work->data, data | flags | work_static(work));
 621}
 622
 623static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq,
 624			 unsigned long extra_flags)
 625{
 626	set_work_data(work, (unsigned long)pwq,
 627		      WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags);
 628}
 629
 630static void set_work_pool_and_keep_pending(struct work_struct *work,
 631					   int pool_id)
 632{
 633	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT,
 634		      WORK_STRUCT_PENDING);
 635}
 636
 637static void set_work_pool_and_clear_pending(struct work_struct *work,
 638					    int pool_id)
 639{
 640	/*
 641	 * The following wmb is paired with the implied mb in
 642	 * test_and_set_bit(PENDING) and ensures all updates to @work made
 643	 * here are visible to and precede any updates by the next PENDING
 644	 * owner.
 645	 */
 646	smp_wmb();
 647	set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0);
 648	/*
 649	 * The following mb guarantees that previous clear of a PENDING bit
 650	 * will not be reordered with any speculative LOADS or STORES from
 651	 * work->current_func, which is executed afterwards.  This possible
 652	 * reordering can lead to a missed execution on attempt to queue
 653	 * the same @work.  E.g. consider this case:
 654	 *
 655	 *   CPU#0                         CPU#1
 656	 *   ----------------------------  --------------------------------
 657	 *
 658	 * 1  STORE event_indicated
 659	 * 2  queue_work_on() {
 660	 * 3    test_and_set_bit(PENDING)
 661	 * 4 }                             set_..._and_clear_pending() {
 662	 * 5                                 set_work_data() # clear bit
 663	 * 6                                 smp_mb()
 664	 * 7                               work->current_func() {
 665	 * 8				      LOAD event_indicated
 666	 *				   }
 667	 *
 668	 * Without an explicit full barrier speculative LOAD on line 8 can
 669	 * be executed before CPU#0 does STORE on line 1.  If that happens,
 670	 * CPU#0 observes the PENDING bit is still set and new execution of
 671	 * a @work is not queued in a hope, that CPU#1 will eventually
 672	 * finish the queued @work.  Meanwhile CPU#1 does not see
 673	 * event_indicated is set, because speculative LOAD was executed
 674	 * before actual STORE.
 675	 */
 676	smp_mb();
 677}
 678
 679static void clear_work_data(struct work_struct *work)
 680{
 681	smp_wmb();	/* see set_work_pool_and_clear_pending() */
 682	set_work_data(work, WORK_STRUCT_NO_POOL, 0);
 683}
 684
 685static struct pool_workqueue *get_work_pwq(struct work_struct *work)
 686{
 687	unsigned long data = atomic_long_read(&work->data);
 688
 689	if (data & WORK_STRUCT_PWQ)
 690		return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
 691	else
 692		return NULL;
 693}
 694
 695/**
 696 * get_work_pool - return the worker_pool a given work was associated with
 697 * @work: the work item of interest
 698 *
 699 * Pools are created and destroyed under wq_pool_mutex, and allows read
 700 * access under RCU read lock.  As such, this function should be
 701 * called under wq_pool_mutex or inside of a rcu_read_lock() region.
 702 *
 703 * All fields of the returned pool are accessible as long as the above
 704 * mentioned locking is in effect.  If the returned pool needs to be used
 705 * beyond the critical section, the caller is responsible for ensuring the
 706 * returned pool is and stays online.
 707 *
 708 * Return: The worker_pool @work was last associated with.  %NULL if none.
 709 */
 710static struct worker_pool *get_work_pool(struct work_struct *work)
 711{
 712	unsigned long data = atomic_long_read(&work->data);
 713	int pool_id;
 714
 715	assert_rcu_or_pool_mutex();
 716
 717	if (data & WORK_STRUCT_PWQ)
 718		return ((struct pool_workqueue *)
 719			(data & WORK_STRUCT_WQ_DATA_MASK))->pool;
 720
 721	pool_id = data >> WORK_OFFQ_POOL_SHIFT;
 722	if (pool_id == WORK_OFFQ_POOL_NONE)
 723		return NULL;
 724
 725	return idr_find(&worker_pool_idr, pool_id);
 726}
 727
 728/**
 729 * get_work_pool_id - return the worker pool ID a given work is associated with
 730 * @work: the work item of interest
 731 *
 732 * Return: The worker_pool ID @work was last associated with.
 733 * %WORK_OFFQ_POOL_NONE if none.
 734 */
 735static int get_work_pool_id(struct work_struct *work)
 736{
 737	unsigned long data = atomic_long_read(&work->data);
 738
 739	if (data & WORK_STRUCT_PWQ)
 740		return ((struct pool_workqueue *)
 741			(data & WORK_STRUCT_WQ_DATA_MASK))->pool->id;
 742
 743	return data >> WORK_OFFQ_POOL_SHIFT;
 744}
 745
 746static void mark_work_canceling(struct work_struct *work)
 747{
 748	unsigned long pool_id = get_work_pool_id(work);
 749
 750	pool_id <<= WORK_OFFQ_POOL_SHIFT;
 751	set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING);
 752}
 753
 754static bool work_is_canceling(struct work_struct *work)
 755{
 756	unsigned long data = atomic_long_read(&work->data);
 757
 758	return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING);
 759}
 760
 761/*
 762 * Policy functions.  These define the policies on how the global worker
 763 * pools are managed.  Unless noted otherwise, these functions assume that
 764 * they're being called with pool->lock held.
 765 */
 766
 767static bool __need_more_worker(struct worker_pool *pool)
 768{
 769	return !atomic_read(&pool->nr_running);
 770}
 771
 772/*
 773 * Need to wake up a worker?  Called from anything but currently
 774 * running workers.
 775 *
 776 * Note that, because unbound workers never contribute to nr_running, this
 777 * function will always return %true for unbound pools as long as the
 778 * worklist isn't empty.
 779 */
 780static bool need_more_worker(struct worker_pool *pool)
 781{
 782	return !list_empty(&pool->worklist) && __need_more_worker(pool);
 783}
 784
 785/* Can I start working?  Called from busy but !running workers. */
 786static bool may_start_working(struct worker_pool *pool)
 787{
 788	return pool->nr_idle;
 789}
 790
 791/* Do I need to keep working?  Called from currently running workers. */
 792static bool keep_working(struct worker_pool *pool)
 793{
 794	return !list_empty(&pool->worklist) &&
 795		atomic_read(&pool->nr_running) <= 1;
 796}
 797
 798/* Do we need a new worker?  Called from manager. */
 799static bool need_to_create_worker(struct worker_pool *pool)
 800{
 801	return need_more_worker(pool) && !may_start_working(pool);
 802}
 803
 804/* Do we have too many workers and should some go away? */
 805static bool too_many_workers(struct worker_pool *pool)
 806{
 807	bool managing = pool->flags & POOL_MANAGER_ACTIVE;
 808	int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
 809	int nr_busy = pool->nr_workers - nr_idle;
 810
 811	return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
 812}
 813
 814/*
 815 * Wake up functions.
 816 */
 817
 818/* Return the first idle worker.  Safe with preemption disabled */
 819static struct worker *first_idle_worker(struct worker_pool *pool)
 820{
 821	if (unlikely(list_empty(&pool->idle_list)))
 822		return NULL;
 823
 824	return list_first_entry(&pool->idle_list, struct worker, entry);
 825}
 826
 827/**
 828 * wake_up_worker - wake up an idle worker
 829 * @pool: worker pool to wake worker from
 830 *
 831 * Wake up the first idle worker of @pool.
 832 *
 833 * CONTEXT:
 834 * spin_lock_irq(pool->lock).
 835 */
 836static void wake_up_worker(struct worker_pool *pool)
 837{
 838	struct worker *worker = first_idle_worker(pool);
 839
 840	if (likely(worker))
 841		wake_up_process(worker->task);
 842}
 843
 844/**
 845 * wq_worker_running - a worker is running again
 846 * @task: task waking up
 
 
 
 
 847 *
 848 * This function is called when a worker returns from schedule()
 
 849 */
 850void wq_worker_running(struct task_struct *task)
 851{
 852	struct worker *worker = kthread_data(task);
 853
 854	if (!worker->sleeping)
 855		return;
 856	if (!(worker->flags & WORKER_NOT_RUNNING))
 857		atomic_inc(&worker->pool->nr_running);
 858	worker->sleeping = 0;
 859}
 860
 861/**
 862 * wq_worker_sleeping - a worker is going to sleep
 863 * @task: task going to sleep
 864 *
 865 * This function is called from schedule() when a busy worker is
 866 * going to sleep.
 
 
 
 
 
 
 
 867 */
 868void wq_worker_sleeping(struct task_struct *task)
 869{
 870	struct worker *next, *worker = kthread_data(task);
 871	struct worker_pool *pool;
 872
 873	/*
 874	 * Rescuers, which may not have all the fields set up like normal
 875	 * workers, also reach here, let's not access anything before
 876	 * checking NOT_RUNNING.
 877	 */
 878	if (worker->flags & WORKER_NOT_RUNNING)
 879		return;
 880
 881	pool = worker->pool;
 882
 883	if (WARN_ON_ONCE(worker->sleeping))
 884		return;
 885
 886	worker->sleeping = 1;
 887	spin_lock_irq(&pool->lock);
 888
 889	/*
 890	 * The counterpart of the following dec_and_test, implied mb,
 891	 * worklist not empty test sequence is in insert_work().
 892	 * Please read comment there.
 893	 *
 894	 * NOT_RUNNING is clear.  This means that we're bound to and
 895	 * running on the local cpu w/ rq lock held and preemption
 896	 * disabled, which in turn means that none else could be
 897	 * manipulating idle_list, so dereferencing idle_list without pool
 898	 * lock is safe.
 899	 */
 900	if (atomic_dec_and_test(&pool->nr_running) &&
 901	    !list_empty(&pool->worklist)) {
 902		next = first_idle_worker(pool);
 903		if (next)
 904			wake_up_process(next->task);
 905	}
 906	spin_unlock_irq(&pool->lock);
 907}
 908
 909/**
 910 * wq_worker_last_func - retrieve worker's last work function
 911 * @task: Task to retrieve last work function of.
 912 *
 913 * Determine the last function a worker executed. This is called from
 914 * the scheduler to get a worker's last known identity.
 915 *
 916 * CONTEXT:
 917 * spin_lock_irq(rq->lock)
 918 *
 919 * This function is called during schedule() when a kworker is going
 920 * to sleep. It's used by psi to identify aggregation workers during
 921 * dequeuing, to allow periodic aggregation to shut-off when that
 922 * worker is the last task in the system or cgroup to go to sleep.
 923 *
 924 * As this function doesn't involve any workqueue-related locking, it
 925 * only returns stable values when called from inside the scheduler's
 926 * queuing and dequeuing paths, when @task, which must be a kworker,
 927 * is guaranteed to not be processing any works.
 928 *
 929 * Return:
 930 * The last work function %current executed as a worker, NULL if it
 931 * hasn't executed any work yet.
 932 */
 933work_func_t wq_worker_last_func(struct task_struct *task)
 934{
 935	struct worker *worker = kthread_data(task);
 936
 937	return worker->last_func;
 938}
 939
 940/**
 941 * worker_set_flags - set worker flags and adjust nr_running accordingly
 942 * @worker: self
 943 * @flags: flags to set
 944 *
 945 * Set @flags in @worker->flags and adjust nr_running accordingly.
 946 *
 947 * CONTEXT:
 948 * spin_lock_irq(pool->lock)
 949 */
 950static inline void worker_set_flags(struct worker *worker, unsigned int flags)
 951{
 952	struct worker_pool *pool = worker->pool;
 953
 954	WARN_ON_ONCE(worker->task != current);
 955
 956	/* If transitioning into NOT_RUNNING, adjust nr_running. */
 957	if ((flags & WORKER_NOT_RUNNING) &&
 958	    !(worker->flags & WORKER_NOT_RUNNING)) {
 959		atomic_dec(&pool->nr_running);
 960	}
 961
 962	worker->flags |= flags;
 963}
 964
 965/**
 966 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
 967 * @worker: self
 968 * @flags: flags to clear
 969 *
 970 * Clear @flags in @worker->flags and adjust nr_running accordingly.
 971 *
 972 * CONTEXT:
 973 * spin_lock_irq(pool->lock)
 974 */
 975static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
 976{
 977	struct worker_pool *pool = worker->pool;
 978	unsigned int oflags = worker->flags;
 979
 980	WARN_ON_ONCE(worker->task != current);
 981
 982	worker->flags &= ~flags;
 983
 984	/*
 985	 * If transitioning out of NOT_RUNNING, increment nr_running.  Note
 986	 * that the nested NOT_RUNNING is not a noop.  NOT_RUNNING is mask
 987	 * of multiple flags, not a single flag.
 988	 */
 989	if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
 990		if (!(worker->flags & WORKER_NOT_RUNNING))
 991			atomic_inc(&pool->nr_running);
 992}
 993
 994/**
 995 * find_worker_executing_work - find worker which is executing a work
 996 * @pool: pool of interest
 997 * @work: work to find worker for
 998 *
 999 * Find a worker which is executing @work on @pool by searching
1000 * @pool->busy_hash which is keyed by the address of @work.  For a worker
1001 * to match, its current execution should match the address of @work and
1002 * its work function.  This is to avoid unwanted dependency between
1003 * unrelated work executions through a work item being recycled while still
1004 * being executed.
1005 *
1006 * This is a bit tricky.  A work item may be freed once its execution
1007 * starts and nothing prevents the freed area from being recycled for
1008 * another work item.  If the same work item address ends up being reused
1009 * before the original execution finishes, workqueue will identify the
1010 * recycled work item as currently executing and make it wait until the
1011 * current execution finishes, introducing an unwanted dependency.
1012 *
1013 * This function checks the work item address and work function to avoid
1014 * false positives.  Note that this isn't complete as one may construct a
1015 * work function which can introduce dependency onto itself through a
1016 * recycled work item.  Well, if somebody wants to shoot oneself in the
1017 * foot that badly, there's only so much we can do, and if such deadlock
1018 * actually occurs, it should be easy to locate the culprit work function.
1019 *
1020 * CONTEXT:
1021 * spin_lock_irq(pool->lock).
1022 *
1023 * Return:
1024 * Pointer to worker which is executing @work if found, %NULL
1025 * otherwise.
1026 */
1027static struct worker *find_worker_executing_work(struct worker_pool *pool,
1028						 struct work_struct *work)
1029{
1030	struct worker *worker;
1031
1032	hash_for_each_possible(pool->busy_hash, worker, hentry,
1033			       (unsigned long)work)
1034		if (worker->current_work == work &&
1035		    worker->current_func == work->func)
1036			return worker;
1037
1038	return NULL;
1039}
1040
1041/**
1042 * move_linked_works - move linked works to a list
1043 * @work: start of series of works to be scheduled
1044 * @head: target list to append @work to
1045 * @nextp: out parameter for nested worklist walking
1046 *
1047 * Schedule linked works starting from @work to @head.  Work series to
1048 * be scheduled starts at @work and includes any consecutive work with
1049 * WORK_STRUCT_LINKED set in its predecessor.
1050 *
1051 * If @nextp is not NULL, it's updated to point to the next work of
1052 * the last scheduled work.  This allows move_linked_works() to be
1053 * nested inside outer list_for_each_entry_safe().
1054 *
1055 * CONTEXT:
1056 * spin_lock_irq(pool->lock).
1057 */
1058static void move_linked_works(struct work_struct *work, struct list_head *head,
1059			      struct work_struct **nextp)
1060{
1061	struct work_struct *n;
1062
1063	/*
1064	 * Linked worklist will always end before the end of the list,
1065	 * use NULL for list head.
1066	 */
1067	list_for_each_entry_safe_from(work, n, NULL, entry) {
1068		list_move_tail(&work->entry, head);
1069		if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1070			break;
1071	}
1072
1073	/*
1074	 * If we're already inside safe list traversal and have moved
1075	 * multiple works to the scheduled queue, the next position
1076	 * needs to be updated.
1077	 */
1078	if (nextp)
1079		*nextp = n;
1080}
1081
1082/**
1083 * get_pwq - get an extra reference on the specified pool_workqueue
1084 * @pwq: pool_workqueue to get
1085 *
1086 * Obtain an extra reference on @pwq.  The caller should guarantee that
1087 * @pwq has positive refcnt and be holding the matching pool->lock.
1088 */
1089static void get_pwq(struct pool_workqueue *pwq)
1090{
1091	lockdep_assert_held(&pwq->pool->lock);
1092	WARN_ON_ONCE(pwq->refcnt <= 0);
1093	pwq->refcnt++;
1094}
1095
1096/**
1097 * put_pwq - put a pool_workqueue reference
1098 * @pwq: pool_workqueue to put
1099 *
1100 * Drop a reference of @pwq.  If its refcnt reaches zero, schedule its
1101 * destruction.  The caller should be holding the matching pool->lock.
1102 */
1103static void put_pwq(struct pool_workqueue *pwq)
1104{
1105	lockdep_assert_held(&pwq->pool->lock);
1106	if (likely(--pwq->refcnt))
1107		return;
1108	if (WARN_ON_ONCE(!(pwq->wq->flags & WQ_UNBOUND)))
1109		return;
1110	/*
1111	 * @pwq can't be released under pool->lock, bounce to
1112	 * pwq_unbound_release_workfn().  This never recurses on the same
1113	 * pool->lock as this path is taken only for unbound workqueues and
1114	 * the release work item is scheduled on a per-cpu workqueue.  To
1115	 * avoid lockdep warning, unbound pool->locks are given lockdep
1116	 * subclass of 1 in get_unbound_pool().
1117	 */
1118	schedule_work(&pwq->unbound_release_work);
1119}
1120
1121/**
1122 * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock
1123 * @pwq: pool_workqueue to put (can be %NULL)
1124 *
1125 * put_pwq() with locking.  This function also allows %NULL @pwq.
1126 */
1127static void put_pwq_unlocked(struct pool_workqueue *pwq)
1128{
1129	if (pwq) {
1130		/*
1131		 * As both pwqs and pools are RCU protected, the
1132		 * following lock operations are safe.
1133		 */
1134		spin_lock_irq(&pwq->pool->lock);
1135		put_pwq(pwq);
1136		spin_unlock_irq(&pwq->pool->lock);
1137	}
1138}
1139
1140static void pwq_activate_delayed_work(struct work_struct *work)
1141{
1142	struct pool_workqueue *pwq = get_work_pwq(work);
1143
1144	trace_workqueue_activate_work(work);
1145	if (list_empty(&pwq->pool->worklist))
1146		pwq->pool->watchdog_ts = jiffies;
1147	move_linked_works(work, &pwq->pool->worklist, NULL);
1148	__clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
1149	pwq->nr_active++;
1150}
1151
1152static void pwq_activate_first_delayed(struct pool_workqueue *pwq)
1153{
1154	struct work_struct *work = list_first_entry(&pwq->delayed_works,
1155						    struct work_struct, entry);
1156
1157	pwq_activate_delayed_work(work);
1158}
1159
1160/**
1161 * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight
1162 * @pwq: pwq of interest
1163 * @color: color of work which left the queue
1164 *
1165 * A work either has completed or is removed from pending queue,
1166 * decrement nr_in_flight of its pwq and handle workqueue flushing.
1167 *
1168 * CONTEXT:
1169 * spin_lock_irq(pool->lock).
1170 */
1171static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, int color)
1172{
1173	/* uncolored work items don't participate in flushing or nr_active */
1174	if (color == WORK_NO_COLOR)
1175		goto out_put;
1176
1177	pwq->nr_in_flight[color]--;
1178
1179	pwq->nr_active--;
1180	if (!list_empty(&pwq->delayed_works)) {
1181		/* one down, submit a delayed one */
1182		if (pwq->nr_active < pwq->max_active)
1183			pwq_activate_first_delayed(pwq);
1184	}
1185
1186	/* is flush in progress and are we at the flushing tip? */
1187	if (likely(pwq->flush_color != color))
1188		goto out_put;
1189
1190	/* are there still in-flight works? */
1191	if (pwq->nr_in_flight[color])
1192		goto out_put;
1193
1194	/* this pwq is done, clear flush_color */
1195	pwq->flush_color = -1;
1196
1197	/*
1198	 * If this was the last pwq, wake up the first flusher.  It
1199	 * will handle the rest.
1200	 */
1201	if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush))
1202		complete(&pwq->wq->first_flusher->done);
1203out_put:
1204	put_pwq(pwq);
1205}
1206
1207/**
1208 * try_to_grab_pending - steal work item from worklist and disable irq
1209 * @work: work item to steal
1210 * @is_dwork: @work is a delayed_work
1211 * @flags: place to store irq state
1212 *
1213 * Try to grab PENDING bit of @work.  This function can handle @work in any
1214 * stable state - idle, on timer or on worklist.
1215 *
1216 * Return:
1217 *  1		if @work was pending and we successfully stole PENDING
1218 *  0		if @work was idle and we claimed PENDING
1219 *  -EAGAIN	if PENDING couldn't be grabbed at the moment, safe to busy-retry
1220 *  -ENOENT	if someone else is canceling @work, this state may persist
1221 *		for arbitrarily long
1222 *
1223 * Note:
1224 * On >= 0 return, the caller owns @work's PENDING bit.  To avoid getting
1225 * interrupted while holding PENDING and @work off queue, irq must be
1226 * disabled on entry.  This, combined with delayed_work->timer being
1227 * irqsafe, ensures that we return -EAGAIN for finite short period of time.
1228 *
1229 * On successful return, >= 0, irq is disabled and the caller is
1230 * responsible for releasing it using local_irq_restore(*@flags).
1231 *
1232 * This function is safe to call from any context including IRQ handler.
1233 */
1234static int try_to_grab_pending(struct work_struct *work, bool is_dwork,
1235			       unsigned long *flags)
1236{
1237	struct worker_pool *pool;
1238	struct pool_workqueue *pwq;
1239
1240	local_irq_save(*flags);
1241
1242	/* try to steal the timer if it exists */
1243	if (is_dwork) {
1244		struct delayed_work *dwork = to_delayed_work(work);
1245
1246		/*
1247		 * dwork->timer is irqsafe.  If del_timer() fails, it's
1248		 * guaranteed that the timer is not queued anywhere and not
1249		 * running on the local CPU.
1250		 */
1251		if (likely(del_timer(&dwork->timer)))
1252			return 1;
1253	}
1254
1255	/* try to claim PENDING the normal way */
1256	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1257		return 0;
1258
1259	rcu_read_lock();
1260	/*
1261	 * The queueing is in progress, or it is already queued. Try to
1262	 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1263	 */
1264	pool = get_work_pool(work);
1265	if (!pool)
1266		goto fail;
1267
1268	spin_lock(&pool->lock);
1269	/*
1270	 * work->data is guaranteed to point to pwq only while the work
1271	 * item is queued on pwq->wq, and both updating work->data to point
1272	 * to pwq on queueing and to pool on dequeueing are done under
1273	 * pwq->pool->lock.  This in turn guarantees that, if work->data
1274	 * points to pwq which is associated with a locked pool, the work
1275	 * item is currently queued on that pool.
1276	 */
1277	pwq = get_work_pwq(work);
1278	if (pwq && pwq->pool == pool) {
1279		debug_work_deactivate(work);
1280
1281		/*
1282		 * A delayed work item cannot be grabbed directly because
1283		 * it might have linked NO_COLOR work items which, if left
1284		 * on the delayed_list, will confuse pwq->nr_active
1285		 * management later on and cause stall.  Make sure the work
1286		 * item is activated before grabbing.
1287		 */
1288		if (*work_data_bits(work) & WORK_STRUCT_DELAYED)
1289			pwq_activate_delayed_work(work);
1290
1291		list_del_init(&work->entry);
1292		pwq_dec_nr_in_flight(pwq, get_work_color(work));
1293
1294		/* work->data points to pwq iff queued, point to pool */
1295		set_work_pool_and_keep_pending(work, pool->id);
1296
1297		spin_unlock(&pool->lock);
1298		rcu_read_unlock();
1299		return 1;
1300	}
1301	spin_unlock(&pool->lock);
1302fail:
1303	rcu_read_unlock();
1304	local_irq_restore(*flags);
1305	if (work_is_canceling(work))
1306		return -ENOENT;
1307	cpu_relax();
1308	return -EAGAIN;
1309}
1310
1311/**
1312 * insert_work - insert a work into a pool
1313 * @pwq: pwq @work belongs to
1314 * @work: work to insert
1315 * @head: insertion point
1316 * @extra_flags: extra WORK_STRUCT_* flags to set
1317 *
1318 * Insert @work which belongs to @pwq after @head.  @extra_flags is or'd to
1319 * work_struct flags.
1320 *
1321 * CONTEXT:
1322 * spin_lock_irq(pool->lock).
1323 */
1324static void insert_work(struct pool_workqueue *pwq, struct work_struct *work,
1325			struct list_head *head, unsigned int extra_flags)
1326{
1327	struct worker_pool *pool = pwq->pool;
1328
1329	/* we own @work, set data and link */
1330	set_work_pwq(work, pwq, extra_flags);
1331	list_add_tail(&work->entry, head);
1332	get_pwq(pwq);
1333
1334	/*
1335	 * Ensure either wq_worker_sleeping() sees the above
1336	 * list_add_tail() or we see zero nr_running to avoid workers lying
1337	 * around lazily while there are works to be processed.
1338	 */
1339	smp_mb();
1340
1341	if (__need_more_worker(pool))
1342		wake_up_worker(pool);
1343}
1344
1345/*
1346 * Test whether @work is being queued from another work executing on the
1347 * same workqueue.
1348 */
1349static bool is_chained_work(struct workqueue_struct *wq)
1350{
1351	struct worker *worker;
1352
1353	worker = current_wq_worker();
1354	/*
1355	 * Return %true iff I'm a worker executing a work item on @wq.  If
1356	 * I'm @worker, it's safe to dereference it without locking.
1357	 */
1358	return worker && worker->current_pwq->wq == wq;
1359}
1360
1361/*
1362 * When queueing an unbound work item to a wq, prefer local CPU if allowed
1363 * by wq_unbound_cpumask.  Otherwise, round robin among the allowed ones to
1364 * avoid perturbing sensitive tasks.
1365 */
1366static int wq_select_unbound_cpu(int cpu)
1367{
1368	static bool printed_dbg_warning;
1369	int new_cpu;
1370
1371	if (likely(!wq_debug_force_rr_cpu)) {
1372		if (cpumask_test_cpu(cpu, wq_unbound_cpumask))
1373			return cpu;
1374	} else if (!printed_dbg_warning) {
1375		pr_warn("workqueue: round-robin CPU selection forced, expect performance impact\n");
1376		printed_dbg_warning = true;
1377	}
1378
1379	if (cpumask_empty(wq_unbound_cpumask))
1380		return cpu;
1381
1382	new_cpu = __this_cpu_read(wq_rr_cpu_last);
1383	new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask);
1384	if (unlikely(new_cpu >= nr_cpu_ids)) {
1385		new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask);
1386		if (unlikely(new_cpu >= nr_cpu_ids))
1387			return cpu;
1388	}
1389	__this_cpu_write(wq_rr_cpu_last, new_cpu);
1390
1391	return new_cpu;
1392}
1393
1394static void __queue_work(int cpu, struct workqueue_struct *wq,
1395			 struct work_struct *work)
1396{
1397	struct pool_workqueue *pwq;
1398	struct worker_pool *last_pool;
1399	struct list_head *worklist;
1400	unsigned int work_flags;
1401	unsigned int req_cpu = cpu;
1402
1403	/*
1404	 * While a work item is PENDING && off queue, a task trying to
1405	 * steal the PENDING will busy-loop waiting for it to either get
1406	 * queued or lose PENDING.  Grabbing PENDING and queueing should
1407	 * happen with IRQ disabled.
1408	 */
1409	lockdep_assert_irqs_disabled();
1410
1411	debug_work_activate(work);
1412
1413	/* if draining, only works from the same workqueue are allowed */
1414	if (unlikely(wq->flags & __WQ_DRAINING) &&
1415	    WARN_ON_ONCE(!is_chained_work(wq)))
1416		return;
1417	rcu_read_lock();
1418retry:
1419	if (req_cpu == WORK_CPU_UNBOUND)
1420		cpu = wq_select_unbound_cpu(raw_smp_processor_id());
1421
1422	/* pwq which will be used unless @work is executing elsewhere */
1423	if (!(wq->flags & WQ_UNBOUND))
1424		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
1425	else
1426		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
1427
1428	/*
1429	 * If @work was previously on a different pool, it might still be
1430	 * running there, in which case the work needs to be queued on that
1431	 * pool to guarantee non-reentrancy.
1432	 */
1433	last_pool = get_work_pool(work);
1434	if (last_pool && last_pool != pwq->pool) {
1435		struct worker *worker;
1436
1437		spin_lock(&last_pool->lock);
1438
1439		worker = find_worker_executing_work(last_pool, work);
1440
1441		if (worker && worker->current_pwq->wq == wq) {
1442			pwq = worker->current_pwq;
1443		} else {
1444			/* meh... not running there, queue here */
1445			spin_unlock(&last_pool->lock);
1446			spin_lock(&pwq->pool->lock);
1447		}
1448	} else {
1449		spin_lock(&pwq->pool->lock);
1450	}
1451
1452	/*
1453	 * pwq is determined and locked.  For unbound pools, we could have
1454	 * raced with pwq release and it could already be dead.  If its
1455	 * refcnt is zero, repeat pwq selection.  Note that pwqs never die
1456	 * without another pwq replacing it in the numa_pwq_tbl or while
1457	 * work items are executing on it, so the retrying is guaranteed to
1458	 * make forward-progress.
1459	 */
1460	if (unlikely(!pwq->refcnt)) {
1461		if (wq->flags & WQ_UNBOUND) {
1462			spin_unlock(&pwq->pool->lock);
1463			cpu_relax();
1464			goto retry;
1465		}
1466		/* oops */
1467		WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt",
1468			  wq->name, cpu);
1469	}
1470
1471	/* pwq determined, queue */
1472	trace_workqueue_queue_work(req_cpu, pwq, work);
1473
1474	if (WARN_ON(!list_empty(&work->entry)))
1475		goto out;
 
 
1476
1477	pwq->nr_in_flight[pwq->work_color]++;
1478	work_flags = work_color_to_flags(pwq->work_color);
1479
1480	if (likely(pwq->nr_active < pwq->max_active)) {
1481		trace_workqueue_activate_work(work);
1482		pwq->nr_active++;
1483		worklist = &pwq->pool->worklist;
1484		if (list_empty(worklist))
1485			pwq->pool->watchdog_ts = jiffies;
1486	} else {
1487		work_flags |= WORK_STRUCT_DELAYED;
1488		worklist = &pwq->delayed_works;
1489	}
1490
1491	insert_work(pwq, work, worklist, work_flags);
1492
1493out:
1494	spin_unlock(&pwq->pool->lock);
1495	rcu_read_unlock();
1496}
1497
1498/**
1499 * queue_work_on - queue work on specific cpu
1500 * @cpu: CPU number to execute work on
1501 * @wq: workqueue to use
1502 * @work: work to queue
1503 *
1504 * We queue the work to a specific CPU, the caller must ensure it
1505 * can't go away.
1506 *
1507 * Return: %false if @work was already on a queue, %true otherwise.
1508 */
1509bool queue_work_on(int cpu, struct workqueue_struct *wq,
1510		   struct work_struct *work)
1511{
1512	bool ret = false;
1513	unsigned long flags;
1514
1515	local_irq_save(flags);
1516
1517	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1518		__queue_work(cpu, wq, work);
1519		ret = true;
1520	}
1521
1522	local_irq_restore(flags);
1523	return ret;
1524}
1525EXPORT_SYMBOL(queue_work_on);
1526
1527/**
1528 * workqueue_select_cpu_near - Select a CPU based on NUMA node
1529 * @node: NUMA node ID that we want to select a CPU from
1530 *
1531 * This function will attempt to find a "random" cpu available on a given
1532 * node. If there are no CPUs available on the given node it will return
1533 * WORK_CPU_UNBOUND indicating that we should just schedule to any
1534 * available CPU if we need to schedule this work.
1535 */
1536static int workqueue_select_cpu_near(int node)
1537{
1538	int cpu;
1539
1540	/* No point in doing this if NUMA isn't enabled for workqueues */
1541	if (!wq_numa_enabled)
1542		return WORK_CPU_UNBOUND;
1543
1544	/* Delay binding to CPU if node is not valid or online */
1545	if (node < 0 || node >= MAX_NUMNODES || !node_online(node))
1546		return WORK_CPU_UNBOUND;
1547
1548	/* Use local node/cpu if we are already there */
1549	cpu = raw_smp_processor_id();
1550	if (node == cpu_to_node(cpu))
1551		return cpu;
1552
1553	/* Use "random" otherwise know as "first" online CPU of node */
1554	cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask);
1555
1556	/* If CPU is valid return that, otherwise just defer */
1557	return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND;
1558}
1559
1560/**
1561 * queue_work_node - queue work on a "random" cpu for a given NUMA node
1562 * @node: NUMA node that we are targeting the work for
1563 * @wq: workqueue to use
1564 * @work: work to queue
1565 *
1566 * We queue the work to a "random" CPU within a given NUMA node. The basic
1567 * idea here is to provide a way to somehow associate work with a given
1568 * NUMA node.
1569 *
1570 * This function will only make a best effort attempt at getting this onto
1571 * the right NUMA node. If no node is requested or the requested node is
1572 * offline then we just fall back to standard queue_work behavior.
1573 *
1574 * Currently the "random" CPU ends up being the first available CPU in the
1575 * intersection of cpu_online_mask and the cpumask of the node, unless we
1576 * are running on the node. In that case we just use the current CPU.
1577 *
1578 * Return: %false if @work was already on a queue, %true otherwise.
1579 */
1580bool queue_work_node(int node, struct workqueue_struct *wq,
1581		     struct work_struct *work)
1582{
1583	unsigned long flags;
1584	bool ret = false;
1585
1586	/*
1587	 * This current implementation is specific to unbound workqueues.
1588	 * Specifically we only return the first available CPU for a given
1589	 * node instead of cycling through individual CPUs within the node.
1590	 *
1591	 * If this is used with a per-cpu workqueue then the logic in
1592	 * workqueue_select_cpu_near would need to be updated to allow for
1593	 * some round robin type logic.
1594	 */
1595	WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND));
1596
1597	local_irq_save(flags);
1598
1599	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1600		int cpu = workqueue_select_cpu_near(node);
1601
1602		__queue_work(cpu, wq, work);
1603		ret = true;
1604	}
1605
1606	local_irq_restore(flags);
1607	return ret;
1608}
1609EXPORT_SYMBOL_GPL(queue_work_node);
1610
1611void delayed_work_timer_fn(struct timer_list *t)
1612{
1613	struct delayed_work *dwork = from_timer(dwork, t, timer);
1614
1615	/* should have been called from irqsafe timer with irq already off */
1616	__queue_work(dwork->cpu, dwork->wq, &dwork->work);
1617}
1618EXPORT_SYMBOL(delayed_work_timer_fn);
1619
1620static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1621				struct delayed_work *dwork, unsigned long delay)
1622{
1623	struct timer_list *timer = &dwork->timer;
1624	struct work_struct *work = &dwork->work;
1625
1626	WARN_ON_ONCE(!wq);
1627	WARN_ON_ONCE(timer->function != delayed_work_timer_fn);
1628	WARN_ON_ONCE(timer_pending(timer));
1629	WARN_ON_ONCE(!list_empty(&work->entry));
1630
1631	/*
1632	 * If @delay is 0, queue @dwork->work immediately.  This is for
1633	 * both optimization and correctness.  The earliest @timer can
1634	 * expire is on the closest next tick and delayed_work users depend
1635	 * on that there's no such delay when @delay is 0.
1636	 */
1637	if (!delay) {
1638		__queue_work(cpu, wq, &dwork->work);
1639		return;
1640	}
1641
 
 
1642	dwork->wq = wq;
1643	dwork->cpu = cpu;
1644	timer->expires = jiffies + delay;
1645
1646	if (unlikely(cpu != WORK_CPU_UNBOUND))
1647		add_timer_on(timer, cpu);
1648	else
1649		add_timer(timer);
1650}
1651
1652/**
1653 * queue_delayed_work_on - queue work on specific CPU after delay
1654 * @cpu: CPU number to execute work on
1655 * @wq: workqueue to use
1656 * @dwork: work to queue
1657 * @delay: number of jiffies to wait before queueing
1658 *
1659 * Return: %false if @work was already on a queue, %true otherwise.  If
1660 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1661 * execution.
1662 */
1663bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1664			   struct delayed_work *dwork, unsigned long delay)
1665{
1666	struct work_struct *work = &dwork->work;
1667	bool ret = false;
1668	unsigned long flags;
1669
1670	/* read the comment in __queue_work() */
1671	local_irq_save(flags);
1672
1673	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1674		__queue_delayed_work(cpu, wq, dwork, delay);
1675		ret = true;
1676	}
1677
1678	local_irq_restore(flags);
1679	return ret;
1680}
1681EXPORT_SYMBOL(queue_delayed_work_on);
1682
1683/**
1684 * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU
1685 * @cpu: CPU number to execute work on
1686 * @wq: workqueue to use
1687 * @dwork: work to queue
1688 * @delay: number of jiffies to wait before queueing
1689 *
1690 * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise,
1691 * modify @dwork's timer so that it expires after @delay.  If @delay is
1692 * zero, @work is guaranteed to be scheduled immediately regardless of its
1693 * current state.
1694 *
1695 * Return: %false if @dwork was idle and queued, %true if @dwork was
1696 * pending and its timer was modified.
1697 *
1698 * This function is safe to call from any context including IRQ handler.
1699 * See try_to_grab_pending() for details.
1700 */
1701bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq,
1702			 struct delayed_work *dwork, unsigned long delay)
1703{
1704	unsigned long flags;
1705	int ret;
1706
1707	do {
1708		ret = try_to_grab_pending(&dwork->work, true, &flags);
1709	} while (unlikely(ret == -EAGAIN));
1710
1711	if (likely(ret >= 0)) {
1712		__queue_delayed_work(cpu, wq, dwork, delay);
1713		local_irq_restore(flags);
1714	}
1715
1716	/* -ENOENT from try_to_grab_pending() becomes %true */
1717	return ret;
1718}
1719EXPORT_SYMBOL_GPL(mod_delayed_work_on);
1720
1721static void rcu_work_rcufn(struct rcu_head *rcu)
1722{
1723	struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu);
1724
1725	/* read the comment in __queue_work() */
1726	local_irq_disable();
1727	__queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work);
1728	local_irq_enable();
1729}
1730
1731/**
1732 * queue_rcu_work - queue work after a RCU grace period
1733 * @wq: workqueue to use
1734 * @rwork: work to queue
1735 *
1736 * Return: %false if @rwork was already pending, %true otherwise.  Note
1737 * that a full RCU grace period is guaranteed only after a %true return.
1738 * While @rwork is guaranteed to be executed after a %false return, the
1739 * execution may happen before a full RCU grace period has passed.
1740 */
1741bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork)
1742{
1743	struct work_struct *work = &rwork->work;
1744
1745	if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
1746		rwork->wq = wq;
1747		call_rcu(&rwork->rcu, rcu_work_rcufn);
1748		return true;
1749	}
1750
1751	return false;
1752}
1753EXPORT_SYMBOL(queue_rcu_work);
1754
1755/**
1756 * worker_enter_idle - enter idle state
1757 * @worker: worker which is entering idle state
1758 *
1759 * @worker is entering idle state.  Update stats and idle timer if
1760 * necessary.
1761 *
1762 * LOCKING:
1763 * spin_lock_irq(pool->lock).
1764 */
1765static void worker_enter_idle(struct worker *worker)
1766{
1767	struct worker_pool *pool = worker->pool;
1768
1769	if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) ||
1770	    WARN_ON_ONCE(!list_empty(&worker->entry) &&
1771			 (worker->hentry.next || worker->hentry.pprev)))
1772		return;
1773
1774	/* can't use worker_set_flags(), also called from create_worker() */
1775	worker->flags |= WORKER_IDLE;
1776	pool->nr_idle++;
1777	worker->last_active = jiffies;
1778
1779	/* idle_list is LIFO */
1780	list_add(&worker->entry, &pool->idle_list);
1781
1782	if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1783		mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
1784
1785	/*
1786	 * Sanity check nr_running.  Because unbind_workers() releases
1787	 * pool->lock between setting %WORKER_UNBOUND and zapping
1788	 * nr_running, the warning may trigger spuriously.  Check iff
1789	 * unbind is not in progress.
1790	 */
1791	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
1792		     pool->nr_workers == pool->nr_idle &&
1793		     atomic_read(&pool->nr_running));
1794}
1795
1796/**
1797 * worker_leave_idle - leave idle state
1798 * @worker: worker which is leaving idle state
1799 *
1800 * @worker is leaving idle state.  Update stats.
1801 *
1802 * LOCKING:
1803 * spin_lock_irq(pool->lock).
1804 */
1805static void worker_leave_idle(struct worker *worker)
1806{
1807	struct worker_pool *pool = worker->pool;
1808
1809	if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE)))
1810		return;
1811	worker_clr_flags(worker, WORKER_IDLE);
1812	pool->nr_idle--;
1813	list_del_init(&worker->entry);
1814}
1815
1816static struct worker *alloc_worker(int node)
1817{
1818	struct worker *worker;
1819
1820	worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node);
1821	if (worker) {
1822		INIT_LIST_HEAD(&worker->entry);
1823		INIT_LIST_HEAD(&worker->scheduled);
1824		INIT_LIST_HEAD(&worker->node);
1825		/* on creation a worker is in !idle && prep state */
1826		worker->flags = WORKER_PREP;
1827	}
1828	return worker;
1829}
1830
1831/**
1832 * worker_attach_to_pool() - attach a worker to a pool
1833 * @worker: worker to be attached
1834 * @pool: the target pool
1835 *
1836 * Attach @worker to @pool.  Once attached, the %WORKER_UNBOUND flag and
1837 * cpu-binding of @worker are kept coordinated with the pool across
1838 * cpu-[un]hotplugs.
1839 */
1840static void worker_attach_to_pool(struct worker *worker,
1841				   struct worker_pool *pool)
1842{
1843	mutex_lock(&wq_pool_attach_mutex);
1844
1845	/*
1846	 * set_cpus_allowed_ptr() will fail if the cpumask doesn't have any
1847	 * online CPUs.  It'll be re-applied when any of the CPUs come up.
1848	 */
1849	set_cpus_allowed_ptr(worker->task, pool->attrs->cpumask);
1850
1851	/*
1852	 * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains
1853	 * stable across this function.  See the comments above the flag
1854	 * definition for details.
1855	 */
1856	if (pool->flags & POOL_DISASSOCIATED)
1857		worker->flags |= WORKER_UNBOUND;
1858
1859	list_add_tail(&worker->node, &pool->workers);
1860	worker->pool = pool;
1861
1862	mutex_unlock(&wq_pool_attach_mutex);
1863}
1864
1865/**
1866 * worker_detach_from_pool() - detach a worker from its pool
1867 * @worker: worker which is attached to its pool
 
1868 *
1869 * Undo the attaching which had been done in worker_attach_to_pool().  The
1870 * caller worker shouldn't access to the pool after detached except it has
1871 * other reference to the pool.
1872 */
1873static void worker_detach_from_pool(struct worker *worker)
 
1874{
1875	struct worker_pool *pool = worker->pool;
1876	struct completion *detach_completion = NULL;
1877
1878	mutex_lock(&wq_pool_attach_mutex);
1879
1880	list_del(&worker->node);
1881	worker->pool = NULL;
1882
1883	if (list_empty(&pool->workers))
1884		detach_completion = pool->detach_completion;
1885	mutex_unlock(&wq_pool_attach_mutex);
1886
1887	/* clear leftover flags without pool->lock after it is detached */
1888	worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND);
1889
1890	if (detach_completion)
1891		complete(detach_completion);
1892}
1893
1894/**
1895 * create_worker - create a new workqueue worker
1896 * @pool: pool the new worker will belong to
1897 *
1898 * Create and start a new worker which is attached to @pool.
1899 *
1900 * CONTEXT:
1901 * Might sleep.  Does GFP_KERNEL allocations.
1902 *
1903 * Return:
1904 * Pointer to the newly created worker.
1905 */
1906static struct worker *create_worker(struct worker_pool *pool)
1907{
1908	struct worker *worker = NULL;
1909	int id = -1;
1910	char id_buf[16];
1911
1912	/* ID is needed to determine kthread name */
1913	id = ida_simple_get(&pool->worker_ida, 0, 0, GFP_KERNEL);
1914	if (id < 0)
1915		goto fail;
1916
1917	worker = alloc_worker(pool->node);
1918	if (!worker)
1919		goto fail;
1920
 
1921	worker->id = id;
1922
1923	if (pool->cpu >= 0)
1924		snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id,
1925			 pool->attrs->nice < 0  ? "H" : "");
1926	else
1927		snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id);
1928
1929	worker->task = kthread_create_on_node(worker_thread, worker, pool->node,
1930					      "kworker/%s", id_buf);
1931	if (IS_ERR(worker->task))
1932		goto fail;
1933
1934	set_user_nice(worker->task, pool->attrs->nice);
1935	kthread_bind_mask(worker->task, pool->attrs->cpumask);
1936
1937	/* successful, attach the worker to the pool */
1938	worker_attach_to_pool(worker, pool);
1939
1940	/* start the newly created worker */
1941	spin_lock_irq(&pool->lock);
1942	worker->pool->nr_workers++;
1943	worker_enter_idle(worker);
1944	wake_up_process(worker->task);
1945	spin_unlock_irq(&pool->lock);
1946
1947	return worker;
1948
1949fail:
1950	if (id >= 0)
1951		ida_simple_remove(&pool->worker_ida, id);
1952	kfree(worker);
1953	return NULL;
1954}
1955
1956/**
1957 * destroy_worker - destroy a workqueue worker
1958 * @worker: worker to be destroyed
1959 *
1960 * Destroy @worker and adjust @pool stats accordingly.  The worker should
1961 * be idle.
1962 *
1963 * CONTEXT:
1964 * spin_lock_irq(pool->lock).
1965 */
1966static void destroy_worker(struct worker *worker)
1967{
1968	struct worker_pool *pool = worker->pool;
1969
1970	lockdep_assert_held(&pool->lock);
1971
1972	/* sanity check frenzy */
1973	if (WARN_ON(worker->current_work) ||
1974	    WARN_ON(!list_empty(&worker->scheduled)) ||
1975	    WARN_ON(!(worker->flags & WORKER_IDLE)))
1976		return;
1977
1978	pool->nr_workers--;
1979	pool->nr_idle--;
1980
1981	list_del_init(&worker->entry);
1982	worker->flags |= WORKER_DIE;
1983	wake_up_process(worker->task);
1984}
1985
1986static void idle_worker_timeout(struct timer_list *t)
1987{
1988	struct worker_pool *pool = from_timer(pool, t, idle_timer);
1989
1990	spin_lock_irq(&pool->lock);
1991
1992	while (too_many_workers(pool)) {
1993		struct worker *worker;
1994		unsigned long expires;
1995
1996		/* idle_list is kept in LIFO order, check the last one */
1997		worker = list_entry(pool->idle_list.prev, struct worker, entry);
1998		expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1999
2000		if (time_before(jiffies, expires)) {
2001			mod_timer(&pool->idle_timer, expires);
2002			break;
2003		}
2004
2005		destroy_worker(worker);
2006	}
2007
2008	spin_unlock_irq(&pool->lock);
2009}
2010
2011static void send_mayday(struct work_struct *work)
2012{
2013	struct pool_workqueue *pwq = get_work_pwq(work);
2014	struct workqueue_struct *wq = pwq->wq;
2015
2016	lockdep_assert_held(&wq_mayday_lock);
2017
2018	if (!wq->rescuer)
2019		return;
2020
2021	/* mayday mayday mayday */
2022	if (list_empty(&pwq->mayday_node)) {
2023		/*
2024		 * If @pwq is for an unbound wq, its base ref may be put at
2025		 * any time due to an attribute change.  Pin @pwq until the
2026		 * rescuer is done with it.
2027		 */
2028		get_pwq(pwq);
2029		list_add_tail(&pwq->mayday_node, &wq->maydays);
2030		wake_up_process(wq->rescuer->task);
2031	}
2032}
2033
2034static void pool_mayday_timeout(struct timer_list *t)
2035{
2036	struct worker_pool *pool = from_timer(pool, t, mayday_timer);
2037	struct work_struct *work;
2038
2039	spin_lock_irq(&pool->lock);
2040	spin_lock(&wq_mayday_lock);		/* for wq->maydays */
2041
2042	if (need_to_create_worker(pool)) {
2043		/*
2044		 * We've been trying to create a new worker but
2045		 * haven't been successful.  We might be hitting an
2046		 * allocation deadlock.  Send distress signals to
2047		 * rescuers.
2048		 */
2049		list_for_each_entry(work, &pool->worklist, entry)
2050			send_mayday(work);
2051	}
2052
2053	spin_unlock(&wq_mayday_lock);
2054	spin_unlock_irq(&pool->lock);
2055
2056	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
2057}
2058
2059/**
2060 * maybe_create_worker - create a new worker if necessary
2061 * @pool: pool to create a new worker for
2062 *
2063 * Create a new worker for @pool if necessary.  @pool is guaranteed to
2064 * have at least one idle worker on return from this function.  If
2065 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
2066 * sent to all rescuers with works scheduled on @pool to resolve
2067 * possible allocation deadlock.
2068 *
2069 * On return, need_to_create_worker() is guaranteed to be %false and
2070 * may_start_working() %true.
2071 *
2072 * LOCKING:
2073 * spin_lock_irq(pool->lock) which may be released and regrabbed
2074 * multiple times.  Does GFP_KERNEL allocations.  Called only from
2075 * manager.
2076 */
2077static void maybe_create_worker(struct worker_pool *pool)
2078__releases(&pool->lock)
2079__acquires(&pool->lock)
2080{
2081restart:
2082	spin_unlock_irq(&pool->lock);
2083
2084	/* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
2085	mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
2086
2087	while (true) {
2088		if (create_worker(pool) || !need_to_create_worker(pool))
2089			break;
2090
2091		schedule_timeout_interruptible(CREATE_COOLDOWN);
2092
2093		if (!need_to_create_worker(pool))
2094			break;
2095	}
2096
2097	del_timer_sync(&pool->mayday_timer);
2098	spin_lock_irq(&pool->lock);
2099	/*
2100	 * This is necessary even after a new worker was just successfully
2101	 * created as @pool->lock was dropped and the new worker might have
2102	 * already become busy.
2103	 */
2104	if (need_to_create_worker(pool))
2105		goto restart;
2106}
2107
2108/**
2109 * manage_workers - manage worker pool
2110 * @worker: self
2111 *
2112 * Assume the manager role and manage the worker pool @worker belongs
2113 * to.  At any given time, there can be only zero or one manager per
2114 * pool.  The exclusion is handled automatically by this function.
2115 *
2116 * The caller can safely start processing works on false return.  On
2117 * true return, it's guaranteed that need_to_create_worker() is false
2118 * and may_start_working() is true.
2119 *
2120 * CONTEXT:
2121 * spin_lock_irq(pool->lock) which may be released and regrabbed
2122 * multiple times.  Does GFP_KERNEL allocations.
2123 *
2124 * Return:
2125 * %false if the pool doesn't need management and the caller can safely
2126 * start processing works, %true if management function was performed and
2127 * the conditions that the caller verified before calling the function may
2128 * no longer be true.
2129 */
2130static bool manage_workers(struct worker *worker)
2131{
2132	struct worker_pool *pool = worker->pool;
2133
2134	if (pool->flags & POOL_MANAGER_ACTIVE)
 
 
 
 
 
 
 
 
 
 
2135		return false;
2136
2137	pool->flags |= POOL_MANAGER_ACTIVE;
2138	pool->manager = worker;
2139
2140	maybe_create_worker(pool);
2141
2142	pool->manager = NULL;
2143	pool->flags &= ~POOL_MANAGER_ACTIVE;
2144	wake_up(&wq_manager_wait);
2145	return true;
2146}
2147
2148/**
2149 * process_one_work - process single work
2150 * @worker: self
2151 * @work: work to process
2152 *
2153 * Process @work.  This function contains all the logics necessary to
2154 * process a single work including synchronization against and
2155 * interaction with other workers on the same cpu, queueing and
2156 * flushing.  As long as context requirement is met, any worker can
2157 * call this function to process a work.
2158 *
2159 * CONTEXT:
2160 * spin_lock_irq(pool->lock) which is released and regrabbed.
2161 */
2162static void process_one_work(struct worker *worker, struct work_struct *work)
2163__releases(&pool->lock)
2164__acquires(&pool->lock)
2165{
2166	struct pool_workqueue *pwq = get_work_pwq(work);
2167	struct worker_pool *pool = worker->pool;
2168	bool cpu_intensive = pwq->wq->flags & WQ_CPU_INTENSIVE;
2169	int work_color;
2170	struct worker *collision;
2171#ifdef CONFIG_LOCKDEP
2172	/*
2173	 * It is permissible to free the struct work_struct from
2174	 * inside the function that is called from it, this we need to
2175	 * take into account for lockdep too.  To avoid bogus "held
2176	 * lock freed" warnings as well as problems when looking into
2177	 * work->lockdep_map, make a copy and use that here.
2178	 */
2179	struct lockdep_map lockdep_map;
2180
2181	lockdep_copy_map(&lockdep_map, &work->lockdep_map);
2182#endif
2183	/* ensure we're on the correct CPU */
2184	WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) &&
2185		     raw_smp_processor_id() != pool->cpu);
2186
2187	/*
2188	 * A single work shouldn't be executed concurrently by
2189	 * multiple workers on a single cpu.  Check whether anyone is
2190	 * already processing the work.  If so, defer the work to the
2191	 * currently executing one.
2192	 */
2193	collision = find_worker_executing_work(pool, work);
2194	if (unlikely(collision)) {
2195		move_linked_works(work, &collision->scheduled, NULL);
2196		return;
2197	}
2198
2199	/* claim and dequeue */
2200	debug_work_deactivate(work);
2201	hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work);
2202	worker->current_work = work;
2203	worker->current_func = work->func;
2204	worker->current_pwq = pwq;
2205	work_color = get_work_color(work);
2206
2207	/*
2208	 * Record wq name for cmdline and debug reporting, may get
2209	 * overridden through set_worker_desc().
2210	 */
2211	strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN);
2212
2213	list_del_init(&work->entry);
2214
2215	/*
2216	 * CPU intensive works don't participate in concurrency management.
2217	 * They're the scheduler's responsibility.  This takes @worker out
2218	 * of concurrency management and the next code block will chain
2219	 * execution of the pending work items.
2220	 */
2221	if (unlikely(cpu_intensive))
2222		worker_set_flags(worker, WORKER_CPU_INTENSIVE);
2223
2224	/*
2225	 * Wake up another worker if necessary.  The condition is always
2226	 * false for normal per-cpu workers since nr_running would always
2227	 * be >= 1 at this point.  This is used to chain execution of the
2228	 * pending work items for WORKER_NOT_RUNNING workers such as the
2229	 * UNBOUND and CPU_INTENSIVE ones.
2230	 */
2231	if (need_more_worker(pool))
2232		wake_up_worker(pool);
2233
2234	/*
2235	 * Record the last pool and clear PENDING which should be the last
2236	 * update to @work.  Also, do this inside @pool->lock so that
2237	 * PENDING and queued state changes happen together while IRQ is
2238	 * disabled.
2239	 */
2240	set_work_pool_and_clear_pending(work, pool->id);
2241
2242	spin_unlock_irq(&pool->lock);
2243
2244	lock_map_acquire(&pwq->wq->lockdep_map);
2245	lock_map_acquire(&lockdep_map);
2246	/*
2247	 * Strictly speaking we should mark the invariant state without holding
2248	 * any locks, that is, before these two lock_map_acquire()'s.
2249	 *
2250	 * However, that would result in:
2251	 *
2252	 *   A(W1)
2253	 *   WFC(C)
2254	 *		A(W1)
2255	 *		C(C)
2256	 *
2257	 * Which would create W1->C->W1 dependencies, even though there is no
2258	 * actual deadlock possible. There are two solutions, using a
2259	 * read-recursive acquire on the work(queue) 'locks', but this will then
2260	 * hit the lockdep limitation on recursive locks, or simply discard
2261	 * these locks.
2262	 *
2263	 * AFAICT there is no possible deadlock scenario between the
2264	 * flush_work() and complete() primitives (except for single-threaded
2265	 * workqueues), so hiding them isn't a problem.
2266	 */
2267	lockdep_invariant_state(true);
2268	trace_workqueue_execute_start(work);
2269	worker->current_func(work);
2270	/*
2271	 * While we must be careful to not use "work" after this, the trace
2272	 * point will only record its address.
2273	 */
2274	trace_workqueue_execute_end(work);
2275	lock_map_release(&lockdep_map);
2276	lock_map_release(&pwq->wq->lockdep_map);
2277
2278	if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2279		pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n"
2280		       "     last function: %ps\n",
2281		       current->comm, preempt_count(), task_pid_nr(current),
2282		       worker->current_func);
2283		debug_show_held_locks(current);
2284		dump_stack();
2285	}
2286
2287	/*
2288	 * The following prevents a kworker from hogging CPU on !PREEMPT
2289	 * kernels, where a requeueing work item waiting for something to
2290	 * happen could deadlock with stop_machine as such work item could
2291	 * indefinitely requeue itself while all other CPUs are trapped in
2292	 * stop_machine. At the same time, report a quiescent RCU state so
2293	 * the same condition doesn't freeze RCU.
2294	 */
2295	cond_resched();
2296
2297	spin_lock_irq(&pool->lock);
2298
2299	/* clear cpu intensive status */
2300	if (unlikely(cpu_intensive))
2301		worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2302
2303	/* tag the worker for identification in schedule() */
2304	worker->last_func = worker->current_func;
2305
2306	/* we're done with it, release */
2307	hash_del(&worker->hentry);
2308	worker->current_work = NULL;
2309	worker->current_func = NULL;
2310	worker->current_pwq = NULL;
 
2311	pwq_dec_nr_in_flight(pwq, work_color);
2312}
2313
2314/**
2315 * process_scheduled_works - process scheduled works
2316 * @worker: self
2317 *
2318 * Process all scheduled works.  Please note that the scheduled list
2319 * may change while processing a work, so this function repeatedly
2320 * fetches a work from the top and executes it.
2321 *
2322 * CONTEXT:
2323 * spin_lock_irq(pool->lock) which may be released and regrabbed
2324 * multiple times.
2325 */
2326static void process_scheduled_works(struct worker *worker)
2327{
2328	while (!list_empty(&worker->scheduled)) {
2329		struct work_struct *work = list_first_entry(&worker->scheduled,
2330						struct work_struct, entry);
2331		process_one_work(worker, work);
2332	}
2333}
2334
2335static void set_pf_worker(bool val)
2336{
2337	mutex_lock(&wq_pool_attach_mutex);
2338	if (val)
2339		current->flags |= PF_WQ_WORKER;
2340	else
2341		current->flags &= ~PF_WQ_WORKER;
2342	mutex_unlock(&wq_pool_attach_mutex);
2343}
2344
2345/**
2346 * worker_thread - the worker thread function
2347 * @__worker: self
2348 *
2349 * The worker thread function.  All workers belong to a worker_pool -
2350 * either a per-cpu one or dynamic unbound one.  These workers process all
2351 * work items regardless of their specific target workqueue.  The only
2352 * exception is work items which belong to workqueues with a rescuer which
2353 * will be explained in rescuer_thread().
2354 *
2355 * Return: 0
2356 */
2357static int worker_thread(void *__worker)
2358{
2359	struct worker *worker = __worker;
2360	struct worker_pool *pool = worker->pool;
2361
2362	/* tell the scheduler that this is a workqueue worker */
2363	set_pf_worker(true);
2364woke_up:
2365	spin_lock_irq(&pool->lock);
2366
2367	/* am I supposed to die? */
2368	if (unlikely(worker->flags & WORKER_DIE)) {
2369		spin_unlock_irq(&pool->lock);
2370		WARN_ON_ONCE(!list_empty(&worker->entry));
2371		set_pf_worker(false);
2372
2373		set_task_comm(worker->task, "kworker/dying");
2374		ida_simple_remove(&pool->worker_ida, worker->id);
2375		worker_detach_from_pool(worker);
2376		kfree(worker);
2377		return 0;
2378	}
2379
2380	worker_leave_idle(worker);
2381recheck:
2382	/* no more worker necessary? */
2383	if (!need_more_worker(pool))
2384		goto sleep;
2385
2386	/* do we need to manage? */
2387	if (unlikely(!may_start_working(pool)) && manage_workers(worker))
2388		goto recheck;
2389
2390	/*
2391	 * ->scheduled list can only be filled while a worker is
2392	 * preparing to process a work or actually processing it.
2393	 * Make sure nobody diddled with it while I was sleeping.
2394	 */
2395	WARN_ON_ONCE(!list_empty(&worker->scheduled));
2396
2397	/*
2398	 * Finish PREP stage.  We're guaranteed to have at least one idle
2399	 * worker or that someone else has already assumed the manager
2400	 * role.  This is where @worker starts participating in concurrency
2401	 * management if applicable and concurrency management is restored
2402	 * after being rebound.  See rebind_workers() for details.
2403	 */
2404	worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND);
2405
2406	do {
2407		struct work_struct *work =
2408			list_first_entry(&pool->worklist,
2409					 struct work_struct, entry);
2410
2411		pool->watchdog_ts = jiffies;
2412
2413		if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2414			/* optimization path, not strictly necessary */
2415			process_one_work(worker, work);
2416			if (unlikely(!list_empty(&worker->scheduled)))
2417				process_scheduled_works(worker);
2418		} else {
2419			move_linked_works(work, &worker->scheduled, NULL);
2420			process_scheduled_works(worker);
2421		}
2422	} while (keep_working(pool));
2423
2424	worker_set_flags(worker, WORKER_PREP);
2425sleep:
2426	/*
2427	 * pool->lock is held and there's no work to process and no need to
2428	 * manage, sleep.  Workers are woken up only while holding
2429	 * pool->lock or from local cpu, so setting the current state
2430	 * before releasing pool->lock is enough to prevent losing any
2431	 * event.
2432	 */
2433	worker_enter_idle(worker);
2434	__set_current_state(TASK_IDLE);
2435	spin_unlock_irq(&pool->lock);
2436	schedule();
2437	goto woke_up;
2438}
2439
2440/**
2441 * rescuer_thread - the rescuer thread function
2442 * @__rescuer: self
2443 *
2444 * Workqueue rescuer thread function.  There's one rescuer for each
2445 * workqueue which has WQ_MEM_RECLAIM set.
2446 *
2447 * Regular work processing on a pool may block trying to create a new
2448 * worker which uses GFP_KERNEL allocation which has slight chance of
2449 * developing into deadlock if some works currently on the same queue
2450 * need to be processed to satisfy the GFP_KERNEL allocation.  This is
2451 * the problem rescuer solves.
2452 *
2453 * When such condition is possible, the pool summons rescuers of all
2454 * workqueues which have works queued on the pool and let them process
2455 * those works so that forward progress can be guaranteed.
2456 *
2457 * This should happen rarely.
2458 *
2459 * Return: 0
2460 */
2461static int rescuer_thread(void *__rescuer)
2462{
2463	struct worker *rescuer = __rescuer;
2464	struct workqueue_struct *wq = rescuer->rescue_wq;
2465	struct list_head *scheduled = &rescuer->scheduled;
2466	bool should_stop;
2467
2468	set_user_nice(current, RESCUER_NICE_LEVEL);
2469
2470	/*
2471	 * Mark rescuer as worker too.  As WORKER_PREP is never cleared, it
2472	 * doesn't participate in concurrency management.
2473	 */
2474	set_pf_worker(true);
2475repeat:
2476	set_current_state(TASK_IDLE);
2477
2478	/*
2479	 * By the time the rescuer is requested to stop, the workqueue
2480	 * shouldn't have any work pending, but @wq->maydays may still have
2481	 * pwq(s) queued.  This can happen by non-rescuer workers consuming
2482	 * all the work items before the rescuer got to them.  Go through
2483	 * @wq->maydays processing before acting on should_stop so that the
2484	 * list is always empty on exit.
2485	 */
2486	should_stop = kthread_should_stop();
2487
2488	/* see whether any pwq is asking for help */
2489	spin_lock_irq(&wq_mayday_lock);
2490
2491	while (!list_empty(&wq->maydays)) {
2492		struct pool_workqueue *pwq = list_first_entry(&wq->maydays,
2493					struct pool_workqueue, mayday_node);
2494		struct worker_pool *pool = pwq->pool;
2495		struct work_struct *work, *n;
2496		bool first = true;
2497
2498		__set_current_state(TASK_RUNNING);
2499		list_del_init(&pwq->mayday_node);
2500
2501		spin_unlock_irq(&wq_mayday_lock);
2502
2503		worker_attach_to_pool(rescuer, pool);
2504
2505		spin_lock_irq(&pool->lock);
 
2506
2507		/*
2508		 * Slurp in all works issued via this workqueue and
2509		 * process'em.
2510		 */
2511		WARN_ON_ONCE(!list_empty(scheduled));
2512		list_for_each_entry_safe(work, n, &pool->worklist, entry) {
2513			if (get_work_pwq(work) == pwq) {
2514				if (first)
2515					pool->watchdog_ts = jiffies;
2516				move_linked_works(work, scheduled, &n);
2517			}
2518			first = false;
2519		}
2520
2521		if (!list_empty(scheduled)) {
2522			process_scheduled_works(rescuer);
2523
2524			/*
2525			 * The above execution of rescued work items could
2526			 * have created more to rescue through
2527			 * pwq_activate_first_delayed() or chained
2528			 * queueing.  Let's put @pwq back on mayday list so
2529			 * that such back-to-back work items, which may be
2530			 * being used to relieve memory pressure, don't
2531			 * incur MAYDAY_INTERVAL delay inbetween.
2532			 */
2533			if (need_to_create_worker(pool)) {
2534				spin_lock(&wq_mayday_lock);
2535				get_pwq(pwq);
2536				list_move_tail(&pwq->mayday_node, &wq->maydays);
2537				spin_unlock(&wq_mayday_lock);
2538			}
2539		}
2540
2541		/*
2542		 * Put the reference grabbed by send_mayday().  @pool won't
2543		 * go away while we're still attached to it.
2544		 */
2545		put_pwq(pwq);
2546
2547		/*
2548		 * Leave this pool.  If need_more_worker() is %true, notify a
2549		 * regular worker; otherwise, we end up with 0 concurrency
2550		 * and stalling the execution.
2551		 */
2552		if (need_more_worker(pool))
2553			wake_up_worker(pool);
2554
 
2555		spin_unlock_irq(&pool->lock);
2556
2557		worker_detach_from_pool(rescuer);
2558
2559		spin_lock_irq(&wq_mayday_lock);
2560	}
2561
2562	spin_unlock_irq(&wq_mayday_lock);
2563
2564	if (should_stop) {
2565		__set_current_state(TASK_RUNNING);
2566		set_pf_worker(false);
2567		return 0;
2568	}
2569
2570	/* rescuers should never participate in concurrency management */
2571	WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING));
2572	schedule();
2573	goto repeat;
2574}
2575
2576/**
2577 * check_flush_dependency - check for flush dependency sanity
2578 * @target_wq: workqueue being flushed
2579 * @target_work: work item being flushed (NULL for workqueue flushes)
2580 *
2581 * %current is trying to flush the whole @target_wq or @target_work on it.
2582 * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not
2583 * reclaiming memory or running on a workqueue which doesn't have
2584 * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to
2585 * a deadlock.
2586 */
2587static void check_flush_dependency(struct workqueue_struct *target_wq,
2588				   struct work_struct *target_work)
2589{
2590	work_func_t target_func = target_work ? target_work->func : NULL;
2591	struct worker *worker;
2592
2593	if (target_wq->flags & WQ_MEM_RECLAIM)
2594		return;
2595
2596	worker = current_wq_worker();
2597
2598	WARN_ONCE(current->flags & PF_MEMALLOC,
2599		  "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps",
2600		  current->pid, current->comm, target_wq->name, target_func);
2601	WARN_ONCE(worker && ((worker->current_pwq->wq->flags &
2602			      (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM),
2603		  "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps",
2604		  worker->current_pwq->wq->name, worker->current_func,
2605		  target_wq->name, target_func);
2606}
2607
2608struct wq_barrier {
2609	struct work_struct	work;
2610	struct completion	done;
2611	struct task_struct	*task;	/* purely informational */
2612};
2613
2614static void wq_barrier_func(struct work_struct *work)
2615{
2616	struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2617	complete(&barr->done);
2618}
2619
2620/**
2621 * insert_wq_barrier - insert a barrier work
2622 * @pwq: pwq to insert barrier into
2623 * @barr: wq_barrier to insert
2624 * @target: target work to attach @barr to
2625 * @worker: worker currently executing @target, NULL if @target is not executing
2626 *
2627 * @barr is linked to @target such that @barr is completed only after
2628 * @target finishes execution.  Please note that the ordering
2629 * guarantee is observed only with respect to @target and on the local
2630 * cpu.
2631 *
2632 * Currently, a queued barrier can't be canceled.  This is because
2633 * try_to_grab_pending() can't determine whether the work to be
2634 * grabbed is at the head of the queue and thus can't clear LINKED
2635 * flag of the previous work while there must be a valid next work
2636 * after a work with LINKED flag set.
2637 *
2638 * Note that when @worker is non-NULL, @target may be modified
2639 * underneath us, so we can't reliably determine pwq from @target.
2640 *
2641 * CONTEXT:
2642 * spin_lock_irq(pool->lock).
2643 */
2644static void insert_wq_barrier(struct pool_workqueue *pwq,
2645			      struct wq_barrier *barr,
2646			      struct work_struct *target, struct worker *worker)
2647{
2648	struct list_head *head;
2649	unsigned int linked = 0;
2650
2651	/*
2652	 * debugobject calls are safe here even with pool->lock locked
2653	 * as we know for sure that this will not trigger any of the
2654	 * checks and call back into the fixup functions where we
2655	 * might deadlock.
2656	 */
2657	INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
2658	__set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
2659
2660	init_completion_map(&barr->done, &target->lockdep_map);
2661
2662	barr->task = current;
2663
2664	/*
2665	 * If @target is currently being executed, schedule the
2666	 * barrier to the worker; otherwise, put it after @target.
2667	 */
2668	if (worker)
2669		head = worker->scheduled.next;
2670	else {
2671		unsigned long *bits = work_data_bits(target);
2672
2673		head = target->entry.next;
2674		/* there can already be other linked works, inherit and set */
2675		linked = *bits & WORK_STRUCT_LINKED;
2676		__set_bit(WORK_STRUCT_LINKED_BIT, bits);
2677	}
2678
2679	debug_work_activate(&barr->work);
2680	insert_work(pwq, &barr->work, head,
2681		    work_color_to_flags(WORK_NO_COLOR) | linked);
2682}
2683
2684/**
2685 * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing
2686 * @wq: workqueue being flushed
2687 * @flush_color: new flush color, < 0 for no-op
2688 * @work_color: new work color, < 0 for no-op
2689 *
2690 * Prepare pwqs for workqueue flushing.
2691 *
2692 * If @flush_color is non-negative, flush_color on all pwqs should be
2693 * -1.  If no pwq has in-flight commands at the specified color, all
2694 * pwq->flush_color's stay at -1 and %false is returned.  If any pwq
2695 * has in flight commands, its pwq->flush_color is set to
2696 * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq
2697 * wakeup logic is armed and %true is returned.
2698 *
2699 * The caller should have initialized @wq->first_flusher prior to
2700 * calling this function with non-negative @flush_color.  If
2701 * @flush_color is negative, no flush color update is done and %false
2702 * is returned.
2703 *
2704 * If @work_color is non-negative, all pwqs should have the same
2705 * work_color which is previous to @work_color and all will be
2706 * advanced to @work_color.
2707 *
2708 * CONTEXT:
2709 * mutex_lock(wq->mutex).
2710 *
2711 * Return:
2712 * %true if @flush_color >= 0 and there's something to flush.  %false
2713 * otherwise.
2714 */
2715static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq,
2716				      int flush_color, int work_color)
2717{
2718	bool wait = false;
2719	struct pool_workqueue *pwq;
2720
2721	if (flush_color >= 0) {
2722		WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush));
2723		atomic_set(&wq->nr_pwqs_to_flush, 1);
2724	}
2725
2726	for_each_pwq(pwq, wq) {
2727		struct worker_pool *pool = pwq->pool;
2728
2729		spin_lock_irq(&pool->lock);
2730
2731		if (flush_color >= 0) {
2732			WARN_ON_ONCE(pwq->flush_color != -1);
2733
2734			if (pwq->nr_in_flight[flush_color]) {
2735				pwq->flush_color = flush_color;
2736				atomic_inc(&wq->nr_pwqs_to_flush);
2737				wait = true;
2738			}
2739		}
2740
2741		if (work_color >= 0) {
2742			WARN_ON_ONCE(work_color != work_next_color(pwq->work_color));
2743			pwq->work_color = work_color;
2744		}
2745
2746		spin_unlock_irq(&pool->lock);
2747	}
2748
2749	if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush))
2750		complete(&wq->first_flusher->done);
2751
2752	return wait;
2753}
2754
2755/**
2756 * flush_workqueue - ensure that any scheduled work has run to completion.
2757 * @wq: workqueue to flush
2758 *
2759 * This function sleeps until all work items which were queued on entry
2760 * have finished execution, but it is not livelocked by new incoming ones.
2761 */
2762void flush_workqueue(struct workqueue_struct *wq)
2763{
2764	struct wq_flusher this_flusher = {
2765		.list = LIST_HEAD_INIT(this_flusher.list),
2766		.flush_color = -1,
2767		.done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map),
2768	};
2769	int next_color;
2770
2771	if (WARN_ON(!wq_online))
2772		return;
2773
2774	lock_map_acquire(&wq->lockdep_map);
2775	lock_map_release(&wq->lockdep_map);
2776
2777	mutex_lock(&wq->mutex);
2778
2779	/*
2780	 * Start-to-wait phase
2781	 */
2782	next_color = work_next_color(wq->work_color);
2783
2784	if (next_color != wq->flush_color) {
2785		/*
2786		 * Color space is not full.  The current work_color
2787		 * becomes our flush_color and work_color is advanced
2788		 * by one.
2789		 */
2790		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow));
2791		this_flusher.flush_color = wq->work_color;
2792		wq->work_color = next_color;
2793
2794		if (!wq->first_flusher) {
2795			/* no flush in progress, become the first flusher */
2796			WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2797
2798			wq->first_flusher = &this_flusher;
2799
2800			if (!flush_workqueue_prep_pwqs(wq, wq->flush_color,
2801						       wq->work_color)) {
2802				/* nothing to flush, done */
2803				wq->flush_color = next_color;
2804				wq->first_flusher = NULL;
2805				goto out_unlock;
2806			}
2807		} else {
2808			/* wait in queue */
2809			WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color);
2810			list_add_tail(&this_flusher.list, &wq->flusher_queue);
2811			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2812		}
2813	} else {
2814		/*
2815		 * Oops, color space is full, wait on overflow queue.
2816		 * The next flush completion will assign us
2817		 * flush_color and transfer to flusher_queue.
2818		 */
2819		list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2820	}
2821
2822	check_flush_dependency(wq, NULL);
2823
2824	mutex_unlock(&wq->mutex);
2825
2826	wait_for_completion(&this_flusher.done);
2827
2828	/*
2829	 * Wake-up-and-cascade phase
2830	 *
2831	 * First flushers are responsible for cascading flushes and
2832	 * handling overflow.  Non-first flushers can simply return.
2833	 */
2834	if (wq->first_flusher != &this_flusher)
2835		return;
2836
2837	mutex_lock(&wq->mutex);
2838
2839	/* we might have raced, check again with mutex held */
2840	if (wq->first_flusher != &this_flusher)
2841		goto out_unlock;
2842
2843	wq->first_flusher = NULL;
2844
2845	WARN_ON_ONCE(!list_empty(&this_flusher.list));
2846	WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color);
2847
2848	while (true) {
2849		struct wq_flusher *next, *tmp;
2850
2851		/* complete all the flushers sharing the current flush color */
2852		list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2853			if (next->flush_color != wq->flush_color)
2854				break;
2855			list_del_init(&next->list);
2856			complete(&next->done);
2857		}
2858
2859		WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) &&
2860			     wq->flush_color != work_next_color(wq->work_color));
2861
2862		/* this flush_color is finished, advance by one */
2863		wq->flush_color = work_next_color(wq->flush_color);
2864
2865		/* one color has been freed, handle overflow queue */
2866		if (!list_empty(&wq->flusher_overflow)) {
2867			/*
2868			 * Assign the same color to all overflowed
2869			 * flushers, advance work_color and append to
2870			 * flusher_queue.  This is the start-to-wait
2871			 * phase for these overflowed flushers.
2872			 */
2873			list_for_each_entry(tmp, &wq->flusher_overflow, list)
2874				tmp->flush_color = wq->work_color;
2875
2876			wq->work_color = work_next_color(wq->work_color);
2877
2878			list_splice_tail_init(&wq->flusher_overflow,
2879					      &wq->flusher_queue);
2880			flush_workqueue_prep_pwqs(wq, -1, wq->work_color);
2881		}
2882
2883		if (list_empty(&wq->flusher_queue)) {
2884			WARN_ON_ONCE(wq->flush_color != wq->work_color);
2885			break;
2886		}
2887
2888		/*
2889		 * Need to flush more colors.  Make the next flusher
2890		 * the new first flusher and arm pwqs.
2891		 */
2892		WARN_ON_ONCE(wq->flush_color == wq->work_color);
2893		WARN_ON_ONCE(wq->flush_color != next->flush_color);
2894
2895		list_del_init(&next->list);
2896		wq->first_flusher = next;
2897
2898		if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1))
2899			break;
2900
2901		/*
2902		 * Meh... this color is already done, clear first
2903		 * flusher and repeat cascading.
2904		 */
2905		wq->first_flusher = NULL;
2906	}
2907
2908out_unlock:
2909	mutex_unlock(&wq->mutex);
2910}
2911EXPORT_SYMBOL(flush_workqueue);
2912
2913/**
2914 * drain_workqueue - drain a workqueue
2915 * @wq: workqueue to drain
2916 *
2917 * Wait until the workqueue becomes empty.  While draining is in progress,
2918 * only chain queueing is allowed.  IOW, only currently pending or running
2919 * work items on @wq can queue further work items on it.  @wq is flushed
2920 * repeatedly until it becomes empty.  The number of flushing is determined
2921 * by the depth of chaining and should be relatively short.  Whine if it
2922 * takes too long.
2923 */
2924void drain_workqueue(struct workqueue_struct *wq)
2925{
2926	unsigned int flush_cnt = 0;
2927	struct pool_workqueue *pwq;
2928
2929	/*
2930	 * __queue_work() needs to test whether there are drainers, is much
2931	 * hotter than drain_workqueue() and already looks at @wq->flags.
2932	 * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers.
2933	 */
2934	mutex_lock(&wq->mutex);
2935	if (!wq->nr_drainers++)
2936		wq->flags |= __WQ_DRAINING;
2937	mutex_unlock(&wq->mutex);
2938reflush:
2939	flush_workqueue(wq);
2940
2941	mutex_lock(&wq->mutex);
2942
2943	for_each_pwq(pwq, wq) {
2944		bool drained;
2945
2946		spin_lock_irq(&pwq->pool->lock);
2947		drained = !pwq->nr_active && list_empty(&pwq->delayed_works);
2948		spin_unlock_irq(&pwq->pool->lock);
2949
2950		if (drained)
2951			continue;
2952
2953		if (++flush_cnt == 10 ||
2954		    (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2955			pr_warn("workqueue %s: drain_workqueue() isn't complete after %u tries\n",
2956				wq->name, flush_cnt);
2957
2958		mutex_unlock(&wq->mutex);
2959		goto reflush;
2960	}
2961
2962	if (!--wq->nr_drainers)
2963		wq->flags &= ~__WQ_DRAINING;
2964	mutex_unlock(&wq->mutex);
2965}
2966EXPORT_SYMBOL_GPL(drain_workqueue);
2967
2968static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2969			     bool from_cancel)
2970{
2971	struct worker *worker = NULL;
2972	struct worker_pool *pool;
2973	struct pool_workqueue *pwq;
2974
2975	might_sleep();
2976
2977	rcu_read_lock();
2978	pool = get_work_pool(work);
2979	if (!pool) {
2980		rcu_read_unlock();
2981		return false;
2982	}
2983
2984	spin_lock_irq(&pool->lock);
2985	/* see the comment in try_to_grab_pending() with the same code */
2986	pwq = get_work_pwq(work);
2987	if (pwq) {
2988		if (unlikely(pwq->pool != pool))
2989			goto already_gone;
2990	} else {
2991		worker = find_worker_executing_work(pool, work);
2992		if (!worker)
2993			goto already_gone;
2994		pwq = worker->current_pwq;
2995	}
2996
2997	check_flush_dependency(pwq->wq, work);
2998
2999	insert_wq_barrier(pwq, barr, work, worker);
3000	spin_unlock_irq(&pool->lock);
3001
3002	/*
3003	 * Force a lock recursion deadlock when using flush_work() inside a
3004	 * single-threaded or rescuer equipped workqueue.
3005	 *
3006	 * For single threaded workqueues the deadlock happens when the work
3007	 * is after the work issuing the flush_work(). For rescuer equipped
3008	 * workqueues the deadlock happens when the rescuer stalls, blocking
3009	 * forward progress.
3010	 */
3011	if (!from_cancel &&
3012	    (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) {
3013		lock_map_acquire(&pwq->wq->lockdep_map);
3014		lock_map_release(&pwq->wq->lockdep_map);
3015	}
3016	rcu_read_unlock();
 
3017	return true;
3018already_gone:
3019	spin_unlock_irq(&pool->lock);
3020	rcu_read_unlock();
3021	return false;
3022}
3023
3024static bool __flush_work(struct work_struct *work, bool from_cancel)
3025{
3026	struct wq_barrier barr;
3027
3028	if (WARN_ON(!wq_online))
3029		return false;
3030
3031	if (WARN_ON(!work->func))
3032		return false;
3033
3034	if (!from_cancel) {
3035		lock_map_acquire(&work->lockdep_map);
3036		lock_map_release(&work->lockdep_map);
3037	}
3038
3039	if (start_flush_work(work, &barr, from_cancel)) {
3040		wait_for_completion(&barr.done);
3041		destroy_work_on_stack(&barr.work);
3042		return true;
3043	} else {
3044		return false;
3045	}
3046}
3047
3048/**
3049 * flush_work - wait for a work to finish executing the last queueing instance
3050 * @work: the work to flush
3051 *
3052 * Wait until @work has finished execution.  @work is guaranteed to be idle
3053 * on return if it hasn't been requeued since flush started.
3054 *
3055 * Return:
3056 * %true if flush_work() waited for the work to finish execution,
3057 * %false if it was already idle.
3058 */
3059bool flush_work(struct work_struct *work)
3060{
3061	return __flush_work(work, false);
 
 
 
 
 
 
 
 
 
 
 
3062}
3063EXPORT_SYMBOL_GPL(flush_work);
3064
3065struct cwt_wait {
3066	wait_queue_entry_t		wait;
3067	struct work_struct	*work;
3068};
3069
3070static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key)
3071{
3072	struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait);
3073
3074	if (cwait->work != key)
3075		return 0;
3076	return autoremove_wake_function(wait, mode, sync, key);
3077}
3078
3079static bool __cancel_work_timer(struct work_struct *work, bool is_dwork)
3080{
3081	static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq);
3082	unsigned long flags;
3083	int ret;
3084
3085	do {
3086		ret = try_to_grab_pending(work, is_dwork, &flags);
3087		/*
3088		 * If someone else is already canceling, wait for it to
3089		 * finish.  flush_work() doesn't work for PREEMPT_NONE
3090		 * because we may get scheduled between @work's completion
3091		 * and the other canceling task resuming and clearing
3092		 * CANCELING - flush_work() will return false immediately
3093		 * as @work is no longer busy, try_to_grab_pending() will
3094		 * return -ENOENT as @work is still being canceled and the
3095		 * other canceling task won't be able to clear CANCELING as
3096		 * we're hogging the CPU.
3097		 *
3098		 * Let's wait for completion using a waitqueue.  As this
3099		 * may lead to the thundering herd problem, use a custom
3100		 * wake function which matches @work along with exclusive
3101		 * wait and wakeup.
3102		 */
3103		if (unlikely(ret == -ENOENT)) {
3104			struct cwt_wait cwait;
3105
3106			init_wait(&cwait.wait);
3107			cwait.wait.func = cwt_wakefn;
3108			cwait.work = work;
3109
3110			prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait,
3111						  TASK_UNINTERRUPTIBLE);
3112			if (work_is_canceling(work))
3113				schedule();
3114			finish_wait(&cancel_waitq, &cwait.wait);
3115		}
3116	} while (unlikely(ret < 0));
3117
3118	/* tell other tasks trying to grab @work to back off */
3119	mark_work_canceling(work);
3120	local_irq_restore(flags);
3121
3122	/*
3123	 * This allows canceling during early boot.  We know that @work
3124	 * isn't executing.
3125	 */
3126	if (wq_online)
3127		__flush_work(work, true);
3128
3129	clear_work_data(work);
3130
3131	/*
3132	 * Paired with prepare_to_wait() above so that either
3133	 * waitqueue_active() is visible here or !work_is_canceling() is
3134	 * visible there.
3135	 */
3136	smp_mb();
3137	if (waitqueue_active(&cancel_waitq))
3138		__wake_up(&cancel_waitq, TASK_NORMAL, 1, work);
3139
3140	return ret;
3141}
3142
3143/**
3144 * cancel_work_sync - cancel a work and wait for it to finish
3145 * @work: the work to cancel
3146 *
3147 * Cancel @work and wait for its execution to finish.  This function
3148 * can be used even if the work re-queues itself or migrates to
3149 * another workqueue.  On return from this function, @work is
3150 * guaranteed to be not pending or executing on any CPU.
3151 *
3152 * cancel_work_sync(&delayed_work->work) must not be used for
3153 * delayed_work's.  Use cancel_delayed_work_sync() instead.
3154 *
3155 * The caller must ensure that the workqueue on which @work was last
3156 * queued can't be destroyed before this function returns.
3157 *
3158 * Return:
3159 * %true if @work was pending, %false otherwise.
3160 */
3161bool cancel_work_sync(struct work_struct *work)
3162{
3163	return __cancel_work_timer(work, false);
3164}
3165EXPORT_SYMBOL_GPL(cancel_work_sync);
3166
3167/**
3168 * flush_delayed_work - wait for a dwork to finish executing the last queueing
3169 * @dwork: the delayed work to flush
3170 *
3171 * Delayed timer is cancelled and the pending work is queued for
3172 * immediate execution.  Like flush_work(), this function only
3173 * considers the last queueing instance of @dwork.
3174 *
3175 * Return:
3176 * %true if flush_work() waited for the work to finish execution,
3177 * %false if it was already idle.
3178 */
3179bool flush_delayed_work(struct delayed_work *dwork)
3180{
3181	local_irq_disable();
3182	if (del_timer_sync(&dwork->timer))
3183		__queue_work(dwork->cpu, dwork->wq, &dwork->work);
3184	local_irq_enable();
3185	return flush_work(&dwork->work);
3186}
3187EXPORT_SYMBOL(flush_delayed_work);
3188
3189/**
3190 * flush_rcu_work - wait for a rwork to finish executing the last queueing
3191 * @rwork: the rcu work to flush
3192 *
3193 * Return:
3194 * %true if flush_rcu_work() waited for the work to finish execution,
3195 * %false if it was already idle.
3196 */
3197bool flush_rcu_work(struct rcu_work *rwork)
3198{
3199	if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) {
3200		rcu_barrier();
3201		flush_work(&rwork->work);
3202		return true;
3203	} else {
3204		return flush_work(&rwork->work);
3205	}
3206}
3207EXPORT_SYMBOL(flush_rcu_work);
3208
3209static bool __cancel_work(struct work_struct *work, bool is_dwork)
3210{
3211	unsigned long flags;
3212	int ret;
3213
3214	do {
3215		ret = try_to_grab_pending(work, is_dwork, &flags);
3216	} while (unlikely(ret == -EAGAIN));
3217
3218	if (unlikely(ret < 0))
3219		return false;
3220
3221	set_work_pool_and_clear_pending(work, get_work_pool_id(work));
3222	local_irq_restore(flags);
3223	return ret;
3224}
3225
3226/**
3227 * cancel_delayed_work - cancel a delayed work
3228 * @dwork: delayed_work to cancel
3229 *
3230 * Kill off a pending delayed_work.
3231 *
3232 * Return: %true if @dwork was pending and canceled; %false if it wasn't
3233 * pending.
3234 *
3235 * Note:
3236 * The work callback function may still be running on return, unless
3237 * it returns %true and the work doesn't re-arm itself.  Explicitly flush or
3238 * use cancel_delayed_work_sync() to wait on it.
3239 *
3240 * This function is safe to call from any context including IRQ handler.
3241 */
3242bool cancel_delayed_work(struct delayed_work *dwork)
3243{
3244	return __cancel_work(&dwork->work, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
3245}
3246EXPORT_SYMBOL(cancel_delayed_work);
3247
3248/**
3249 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
3250 * @dwork: the delayed work cancel
3251 *
3252 * This is cancel_work_sync() for delayed works.
3253 *
3254 * Return:
3255 * %true if @dwork was pending, %false otherwise.
3256 */
3257bool cancel_delayed_work_sync(struct delayed_work *dwork)
3258{
3259	return __cancel_work_timer(&dwork->work, true);
3260}
3261EXPORT_SYMBOL(cancel_delayed_work_sync);
3262
3263/**
3264 * schedule_on_each_cpu - execute a function synchronously on each online CPU
3265 * @func: the function to call
3266 *
3267 * schedule_on_each_cpu() executes @func on each online CPU using the
3268 * system workqueue and blocks until all CPUs have completed.
3269 * schedule_on_each_cpu() is very slow.
3270 *
3271 * Return:
3272 * 0 on success, -errno on failure.
3273 */
3274int schedule_on_each_cpu(work_func_t func)
3275{
3276	int cpu;
3277	struct work_struct __percpu *works;
3278
3279	works = alloc_percpu(struct work_struct);
3280	if (!works)
3281		return -ENOMEM;
3282
3283	get_online_cpus();
3284
3285	for_each_online_cpu(cpu) {
3286		struct work_struct *work = per_cpu_ptr(works, cpu);
3287
3288		INIT_WORK(work, func);
3289		schedule_work_on(cpu, work);
3290	}
3291
3292	for_each_online_cpu(cpu)
3293		flush_work(per_cpu_ptr(works, cpu));
3294
3295	put_online_cpus();
3296	free_percpu(works);
3297	return 0;
3298}
3299
3300/**
3301 * execute_in_process_context - reliably execute the routine with user context
3302 * @fn:		the function to execute
3303 * @ew:		guaranteed storage for the execute work structure (must
3304 *		be available when the work executes)
3305 *
3306 * Executes the function immediately if process context is available,
3307 * otherwise schedules the function for delayed execution.
3308 *
3309 * Return:	0 - function was executed
3310 *		1 - function was scheduled for execution
3311 */
3312int execute_in_process_context(work_func_t fn, struct execute_work *ew)
3313{
3314	if (!in_interrupt()) {
3315		fn(&ew->work);
3316		return 0;
3317	}
3318
3319	INIT_WORK(&ew->work, fn);
3320	schedule_work(&ew->work);
3321
3322	return 1;
3323}
3324EXPORT_SYMBOL_GPL(execute_in_process_context);
3325
3326/**
3327 * free_workqueue_attrs - free a workqueue_attrs
3328 * @attrs: workqueue_attrs to free
3329 *
3330 * Undo alloc_workqueue_attrs().
3331 */
3332void free_workqueue_attrs(struct workqueue_attrs *attrs)
3333{
3334	if (attrs) {
3335		free_cpumask_var(attrs->cpumask);
3336		kfree(attrs);
3337	}
3338}
3339
3340/**
3341 * alloc_workqueue_attrs - allocate a workqueue_attrs
 
3342 *
3343 * Allocate a new workqueue_attrs, initialize with default settings and
3344 * return it.
3345 *
3346 * Return: The allocated new workqueue_attr on success. %NULL on failure.
3347 */
3348struct workqueue_attrs *alloc_workqueue_attrs(void)
3349{
3350	struct workqueue_attrs *attrs;
3351
3352	attrs = kzalloc(sizeof(*attrs), GFP_KERNEL);
3353	if (!attrs)
3354		goto fail;
3355	if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL))
3356		goto fail;
3357
3358	cpumask_copy(attrs->cpumask, cpu_possible_mask);
3359	return attrs;
3360fail:
3361	free_workqueue_attrs(attrs);
3362	return NULL;
3363}
3364
3365static void copy_workqueue_attrs(struct workqueue_attrs *to,
3366				 const struct workqueue_attrs *from)
3367{
3368	to->nice = from->nice;
3369	cpumask_copy(to->cpumask, from->cpumask);
3370	/*
3371	 * Unlike hash and equality test, this function doesn't ignore
3372	 * ->no_numa as it is used for both pool and wq attrs.  Instead,
3373	 * get_unbound_pool() explicitly clears ->no_numa after copying.
3374	 */
3375	to->no_numa = from->no_numa;
3376}
3377
3378/* hash value of the content of @attr */
3379static u32 wqattrs_hash(const struct workqueue_attrs *attrs)
3380{
3381	u32 hash = 0;
3382
3383	hash = jhash_1word(attrs->nice, hash);
3384	hash = jhash(cpumask_bits(attrs->cpumask),
3385		     BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash);
3386	return hash;
3387}
3388
3389/* content equality test */
3390static bool wqattrs_equal(const struct workqueue_attrs *a,
3391			  const struct workqueue_attrs *b)
3392{
3393	if (a->nice != b->nice)
3394		return false;
3395	if (!cpumask_equal(a->cpumask, b->cpumask))
3396		return false;
3397	return true;
3398}
3399
3400/**
3401 * init_worker_pool - initialize a newly zalloc'd worker_pool
3402 * @pool: worker_pool to initialize
3403 *
3404 * Initialize a newly zalloc'd @pool.  It also allocates @pool->attrs.
3405 *
3406 * Return: 0 on success, -errno on failure.  Even on failure, all fields
3407 * inside @pool proper are initialized and put_unbound_pool() can be called
3408 * on @pool safely to release it.
3409 */
3410static int init_worker_pool(struct worker_pool *pool)
3411{
3412	spin_lock_init(&pool->lock);
3413	pool->id = -1;
3414	pool->cpu = -1;
3415	pool->node = NUMA_NO_NODE;
3416	pool->flags |= POOL_DISASSOCIATED;
3417	pool->watchdog_ts = jiffies;
3418	INIT_LIST_HEAD(&pool->worklist);
3419	INIT_LIST_HEAD(&pool->idle_list);
3420	hash_init(pool->busy_hash);
3421
3422	timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE);
 
 
3423
3424	timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0);
 
3425
 
 
3426	INIT_LIST_HEAD(&pool->workers);
3427
3428	ida_init(&pool->worker_ida);
3429	INIT_HLIST_NODE(&pool->hash_node);
3430	pool->refcnt = 1;
3431
3432	/* shouldn't fail above this point */
3433	pool->attrs = alloc_workqueue_attrs();
3434	if (!pool->attrs)
3435		return -ENOMEM;
3436	return 0;
3437}
3438
3439#ifdef CONFIG_LOCKDEP
3440static void wq_init_lockdep(struct workqueue_struct *wq)
3441{
3442	char *lock_name;
3443
3444	lockdep_register_key(&wq->key);
3445	lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name);
3446	if (!lock_name)
3447		lock_name = wq->name;
3448
3449	wq->lock_name = lock_name;
3450	lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0);
3451}
3452
3453static void wq_unregister_lockdep(struct workqueue_struct *wq)
3454{
3455	lockdep_unregister_key(&wq->key);
3456}
3457
3458static void wq_free_lockdep(struct workqueue_struct *wq)
3459{
3460	if (wq->lock_name != wq->name)
3461		kfree(wq->lock_name);
3462}
3463#else
3464static void wq_init_lockdep(struct workqueue_struct *wq)
3465{
3466}
3467
3468static void wq_unregister_lockdep(struct workqueue_struct *wq)
3469{
3470}
3471
3472static void wq_free_lockdep(struct workqueue_struct *wq)
3473{
3474}
3475#endif
3476
3477static void rcu_free_wq(struct rcu_head *rcu)
3478{
3479	struct workqueue_struct *wq =
3480		container_of(rcu, struct workqueue_struct, rcu);
3481
3482	wq_free_lockdep(wq);
3483
3484	if (!(wq->flags & WQ_UNBOUND))
3485		free_percpu(wq->cpu_pwqs);
3486	else
3487		free_workqueue_attrs(wq->unbound_attrs);
3488
3489	kfree(wq->rescuer);
3490	kfree(wq);
3491}
3492
3493static void rcu_free_pool(struct rcu_head *rcu)
3494{
3495	struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu);
3496
3497	ida_destroy(&pool->worker_ida);
3498	free_workqueue_attrs(pool->attrs);
3499	kfree(pool);
3500}
3501
3502/**
3503 * put_unbound_pool - put a worker_pool
3504 * @pool: worker_pool to put
3505 *
3506 * Put @pool.  If its refcnt reaches zero, it gets destroyed in RCU
3507 * safe manner.  get_unbound_pool() calls this function on its failure path
3508 * and this function should be able to release pools which went through,
3509 * successfully or not, init_worker_pool().
3510 *
3511 * Should be called with wq_pool_mutex held.
3512 */
3513static void put_unbound_pool(struct worker_pool *pool)
3514{
3515	DECLARE_COMPLETION_ONSTACK(detach_completion);
3516	struct worker *worker;
3517
3518	lockdep_assert_held(&wq_pool_mutex);
3519
3520	if (--pool->refcnt)
3521		return;
3522
3523	/* sanity checks */
3524	if (WARN_ON(!(pool->cpu < 0)) ||
3525	    WARN_ON(!list_empty(&pool->worklist)))
3526		return;
3527
3528	/* release id and unhash */
3529	if (pool->id >= 0)
3530		idr_remove(&worker_pool_idr, pool->id);
3531	hash_del(&pool->hash_node);
3532
3533	/*
3534	 * Become the manager and destroy all workers.  This prevents
3535	 * @pool's workers from blocking on attach_mutex.  We're the last
3536	 * manager and @pool gets freed with the flag set.
3537	 */
3538	spin_lock_irq(&pool->lock);
3539	wait_event_lock_irq(wq_manager_wait,
3540			    !(pool->flags & POOL_MANAGER_ACTIVE), pool->lock);
3541	pool->flags |= POOL_MANAGER_ACTIVE;
3542
 
3543	while ((worker = first_idle_worker(pool)))
3544		destroy_worker(worker);
3545	WARN_ON(pool->nr_workers || pool->nr_idle);
3546	spin_unlock_irq(&pool->lock);
3547
3548	mutex_lock(&wq_pool_attach_mutex);
3549	if (!list_empty(&pool->workers))
3550		pool->detach_completion = &detach_completion;
3551	mutex_unlock(&wq_pool_attach_mutex);
3552
3553	if (pool->detach_completion)
3554		wait_for_completion(pool->detach_completion);
3555
 
 
3556	/* shut down the timers */
3557	del_timer_sync(&pool->idle_timer);
3558	del_timer_sync(&pool->mayday_timer);
3559
3560	/* RCU protected to allow dereferences from get_work_pool() */
3561	call_rcu(&pool->rcu, rcu_free_pool);
3562}
3563
3564/**
3565 * get_unbound_pool - get a worker_pool with the specified attributes
3566 * @attrs: the attributes of the worker_pool to get
3567 *
3568 * Obtain a worker_pool which has the same attributes as @attrs, bump the
3569 * reference count and return it.  If there already is a matching
3570 * worker_pool, it will be used; otherwise, this function attempts to
3571 * create a new one.
3572 *
3573 * Should be called with wq_pool_mutex held.
3574 *
3575 * Return: On success, a worker_pool with the same attributes as @attrs.
3576 * On failure, %NULL.
3577 */
3578static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs)
3579{
3580	u32 hash = wqattrs_hash(attrs);
3581	struct worker_pool *pool;
3582	int node;
3583	int target_node = NUMA_NO_NODE;
3584
3585	lockdep_assert_held(&wq_pool_mutex);
3586
3587	/* do we already have a matching pool? */
3588	hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) {
3589		if (wqattrs_equal(pool->attrs, attrs)) {
3590			pool->refcnt++;
3591			return pool;
3592		}
3593	}
3594
3595	/* if cpumask is contained inside a NUMA node, we belong to that node */
3596	if (wq_numa_enabled) {
3597		for_each_node(node) {
3598			if (cpumask_subset(attrs->cpumask,
3599					   wq_numa_possible_cpumask[node])) {
3600				target_node = node;
3601				break;
3602			}
3603		}
3604	}
3605
3606	/* nope, create a new one */
3607	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, target_node);
3608	if (!pool || init_worker_pool(pool) < 0)
3609		goto fail;
3610
3611	lockdep_set_subclass(&pool->lock, 1);	/* see put_pwq() */
3612	copy_workqueue_attrs(pool->attrs, attrs);
3613	pool->node = target_node;
3614
3615	/*
3616	 * no_numa isn't a worker_pool attribute, always clear it.  See
3617	 * 'struct workqueue_attrs' comments for detail.
3618	 */
3619	pool->attrs->no_numa = false;
3620
3621	if (worker_pool_assign_id(pool) < 0)
3622		goto fail;
3623
3624	/* create and start the initial worker */
3625	if (wq_online && !create_worker(pool))
3626		goto fail;
3627
3628	/* install */
3629	hash_add(unbound_pool_hash, &pool->hash_node, hash);
3630
3631	return pool;
3632fail:
3633	if (pool)
3634		put_unbound_pool(pool);
3635	return NULL;
3636}
3637
3638static void rcu_free_pwq(struct rcu_head *rcu)
3639{
3640	kmem_cache_free(pwq_cache,
3641			container_of(rcu, struct pool_workqueue, rcu));
3642}
3643
3644/*
3645 * Scheduled on system_wq by put_pwq() when an unbound pwq hits zero refcnt
3646 * and needs to be destroyed.
3647 */
3648static void pwq_unbound_release_workfn(struct work_struct *work)
3649{
3650	struct pool_workqueue *pwq = container_of(work, struct pool_workqueue,
3651						  unbound_release_work);
3652	struct workqueue_struct *wq = pwq->wq;
3653	struct worker_pool *pool = pwq->pool;
3654	bool is_last;
3655
3656	if (WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)))
3657		return;
3658
3659	mutex_lock(&wq->mutex);
3660	list_del_rcu(&pwq->pwqs_node);
3661	is_last = list_empty(&wq->pwqs);
3662	mutex_unlock(&wq->mutex);
3663
3664	mutex_lock(&wq_pool_mutex);
3665	put_unbound_pool(pool);
3666	mutex_unlock(&wq_pool_mutex);
3667
3668	call_rcu(&pwq->rcu, rcu_free_pwq);
3669
3670	/*
3671	 * If we're the last pwq going away, @wq is already dead and no one
3672	 * is gonna access it anymore.  Schedule RCU free.
3673	 */
3674	if (is_last) {
3675		wq_unregister_lockdep(wq);
3676		call_rcu(&wq->rcu, rcu_free_wq);
3677	}
3678}
3679
3680/**
3681 * pwq_adjust_max_active - update a pwq's max_active to the current setting
3682 * @pwq: target pool_workqueue
3683 *
3684 * If @pwq isn't freezing, set @pwq->max_active to the associated
3685 * workqueue's saved_max_active and activate delayed work items
3686 * accordingly.  If @pwq is freezing, clear @pwq->max_active to zero.
3687 */
3688static void pwq_adjust_max_active(struct pool_workqueue *pwq)
3689{
3690	struct workqueue_struct *wq = pwq->wq;
3691	bool freezable = wq->flags & WQ_FREEZABLE;
3692	unsigned long flags;
3693
3694	/* for @wq->saved_max_active */
3695	lockdep_assert_held(&wq->mutex);
3696
3697	/* fast exit for non-freezable wqs */
3698	if (!freezable && pwq->max_active == wq->saved_max_active)
3699		return;
3700
3701	/* this function can be called during early boot w/ irq disabled */
3702	spin_lock_irqsave(&pwq->pool->lock, flags);
3703
3704	/*
3705	 * During [un]freezing, the caller is responsible for ensuring that
3706	 * this function is called at least once after @workqueue_freezing
3707	 * is updated and visible.
3708	 */
3709	if (!freezable || !workqueue_freezing) {
3710		pwq->max_active = wq->saved_max_active;
3711
3712		while (!list_empty(&pwq->delayed_works) &&
3713		       pwq->nr_active < pwq->max_active)
3714			pwq_activate_first_delayed(pwq);
3715
3716		/*
3717		 * Need to kick a worker after thawed or an unbound wq's
3718		 * max_active is bumped.  It's a slow path.  Do it always.
3719		 */
3720		wake_up_worker(pwq->pool);
3721	} else {
3722		pwq->max_active = 0;
3723	}
3724
3725	spin_unlock_irqrestore(&pwq->pool->lock, flags);
3726}
3727
3728/* initialize newly alloced @pwq which is associated with @wq and @pool */
3729static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq,
3730		     struct worker_pool *pool)
3731{
3732	BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK);
3733
3734	memset(pwq, 0, sizeof(*pwq));
3735
3736	pwq->pool = pool;
3737	pwq->wq = wq;
3738	pwq->flush_color = -1;
3739	pwq->refcnt = 1;
3740	INIT_LIST_HEAD(&pwq->delayed_works);
3741	INIT_LIST_HEAD(&pwq->pwqs_node);
3742	INIT_LIST_HEAD(&pwq->mayday_node);
3743	INIT_WORK(&pwq->unbound_release_work, pwq_unbound_release_workfn);
3744}
3745
3746/* sync @pwq with the current state of its associated wq and link it */
3747static void link_pwq(struct pool_workqueue *pwq)
3748{
3749	struct workqueue_struct *wq = pwq->wq;
3750
3751	lockdep_assert_held(&wq->mutex);
3752
3753	/* may be called multiple times, ignore if already linked */
3754	if (!list_empty(&pwq->pwqs_node))
3755		return;
3756
3757	/* set the matching work_color */
3758	pwq->work_color = wq->work_color;
3759
3760	/* sync max_active to the current setting */
3761	pwq_adjust_max_active(pwq);
3762
3763	/* link in @pwq */
3764	list_add_rcu(&pwq->pwqs_node, &wq->pwqs);
3765}
3766
3767/* obtain a pool matching @attr and create a pwq associating the pool and @wq */
3768static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq,
3769					const struct workqueue_attrs *attrs)
3770{
3771	struct worker_pool *pool;
3772	struct pool_workqueue *pwq;
3773
3774	lockdep_assert_held(&wq_pool_mutex);
3775
3776	pool = get_unbound_pool(attrs);
3777	if (!pool)
3778		return NULL;
3779
3780	pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node);
3781	if (!pwq) {
3782		put_unbound_pool(pool);
3783		return NULL;
3784	}
3785
3786	init_pwq(pwq, wq, pool);
3787	return pwq;
3788}
3789
3790/**
3791 * wq_calc_node_cpumask - calculate a wq_attrs' cpumask for the specified node
3792 * @attrs: the wq_attrs of the default pwq of the target workqueue
3793 * @node: the target NUMA node
3794 * @cpu_going_down: if >= 0, the CPU to consider as offline
3795 * @cpumask: outarg, the resulting cpumask
3796 *
3797 * Calculate the cpumask a workqueue with @attrs should use on @node.  If
3798 * @cpu_going_down is >= 0, that cpu is considered offline during
3799 * calculation.  The result is stored in @cpumask.
3800 *
3801 * If NUMA affinity is not enabled, @attrs->cpumask is always used.  If
3802 * enabled and @node has online CPUs requested by @attrs, the returned
3803 * cpumask is the intersection of the possible CPUs of @node and
3804 * @attrs->cpumask.
3805 *
3806 * The caller is responsible for ensuring that the cpumask of @node stays
3807 * stable.
3808 *
3809 * Return: %true if the resulting @cpumask is different from @attrs->cpumask,
3810 * %false if equal.
3811 */
3812static bool wq_calc_node_cpumask(const struct workqueue_attrs *attrs, int node,
3813				 int cpu_going_down, cpumask_t *cpumask)
3814{
3815	if (!wq_numa_enabled || attrs->no_numa)
3816		goto use_dfl;
3817
3818	/* does @node have any online CPUs @attrs wants? */
3819	cpumask_and(cpumask, cpumask_of_node(node), attrs->cpumask);
3820	if (cpu_going_down >= 0)
3821		cpumask_clear_cpu(cpu_going_down, cpumask);
3822
3823	if (cpumask_empty(cpumask))
3824		goto use_dfl;
3825
3826	/* yeap, return possible CPUs in @node that @attrs wants */
3827	cpumask_and(cpumask, attrs->cpumask, wq_numa_possible_cpumask[node]);
3828
3829	if (cpumask_empty(cpumask)) {
3830		pr_warn_once("WARNING: workqueue cpumask: online intersect > "
3831				"possible intersect\n");
3832		return false;
3833	}
3834
3835	return !cpumask_equal(cpumask, attrs->cpumask);
3836
3837use_dfl:
3838	cpumask_copy(cpumask, attrs->cpumask);
3839	return false;
3840}
3841
3842/* install @pwq into @wq's numa_pwq_tbl[] for @node and return the old pwq */
3843static struct pool_workqueue *numa_pwq_tbl_install(struct workqueue_struct *wq,
3844						   int node,
3845						   struct pool_workqueue *pwq)
3846{
3847	struct pool_workqueue *old_pwq;
3848
3849	lockdep_assert_held(&wq_pool_mutex);
3850	lockdep_assert_held(&wq->mutex);
3851
3852	/* link_pwq() can handle duplicate calls */
3853	link_pwq(pwq);
3854
3855	old_pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
3856	rcu_assign_pointer(wq->numa_pwq_tbl[node], pwq);
3857	return old_pwq;
3858}
3859
3860/* context to store the prepared attrs & pwqs before applying */
3861struct apply_wqattrs_ctx {
3862	struct workqueue_struct	*wq;		/* target workqueue */
3863	struct workqueue_attrs	*attrs;		/* attrs to apply */
3864	struct list_head	list;		/* queued for batching commit */
3865	struct pool_workqueue	*dfl_pwq;
3866	struct pool_workqueue	*pwq_tbl[];
3867};
3868
3869/* free the resources after success or abort */
3870static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx)
3871{
3872	if (ctx) {
3873		int node;
3874
3875		for_each_node(node)
3876			put_pwq_unlocked(ctx->pwq_tbl[node]);
3877		put_pwq_unlocked(ctx->dfl_pwq);
3878
3879		free_workqueue_attrs(ctx->attrs);
3880
3881		kfree(ctx);
3882	}
3883}
3884
3885/* allocate the attrs and pwqs for later installation */
3886static struct apply_wqattrs_ctx *
3887apply_wqattrs_prepare(struct workqueue_struct *wq,
3888		      const struct workqueue_attrs *attrs)
3889{
3890	struct apply_wqattrs_ctx *ctx;
3891	struct workqueue_attrs *new_attrs, *tmp_attrs;
3892	int node;
3893
3894	lockdep_assert_held(&wq_pool_mutex);
3895
3896	ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_node_ids), GFP_KERNEL);
 
3897
3898	new_attrs = alloc_workqueue_attrs();
3899	tmp_attrs = alloc_workqueue_attrs();
3900	if (!ctx || !new_attrs || !tmp_attrs)
3901		goto out_free;
3902
3903	/*
3904	 * Calculate the attrs of the default pwq.
3905	 * If the user configured cpumask doesn't overlap with the
3906	 * wq_unbound_cpumask, we fallback to the wq_unbound_cpumask.
3907	 */
3908	copy_workqueue_attrs(new_attrs, attrs);
3909	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, wq_unbound_cpumask);
3910	if (unlikely(cpumask_empty(new_attrs->cpumask)))
3911		cpumask_copy(new_attrs->cpumask, wq_unbound_cpumask);
3912
3913	/*
3914	 * We may create multiple pwqs with differing cpumasks.  Make a
3915	 * copy of @new_attrs which will be modified and used to obtain
3916	 * pools.
3917	 */
3918	copy_workqueue_attrs(tmp_attrs, new_attrs);
3919
3920	/*
3921	 * If something goes wrong during CPU up/down, we'll fall back to
3922	 * the default pwq covering whole @attrs->cpumask.  Always create
3923	 * it even if we don't use it immediately.
3924	 */
3925	ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs);
3926	if (!ctx->dfl_pwq)
3927		goto out_free;
3928
3929	for_each_node(node) {
3930		if (wq_calc_node_cpumask(new_attrs, node, -1, tmp_attrs->cpumask)) {
3931			ctx->pwq_tbl[node] = alloc_unbound_pwq(wq, tmp_attrs);
3932			if (!ctx->pwq_tbl[node])
3933				goto out_free;
3934		} else {
3935			ctx->dfl_pwq->refcnt++;
3936			ctx->pwq_tbl[node] = ctx->dfl_pwq;
3937		}
3938	}
3939
3940	/* save the user configured attrs and sanitize it. */
3941	copy_workqueue_attrs(new_attrs, attrs);
3942	cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask);
3943	ctx->attrs = new_attrs;
3944
3945	ctx->wq = wq;
3946	free_workqueue_attrs(tmp_attrs);
3947	return ctx;
3948
3949out_free:
3950	free_workqueue_attrs(tmp_attrs);
3951	free_workqueue_attrs(new_attrs);
3952	apply_wqattrs_cleanup(ctx);
3953	return NULL;
3954}
3955
3956/* set attrs and install prepared pwqs, @ctx points to old pwqs on return */
3957static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx)
3958{
3959	int node;
3960
3961	/* all pwqs have been created successfully, let's install'em */
3962	mutex_lock(&ctx->wq->mutex);
3963
3964	copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs);
3965
3966	/* save the previous pwq and install the new one */
3967	for_each_node(node)
3968		ctx->pwq_tbl[node] = numa_pwq_tbl_install(ctx->wq, node,
3969							  ctx->pwq_tbl[node]);
3970
3971	/* @dfl_pwq might not have been used, ensure it's linked */
3972	link_pwq(ctx->dfl_pwq);
3973	swap(ctx->wq->dfl_pwq, ctx->dfl_pwq);
3974
3975	mutex_unlock(&ctx->wq->mutex);
3976}
3977
3978static void apply_wqattrs_lock(void)
3979{
3980	/* CPUs should stay stable across pwq creations and installations */
3981	get_online_cpus();
3982	mutex_lock(&wq_pool_mutex);
3983}
3984
3985static void apply_wqattrs_unlock(void)
3986{
3987	mutex_unlock(&wq_pool_mutex);
3988	put_online_cpus();
3989}
3990
3991static int apply_workqueue_attrs_locked(struct workqueue_struct *wq,
3992					const struct workqueue_attrs *attrs)
3993{
3994	struct apply_wqattrs_ctx *ctx;
3995
3996	/* only unbound workqueues can change attributes */
3997	if (WARN_ON(!(wq->flags & WQ_UNBOUND)))
3998		return -EINVAL;
3999
4000	/* creating multiple pwqs breaks ordering guarantee */
4001	if (!list_empty(&wq->pwqs)) {
4002		if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4003			return -EINVAL;
4004
4005		wq->flags &= ~__WQ_ORDERED;
4006	}
4007
4008	ctx = apply_wqattrs_prepare(wq, attrs);
4009	if (!ctx)
4010		return -ENOMEM;
4011
4012	/* the ctx has been prepared successfully, let's commit it */
4013	apply_wqattrs_commit(ctx);
4014	apply_wqattrs_cleanup(ctx);
4015
4016	return 0;
4017}
4018
4019/**
4020 * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue
4021 * @wq: the target workqueue
4022 * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs()
4023 *
4024 * Apply @attrs to an unbound workqueue @wq.  Unless disabled, on NUMA
4025 * machines, this function maps a separate pwq to each NUMA node with
4026 * possibles CPUs in @attrs->cpumask so that work items are affine to the
4027 * NUMA node it was issued on.  Older pwqs are released as in-flight work
4028 * items finish.  Note that a work item which repeatedly requeues itself
4029 * back-to-back will stay on its current pwq.
4030 *
4031 * Performs GFP_KERNEL allocations.
4032 *
4033 * Assumes caller has CPU hotplug read exclusion, i.e. get_online_cpus().
4034 *
4035 * Return: 0 on success and -errno on failure.
4036 */
4037int apply_workqueue_attrs(struct workqueue_struct *wq,
4038			  const struct workqueue_attrs *attrs)
4039{
4040	int ret;
4041
4042	lockdep_assert_cpus_held();
4043
4044	mutex_lock(&wq_pool_mutex);
4045	ret = apply_workqueue_attrs_locked(wq, attrs);
4046	mutex_unlock(&wq_pool_mutex);
4047
4048	return ret;
4049}
4050
4051/**
4052 * wq_update_unbound_numa - update NUMA affinity of a wq for CPU hot[un]plug
4053 * @wq: the target workqueue
4054 * @cpu: the CPU coming up or going down
4055 * @online: whether @cpu is coming up or going down
4056 *
4057 * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and
4058 * %CPU_DOWN_FAILED.  @cpu is being hot[un]plugged, update NUMA affinity of
4059 * @wq accordingly.
4060 *
4061 * If NUMA affinity can't be adjusted due to memory allocation failure, it
4062 * falls back to @wq->dfl_pwq which may not be optimal but is always
4063 * correct.
4064 *
4065 * Note that when the last allowed CPU of a NUMA node goes offline for a
4066 * workqueue with a cpumask spanning multiple nodes, the workers which were
4067 * already executing the work items for the workqueue will lose their CPU
4068 * affinity and may execute on any CPU.  This is similar to how per-cpu
4069 * workqueues behave on CPU_DOWN.  If a workqueue user wants strict
4070 * affinity, it's the user's responsibility to flush the work item from
4071 * CPU_DOWN_PREPARE.
4072 */
4073static void wq_update_unbound_numa(struct workqueue_struct *wq, int cpu,
4074				   bool online)
4075{
4076	int node = cpu_to_node(cpu);
4077	int cpu_off = online ? -1 : cpu;
4078	struct pool_workqueue *old_pwq = NULL, *pwq;
4079	struct workqueue_attrs *target_attrs;
4080	cpumask_t *cpumask;
4081
4082	lockdep_assert_held(&wq_pool_mutex);
4083
4084	if (!wq_numa_enabled || !(wq->flags & WQ_UNBOUND) ||
4085	    wq->unbound_attrs->no_numa)
4086		return;
4087
4088	/*
4089	 * We don't wanna alloc/free wq_attrs for each wq for each CPU.
4090	 * Let's use a preallocated one.  The following buf is protected by
4091	 * CPU hotplug exclusion.
4092	 */
4093	target_attrs = wq_update_unbound_numa_attrs_buf;
4094	cpumask = target_attrs->cpumask;
4095
4096	copy_workqueue_attrs(target_attrs, wq->unbound_attrs);
4097	pwq = unbound_pwq_by_node(wq, node);
4098
4099	/*
4100	 * Let's determine what needs to be done.  If the target cpumask is
4101	 * different from the default pwq's, we need to compare it to @pwq's
4102	 * and create a new one if they don't match.  If the target cpumask
4103	 * equals the default pwq's, the default pwq should be used.
4104	 */
4105	if (wq_calc_node_cpumask(wq->dfl_pwq->pool->attrs, node, cpu_off, cpumask)) {
4106		if (cpumask_equal(cpumask, pwq->pool->attrs->cpumask))
4107			return;
4108	} else {
4109		goto use_dfl_pwq;
4110	}
4111
4112	/* create a new pwq */
4113	pwq = alloc_unbound_pwq(wq, target_attrs);
4114	if (!pwq) {
4115		pr_warn("workqueue: allocation failed while updating NUMA affinity of \"%s\"\n",
4116			wq->name);
4117		goto use_dfl_pwq;
4118	}
4119
4120	/* Install the new pwq. */
4121	mutex_lock(&wq->mutex);
4122	old_pwq = numa_pwq_tbl_install(wq, node, pwq);
4123	goto out_unlock;
4124
4125use_dfl_pwq:
4126	mutex_lock(&wq->mutex);
4127	spin_lock_irq(&wq->dfl_pwq->pool->lock);
4128	get_pwq(wq->dfl_pwq);
4129	spin_unlock_irq(&wq->dfl_pwq->pool->lock);
4130	old_pwq = numa_pwq_tbl_install(wq, node, wq->dfl_pwq);
4131out_unlock:
4132	mutex_unlock(&wq->mutex);
4133	put_pwq_unlocked(old_pwq);
4134}
4135
4136static int alloc_and_link_pwqs(struct workqueue_struct *wq)
4137{
4138	bool highpri = wq->flags & WQ_HIGHPRI;
4139	int cpu, ret;
4140
4141	if (!(wq->flags & WQ_UNBOUND)) {
4142		wq->cpu_pwqs = alloc_percpu(struct pool_workqueue);
4143		if (!wq->cpu_pwqs)
4144			return -ENOMEM;
4145
4146		for_each_possible_cpu(cpu) {
4147			struct pool_workqueue *pwq =
4148				per_cpu_ptr(wq->cpu_pwqs, cpu);
4149			struct worker_pool *cpu_pools =
4150				per_cpu(cpu_worker_pools, cpu);
4151
4152			init_pwq(pwq, wq, &cpu_pools[highpri]);
4153
4154			mutex_lock(&wq->mutex);
4155			link_pwq(pwq);
4156			mutex_unlock(&wq->mutex);
4157		}
4158		return 0;
4159	}
4160
4161	get_online_cpus();
4162	if (wq->flags & __WQ_ORDERED) {
4163		ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]);
4164		/* there should only be single pwq for ordering guarantee */
4165		WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node ||
4166			      wq->pwqs.prev != &wq->dfl_pwq->pwqs_node),
4167		     "ordering guarantee broken for workqueue %s\n", wq->name);
 
4168	} else {
4169		ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]);
4170	}
4171	put_online_cpus();
4172
4173	return ret;
4174}
4175
4176static int wq_clamp_max_active(int max_active, unsigned int flags,
4177			       const char *name)
4178{
4179	int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
4180
4181	if (max_active < 1 || max_active > lim)
4182		pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n",
4183			max_active, name, 1, lim);
4184
4185	return clamp_val(max_active, 1, lim);
4186}
4187
4188/*
4189 * Workqueues which may be used during memory reclaim should have a rescuer
4190 * to guarantee forward progress.
4191 */
4192static int init_rescuer(struct workqueue_struct *wq)
4193{
4194	struct worker *rescuer;
4195	int ret;
4196
4197	if (!(wq->flags & WQ_MEM_RECLAIM))
4198		return 0;
4199
4200	rescuer = alloc_worker(NUMA_NO_NODE);
4201	if (!rescuer)
4202		return -ENOMEM;
4203
4204	rescuer->rescue_wq = wq;
4205	rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name);
4206	ret = PTR_ERR_OR_ZERO(rescuer->task);
4207	if (ret) {
4208		kfree(rescuer);
4209		return ret;
4210	}
4211
4212	wq->rescuer = rescuer;
4213	kthread_bind_mask(rescuer->task, cpu_possible_mask);
4214	wake_up_process(rescuer->task);
4215
4216	return 0;
4217}
4218
4219__printf(1, 4)
4220struct workqueue_struct *alloc_workqueue(const char *fmt,
4221					 unsigned int flags,
4222					 int max_active, ...)
4223{
4224	size_t tbl_size = 0;
4225	va_list args;
4226	struct workqueue_struct *wq;
4227	struct pool_workqueue *pwq;
4228
4229	/*
4230	 * Unbound && max_active == 1 used to imply ordered, which is no
4231	 * longer the case on NUMA machines due to per-node pools.  While
4232	 * alloc_ordered_workqueue() is the right way to create an ordered
4233	 * workqueue, keep the previous behavior to avoid subtle breakages
4234	 * on NUMA.
4235	 */
4236	if ((flags & WQ_UNBOUND) && max_active == 1)
4237		flags |= __WQ_ORDERED;
4238
4239	/* see the comment above the definition of WQ_POWER_EFFICIENT */
4240	if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient)
4241		flags |= WQ_UNBOUND;
4242
4243	/* allocate wq and format name */
4244	if (flags & WQ_UNBOUND)
4245		tbl_size = nr_node_ids * sizeof(wq->numa_pwq_tbl[0]);
4246
4247	wq = kzalloc(sizeof(*wq) + tbl_size, GFP_KERNEL);
4248	if (!wq)
4249		return NULL;
4250
4251	if (flags & WQ_UNBOUND) {
4252		wq->unbound_attrs = alloc_workqueue_attrs();
4253		if (!wq->unbound_attrs)
4254			goto err_free_wq;
4255	}
4256
4257	va_start(args, max_active);
4258	vsnprintf(wq->name, sizeof(wq->name), fmt, args);
4259	va_end(args);
4260
4261	max_active = max_active ?: WQ_DFL_ACTIVE;
4262	max_active = wq_clamp_max_active(max_active, flags, wq->name);
4263
4264	/* init wq */
4265	wq->flags = flags;
4266	wq->saved_max_active = max_active;
4267	mutex_init(&wq->mutex);
4268	atomic_set(&wq->nr_pwqs_to_flush, 0);
4269	INIT_LIST_HEAD(&wq->pwqs);
4270	INIT_LIST_HEAD(&wq->flusher_queue);
4271	INIT_LIST_HEAD(&wq->flusher_overflow);
4272	INIT_LIST_HEAD(&wq->maydays);
4273
4274	wq_init_lockdep(wq);
4275	INIT_LIST_HEAD(&wq->list);
4276
4277	if (alloc_and_link_pwqs(wq) < 0)
4278		goto err_unreg_lockdep;
4279
4280	if (wq_online && init_rescuer(wq) < 0)
4281		goto err_destroy;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4282
4283	if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq))
4284		goto err_destroy;
4285
4286	/*
4287	 * wq_pool_mutex protects global freeze state and workqueues list.
4288	 * Grab it, adjust max_active and add the new @wq to workqueues
4289	 * list.
4290	 */
4291	mutex_lock(&wq_pool_mutex);
4292
4293	mutex_lock(&wq->mutex);
4294	for_each_pwq(pwq, wq)
4295		pwq_adjust_max_active(pwq);
4296	mutex_unlock(&wq->mutex);
4297
4298	list_add_tail_rcu(&wq->list, &workqueues);
4299
4300	mutex_unlock(&wq_pool_mutex);
4301
4302	return wq;
4303
4304err_unreg_lockdep:
4305	wq_unregister_lockdep(wq);
4306	wq_free_lockdep(wq);
4307err_free_wq:
4308	free_workqueue_attrs(wq->unbound_attrs);
4309	kfree(wq);
4310	return NULL;
4311err_destroy:
4312	destroy_workqueue(wq);
4313	return NULL;
4314}
4315EXPORT_SYMBOL_GPL(alloc_workqueue);
4316
4317/**
4318 * destroy_workqueue - safely terminate a workqueue
4319 * @wq: target workqueue
4320 *
4321 * Safely destroy a workqueue. All work currently pending will be done first.
4322 */
4323void destroy_workqueue(struct workqueue_struct *wq)
4324{
4325	struct pool_workqueue *pwq;
4326	int node;
4327
4328	/* drain it before proceeding with destruction */
4329	drain_workqueue(wq);
4330
4331	/* sanity checks */
4332	mutex_lock(&wq->mutex);
4333	for_each_pwq(pwq, wq) {
4334		int i;
4335
4336		for (i = 0; i < WORK_NR_COLORS; i++) {
4337			if (WARN_ON(pwq->nr_in_flight[i])) {
4338				mutex_unlock(&wq->mutex);
4339				show_workqueue_state();
4340				return;
4341			}
4342		}
4343
4344		if (WARN_ON((pwq != wq->dfl_pwq) && (pwq->refcnt > 1)) ||
4345		    WARN_ON(pwq->nr_active) ||
4346		    WARN_ON(!list_empty(&pwq->delayed_works))) {
4347			mutex_unlock(&wq->mutex);
4348			show_workqueue_state();
4349			return;
4350		}
4351	}
4352	mutex_unlock(&wq->mutex);
4353
4354	/*
4355	 * wq list is used to freeze wq, remove from list after
4356	 * flushing is complete in case freeze races us.
4357	 */
4358	mutex_lock(&wq_pool_mutex);
4359	list_del_rcu(&wq->list);
4360	mutex_unlock(&wq_pool_mutex);
4361
4362	workqueue_sysfs_unregister(wq);
4363
4364	if (wq->rescuer)
4365		kthread_stop(wq->rescuer->task);
4366
4367	if (!(wq->flags & WQ_UNBOUND)) {
4368		wq_unregister_lockdep(wq);
4369		/*
4370		 * The base ref is never dropped on per-cpu pwqs.  Directly
4371		 * schedule RCU free.
4372		 */
4373		call_rcu(&wq->rcu, rcu_free_wq);
4374	} else {
4375		/*
4376		 * We're the sole accessor of @wq at this point.  Directly
4377		 * access numa_pwq_tbl[] and dfl_pwq to put the base refs.
4378		 * @wq will be freed when the last pwq is released.
4379		 */
4380		for_each_node(node) {
4381			pwq = rcu_access_pointer(wq->numa_pwq_tbl[node]);
4382			RCU_INIT_POINTER(wq->numa_pwq_tbl[node], NULL);
4383			put_pwq_unlocked(pwq);
4384		}
4385
4386		/*
4387		 * Put dfl_pwq.  @wq may be freed any time after dfl_pwq is
4388		 * put.  Don't access it afterwards.
4389		 */
4390		pwq = wq->dfl_pwq;
4391		wq->dfl_pwq = NULL;
4392		put_pwq_unlocked(pwq);
4393	}
4394}
4395EXPORT_SYMBOL_GPL(destroy_workqueue);
4396
4397/**
4398 * workqueue_set_max_active - adjust max_active of a workqueue
4399 * @wq: target workqueue
4400 * @max_active: new max_active value.
4401 *
4402 * Set max_active of @wq to @max_active.
4403 *
4404 * CONTEXT:
4405 * Don't call from IRQ context.
4406 */
4407void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
4408{
4409	struct pool_workqueue *pwq;
4410
4411	/* disallow meddling with max_active for ordered workqueues */
4412	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
4413		return;
4414
4415	max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
4416
4417	mutex_lock(&wq->mutex);
4418
4419	wq->flags &= ~__WQ_ORDERED;
4420	wq->saved_max_active = max_active;
4421
4422	for_each_pwq(pwq, wq)
4423		pwq_adjust_max_active(pwq);
4424
4425	mutex_unlock(&wq->mutex);
4426}
4427EXPORT_SYMBOL_GPL(workqueue_set_max_active);
4428
4429/**
4430 * current_work - retrieve %current task's work struct
4431 *
4432 * Determine if %current task is a workqueue worker and what it's working on.
4433 * Useful to find out the context that the %current task is running in.
4434 *
4435 * Return: work struct if %current task is a workqueue worker, %NULL otherwise.
4436 */
4437struct work_struct *current_work(void)
4438{
4439	struct worker *worker = current_wq_worker();
4440
4441	return worker ? worker->current_work : NULL;
4442}
4443EXPORT_SYMBOL(current_work);
4444
4445/**
4446 * current_is_workqueue_rescuer - is %current workqueue rescuer?
4447 *
4448 * Determine whether %current is a workqueue rescuer.  Can be used from
4449 * work functions to determine whether it's being run off the rescuer task.
4450 *
4451 * Return: %true if %current is a workqueue rescuer. %false otherwise.
4452 */
4453bool current_is_workqueue_rescuer(void)
4454{
4455	struct worker *worker = current_wq_worker();
4456
4457	return worker && worker->rescue_wq;
4458}
4459
4460/**
4461 * workqueue_congested - test whether a workqueue is congested
4462 * @cpu: CPU in question
4463 * @wq: target workqueue
4464 *
4465 * Test whether @wq's cpu workqueue for @cpu is congested.  There is
4466 * no synchronization around this function and the test result is
4467 * unreliable and only useful as advisory hints or for debugging.
4468 *
4469 * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU.
4470 * Note that both per-cpu and unbound workqueues may be associated with
4471 * multiple pool_workqueues which have separate congested states.  A
4472 * workqueue being congested on one CPU doesn't mean the workqueue is also
4473 * contested on other CPUs / NUMA nodes.
4474 *
4475 * Return:
4476 * %true if congested, %false otherwise.
4477 */
4478bool workqueue_congested(int cpu, struct workqueue_struct *wq)
4479{
4480	struct pool_workqueue *pwq;
4481	bool ret;
4482
4483	rcu_read_lock();
4484	preempt_disable();
4485
4486	if (cpu == WORK_CPU_UNBOUND)
4487		cpu = smp_processor_id();
4488
4489	if (!(wq->flags & WQ_UNBOUND))
4490		pwq = per_cpu_ptr(wq->cpu_pwqs, cpu);
4491	else
4492		pwq = unbound_pwq_by_node(wq, cpu_to_node(cpu));
4493
4494	ret = !list_empty(&pwq->delayed_works);
4495	preempt_enable();
4496	rcu_read_unlock();
4497
4498	return ret;
4499}
4500EXPORT_SYMBOL_GPL(workqueue_congested);
4501
4502/**
4503 * work_busy - test whether a work is currently pending or running
4504 * @work: the work to be tested
4505 *
4506 * Test whether @work is currently pending or running.  There is no
4507 * synchronization around this function and the test result is
4508 * unreliable and only useful as advisory hints or for debugging.
4509 *
4510 * Return:
4511 * OR'd bitmask of WORK_BUSY_* bits.
4512 */
4513unsigned int work_busy(struct work_struct *work)
4514{
4515	struct worker_pool *pool;
4516	unsigned long flags;
4517	unsigned int ret = 0;
4518
4519	if (work_pending(work))
4520		ret |= WORK_BUSY_PENDING;
4521
4522	rcu_read_lock();
4523	pool = get_work_pool(work);
4524	if (pool) {
4525		spin_lock_irqsave(&pool->lock, flags);
4526		if (find_worker_executing_work(pool, work))
4527			ret |= WORK_BUSY_RUNNING;
4528		spin_unlock_irqrestore(&pool->lock, flags);
4529	}
4530	rcu_read_unlock();
4531
4532	return ret;
4533}
4534EXPORT_SYMBOL_GPL(work_busy);
4535
4536/**
4537 * set_worker_desc - set description for the current work item
4538 * @fmt: printf-style format string
4539 * @...: arguments for the format string
4540 *
4541 * This function can be called by a running work function to describe what
4542 * the work item is about.  If the worker task gets dumped, this
4543 * information will be printed out together to help debugging.  The
4544 * description can be at most WORKER_DESC_LEN including the trailing '\0'.
4545 */
4546void set_worker_desc(const char *fmt, ...)
4547{
4548	struct worker *worker = current_wq_worker();
4549	va_list args;
4550
4551	if (worker) {
4552		va_start(args, fmt);
4553		vsnprintf(worker->desc, sizeof(worker->desc), fmt, args);
4554		va_end(args);
 
4555	}
4556}
4557EXPORT_SYMBOL_GPL(set_worker_desc);
4558
4559/**
4560 * print_worker_info - print out worker information and description
4561 * @log_lvl: the log level to use when printing
4562 * @task: target task
4563 *
4564 * If @task is a worker and currently executing a work item, print out the
4565 * name of the workqueue being serviced and worker description set with
4566 * set_worker_desc() by the currently executing work item.
4567 *
4568 * This function can be safely called on any task as long as the
4569 * task_struct itself is accessible.  While safe, this function isn't
4570 * synchronized and may print out mixups or garbages of limited length.
4571 */
4572void print_worker_info(const char *log_lvl, struct task_struct *task)
4573{
4574	work_func_t *fn = NULL;
4575	char name[WQ_NAME_LEN] = { };
4576	char desc[WORKER_DESC_LEN] = { };
4577	struct pool_workqueue *pwq = NULL;
4578	struct workqueue_struct *wq = NULL;
 
4579	struct worker *worker;
4580
4581	if (!(task->flags & PF_WQ_WORKER))
4582		return;
4583
4584	/*
4585	 * This function is called without any synchronization and @task
4586	 * could be in any state.  Be careful with dereferences.
4587	 */
4588	worker = kthread_probe_data(task);
4589
4590	/*
4591	 * Carefully copy the associated workqueue's workfn, name and desc.
4592	 * Keep the original last '\0' in case the original is garbage.
4593	 */
4594	probe_kernel_read(&fn, &worker->current_func, sizeof(fn));
4595	probe_kernel_read(&pwq, &worker->current_pwq, sizeof(pwq));
4596	probe_kernel_read(&wq, &pwq->wq, sizeof(wq));
4597	probe_kernel_read(name, wq->name, sizeof(name) - 1);
4598	probe_kernel_read(desc, worker->desc, sizeof(desc) - 1);
 
 
 
 
4599
4600	if (fn || name[0] || desc[0]) {
4601		printk("%sWorkqueue: %s %ps", log_lvl, name, fn);
4602		if (strcmp(name, desc))
4603			pr_cont(" (%s)", desc);
4604		pr_cont("\n");
4605	}
4606}
4607
4608static void pr_cont_pool_info(struct worker_pool *pool)
4609{
4610	pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask);
4611	if (pool->node != NUMA_NO_NODE)
4612		pr_cont(" node=%d", pool->node);
4613	pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice);
4614}
4615
4616static void pr_cont_work(bool comma, struct work_struct *work)
4617{
4618	if (work->func == wq_barrier_func) {
4619		struct wq_barrier *barr;
4620
4621		barr = container_of(work, struct wq_barrier, work);
4622
4623		pr_cont("%s BAR(%d)", comma ? "," : "",
4624			task_pid_nr(barr->task));
4625	} else {
4626		pr_cont("%s %ps", comma ? "," : "", work->func);
4627	}
4628}
4629
4630static void show_pwq(struct pool_workqueue *pwq)
4631{
4632	struct worker_pool *pool = pwq->pool;
4633	struct work_struct *work;
4634	struct worker *worker;
4635	bool has_in_flight = false, has_pending = false;
4636	int bkt;
4637
4638	pr_info("  pwq %d:", pool->id);
4639	pr_cont_pool_info(pool);
4640
4641	pr_cont(" active=%d/%d%s\n", pwq->nr_active, pwq->max_active,
4642		!list_empty(&pwq->mayday_node) ? " MAYDAY" : "");
4643
4644	hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4645		if (worker->current_pwq == pwq) {
4646			has_in_flight = true;
4647			break;
4648		}
4649	}
4650	if (has_in_flight) {
4651		bool comma = false;
4652
4653		pr_info("    in-flight:");
4654		hash_for_each(pool->busy_hash, bkt, worker, hentry) {
4655			if (worker->current_pwq != pwq)
4656				continue;
4657
4658			pr_cont("%s %d%s:%ps", comma ? "," : "",
4659				task_pid_nr(worker->task),
4660				worker == pwq->wq->rescuer ? "(RESCUER)" : "",
4661				worker->current_func);
4662			list_for_each_entry(work, &worker->scheduled, entry)
4663				pr_cont_work(false, work);
4664			comma = true;
4665		}
4666		pr_cont("\n");
4667	}
4668
4669	list_for_each_entry(work, &pool->worklist, entry) {
4670		if (get_work_pwq(work) == pwq) {
4671			has_pending = true;
4672			break;
4673		}
4674	}
4675	if (has_pending) {
4676		bool comma = false;
4677
4678		pr_info("    pending:");
4679		list_for_each_entry(work, &pool->worklist, entry) {
4680			if (get_work_pwq(work) != pwq)
4681				continue;
4682
4683			pr_cont_work(comma, work);
4684			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4685		}
4686		pr_cont("\n");
4687	}
4688
4689	if (!list_empty(&pwq->delayed_works)) {
4690		bool comma = false;
4691
4692		pr_info("    delayed:");
4693		list_for_each_entry(work, &pwq->delayed_works, entry) {
4694			pr_cont_work(comma, work);
4695			comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED);
4696		}
4697		pr_cont("\n");
4698	}
4699}
4700
4701/**
4702 * show_workqueue_state - dump workqueue state
4703 *
4704 * Called from a sysrq handler or try_to_freeze_tasks() and prints out
4705 * all busy workqueues and pools.
4706 */
4707void show_workqueue_state(void)
4708{
4709	struct workqueue_struct *wq;
4710	struct worker_pool *pool;
4711	unsigned long flags;
4712	int pi;
4713
4714	rcu_read_lock();
4715
4716	pr_info("Showing busy workqueues and worker pools:\n");
4717
4718	list_for_each_entry_rcu(wq, &workqueues, list) {
4719		struct pool_workqueue *pwq;
4720		bool idle = true;
4721
4722		for_each_pwq(pwq, wq) {
4723			if (pwq->nr_active || !list_empty(&pwq->delayed_works)) {
4724				idle = false;
4725				break;
4726			}
4727		}
4728		if (idle)
4729			continue;
4730
4731		pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags);
4732
4733		for_each_pwq(pwq, wq) {
4734			spin_lock_irqsave(&pwq->pool->lock, flags);
4735			if (pwq->nr_active || !list_empty(&pwq->delayed_works))
4736				show_pwq(pwq);
4737			spin_unlock_irqrestore(&pwq->pool->lock, flags);
4738			/*
4739			 * We could be printing a lot from atomic context, e.g.
4740			 * sysrq-t -> show_workqueue_state(). Avoid triggering
4741			 * hard lockup.
4742			 */
4743			touch_nmi_watchdog();
4744		}
4745	}
4746
4747	for_each_pool(pool, pi) {
4748		struct worker *worker;
4749		bool first = true;
4750
4751		spin_lock_irqsave(&pool->lock, flags);
4752		if (pool->nr_workers == pool->nr_idle)
4753			goto next_pool;
4754
4755		pr_info("pool %d:", pool->id);
4756		pr_cont_pool_info(pool);
4757		pr_cont(" hung=%us workers=%d",
4758			jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000,
4759			pool->nr_workers);
4760		if (pool->manager)
4761			pr_cont(" manager: %d",
4762				task_pid_nr(pool->manager->task));
4763		list_for_each_entry(worker, &pool->idle_list, entry) {
4764			pr_cont(" %s%d", first ? "idle: " : "",
4765				task_pid_nr(worker->task));
4766			first = false;
4767		}
4768		pr_cont("\n");
4769	next_pool:
4770		spin_unlock_irqrestore(&pool->lock, flags);
4771		/*
4772		 * We could be printing a lot from atomic context, e.g.
4773		 * sysrq-t -> show_workqueue_state(). Avoid triggering
4774		 * hard lockup.
4775		 */
4776		touch_nmi_watchdog();
4777	}
4778
4779	rcu_read_unlock();
4780}
4781
4782/* used to show worker information through /proc/PID/{comm,stat,status} */
4783void wq_worker_comm(char *buf, size_t size, struct task_struct *task)
4784{
4785	int off;
4786
4787	/* always show the actual comm */
4788	off = strscpy(buf, task->comm, size);
4789	if (off < 0)
4790		return;
4791
4792	/* stabilize PF_WQ_WORKER and worker pool association */
4793	mutex_lock(&wq_pool_attach_mutex);
4794
4795	if (task->flags & PF_WQ_WORKER) {
4796		struct worker *worker = kthread_data(task);
4797		struct worker_pool *pool = worker->pool;
4798
4799		if (pool) {
4800			spin_lock_irq(&pool->lock);
4801			/*
4802			 * ->desc tracks information (wq name or
4803			 * set_worker_desc()) for the latest execution.  If
4804			 * current, prepend '+', otherwise '-'.
4805			 */
4806			if (worker->desc[0] != '\0') {
4807				if (worker->current_work)
4808					scnprintf(buf + off, size - off, "+%s",
4809						  worker->desc);
4810				else
4811					scnprintf(buf + off, size - off, "-%s",
4812						  worker->desc);
4813			}
4814			spin_unlock_irq(&pool->lock);
4815		}
4816	}
4817
4818	mutex_unlock(&wq_pool_attach_mutex);
4819}
4820
4821#ifdef CONFIG_SMP
4822
4823/*
4824 * CPU hotplug.
4825 *
4826 * There are two challenges in supporting CPU hotplug.  Firstly, there
4827 * are a lot of assumptions on strong associations among work, pwq and
4828 * pool which make migrating pending and scheduled works very
4829 * difficult to implement without impacting hot paths.  Secondly,
4830 * worker pools serve mix of short, long and very long running works making
4831 * blocked draining impractical.
4832 *
4833 * This is solved by allowing the pools to be disassociated from the CPU
4834 * running as an unbound one and allowing it to be reattached later if the
4835 * cpu comes back online.
4836 */
4837
4838static void unbind_workers(int cpu)
4839{
 
4840	struct worker_pool *pool;
4841	struct worker *worker;
4842
4843	for_each_cpu_worker_pool(pool, cpu) {
4844		mutex_lock(&wq_pool_attach_mutex);
4845		spin_lock_irq(&pool->lock);
4846
4847		/*
4848		 * We've blocked all attach/detach operations. Make all workers
4849		 * unbound and set DISASSOCIATED.  Before this, all workers
4850		 * except for the ones which are still executing works from
4851		 * before the last CPU down must be on the cpu.  After
4852		 * this, they may become diasporas.
4853		 */
4854		for_each_pool_worker(worker, pool)
4855			worker->flags |= WORKER_UNBOUND;
4856
4857		pool->flags |= POOL_DISASSOCIATED;
4858
4859		spin_unlock_irq(&pool->lock);
4860		mutex_unlock(&wq_pool_attach_mutex);
4861
4862		/*
4863		 * Call schedule() so that we cross rq->lock and thus can
4864		 * guarantee sched callbacks see the %WORKER_UNBOUND flag.
4865		 * This is necessary as scheduler callbacks may be invoked
4866		 * from other cpus.
4867		 */
4868		schedule();
4869
4870		/*
4871		 * Sched callbacks are disabled now.  Zap nr_running.
4872		 * After this, nr_running stays zero and need_more_worker()
4873		 * and keep_working() are always true as long as the
4874		 * worklist is not empty.  This pool now behaves as an
4875		 * unbound (in terms of concurrency management) pool which
4876		 * are served by workers tied to the pool.
4877		 */
4878		atomic_set(&pool->nr_running, 0);
4879
4880		/*
4881		 * With concurrency management just turned off, a busy
4882		 * worker blocking could lead to lengthy stalls.  Kick off
4883		 * unbound chain execution of currently pending work items.
4884		 */
4885		spin_lock_irq(&pool->lock);
4886		wake_up_worker(pool);
4887		spin_unlock_irq(&pool->lock);
4888	}
4889}
4890
4891/**
4892 * rebind_workers - rebind all workers of a pool to the associated CPU
4893 * @pool: pool of interest
4894 *
4895 * @pool->cpu is coming online.  Rebind all workers to the CPU.
4896 */
4897static void rebind_workers(struct worker_pool *pool)
4898{
4899	struct worker *worker;
4900
4901	lockdep_assert_held(&wq_pool_attach_mutex);
4902
4903	/*
4904	 * Restore CPU affinity of all workers.  As all idle workers should
4905	 * be on the run-queue of the associated CPU before any local
4906	 * wake-ups for concurrency management happen, restore CPU affinity
4907	 * of all workers first and then clear UNBOUND.  As we're called
4908	 * from CPU_ONLINE, the following shouldn't fail.
4909	 */
4910	for_each_pool_worker(worker, pool)
4911		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task,
4912						  pool->attrs->cpumask) < 0);
4913
4914	spin_lock_irq(&pool->lock);
4915
 
 
 
 
 
 
 
 
 
 
4916	pool->flags &= ~POOL_DISASSOCIATED;
4917
4918	for_each_pool_worker(worker, pool) {
4919		unsigned int worker_flags = worker->flags;
4920
4921		/*
4922		 * A bound idle worker should actually be on the runqueue
4923		 * of the associated CPU for local wake-ups targeting it to
4924		 * work.  Kick all idle workers so that they migrate to the
4925		 * associated CPU.  Doing this in the same loop as
4926		 * replacing UNBOUND with REBOUND is safe as no worker will
4927		 * be bound before @pool->lock is released.
4928		 */
4929		if (worker_flags & WORKER_IDLE)
4930			wake_up_process(worker->task);
4931
4932		/*
4933		 * We want to clear UNBOUND but can't directly call
4934		 * worker_clr_flags() or adjust nr_running.  Atomically
4935		 * replace UNBOUND with another NOT_RUNNING flag REBOUND.
4936		 * @worker will clear REBOUND using worker_clr_flags() when
4937		 * it initiates the next execution cycle thus restoring
4938		 * concurrency management.  Note that when or whether
4939		 * @worker clears REBOUND doesn't affect correctness.
4940		 *
4941		 * WRITE_ONCE() is necessary because @worker->flags may be
4942		 * tested without holding any lock in
4943		 * wq_worker_running().  Without it, NOT_RUNNING test may
4944		 * fail incorrectly leading to premature concurrency
4945		 * management operations.
4946		 */
4947		WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND));
4948		worker_flags |= WORKER_REBOUND;
4949		worker_flags &= ~WORKER_UNBOUND;
4950		WRITE_ONCE(worker->flags, worker_flags);
4951	}
4952
4953	spin_unlock_irq(&pool->lock);
4954}
4955
4956/**
4957 * restore_unbound_workers_cpumask - restore cpumask of unbound workers
4958 * @pool: unbound pool of interest
4959 * @cpu: the CPU which is coming up
4960 *
4961 * An unbound pool may end up with a cpumask which doesn't have any online
4962 * CPUs.  When a worker of such pool get scheduled, the scheduler resets
4963 * its cpus_allowed.  If @cpu is in @pool's cpumask which didn't have any
4964 * online CPU before, cpus_allowed of all its workers should be restored.
4965 */
4966static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu)
4967{
4968	static cpumask_t cpumask;
4969	struct worker *worker;
4970
4971	lockdep_assert_held(&wq_pool_attach_mutex);
4972
4973	/* is @cpu allowed for @pool? */
4974	if (!cpumask_test_cpu(cpu, pool->attrs->cpumask))
4975		return;
4976
 
4977	cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask);
 
 
4978
4979	/* as we're called from CPU_ONLINE, the following shouldn't fail */
4980	for_each_pool_worker(worker, pool)
4981		WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0);
4982}
4983
4984int workqueue_prepare_cpu(unsigned int cpu)
4985{
4986	struct worker_pool *pool;
4987
4988	for_each_cpu_worker_pool(pool, cpu) {
4989		if (pool->nr_workers)
4990			continue;
4991		if (!create_worker(pool))
4992			return -ENOMEM;
4993	}
4994	return 0;
4995}
4996
4997int workqueue_online_cpu(unsigned int cpu)
 
 
 
 
 
 
4998{
 
4999	struct worker_pool *pool;
5000	struct workqueue_struct *wq;
5001	int pi;
5002
5003	mutex_lock(&wq_pool_mutex);
 
 
 
 
 
 
 
 
5004
5005	for_each_pool(pool, pi) {
5006		mutex_lock(&wq_pool_attach_mutex);
 
5007
5008		if (pool->cpu == cpu)
5009			rebind_workers(pool);
5010		else if (pool->cpu < 0)
5011			restore_unbound_workers_cpumask(pool, cpu);
5012
5013		mutex_unlock(&wq_pool_attach_mutex);
5014	}
 
 
5015
5016	/* update NUMA affinity of unbound workqueues */
5017	list_for_each_entry(wq, &workqueues, list)
5018		wq_update_unbound_numa(wq, cpu, true);
5019
5020	mutex_unlock(&wq_pool_mutex);
5021	return 0;
 
 
 
 
 
 
5022}
5023
5024int workqueue_offline_cpu(unsigned int cpu)
 
 
 
 
 
 
5025{
 
 
5026	struct workqueue_struct *wq;
5027
5028	/* unbinding per-cpu workers should happen on the local CPU */
5029	if (WARN_ON(cpu != smp_processor_id()))
5030		return -1;
5031
5032	unbind_workers(cpu);
5033
5034	/* update NUMA affinity of unbound workqueues */
5035	mutex_lock(&wq_pool_mutex);
5036	list_for_each_entry(wq, &workqueues, list)
5037		wq_update_unbound_numa(wq, cpu, false);
5038	mutex_unlock(&wq_pool_mutex);
5039
5040	return 0;
 
 
 
 
 
5041}
5042
 
 
5043struct work_for_cpu {
5044	struct work_struct work;
5045	long (*fn)(void *);
5046	void *arg;
5047	long ret;
5048};
5049
5050static void work_for_cpu_fn(struct work_struct *work)
5051{
5052	struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work);
5053
5054	wfc->ret = wfc->fn(wfc->arg);
5055}
5056
5057/**
5058 * work_on_cpu - run a function in thread context on a particular cpu
5059 * @cpu: the cpu to run on
5060 * @fn: the function to run
5061 * @arg: the function arg
5062 *
5063 * It is up to the caller to ensure that the cpu doesn't go offline.
5064 * The caller must not hold any locks which would prevent @fn from completing.
5065 *
5066 * Return: The value @fn returns.
5067 */
5068long work_on_cpu(int cpu, long (*fn)(void *), void *arg)
5069{
5070	struct work_for_cpu wfc = { .fn = fn, .arg = arg };
5071
5072	INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn);
5073	schedule_work_on(cpu, &wfc.work);
5074	flush_work(&wfc.work);
5075	destroy_work_on_stack(&wfc.work);
5076	return wfc.ret;
5077}
5078EXPORT_SYMBOL_GPL(work_on_cpu);
5079
5080/**
5081 * work_on_cpu_safe - run a function in thread context on a particular cpu
5082 * @cpu: the cpu to run on
5083 * @fn:  the function to run
5084 * @arg: the function argument
5085 *
5086 * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold
5087 * any locks which would prevent @fn from completing.
5088 *
5089 * Return: The value @fn returns.
5090 */
5091long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg)
5092{
5093	long ret = -ENODEV;
5094
5095	get_online_cpus();
5096	if (cpu_online(cpu))
5097		ret = work_on_cpu(cpu, fn, arg);
5098	put_online_cpus();
5099	return ret;
5100}
5101EXPORT_SYMBOL_GPL(work_on_cpu_safe);
5102#endif /* CONFIG_SMP */
5103
5104#ifdef CONFIG_FREEZER
5105
5106/**
5107 * freeze_workqueues_begin - begin freezing workqueues
5108 *
5109 * Start freezing workqueues.  After this function returns, all freezable
5110 * workqueues will queue new works to their delayed_works list instead of
5111 * pool->worklist.
5112 *
5113 * CONTEXT:
5114 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5115 */
5116void freeze_workqueues_begin(void)
5117{
5118	struct workqueue_struct *wq;
5119	struct pool_workqueue *pwq;
5120
5121	mutex_lock(&wq_pool_mutex);
5122
5123	WARN_ON_ONCE(workqueue_freezing);
5124	workqueue_freezing = true;
5125
5126	list_for_each_entry(wq, &workqueues, list) {
5127		mutex_lock(&wq->mutex);
5128		for_each_pwq(pwq, wq)
5129			pwq_adjust_max_active(pwq);
5130		mutex_unlock(&wq->mutex);
5131	}
5132
5133	mutex_unlock(&wq_pool_mutex);
5134}
5135
5136/**
5137 * freeze_workqueues_busy - are freezable workqueues still busy?
5138 *
5139 * Check whether freezing is complete.  This function must be called
5140 * between freeze_workqueues_begin() and thaw_workqueues().
5141 *
5142 * CONTEXT:
5143 * Grabs and releases wq_pool_mutex.
5144 *
5145 * Return:
5146 * %true if some freezable workqueues are still busy.  %false if freezing
5147 * is complete.
5148 */
5149bool freeze_workqueues_busy(void)
5150{
5151	bool busy = false;
5152	struct workqueue_struct *wq;
5153	struct pool_workqueue *pwq;
5154
5155	mutex_lock(&wq_pool_mutex);
5156
5157	WARN_ON_ONCE(!workqueue_freezing);
5158
5159	list_for_each_entry(wq, &workqueues, list) {
5160		if (!(wq->flags & WQ_FREEZABLE))
5161			continue;
5162		/*
5163		 * nr_active is monotonically decreasing.  It's safe
5164		 * to peek without lock.
5165		 */
5166		rcu_read_lock();
5167		for_each_pwq(pwq, wq) {
5168			WARN_ON_ONCE(pwq->nr_active < 0);
5169			if (pwq->nr_active) {
5170				busy = true;
5171				rcu_read_unlock();
5172				goto out_unlock;
5173			}
5174		}
5175		rcu_read_unlock();
5176	}
5177out_unlock:
5178	mutex_unlock(&wq_pool_mutex);
5179	return busy;
5180}
5181
5182/**
5183 * thaw_workqueues - thaw workqueues
5184 *
5185 * Thaw workqueues.  Normal queueing is restored and all collected
5186 * frozen works are transferred to their respective pool worklists.
5187 *
5188 * CONTEXT:
5189 * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's.
5190 */
5191void thaw_workqueues(void)
5192{
5193	struct workqueue_struct *wq;
5194	struct pool_workqueue *pwq;
5195
5196	mutex_lock(&wq_pool_mutex);
5197
5198	if (!workqueue_freezing)
5199		goto out_unlock;
5200
5201	workqueue_freezing = false;
5202
5203	/* restore max_active and repopulate worklist */
5204	list_for_each_entry(wq, &workqueues, list) {
5205		mutex_lock(&wq->mutex);
5206		for_each_pwq(pwq, wq)
5207			pwq_adjust_max_active(pwq);
5208		mutex_unlock(&wq->mutex);
5209	}
5210
5211out_unlock:
5212	mutex_unlock(&wq_pool_mutex);
5213}
5214#endif /* CONFIG_FREEZER */
5215
5216static int workqueue_apply_unbound_cpumask(void)
5217{
5218	LIST_HEAD(ctxs);
5219	int ret = 0;
5220	struct workqueue_struct *wq;
5221	struct apply_wqattrs_ctx *ctx, *n;
5222
5223	lockdep_assert_held(&wq_pool_mutex);
5224
5225	list_for_each_entry(wq, &workqueues, list) {
5226		if (!(wq->flags & WQ_UNBOUND))
5227			continue;
5228		/* creating multiple pwqs breaks ordering guarantee */
5229		if (wq->flags & __WQ_ORDERED)
5230			continue;
5231
5232		ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs);
5233		if (!ctx) {
5234			ret = -ENOMEM;
5235			break;
5236		}
5237
5238		list_add_tail(&ctx->list, &ctxs);
5239	}
5240
5241	list_for_each_entry_safe(ctx, n, &ctxs, list) {
5242		if (!ret)
5243			apply_wqattrs_commit(ctx);
5244		apply_wqattrs_cleanup(ctx);
5245	}
5246
5247	return ret;
5248}
5249
5250/**
5251 *  workqueue_set_unbound_cpumask - Set the low-level unbound cpumask
5252 *  @cpumask: the cpumask to set
5253 *
5254 *  The low-level workqueues cpumask is a global cpumask that limits
5255 *  the affinity of all unbound workqueues.  This function check the @cpumask
5256 *  and apply it to all unbound workqueues and updates all pwqs of them.
5257 *
5258 *  Retun:	0	- Success
5259 *  		-EINVAL	- Invalid @cpumask
5260 *  		-ENOMEM	- Failed to allocate memory for attrs or pwqs.
5261 */
5262int workqueue_set_unbound_cpumask(cpumask_var_t cpumask)
5263{
5264	int ret = -EINVAL;
5265	cpumask_var_t saved_cpumask;
5266
5267	if (!zalloc_cpumask_var(&saved_cpumask, GFP_KERNEL))
5268		return -ENOMEM;
5269
5270	/*
5271	 * Not excluding isolated cpus on purpose.
5272	 * If the user wishes to include them, we allow that.
5273	 */
5274	cpumask_and(cpumask, cpumask, cpu_possible_mask);
5275	if (!cpumask_empty(cpumask)) {
5276		apply_wqattrs_lock();
5277
5278		/* save the old wq_unbound_cpumask. */
5279		cpumask_copy(saved_cpumask, wq_unbound_cpumask);
5280
5281		/* update wq_unbound_cpumask at first and apply it to wqs. */
5282		cpumask_copy(wq_unbound_cpumask, cpumask);
5283		ret = workqueue_apply_unbound_cpumask();
5284
5285		/* restore the wq_unbound_cpumask when failed. */
5286		if (ret < 0)
5287			cpumask_copy(wq_unbound_cpumask, saved_cpumask);
5288
5289		apply_wqattrs_unlock();
5290	}
5291
5292	free_cpumask_var(saved_cpumask);
5293	return ret;
5294}
5295
5296#ifdef CONFIG_SYSFS
5297/*
5298 * Workqueues with WQ_SYSFS flag set is visible to userland via
5299 * /sys/bus/workqueue/devices/WQ_NAME.  All visible workqueues have the
5300 * following attributes.
5301 *
5302 *  per_cpu	RO bool	: whether the workqueue is per-cpu or unbound
5303 *  max_active	RW int	: maximum number of in-flight work items
5304 *
5305 * Unbound workqueues have the following extra attributes.
5306 *
5307 *  pool_ids	RO int	: the associated pool IDs for each node
5308 *  nice	RW int	: nice value of the workers
5309 *  cpumask	RW mask	: bitmask of allowed CPUs for the workers
5310 *  numa	RW bool	: whether enable NUMA affinity
5311 */
5312struct wq_device {
5313	struct workqueue_struct		*wq;
5314	struct device			dev;
5315};
5316
5317static struct workqueue_struct *dev_to_wq(struct device *dev)
5318{
5319	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5320
5321	return wq_dev->wq;
5322}
5323
5324static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr,
5325			    char *buf)
5326{
5327	struct workqueue_struct *wq = dev_to_wq(dev);
5328
5329	return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND));
5330}
5331static DEVICE_ATTR_RO(per_cpu);
5332
5333static ssize_t max_active_show(struct device *dev,
5334			       struct device_attribute *attr, char *buf)
5335{
5336	struct workqueue_struct *wq = dev_to_wq(dev);
5337
5338	return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active);
5339}
5340
5341static ssize_t max_active_store(struct device *dev,
5342				struct device_attribute *attr, const char *buf,
5343				size_t count)
5344{
5345	struct workqueue_struct *wq = dev_to_wq(dev);
5346	int val;
5347
5348	if (sscanf(buf, "%d", &val) != 1 || val <= 0)
5349		return -EINVAL;
5350
5351	workqueue_set_max_active(wq, val);
5352	return count;
5353}
5354static DEVICE_ATTR_RW(max_active);
5355
5356static struct attribute *wq_sysfs_attrs[] = {
5357	&dev_attr_per_cpu.attr,
5358	&dev_attr_max_active.attr,
5359	NULL,
5360};
5361ATTRIBUTE_GROUPS(wq_sysfs);
5362
5363static ssize_t wq_pool_ids_show(struct device *dev,
5364				struct device_attribute *attr, char *buf)
5365{
5366	struct workqueue_struct *wq = dev_to_wq(dev);
5367	const char *delim = "";
5368	int node, written = 0;
5369
5370	get_online_cpus();
5371	rcu_read_lock();
5372	for_each_node(node) {
5373		written += scnprintf(buf + written, PAGE_SIZE - written,
5374				     "%s%d:%d", delim, node,
5375				     unbound_pwq_by_node(wq, node)->pool->id);
5376		delim = " ";
5377	}
5378	written += scnprintf(buf + written, PAGE_SIZE - written, "\n");
5379	rcu_read_unlock();
5380	put_online_cpus();
5381
5382	return written;
5383}
5384
5385static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr,
5386			    char *buf)
5387{
5388	struct workqueue_struct *wq = dev_to_wq(dev);
5389	int written;
5390
5391	mutex_lock(&wq->mutex);
5392	written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice);
5393	mutex_unlock(&wq->mutex);
5394
5395	return written;
5396}
5397
5398/* prepare workqueue_attrs for sysfs store operations */
5399static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq)
5400{
5401	struct workqueue_attrs *attrs;
5402
5403	lockdep_assert_held(&wq_pool_mutex);
5404
5405	attrs = alloc_workqueue_attrs();
5406	if (!attrs)
5407		return NULL;
5408
5409	copy_workqueue_attrs(attrs, wq->unbound_attrs);
5410	return attrs;
5411}
5412
5413static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr,
5414			     const char *buf, size_t count)
5415{
5416	struct workqueue_struct *wq = dev_to_wq(dev);
5417	struct workqueue_attrs *attrs;
5418	int ret = -ENOMEM;
5419
5420	apply_wqattrs_lock();
5421
5422	attrs = wq_sysfs_prep_attrs(wq);
5423	if (!attrs)
5424		goto out_unlock;
5425
5426	if (sscanf(buf, "%d", &attrs->nice) == 1 &&
5427	    attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE)
5428		ret = apply_workqueue_attrs_locked(wq, attrs);
5429	else
5430		ret = -EINVAL;
5431
5432out_unlock:
5433	apply_wqattrs_unlock();
5434	free_workqueue_attrs(attrs);
5435	return ret ?: count;
5436}
5437
5438static ssize_t wq_cpumask_show(struct device *dev,
5439			       struct device_attribute *attr, char *buf)
5440{
5441	struct workqueue_struct *wq = dev_to_wq(dev);
5442	int written;
5443
5444	mutex_lock(&wq->mutex);
5445	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5446			    cpumask_pr_args(wq->unbound_attrs->cpumask));
5447	mutex_unlock(&wq->mutex);
5448	return written;
5449}
5450
5451static ssize_t wq_cpumask_store(struct device *dev,
5452				struct device_attribute *attr,
5453				const char *buf, size_t count)
5454{
5455	struct workqueue_struct *wq = dev_to_wq(dev);
5456	struct workqueue_attrs *attrs;
5457	int ret = -ENOMEM;
5458
5459	apply_wqattrs_lock();
5460
5461	attrs = wq_sysfs_prep_attrs(wq);
5462	if (!attrs)
5463		goto out_unlock;
5464
5465	ret = cpumask_parse(buf, attrs->cpumask);
5466	if (!ret)
5467		ret = apply_workqueue_attrs_locked(wq, attrs);
5468
5469out_unlock:
5470	apply_wqattrs_unlock();
5471	free_workqueue_attrs(attrs);
5472	return ret ?: count;
5473}
5474
5475static ssize_t wq_numa_show(struct device *dev, struct device_attribute *attr,
5476			    char *buf)
5477{
5478	struct workqueue_struct *wq = dev_to_wq(dev);
5479	int written;
5480
5481	mutex_lock(&wq->mutex);
5482	written = scnprintf(buf, PAGE_SIZE, "%d\n",
5483			    !wq->unbound_attrs->no_numa);
5484	mutex_unlock(&wq->mutex);
5485
5486	return written;
5487}
5488
5489static ssize_t wq_numa_store(struct device *dev, struct device_attribute *attr,
5490			     const char *buf, size_t count)
5491{
5492	struct workqueue_struct *wq = dev_to_wq(dev);
5493	struct workqueue_attrs *attrs;
5494	int v, ret = -ENOMEM;
5495
5496	apply_wqattrs_lock();
5497
5498	attrs = wq_sysfs_prep_attrs(wq);
5499	if (!attrs)
5500		goto out_unlock;
5501
5502	ret = -EINVAL;
5503	if (sscanf(buf, "%d", &v) == 1) {
5504		attrs->no_numa = !v;
5505		ret = apply_workqueue_attrs_locked(wq, attrs);
5506	}
5507
5508out_unlock:
5509	apply_wqattrs_unlock();
5510	free_workqueue_attrs(attrs);
5511	return ret ?: count;
5512}
5513
5514static struct device_attribute wq_sysfs_unbound_attrs[] = {
5515	__ATTR(pool_ids, 0444, wq_pool_ids_show, NULL),
5516	__ATTR(nice, 0644, wq_nice_show, wq_nice_store),
5517	__ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store),
5518	__ATTR(numa, 0644, wq_numa_show, wq_numa_store),
5519	__ATTR_NULL,
5520};
5521
5522static struct bus_type wq_subsys = {
5523	.name				= "workqueue",
5524	.dev_groups			= wq_sysfs_groups,
5525};
5526
5527static ssize_t wq_unbound_cpumask_show(struct device *dev,
5528		struct device_attribute *attr, char *buf)
5529{
5530	int written;
5531
5532	mutex_lock(&wq_pool_mutex);
5533	written = scnprintf(buf, PAGE_SIZE, "%*pb\n",
5534			    cpumask_pr_args(wq_unbound_cpumask));
5535	mutex_unlock(&wq_pool_mutex);
5536
5537	return written;
5538}
5539
5540static ssize_t wq_unbound_cpumask_store(struct device *dev,
5541		struct device_attribute *attr, const char *buf, size_t count)
5542{
5543	cpumask_var_t cpumask;
5544	int ret;
5545
5546	if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL))
5547		return -ENOMEM;
5548
5549	ret = cpumask_parse(buf, cpumask);
5550	if (!ret)
5551		ret = workqueue_set_unbound_cpumask(cpumask);
5552
5553	free_cpumask_var(cpumask);
5554	return ret ? ret : count;
5555}
5556
5557static struct device_attribute wq_sysfs_cpumask_attr =
5558	__ATTR(cpumask, 0644, wq_unbound_cpumask_show,
5559	       wq_unbound_cpumask_store);
5560
5561static int __init wq_sysfs_init(void)
5562{
5563	int err;
5564
5565	err = subsys_virtual_register(&wq_subsys, NULL);
5566	if (err)
5567		return err;
5568
5569	return device_create_file(wq_subsys.dev_root, &wq_sysfs_cpumask_attr);
5570}
5571core_initcall(wq_sysfs_init);
5572
5573static void wq_device_release(struct device *dev)
5574{
5575	struct wq_device *wq_dev = container_of(dev, struct wq_device, dev);
5576
5577	kfree(wq_dev);
5578}
5579
5580/**
5581 * workqueue_sysfs_register - make a workqueue visible in sysfs
5582 * @wq: the workqueue to register
5583 *
5584 * Expose @wq in sysfs under /sys/bus/workqueue/devices.
5585 * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set
5586 * which is the preferred method.
5587 *
5588 * Workqueue user should use this function directly iff it wants to apply
5589 * workqueue_attrs before making the workqueue visible in sysfs; otherwise,
5590 * apply_workqueue_attrs() may race against userland updating the
5591 * attributes.
5592 *
5593 * Return: 0 on success, -errno on failure.
5594 */
5595int workqueue_sysfs_register(struct workqueue_struct *wq)
5596{
5597	struct wq_device *wq_dev;
5598	int ret;
5599
5600	/*
5601	 * Adjusting max_active or creating new pwqs by applying
5602	 * attributes breaks ordering guarantee.  Disallow exposing ordered
5603	 * workqueues.
5604	 */
5605	if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT))
5606		return -EINVAL;
5607
5608	wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL);
5609	if (!wq_dev)
5610		return -ENOMEM;
5611
5612	wq_dev->wq = wq;
5613	wq_dev->dev.bus = &wq_subsys;
5614	wq_dev->dev.release = wq_device_release;
5615	dev_set_name(&wq_dev->dev, "%s", wq->name);
5616
5617	/*
5618	 * unbound_attrs are created separately.  Suppress uevent until
5619	 * everything is ready.
5620	 */
5621	dev_set_uevent_suppress(&wq_dev->dev, true);
5622
5623	ret = device_register(&wq_dev->dev);
5624	if (ret) {
5625		put_device(&wq_dev->dev);
5626		wq->wq_dev = NULL;
5627		return ret;
5628	}
5629
5630	if (wq->flags & WQ_UNBOUND) {
5631		struct device_attribute *attr;
5632
5633		for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) {
5634			ret = device_create_file(&wq_dev->dev, attr);
5635			if (ret) {
5636				device_unregister(&wq_dev->dev);
5637				wq->wq_dev = NULL;
5638				return ret;
5639			}
5640		}
5641	}
5642
5643	dev_set_uevent_suppress(&wq_dev->dev, false);
5644	kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD);
5645	return 0;
5646}
5647
5648/**
5649 * workqueue_sysfs_unregister - undo workqueue_sysfs_register()
5650 * @wq: the workqueue to unregister
5651 *
5652 * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister.
5653 */
5654static void workqueue_sysfs_unregister(struct workqueue_struct *wq)
5655{
5656	struct wq_device *wq_dev = wq->wq_dev;
5657
5658	if (!wq->wq_dev)
5659		return;
5660
5661	wq->wq_dev = NULL;
5662	device_unregister(&wq_dev->dev);
5663}
5664#else	/* CONFIG_SYSFS */
5665static void workqueue_sysfs_unregister(struct workqueue_struct *wq)	{ }
5666#endif	/* CONFIG_SYSFS */
5667
5668/*
5669 * Workqueue watchdog.
5670 *
5671 * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal
5672 * flush dependency, a concurrency managed work item which stays RUNNING
5673 * indefinitely.  Workqueue stalls can be very difficult to debug as the
5674 * usual warning mechanisms don't trigger and internal workqueue state is
5675 * largely opaque.
5676 *
5677 * Workqueue watchdog monitors all worker pools periodically and dumps
5678 * state if some pools failed to make forward progress for a while where
5679 * forward progress is defined as the first item on ->worklist changing.
5680 *
5681 * This mechanism is controlled through the kernel parameter
5682 * "workqueue.watchdog_thresh" which can be updated at runtime through the
5683 * corresponding sysfs parameter file.
5684 */
5685#ifdef CONFIG_WQ_WATCHDOG
5686
 
 
5687static unsigned long wq_watchdog_thresh = 30;
5688static struct timer_list wq_watchdog_timer;
 
5689
5690static unsigned long wq_watchdog_touched = INITIAL_JIFFIES;
5691static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES;
5692
5693static void wq_watchdog_reset_touched(void)
5694{
5695	int cpu;
5696
5697	wq_watchdog_touched = jiffies;
5698	for_each_possible_cpu(cpu)
5699		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5700}
5701
5702static void wq_watchdog_timer_fn(struct timer_list *unused)
5703{
5704	unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ;
5705	bool lockup_detected = false;
5706	struct worker_pool *pool;
5707	int pi;
5708
5709	if (!thresh)
5710		return;
5711
5712	rcu_read_lock();
5713
5714	for_each_pool(pool, pi) {
5715		unsigned long pool_ts, touched, ts;
5716
5717		if (list_empty(&pool->worklist))
5718			continue;
5719
5720		/* get the latest of pool and touched timestamps */
5721		pool_ts = READ_ONCE(pool->watchdog_ts);
5722		touched = READ_ONCE(wq_watchdog_touched);
5723
5724		if (time_after(pool_ts, touched))
5725			ts = pool_ts;
5726		else
5727			ts = touched;
5728
5729		if (pool->cpu >= 0) {
5730			unsigned long cpu_touched =
5731				READ_ONCE(per_cpu(wq_watchdog_touched_cpu,
5732						  pool->cpu));
5733			if (time_after(cpu_touched, ts))
5734				ts = cpu_touched;
5735		}
5736
5737		/* did we stall? */
5738		if (time_after(jiffies, ts + thresh)) {
5739			lockup_detected = true;
5740			pr_emerg("BUG: workqueue lockup - pool");
5741			pr_cont_pool_info(pool);
5742			pr_cont(" stuck for %us!\n",
5743				jiffies_to_msecs(jiffies - pool_ts) / 1000);
5744		}
5745	}
5746
5747	rcu_read_unlock();
5748
5749	if (lockup_detected)
5750		show_workqueue_state();
5751
5752	wq_watchdog_reset_touched();
5753	mod_timer(&wq_watchdog_timer, jiffies + thresh);
5754}
5755
5756notrace void wq_watchdog_touch(int cpu)
5757{
5758	if (cpu >= 0)
5759		per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies;
5760	else
5761		wq_watchdog_touched = jiffies;
5762}
5763
5764static void wq_watchdog_set_thresh(unsigned long thresh)
5765{
5766	wq_watchdog_thresh = 0;
5767	del_timer_sync(&wq_watchdog_timer);
5768
5769	if (thresh) {
5770		wq_watchdog_thresh = thresh;
5771		wq_watchdog_reset_touched();
5772		mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ);
5773	}
5774}
5775
5776static int wq_watchdog_param_set_thresh(const char *val,
5777					const struct kernel_param *kp)
5778{
5779	unsigned long thresh;
5780	int ret;
5781
5782	ret = kstrtoul(val, 0, &thresh);
5783	if (ret)
5784		return ret;
5785
5786	if (system_wq)
5787		wq_watchdog_set_thresh(thresh);
5788	else
5789		wq_watchdog_thresh = thresh;
5790
5791	return 0;
5792}
5793
5794static const struct kernel_param_ops wq_watchdog_thresh_ops = {
5795	.set	= wq_watchdog_param_set_thresh,
5796	.get	= param_get_ulong,
5797};
5798
5799module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh,
5800		0644);
5801
5802static void wq_watchdog_init(void)
5803{
5804	timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE);
5805	wq_watchdog_set_thresh(wq_watchdog_thresh);
5806}
5807
5808#else	/* CONFIG_WQ_WATCHDOG */
5809
5810static inline void wq_watchdog_init(void) { }
5811
5812#endif	/* CONFIG_WQ_WATCHDOG */
5813
5814static void __init wq_numa_init(void)
5815{
5816	cpumask_var_t *tbl;
5817	int node, cpu;
5818
5819	if (num_possible_nodes() <= 1)
5820		return;
5821
5822	if (wq_disable_numa) {
5823		pr_info("workqueue: NUMA affinity support disabled\n");
5824		return;
5825	}
5826
5827	wq_update_unbound_numa_attrs_buf = alloc_workqueue_attrs();
5828	BUG_ON(!wq_update_unbound_numa_attrs_buf);
5829
5830	/*
5831	 * We want masks of possible CPUs of each node which isn't readily
5832	 * available.  Build one from cpu_to_node() which should have been
5833	 * fully initialized by now.
5834	 */
5835	tbl = kcalloc(nr_node_ids, sizeof(tbl[0]), GFP_KERNEL);
5836	BUG_ON(!tbl);
5837
5838	for_each_node(node)
5839		BUG_ON(!zalloc_cpumask_var_node(&tbl[node], GFP_KERNEL,
5840				node_online(node) ? node : NUMA_NO_NODE));
5841
5842	for_each_possible_cpu(cpu) {
5843		node = cpu_to_node(cpu);
5844		if (WARN_ON(node == NUMA_NO_NODE)) {
5845			pr_warn("workqueue: NUMA node mapping not available for cpu%d, disabling NUMA support\n", cpu);
5846			/* happens iff arch is bonkers, let's just proceed */
5847			return;
5848		}
5849		cpumask_set_cpu(cpu, tbl[node]);
5850	}
5851
5852	wq_numa_possible_cpumask = tbl;
5853	wq_numa_enabled = true;
5854}
5855
5856/**
5857 * workqueue_init_early - early init for workqueue subsystem
5858 *
5859 * This is the first half of two-staged workqueue subsystem initialization
5860 * and invoked as soon as the bare basics - memory allocation, cpumasks and
5861 * idr are up.  It sets up all the data structures and system workqueues
5862 * and allows early boot code to create workqueues and queue/cancel work
5863 * items.  Actual work item execution starts only after kthreads can be
5864 * created and scheduled right before early initcalls.
5865 */
5866int __init workqueue_init_early(void)
5867{
5868	int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL };
5869	int hk_flags = HK_FLAG_DOMAIN | HK_FLAG_WQ;
5870	int i, cpu;
5871
5872	WARN_ON(__alignof__(struct pool_workqueue) < __alignof__(long long));
5873
5874	BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL));
5875	cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(hk_flags));
5876
5877	pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC);
5878
 
 
 
 
 
5879	/* initialize CPU pools */
5880	for_each_possible_cpu(cpu) {
5881		struct worker_pool *pool;
5882
5883		i = 0;
5884		for_each_cpu_worker_pool(pool, cpu) {
5885			BUG_ON(init_worker_pool(pool));
5886			pool->cpu = cpu;
5887			cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu));
5888			pool->attrs->nice = std_nice[i++];
5889			pool->node = cpu_to_node(cpu);
5890
5891			/* alloc pool ID */
5892			mutex_lock(&wq_pool_mutex);
5893			BUG_ON(worker_pool_assign_id(pool));
5894			mutex_unlock(&wq_pool_mutex);
5895		}
5896	}
5897
 
 
 
 
 
 
 
 
 
 
5898	/* create default unbound and ordered wq attrs */
5899	for (i = 0; i < NR_STD_WORKER_POOLS; i++) {
5900		struct workqueue_attrs *attrs;
5901
5902		BUG_ON(!(attrs = alloc_workqueue_attrs()));
5903		attrs->nice = std_nice[i];
5904		unbound_std_wq_attrs[i] = attrs;
5905
5906		/*
5907		 * An ordered wq should have only one pwq as ordering is
5908		 * guaranteed by max_active which is enforced by pwqs.
5909		 * Turn off NUMA so that dfl_pwq is used for all nodes.
5910		 */
5911		BUG_ON(!(attrs = alloc_workqueue_attrs()));
5912		attrs->nice = std_nice[i];
5913		attrs->no_numa = true;
5914		ordered_wq_attrs[i] = attrs;
5915	}
5916
5917	system_wq = alloc_workqueue("events", 0, 0);
5918	system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0);
5919	system_long_wq = alloc_workqueue("events_long", 0, 0);
5920	system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
5921					    WQ_UNBOUND_MAX_ACTIVE);
5922	system_freezable_wq = alloc_workqueue("events_freezable",
5923					      WQ_FREEZABLE, 0);
5924	system_power_efficient_wq = alloc_workqueue("events_power_efficient",
5925					      WQ_POWER_EFFICIENT, 0);
5926	system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient",
5927					      WQ_FREEZABLE | WQ_POWER_EFFICIENT,
5928					      0);
5929	BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq ||
5930	       !system_unbound_wq || !system_freezable_wq ||
5931	       !system_power_efficient_wq ||
5932	       !system_freezable_power_efficient_wq);
5933
5934	return 0;
5935}
5936
5937/**
5938 * workqueue_init - bring workqueue subsystem fully online
5939 *
5940 * This is the latter half of two-staged workqueue subsystem initialization
5941 * and invoked as soon as kthreads can be created and scheduled.
5942 * Workqueues have been created and work items queued on them, but there
5943 * are no kworkers executing the work items yet.  Populate the worker pools
5944 * with the initial workers and enable future kworker creations.
5945 */
5946int __init workqueue_init(void)
5947{
5948	struct workqueue_struct *wq;
5949	struct worker_pool *pool;
5950	int cpu, bkt;
5951
5952	/*
5953	 * It'd be simpler to initialize NUMA in workqueue_init_early() but
5954	 * CPU to node mapping may not be available that early on some
5955	 * archs such as power and arm64.  As per-cpu pools created
5956	 * previously could be missing node hint and unbound pools NUMA
5957	 * affinity, fix them up.
5958	 *
5959	 * Also, while iterating workqueues, create rescuers if requested.
5960	 */
5961	wq_numa_init();
5962
5963	mutex_lock(&wq_pool_mutex);
5964
5965	for_each_possible_cpu(cpu) {
5966		for_each_cpu_worker_pool(pool, cpu) {
5967			pool->node = cpu_to_node(cpu);
5968		}
5969	}
5970
5971	list_for_each_entry(wq, &workqueues, list) {
5972		wq_update_unbound_numa(wq, smp_processor_id(), true);
5973		WARN(init_rescuer(wq),
5974		     "workqueue: failed to create early rescuer for %s",
5975		     wq->name);
5976	}
5977
5978	mutex_unlock(&wq_pool_mutex);
5979
5980	/* create the initial workers */
5981	for_each_online_cpu(cpu) {
5982		for_each_cpu_worker_pool(pool, cpu) {
5983			pool->flags &= ~POOL_DISASSOCIATED;
5984			BUG_ON(!create_worker(pool));
5985		}
5986	}
5987
5988	hash_for_each(unbound_pool_hash, bkt, pool, hash_node)
5989		BUG_ON(!create_worker(pool));
5990
5991	wq_online = true;
5992	wq_watchdog_init();
5993
5994	return 0;
5995}
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);