1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 3c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 41da177e4SLinus Torvalds * 5c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 81da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 9e1f8e874SFrancois Cami * Andrew Morton 101da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 111da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1289ada679SChristoph Lameter * 13cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 14c54fce6eSTejun Heo * 15c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 16c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 17c54fce6eSTejun Heo * 18c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 19c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 20b11895c4SLibin * automatically managed. There are two worker pools for each CPU (one for 21b11895c4SLibin * normal work items and the other for high priority ones) and some extra 22b11895c4SLibin * pools for workqueues which are not bound to any specific CPU - the 23b11895c4SLibin * number of these backing pools is dynamic. 24c54fce6eSTejun Heo * 259a261491SBenjamin Peterson * Please read Documentation/core-api/workqueue.rst for details. 261da177e4SLinus Torvalds */ 271da177e4SLinus Torvalds 289984de1aSPaul Gortmaker #include <linux/export.h> 291da177e4SLinus Torvalds #include <linux/kernel.h> 301da177e4SLinus Torvalds #include <linux/sched.h> 311da177e4SLinus Torvalds #include <linux/init.h> 321da177e4SLinus Torvalds #include <linux/signal.h> 331da177e4SLinus Torvalds #include <linux/completion.h> 341da177e4SLinus Torvalds #include <linux/workqueue.h> 351da177e4SLinus Torvalds #include <linux/slab.h> 361da177e4SLinus Torvalds #include <linux/cpu.h> 371da177e4SLinus Torvalds #include <linux/notifier.h> 381da177e4SLinus Torvalds #include <linux/kthread.h> 391fa44ecaSJames Bottomley #include <linux/hardirq.h> 4046934023SChristoph Lameter #include <linux/mempolicy.h> 41341a5958SRafael J. Wysocki #include <linux/freezer.h> 42d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 434e6045f1SJohannes Berg #include <linux/lockdep.h> 44c34056a3STejun Heo #include <linux/idr.h> 4529c91e99STejun Heo #include <linux/jhash.h> 4642f8570fSSasha Levin #include <linux/hashtable.h> 4776af4d93STejun Heo #include <linux/rculist.h> 48bce90380STejun Heo #include <linux/nodemask.h> 494c16bd32STejun Heo #include <linux/moduleparam.h> 503d1cb205STejun Heo #include <linux/uaccess.h> 51c98a9805STal Shorer #include <linux/sched/isolation.h> 52cd2440d6SPetr Mladek #include <linux/sched/debug.h> 5362635ea8SSergey Senozhatsky #include <linux/nmi.h> 54940d71c6SSergey Senozhatsky #include <linux/kvm_para.h> 55aa6fde93STejun Heo #include <linux/delay.h> 56e22bee78STejun Heo 57ea138446STejun Heo #include "workqueue_internal.h" 581da177e4SLinus Torvalds 59c8e55f36STejun Heo enum { 60bc2ae0f5STejun Heo /* 6124647570STejun Heo * worker_pool flags 62bc2ae0f5STejun Heo * 6324647570STejun Heo * A bound pool is either associated or disassociated with its CPU. 64bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 65bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 66bc2ae0f5STejun Heo * is in effect. 67bc2ae0f5STejun Heo * 68bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 69bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 7024647570STejun Heo * be executing on any CPU. The pool behaves as an unbound one. 71bc2ae0f5STejun Heo * 72bc3a1afcSTejun Heo * Note that DISASSOCIATED should be flipped only while holding 731258fae7STejun Heo * wq_pool_attach_mutex to avoid changing binding state while 744736cbf7SLai Jiangshan * worker_attach_to_pool() is in progress. 75bc2ae0f5STejun Heo */ 76692b4825STejun Heo POOL_MANAGER_ACTIVE = 1 << 0, /* being managed */ 7724647570STejun Heo POOL_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ 78db7bccf4STejun Heo 79c8e55f36STejun Heo /* worker flags */ 80c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 81c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 82e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 83fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 84f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 85a9ab775bSTejun Heo WORKER_REBOUND = 1 << 8, /* worker was rebound */ 86e22bee78STejun Heo 87a9ab775bSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_CPU_INTENSIVE | 88a9ab775bSTejun Heo WORKER_UNBOUND | WORKER_REBOUND, 89db7bccf4STejun Heo 90e34cdddbSTejun Heo NR_STD_WORKER_POOLS = 2, /* # standard pools per cpu */ 914ce62e9eSTejun Heo 9229c91e99STejun Heo UNBOUND_POOL_HASH_ORDER = 6, /* hashed by pool->attrs */ 93c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 94db7bccf4STejun Heo 95e22bee78STejun Heo MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ 96e22bee78STejun Heo IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ 97e22bee78STejun Heo 983233cdbdSTejun Heo MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, 993233cdbdSTejun Heo /* call for help after 10ms 1003233cdbdSTejun Heo (min two ticks) */ 101e22bee78STejun Heo MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ 102e22bee78STejun Heo CREATE_COOLDOWN = HZ, /* time to breath after fail */ 1031da177e4SLinus Torvalds 1041da177e4SLinus Torvalds /* 105e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 1068698a745SDongsheng Yang * all cpus. Give MIN_NICE. 107e22bee78STejun Heo */ 1088698a745SDongsheng Yang RESCUER_NICE_LEVEL = MIN_NICE, 1098698a745SDongsheng Yang HIGHPRI_NICE_LEVEL = MIN_NICE, 110ecf6881fSTejun Heo 111ecf6881fSTejun Heo WQ_NAME_LEN = 24, 112c8e55f36STejun Heo }; 113c8e55f36STejun Heo 1141da177e4SLinus Torvalds /* 1154690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1164690c4abSTejun Heo * 117e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 118e41e704bSTejun Heo * everyone else. 1194690c4abSTejun Heo * 120e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 121e22bee78STejun Heo * only be modified and accessed from the local cpu. 122e22bee78STejun Heo * 123d565ed63STejun Heo * L: pool->lock protected. Access with pool->lock held. 1244690c4abSTejun Heo * 125bdf8b9bfSTejun Heo * K: Only modified by worker while holding pool->lock. Can be safely read by 126bdf8b9bfSTejun Heo * self, while holding pool->lock or from IRQ context if %current is the 127bdf8b9bfSTejun Heo * kworker. 128bdf8b9bfSTejun Heo * 129bdf8b9bfSTejun Heo * S: Only modified by worker self. 130bdf8b9bfSTejun Heo * 1311258fae7STejun Heo * A: wq_pool_attach_mutex protected. 132822d8405STejun Heo * 13368e13a67SLai Jiangshan * PL: wq_pool_mutex protected. 13476af4d93STejun Heo * 13524acfb71SThomas Gleixner * PR: wq_pool_mutex protected for writes. RCU protected for reads. 1365bcab335STejun Heo * 1375b95e1afSLai Jiangshan * PW: wq_pool_mutex and wq->mutex protected for writes. Either for reads. 1385b95e1afSLai Jiangshan * 1395b95e1afSLai Jiangshan * PWR: wq_pool_mutex and wq->mutex protected for writes. Either or 14024acfb71SThomas Gleixner * RCU for reads. 1415b95e1afSLai Jiangshan * 1423c25a55dSLai Jiangshan * WQ: wq->mutex protected. 1433c25a55dSLai Jiangshan * 14424acfb71SThomas Gleixner * WR: wq->mutex protected for writes. RCU protected for reads. 1452e109a28STejun Heo * 1462e109a28STejun Heo * MD: wq_mayday_lock protected. 147cd2440d6SPetr Mladek * 148cd2440d6SPetr Mladek * WD: Used internally by the watchdog. 1494690c4abSTejun Heo */ 1504690c4abSTejun Heo 1512eaebdb3STejun Heo /* struct worker is defined in workqueue_internal.h */ 152c34056a3STejun Heo 153bd7bdd43STejun Heo struct worker_pool { 154a9b8a985SSebastian Andrzej Siewior raw_spinlock_t lock; /* the pool lock */ 155d84ff051STejun Heo int cpu; /* I: the associated cpu */ 156f3f90ad4STejun Heo int node; /* I: the associated node ID */ 1579daf9e67STejun Heo int id; /* I: pool ID */ 158bc8b50c2STejun Heo unsigned int flags; /* L: flags */ 159bd7bdd43STejun Heo 16082607adcSTejun Heo unsigned long watchdog_ts; /* L: watchdog timestamp */ 161cd2440d6SPetr Mladek bool cpu_stall; /* WD: stalled cpu bound pool */ 16282607adcSTejun Heo 163bc35f7efSLai Jiangshan /* 164bc35f7efSLai Jiangshan * The counter is incremented in a process context on the associated CPU 165bc35f7efSLai Jiangshan * w/ preemption disabled, and decremented or reset in the same context 166bc35f7efSLai Jiangshan * but w/ pool->lock held. The readers grab pool->lock and are 167bc35f7efSLai Jiangshan * guaranteed to see if the counter reached zero. 168bc35f7efSLai Jiangshan */ 169bc35f7efSLai Jiangshan int nr_running; 17084f91c62SLai Jiangshan 171bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 172ea1abd61SLai Jiangshan 1735826cc8fSLai Jiangshan int nr_workers; /* L: total number of workers */ 1745826cc8fSLai Jiangshan int nr_idle; /* L: currently idle workers */ 175bd7bdd43STejun Heo 1762c1f1a91SLai Jiangshan struct list_head idle_list; /* L: list of idle workers */ 177bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 1783f959aa3SValentin Schneider struct work_struct idle_cull_work; /* L: worker idle cleanup */ 1793f959aa3SValentin Schneider 180bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 181bd7bdd43STejun Heo 182c5aa87bbSTejun Heo /* a workers is either on busy_hash or idle_list, or the manager */ 183c9e7cf27STejun Heo DECLARE_HASHTABLE(busy_hash, BUSY_WORKER_HASH_ORDER); 184c9e7cf27STejun Heo /* L: hash of busy workers */ 185c9e7cf27STejun Heo 1862607d7a6STejun Heo struct worker *manager; /* L: purely informational */ 18792f9c5c4SLai Jiangshan struct list_head workers; /* A: attached workers */ 188e02b9312SValentin Schneider struct list_head dying_workers; /* A: workers about to die */ 18960f5a4bcSLai Jiangshan struct completion *detach_completion; /* all workers detached */ 190e19e397aSTejun Heo 1917cda9aaeSLai Jiangshan struct ida worker_ida; /* worker IDs for task name */ 192e19e397aSTejun Heo 1937a4e344cSTejun Heo struct workqueue_attrs *attrs; /* I: worker attributes */ 19468e13a67SLai Jiangshan struct hlist_node hash_node; /* PL: unbound_pool_hash node */ 19568e13a67SLai Jiangshan int refcnt; /* PL: refcnt for unbound pools */ 1967a4e344cSTejun Heo 197e19e397aSTejun Heo /* 19824acfb71SThomas Gleixner * Destruction of pool is RCU protected to allow dereferences 19929c91e99STejun Heo * from get_work_pool(). 20029c91e99STejun Heo */ 20129c91e99STejun Heo struct rcu_head rcu; 20284f91c62SLai Jiangshan }; 2038b03ae3cSTejun Heo 2048b03ae3cSTejun Heo /* 205725e8ec5STejun Heo * Per-pool_workqueue statistics. These can be monitored using 206725e8ec5STejun Heo * tools/workqueue/wq_monitor.py. 207725e8ec5STejun Heo */ 208725e8ec5STejun Heo enum pool_workqueue_stats { 209725e8ec5STejun Heo PWQ_STAT_STARTED, /* work items started execution */ 210725e8ec5STejun Heo PWQ_STAT_COMPLETED, /* work items completed execution */ 2118a1dd1e5STejun Heo PWQ_STAT_CPU_TIME, /* total CPU time consumed */ 212616db877STejun Heo PWQ_STAT_CPU_INTENSIVE, /* wq_cpu_intensive_thresh_us violations */ 213725e8ec5STejun Heo PWQ_STAT_CM_WAKEUP, /* concurrency-management worker wakeups */ 214725e8ec5STejun Heo PWQ_STAT_MAYDAY, /* maydays to rescuer */ 215725e8ec5STejun Heo PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 216725e8ec5STejun Heo 217725e8ec5STejun Heo PWQ_NR_STATS, 218725e8ec5STejun Heo }; 219725e8ec5STejun Heo 220725e8ec5STejun Heo /* 221112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 222112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 223112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 224112202d9STejun Heo * number of flag bits. 2251da177e4SLinus Torvalds */ 226112202d9STejun Heo struct pool_workqueue { 227bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2284690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 22973f53c4aSTejun Heo int work_color; /* L: current color */ 23073f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2318864b4e5STejun Heo int refcnt; /* L: reference count */ 23273f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 23373f53c4aSTejun Heo /* L: nr of in_flight works */ 234018f3a13SLai Jiangshan 235018f3a13SLai Jiangshan /* 236018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 237018f3a13SLai Jiangshan * 238018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 239018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 240018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 241018f3a13SLai Jiangshan * 242018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 243018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 244018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 245018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 246018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 247018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 248018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 249018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 250018f3a13SLai Jiangshan */ 2511e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 252a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 253f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2543c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2552e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2568864b4e5STejun Heo 257725e8ec5STejun Heo u64 stats[PWQ_NR_STATS]; 258725e8ec5STejun Heo 2598864b4e5STejun Heo /* 260967b494eSTejun Heo * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 261687a9aa5STejun Heo * and pwq_release_workfn() for details. pool_workqueue itself is also 262687a9aa5STejun Heo * RCU protected so that the first pwq can be determined without 263967b494eSTejun Heo * grabbing wq->mutex. 2648864b4e5STejun Heo */ 265687a9aa5STejun Heo struct kthread_work release_work; 2668864b4e5STejun Heo struct rcu_head rcu; 267e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2681da177e4SLinus Torvalds 2691da177e4SLinus Torvalds /* 27073f53c4aSTejun Heo * Structure used to wait for workqueue flush. 27173f53c4aSTejun Heo */ 27273f53c4aSTejun Heo struct wq_flusher { 2733c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2743c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 27573f53c4aSTejun Heo struct completion done; /* flush completion */ 27673f53c4aSTejun Heo }; 2771da177e4SLinus Torvalds 278226223abSTejun Heo struct wq_device; 279226223abSTejun Heo 28073f53c4aSTejun Heo /* 281c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 282c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2831da177e4SLinus Torvalds */ 2841da177e4SLinus Torvalds struct workqueue_struct { 2853c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 286e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 28773f53c4aSTejun Heo 2883c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2893c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2903c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 291112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2923c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2933c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2943c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 29573f53c4aSTejun Heo 2962e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 29730ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 298e22bee78STejun Heo 29987fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 300a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 301226223abSTejun Heo 3025b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 3035b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 3046029a918STejun Heo 305226223abSTejun Heo #ifdef CONFIG_SYSFS 306226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 307226223abSTejun Heo #endif 3084e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 309669de8bdSBart Van Assche char *lock_name; 310669de8bdSBart Van Assche struct lock_class_key key; 3114e6045f1SJohannes Berg struct lockdep_map lockdep_map; 3124e6045f1SJohannes Berg #endif 313ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 3142728fd2fSTejun Heo 315e2dca7adSTejun Heo /* 31624acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 31724acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 318e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 319e2dca7adSTejun Heo */ 320e2dca7adSTejun Heo struct rcu_head rcu; 321e2dca7adSTejun Heo 3222728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 3232728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 324636b927eSTejun Heo struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ 3251da177e4SLinus Torvalds }; 3261da177e4SLinus Torvalds 327e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 328e904e6c2STejun Heo 32984193c07STejun Heo /* 33084193c07STejun Heo * Each pod type describes how CPUs should be grouped for unbound workqueues. 33184193c07STejun Heo * See the comment above workqueue_attrs->affn_scope. 33284193c07STejun Heo */ 33384193c07STejun Heo struct wq_pod_type { 33484193c07STejun Heo int nr_pods; /* number of pods */ 33584193c07STejun Heo cpumask_var_t *pod_cpus; /* pod -> cpus */ 33684193c07STejun Heo int *pod_node; /* pod -> node */ 33784193c07STejun Heo int *cpu_pod; /* cpu -> pod */ 33884193c07STejun Heo }; 33984193c07STejun Heo 34084193c07STejun Heo static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 34163c5484eSTejun Heo static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_DFL; 34263c5484eSTejun Heo 34363c5484eSTejun Heo static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { 34463c5484eSTejun Heo [WQ_AFFN_CPU] = "cpu", 34563c5484eSTejun Heo [WQ_AFFN_SMT] = "smt", 34663c5484eSTejun Heo [WQ_AFFN_CACHE] = "cache", 34763c5484eSTejun Heo [WQ_AFFN_NUMA] = "numa", 34863c5484eSTejun Heo [WQ_AFFN_SYSTEM] = "system", 34963c5484eSTejun Heo }; 350bce90380STejun Heo 351616db877STejun Heo /* 352616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 353616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 354616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 355aa6fde93STejun Heo * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 356aa6fde93STejun Heo * The actual value is initialized in wq_cpu_intensive_thresh_init(). 357616db877STejun Heo */ 358aa6fde93STejun Heo static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 359616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 360616db877STejun Heo 361cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 362552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 363cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 364cee22a15SViresh Kumar 365863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3663347fa09STejun Heo 367fef59c9cSTejun Heo /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 368fef59c9cSTejun Heo static struct workqueue_attrs *wq_update_pod_attrs_buf; 3694c16bd32STejun Heo 37068e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3711258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 372a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 373d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 374d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3755bcab335STejun Heo 376e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 37768e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3787d19c5ceSTejun Heo 37999c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 380ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 381ef557180SMike Galbraith 382ace3c549Stiozhang /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 383ace3c549Stiozhang static struct cpumask wq_cmdline_cpumask __initdata; 384ace3c549Stiozhang 385ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 386ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 387b05a7928SFrederic Weisbecker 388f303fccbSTejun Heo /* 389f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 390f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 391f303fccbSTejun Heo * to uncover usages which depend on it. 392f303fccbSTejun Heo */ 393f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 394f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 395f303fccbSTejun Heo #else 396f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 397f303fccbSTejun Heo #endif 398f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 399f303fccbSTejun Heo 4007d19c5ceSTejun Heo /* the per-cpu worker pools */ 40125528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 4027d19c5ceSTejun Heo 40368e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 4047d19c5ceSTejun Heo 40568e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 40629c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 40729c91e99STejun Heo 408c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 40929c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 41029c91e99STejun Heo 4118a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 4128a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 4138a2b7538STejun Heo 414967b494eSTejun Heo /* 415967b494eSTejun Heo * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 416967b494eSTejun Heo * process context while holding a pool lock. Bounce to a dedicated kthread 417967b494eSTejun Heo * worker to avoid A-A deadlocks. 418967b494eSTejun Heo */ 419967b494eSTejun Heo static struct kthread_worker *pwq_release_worker; 420967b494eSTejun Heo 421d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 422ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 423044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4241aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 425044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 426d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 427044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 428f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 429044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 43024d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4310668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4320668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4330668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4340668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 435d320c038STejun Heo 4367d19c5ceSTejun Heo static int worker_thread(void *__worker); 4376ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 438c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 43955df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4407d19c5ceSTejun Heo 44197bd2347STejun Heo #define CREATE_TRACE_POINTS 44297bd2347STejun Heo #include <trace/events/workqueue.h> 44397bd2347STejun Heo 44468e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 44524acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 446f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 44724acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4485bcab335STejun Heo 4495b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 45024acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 451f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 452f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 45324acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4545b95e1afSLai Jiangshan 455f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 456f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 457f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4587a62c2c8STejun Heo (pool)++) 4594ce62e9eSTejun Heo 46049e3cf44STejun Heo /** 46117116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 46217116969STejun Heo * @pool: iteration cursor 463611c92a0STejun Heo * @pi: integer used for iteration 464fa1b54e6STejun Heo * 46524acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 46668e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 46768e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 468fa1b54e6STejun Heo * 469fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 470fa1b54e6STejun Heo * ignored. 47117116969STejun Heo */ 472611c92a0STejun Heo #define for_each_pool(pool, pi) \ 473611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 47468e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 475fa1b54e6STejun Heo else 47617116969STejun Heo 47717116969STejun Heo /** 478822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 479822d8405STejun Heo * @worker: iteration cursor 480822d8405STejun Heo * @pool: worker_pool to iterate workers of 481822d8405STejun Heo * 4821258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 483822d8405STejun Heo * 484822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 485822d8405STejun Heo * ignored. 486822d8405STejun Heo */ 487da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 488da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4891258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 490822d8405STejun Heo else 491822d8405STejun Heo 492822d8405STejun Heo /** 49349e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 49449e3cf44STejun Heo * @pwq: iteration cursor 49549e3cf44STejun Heo * @wq: the target workqueue 49676af4d93STejun Heo * 49724acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 498794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 499794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 50076af4d93STejun Heo * 50176af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 50276af4d93STejun Heo * ignored. 50349e3cf44STejun Heo */ 50449e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 50549e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 5065a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 507f3421797STejun Heo 508dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 509dc186ad7SThomas Gleixner 510f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 511dc186ad7SThomas Gleixner 51299777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 51399777288SStanislaw Gruszka { 51499777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 51599777288SStanislaw Gruszka } 51699777288SStanislaw Gruszka 517b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 518b9fdac7fSDu, Changbin { 519b9fdac7fSDu, Changbin struct work_struct *work = addr; 520b9fdac7fSDu, Changbin 521b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 522b9fdac7fSDu, Changbin } 523b9fdac7fSDu, Changbin 524dc186ad7SThomas Gleixner /* 525dc186ad7SThomas Gleixner * fixup_init is called when: 526dc186ad7SThomas Gleixner * - an active object is initialized 527dc186ad7SThomas Gleixner */ 52802a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 529dc186ad7SThomas Gleixner { 530dc186ad7SThomas Gleixner struct work_struct *work = addr; 531dc186ad7SThomas Gleixner 532dc186ad7SThomas Gleixner switch (state) { 533dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 534dc186ad7SThomas Gleixner cancel_work_sync(work); 535dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 53602a982a6SDu, Changbin return true; 537dc186ad7SThomas Gleixner default: 53802a982a6SDu, Changbin return false; 539dc186ad7SThomas Gleixner } 540dc186ad7SThomas Gleixner } 541dc186ad7SThomas Gleixner 542dc186ad7SThomas Gleixner /* 543dc186ad7SThomas Gleixner * fixup_free is called when: 544dc186ad7SThomas Gleixner * - an active object is freed 545dc186ad7SThomas Gleixner */ 54602a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 547dc186ad7SThomas Gleixner { 548dc186ad7SThomas Gleixner struct work_struct *work = addr; 549dc186ad7SThomas Gleixner 550dc186ad7SThomas Gleixner switch (state) { 551dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 552dc186ad7SThomas Gleixner cancel_work_sync(work); 553dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 55402a982a6SDu, Changbin return true; 555dc186ad7SThomas Gleixner default: 55602a982a6SDu, Changbin return false; 557dc186ad7SThomas Gleixner } 558dc186ad7SThomas Gleixner } 559dc186ad7SThomas Gleixner 560f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 561dc186ad7SThomas Gleixner .name = "work_struct", 56299777288SStanislaw Gruszka .debug_hint = work_debug_hint, 563b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 564dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 565dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 566dc186ad7SThomas Gleixner }; 567dc186ad7SThomas Gleixner 568dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 569dc186ad7SThomas Gleixner { 570dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 571dc186ad7SThomas Gleixner } 572dc186ad7SThomas Gleixner 573dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 574dc186ad7SThomas Gleixner { 575dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 576dc186ad7SThomas Gleixner } 577dc186ad7SThomas Gleixner 578dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 579dc186ad7SThomas Gleixner { 580dc186ad7SThomas Gleixner if (onstack) 581dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 582dc186ad7SThomas Gleixner else 583dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 584dc186ad7SThomas Gleixner } 585dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 586dc186ad7SThomas Gleixner 587dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 588dc186ad7SThomas Gleixner { 589dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 590dc186ad7SThomas Gleixner } 591dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 592dc186ad7SThomas Gleixner 593ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 594ea2e64f2SThomas Gleixner { 595ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 596ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 597ea2e64f2SThomas Gleixner } 598ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 599ea2e64f2SThomas Gleixner 600dc186ad7SThomas Gleixner #else 601dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 602dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 603dc186ad7SThomas Gleixner #endif 604dc186ad7SThomas Gleixner 6054e8b22bdSLi Bin /** 60667dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 6074e8b22bdSLi Bin * @pool: the pool pointer of interest 6084e8b22bdSLi Bin * 6094e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 6104e8b22bdSLi Bin * successfully, -errno on failure. 6114e8b22bdSLi Bin */ 6129daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 6139daf9e67STejun Heo { 6149daf9e67STejun Heo int ret; 6159daf9e67STejun Heo 61668e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 6175bcab335STejun Heo 6184e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 6194e8b22bdSLi Bin GFP_KERNEL); 620229641a6STejun Heo if (ret >= 0) { 621e68035fbSTejun Heo pool->id = ret; 622229641a6STejun Heo return 0; 623229641a6STejun Heo } 6249daf9e67STejun Heo return ret; 6259daf9e67STejun Heo } 6269daf9e67STejun Heo 62773f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 62873f53c4aSTejun Heo { 62973f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 63073f53c4aSTejun Heo } 63173f53c4aSTejun Heo 632c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 63373f53c4aSTejun Heo { 634c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 63573f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 63673f53c4aSTejun Heo } 63773f53c4aSTejun Heo 63873f53c4aSTejun Heo static int work_next_color(int color) 63973f53c4aSTejun Heo { 64073f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 641a848e3b6SOleg Nesterov } 642a848e3b6SOleg Nesterov 6434594bf15SDavid Howells /* 644112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 645112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6467c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6477a22ad75STejun Heo * 648112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 649112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 650bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 651bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6527a22ad75STejun Heo * 653112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6547c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 655112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6567c3eed5cSTejun Heo * available only while the work item is queued. 657bbb68dfaSTejun Heo * 658bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 659bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 660bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 661bbb68dfaSTejun Heo * try to steal the PENDING bit. 6624594bf15SDavid Howells */ 6637a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6647a22ad75STejun Heo unsigned long flags) 6657a22ad75STejun Heo { 6666183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6677a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6687a22ad75STejun Heo } 6697a22ad75STejun Heo 670112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6714690c4abSTejun Heo unsigned long extra_flags) 672365970a1SDavid Howells { 673112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 674112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 675365970a1SDavid Howells } 676365970a1SDavid Howells 6774468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6784468a00fSLai Jiangshan int pool_id) 6794468a00fSLai Jiangshan { 6804468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6814468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6824468a00fSLai Jiangshan } 6834468a00fSLai Jiangshan 6847c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6857c3eed5cSTejun Heo int pool_id) 6864d707b9fSOleg Nesterov { 68723657bb1STejun Heo /* 68823657bb1STejun Heo * The following wmb is paired with the implied mb in 68923657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 69023657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 69123657bb1STejun Heo * owner. 69223657bb1STejun Heo */ 69323657bb1STejun Heo smp_wmb(); 6947c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 695346c09f8SRoman Pen /* 696346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 697346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 698346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 6998bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 700346c09f8SRoman Pen * the same @work. E.g. consider this case: 701346c09f8SRoman Pen * 702346c09f8SRoman Pen * CPU#0 CPU#1 703346c09f8SRoman Pen * ---------------------------- -------------------------------- 704346c09f8SRoman Pen * 705346c09f8SRoman Pen * 1 STORE event_indicated 706346c09f8SRoman Pen * 2 queue_work_on() { 707346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 708346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 709346c09f8SRoman Pen * 5 set_work_data() # clear bit 710346c09f8SRoman Pen * 6 smp_mb() 711346c09f8SRoman Pen * 7 work->current_func() { 712346c09f8SRoman Pen * 8 LOAD event_indicated 713346c09f8SRoman Pen * } 714346c09f8SRoman Pen * 715346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 716346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 717346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 718346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 719346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 720346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 721346c09f8SRoman Pen * before actual STORE. 722346c09f8SRoman Pen */ 723346c09f8SRoman Pen smp_mb(); 7244d707b9fSOleg Nesterov } 7254d707b9fSOleg Nesterov 7267a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 727365970a1SDavid Howells { 7287c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7297c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7307a22ad75STejun Heo } 7317a22ad75STejun Heo 732afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 733afa4bb77SLinus Torvalds { 734afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 735afa4bb77SLinus Torvalds } 736afa4bb77SLinus Torvalds 737112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7387a22ad75STejun Heo { 739e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7407a22ad75STejun Heo 741112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 742afa4bb77SLinus Torvalds return work_struct_pwq(data); 743e120153dSTejun Heo else 744e120153dSTejun Heo return NULL; 7457a22ad75STejun Heo } 7467a22ad75STejun Heo 7477c3eed5cSTejun Heo /** 7487c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7497c3eed5cSTejun Heo * @work: the work item of interest 7507c3eed5cSTejun Heo * 75168e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 75224acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 75324acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 754fa1b54e6STejun Heo * 755fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 756fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 757fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 758fa1b54e6STejun Heo * returned pool is and stays online. 759d185af30SYacine Belkadi * 760d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7617c3eed5cSTejun Heo */ 7627c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7637a22ad75STejun Heo { 764e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7657c3eed5cSTejun Heo int pool_id; 7667a22ad75STejun Heo 76768e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 768fa1b54e6STejun Heo 769112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 770afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7717a22ad75STejun Heo 7727c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7737c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7747a22ad75STejun Heo return NULL; 7757a22ad75STejun Heo 776fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7777c3eed5cSTejun Heo } 7787c3eed5cSTejun Heo 7797c3eed5cSTejun Heo /** 7807c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7817c3eed5cSTejun Heo * @work: the work item of interest 7827c3eed5cSTejun Heo * 783d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7847c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7857c3eed5cSTejun Heo */ 7867c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7877c3eed5cSTejun Heo { 78854d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7897c3eed5cSTejun Heo 790112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 791afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 79254d5b7d0SLai Jiangshan 79354d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7947c3eed5cSTejun Heo } 7957c3eed5cSTejun Heo 796bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 797bbb68dfaSTejun Heo { 7987c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 799bbb68dfaSTejun Heo 8007c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8017c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 802bbb68dfaSTejun Heo } 803bbb68dfaSTejun Heo 804bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 805bbb68dfaSTejun Heo { 806bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 807bbb68dfaSTejun Heo 808112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 809bbb68dfaSTejun Heo } 810bbb68dfaSTejun Heo 811e22bee78STejun Heo /* 8123270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8133270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 814d565ed63STejun Heo * they're being called with pool->lock held. 815e22bee78STejun Heo */ 816e22bee78STejun Heo 817e22bee78STejun Heo /* 818e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 819e22bee78STejun Heo * running workers. 820974271c4STejun Heo * 821974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 822706026c2STejun Heo * function will always return %true for unbound pools as long as the 823974271c4STejun Heo * worklist isn't empty. 824e22bee78STejun Heo */ 82563d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 826e22bee78STejun Heo { 8270219a352STejun Heo return !list_empty(&pool->worklist) && !pool->nr_running; 828e22bee78STejun Heo } 829e22bee78STejun Heo 830e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 83163d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 832e22bee78STejun Heo { 83363d95a91STejun Heo return pool->nr_idle; 834e22bee78STejun Heo } 835e22bee78STejun Heo 836e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 83763d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 838e22bee78STejun Heo { 839bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 840e22bee78STejun Heo } 841e22bee78STejun Heo 842e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 84363d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 844e22bee78STejun Heo { 84563d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 846e22bee78STejun Heo } 847e22bee78STejun Heo 848e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 84963d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 850e22bee78STejun Heo { 851692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 85263d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 85363d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 854e22bee78STejun Heo 855e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 856e22bee78STejun Heo } 857e22bee78STejun Heo 8584690c4abSTejun Heo /** 859e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 860cb444766STejun Heo * @worker: self 861d302f017STejun Heo * @flags: flags to set 862d302f017STejun Heo * 863228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 864d302f017STejun Heo */ 865228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 866d302f017STejun Heo { 867bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 868e22bee78STejun Heo 869bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 870cb444766STejun Heo 871228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 872e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 873e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 874bc35f7efSLai Jiangshan pool->nr_running--; 875e22bee78STejun Heo } 876e22bee78STejun Heo 877d302f017STejun Heo worker->flags |= flags; 878d302f017STejun Heo } 879d302f017STejun Heo 880d302f017STejun Heo /** 881e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 882cb444766STejun Heo * @worker: self 883d302f017STejun Heo * @flags: flags to clear 884d302f017STejun Heo * 885e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 886d302f017STejun Heo */ 887d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 888d302f017STejun Heo { 88963d95a91STejun Heo struct worker_pool *pool = worker->pool; 890e22bee78STejun Heo unsigned int oflags = worker->flags; 891e22bee78STejun Heo 892bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 893cb444766STejun Heo 894d302f017STejun Heo worker->flags &= ~flags; 895e22bee78STejun Heo 89642c025f3STejun Heo /* 89742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 89842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 89942c025f3STejun Heo * of multiple flags, not a single flag. 90042c025f3STejun Heo */ 901e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 902e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 903bc35f7efSLai Jiangshan pool->nr_running++; 904d302f017STejun Heo } 905d302f017STejun Heo 906797e8345STejun Heo /* Return the first idle worker. Called with pool->lock held. */ 907797e8345STejun Heo static struct worker *first_idle_worker(struct worker_pool *pool) 908797e8345STejun Heo { 909797e8345STejun Heo if (unlikely(list_empty(&pool->idle_list))) 910797e8345STejun Heo return NULL; 911797e8345STejun Heo 912797e8345STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 913797e8345STejun Heo } 914797e8345STejun Heo 915797e8345STejun Heo /** 916797e8345STejun Heo * worker_enter_idle - enter idle state 917797e8345STejun Heo * @worker: worker which is entering idle state 918797e8345STejun Heo * 919797e8345STejun Heo * @worker is entering idle state. Update stats and idle timer if 920797e8345STejun Heo * necessary. 921797e8345STejun Heo * 922797e8345STejun Heo * LOCKING: 923797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 924797e8345STejun Heo */ 925797e8345STejun Heo static void worker_enter_idle(struct worker *worker) 926797e8345STejun Heo { 927797e8345STejun Heo struct worker_pool *pool = worker->pool; 928797e8345STejun Heo 929797e8345STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 930797e8345STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 931797e8345STejun Heo (worker->hentry.next || worker->hentry.pprev))) 932797e8345STejun Heo return; 933797e8345STejun Heo 934797e8345STejun Heo /* can't use worker_set_flags(), also called from create_worker() */ 935797e8345STejun Heo worker->flags |= WORKER_IDLE; 936797e8345STejun Heo pool->nr_idle++; 937797e8345STejun Heo worker->last_active = jiffies; 938797e8345STejun Heo 939797e8345STejun Heo /* idle_list is LIFO */ 940797e8345STejun Heo list_add(&worker->entry, &pool->idle_list); 941797e8345STejun Heo 942797e8345STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 943797e8345STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 944797e8345STejun Heo 945797e8345STejun Heo /* Sanity check nr_running. */ 946797e8345STejun Heo WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 947797e8345STejun Heo } 948797e8345STejun Heo 949797e8345STejun Heo /** 950797e8345STejun Heo * worker_leave_idle - leave idle state 951797e8345STejun Heo * @worker: worker which is leaving idle state 952797e8345STejun Heo * 953797e8345STejun Heo * @worker is leaving idle state. Update stats. 954797e8345STejun Heo * 955797e8345STejun Heo * LOCKING: 956797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 957797e8345STejun Heo */ 958797e8345STejun Heo static void worker_leave_idle(struct worker *worker) 959797e8345STejun Heo { 960797e8345STejun Heo struct worker_pool *pool = worker->pool; 961797e8345STejun Heo 962797e8345STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 963797e8345STejun Heo return; 964797e8345STejun Heo worker_clr_flags(worker, WORKER_IDLE); 965797e8345STejun Heo pool->nr_idle--; 966797e8345STejun Heo list_del_init(&worker->entry); 967797e8345STejun Heo } 968797e8345STejun Heo 969797e8345STejun Heo /** 970797e8345STejun Heo * find_worker_executing_work - find worker which is executing a work 971797e8345STejun Heo * @pool: pool of interest 972797e8345STejun Heo * @work: work to find worker for 973797e8345STejun Heo * 974797e8345STejun Heo * Find a worker which is executing @work on @pool by searching 975797e8345STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 976797e8345STejun Heo * to match, its current execution should match the address of @work and 977797e8345STejun Heo * its work function. This is to avoid unwanted dependency between 978797e8345STejun Heo * unrelated work executions through a work item being recycled while still 979797e8345STejun Heo * being executed. 980797e8345STejun Heo * 981797e8345STejun Heo * This is a bit tricky. A work item may be freed once its execution 982797e8345STejun Heo * starts and nothing prevents the freed area from being recycled for 983797e8345STejun Heo * another work item. If the same work item address ends up being reused 984797e8345STejun Heo * before the original execution finishes, workqueue will identify the 985797e8345STejun Heo * recycled work item as currently executing and make it wait until the 986797e8345STejun Heo * current execution finishes, introducing an unwanted dependency. 987797e8345STejun Heo * 988797e8345STejun Heo * This function checks the work item address and work function to avoid 989797e8345STejun Heo * false positives. Note that this isn't complete as one may construct a 990797e8345STejun Heo * work function which can introduce dependency onto itself through a 991797e8345STejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 992797e8345STejun Heo * foot that badly, there's only so much we can do, and if such deadlock 993797e8345STejun Heo * actually occurs, it should be easy to locate the culprit work function. 994797e8345STejun Heo * 995797e8345STejun Heo * CONTEXT: 996797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 997797e8345STejun Heo * 998797e8345STejun Heo * Return: 999797e8345STejun Heo * Pointer to worker which is executing @work if found, %NULL 1000797e8345STejun Heo * otherwise. 1001797e8345STejun Heo */ 1002797e8345STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 1003797e8345STejun Heo struct work_struct *work) 1004797e8345STejun Heo { 1005797e8345STejun Heo struct worker *worker; 1006797e8345STejun Heo 1007797e8345STejun Heo hash_for_each_possible(pool->busy_hash, worker, hentry, 1008797e8345STejun Heo (unsigned long)work) 1009797e8345STejun Heo if (worker->current_work == work && 1010797e8345STejun Heo worker->current_func == work->func) 1011797e8345STejun Heo return worker; 1012797e8345STejun Heo 1013797e8345STejun Heo return NULL; 1014797e8345STejun Heo } 1015797e8345STejun Heo 1016797e8345STejun Heo /** 1017797e8345STejun Heo * move_linked_works - move linked works to a list 1018797e8345STejun Heo * @work: start of series of works to be scheduled 1019797e8345STejun Heo * @head: target list to append @work to 1020797e8345STejun Heo * @nextp: out parameter for nested worklist walking 1021797e8345STejun Heo * 1022873eaca6STejun Heo * Schedule linked works starting from @work to @head. Work series to be 1023873eaca6STejun Heo * scheduled starts at @work and includes any consecutive work with 1024873eaca6STejun Heo * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1025873eaca6STejun Heo * @nextp. 1026797e8345STejun Heo * 1027797e8345STejun Heo * CONTEXT: 1028797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1029797e8345STejun Heo */ 1030797e8345STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1031797e8345STejun Heo struct work_struct **nextp) 1032797e8345STejun Heo { 1033797e8345STejun Heo struct work_struct *n; 1034797e8345STejun Heo 1035797e8345STejun Heo /* 1036797e8345STejun Heo * Linked worklist will always end before the end of the list, 1037797e8345STejun Heo * use NULL for list head. 1038797e8345STejun Heo */ 1039797e8345STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1040797e8345STejun Heo list_move_tail(&work->entry, head); 1041797e8345STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1042797e8345STejun Heo break; 1043797e8345STejun Heo } 1044797e8345STejun Heo 1045797e8345STejun Heo /* 1046797e8345STejun Heo * If we're already inside safe list traversal and have moved 1047797e8345STejun Heo * multiple works to the scheduled queue, the next position 1048797e8345STejun Heo * needs to be updated. 1049797e8345STejun Heo */ 1050797e8345STejun Heo if (nextp) 1051797e8345STejun Heo *nextp = n; 1052797e8345STejun Heo } 1053797e8345STejun Heo 1054797e8345STejun Heo /** 1055873eaca6STejun Heo * assign_work - assign a work item and its linked work items to a worker 1056873eaca6STejun Heo * @work: work to assign 1057873eaca6STejun Heo * @worker: worker to assign to 1058873eaca6STejun Heo * @nextp: out parameter for nested worklist walking 1059873eaca6STejun Heo * 1060873eaca6STejun Heo * Assign @work and its linked work items to @worker. If @work is already being 1061873eaca6STejun Heo * executed by another worker in the same pool, it'll be punted there. 1062873eaca6STejun Heo * 1063873eaca6STejun Heo * If @nextp is not NULL, it's updated to point to the next work of the last 1064873eaca6STejun Heo * scheduled work. This allows assign_work() to be nested inside 1065873eaca6STejun Heo * list_for_each_entry_safe(). 1066873eaca6STejun Heo * 1067873eaca6STejun Heo * Returns %true if @work was successfully assigned to @worker. %false if @work 1068873eaca6STejun Heo * was punted to another worker already executing it. 1069873eaca6STejun Heo */ 1070873eaca6STejun Heo static bool assign_work(struct work_struct *work, struct worker *worker, 1071873eaca6STejun Heo struct work_struct **nextp) 1072873eaca6STejun Heo { 1073873eaca6STejun Heo struct worker_pool *pool = worker->pool; 1074873eaca6STejun Heo struct worker *collision; 1075873eaca6STejun Heo 1076873eaca6STejun Heo lockdep_assert_held(&pool->lock); 1077873eaca6STejun Heo 1078873eaca6STejun Heo /* 1079873eaca6STejun Heo * A single work shouldn't be executed concurrently by multiple workers. 1080873eaca6STejun Heo * __queue_work() ensures that @work doesn't jump to a different pool 1081873eaca6STejun Heo * while still running in the previous pool. Here, we should ensure that 1082873eaca6STejun Heo * @work is not executed concurrently by multiple workers from the same 1083873eaca6STejun Heo * pool. Check whether anyone is already processing the work. If so, 1084873eaca6STejun Heo * defer the work to the currently executing one. 1085873eaca6STejun Heo */ 1086873eaca6STejun Heo collision = find_worker_executing_work(pool, work); 1087873eaca6STejun Heo if (unlikely(collision)) { 1088873eaca6STejun Heo move_linked_works(work, &collision->scheduled, nextp); 1089873eaca6STejun Heo return false; 1090873eaca6STejun Heo } 1091873eaca6STejun Heo 1092873eaca6STejun Heo move_linked_works(work, &worker->scheduled, nextp); 1093873eaca6STejun Heo return true; 1094873eaca6STejun Heo } 1095873eaca6STejun Heo 1096873eaca6STejun Heo /** 10970219a352STejun Heo * kick_pool - wake up an idle worker if necessary 10980219a352STejun Heo * @pool: pool to kick 1099797e8345STejun Heo * 11000219a352STejun Heo * @pool may have pending work items. Wake up worker if necessary. Returns 11010219a352STejun Heo * whether a worker was woken up. 1102797e8345STejun Heo */ 11030219a352STejun Heo static bool kick_pool(struct worker_pool *pool) 1104797e8345STejun Heo { 1105797e8345STejun Heo struct worker *worker = first_idle_worker(pool); 1106797e8345STejun Heo 11070219a352STejun Heo lockdep_assert_held(&pool->lock); 11080219a352STejun Heo 11090219a352STejun Heo if (!need_more_worker(pool) || !worker) 11100219a352STejun Heo return false; 11110219a352STejun Heo 1112797e8345STejun Heo wake_up_process(worker->task); 11130219a352STejun Heo return true; 1114797e8345STejun Heo } 1115797e8345STejun Heo 111663638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 111763638450STejun Heo 111863638450STejun Heo /* 111963638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 112063638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 112163638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 112263638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 112363638450STejun Heo * should be using an unbound workqueue instead. 112463638450STejun Heo * 112563638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 112663638450STejun Heo * and report them so that they can be examined and converted to use unbound 112763638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 112863638450STejun Heo * function is tracked and reported with exponential backoff. 112963638450STejun Heo */ 113063638450STejun Heo #define WCI_MAX_ENTS 128 113163638450STejun Heo 113263638450STejun Heo struct wci_ent { 113363638450STejun Heo work_func_t func; 113463638450STejun Heo atomic64_t cnt; 113563638450STejun Heo struct hlist_node hash_node; 113663638450STejun Heo }; 113763638450STejun Heo 113863638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 113963638450STejun Heo static int wci_nr_ents; 114063638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 114163638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 114263638450STejun Heo 114363638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 114463638450STejun Heo { 114563638450STejun Heo struct wci_ent *ent; 114663638450STejun Heo 114763638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 114863638450STejun Heo (unsigned long)func) { 114963638450STejun Heo if (ent->func == func) 115063638450STejun Heo return ent; 115163638450STejun Heo } 115263638450STejun Heo return NULL; 115363638450STejun Heo } 115463638450STejun Heo 115563638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 115663638450STejun Heo { 115763638450STejun Heo struct wci_ent *ent; 115863638450STejun Heo 115963638450STejun Heo restart: 116063638450STejun Heo ent = wci_find_ent(func); 116163638450STejun Heo if (ent) { 116263638450STejun Heo u64 cnt; 116363638450STejun Heo 116463638450STejun Heo /* 116563638450STejun Heo * Start reporting from the fourth time and back off 116663638450STejun Heo * exponentially. 116763638450STejun Heo */ 116863638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 116963638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 117063638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 117163638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 117263638450STejun Heo atomic64_read(&ent->cnt)); 117363638450STejun Heo return; 117463638450STejun Heo } 117563638450STejun Heo 117663638450STejun Heo /* 117763638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 117863638450STejun Heo * is exhausted, something went really wrong and we probably made enough 117963638450STejun Heo * noise already. 118063638450STejun Heo */ 118163638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 118263638450STejun Heo return; 118363638450STejun Heo 118463638450STejun Heo raw_spin_lock(&wci_lock); 118563638450STejun Heo 118663638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 118763638450STejun Heo raw_spin_unlock(&wci_lock); 118863638450STejun Heo return; 118963638450STejun Heo } 119063638450STejun Heo 119163638450STejun Heo if (wci_find_ent(func)) { 119263638450STejun Heo raw_spin_unlock(&wci_lock); 119363638450STejun Heo goto restart; 119463638450STejun Heo } 119563638450STejun Heo 119663638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 119763638450STejun Heo ent->func = func; 119863638450STejun Heo atomic64_set(&ent->cnt, 1); 119963638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 120063638450STejun Heo 120163638450STejun Heo raw_spin_unlock(&wci_lock); 120263638450STejun Heo } 120363638450STejun Heo 120463638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 120563638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 120663638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 120763638450STejun Heo 1208c54d5046STejun Heo /** 12091da177e4SLinus Torvalds * wq_worker_running - a worker is running again 12101da177e4SLinus Torvalds * @task: task waking up 12111da177e4SLinus Torvalds * 12121da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 12131da177e4SLinus Torvalds */ 12141da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 12151da177e4SLinus Torvalds { 12161da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 12171da177e4SLinus Torvalds 1218c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 12191da177e4SLinus Torvalds return; 12201da177e4SLinus Torvalds 12211da177e4SLinus Torvalds /* 12221da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 12231da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 12241da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 12251da177e4SLinus Torvalds * pool. Protect against such race. 12261da177e4SLinus Torvalds */ 12271da177e4SLinus Torvalds preempt_disable(); 12281da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 12291da177e4SLinus Torvalds worker->pool->nr_running++; 12301da177e4SLinus Torvalds preempt_enable(); 1231616db877STejun Heo 1232616db877STejun Heo /* 1233616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1234616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1235616db877STejun Heo */ 1236616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1237616db877STejun Heo 1238c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 12391da177e4SLinus Torvalds } 12401da177e4SLinus Torvalds 12411da177e4SLinus Torvalds /** 12421da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 12431da177e4SLinus Torvalds * @task: task going to sleep 12441da177e4SLinus Torvalds * 12451da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 12461da177e4SLinus Torvalds * going to sleep. 12471da177e4SLinus Torvalds */ 12481da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 12491da177e4SLinus Torvalds { 12501da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 12511da177e4SLinus Torvalds struct worker_pool *pool; 12521da177e4SLinus Torvalds 12531da177e4SLinus Torvalds /* 12541da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 12551da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 12561da177e4SLinus Torvalds * checking NOT_RUNNING. 12571da177e4SLinus Torvalds */ 12581da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 12591da177e4SLinus Torvalds return; 12601da177e4SLinus Torvalds 12611da177e4SLinus Torvalds pool = worker->pool; 12621da177e4SLinus Torvalds 12631da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1264c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 12651da177e4SLinus Torvalds return; 12661da177e4SLinus Torvalds 1267c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 12681da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 12691da177e4SLinus Torvalds 12701da177e4SLinus Torvalds /* 12711da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 12721da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 12731da177e4SLinus Torvalds * and nr_running has been reset. 12741da177e4SLinus Torvalds */ 12751da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 12761da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12771da177e4SLinus Torvalds return; 12781da177e4SLinus Torvalds } 12791da177e4SLinus Torvalds 12801da177e4SLinus Torvalds pool->nr_running--; 12810219a352STejun Heo if (kick_pool(pool)) 1282725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 12830219a352STejun Heo 12841da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 12851da177e4SLinus Torvalds } 12861da177e4SLinus Torvalds 12871da177e4SLinus Torvalds /** 1288616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1289616db877STejun Heo * @task: task currently running 1290616db877STejun Heo * 1291616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1292616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1293616db877STejun Heo */ 1294616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1295616db877STejun Heo { 1296616db877STejun Heo struct worker *worker = kthread_data(task); 1297616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1298616db877STejun Heo struct worker_pool *pool = worker->pool; 1299616db877STejun Heo 1300616db877STejun Heo if (!pwq) 1301616db877STejun Heo return; 1302616db877STejun Heo 13038a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 13048a1dd1e5STejun Heo 130518c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 130618c8ae81SZqiang return; 130718c8ae81SZqiang 1308616db877STejun Heo /* 1309616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1310616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1311616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1312c8f6219bSZqiang * 1313c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1314c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1315c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1316c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1317c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1318c8f6219bSZqiang * We probably want to make this prettier in the future. 1319616db877STejun Heo */ 1320c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1321616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1322616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1323616db877STejun Heo return; 1324616db877STejun Heo 1325616db877STejun Heo raw_spin_lock(&pool->lock); 1326616db877STejun Heo 1327616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 132863638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1329616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1330616db877STejun Heo 13310219a352STejun Heo if (kick_pool(pool)) 1332616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1333616db877STejun Heo 1334616db877STejun Heo raw_spin_unlock(&pool->lock); 1335616db877STejun Heo } 1336616db877STejun Heo 1337616db877STejun Heo /** 13381da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 13391da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 13401da177e4SLinus Torvalds * 13411da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 13421da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 13431da177e4SLinus Torvalds * 13441da177e4SLinus Torvalds * CONTEXT: 13459b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 13461da177e4SLinus Torvalds * 13471da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1348f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1349f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 13501da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 13511da177e4SLinus Torvalds * 13521da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 13531da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 13541da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 13551da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1356365970a1SDavid Howells * 1357365970a1SDavid Howells * Return: 1358365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1359365970a1SDavid Howells * hasn't executed any work yet. 1360365970a1SDavid Howells */ 1361365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1362365970a1SDavid Howells { 1363365970a1SDavid Howells struct worker *worker = kthread_data(task); 1364365970a1SDavid Howells 1365365970a1SDavid Howells return worker->last_func; 1366365970a1SDavid Howells } 1367365970a1SDavid Howells 1368d302f017STejun Heo /** 13698864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 13708864b4e5STejun Heo * @pwq: pool_workqueue to get 13718864b4e5STejun Heo * 13728864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 13738864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 13748864b4e5STejun Heo */ 13758864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 13768864b4e5STejun Heo { 13778864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13788864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 13798864b4e5STejun Heo pwq->refcnt++; 13808864b4e5STejun Heo } 13818864b4e5STejun Heo 13828864b4e5STejun Heo /** 13838864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 13848864b4e5STejun Heo * @pwq: pool_workqueue to put 13858864b4e5STejun Heo * 13868864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 13878864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 13888864b4e5STejun Heo */ 13898864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 13908864b4e5STejun Heo { 13918864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 13928864b4e5STejun Heo if (likely(--pwq->refcnt)) 13938864b4e5STejun Heo return; 13948864b4e5STejun Heo /* 1395967b494eSTejun Heo * @pwq can't be released under pool->lock, bounce to a dedicated 1396967b494eSTejun Heo * kthread_worker to avoid A-A deadlocks. 13978864b4e5STejun Heo */ 1398687a9aa5STejun Heo kthread_queue_work(pwq_release_worker, &pwq->release_work); 13998864b4e5STejun Heo } 14008864b4e5STejun Heo 1401dce90d47STejun Heo /** 1402dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1403dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1404dce90d47STejun Heo * 1405dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1406dce90d47STejun Heo */ 1407dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1408dce90d47STejun Heo { 1409dce90d47STejun Heo if (pwq) { 1410dce90d47STejun Heo /* 141124acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1412dce90d47STejun Heo * following lock operations are safe. 1413dce90d47STejun Heo */ 1414a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1415dce90d47STejun Heo put_pwq(pwq); 1416a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1417dce90d47STejun Heo } 1418dce90d47STejun Heo } 1419dce90d47STejun Heo 1420f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1421bf4ede01STejun Heo { 1422112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1423bf4ede01STejun Heo 1424bf4ede01STejun Heo trace_workqueue_activate_work(work); 142582607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 142682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1427112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1428f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1429112202d9STejun Heo pwq->nr_active++; 1430bf4ede01STejun Heo } 1431bf4ede01STejun Heo 1432f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 14333aa62497SLai Jiangshan { 1434f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 14353aa62497SLai Jiangshan struct work_struct, entry); 14363aa62497SLai Jiangshan 1437f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 14383aa62497SLai Jiangshan } 14393aa62497SLai Jiangshan 1440bf4ede01STejun Heo /** 1441112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1442112202d9STejun Heo * @pwq: pwq of interest 1443c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1444bf4ede01STejun Heo * 1445bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1446112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1447bf4ede01STejun Heo * 1448bf4ede01STejun Heo * CONTEXT: 1449a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1450bf4ede01STejun Heo */ 1451c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1452bf4ede01STejun Heo { 1453c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1454c4560c2cSLai Jiangshan 1455018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1456112202d9STejun Heo pwq->nr_active--; 1457f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1458f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1459112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1460f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1461bf4ede01STejun Heo } 1462018f3a13SLai Jiangshan } 1463018f3a13SLai Jiangshan 1464018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1465bf4ede01STejun Heo 1466bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1467112202d9STejun Heo if (likely(pwq->flush_color != color)) 14688864b4e5STejun Heo goto out_put; 1469bf4ede01STejun Heo 1470bf4ede01STejun Heo /* are there still in-flight works? */ 1471112202d9STejun Heo if (pwq->nr_in_flight[color]) 14728864b4e5STejun Heo goto out_put; 1473bf4ede01STejun Heo 1474112202d9STejun Heo /* this pwq is done, clear flush_color */ 1475112202d9STejun Heo pwq->flush_color = -1; 1476bf4ede01STejun Heo 1477bf4ede01STejun Heo /* 1478112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1479bf4ede01STejun Heo * will handle the rest. 1480bf4ede01STejun Heo */ 1481112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1482112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 14838864b4e5STejun Heo out_put: 14848864b4e5STejun Heo put_pwq(pwq); 1485bf4ede01STejun Heo } 1486bf4ede01STejun Heo 148736e227d2STejun Heo /** 1488bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 148936e227d2STejun Heo * @work: work item to steal 149036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1491bbb68dfaSTejun Heo * @flags: place to store irq state 149236e227d2STejun Heo * 149336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1494d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 149536e227d2STejun Heo * 1496d185af30SYacine Belkadi * Return: 14973eb6b31bSMauro Carvalho Chehab * 14983eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 149936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 150036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 150136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1502bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1503bbb68dfaSTejun Heo * for arbitrarily long 15043eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 150536e227d2STejun Heo * 1506d185af30SYacine Belkadi * Note: 1507bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1508e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1509e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1510e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1511bbb68dfaSTejun Heo * 1512bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1513bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1514bbb68dfaSTejun Heo * 1515e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1516bf4ede01STejun Heo */ 1517bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1518bbb68dfaSTejun Heo unsigned long *flags) 1519bf4ede01STejun Heo { 1520d565ed63STejun Heo struct worker_pool *pool; 1521112202d9STejun Heo struct pool_workqueue *pwq; 1522bf4ede01STejun Heo 1523bbb68dfaSTejun Heo local_irq_save(*flags); 1524bbb68dfaSTejun Heo 152536e227d2STejun Heo /* try to steal the timer if it exists */ 152636e227d2STejun Heo if (is_dwork) { 152736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 152836e227d2STejun Heo 1529e0aecdd8STejun Heo /* 1530e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1531e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1532e0aecdd8STejun Heo * running on the local CPU. 1533e0aecdd8STejun Heo */ 153436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 153536e227d2STejun Heo return 1; 153636e227d2STejun Heo } 153736e227d2STejun Heo 153836e227d2STejun Heo /* try to claim PENDING the normal way */ 1539bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1540bf4ede01STejun Heo return 0; 1541bf4ede01STejun Heo 154224acfb71SThomas Gleixner rcu_read_lock(); 1543bf4ede01STejun Heo /* 1544bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1545bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1546bf4ede01STejun Heo */ 1547d565ed63STejun Heo pool = get_work_pool(work); 1548d565ed63STejun Heo if (!pool) 1549bbb68dfaSTejun Heo goto fail; 1550bf4ede01STejun Heo 1551a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1552bf4ede01STejun Heo /* 1553112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1554112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1555112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1556112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1557112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 15580b3dae68SLai Jiangshan * item is currently queued on that pool. 1559bf4ede01STejun Heo */ 1560112202d9STejun Heo pwq = get_work_pwq(work); 1561112202d9STejun Heo if (pwq && pwq->pool == pool) { 1562bf4ede01STejun Heo debug_work_deactivate(work); 15633aa62497SLai Jiangshan 15643aa62497SLai Jiangshan /* 1565018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1566018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1567018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1568018f3a13SLai Jiangshan * 1569f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1570d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1571f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 157216062836STejun Heo * management later on and cause stall. Make sure the work 157316062836STejun Heo * item is activated before grabbing. 15743aa62497SLai Jiangshan */ 1575f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1576f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 15773aa62497SLai Jiangshan 1578bf4ede01STejun Heo list_del_init(&work->entry); 1579c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 158036e227d2STejun Heo 1581112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 15824468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 15834468a00fSLai Jiangshan 1584a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 158524acfb71SThomas Gleixner rcu_read_unlock(); 158636e227d2STejun Heo return 1; 1587bf4ede01STejun Heo } 1588a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1589bbb68dfaSTejun Heo fail: 159024acfb71SThomas Gleixner rcu_read_unlock(); 1591bbb68dfaSTejun Heo local_irq_restore(*flags); 1592bbb68dfaSTejun Heo if (work_is_canceling(work)) 1593bbb68dfaSTejun Heo return -ENOENT; 1594bbb68dfaSTejun Heo cpu_relax(); 159536e227d2STejun Heo return -EAGAIN; 1596bf4ede01STejun Heo } 1597bf4ede01STejun Heo 1598bf4ede01STejun Heo /** 1599706026c2STejun Heo * insert_work - insert a work into a pool 1600112202d9STejun Heo * @pwq: pwq @work belongs to 16014690c4abSTejun Heo * @work: work to insert 16024690c4abSTejun Heo * @head: insertion point 16034690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 16044690c4abSTejun Heo * 1605112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1606706026c2STejun Heo * work_struct flags. 16074690c4abSTejun Heo * 16084690c4abSTejun Heo * CONTEXT: 1609a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1610365970a1SDavid Howells */ 1611112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1612112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1613b89deed3SOleg Nesterov { 1614fe089f87STejun Heo debug_work_activate(work); 1615e1d8aa9fSFrederic Weisbecker 1616e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1617f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1618e89a85d6SWalter Wu 16194690c4abSTejun Heo /* we own @work, set data and link */ 1620112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 16211a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 16228864b4e5STejun Heo get_pwq(pwq); 1623b89deed3SOleg Nesterov } 1624b89deed3SOleg Nesterov 1625c8efcc25STejun Heo /* 1626c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 16278d03ecfeSTejun Heo * same workqueue. 1628c8efcc25STejun Heo */ 1629c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1630c8efcc25STejun Heo { 1631c8efcc25STejun Heo struct worker *worker; 1632c8efcc25STejun Heo 16338d03ecfeSTejun Heo worker = current_wq_worker(); 1634c8efcc25STejun Heo /* 1635bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 16368d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1637c8efcc25STejun Heo */ 1638112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1639c8efcc25STejun Heo } 1640c8efcc25STejun Heo 1641ef557180SMike Galbraith /* 1642ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1643ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1644ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1645ef557180SMike Galbraith */ 1646ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1647ef557180SMike Galbraith { 1648ef557180SMike Galbraith int new_cpu; 1649ef557180SMike Galbraith 1650f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1651ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1652ef557180SMike Galbraith return cpu; 1653a8ec5880SAmmar Faizi } else { 1654a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1655f303fccbSTejun Heo } 1656f303fccbSTejun Heo 1657ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1658ef557180SMike Galbraith return cpu; 1659ef557180SMike Galbraith 1660ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1661ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1662ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1663ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1664ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1665ef557180SMike Galbraith return cpu; 1666ef557180SMike Galbraith } 1667ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1668ef557180SMike Galbraith 1669ef557180SMike Galbraith return new_cpu; 1670ef557180SMike Galbraith } 1671ef557180SMike Galbraith 1672d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 16731da177e4SLinus Torvalds struct work_struct *work) 16741da177e4SLinus Torvalds { 1675112202d9STejun Heo struct pool_workqueue *pwq; 1676fe089f87STejun Heo struct worker_pool *last_pool, *pool; 16778a2e8e5dSTejun Heo unsigned int work_flags; 1678b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 16798930cabaSTejun Heo 16808930cabaSTejun Heo /* 16818930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 16828930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 16838930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 16848930cabaSTejun Heo * happen with IRQ disabled. 16858930cabaSTejun Heo */ 16868e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 16871da177e4SLinus Torvalds 16881e19ffc6STejun Heo 168933e3f0a3SRichard Clark /* 169033e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 169133e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 169233e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 169333e3f0a3SRichard Clark */ 169433e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 169533e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1696e41e704bSTejun Heo return; 169724acfb71SThomas Gleixner rcu_read_lock(); 16989e8cd2f5STejun Heo retry: 1699aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1700636b927eSTejun Heo if (req_cpu == WORK_CPU_UNBOUND) { 1701636b927eSTejun Heo if (wq->flags & WQ_UNBOUND) 1702ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1703636b927eSTejun Heo else 1704aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1705aa202f1fSHillf Danton } 1706f3421797STejun Heo 1707636b927eSTejun Heo pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 1708fe089f87STejun Heo pool = pwq->pool; 1709fe089f87STejun Heo 171018aa9effSTejun Heo /* 1711c9178087STejun Heo * If @work was previously on a different pool, it might still be 1712c9178087STejun Heo * running there, in which case the work needs to be queued on that 1713c9178087STejun Heo * pool to guarantee non-reentrancy. 171418aa9effSTejun Heo */ 1715c9e7cf27STejun Heo last_pool = get_work_pool(work); 1716fe089f87STejun Heo if (last_pool && last_pool != pool) { 171718aa9effSTejun Heo struct worker *worker; 171818aa9effSTejun Heo 1719a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 172018aa9effSTejun Heo 1721c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 172218aa9effSTejun Heo 1723112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1724c9178087STejun Heo pwq = worker->current_pwq; 1725fe089f87STejun Heo pool = pwq->pool; 1726fe089f87STejun Heo WARN_ON_ONCE(pool != last_pool); 17278594fadeSLai Jiangshan } else { 172818aa9effSTejun Heo /* meh... not running there, queue here */ 1729a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1730fe089f87STejun Heo raw_spin_lock(&pool->lock); 173118aa9effSTejun Heo } 17328930cabaSTejun Heo } else { 1733fe089f87STejun Heo raw_spin_lock(&pool->lock); 17348930cabaSTejun Heo } 1735502ca9d8STejun Heo 17369e8cd2f5STejun Heo /* 1737636b927eSTejun Heo * pwq is determined and locked. For unbound pools, we could have raced 1738636b927eSTejun Heo * with pwq release and it could already be dead. If its refcnt is zero, 1739636b927eSTejun Heo * repeat pwq selection. Note that unbound pwqs never die without 1740636b927eSTejun Heo * another pwq replacing it in cpu_pwq or while work items are executing 1741636b927eSTejun Heo * on it, so the retrying is guaranteed to make forward-progress. 17429e8cd2f5STejun Heo */ 17439e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 17449e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1745fe089f87STejun Heo raw_spin_unlock(&pool->lock); 17469e8cd2f5STejun Heo cpu_relax(); 17479e8cd2f5STejun Heo goto retry; 17489e8cd2f5STejun Heo } 17499e8cd2f5STejun Heo /* oops */ 17509e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 17519e8cd2f5STejun Heo wq->name, cpu); 17529e8cd2f5STejun Heo } 17539e8cd2f5STejun Heo 1754112202d9STejun Heo /* pwq determined, queue */ 1755112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1756502ca9d8STejun Heo 175724acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 175824acfb71SThomas Gleixner goto out; 17591e19ffc6STejun Heo 1760112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1761112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 17621e19ffc6STejun Heo 1763112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1764fe089f87STejun Heo if (list_empty(&pool->worklist)) 1765fe089f87STejun Heo pool->watchdog_ts = jiffies; 1766fe089f87STejun Heo 1767cdadf009STejun Heo trace_workqueue_activate_work(work); 1768112202d9STejun Heo pwq->nr_active++; 1769fe089f87STejun Heo insert_work(pwq, work, &pool->worklist, work_flags); 17700219a352STejun Heo kick_pool(pool); 17718a2e8e5dSTejun Heo } else { 1772f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1773fe089f87STejun Heo insert_work(pwq, work, &pwq->inactive_works, work_flags); 17748a2e8e5dSTejun Heo } 17751e19ffc6STejun Heo 177624acfb71SThomas Gleixner out: 1777fe089f87STejun Heo raw_spin_unlock(&pool->lock); 177824acfb71SThomas Gleixner rcu_read_unlock(); 17791da177e4SLinus Torvalds } 17801da177e4SLinus Torvalds 17810fcb78c2SRolf Eike Beer /** 1782c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1783c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1784c1a220e7SZhang Rui * @wq: workqueue to use 1785c1a220e7SZhang Rui * @work: work to queue 1786c1a220e7SZhang Rui * 1787c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1788443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1789443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1790854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1791854f5cc5SPaul E. McKenney * online will get a splat. 1792d185af30SYacine Belkadi * 1793d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1794c1a220e7SZhang Rui */ 1795d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1796d4283e93STejun Heo struct work_struct *work) 1797c1a220e7SZhang Rui { 1798d4283e93STejun Heo bool ret = false; 17998930cabaSTejun Heo unsigned long flags; 18008930cabaSTejun Heo 18018930cabaSTejun Heo local_irq_save(flags); 1802c1a220e7SZhang Rui 180322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18044690c4abSTejun Heo __queue_work(cpu, wq, work); 1805d4283e93STejun Heo ret = true; 1806c1a220e7SZhang Rui } 18078930cabaSTejun Heo 18088930cabaSTejun Heo local_irq_restore(flags); 1809c1a220e7SZhang Rui return ret; 1810c1a220e7SZhang Rui } 1811ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1812c1a220e7SZhang Rui 18138204e0c1SAlexander Duyck /** 1814fef59c9cSTejun Heo * select_numa_node_cpu - Select a CPU based on NUMA node 18158204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 18168204e0c1SAlexander Duyck * 18178204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 18188204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 18198204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 18208204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 18218204e0c1SAlexander Duyck */ 1822fef59c9cSTejun Heo static int select_numa_node_cpu(int node) 18238204e0c1SAlexander Duyck { 18248204e0c1SAlexander Duyck int cpu; 18258204e0c1SAlexander Duyck 18268204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 18278204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 18288204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 18298204e0c1SAlexander Duyck 18308204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 18318204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 18328204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 18338204e0c1SAlexander Duyck return cpu; 18348204e0c1SAlexander Duyck 18358204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 18368204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 18378204e0c1SAlexander Duyck 18388204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 18398204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 18408204e0c1SAlexander Duyck } 18418204e0c1SAlexander Duyck 18428204e0c1SAlexander Duyck /** 18438204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 18448204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 18458204e0c1SAlexander Duyck * @wq: workqueue to use 18468204e0c1SAlexander Duyck * @work: work to queue 18478204e0c1SAlexander Duyck * 18488204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 18498204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 18508204e0c1SAlexander Duyck * NUMA node. 18518204e0c1SAlexander Duyck * 18528204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 18538204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 18548204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 18558204e0c1SAlexander Duyck * 18568204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 18578204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 18588204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 18598204e0c1SAlexander Duyck * 18608204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18618204e0c1SAlexander Duyck */ 18628204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18638204e0c1SAlexander Duyck struct work_struct *work) 18648204e0c1SAlexander Duyck { 18658204e0c1SAlexander Duyck unsigned long flags; 18668204e0c1SAlexander Duyck bool ret = false; 18678204e0c1SAlexander Duyck 18688204e0c1SAlexander Duyck /* 18698204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 18708204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 18718204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 18728204e0c1SAlexander Duyck * 18738204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 18748204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 18758204e0c1SAlexander Duyck * some round robin type logic. 18768204e0c1SAlexander Duyck */ 18778204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 18788204e0c1SAlexander Duyck 18798204e0c1SAlexander Duyck local_irq_save(flags); 18808204e0c1SAlexander Duyck 18818204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1882fef59c9cSTejun Heo int cpu = select_numa_node_cpu(node); 18838204e0c1SAlexander Duyck 18848204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 18858204e0c1SAlexander Duyck ret = true; 18868204e0c1SAlexander Duyck } 18878204e0c1SAlexander Duyck 18888204e0c1SAlexander Duyck local_irq_restore(flags); 18898204e0c1SAlexander Duyck return ret; 18908204e0c1SAlexander Duyck } 18918204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 18928204e0c1SAlexander Duyck 18938c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 18941da177e4SLinus Torvalds { 18958c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 18961da177e4SLinus Torvalds 1897e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 189860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 18991da177e4SLinus Torvalds } 19001438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 19011da177e4SLinus Torvalds 19027beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 190352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19041da177e4SLinus Torvalds { 19057beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 19067beb2edfSTejun Heo struct work_struct *work = &dwork->work; 19071da177e4SLinus Torvalds 1908637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 19094b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1910fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1911fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 19127beb2edfSTejun Heo 19138852aac2STejun Heo /* 19148852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 19158852aac2STejun Heo * both optimization and correctness. The earliest @timer can 19168852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 19178852aac2STejun Heo * on that there's no such delay when @delay is 0. 19188852aac2STejun Heo */ 19198852aac2STejun Heo if (!delay) { 19208852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 19218852aac2STejun Heo return; 19228852aac2STejun Heo } 19238852aac2STejun Heo 192460c057bcSLai Jiangshan dwork->wq = wq; 19251265057fSTejun Heo dwork->cpu = cpu; 19267beb2edfSTejun Heo timer->expires = jiffies + delay; 19277beb2edfSTejun Heo 1928041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 19297beb2edfSTejun Heo add_timer_on(timer, cpu); 1930041bd12eSTejun Heo else 1931041bd12eSTejun Heo add_timer(timer); 19327beb2edfSTejun Heo } 19331da177e4SLinus Torvalds 19340fcb78c2SRolf Eike Beer /** 19350fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 19360fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 19370fcb78c2SRolf Eike Beer * @wq: workqueue to use 1938af9997e4SRandy Dunlap * @dwork: work to queue 19390fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 19400fcb78c2SRolf Eike Beer * 1941d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1942715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1943715f1300STejun Heo * execution. 19440fcb78c2SRolf Eike Beer */ 1945d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 194652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19477a6bc1cdSVenkatesh Pallipadi { 194852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1949d4283e93STejun Heo bool ret = false; 19508930cabaSTejun Heo unsigned long flags; 19518930cabaSTejun Heo 19528930cabaSTejun Heo /* read the comment in __queue_work() */ 19538930cabaSTejun Heo local_irq_save(flags); 19547a6bc1cdSVenkatesh Pallipadi 195522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 19567beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1957d4283e93STejun Heo ret = true; 19587a6bc1cdSVenkatesh Pallipadi } 19598930cabaSTejun Heo 19608930cabaSTejun Heo local_irq_restore(flags); 19617a6bc1cdSVenkatesh Pallipadi return ret; 19627a6bc1cdSVenkatesh Pallipadi } 1963ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19641da177e4SLinus Torvalds 1965c8e55f36STejun Heo /** 19668376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19678376fe22STejun Heo * @cpu: CPU number to execute work on 19688376fe22STejun Heo * @wq: workqueue to use 19698376fe22STejun Heo * @dwork: work to queue 19708376fe22STejun Heo * @delay: number of jiffies to wait before queueing 19718376fe22STejun Heo * 19728376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 19738376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 19748376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 19758376fe22STejun Heo * current state. 19768376fe22STejun Heo * 1977d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 19788376fe22STejun Heo * pending and its timer was modified. 19798376fe22STejun Heo * 1980e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 19818376fe22STejun Heo * See try_to_grab_pending() for details. 19828376fe22STejun Heo */ 19838376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 19848376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 19858376fe22STejun Heo { 19868376fe22STejun Heo unsigned long flags; 19878376fe22STejun Heo int ret; 19888376fe22STejun Heo 19898376fe22STejun Heo do { 19908376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 19918376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 19928376fe22STejun Heo 19938376fe22STejun Heo if (likely(ret >= 0)) { 19948376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 19958376fe22STejun Heo local_irq_restore(flags); 19968376fe22STejun Heo } 19978376fe22STejun Heo 19988376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 19998376fe22STejun Heo return ret; 20008376fe22STejun Heo } 20018376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 20028376fe22STejun Heo 200305f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 200405f0fe6bSTejun Heo { 200505f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 200605f0fe6bSTejun Heo 200705f0fe6bSTejun Heo /* read the comment in __queue_work() */ 200805f0fe6bSTejun Heo local_irq_disable(); 200905f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 201005f0fe6bSTejun Heo local_irq_enable(); 201105f0fe6bSTejun Heo } 201205f0fe6bSTejun Heo 201305f0fe6bSTejun Heo /** 201405f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 201505f0fe6bSTejun Heo * @wq: workqueue to use 201605f0fe6bSTejun Heo * @rwork: work to queue 201705f0fe6bSTejun Heo * 201805f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 201905f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 2020bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 202105f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 202205f0fe6bSTejun Heo */ 202305f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 202405f0fe6bSTejun Heo { 202505f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 202605f0fe6bSTejun Heo 202705f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 202805f0fe6bSTejun Heo rwork->wq = wq; 2029a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 203005f0fe6bSTejun Heo return true; 203105f0fe6bSTejun Heo } 203205f0fe6bSTejun Heo 203305f0fe6bSTejun Heo return false; 203405f0fe6bSTejun Heo } 203505f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 203605f0fe6bSTejun Heo 2037f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2038c34056a3STejun Heo { 2039c34056a3STejun Heo struct worker *worker; 2040c34056a3STejun Heo 2041f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2042c8e55f36STejun Heo if (worker) { 2043c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2044affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2045da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2046e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2047e22bee78STejun Heo worker->flags = WORKER_PREP; 2048c8e55f36STejun Heo } 2049c34056a3STejun Heo return worker; 2050c34056a3STejun Heo } 2051c34056a3STejun Heo 2052*9546b29eSTejun Heo static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 2053*9546b29eSTejun Heo { 2054*9546b29eSTejun Heo return pool->attrs->__pod_cpumask; 2055*9546b29eSTejun Heo } 2056*9546b29eSTejun Heo 2057c34056a3STejun Heo /** 20584736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20594736cbf7SLai Jiangshan * @worker: worker to be attached 20604736cbf7SLai Jiangshan * @pool: the target pool 20614736cbf7SLai Jiangshan * 20624736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20634736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20644736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20654736cbf7SLai Jiangshan */ 20664736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 20674736cbf7SLai Jiangshan struct worker_pool *pool) 20684736cbf7SLai Jiangshan { 20691258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 20704736cbf7SLai Jiangshan 20714736cbf7SLai Jiangshan /* 20721258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 20731258fae7STejun Heo * stable across this function. See the comments above the flag 20741258fae7STejun Heo * definition for details. 20754736cbf7SLai Jiangshan */ 20764736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 20774736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 20785c25b5ffSPeter Zijlstra else 20795c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 20804736cbf7SLai Jiangshan 2081640f17c8SPeter Zijlstra if (worker->rescue_wq) 2082*9546b29eSTejun Heo set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2083640f17c8SPeter Zijlstra 20844736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2085a2d812a2STejun Heo worker->pool = pool; 20864736cbf7SLai Jiangshan 20871258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 20884736cbf7SLai Jiangshan } 20894736cbf7SLai Jiangshan 20904736cbf7SLai Jiangshan /** 209160f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 209260f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 209360f5a4bcSLai Jiangshan * 20944736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 20954736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 20964736cbf7SLai Jiangshan * other reference to the pool. 209760f5a4bcSLai Jiangshan */ 2098a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 209960f5a4bcSLai Jiangshan { 2100a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 210160f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 210260f5a4bcSLai Jiangshan 21031258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2104a2d812a2STejun Heo 21055c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2106da028469SLai Jiangshan list_del(&worker->node); 2107a2d812a2STejun Heo worker->pool = NULL; 2108a2d812a2STejun Heo 2109e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 211060f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 21111258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 211260f5a4bcSLai Jiangshan 2113b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2114b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2115b62c0751SLai Jiangshan 211660f5a4bcSLai Jiangshan if (detach_completion) 211760f5a4bcSLai Jiangshan complete(detach_completion); 211860f5a4bcSLai Jiangshan } 211960f5a4bcSLai Jiangshan 212060f5a4bcSLai Jiangshan /** 2121c34056a3STejun Heo * create_worker - create a new workqueue worker 212263d95a91STejun Heo * @pool: pool the new worker will belong to 2123c34056a3STejun Heo * 2124051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2125c34056a3STejun Heo * 2126c34056a3STejun Heo * CONTEXT: 2127c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2128c34056a3STejun Heo * 2129d185af30SYacine Belkadi * Return: 2130c34056a3STejun Heo * Pointer to the newly created worker. 2131c34056a3STejun Heo */ 2132bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2133c34056a3STejun Heo { 2134e441b56fSZhen Lei struct worker *worker; 2135e441b56fSZhen Lei int id; 2136e3c916a4STejun Heo char id_buf[16]; 2137c34056a3STejun Heo 21387cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2139e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21403f0ea0b8SPetr Mladek if (id < 0) { 21413f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21423f0ea0b8SPetr Mladek ERR_PTR(id)); 2143e441b56fSZhen Lei return NULL; 21443f0ea0b8SPetr Mladek } 2145c34056a3STejun Heo 2146f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21473f0ea0b8SPetr Mladek if (!worker) { 21483f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2149c34056a3STejun Heo goto fail; 21503f0ea0b8SPetr Mladek } 2151c34056a3STejun Heo 2152c34056a3STejun Heo worker->id = id; 2153c34056a3STejun Heo 215429c91e99STejun Heo if (pool->cpu >= 0) 2155e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2156e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2157f3421797STejun Heo else 2158e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2159e3c916a4STejun Heo 2160f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2161e3c916a4STejun Heo "kworker/%s", id_buf); 21623f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 216360f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 216460f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 216560f54038SPetr Mladek id_buf); 216660f54038SPetr Mladek } else { 21673f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 21683f0ea0b8SPetr Mladek worker->task); 216960f54038SPetr Mladek } 2170c34056a3STejun Heo goto fail; 21713f0ea0b8SPetr Mladek } 2172c34056a3STejun Heo 217391151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 2174*9546b29eSTejun Heo kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 217591151228SOleg Nesterov 2176da028469SLai Jiangshan /* successful, attach the worker to the pool */ 21774736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2178822d8405STejun Heo 2179051e1850SLai Jiangshan /* start the newly created worker */ 2180a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 21810219a352STejun Heo 2182051e1850SLai Jiangshan worker->pool->nr_workers++; 2183051e1850SLai Jiangshan worker_enter_idle(worker); 21840219a352STejun Heo kick_pool(pool); 21850219a352STejun Heo 21860219a352STejun Heo /* 21870219a352STejun Heo * @worker is waiting on a completion in kthread() and will trigger hung 21880219a352STejun Heo * check if not woken up soon. As kick_pool() might not have waken it 21890219a352STejun Heo * up, wake it up explicitly once more. 21900219a352STejun Heo */ 2191051e1850SLai Jiangshan wake_up_process(worker->task); 21920219a352STejun Heo 2193a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2194051e1850SLai Jiangshan 2195c34056a3STejun Heo return worker; 2196822d8405STejun Heo 2197c34056a3STejun Heo fail: 2198e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2199c34056a3STejun Heo kfree(worker); 2200c34056a3STejun Heo return NULL; 2201c34056a3STejun Heo } 2202c34056a3STejun Heo 2203793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2204793777bcSValentin Schneider { 2205793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2206793777bcSValentin Schneider 2207793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2208793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2209793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2210793777bcSValentin Schneider else 2211793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2212793777bcSValentin Schneider } 2213793777bcSValentin Schneider 2214e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2215e02b9312SValentin Schneider { 2216e02b9312SValentin Schneider struct worker *worker, *tmp; 2217e02b9312SValentin Schneider 2218e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2219e02b9312SValentin Schneider list_del_init(&worker->entry); 2220e02b9312SValentin Schneider unbind_worker(worker); 2221e02b9312SValentin Schneider /* 2222e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2223e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2224e02b9312SValentin Schneider * wouldn't have gotten here. 2225c34056a3STejun Heo * 2226e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2227e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2228e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2229e02b9312SValentin Schneider * outside of pool->lock. 2230e02b9312SValentin Schneider */ 2231e02b9312SValentin Schneider wake_up_process(worker->task); 2232e02b9312SValentin Schneider } 2233e02b9312SValentin Schneider } 2234e02b9312SValentin Schneider 2235e02b9312SValentin Schneider /** 2236e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2237e02b9312SValentin Schneider * @worker: worker to be destroyed 2238e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2239e02b9312SValentin Schneider * 2240e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2241e02b9312SValentin Schneider * should be idle. 2242c8e55f36STejun Heo * 2243c8e55f36STejun Heo * CONTEXT: 2244a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2245c34056a3STejun Heo */ 2246e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2247c34056a3STejun Heo { 2248bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2249c34056a3STejun Heo 2250cd549687STejun Heo lockdep_assert_held(&pool->lock); 2251e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2252cd549687STejun Heo 2253c34056a3STejun Heo /* sanity check frenzy */ 22546183c009STejun Heo if (WARN_ON(worker->current_work) || 225573eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 225673eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22576183c009STejun Heo return; 2258c34056a3STejun Heo 2259bd7bdd43STejun Heo pool->nr_workers--; 2260bd7bdd43STejun Heo pool->nr_idle--; 2261c8e55f36STejun Heo 2262cb444766STejun Heo worker->flags |= WORKER_DIE; 2263e02b9312SValentin Schneider 2264e02b9312SValentin Schneider list_move(&worker->entry, list); 2265e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2266c34056a3STejun Heo } 2267c34056a3STejun Heo 22683f959aa3SValentin Schneider /** 22693f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 22703f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 22713f959aa3SValentin Schneider * 22723f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 22733f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 22743f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 22753f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 22763f959aa3SValentin Schneider * it expire and re-evaluate things from there. 22773f959aa3SValentin Schneider */ 227832a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2279e22bee78STejun Heo { 228032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 22813f959aa3SValentin Schneider bool do_cull = false; 22823f959aa3SValentin Schneider 22833f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 22843f959aa3SValentin Schneider return; 22853f959aa3SValentin Schneider 22863f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 22873f959aa3SValentin Schneider 22883f959aa3SValentin Schneider if (too_many_workers(pool)) { 22893f959aa3SValentin Schneider struct worker *worker; 22903f959aa3SValentin Schneider unsigned long expires; 22913f959aa3SValentin Schneider 22923f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 22933f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 22943f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 22953f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 22963f959aa3SValentin Schneider 22973f959aa3SValentin Schneider if (!do_cull) 22983f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 22993f959aa3SValentin Schneider } 23003f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 23013f959aa3SValentin Schneider 23023f959aa3SValentin Schneider if (do_cull) 23033f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 23043f959aa3SValentin Schneider } 23053f959aa3SValentin Schneider 23063f959aa3SValentin Schneider /** 23073f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 23083f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 23093f959aa3SValentin Schneider * 23103f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 23113f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2312e02b9312SValentin Schneider * 2313e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2314e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2315e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 23163f959aa3SValentin Schneider */ 23173f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 23183f959aa3SValentin Schneider { 23193f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 23209680540cSYang Yingliang LIST_HEAD(cull_list); 2321e22bee78STejun Heo 2322e02b9312SValentin Schneider /* 2323e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2324e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2325e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2326e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2327e02b9312SValentin Schneider */ 2328e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2329a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2330e22bee78STejun Heo 23313347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2332e22bee78STejun Heo struct worker *worker; 2333e22bee78STejun Heo unsigned long expires; 2334e22bee78STejun Heo 233563d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2336e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2337e22bee78STejun Heo 23383347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 233963d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23403347fc9fSLai Jiangshan break; 2341e22bee78STejun Heo } 23423347fc9fSLai Jiangshan 2343e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2344e22bee78STejun Heo } 2345e22bee78STejun Heo 2346a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2347e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2348e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2349e22bee78STejun Heo } 2350e22bee78STejun Heo 2351493a1724STejun Heo static void send_mayday(struct work_struct *work) 2352e22bee78STejun Heo { 2353112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2354112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2355493a1724STejun Heo 23562e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2357e22bee78STejun Heo 2358493008a8STejun Heo if (!wq->rescuer) 2359493a1724STejun Heo return; 2360e22bee78STejun Heo 2361e22bee78STejun Heo /* mayday mayday mayday */ 2362493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 236377668c8bSLai Jiangshan /* 236477668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 236577668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 236677668c8bSLai Jiangshan * rescuer is done with it. 236777668c8bSLai Jiangshan */ 236877668c8bSLai Jiangshan get_pwq(pwq); 2369493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2370e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2371725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2372493a1724STejun Heo } 2373e22bee78STejun Heo } 2374e22bee78STejun Heo 237532a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2376e22bee78STejun Heo { 237732a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2378e22bee78STejun Heo struct work_struct *work; 2379e22bee78STejun Heo 2380a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2381a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2382e22bee78STejun Heo 238363d95a91STejun Heo if (need_to_create_worker(pool)) { 2384e22bee78STejun Heo /* 2385e22bee78STejun Heo * We've been trying to create a new worker but 2386e22bee78STejun Heo * haven't been successful. We might be hitting an 2387e22bee78STejun Heo * allocation deadlock. Send distress signals to 2388e22bee78STejun Heo * rescuers. 2389e22bee78STejun Heo */ 239063d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2391e22bee78STejun Heo send_mayday(work); 2392e22bee78STejun Heo } 2393e22bee78STejun Heo 2394a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2395a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2396e22bee78STejun Heo 239763d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2398e22bee78STejun Heo } 2399e22bee78STejun Heo 2400e22bee78STejun Heo /** 2401e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 240263d95a91STejun Heo * @pool: pool to create a new worker for 2403e22bee78STejun Heo * 240463d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2405e22bee78STejun Heo * have at least one idle worker on return from this function. If 2406e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 240763d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2408e22bee78STejun Heo * possible allocation deadlock. 2409e22bee78STejun Heo * 2410c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2411c5aa87bbSTejun Heo * may_start_working() %true. 2412e22bee78STejun Heo * 2413e22bee78STejun Heo * LOCKING: 2414a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2415e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2416e22bee78STejun Heo * manager. 2417e22bee78STejun Heo */ 241829187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2419d565ed63STejun Heo __releases(&pool->lock) 2420d565ed63STejun Heo __acquires(&pool->lock) 2421e22bee78STejun Heo { 2422e22bee78STejun Heo restart: 2423a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24249f9c2364STejun Heo 2425e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 242663d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2427e22bee78STejun Heo 2428e22bee78STejun Heo while (true) { 2429051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2430e22bee78STejun Heo break; 2431e22bee78STejun Heo 2432e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24339f9c2364STejun Heo 243463d95a91STejun Heo if (!need_to_create_worker(pool)) 2435e22bee78STejun Heo break; 2436e22bee78STejun Heo } 2437e22bee78STejun Heo 243863d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2439a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2440051e1850SLai Jiangshan /* 2441051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2442051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2443051e1850SLai Jiangshan * already become busy. 2444051e1850SLai Jiangshan */ 244563d95a91STejun Heo if (need_to_create_worker(pool)) 2446e22bee78STejun Heo goto restart; 2447e22bee78STejun Heo } 2448e22bee78STejun Heo 2449e22bee78STejun Heo /** 2450e22bee78STejun Heo * manage_workers - manage worker pool 2451e22bee78STejun Heo * @worker: self 2452e22bee78STejun Heo * 2453706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2454e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2455706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2456e22bee78STejun Heo * 2457e22bee78STejun Heo * The caller can safely start processing works on false return. On 2458e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2459e22bee78STejun Heo * and may_start_working() is true. 2460e22bee78STejun Heo * 2461e22bee78STejun Heo * CONTEXT: 2462a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2463e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2464e22bee78STejun Heo * 2465d185af30SYacine Belkadi * Return: 246629187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 246729187a9eSTejun Heo * start processing works, %true if management function was performed and 246829187a9eSTejun Heo * the conditions that the caller verified before calling the function may 246929187a9eSTejun Heo * no longer be true. 2470e22bee78STejun Heo */ 2471e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2472e22bee78STejun Heo { 247363d95a91STejun Heo struct worker_pool *pool = worker->pool; 2474e22bee78STejun Heo 2475692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 247629187a9eSTejun Heo return false; 2477692b4825STejun Heo 2478692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 24792607d7a6STejun Heo pool->manager = worker; 2480e22bee78STejun Heo 248129187a9eSTejun Heo maybe_create_worker(pool); 2482e22bee78STejun Heo 24832607d7a6STejun Heo pool->manager = NULL; 2484692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2485d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 248629187a9eSTejun Heo return true; 2487e22bee78STejun Heo } 2488e22bee78STejun Heo 2489a62428c0STejun Heo /** 2490a62428c0STejun Heo * process_one_work - process single work 2491c34056a3STejun Heo * @worker: self 2492a62428c0STejun Heo * @work: work to process 2493a62428c0STejun Heo * 2494a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2495a62428c0STejun Heo * process a single work including synchronization against and 2496a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2497a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2498a62428c0STejun Heo * call this function to process a work. 2499a62428c0STejun Heo * 2500a62428c0STejun Heo * CONTEXT: 2501a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2502a62428c0STejun Heo */ 2503c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2504d565ed63STejun Heo __releases(&pool->lock) 2505d565ed63STejun Heo __acquires(&pool->lock) 25061da177e4SLinus Torvalds { 2507112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2508bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2509c4560c2cSLai Jiangshan unsigned long work_data; 25104e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 25114e6045f1SJohannes Berg /* 2512a62428c0STejun Heo * It is permissible to free the struct work_struct from 2513a62428c0STejun Heo * inside the function that is called from it, this we need to 2514a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2515a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2516a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 25174e6045f1SJohannes Berg */ 25184d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 25194d82a1deSPeter Zijlstra 25204d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 25214e6045f1SJohannes Berg #endif 2522807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 252385327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2524ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 252525511a47STejun Heo 25268930cabaSTejun Heo /* claim and dequeue */ 2527dc186ad7SThomas Gleixner debug_work_deactivate(work); 2528c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2529c34056a3STejun Heo worker->current_work = work; 2530a2c1c57bSTejun Heo worker->current_func = work->func; 2531112202d9STejun Heo worker->current_pwq = pwq; 2532616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2533c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2534d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25357a22ad75STejun Heo 25368bf89593STejun Heo /* 25378bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25388bf89593STejun Heo * overridden through set_worker_desc(). 25398bf89593STejun Heo */ 25408bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25418bf89593STejun Heo 2542a62428c0STejun Heo list_del_init(&work->entry); 2543a62428c0STejun Heo 2544649027d7STejun Heo /* 2545228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2546228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2547228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2548228f1d00SLai Jiangshan * execution of the pending work items. 2549fb0e7bebSTejun Heo */ 2550616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2551228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2552fb0e7bebSTejun Heo 2553974271c4STejun Heo /* 25540219a352STejun Heo * Kick @pool if necessary. It's always noop for per-cpu worker pools 25550219a352STejun Heo * since nr_running would always be >= 1 at this point. This is used to 25560219a352STejun Heo * chain execution of the pending work items for WORKER_NOT_RUNNING 25570219a352STejun Heo * workers such as the UNBOUND and CPU_INTENSIVE ones. 2558974271c4STejun Heo */ 25590219a352STejun Heo kick_pool(pool); 2560974271c4STejun Heo 25618930cabaSTejun Heo /* 25627c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2563d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 256423657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 256523657bb1STejun Heo * disabled. 25668930cabaSTejun Heo */ 25677c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 25681da177e4SLinus Torvalds 2569a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2570365970a1SDavid Howells 2571a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 25723295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2573e6f3faa7SPeter Zijlstra /* 2574f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2575f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2576e6f3faa7SPeter Zijlstra * 2577e6f3faa7SPeter Zijlstra * However, that would result in: 2578e6f3faa7SPeter Zijlstra * 2579e6f3faa7SPeter Zijlstra * A(W1) 2580e6f3faa7SPeter Zijlstra * WFC(C) 2581e6f3faa7SPeter Zijlstra * A(W1) 2582e6f3faa7SPeter Zijlstra * C(C) 2583e6f3faa7SPeter Zijlstra * 2584e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2585e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2586e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2587f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2588e6f3faa7SPeter Zijlstra * these locks. 2589e6f3faa7SPeter Zijlstra * 2590e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2591e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2592e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2593e6f3faa7SPeter Zijlstra */ 2594f52be570SPeter Zijlstra lockdep_invariant_state(true); 2595725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2596e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2597a2c1c57bSTejun Heo worker->current_func(work); 2598e36c886aSArjan van de Ven /* 2599e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2600e36c886aSArjan van de Ven * point will only record its address. 2601e36c886aSArjan van de Ven */ 26021c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2603725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 26043295f0efSIngo Molnar lock_map_release(&lockdep_map); 2605112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 26061da177e4SLinus Torvalds 2607d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2608044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2609d75f773cSSakari Ailus " last function: %ps\n", 2610a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2611a2c1c57bSTejun Heo worker->current_func); 2612d5abe669SPeter Zijlstra debug_show_held_locks(current); 2613d5abe669SPeter Zijlstra dump_stack(); 2614d5abe669SPeter Zijlstra } 2615d5abe669SPeter Zijlstra 2616b22ce278STejun Heo /* 2617025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2618b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2619b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2620b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2621789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2622789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2623b22ce278STejun Heo */ 2624a7e6425eSPaul E. McKenney cond_resched(); 2625b22ce278STejun Heo 2626a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2627a62428c0STejun Heo 2628616db877STejun Heo /* 2629616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2630616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2631616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2632616db877STejun Heo */ 2633fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2634fb0e7bebSTejun Heo 26351b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26361b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26371b69ac6bSJohannes Weiner 2638a62428c0STejun Heo /* we're done with it, release */ 263942f8570fSSasha Levin hash_del(&worker->hentry); 2640c34056a3STejun Heo worker->current_work = NULL; 2641a2c1c57bSTejun Heo worker->current_func = NULL; 2642112202d9STejun Heo worker->current_pwq = NULL; 2643d812796eSLai Jiangshan worker->current_color = INT_MAX; 2644c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26451da177e4SLinus Torvalds } 26461da177e4SLinus Torvalds 2647affee4b2STejun Heo /** 2648affee4b2STejun Heo * process_scheduled_works - process scheduled works 2649affee4b2STejun Heo * @worker: self 2650affee4b2STejun Heo * 2651affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2652affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2653affee4b2STejun Heo * fetches a work from the top and executes it. 2654affee4b2STejun Heo * 2655affee4b2STejun Heo * CONTEXT: 2656a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2657affee4b2STejun Heo * multiple times. 2658affee4b2STejun Heo */ 2659affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26601da177e4SLinus Torvalds { 2661c0ab017dSTejun Heo struct work_struct *work; 2662c0ab017dSTejun Heo bool first = true; 2663c0ab017dSTejun Heo 2664c0ab017dSTejun Heo while ((work = list_first_entry_or_null(&worker->scheduled, 2665c0ab017dSTejun Heo struct work_struct, entry))) { 2666c0ab017dSTejun Heo if (first) { 2667c0ab017dSTejun Heo worker->pool->watchdog_ts = jiffies; 2668c0ab017dSTejun Heo first = false; 2669c0ab017dSTejun Heo } 2670c34056a3STejun Heo process_one_work(worker, work); 2671a62428c0STejun Heo } 26721da177e4SLinus Torvalds } 26731da177e4SLinus Torvalds 2674197f6accSTejun Heo static void set_pf_worker(bool val) 2675197f6accSTejun Heo { 2676197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2677197f6accSTejun Heo if (val) 2678197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2679197f6accSTejun Heo else 2680197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2681197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2682197f6accSTejun Heo } 2683197f6accSTejun Heo 26844690c4abSTejun Heo /** 26854690c4abSTejun Heo * worker_thread - the worker thread function 2686c34056a3STejun Heo * @__worker: self 26874690c4abSTejun Heo * 2688c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2689c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2690c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2691c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2692c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2693d185af30SYacine Belkadi * 2694d185af30SYacine Belkadi * Return: 0 26954690c4abSTejun Heo */ 2696c34056a3STejun Heo static int worker_thread(void *__worker) 26971da177e4SLinus Torvalds { 2698c34056a3STejun Heo struct worker *worker = __worker; 2699bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 27001da177e4SLinus Torvalds 2701e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2702197f6accSTejun Heo set_pf_worker(true); 2703c8e55f36STejun Heo woke_up: 2704a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2705affee4b2STejun Heo 2706a9ab775bSTejun Heo /* am I supposed to die? */ 2707a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2708a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2709197f6accSTejun Heo set_pf_worker(false); 271060f5a4bcSLai Jiangshan 271160f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2712e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2713a2d812a2STejun Heo worker_detach_from_pool(worker); 2714e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 271560f5a4bcSLai Jiangshan kfree(worker); 2716c8e55f36STejun Heo return 0; 2717c8e55f36STejun Heo } 2718c8e55f36STejun Heo 2719c8e55f36STejun Heo worker_leave_idle(worker); 2720db7bccf4STejun Heo recheck: 2721e22bee78STejun Heo /* no more worker necessary? */ 272263d95a91STejun Heo if (!need_more_worker(pool)) 2723e22bee78STejun Heo goto sleep; 2724e22bee78STejun Heo 2725e22bee78STejun Heo /* do we need to manage? */ 272663d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2727e22bee78STejun Heo goto recheck; 2728e22bee78STejun Heo 2729c8e55f36STejun Heo /* 2730c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2731c8e55f36STejun Heo * preparing to process a work or actually processing it. 2732c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2733c8e55f36STejun Heo */ 27346183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2735c8e55f36STejun Heo 2736e22bee78STejun Heo /* 2737a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2738a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2739a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2740a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2741a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2742e22bee78STejun Heo */ 2743a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2744e22bee78STejun Heo 2745e22bee78STejun Heo do { 2746affee4b2STejun Heo struct work_struct *work = 2747bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2748affee4b2STejun Heo struct work_struct, entry); 2749affee4b2STejun Heo 2750873eaca6STejun Heo if (assign_work(work, worker, NULL)) 2751affee4b2STejun Heo process_scheduled_works(worker); 275263d95a91STejun Heo } while (keep_working(pool)); 2753affee4b2STejun Heo 2754228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2755d313dd85STejun Heo sleep: 2756c8e55f36STejun Heo /* 2757d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2758d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2759d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2760d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2761d565ed63STejun Heo * event. 2762c8e55f36STejun Heo */ 2763c8e55f36STejun Heo worker_enter_idle(worker); 2764c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2765a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27661da177e4SLinus Torvalds schedule(); 2767c8e55f36STejun Heo goto woke_up; 27681da177e4SLinus Torvalds } 27691da177e4SLinus Torvalds 2770e22bee78STejun Heo /** 2771e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2772111c225aSTejun Heo * @__rescuer: self 2773e22bee78STejun Heo * 2774e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2775493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2776e22bee78STejun Heo * 2777706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2778e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2779e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2780e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2781e22bee78STejun Heo * the problem rescuer solves. 2782e22bee78STejun Heo * 2783706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2784706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2785e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2786e22bee78STejun Heo * 2787e22bee78STejun Heo * This should happen rarely. 2788d185af30SYacine Belkadi * 2789d185af30SYacine Belkadi * Return: 0 2790e22bee78STejun Heo */ 2791111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2792e22bee78STejun Heo { 2793111c225aSTejun Heo struct worker *rescuer = __rescuer; 2794111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 27954d595b86SLai Jiangshan bool should_stop; 2796e22bee78STejun Heo 2797e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2798111c225aSTejun Heo 2799111c225aSTejun Heo /* 2800111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2801111c225aSTejun Heo * doesn't participate in concurrency management. 2802111c225aSTejun Heo */ 2803197f6accSTejun Heo set_pf_worker(true); 2804e22bee78STejun Heo repeat: 2805c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 28061da177e4SLinus Torvalds 28074d595b86SLai Jiangshan /* 28084d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 28094d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28104d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28114d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28124d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28134d595b86SLai Jiangshan * list is always empty on exit. 28144d595b86SLai Jiangshan */ 28154d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28161da177e4SLinus Torvalds 2817493a1724STejun Heo /* see whether any pwq is asking for help */ 2818a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2819493a1724STejun Heo 2820493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2821493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2822493a1724STejun Heo struct pool_workqueue, mayday_node); 2823112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2824e22bee78STejun Heo struct work_struct *work, *n; 2825e22bee78STejun Heo 2826e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2827493a1724STejun Heo list_del_init(&pwq->mayday_node); 2828493a1724STejun Heo 2829a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2830e22bee78STejun Heo 283151697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 283251697d39SLai Jiangshan 2833a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2834e22bee78STejun Heo 2835e22bee78STejun Heo /* 2836e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2837e22bee78STejun Heo * process'em. 2838e22bee78STejun Heo */ 2839873eaca6STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 284082607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 2841873eaca6STejun Heo if (get_work_pwq(work) == pwq && 2842873eaca6STejun Heo assign_work(work, rescuer, &n)) 2843725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 284482607adcSTejun Heo } 2845e22bee78STejun Heo 2846873eaca6STejun Heo if (!list_empty(&rescuer->scheduled)) { 2847e22bee78STejun Heo process_scheduled_works(rescuer); 28487576958aSTejun Heo 28497576958aSTejun Heo /* 2850008847f6SNeilBrown * The above execution of rescued work items could 2851008847f6SNeilBrown * have created more to rescue through 2852f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2853008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2854008847f6SNeilBrown * that such back-to-back work items, which may be 2855008847f6SNeilBrown * being used to relieve memory pressure, don't 2856008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2857008847f6SNeilBrown */ 28584f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2859a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2860e66b39afSTejun Heo /* 2861e66b39afSTejun Heo * Queue iff we aren't racing destruction 2862e66b39afSTejun Heo * and somebody else hasn't queued it already. 2863e66b39afSTejun Heo */ 2864e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2865008847f6SNeilBrown get_pwq(pwq); 2866e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2867e66b39afSTejun Heo } 2868a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2869008847f6SNeilBrown } 2870008847f6SNeilBrown } 2871008847f6SNeilBrown 2872008847f6SNeilBrown /* 287377668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 287413b1d625SLai Jiangshan * go away while we're still attached to it. 287577668c8bSLai Jiangshan */ 287677668c8bSLai Jiangshan put_pwq(pwq); 287777668c8bSLai Jiangshan 287877668c8bSLai Jiangshan /* 28790219a352STejun Heo * Leave this pool. Notify regular workers; otherwise, we end up 28800219a352STejun Heo * with 0 concurrency and stalling the execution. 28817576958aSTejun Heo */ 28820219a352STejun Heo kick_pool(pool); 28837576958aSTejun Heo 2884a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 288513b1d625SLai Jiangshan 2886a2d812a2STejun Heo worker_detach_from_pool(rescuer); 288713b1d625SLai Jiangshan 2888a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 28891da177e4SLinus Torvalds } 28901da177e4SLinus Torvalds 2891a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2892493a1724STejun Heo 28934d595b86SLai Jiangshan if (should_stop) { 28944d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2895197f6accSTejun Heo set_pf_worker(false); 28964d595b86SLai Jiangshan return 0; 28974d595b86SLai Jiangshan } 28984d595b86SLai Jiangshan 2899111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2900111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2901e22bee78STejun Heo schedule(); 2902e22bee78STejun Heo goto repeat; 29031da177e4SLinus Torvalds } 29041da177e4SLinus Torvalds 2905fca839c0STejun Heo /** 2906fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2907fca839c0STejun Heo * @target_wq: workqueue being flushed 2908fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2909fca839c0STejun Heo * 2910fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2911fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2912fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2913fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2914fca839c0STejun Heo * a deadlock. 2915fca839c0STejun Heo */ 2916fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2917fca839c0STejun Heo struct work_struct *target_work) 2918fca839c0STejun Heo { 2919fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2920fca839c0STejun Heo struct worker *worker; 2921fca839c0STejun Heo 2922fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2923fca839c0STejun Heo return; 2924fca839c0STejun Heo 2925fca839c0STejun Heo worker = current_wq_worker(); 2926fca839c0STejun Heo 2927fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2928d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2929fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 293023d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 293123d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2932d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2933fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2934fca839c0STejun Heo target_wq->name, target_func); 2935fca839c0STejun Heo } 2936fca839c0STejun Heo 2937fc2e4d70SOleg Nesterov struct wq_barrier { 2938fc2e4d70SOleg Nesterov struct work_struct work; 2939fc2e4d70SOleg Nesterov struct completion done; 29402607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2941fc2e4d70SOleg Nesterov }; 2942fc2e4d70SOleg Nesterov 2943fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2944fc2e4d70SOleg Nesterov { 2945fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2946fc2e4d70SOleg Nesterov complete(&barr->done); 2947fc2e4d70SOleg Nesterov } 2948fc2e4d70SOleg Nesterov 29494690c4abSTejun Heo /** 29504690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2951112202d9STejun Heo * @pwq: pwq to insert barrier into 29524690c4abSTejun Heo * @barr: wq_barrier to insert 2953affee4b2STejun Heo * @target: target work to attach @barr to 2954affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29554690c4abSTejun Heo * 2956affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2957affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2958affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2959affee4b2STejun Heo * cpu. 2960affee4b2STejun Heo * 2961affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2962affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2963affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2964affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2965affee4b2STejun Heo * after a work with LINKED flag set. 2966affee4b2STejun Heo * 2967affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2968112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 29694690c4abSTejun Heo * 29704690c4abSTejun Heo * CONTEXT: 2971a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 29724690c4abSTejun Heo */ 2973112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 2974affee4b2STejun Heo struct wq_barrier *barr, 2975affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2976fc2e4d70SOleg Nesterov { 2977d812796eSLai Jiangshan unsigned int work_flags = 0; 2978d812796eSLai Jiangshan unsigned int work_color; 2979affee4b2STejun Heo struct list_head *head; 2980affee4b2STejun Heo 2981dc186ad7SThomas Gleixner /* 2982d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 2983dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2984dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2985dc186ad7SThomas Gleixner * might deadlock. 2986dc186ad7SThomas Gleixner */ 2987ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 298822df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 298952fa5bc5SBoqun Feng 2990fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 2991fd1a5b04SByungchul Park 29922607d7a6STejun Heo barr->task = current; 299383c22520SOleg Nesterov 2994018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 2995018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 2996018f3a13SLai Jiangshan 2997affee4b2STejun Heo /* 2998affee4b2STejun Heo * If @target is currently being executed, schedule the 2999affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 3000affee4b2STejun Heo */ 3001d812796eSLai Jiangshan if (worker) { 3002affee4b2STejun Heo head = worker->scheduled.next; 3003d812796eSLai Jiangshan work_color = worker->current_color; 3004d812796eSLai Jiangshan } else { 3005affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 3006affee4b2STejun Heo 3007affee4b2STejun Heo head = target->entry.next; 3008affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3009d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3010d812796eSLai Jiangshan work_color = get_work_color(*bits); 3011affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3012affee4b2STejun Heo } 3013affee4b2STejun Heo 3014d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3015d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3016d812796eSLai Jiangshan 3017d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3018fc2e4d70SOleg Nesterov } 3019fc2e4d70SOleg Nesterov 302073f53c4aSTejun Heo /** 3021112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 302273f53c4aSTejun Heo * @wq: workqueue being flushed 302373f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 302473f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 302573f53c4aSTejun Heo * 3026112202d9STejun Heo * Prepare pwqs for workqueue flushing. 302773f53c4aSTejun Heo * 3028112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3029112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3030112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3031112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3032112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 303373f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 303473f53c4aSTejun Heo * 303573f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 303673f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 303773f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 303873f53c4aSTejun Heo * is returned. 303973f53c4aSTejun Heo * 3040112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 304173f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 304273f53c4aSTejun Heo * advanced to @work_color. 304373f53c4aSTejun Heo * 304473f53c4aSTejun Heo * CONTEXT: 30453c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 304673f53c4aSTejun Heo * 3047d185af30SYacine Belkadi * Return: 304873f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 304973f53c4aSTejun Heo * otherwise. 305073f53c4aSTejun Heo */ 3051112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 305273f53c4aSTejun Heo int flush_color, int work_color) 30531da177e4SLinus Torvalds { 305473f53c4aSTejun Heo bool wait = false; 305549e3cf44STejun Heo struct pool_workqueue *pwq; 30561da177e4SLinus Torvalds 305773f53c4aSTejun Heo if (flush_color >= 0) { 30586183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3059112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3060dc186ad7SThomas Gleixner } 306114441960SOleg Nesterov 306249e3cf44STejun Heo for_each_pwq(pwq, wq) { 3063112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30641da177e4SLinus Torvalds 3065a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 306673f53c4aSTejun Heo 306773f53c4aSTejun Heo if (flush_color >= 0) { 30686183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 306973f53c4aSTejun Heo 3070112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3071112202d9STejun Heo pwq->flush_color = flush_color; 3072112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 307373f53c4aSTejun Heo wait = true; 30741da177e4SLinus Torvalds } 307573f53c4aSTejun Heo } 307673f53c4aSTejun Heo 307773f53c4aSTejun Heo if (work_color >= 0) { 30786183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3079112202d9STejun Heo pwq->work_color = work_color; 308073f53c4aSTejun Heo } 308173f53c4aSTejun Heo 3082a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 30831da177e4SLinus Torvalds } 30841da177e4SLinus Torvalds 3085112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 308673f53c4aSTejun Heo complete(&wq->first_flusher->done); 308773f53c4aSTejun Heo 308873f53c4aSTejun Heo return wait; 308983c22520SOleg Nesterov } 30901da177e4SLinus Torvalds 30910fcb78c2SRolf Eike Beer /** 3092c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 30930fcb78c2SRolf Eike Beer * @wq: workqueue to flush 30941da177e4SLinus Torvalds * 3095c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3096c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 30971da177e4SLinus Torvalds */ 3098c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 30991da177e4SLinus Torvalds { 310073f53c4aSTejun Heo struct wq_flusher this_flusher = { 310173f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 310273f53c4aSTejun Heo .flush_color = -1, 3103fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 310473f53c4aSTejun Heo }; 310573f53c4aSTejun Heo int next_color; 3106b1f4ec17SOleg Nesterov 31073347fa09STejun Heo if (WARN_ON(!wq_online)) 31083347fa09STejun Heo return; 31093347fa09STejun Heo 311087915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 311187915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 311287915adcSJohannes Berg 31133c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 311473f53c4aSTejun Heo 311573f53c4aSTejun Heo /* 311673f53c4aSTejun Heo * Start-to-wait phase 311773f53c4aSTejun Heo */ 311873f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 311973f53c4aSTejun Heo 312073f53c4aSTejun Heo if (next_color != wq->flush_color) { 312173f53c4aSTejun Heo /* 312273f53c4aSTejun Heo * Color space is not full. The current work_color 312373f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 312473f53c4aSTejun Heo * by one. 312573f53c4aSTejun Heo */ 31266183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 312773f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 312873f53c4aSTejun Heo wq->work_color = next_color; 312973f53c4aSTejun Heo 313073f53c4aSTejun Heo if (!wq->first_flusher) { 313173f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31326183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 313373f53c4aSTejun Heo 313473f53c4aSTejun Heo wq->first_flusher = &this_flusher; 313573f53c4aSTejun Heo 3136112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 313773f53c4aSTejun Heo wq->work_color)) { 313873f53c4aSTejun Heo /* nothing to flush, done */ 313973f53c4aSTejun Heo wq->flush_color = next_color; 314073f53c4aSTejun Heo wq->first_flusher = NULL; 314173f53c4aSTejun Heo goto out_unlock; 314273f53c4aSTejun Heo } 314373f53c4aSTejun Heo } else { 314473f53c4aSTejun Heo /* wait in queue */ 31456183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 314673f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3147112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 314873f53c4aSTejun Heo } 314973f53c4aSTejun Heo } else { 315073f53c4aSTejun Heo /* 315173f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 315273f53c4aSTejun Heo * The next flush completion will assign us 315373f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 315473f53c4aSTejun Heo */ 315573f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 315673f53c4aSTejun Heo } 315773f53c4aSTejun Heo 3158fca839c0STejun Heo check_flush_dependency(wq, NULL); 3159fca839c0STejun Heo 31603c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 316173f53c4aSTejun Heo 316273f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 316373f53c4aSTejun Heo 316473f53c4aSTejun Heo /* 316573f53c4aSTejun Heo * Wake-up-and-cascade phase 316673f53c4aSTejun Heo * 316773f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 316873f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 316973f53c4aSTejun Heo */ 317000d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 317173f53c4aSTejun Heo return; 317273f53c4aSTejun Heo 31733c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 317473f53c4aSTejun Heo 31754ce48b37STejun Heo /* we might have raced, check again with mutex held */ 31764ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 31774ce48b37STejun Heo goto out_unlock; 31784ce48b37STejun Heo 317900d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 318073f53c4aSTejun Heo 31816183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 31826183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 318373f53c4aSTejun Heo 318473f53c4aSTejun Heo while (true) { 318573f53c4aSTejun Heo struct wq_flusher *next, *tmp; 318673f53c4aSTejun Heo 318773f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 318873f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 318973f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 319073f53c4aSTejun Heo break; 319173f53c4aSTejun Heo list_del_init(&next->list); 319273f53c4aSTejun Heo complete(&next->done); 319373f53c4aSTejun Heo } 319473f53c4aSTejun Heo 31956183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 319673f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 319773f53c4aSTejun Heo 319873f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 319973f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 320073f53c4aSTejun Heo 320173f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 320273f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 320373f53c4aSTejun Heo /* 320473f53c4aSTejun Heo * Assign the same color to all overflowed 320573f53c4aSTejun Heo * flushers, advance work_color and append to 320673f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 320773f53c4aSTejun Heo * phase for these overflowed flushers. 320873f53c4aSTejun Heo */ 320973f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 321073f53c4aSTejun Heo tmp->flush_color = wq->work_color; 321173f53c4aSTejun Heo 321273f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 321373f53c4aSTejun Heo 321473f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 321573f53c4aSTejun Heo &wq->flusher_queue); 3216112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 321773f53c4aSTejun Heo } 321873f53c4aSTejun Heo 321973f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32206183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 322173f53c4aSTejun Heo break; 322273f53c4aSTejun Heo } 322373f53c4aSTejun Heo 322473f53c4aSTejun Heo /* 322573f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3226112202d9STejun Heo * the new first flusher and arm pwqs. 322773f53c4aSTejun Heo */ 32286183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32296183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 323073f53c4aSTejun Heo 323173f53c4aSTejun Heo list_del_init(&next->list); 323273f53c4aSTejun Heo wq->first_flusher = next; 323373f53c4aSTejun Heo 3234112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 323573f53c4aSTejun Heo break; 323673f53c4aSTejun Heo 323773f53c4aSTejun Heo /* 323873f53c4aSTejun Heo * Meh... this color is already done, clear first 323973f53c4aSTejun Heo * flusher and repeat cascading. 324073f53c4aSTejun Heo */ 324173f53c4aSTejun Heo wq->first_flusher = NULL; 324273f53c4aSTejun Heo } 324373f53c4aSTejun Heo 324473f53c4aSTejun Heo out_unlock: 32453c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32461da177e4SLinus Torvalds } 3247c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32481da177e4SLinus Torvalds 32499c5a2ba7STejun Heo /** 32509c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32519c5a2ba7STejun Heo * @wq: workqueue to drain 32529c5a2ba7STejun Heo * 32539c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32549c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32559c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3256b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32579c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32589c5a2ba7STejun Heo * takes too long. 32599c5a2ba7STejun Heo */ 32609c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32619c5a2ba7STejun Heo { 32629c5a2ba7STejun Heo unsigned int flush_cnt = 0; 326349e3cf44STejun Heo struct pool_workqueue *pwq; 32649c5a2ba7STejun Heo 32659c5a2ba7STejun Heo /* 32669c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 32679c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3268618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 32699c5a2ba7STejun Heo */ 327087fc741eSLai Jiangshan mutex_lock(&wq->mutex); 32719c5a2ba7STejun Heo if (!wq->nr_drainers++) 3272618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 327387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 32749c5a2ba7STejun Heo reflush: 3275c4f135d6STetsuo Handa __flush_workqueue(wq); 32769c5a2ba7STejun Heo 3277b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 327876af4d93STejun Heo 327949e3cf44STejun Heo for_each_pwq(pwq, wq) { 3280fa2563e4SThomas Tuttle bool drained; 32819c5a2ba7STejun Heo 3282a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3283f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3284a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3285fa2563e4SThomas Tuttle 3286fa2563e4SThomas Tuttle if (drained) 32879c5a2ba7STejun Heo continue; 32889c5a2ba7STejun Heo 32899c5a2ba7STejun Heo if (++flush_cnt == 10 || 32909c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3291e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3292e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 329376af4d93STejun Heo 3294b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 32959c5a2ba7STejun Heo goto reflush; 32969c5a2ba7STejun Heo } 32979c5a2ba7STejun Heo 32989c5a2ba7STejun Heo if (!--wq->nr_drainers) 3299618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 330087fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33019c5a2ba7STejun Heo } 33029c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 33039c5a2ba7STejun Heo 3304d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3305d6e89786SJohannes Berg bool from_cancel) 3306baf59022STejun Heo { 3307baf59022STejun Heo struct worker *worker = NULL; 3308c9e7cf27STejun Heo struct worker_pool *pool; 3309112202d9STejun Heo struct pool_workqueue *pwq; 3310baf59022STejun Heo 3311baf59022STejun Heo might_sleep(); 3312baf59022STejun Heo 331324acfb71SThomas Gleixner rcu_read_lock(); 3314fa1b54e6STejun Heo pool = get_work_pool(work); 3315fa1b54e6STejun Heo if (!pool) { 331624acfb71SThomas Gleixner rcu_read_unlock(); 3317fa1b54e6STejun Heo return false; 3318fa1b54e6STejun Heo } 3319fa1b54e6STejun Heo 3320a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33210b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3322112202d9STejun Heo pwq = get_work_pwq(work); 3323112202d9STejun Heo if (pwq) { 3324112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3325baf59022STejun Heo goto already_gone; 3326606a5020STejun Heo } else { 3327c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3328baf59022STejun Heo if (!worker) 3329baf59022STejun Heo goto already_gone; 3330112202d9STejun Heo pwq = worker->current_pwq; 3331606a5020STejun Heo } 3332baf59022STejun Heo 3333fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3334fca839c0STejun Heo 3335112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3336a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3337baf59022STejun Heo 3338e159489bSTejun Heo /* 3339a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3340a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3341a1d14934SPeter Zijlstra * 3342a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3343a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3344a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3345a1d14934SPeter Zijlstra * forward progress. 3346e159489bSTejun Heo */ 3347d6e89786SJohannes Berg if (!from_cancel && 3348d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3349112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3350112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3351a1d14934SPeter Zijlstra } 335224acfb71SThomas Gleixner rcu_read_unlock(); 3353baf59022STejun Heo return true; 3354baf59022STejun Heo already_gone: 3355a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 335624acfb71SThomas Gleixner rcu_read_unlock(); 3357baf59022STejun Heo return false; 3358baf59022STejun Heo } 3359baf59022STejun Heo 3360d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3361d6e89786SJohannes Berg { 3362d6e89786SJohannes Berg struct wq_barrier barr; 3363d6e89786SJohannes Berg 3364d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3365d6e89786SJohannes Berg return false; 3366d6e89786SJohannes Berg 33674d43d395STetsuo Handa if (WARN_ON(!work->func)) 33684d43d395STetsuo Handa return false; 33694d43d395STetsuo Handa 337087915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 337187915adcSJohannes Berg lock_map_release(&work->lockdep_map); 337287915adcSJohannes Berg 3373d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3374d6e89786SJohannes Berg wait_for_completion(&barr.done); 3375d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3376d6e89786SJohannes Berg return true; 3377d6e89786SJohannes Berg } else { 3378d6e89786SJohannes Berg return false; 3379d6e89786SJohannes Berg } 3380d6e89786SJohannes Berg } 3381d6e89786SJohannes Berg 3382db700897SOleg Nesterov /** 3383401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3384401a8d04STejun Heo * @work: the work to flush 3385db700897SOleg Nesterov * 3386606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3387606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3388401a8d04STejun Heo * 3389d185af30SYacine Belkadi * Return: 3390401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3391401a8d04STejun Heo * %false if it was already idle. 3392db700897SOleg Nesterov */ 3393401a8d04STejun Heo bool flush_work(struct work_struct *work) 3394db700897SOleg Nesterov { 3395d6e89786SJohannes Berg return __flush_work(work, false); 3396606a5020STejun Heo } 3397db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3398db700897SOleg Nesterov 33998603e1b3STejun Heo struct cwt_wait { 3400ac6424b9SIngo Molnar wait_queue_entry_t wait; 34018603e1b3STejun Heo struct work_struct *work; 34028603e1b3STejun Heo }; 34038603e1b3STejun Heo 3404ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 34058603e1b3STejun Heo { 34068603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34078603e1b3STejun Heo 34088603e1b3STejun Heo if (cwait->work != key) 34098603e1b3STejun Heo return 0; 34108603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34118603e1b3STejun Heo } 34128603e1b3STejun Heo 341336e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3414401a8d04STejun Heo { 34158603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3416bbb68dfaSTejun Heo unsigned long flags; 34171f1f642eSOleg Nesterov int ret; 34181f1f642eSOleg Nesterov 34191f1f642eSOleg Nesterov do { 3420bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3421bbb68dfaSTejun Heo /* 34228603e1b3STejun Heo * If someone else is already canceling, wait for it to 34238603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34248603e1b3STejun Heo * because we may get scheduled between @work's completion 34258603e1b3STejun Heo * and the other canceling task resuming and clearing 34268603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34278603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34288603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34298603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34308603e1b3STejun Heo * we're hogging the CPU. 34318603e1b3STejun Heo * 34328603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34338603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34348603e1b3STejun Heo * wake function which matches @work along with exclusive 34358603e1b3STejun Heo * wait and wakeup. 3436bbb68dfaSTejun Heo */ 34378603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34388603e1b3STejun Heo struct cwt_wait cwait; 34398603e1b3STejun Heo 34408603e1b3STejun Heo init_wait(&cwait.wait); 34418603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34428603e1b3STejun Heo cwait.work = work; 34438603e1b3STejun Heo 34448603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34458603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34468603e1b3STejun Heo if (work_is_canceling(work)) 34478603e1b3STejun Heo schedule(); 34488603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34498603e1b3STejun Heo } 34501f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34511f1f642eSOleg Nesterov 3452bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3453bbb68dfaSTejun Heo mark_work_canceling(work); 3454bbb68dfaSTejun Heo local_irq_restore(flags); 3455bbb68dfaSTejun Heo 34563347fa09STejun Heo /* 34573347fa09STejun Heo * This allows canceling during early boot. We know that @work 34583347fa09STejun Heo * isn't executing. 34593347fa09STejun Heo */ 34603347fa09STejun Heo if (wq_online) 3461d6e89786SJohannes Berg __flush_work(work, true); 34623347fa09STejun Heo 34637a22ad75STejun Heo clear_work_data(work); 34648603e1b3STejun Heo 34658603e1b3STejun Heo /* 34668603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 34678603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 34688603e1b3STejun Heo * visible there. 34698603e1b3STejun Heo */ 34708603e1b3STejun Heo smp_mb(); 34718603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 34728603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 34738603e1b3STejun Heo 34741f1f642eSOleg Nesterov return ret; 34751f1f642eSOleg Nesterov } 34761f1f642eSOleg Nesterov 34776e84d644SOleg Nesterov /** 3478401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3479401a8d04STejun Heo * @work: the work to cancel 34806e84d644SOleg Nesterov * 3481401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3482401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3483401a8d04STejun Heo * another workqueue. On return from this function, @work is 3484401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 34851f1f642eSOleg Nesterov * 3486401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3487401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 34886e84d644SOleg Nesterov * 3489401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 34906e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3491401a8d04STejun Heo * 3492d185af30SYacine Belkadi * Return: 3493401a8d04STejun Heo * %true if @work was pending, %false otherwise. 34946e84d644SOleg Nesterov */ 3495401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 34966e84d644SOleg Nesterov { 349736e227d2STejun Heo return __cancel_work_timer(work, false); 3498b89deed3SOleg Nesterov } 349928e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3500b89deed3SOleg Nesterov 35016e84d644SOleg Nesterov /** 3502401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3503401a8d04STejun Heo * @dwork: the delayed work to flush 35046e84d644SOleg Nesterov * 3505401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3506401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3507401a8d04STejun Heo * considers the last queueing instance of @dwork. 35081f1f642eSOleg Nesterov * 3509d185af30SYacine Belkadi * Return: 3510401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3511401a8d04STejun Heo * %false if it was already idle. 35126e84d644SOleg Nesterov */ 3513401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3514401a8d04STejun Heo { 35158930cabaSTejun Heo local_irq_disable(); 3516401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 351760c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35188930cabaSTejun Heo local_irq_enable(); 3519401a8d04STejun Heo return flush_work(&dwork->work); 3520401a8d04STejun Heo } 3521401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3522401a8d04STejun Heo 352305f0fe6bSTejun Heo /** 352405f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 352505f0fe6bSTejun Heo * @rwork: the rcu work to flush 352605f0fe6bSTejun Heo * 352705f0fe6bSTejun Heo * Return: 352805f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 352905f0fe6bSTejun Heo * %false if it was already idle. 353005f0fe6bSTejun Heo */ 353105f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 353205f0fe6bSTejun Heo { 353305f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 353405f0fe6bSTejun Heo rcu_barrier(); 353505f0fe6bSTejun Heo flush_work(&rwork->work); 353605f0fe6bSTejun Heo return true; 353705f0fe6bSTejun Heo } else { 353805f0fe6bSTejun Heo return flush_work(&rwork->work); 353905f0fe6bSTejun Heo } 354005f0fe6bSTejun Heo } 354105f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 354205f0fe6bSTejun Heo 3543f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3544f72b8792SJens Axboe { 3545f72b8792SJens Axboe unsigned long flags; 3546f72b8792SJens Axboe int ret; 3547f72b8792SJens Axboe 3548f72b8792SJens Axboe do { 3549f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3550f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3551f72b8792SJens Axboe 3552f72b8792SJens Axboe if (unlikely(ret < 0)) 3553f72b8792SJens Axboe return false; 3554f72b8792SJens Axboe 3555f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3556f72b8792SJens Axboe local_irq_restore(flags); 3557f72b8792SJens Axboe return ret; 3558f72b8792SJens Axboe } 3559f72b8792SJens Axboe 356073b4b532SAndrey Grodzovsky /* 356173b4b532SAndrey Grodzovsky * See cancel_delayed_work() 356273b4b532SAndrey Grodzovsky */ 356373b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 356473b4b532SAndrey Grodzovsky { 356573b4b532SAndrey Grodzovsky return __cancel_work(work, false); 356673b4b532SAndrey Grodzovsky } 356773b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 356873b4b532SAndrey Grodzovsky 3569401a8d04STejun Heo /** 357057b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 357157b30ae7STejun Heo * @dwork: delayed_work to cancel 357209383498STejun Heo * 3573d185af30SYacine Belkadi * Kill off a pending delayed_work. 3574d185af30SYacine Belkadi * 3575d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3576d185af30SYacine Belkadi * pending. 3577d185af30SYacine Belkadi * 3578d185af30SYacine Belkadi * Note: 3579d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3580d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3581d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 358209383498STejun Heo * 358357b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 358409383498STejun Heo */ 358557b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 358609383498STejun Heo { 3587f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 358809383498STejun Heo } 358957b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 359009383498STejun Heo 359109383498STejun Heo /** 3592401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3593401a8d04STejun Heo * @dwork: the delayed work cancel 3594401a8d04STejun Heo * 3595401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3596401a8d04STejun Heo * 3597d185af30SYacine Belkadi * Return: 3598401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3599401a8d04STejun Heo */ 3600401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 36016e84d644SOleg Nesterov { 360236e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 36036e84d644SOleg Nesterov } 3604f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 36051da177e4SLinus Torvalds 36060fcb78c2SRolf Eike Beer /** 360731ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3608b6136773SAndrew Morton * @func: the function to call 3609b6136773SAndrew Morton * 361031ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 361131ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3612b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 361331ddd871STejun Heo * 3614d185af30SYacine Belkadi * Return: 361531ddd871STejun Heo * 0 on success, -errno on failure. 3616b6136773SAndrew Morton */ 361765f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 361815316ba8SChristoph Lameter { 361915316ba8SChristoph Lameter int cpu; 362038f51568SNamhyung Kim struct work_struct __percpu *works; 362115316ba8SChristoph Lameter 3622b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3623b6136773SAndrew Morton if (!works) 362415316ba8SChristoph Lameter return -ENOMEM; 3625b6136773SAndrew Morton 3626ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 362793981800STejun Heo 362815316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36299bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36309bfb1839SIngo Molnar 36319bfb1839SIngo Molnar INIT_WORK(work, func); 36328de6d308SOleg Nesterov schedule_work_on(cpu, work); 363315316ba8SChristoph Lameter } 363493981800STejun Heo 363593981800STejun Heo for_each_online_cpu(cpu) 36368616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 363793981800STejun Heo 3638ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3639b6136773SAndrew Morton free_percpu(works); 364015316ba8SChristoph Lameter return 0; 364115316ba8SChristoph Lameter } 364215316ba8SChristoph Lameter 3643eef6a7d5SAlan Stern /** 36441fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36451fa44ecaSJames Bottomley * @fn: the function to execute 36461fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36471fa44ecaSJames Bottomley * be available when the work executes) 36481fa44ecaSJames Bottomley * 36491fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36501fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36511fa44ecaSJames Bottomley * 3652d185af30SYacine Belkadi * Return: 0 - function was executed 36531fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36541fa44ecaSJames Bottomley */ 365565f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36561fa44ecaSJames Bottomley { 36571fa44ecaSJames Bottomley if (!in_interrupt()) { 365865f27f38SDavid Howells fn(&ew->work); 36591fa44ecaSJames Bottomley return 0; 36601fa44ecaSJames Bottomley } 36611fa44ecaSJames Bottomley 366265f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36631fa44ecaSJames Bottomley schedule_work(&ew->work); 36641fa44ecaSJames Bottomley 36651fa44ecaSJames Bottomley return 1; 36661fa44ecaSJames Bottomley } 36671fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 36681fa44ecaSJames Bottomley 36697a4e344cSTejun Heo /** 36707a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 36717a4e344cSTejun Heo * @attrs: workqueue_attrs to free 36727a4e344cSTejun Heo * 36737a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 36747a4e344cSTejun Heo */ 3675513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 36767a4e344cSTejun Heo { 36777a4e344cSTejun Heo if (attrs) { 36787a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 3679*9546b29eSTejun Heo free_cpumask_var(attrs->__pod_cpumask); 36807a4e344cSTejun Heo kfree(attrs); 36817a4e344cSTejun Heo } 36827a4e344cSTejun Heo } 36837a4e344cSTejun Heo 36847a4e344cSTejun Heo /** 36857a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 36867a4e344cSTejun Heo * 36877a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3688d185af30SYacine Belkadi * return it. 3689d185af30SYacine Belkadi * 3690d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 36917a4e344cSTejun Heo */ 3692513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 36937a4e344cSTejun Heo { 36947a4e344cSTejun Heo struct workqueue_attrs *attrs; 36957a4e344cSTejun Heo 3696be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 36977a4e344cSTejun Heo if (!attrs) 36987a4e344cSTejun Heo goto fail; 3699be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 37007a4e344cSTejun Heo goto fail; 3701*9546b29eSTejun Heo if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 3702*9546b29eSTejun Heo goto fail; 37037a4e344cSTejun Heo 370413e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 370563c5484eSTejun Heo attrs->affn_scope = wq_affn_dfl; 37067a4e344cSTejun Heo return attrs; 37077a4e344cSTejun Heo fail: 37087a4e344cSTejun Heo free_workqueue_attrs(attrs); 37097a4e344cSTejun Heo return NULL; 37107a4e344cSTejun Heo } 37117a4e344cSTejun Heo 371229c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 371329c91e99STejun Heo const struct workqueue_attrs *from) 371429c91e99STejun Heo { 371529c91e99STejun Heo to->nice = from->nice; 371629c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 3717*9546b29eSTejun Heo cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 371884193c07STejun Heo 37192865a8fbSShaohua Li /* 372084193c07STejun Heo * Unlike hash and equality test, copying shouldn't ignore wq-only 372184193c07STejun Heo * fields as copying is used for both pool and wq attrs. Instead, 372284193c07STejun Heo * get_unbound_pool() explicitly clears the fields. 37232865a8fbSShaohua Li */ 372484193c07STejun Heo to->affn_scope = from->affn_scope; 3725af73f5c9STejun Heo to->ordered = from->ordered; 372629c91e99STejun Heo } 372729c91e99STejun Heo 37285de7a03cSTejun Heo /* 37295de7a03cSTejun Heo * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 37305de7a03cSTejun Heo * comments in 'struct workqueue_attrs' definition. 37315de7a03cSTejun Heo */ 37325de7a03cSTejun Heo static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 37335de7a03cSTejun Heo { 373484193c07STejun Heo attrs->affn_scope = WQ_AFFN_NR_TYPES; 37355de7a03cSTejun Heo attrs->ordered = false; 37365de7a03cSTejun Heo } 37375de7a03cSTejun Heo 373829c91e99STejun Heo /* hash value of the content of @attr */ 373929c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 374029c91e99STejun Heo { 374129c91e99STejun Heo u32 hash = 0; 374229c91e99STejun Heo 374329c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 374413e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 374513e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 3746*9546b29eSTejun Heo hash = jhash(cpumask_bits(attrs->__pod_cpumask), 3747*9546b29eSTejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 374829c91e99STejun Heo return hash; 374929c91e99STejun Heo } 375029c91e99STejun Heo 375129c91e99STejun Heo /* content equality test */ 375229c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 375329c91e99STejun Heo const struct workqueue_attrs *b) 375429c91e99STejun Heo { 375529c91e99STejun Heo if (a->nice != b->nice) 375629c91e99STejun Heo return false; 375729c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 375829c91e99STejun Heo return false; 3759*9546b29eSTejun Heo if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 3760*9546b29eSTejun Heo return false; 376129c91e99STejun Heo return true; 376229c91e99STejun Heo } 376329c91e99STejun Heo 37640f36ee24STejun Heo /* Update @attrs with actually available CPUs */ 37650f36ee24STejun Heo static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 37660f36ee24STejun Heo const cpumask_t *unbound_cpumask) 37670f36ee24STejun Heo { 37680f36ee24STejun Heo /* 37690f36ee24STejun Heo * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 37700f36ee24STejun Heo * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 37710f36ee24STejun Heo * @unbound_cpumask. 37720f36ee24STejun Heo */ 37730f36ee24STejun Heo cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 37740f36ee24STejun Heo if (unlikely(cpumask_empty(attrs->cpumask))) 37750f36ee24STejun Heo cpumask_copy(attrs->cpumask, unbound_cpumask); 37760f36ee24STejun Heo } 37770f36ee24STejun Heo 377884193c07STejun Heo /* find wq_pod_type to use for @attrs */ 377984193c07STejun Heo static const struct wq_pod_type * 378084193c07STejun Heo wqattrs_pod_type(const struct workqueue_attrs *attrs) 378184193c07STejun Heo { 378284193c07STejun Heo struct wq_pod_type *pt = &wq_pod_types[attrs->affn_scope]; 378384193c07STejun Heo 378484193c07STejun Heo if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 378584193c07STejun Heo likely(pt->nr_pods)) 378684193c07STejun Heo return pt; 378784193c07STejun Heo 378884193c07STejun Heo /* 378984193c07STejun Heo * Before workqueue_init_topology(), only SYSTEM is available which is 379084193c07STejun Heo * initialized in workqueue_init_early(). 379184193c07STejun Heo */ 379284193c07STejun Heo pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 379384193c07STejun Heo BUG_ON(!pt->nr_pods); 379484193c07STejun Heo return pt; 379584193c07STejun Heo } 379684193c07STejun Heo 37977a4e344cSTejun Heo /** 37987a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 37997a4e344cSTejun Heo * @pool: worker_pool to initialize 38007a4e344cSTejun Heo * 3801402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3802d185af30SYacine Belkadi * 3803d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 380429c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 380529c91e99STejun Heo * on @pool safely to release it. 38067a4e344cSTejun Heo */ 38077a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 38084e1a1f9aSTejun Heo { 3809a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 381029c91e99STejun Heo pool->id = -1; 381129c91e99STejun Heo pool->cpu = -1; 3812f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 38134e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 381482607adcSTejun Heo pool->watchdog_ts = jiffies; 38154e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 38164e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 38174e1a1f9aSTejun Heo hash_init(pool->busy_hash); 38184e1a1f9aSTejun Heo 381932a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 38203f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 38214e1a1f9aSTejun Heo 382232a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 38234e1a1f9aSTejun Heo 3824da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3825e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 38267a4e344cSTejun Heo 38277cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 382829c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 382929c91e99STejun Heo pool->refcnt = 1; 383029c91e99STejun Heo 383129c91e99STejun Heo /* shouldn't fail above this point */ 3832be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 38337a4e344cSTejun Heo if (!pool->attrs) 38347a4e344cSTejun Heo return -ENOMEM; 38355de7a03cSTejun Heo 38365de7a03cSTejun Heo wqattrs_clear_for_pool(pool->attrs); 38375de7a03cSTejun Heo 38387a4e344cSTejun Heo return 0; 38394e1a1f9aSTejun Heo } 38404e1a1f9aSTejun Heo 3841669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3842669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3843669de8bdSBart Van Assche { 3844669de8bdSBart Van Assche char *lock_name; 3845669de8bdSBart Van Assche 3846669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3847669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3848669de8bdSBart Van Assche if (!lock_name) 3849669de8bdSBart Van Assche lock_name = wq->name; 385069a106c0SQian Cai 385169a106c0SQian Cai wq->lock_name = lock_name; 3852669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3853669de8bdSBart Van Assche } 3854669de8bdSBart Van Assche 3855669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3856669de8bdSBart Van Assche { 3857669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3858669de8bdSBart Van Assche } 3859669de8bdSBart Van Assche 3860669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3861669de8bdSBart Van Assche { 3862669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3863669de8bdSBart Van Assche kfree(wq->lock_name); 3864669de8bdSBart Van Assche } 3865669de8bdSBart Van Assche #else 3866669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3867669de8bdSBart Van Assche { 3868669de8bdSBart Van Assche } 3869669de8bdSBart Van Assche 3870669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3871669de8bdSBart Van Assche { 3872669de8bdSBart Van Assche } 3873669de8bdSBart Van Assche 3874669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3875669de8bdSBart Van Assche { 3876669de8bdSBart Van Assche } 3877669de8bdSBart Van Assche #endif 3878669de8bdSBart Van Assche 3879e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3880e2dca7adSTejun Heo { 3881e2dca7adSTejun Heo struct workqueue_struct *wq = 3882e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3883e2dca7adSTejun Heo 3884669de8bdSBart Van Assche wq_free_lockdep(wq); 3885ee1ceef7STejun Heo free_percpu(wq->cpu_pwq); 3886e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3887e2dca7adSTejun Heo kfree(wq); 3888e2dca7adSTejun Heo } 3889e2dca7adSTejun Heo 389029c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 389129c91e99STejun Heo { 389229c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 389329c91e99STejun Heo 38947cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 389529c91e99STejun Heo free_workqueue_attrs(pool->attrs); 389629c91e99STejun Heo kfree(pool); 389729c91e99STejun Heo } 389829c91e99STejun Heo 389929c91e99STejun Heo /** 390029c91e99STejun Heo * put_unbound_pool - put a worker_pool 390129c91e99STejun Heo * @pool: worker_pool to put 390229c91e99STejun Heo * 390324acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3904c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3905c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3906c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3907a892caccSTejun Heo * 3908a892caccSTejun Heo * Should be called with wq_pool_mutex held. 390929c91e99STejun Heo */ 391029c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 391129c91e99STejun Heo { 391260f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 391329c91e99STejun Heo struct worker *worker; 39149680540cSYang Yingliang LIST_HEAD(cull_list); 3915e02b9312SValentin Schneider 3916a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3917a892caccSTejun Heo 3918a892caccSTejun Heo if (--pool->refcnt) 391929c91e99STejun Heo return; 392029c91e99STejun Heo 392129c91e99STejun Heo /* sanity checks */ 392261d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3923a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 392429c91e99STejun Heo return; 392529c91e99STejun Heo 392629c91e99STejun Heo /* release id and unhash */ 392729c91e99STejun Heo if (pool->id >= 0) 392829c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 392929c91e99STejun Heo hash_del(&pool->hash_node); 393029c91e99STejun Heo 3931c5aa87bbSTejun Heo /* 3932692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3933692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3934692b4825STejun Heo * manager and @pool gets freed with the flag set. 39359ab03be4SValentin Schneider * 39369ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 39379ab03be4SValentin Schneider * only get here with 39389ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 39399ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 39409ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 39419ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 39429ab03be4SValentin Schneider * drops pool->lock 3943c5aa87bbSTejun Heo */ 39449ab03be4SValentin Schneider while (true) { 39459ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 39469ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3947d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3948e02b9312SValentin Schneider 3949e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 39509ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 39519ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 3952692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 39539ab03be4SValentin Schneider break; 39549ab03be4SValentin Schneider } 39559ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 3956e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 39579ab03be4SValentin Schneider } 3958692b4825STejun Heo 39591037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 3960e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 396129c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 3962a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 396360f5a4bcSLai Jiangshan 3964e02b9312SValentin Schneider wake_dying_workers(&cull_list); 3965e02b9312SValentin Schneider 3966e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 396760f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 39681258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 396960f5a4bcSLai Jiangshan 397060f5a4bcSLai Jiangshan if (pool->detach_completion) 397160f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 397260f5a4bcSLai Jiangshan 397329c91e99STejun Heo /* shut down the timers */ 397429c91e99STejun Heo del_timer_sync(&pool->idle_timer); 39753f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 397629c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 397729c91e99STejun Heo 397824acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 397925b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 398029c91e99STejun Heo } 398129c91e99STejun Heo 398229c91e99STejun Heo /** 398329c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 398429c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 398529c91e99STejun Heo * 398629c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 398729c91e99STejun Heo * reference count and return it. If there already is a matching 398829c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 3989d185af30SYacine Belkadi * create a new one. 3990a892caccSTejun Heo * 3991a892caccSTejun Heo * Should be called with wq_pool_mutex held. 3992d185af30SYacine Belkadi * 3993d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 3994d185af30SYacine Belkadi * On failure, %NULL. 399529c91e99STejun Heo */ 399629c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 399729c91e99STejun Heo { 399884193c07STejun Heo struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 399929c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 400029c91e99STejun Heo struct worker_pool *pool; 400184193c07STejun Heo int pod, node = NUMA_NO_NODE; 400229c91e99STejun Heo 4003a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 400429c91e99STejun Heo 400529c91e99STejun Heo /* do we already have a matching pool? */ 400629c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 400729c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 400829c91e99STejun Heo pool->refcnt++; 40093fb1823cSLai Jiangshan return pool; 401029c91e99STejun Heo } 401129c91e99STejun Heo } 401229c91e99STejun Heo 4013*9546b29eSTejun Heo /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 401484193c07STejun Heo for (pod = 0; pod < pt->nr_pods; pod++) { 4015*9546b29eSTejun Heo if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 401684193c07STejun Heo node = pt->pod_node[pod]; 4017e2273584SXunlei Pang break; 4018e2273584SXunlei Pang } 4019e2273584SXunlei Pang } 4020e2273584SXunlei Pang 402129c91e99STejun Heo /* nope, create a new one */ 402284193c07STejun Heo pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 402329c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 402429c91e99STejun Heo goto fail; 402529c91e99STejun Heo 402684193c07STejun Heo pool->node = node; 40275de7a03cSTejun Heo copy_workqueue_attrs(pool->attrs, attrs); 40285de7a03cSTejun Heo wqattrs_clear_for_pool(pool->attrs); 40292865a8fbSShaohua Li 403029c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 403129c91e99STejun Heo goto fail; 403229c91e99STejun Heo 403329c91e99STejun Heo /* create and start the initial worker */ 40343347fa09STejun Heo if (wq_online && !create_worker(pool)) 403529c91e99STejun Heo goto fail; 403629c91e99STejun Heo 403729c91e99STejun Heo /* install */ 403829c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 40393fb1823cSLai Jiangshan 404029c91e99STejun Heo return pool; 404129c91e99STejun Heo fail: 404229c91e99STejun Heo if (pool) 404329c91e99STejun Heo put_unbound_pool(pool); 404429c91e99STejun Heo return NULL; 404529c91e99STejun Heo } 404629c91e99STejun Heo 40478864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 40488864b4e5STejun Heo { 40498864b4e5STejun Heo kmem_cache_free(pwq_cache, 40508864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 40518864b4e5STejun Heo } 40528864b4e5STejun Heo 40538864b4e5STejun Heo /* 4054967b494eSTejun Heo * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 4055967b494eSTejun Heo * refcnt and needs to be destroyed. 40568864b4e5STejun Heo */ 4057687a9aa5STejun Heo static void pwq_release_workfn(struct kthread_work *work) 40588864b4e5STejun Heo { 40598864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 4060687a9aa5STejun Heo release_work); 40618864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 40628864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4063b42b0bddSYang Yingliang bool is_last = false; 40648864b4e5STejun Heo 4065b42b0bddSYang Yingliang /* 4066687a9aa5STejun Heo * When @pwq is not linked, it doesn't hold any reference to the 4067b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4068b42b0bddSYang Yingliang */ 4069b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 40703c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 40718864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4072bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 40733c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4074b42b0bddSYang Yingliang } 40758864b4e5STejun Heo 4076687a9aa5STejun Heo if (wq->flags & WQ_UNBOUND) { 4077a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 40788864b4e5STejun Heo put_unbound_pool(pool); 4079a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4080687a9aa5STejun Heo } 4081a892caccSTejun Heo 408225b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 40838864b4e5STejun Heo 40848864b4e5STejun Heo /* 40858864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4086e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 40878864b4e5STejun Heo */ 4088669de8bdSBart Van Assche if (is_last) { 4089669de8bdSBart Van Assche wq_unregister_lockdep(wq); 409025b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 40916029a918STejun Heo } 4092669de8bdSBart Van Assche } 40938864b4e5STejun Heo 40940fbd95aaSTejun Heo /** 4095699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 40960fbd95aaSTejun Heo * @pwq: target pool_workqueue 40970fbd95aaSTejun Heo * 4098699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4099f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4100699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 41010fbd95aaSTejun Heo */ 4102699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 41030fbd95aaSTejun Heo { 4104699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4105699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 41063347fa09STejun Heo unsigned long flags; 4107699ce097STejun Heo 4108699ce097STejun Heo /* for @wq->saved_max_active */ 4109a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4110699ce097STejun Heo 4111699ce097STejun Heo /* fast exit for non-freezable wqs */ 4112699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4113699ce097STejun Heo return; 4114699ce097STejun Heo 41153347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4116a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4117699ce097STejun Heo 411874b414eaSLai Jiangshan /* 411974b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 412074b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 412174b414eaSLai Jiangshan * is updated and visible. 412274b414eaSLai Jiangshan */ 412374b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 4124699ce097STejun Heo pwq->max_active = wq->saved_max_active; 41250fbd95aaSTejun Heo 4126f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 41270219a352STejun Heo pwq->nr_active < pwq->max_active) 4128f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 4129951a078aSLai Jiangshan 41300219a352STejun Heo kick_pool(pwq->pool); 4131699ce097STejun Heo } else { 4132699ce097STejun Heo pwq->max_active = 0; 4133699ce097STejun Heo } 4134699ce097STejun Heo 4135a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 41360fbd95aaSTejun Heo } 41370fbd95aaSTejun Heo 413867dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4139f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4140f147f29eSTejun Heo struct worker_pool *pool) 4141d2c1d404STejun Heo { 4142d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4143d2c1d404STejun Heo 4144e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4145e50aba9aSTejun Heo 4146d2c1d404STejun Heo pwq->pool = pool; 4147d2c1d404STejun Heo pwq->wq = wq; 4148d2c1d404STejun Heo pwq->flush_color = -1; 41498864b4e5STejun Heo pwq->refcnt = 1; 4150f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41511befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4152d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 4153687a9aa5STejun Heo kthread_init_work(&pwq->release_work, pwq_release_workfn); 4154f147f29eSTejun Heo } 4155d2c1d404STejun Heo 4156f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 41571befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4158f147f29eSTejun Heo { 4159f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4160f147f29eSTejun Heo 4161f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 416275ccf595STejun Heo 41631befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 41641befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 41651befcf30STejun Heo return; 41661befcf30STejun Heo 416729b1cb41SLai Jiangshan /* set the matching work_color */ 416875ccf595STejun Heo pwq->work_color = wq->work_color; 4169983ca25eSTejun Heo 4170983ca25eSTejun Heo /* sync max_active to the current setting */ 4171983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4172983ca25eSTejun Heo 4173983ca25eSTejun Heo /* link in @pwq */ 41749e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4175df2d5ae4STejun Heo } 41766029a918STejun Heo 4177f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4178f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4179f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4180f147f29eSTejun Heo { 4181f147f29eSTejun Heo struct worker_pool *pool; 4182f147f29eSTejun Heo struct pool_workqueue *pwq; 4183f147f29eSTejun Heo 4184f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4185f147f29eSTejun Heo 4186f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4187f147f29eSTejun Heo if (!pool) 4188f147f29eSTejun Heo return NULL; 4189f147f29eSTejun Heo 4190e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4191f147f29eSTejun Heo if (!pwq) { 4192f147f29eSTejun Heo put_unbound_pool(pool); 4193f147f29eSTejun Heo return NULL; 4194f147f29eSTejun Heo } 4195f147f29eSTejun Heo 4196f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4197f147f29eSTejun Heo return pwq; 4198d2c1d404STejun Heo } 4199d2c1d404STejun Heo 42004c16bd32STejun Heo /** 4201fef59c9cSTejun Heo * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 4202042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 420384193c07STejun Heo * @cpu: the target CPU 42044c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 42054c16bd32STejun Heo * 4206fef59c9cSTejun Heo * Calculate the cpumask a workqueue with @attrs should use on @pod. If 4207fef59c9cSTejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during calculation. 4208*9546b29eSTejun Heo * The result is stored in @attrs->__pod_cpumask. 42094c16bd32STejun Heo * 4210fef59c9cSTejun Heo * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 4211fef59c9cSTejun Heo * and @pod has online CPUs requested by @attrs, the returned cpumask is the 4212fef59c9cSTejun Heo * intersection of the possible CPUs of @pod and @attrs->cpumask. 42134c16bd32STejun Heo * 4214fef59c9cSTejun Heo * The caller is responsible for ensuring that the cpumask of @pod stays stable. 42154c16bd32STejun Heo */ 4216*9546b29eSTejun Heo static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, 4217*9546b29eSTejun Heo int cpu_going_down) 42184c16bd32STejun Heo { 421984193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 422084193c07STejun Heo int pod = pt->cpu_pod[cpu]; 42214c16bd32STejun Heo 4222fef59c9cSTejun Heo /* does @pod have any online CPUs @attrs wants? */ 4223*9546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 4224*9546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); 42254c16bd32STejun Heo if (cpu_going_down >= 0) 4226*9546b29eSTejun Heo cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); 42274c16bd32STejun Heo 4228*9546b29eSTejun Heo if (cpumask_empty(attrs->__pod_cpumask)) { 4229*9546b29eSTejun Heo cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 423084193c07STejun Heo return; 423184193c07STejun Heo } 42324c16bd32STejun Heo 4233fef59c9cSTejun Heo /* yeap, return possible CPUs in @pod that @attrs wants */ 4234*9546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); 42351ad0f0a7SMichael Bringmann 4236*9546b29eSTejun Heo if (cpumask_empty(attrs->__pod_cpumask)) 42371ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42381ad0f0a7SMichael Bringmann "possible intersect\n"); 42394c16bd32STejun Heo } 42404c16bd32STejun Heo 4241636b927eSTejun Heo /* install @pwq into @wq's cpu_pwq and return the old pwq */ 4242636b927eSTejun Heo static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 4243636b927eSTejun Heo int cpu, struct pool_workqueue *pwq) 42441befcf30STejun Heo { 42451befcf30STejun Heo struct pool_workqueue *old_pwq; 42461befcf30STejun Heo 42475b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42481befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42491befcf30STejun Heo 42501befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42511befcf30STejun Heo link_pwq(pwq); 42521befcf30STejun Heo 4253636b927eSTejun Heo old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4254636b927eSTejun Heo rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); 42551befcf30STejun Heo return old_pwq; 42561befcf30STejun Heo } 42571befcf30STejun Heo 42582d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 42592d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 42602d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 42612d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4262042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 42632d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 42642d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 42652d5f0764SLai Jiangshan }; 42662d5f0764SLai Jiangshan 42672d5f0764SLai Jiangshan /* free the resources after success or abort */ 42682d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 42692d5f0764SLai Jiangshan { 42702d5f0764SLai Jiangshan if (ctx) { 4271636b927eSTejun Heo int cpu; 42722d5f0764SLai Jiangshan 4273636b927eSTejun Heo for_each_possible_cpu(cpu) 4274636b927eSTejun Heo put_pwq_unlocked(ctx->pwq_tbl[cpu]); 42752d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 42762d5f0764SLai Jiangshan 42772d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 42782d5f0764SLai Jiangshan 42792d5f0764SLai Jiangshan kfree(ctx); 42802d5f0764SLai Jiangshan } 42812d5f0764SLai Jiangshan } 42822d5f0764SLai Jiangshan 42832d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 42842d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 42852d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 428699c621efSLai Jiangshan const struct workqueue_attrs *attrs, 428799c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 42882d5f0764SLai Jiangshan { 42892d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4290*9546b29eSTejun Heo struct workqueue_attrs *new_attrs; 4291636b927eSTejun Heo int cpu; 42922d5f0764SLai Jiangshan 42932d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42942d5f0764SLai Jiangshan 429584193c07STejun Heo if (WARN_ON(attrs->affn_scope < 0 || 429684193c07STejun Heo attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 429784193c07STejun Heo return ERR_PTR(-EINVAL); 429884193c07STejun Heo 4299636b927eSTejun Heo ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 43002d5f0764SLai Jiangshan 4301be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 4302*9546b29eSTejun Heo if (!ctx || !new_attrs) 43032d5f0764SLai Jiangshan goto out_free; 43042d5f0764SLai Jiangshan 4305042f7df1SLai Jiangshan /* 43062d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 43072d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 43082d5f0764SLai Jiangshan * it even if we don't use it immediately. 43092d5f0764SLai Jiangshan */ 43100f36ee24STejun Heo copy_workqueue_attrs(new_attrs, attrs); 43110f36ee24STejun Heo wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 4312*9546b29eSTejun Heo cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 43132d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 43142d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 43152d5f0764SLai Jiangshan goto out_free; 43162d5f0764SLai Jiangshan 4317636b927eSTejun Heo for_each_possible_cpu(cpu) { 4318af73f5c9STejun Heo if (new_attrs->ordered) { 43192d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 4320636b927eSTejun Heo ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 4321636b927eSTejun Heo } else { 4322*9546b29eSTejun Heo wq_calc_pod_cpumask(new_attrs, cpu, -1); 4323*9546b29eSTejun Heo ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 4324636b927eSTejun Heo if (!ctx->pwq_tbl[cpu]) 4325636b927eSTejun Heo goto out_free; 43262d5f0764SLai Jiangshan } 43272d5f0764SLai Jiangshan } 43282d5f0764SLai Jiangshan 4329042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4330042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4331042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 4332*9546b29eSTejun Heo cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 43332d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4334042f7df1SLai Jiangshan 43352d5f0764SLai Jiangshan ctx->wq = wq; 43362d5f0764SLai Jiangshan return ctx; 43372d5f0764SLai Jiangshan 43382d5f0764SLai Jiangshan out_free: 43392d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43402d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 434184193c07STejun Heo return ERR_PTR(-ENOMEM); 43422d5f0764SLai Jiangshan } 43432d5f0764SLai Jiangshan 43442d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43452d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43462d5f0764SLai Jiangshan { 4347636b927eSTejun Heo int cpu; 43482d5f0764SLai Jiangshan 43492d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43502d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43512d5f0764SLai Jiangshan 43522d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 43532d5f0764SLai Jiangshan 43542d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 4355636b927eSTejun Heo for_each_possible_cpu(cpu) 4356636b927eSTejun Heo ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 4357636b927eSTejun Heo ctx->pwq_tbl[cpu]); 43582d5f0764SLai Jiangshan 43592d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 43602d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 43612d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 43622d5f0764SLai Jiangshan 43632d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 43642d5f0764SLai Jiangshan } 43652d5f0764SLai Jiangshan 4366a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4367a0111cf6SLai Jiangshan { 4368a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4369ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4370a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4371a0111cf6SLai Jiangshan } 4372a0111cf6SLai Jiangshan 4373a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4374a0111cf6SLai Jiangshan { 4375a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4376ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4377a0111cf6SLai Jiangshan } 4378a0111cf6SLai Jiangshan 4379a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4380a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4381a0111cf6SLai Jiangshan { 4382a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4383a0111cf6SLai Jiangshan 4384a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4385a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4386a0111cf6SLai Jiangshan return -EINVAL; 4387a0111cf6SLai Jiangshan 4388a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 43890a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 43900a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4391a0111cf6SLai Jiangshan return -EINVAL; 4392a0111cf6SLai Jiangshan 43930a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 43940a94efb5STejun Heo } 43950a94efb5STejun Heo 439699c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 439784193c07STejun Heo if (IS_ERR(ctx)) 439884193c07STejun Heo return PTR_ERR(ctx); 4399a0111cf6SLai Jiangshan 4400a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4401a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4402a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4403a0111cf6SLai Jiangshan 44046201171eSwanghaibin return 0; 4405a0111cf6SLai Jiangshan } 4406a0111cf6SLai Jiangshan 44079e8cd2f5STejun Heo /** 44089e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 44099e8cd2f5STejun Heo * @wq: the target workqueue 44109e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 44119e8cd2f5STejun Heo * 4412fef59c9cSTejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 4413fef59c9cSTejun Heo * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 4414fef59c9cSTejun Heo * work items are affine to the pod it was issued on. Older pwqs are released as 4415fef59c9cSTejun Heo * in-flight work items finish. Note that a work item which repeatedly requeues 4416fef59c9cSTejun Heo * itself back-to-back will stay on its current pwq. 44179e8cd2f5STejun Heo * 4418d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4419d185af30SYacine Belkadi * 4420ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4421509b3204SDaniel Jordan * 4422d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44239e8cd2f5STejun Heo */ 4424513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44259e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44269e8cd2f5STejun Heo { 4427a0111cf6SLai Jiangshan int ret; 44289e8cd2f5STejun Heo 4429509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4430509b3204SDaniel Jordan 4431509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4432a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4433509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44342d5f0764SLai Jiangshan 44352d5f0764SLai Jiangshan return ret; 44369e8cd2f5STejun Heo } 44379e8cd2f5STejun Heo 44384c16bd32STejun Heo /** 4439fef59c9cSTejun Heo * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug 44404c16bd32STejun Heo * @wq: the target workqueue 44414cbfd3deSTejun Heo * @cpu: the CPU to update pool association for 44424cbfd3deSTejun Heo * @hotplug_cpu: the CPU coming up or going down 44434c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44444c16bd32STejun Heo * 44454c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4446fef59c9cSTejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of 44474c16bd32STejun Heo * @wq accordingly. 44484c16bd32STejun Heo * 44494c16bd32STejun Heo * 4450fef59c9cSTejun Heo * If pod affinity can't be adjusted due to memory allocation failure, it falls 4451fef59c9cSTejun Heo * back to @wq->dfl_pwq which may not be optimal but is always correct. 4452fef59c9cSTejun Heo * 4453fef59c9cSTejun Heo * Note that when the last allowed CPU of a pod goes offline for a workqueue 4454fef59c9cSTejun Heo * with a cpumask spanning multiple pods, the workers which were already 4455fef59c9cSTejun Heo * executing the work items for the workqueue will lose their CPU affinity and 4456fef59c9cSTejun Heo * may execute on any CPU. This is similar to how per-cpu workqueues behave on 4457fef59c9cSTejun Heo * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 4458fef59c9cSTejun Heo * responsibility to flush the work item from CPU_DOWN_PREPARE. 44594c16bd32STejun Heo */ 4460fef59c9cSTejun Heo static void wq_update_pod(struct workqueue_struct *wq, int cpu, 44614cbfd3deSTejun Heo int hotplug_cpu, bool online) 44624c16bd32STejun Heo { 44634cbfd3deSTejun Heo int off_cpu = online ? -1 : hotplug_cpu; 44644c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 44654c16bd32STejun Heo struct workqueue_attrs *target_attrs; 44664c16bd32STejun Heo 44674c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 44684c16bd32STejun Heo 446984193c07STejun Heo if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 44704c16bd32STejun Heo return; 44714c16bd32STejun Heo 44724c16bd32STejun Heo /* 44734c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 44744c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 44754c16bd32STejun Heo * CPU hotplug exclusion. 44764c16bd32STejun Heo */ 4477fef59c9cSTejun Heo target_attrs = wq_update_pod_attrs_buf; 44784c16bd32STejun Heo 44794c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 44800f36ee24STejun Heo wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 44814c16bd32STejun Heo 4482636b927eSTejun Heo /* nothing to do if the target cpumask matches the current pwq */ 4483*9546b29eSTejun Heo wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); 4484636b927eSTejun Heo pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), 4485636b927eSTejun Heo lockdep_is_held(&wq_pool_mutex)); 4486*9546b29eSTejun Heo if (wqattrs_equal(target_attrs, pwq->pool->attrs)) 4487f7142ed4SLai Jiangshan return; 44884c16bd32STejun Heo 44894c16bd32STejun Heo /* create a new pwq */ 44904c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 44914c16bd32STejun Heo if (!pwq) { 4492fef59c9cSTejun Heo pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 44934c16bd32STejun Heo wq->name); 449477f300b1SDaeseok Youn goto use_dfl_pwq; 44954c16bd32STejun Heo } 44964c16bd32STejun Heo 4497f7142ed4SLai Jiangshan /* Install the new pwq. */ 44984c16bd32STejun Heo mutex_lock(&wq->mutex); 4499636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, pwq); 45004c16bd32STejun Heo goto out_unlock; 45014c16bd32STejun Heo 45024c16bd32STejun Heo use_dfl_pwq: 4503f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4504a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45054c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4506a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4507636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); 45084c16bd32STejun Heo out_unlock: 45094c16bd32STejun Heo mutex_unlock(&wq->mutex); 45104c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45114c16bd32STejun Heo } 45124c16bd32STejun Heo 451330cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45141da177e4SLinus Torvalds { 451549e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45168a2b7538STejun Heo int cpu, ret; 4517e1d8aa9fSFrederic Weisbecker 4518687a9aa5STejun Heo wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4519ee1ceef7STejun Heo if (!wq->cpu_pwq) 4520687a9aa5STejun Heo goto enomem; 452130cdf249STejun Heo 4522636b927eSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 452330cdf249STejun Heo for_each_possible_cpu(cpu) { 4524687a9aa5STejun Heo struct pool_workqueue **pwq_p = 4525ee1ceef7STejun Heo per_cpu_ptr(wq->cpu_pwq, cpu); 4526687a9aa5STejun Heo struct worker_pool *pool = 4527687a9aa5STejun Heo &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 452830cdf249STejun Heo 4529687a9aa5STejun Heo *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4530687a9aa5STejun Heo pool->node); 4531687a9aa5STejun Heo if (!*pwq_p) 4532687a9aa5STejun Heo goto enomem; 4533687a9aa5STejun Heo 4534687a9aa5STejun Heo init_pwq(*pwq_p, wq, pool); 4535f147f29eSTejun Heo 4536f147f29eSTejun Heo mutex_lock(&wq->mutex); 4537687a9aa5STejun Heo link_pwq(*pwq_p); 4538f147f29eSTejun Heo mutex_unlock(&wq->mutex); 453930cdf249STejun Heo } 454030cdf249STejun Heo return 0; 4541509b3204SDaniel Jordan } 4542509b3204SDaniel Jordan 4543ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4544509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45458a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45468a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45478a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45488a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45498a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45509e8cd2f5STejun Heo } else { 4551509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 45529e8cd2f5STejun Heo } 4553ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4554509b3204SDaniel Jordan 4555509b3204SDaniel Jordan return ret; 4556687a9aa5STejun Heo 4557687a9aa5STejun Heo enomem: 4558687a9aa5STejun Heo if (wq->cpu_pwq) { 4559687a9aa5STejun Heo for_each_possible_cpu(cpu) 4560687a9aa5STejun Heo kfree(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4561687a9aa5STejun Heo free_percpu(wq->cpu_pwq); 4562687a9aa5STejun Heo wq->cpu_pwq = NULL; 4563687a9aa5STejun Heo } 4564687a9aa5STejun Heo return -ENOMEM; 45650f900049STejun Heo } 45660f900049STejun Heo 4567f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4568f3421797STejun Heo const char *name) 4569b71ab8c2STejun Heo { 4570636b927eSTejun Heo if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 4571044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4572636b927eSTejun Heo max_active, name, 1, WQ_MAX_ACTIVE); 4573b71ab8c2STejun Heo 4574636b927eSTejun Heo return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 4575b71ab8c2STejun Heo } 4576b71ab8c2STejun Heo 4577983c7515STejun Heo /* 4578983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4579983c7515STejun Heo * to guarantee forward progress. 4580983c7515STejun Heo */ 4581983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4582983c7515STejun Heo { 4583983c7515STejun Heo struct worker *rescuer; 4584b92b36eaSDan Carpenter int ret; 4585983c7515STejun Heo 4586983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4587983c7515STejun Heo return 0; 4588983c7515STejun Heo 4589983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 45904c0736a7SPetr Mladek if (!rescuer) { 45914c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 45924c0736a7SPetr Mladek wq->name); 4593983c7515STejun Heo return -ENOMEM; 45944c0736a7SPetr Mladek } 4595983c7515STejun Heo 4596983c7515STejun Heo rescuer->rescue_wq = wq; 4597983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4598f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4599b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 46004c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 46014c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4602983c7515STejun Heo kfree(rescuer); 4603b92b36eaSDan Carpenter return ret; 4604983c7515STejun Heo } 4605983c7515STejun Heo 4606983c7515STejun Heo wq->rescuer = rescuer; 4607983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4608983c7515STejun Heo wake_up_process(rescuer->task); 4609983c7515STejun Heo 4610983c7515STejun Heo return 0; 4611983c7515STejun Heo } 4612983c7515STejun Heo 4613a2775bbcSMathieu Malaterre __printf(1, 4) 4614669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 461597e37d7bSTejun Heo unsigned int flags, 4616669de8bdSBart Van Assche int max_active, ...) 46173af24433SOleg Nesterov { 4618ecf6881fSTejun Heo va_list args; 46193af24433SOleg Nesterov struct workqueue_struct *wq; 462049e3cf44STejun Heo struct pool_workqueue *pwq; 4621b196be89STejun Heo 46225c0338c6STejun Heo /* 4623fef59c9cSTejun Heo * Unbound && max_active == 1 used to imply ordered, which is no longer 4624fef59c9cSTejun Heo * the case on many machines due to per-pod pools. While 46255c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 4626fef59c9cSTejun Heo * workqueue, keep the previous behavior to avoid subtle breakages. 46275c0338c6STejun Heo */ 46285c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46295c0338c6STejun Heo flags |= __WQ_ORDERED; 46305c0338c6STejun Heo 4631cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4632cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4633cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4634cee22a15SViresh Kumar 4635ecf6881fSTejun Heo /* allocate wq and format name */ 4636636b927eSTejun Heo wq = kzalloc(sizeof(*wq), GFP_KERNEL); 4637b196be89STejun Heo if (!wq) 4638d2c1d404STejun Heo return NULL; 4639b196be89STejun Heo 46406029a918STejun Heo if (flags & WQ_UNBOUND) { 4641be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46426029a918STejun Heo if (!wq->unbound_attrs) 46436029a918STejun Heo goto err_free_wq; 46446029a918STejun Heo } 46456029a918STejun Heo 4646669de8bdSBart Van Assche va_start(args, max_active); 4647ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4648b196be89STejun Heo va_end(args); 46493af24433SOleg Nesterov 4650d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4651b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 46523af24433SOleg Nesterov 4653b196be89STejun Heo /* init wq */ 465497e37d7bSTejun Heo wq->flags = flags; 4655a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 46563c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4657112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 465830cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 465973f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 466073f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4661493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 46623af24433SOleg Nesterov 4663669de8bdSBart Van Assche wq_init_lockdep(wq); 4664cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 46653af24433SOleg Nesterov 466630cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 466782efcab3SBart Van Assche goto err_unreg_lockdep; 46681537663fSTejun Heo 466940c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4670d2c1d404STejun Heo goto err_destroy; 4671e22bee78STejun Heo 4672226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4673226223abSTejun Heo goto err_destroy; 4674226223abSTejun Heo 46756af8bf3dSOleg Nesterov /* 467668e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 467768e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 467868e13a67SLai Jiangshan * list. 46796af8bf3dSOleg Nesterov */ 468068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4681a0a1a5fdSTejun Heo 4682a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 468349e3cf44STejun Heo for_each_pwq(pwq, wq) 4684699ce097STejun Heo pwq_adjust_max_active(pwq); 4685a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4686a0a1a5fdSTejun Heo 4687e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4688a0a1a5fdSTejun Heo 468968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 46903af24433SOleg Nesterov 46913af24433SOleg Nesterov return wq; 4692d2c1d404STejun Heo 469382efcab3SBart Van Assche err_unreg_lockdep: 4694009bb421SBart Van Assche wq_unregister_lockdep(wq); 4695009bb421SBart Van Assche wq_free_lockdep(wq); 469682efcab3SBart Van Assche err_free_wq: 46976029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 46984690c4abSTejun Heo kfree(wq); 4699d2c1d404STejun Heo return NULL; 4700d2c1d404STejun Heo err_destroy: 4701d2c1d404STejun Heo destroy_workqueue(wq); 47024690c4abSTejun Heo return NULL; 47031da177e4SLinus Torvalds } 4704669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47051da177e4SLinus Torvalds 4706c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4707c29eb853STejun Heo { 4708c29eb853STejun Heo int i; 4709c29eb853STejun Heo 4710c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4711c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4712c29eb853STejun Heo return true; 4713c29eb853STejun Heo 4714c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4715c29eb853STejun Heo return true; 4716f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4717c29eb853STejun Heo return true; 4718c29eb853STejun Heo 4719c29eb853STejun Heo return false; 4720c29eb853STejun Heo } 4721c29eb853STejun Heo 47223af24433SOleg Nesterov /** 47233af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47243af24433SOleg Nesterov * @wq: target workqueue 47253af24433SOleg Nesterov * 47263af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47273af24433SOleg Nesterov */ 47283af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47293af24433SOleg Nesterov { 473049e3cf44STejun Heo struct pool_workqueue *pwq; 4731636b927eSTejun Heo int cpu; 47323af24433SOleg Nesterov 4733def98c84STejun Heo /* 4734def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4735def98c84STejun Heo * lead to sysfs name conflicts. 4736def98c84STejun Heo */ 4737def98c84STejun Heo workqueue_sysfs_unregister(wq); 4738def98c84STejun Heo 473933e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 474033e3f0a3SRichard Clark mutex_lock(&wq->mutex); 474133e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 474233e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 474333e3f0a3SRichard Clark 47449c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47459c5a2ba7STejun Heo drain_workqueue(wq); 4746c8efcc25STejun Heo 4747def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4748def98c84STejun Heo if (wq->rescuer) { 4749def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4750def98c84STejun Heo 4751def98c84STejun Heo /* this prevents new queueing */ 4752a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4753def98c84STejun Heo wq->rescuer = NULL; 4754a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4755def98c84STejun Heo 4756def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4757def98c84STejun Heo kthread_stop(rescuer->task); 47588efe1223STejun Heo kfree(rescuer); 4759def98c84STejun Heo } 4760def98c84STejun Heo 4761c29eb853STejun Heo /* 4762c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4763c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4764c29eb853STejun Heo */ 4765c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4766b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 476749e3cf44STejun Heo for_each_pwq(pwq, wq) { 4768a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4769c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 47701d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4771e66b39afSTejun Heo __func__, wq->name); 4772c29eb853STejun Heo show_pwq(pwq); 4773a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4774b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4775c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 477655df0933SImran Khan show_one_workqueue(wq); 47776183c009STejun Heo return; 477876af4d93STejun Heo } 4779a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 478076af4d93STejun Heo } 4781b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 47826183c009STejun Heo 4783a0a1a5fdSTejun Heo /* 4784a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4785a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4786a0a1a5fdSTejun Heo */ 4787e2dca7adSTejun Heo list_del_rcu(&wq->list); 478868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47893af24433SOleg Nesterov 47908864b4e5STejun Heo /* 4791636b927eSTejun Heo * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 4792636b927eSTejun Heo * to put the base refs. @wq will be auto-destroyed from the last 4793636b927eSTejun Heo * pwq_put. RCU read lock prevents @wq from going away from under us. 47948864b4e5STejun Heo */ 4795636b927eSTejun Heo rcu_read_lock(); 4796636b927eSTejun Heo 4797636b927eSTejun Heo for_each_possible_cpu(cpu) { 4798636b927eSTejun Heo pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4799636b927eSTejun Heo RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); 48004c16bd32STejun Heo put_pwq_unlocked(pwq); 48014c16bd32STejun Heo } 48024c16bd32STejun Heo 4803636b927eSTejun Heo put_pwq_unlocked(wq->dfl_pwq); 48044c16bd32STejun Heo wq->dfl_pwq = NULL; 4805636b927eSTejun Heo 4806636b927eSTejun Heo rcu_read_unlock(); 48073af24433SOleg Nesterov } 48083af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48093af24433SOleg Nesterov 4810dcd989cbSTejun Heo /** 4811dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4812dcd989cbSTejun Heo * @wq: target workqueue 4813dcd989cbSTejun Heo * @max_active: new max_active value. 4814dcd989cbSTejun Heo * 4815dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4816dcd989cbSTejun Heo * 4817dcd989cbSTejun Heo * CONTEXT: 4818dcd989cbSTejun Heo * Don't call from IRQ context. 4819dcd989cbSTejun Heo */ 4820dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4821dcd989cbSTejun Heo { 482249e3cf44STejun Heo struct pool_workqueue *pwq; 4823dcd989cbSTejun Heo 48248719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48250a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48268719dceaSTejun Heo return; 48278719dceaSTejun Heo 4828f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4829dcd989cbSTejun Heo 4830a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4831dcd989cbSTejun Heo 48320a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4833dcd989cbSTejun Heo wq->saved_max_active = max_active; 4834dcd989cbSTejun Heo 4835699ce097STejun Heo for_each_pwq(pwq, wq) 4836699ce097STejun Heo pwq_adjust_max_active(pwq); 4837dcd989cbSTejun Heo 4838a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4839dcd989cbSTejun Heo } 4840dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4841dcd989cbSTejun Heo 4842dcd989cbSTejun Heo /** 484327d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 484427d4ee03SLukas Wunner * 484527d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 484627d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 484727d4ee03SLukas Wunner * 484827d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 484927d4ee03SLukas Wunner */ 485027d4ee03SLukas Wunner struct work_struct *current_work(void) 485127d4ee03SLukas Wunner { 485227d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 485327d4ee03SLukas Wunner 485427d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 485527d4ee03SLukas Wunner } 485627d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 485727d4ee03SLukas Wunner 485827d4ee03SLukas Wunner /** 4859e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4860e6267616STejun Heo * 4861e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4862e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4863d185af30SYacine Belkadi * 4864d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4865e6267616STejun Heo */ 4866e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4867e6267616STejun Heo { 4868e6267616STejun Heo struct worker *worker = current_wq_worker(); 4869e6267616STejun Heo 48706a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4871e6267616STejun Heo } 4872e6267616STejun Heo 4873e6267616STejun Heo /** 4874dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4875dcd989cbSTejun Heo * @cpu: CPU in question 4876dcd989cbSTejun Heo * @wq: target workqueue 4877dcd989cbSTejun Heo * 4878dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4879dcd989cbSTejun Heo * no synchronization around this function and the test result is 4880dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4881dcd989cbSTejun Heo * 4882d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4883636b927eSTejun Heo * 4884636b927eSTejun Heo * With the exception of ordered workqueues, all workqueues have per-cpu 4885636b927eSTejun Heo * pool_workqueues, each with its own congested state. A workqueue being 4886636b927eSTejun Heo * congested on one CPU doesn't mean that the workqueue is contested on any 4887636b927eSTejun Heo * other CPUs. 4888d3251859STejun Heo * 4889d185af30SYacine Belkadi * Return: 4890dcd989cbSTejun Heo * %true if congested, %false otherwise. 4891dcd989cbSTejun Heo */ 4892d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4893dcd989cbSTejun Heo { 48947fb98ea7STejun Heo struct pool_workqueue *pwq; 489576af4d93STejun Heo bool ret; 489676af4d93STejun Heo 489724acfb71SThomas Gleixner rcu_read_lock(); 489824acfb71SThomas Gleixner preempt_disable(); 48997fb98ea7STejun Heo 4900d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4901d3251859STejun Heo cpu = smp_processor_id(); 4902d3251859STejun Heo 4903687a9aa5STejun Heo pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4904f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 4905636b927eSTejun Heo 490624acfb71SThomas Gleixner preempt_enable(); 490724acfb71SThomas Gleixner rcu_read_unlock(); 490876af4d93STejun Heo 490976af4d93STejun Heo return ret; 4910dcd989cbSTejun Heo } 4911dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4912dcd989cbSTejun Heo 4913dcd989cbSTejun Heo /** 4914dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4915dcd989cbSTejun Heo * @work: the work to be tested 4916dcd989cbSTejun Heo * 4917dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4918dcd989cbSTejun Heo * synchronization around this function and the test result is 4919dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4920dcd989cbSTejun Heo * 4921d185af30SYacine Belkadi * Return: 4922dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4923dcd989cbSTejun Heo */ 4924dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4925dcd989cbSTejun Heo { 4926fa1b54e6STejun Heo struct worker_pool *pool; 4927dcd989cbSTejun Heo unsigned long flags; 4928dcd989cbSTejun Heo unsigned int ret = 0; 4929dcd989cbSTejun Heo 4930dcd989cbSTejun Heo if (work_pending(work)) 4931dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4932038366c5SLai Jiangshan 493324acfb71SThomas Gleixner rcu_read_lock(); 4934fa1b54e6STejun Heo pool = get_work_pool(work); 4935038366c5SLai Jiangshan if (pool) { 4936a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4937c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4938dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4939a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4940038366c5SLai Jiangshan } 494124acfb71SThomas Gleixner rcu_read_unlock(); 4942dcd989cbSTejun Heo 4943dcd989cbSTejun Heo return ret; 4944dcd989cbSTejun Heo } 4945dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4946dcd989cbSTejun Heo 49473d1cb205STejun Heo /** 49483d1cb205STejun Heo * set_worker_desc - set description for the current work item 49493d1cb205STejun Heo * @fmt: printf-style format string 49503d1cb205STejun Heo * @...: arguments for the format string 49513d1cb205STejun Heo * 49523d1cb205STejun Heo * This function can be called by a running work function to describe what 49533d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 49543d1cb205STejun Heo * information will be printed out together to help debugging. The 49553d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 49563d1cb205STejun Heo */ 49573d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 49583d1cb205STejun Heo { 49593d1cb205STejun Heo struct worker *worker = current_wq_worker(); 49603d1cb205STejun Heo va_list args; 49613d1cb205STejun Heo 49623d1cb205STejun Heo if (worker) { 49633d1cb205STejun Heo va_start(args, fmt); 49643d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 49653d1cb205STejun Heo va_end(args); 49663d1cb205STejun Heo } 49673d1cb205STejun Heo } 49685c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 49693d1cb205STejun Heo 49703d1cb205STejun Heo /** 49713d1cb205STejun Heo * print_worker_info - print out worker information and description 49723d1cb205STejun Heo * @log_lvl: the log level to use when printing 49733d1cb205STejun Heo * @task: target task 49743d1cb205STejun Heo * 49753d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 49763d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 49773d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 49783d1cb205STejun Heo * 49793d1cb205STejun Heo * This function can be safely called on any task as long as the 49803d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 49813d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 49823d1cb205STejun Heo */ 49833d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 49843d1cb205STejun Heo { 49853d1cb205STejun Heo work_func_t *fn = NULL; 49863d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 49873d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 49883d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 49893d1cb205STejun Heo struct workqueue_struct *wq = NULL; 49903d1cb205STejun Heo struct worker *worker; 49913d1cb205STejun Heo 49923d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 49933d1cb205STejun Heo return; 49943d1cb205STejun Heo 49953d1cb205STejun Heo /* 49963d1cb205STejun Heo * This function is called without any synchronization and @task 49973d1cb205STejun Heo * could be in any state. Be careful with dereferences. 49983d1cb205STejun Heo */ 4999e700591aSPetr Mladek worker = kthread_probe_data(task); 50003d1cb205STejun Heo 50013d1cb205STejun Heo /* 50028bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50038bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50043d1cb205STejun Heo */ 5005fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5006fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5007fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5008fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5009fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50103d1cb205STejun Heo 50113d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5012d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50138bf89593STejun Heo if (strcmp(name, desc)) 50143d1cb205STejun Heo pr_cont(" (%s)", desc); 50153d1cb205STejun Heo pr_cont("\n"); 50163d1cb205STejun Heo } 50173d1cb205STejun Heo } 50183d1cb205STejun Heo 50193494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50203494fc30STejun Heo { 50213494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50223494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50233494fc30STejun Heo pr_cont(" node=%d", pool->node); 50243494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50253494fc30STejun Heo } 50263494fc30STejun Heo 5027c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5028c76feb0dSPaul E. McKenney bool comma; 5029c76feb0dSPaul E. McKenney work_func_t func; 5030c76feb0dSPaul E. McKenney long ctr; 5031c76feb0dSPaul E. McKenney }; 5032c76feb0dSPaul E. McKenney 5033c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5034c76feb0dSPaul E. McKenney { 5035c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5036c76feb0dSPaul E. McKenney goto out_record; 5037c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5038c76feb0dSPaul E. McKenney pcwsp->ctr++; 5039c76feb0dSPaul E. McKenney return; 5040c76feb0dSPaul E. McKenney } 5041c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5042c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5043c76feb0dSPaul E. McKenney else 5044c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5045c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5046c76feb0dSPaul E. McKenney out_record: 5047c76feb0dSPaul E. McKenney if ((long)func == -1L) 5048c76feb0dSPaul E. McKenney return; 5049c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5050c76feb0dSPaul E. McKenney pcwsp->func = func; 5051c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5052c76feb0dSPaul E. McKenney } 5053c76feb0dSPaul E. McKenney 5054c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 50553494fc30STejun Heo { 50563494fc30STejun Heo if (work->func == wq_barrier_func) { 50573494fc30STejun Heo struct wq_barrier *barr; 50583494fc30STejun Heo 50593494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 50603494fc30STejun Heo 5061c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 50623494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 50633494fc30STejun Heo task_pid_nr(barr->task)); 50643494fc30STejun Heo } else { 5065c76feb0dSPaul E. McKenney if (!comma) 5066c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5067c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 50683494fc30STejun Heo } 50693494fc30STejun Heo } 50703494fc30STejun Heo 50713494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 50723494fc30STejun Heo { 5073c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 50743494fc30STejun Heo struct worker_pool *pool = pwq->pool; 50753494fc30STejun Heo struct work_struct *work; 50763494fc30STejun Heo struct worker *worker; 50773494fc30STejun Heo bool has_in_flight = false, has_pending = false; 50783494fc30STejun Heo int bkt; 50793494fc30STejun Heo 50803494fc30STejun Heo pr_info(" pwq %d:", pool->id); 50813494fc30STejun Heo pr_cont_pool_info(pool); 50823494fc30STejun Heo 5083e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5084e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 50853494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 50863494fc30STejun Heo 50873494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50883494fc30STejun Heo if (worker->current_pwq == pwq) { 50893494fc30STejun Heo has_in_flight = true; 50903494fc30STejun Heo break; 50913494fc30STejun Heo } 50923494fc30STejun Heo } 50933494fc30STejun Heo if (has_in_flight) { 50943494fc30STejun Heo bool comma = false; 50953494fc30STejun Heo 50963494fc30STejun Heo pr_info(" in-flight:"); 50973494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 50983494fc30STejun Heo if (worker->current_pwq != pwq) 50993494fc30STejun Heo continue; 51003494fc30STejun Heo 5101d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51023494fc30STejun Heo task_pid_nr(worker->task), 510330ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51043494fc30STejun Heo worker->current_func); 51053494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5106c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5107c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51083494fc30STejun Heo comma = true; 51093494fc30STejun Heo } 51103494fc30STejun Heo pr_cont("\n"); 51113494fc30STejun Heo } 51123494fc30STejun Heo 51133494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51143494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51153494fc30STejun Heo has_pending = true; 51163494fc30STejun Heo break; 51173494fc30STejun Heo } 51183494fc30STejun Heo } 51193494fc30STejun Heo if (has_pending) { 51203494fc30STejun Heo bool comma = false; 51213494fc30STejun Heo 51223494fc30STejun Heo pr_info(" pending:"); 51233494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51243494fc30STejun Heo if (get_work_pwq(work) != pwq) 51253494fc30STejun Heo continue; 51263494fc30STejun Heo 5127c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51283494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51293494fc30STejun Heo } 5130c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51313494fc30STejun Heo pr_cont("\n"); 51323494fc30STejun Heo } 51333494fc30STejun Heo 5134f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51353494fc30STejun Heo bool comma = false; 51363494fc30STejun Heo 5137f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5138f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5139c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51403494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51413494fc30STejun Heo } 5142c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51433494fc30STejun Heo pr_cont("\n"); 51443494fc30STejun Heo } 51453494fc30STejun Heo } 51463494fc30STejun Heo 51473494fc30STejun Heo /** 514855df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 514955df0933SImran Khan * @wq: workqueue whose state will be printed 51503494fc30STejun Heo */ 515155df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 51523494fc30STejun Heo { 51533494fc30STejun Heo struct pool_workqueue *pwq; 51543494fc30STejun Heo bool idle = true; 515555df0933SImran Khan unsigned long flags; 51563494fc30STejun Heo 51573494fc30STejun Heo for_each_pwq(pwq, wq) { 5158f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 51593494fc30STejun Heo idle = false; 51603494fc30STejun Heo break; 51613494fc30STejun Heo } 51623494fc30STejun Heo } 516355df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 516455df0933SImran Khan return; 51653494fc30STejun Heo 51663494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 51673494fc30STejun Heo 51683494fc30STejun Heo for_each_pwq(pwq, wq) { 5169a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 517057116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 517157116ce1SJohan Hovold /* 517257116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 517357116ce1SJohan Hovold * drivers that queue work while holding locks 517457116ce1SJohan Hovold * also taken in their write paths. 517557116ce1SJohan Hovold */ 517657116ce1SJohan Hovold printk_deferred_enter(); 51773494fc30STejun Heo show_pwq(pwq); 517857116ce1SJohan Hovold printk_deferred_exit(); 517957116ce1SJohan Hovold } 5180a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 518162635ea8SSergey Senozhatsky /* 518262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 518355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 518462635ea8SSergey Senozhatsky * hard lockup. 518562635ea8SSergey Senozhatsky */ 518662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 51873494fc30STejun Heo } 518855df0933SImran Khan 51893494fc30STejun Heo } 51903494fc30STejun Heo 519155df0933SImran Khan /** 519255df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 519355df0933SImran Khan * @pool: worker pool whose state will be printed 519455df0933SImran Khan */ 519555df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 519655df0933SImran Khan { 51973494fc30STejun Heo struct worker *worker; 51983494fc30STejun Heo bool first = true; 519955df0933SImran Khan unsigned long flags; 5200335a42ebSPetr Mladek unsigned long hung = 0; 52013494fc30STejun Heo 5202a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52033494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52043494fc30STejun Heo goto next_pool; 5205335a42ebSPetr Mladek 5206335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5207335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5208335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5209335a42ebSPetr Mladek 521057116ce1SJohan Hovold /* 521157116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 521257116ce1SJohan Hovold * queue work while holding locks also taken in their write 521357116ce1SJohan Hovold * paths. 521457116ce1SJohan Hovold */ 521557116ce1SJohan Hovold printk_deferred_enter(); 52163494fc30STejun Heo pr_info("pool %d:", pool->id); 52173494fc30STejun Heo pr_cont_pool_info(pool); 5218335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52193494fc30STejun Heo if (pool->manager) 52203494fc30STejun Heo pr_cont(" manager: %d", 52213494fc30STejun Heo task_pid_nr(pool->manager->task)); 52223494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52233494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52243494fc30STejun Heo task_pid_nr(worker->task)); 52253494fc30STejun Heo first = false; 52263494fc30STejun Heo } 52273494fc30STejun Heo pr_cont("\n"); 522857116ce1SJohan Hovold printk_deferred_exit(); 52293494fc30STejun Heo next_pool: 5230a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 523162635ea8SSergey Senozhatsky /* 523262635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 523355df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 523462635ea8SSergey Senozhatsky * hard lockup. 523562635ea8SSergey Senozhatsky */ 523662635ea8SSergey Senozhatsky touch_nmi_watchdog(); 523755df0933SImran Khan 52383494fc30STejun Heo } 52393494fc30STejun Heo 524055df0933SImran Khan /** 524155df0933SImran Khan * show_all_workqueues - dump workqueue state 524255df0933SImran Khan * 5243704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 524455df0933SImran Khan */ 524555df0933SImran Khan void show_all_workqueues(void) 524655df0933SImran Khan { 524755df0933SImran Khan struct workqueue_struct *wq; 524855df0933SImran Khan struct worker_pool *pool; 524955df0933SImran Khan int pi; 525055df0933SImran Khan 525155df0933SImran Khan rcu_read_lock(); 525255df0933SImran Khan 525355df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 525455df0933SImran Khan 525555df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 525655df0933SImran Khan show_one_workqueue(wq); 525755df0933SImran Khan 525855df0933SImran Khan for_each_pool(pool, pi) 525955df0933SImran Khan show_one_worker_pool(pool); 526055df0933SImran Khan 526124acfb71SThomas Gleixner rcu_read_unlock(); 52623494fc30STejun Heo } 52633494fc30STejun Heo 5264704bc669SJungseung Lee /** 5265704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5266704bc669SJungseung Lee * 5267704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5268704bc669SJungseung Lee * still busy. 5269704bc669SJungseung Lee */ 5270704bc669SJungseung Lee void show_freezable_workqueues(void) 5271704bc669SJungseung Lee { 5272704bc669SJungseung Lee struct workqueue_struct *wq; 5273704bc669SJungseung Lee 5274704bc669SJungseung Lee rcu_read_lock(); 5275704bc669SJungseung Lee 5276704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5277704bc669SJungseung Lee 5278704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5279704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5280704bc669SJungseung Lee continue; 5281704bc669SJungseung Lee show_one_workqueue(wq); 5282704bc669SJungseung Lee } 5283704bc669SJungseung Lee 5284704bc669SJungseung Lee rcu_read_unlock(); 5285704bc669SJungseung Lee } 5286704bc669SJungseung Lee 52876b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 52886b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 52896b59808bSTejun Heo { 52906b59808bSTejun Heo int off; 52916b59808bSTejun Heo 52926b59808bSTejun Heo /* always show the actual comm */ 52936b59808bSTejun Heo off = strscpy(buf, task->comm, size); 52946b59808bSTejun Heo if (off < 0) 52956b59808bSTejun Heo return; 52966b59808bSTejun Heo 5297197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 52986b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 52996b59808bSTejun Heo 5300197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5301197f6accSTejun Heo struct worker *worker = kthread_data(task); 5302197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53036b59808bSTejun Heo 53046b59808bSTejun Heo if (pool) { 5305a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53066b59808bSTejun Heo /* 5307197f6accSTejun Heo * ->desc tracks information (wq name or 5308197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5309197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53106b59808bSTejun Heo */ 53116b59808bSTejun Heo if (worker->desc[0] != '\0') { 53126b59808bSTejun Heo if (worker->current_work) 53136b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53146b59808bSTejun Heo worker->desc); 53156b59808bSTejun Heo else 53166b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53176b59808bSTejun Heo worker->desc); 53186b59808bSTejun Heo } 5319a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53206b59808bSTejun Heo } 5321197f6accSTejun Heo } 53226b59808bSTejun Heo 53236b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53246b59808bSTejun Heo } 53256b59808bSTejun Heo 532666448bc2SMathieu Malaterre #ifdef CONFIG_SMP 532766448bc2SMathieu Malaterre 5328db7bccf4STejun Heo /* 5329db7bccf4STejun Heo * CPU hotplug. 5330db7bccf4STejun Heo * 5331e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5332112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5333706026c2STejun Heo * pool which make migrating pending and scheduled works very 5334e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 533594cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5336e22bee78STejun Heo * blocked draining impractical. 5337e22bee78STejun Heo * 533824647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5339628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5340628c78e7STejun Heo * cpu comes back online. 5341db7bccf4STejun Heo */ 5342db7bccf4STejun Heo 5343e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5344db7bccf4STejun Heo { 53454ce62e9eSTejun Heo struct worker_pool *pool; 5346db7bccf4STejun Heo struct worker *worker; 5347db7bccf4STejun Heo 5348f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53491258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5350a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5351e22bee78STejun Heo 5352f2d5a0eeSTejun Heo /* 535392f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 535494cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 535511b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5356b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5357b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5358b4ac9384SLai Jiangshan * is on the same cpu. 5359f2d5a0eeSTejun Heo */ 5360da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5361403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5362db7bccf4STejun Heo 536324647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5364f2d5a0eeSTejun Heo 5365e22bee78STejun Heo /* 5366989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5367989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5368989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5369989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5370989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5371eb283428SLai Jiangshan * are served by workers tied to the pool. 5372e22bee78STejun Heo */ 5373bc35f7efSLai Jiangshan pool->nr_running = 0; 5374eb283428SLai Jiangshan 5375eb283428SLai Jiangshan /* 5376eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5377eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5378eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5379eb283428SLai Jiangshan */ 53800219a352STejun Heo kick_pool(pool); 5381989442d7SLai Jiangshan 5382a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5383989442d7SLai Jiangshan 5384793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5385793777bcSValentin Schneider unbind_worker(worker); 5386989442d7SLai Jiangshan 5387989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5388eb283428SLai Jiangshan } 5389db7bccf4STejun Heo } 5390db7bccf4STejun Heo 5391bd7c089eSTejun Heo /** 5392bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5393bd7c089eSTejun Heo * @pool: pool of interest 5394bd7c089eSTejun Heo * 5395a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5396bd7c089eSTejun Heo */ 5397bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5398bd7c089eSTejun Heo { 5399a9ab775bSTejun Heo struct worker *worker; 5400bd7c089eSTejun Heo 54011258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5402bd7c089eSTejun Heo 5403bd7c089eSTejun Heo /* 5404a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5405a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5406402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5407a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5408a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5409bd7c089eSTejun Heo */ 5410c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5411c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5412c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 5413*9546b29eSTejun Heo pool_allowed_cpus(pool)) < 0); 5414c63a2e52SValentin Schneider } 5415a9ab775bSTejun Heo 5416a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5417f7c17d26SWanpeng Li 54183de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5419a9ab775bSTejun Heo 5420da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5421a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5422a9ab775bSTejun Heo 5423a9ab775bSTejun Heo /* 5424a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5425a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5426a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5427a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5428a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5429a9ab775bSTejun Heo * concurrency management. Note that when or whether 5430a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5431a9ab775bSTejun Heo * 5432c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5433a9ab775bSTejun Heo * tested without holding any lock in 54346d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5435a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5436a9ab775bSTejun Heo * management operations. 5437bd7c089eSTejun Heo */ 5438a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5439a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5440a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5441c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5442bd7c089eSTejun Heo } 5443a9ab775bSTejun Heo 5444a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5445bd7c089eSTejun Heo } 5446bd7c089eSTejun Heo 54477dbc725eSTejun Heo /** 54487dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54497dbc725eSTejun Heo * @pool: unbound pool of interest 54507dbc725eSTejun Heo * @cpu: the CPU which is coming up 54517dbc725eSTejun Heo * 54527dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 54537dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 54547dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 54557dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 54567dbc725eSTejun Heo */ 54577dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 54587dbc725eSTejun Heo { 54597dbc725eSTejun Heo static cpumask_t cpumask; 54607dbc725eSTejun Heo struct worker *worker; 54617dbc725eSTejun Heo 54621258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 54637dbc725eSTejun Heo 54647dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 54657dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 54667dbc725eSTejun Heo return; 54677dbc725eSTejun Heo 54687dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 54697dbc725eSTejun Heo 54707dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5471da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5472d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 54737dbc725eSTejun Heo } 54747dbc725eSTejun Heo 54757ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 54761da177e4SLinus Torvalds { 54774ce62e9eSTejun Heo struct worker_pool *pool; 54781da177e4SLinus Torvalds 5479f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 54803ce63377STejun Heo if (pool->nr_workers) 54813ce63377STejun Heo continue; 5482051e1850SLai Jiangshan if (!create_worker(pool)) 54837ee681b2SThomas Gleixner return -ENOMEM; 54843af24433SOleg Nesterov } 54857ee681b2SThomas Gleixner return 0; 54867ee681b2SThomas Gleixner } 54871da177e4SLinus Torvalds 54887ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 54897ee681b2SThomas Gleixner { 54907ee681b2SThomas Gleixner struct worker_pool *pool; 54917ee681b2SThomas Gleixner struct workqueue_struct *wq; 54927ee681b2SThomas Gleixner int pi; 54937ee681b2SThomas Gleixner 549468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 54957dbc725eSTejun Heo 54967dbc725eSTejun Heo for_each_pool(pool, pi) { 54971258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 549894cf58bbSTejun Heo 5499f05b558dSLai Jiangshan if (pool->cpu == cpu) 550094cf58bbSTejun Heo rebind_workers(pool); 5501f05b558dSLai Jiangshan else if (pool->cpu < 0) 55027dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 550394cf58bbSTejun Heo 55041258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 550594cf58bbSTejun Heo } 55067dbc725eSTejun Heo 5507fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 55084cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 550984193c07STejun Heo struct workqueue_attrs *attrs = wq->unbound_attrs; 551084193c07STejun Heo 551184193c07STejun Heo if (attrs) { 551284193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 55134cbfd3deSTejun Heo int tcpu; 55144cbfd3deSTejun Heo 551584193c07STejun Heo for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5516fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, true); 55174cbfd3deSTejun Heo } 55184cbfd3deSTejun Heo } 55194c16bd32STejun Heo 552068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55217ee681b2SThomas Gleixner return 0; 552265758202STejun Heo } 552365758202STejun Heo 55247ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 552565758202STejun Heo { 55264c16bd32STejun Heo struct workqueue_struct *wq; 55278db25e78STejun Heo 55284c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5529e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5530e8b3f8dbSLai Jiangshan return -1; 5531e8b3f8dbSLai Jiangshan 5532e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55334c16bd32STejun Heo 5534fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 55354c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 55364cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 553784193c07STejun Heo struct workqueue_attrs *attrs = wq->unbound_attrs; 553884193c07STejun Heo 553984193c07STejun Heo if (attrs) { 554084193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 55414cbfd3deSTejun Heo int tcpu; 55424cbfd3deSTejun Heo 554384193c07STejun Heo for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5544fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, false); 55454cbfd3deSTejun Heo } 55464cbfd3deSTejun Heo } 55474c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55484c16bd32STejun Heo 55497ee681b2SThomas Gleixner return 0; 555065758202STejun Heo } 555165758202STejun Heo 55522d3854a3SRusty Russell struct work_for_cpu { 5553ed48ece2STejun Heo struct work_struct work; 55542d3854a3SRusty Russell long (*fn)(void *); 55552d3854a3SRusty Russell void *arg; 55562d3854a3SRusty Russell long ret; 55572d3854a3SRusty Russell }; 55582d3854a3SRusty Russell 5559ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 55602d3854a3SRusty Russell { 5561ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5562ed48ece2STejun Heo 55632d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 55642d3854a3SRusty Russell } 55652d3854a3SRusty Russell 55662d3854a3SRusty Russell /** 556722aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 55682d3854a3SRusty Russell * @cpu: the cpu to run on 55692d3854a3SRusty Russell * @fn: the function to run 55702d3854a3SRusty Russell * @arg: the function arg 55712d3854a3SRusty Russell * 557231ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 55736b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5574d185af30SYacine Belkadi * 5575d185af30SYacine Belkadi * Return: The value @fn returns. 55762d3854a3SRusty Russell */ 5577d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 55782d3854a3SRusty Russell { 5579ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 55802d3854a3SRusty Russell 5581ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5582ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 558312997d1aSBjorn Helgaas flush_work(&wfc.work); 5584440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 55852d3854a3SRusty Russell return wfc.ret; 55862d3854a3SRusty Russell } 55872d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 55880e8d6a93SThomas Gleixner 55890e8d6a93SThomas Gleixner /** 55900e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 55910e8d6a93SThomas Gleixner * @cpu: the cpu to run on 55920e8d6a93SThomas Gleixner * @fn: the function to run 55930e8d6a93SThomas Gleixner * @arg: the function argument 55940e8d6a93SThomas Gleixner * 55950e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 55960e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 55970e8d6a93SThomas Gleixner * 55980e8d6a93SThomas Gleixner * Return: The value @fn returns. 55990e8d6a93SThomas Gleixner */ 56000e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 56010e8d6a93SThomas Gleixner { 56020e8d6a93SThomas Gleixner long ret = -ENODEV; 56030e8d6a93SThomas Gleixner 5604ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 56050e8d6a93SThomas Gleixner if (cpu_online(cpu)) 56060e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5607ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56080e8d6a93SThomas Gleixner return ret; 56090e8d6a93SThomas Gleixner } 56100e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56112d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56122d3854a3SRusty Russell 5613a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5614e7577c50SRusty Russell 5615a0a1a5fdSTejun Heo /** 5616a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5617a0a1a5fdSTejun Heo * 561858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5619f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5620706026c2STejun Heo * pool->worklist. 5621a0a1a5fdSTejun Heo * 5622a0a1a5fdSTejun Heo * CONTEXT: 5623a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5624a0a1a5fdSTejun Heo */ 5625a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5626a0a1a5fdSTejun Heo { 562724b8a847STejun Heo struct workqueue_struct *wq; 562824b8a847STejun Heo struct pool_workqueue *pwq; 5629a0a1a5fdSTejun Heo 563068e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5631a0a1a5fdSTejun Heo 56326183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5633a0a1a5fdSTejun Heo workqueue_freezing = true; 5634a0a1a5fdSTejun Heo 563524b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5636a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5637699ce097STejun Heo for_each_pwq(pwq, wq) 5638699ce097STejun Heo pwq_adjust_max_active(pwq); 5639a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5640a1056305STejun Heo } 56415bcab335STejun Heo 564268e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5643a0a1a5fdSTejun Heo } 5644a0a1a5fdSTejun Heo 5645a0a1a5fdSTejun Heo /** 564658a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5647a0a1a5fdSTejun Heo * 5648a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5649a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5650a0a1a5fdSTejun Heo * 5651a0a1a5fdSTejun Heo * CONTEXT: 565268e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5653a0a1a5fdSTejun Heo * 5654d185af30SYacine Belkadi * Return: 565558a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 565658a69cb4STejun Heo * is complete. 5657a0a1a5fdSTejun Heo */ 5658a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5659a0a1a5fdSTejun Heo { 5660a0a1a5fdSTejun Heo bool busy = false; 566124b8a847STejun Heo struct workqueue_struct *wq; 566224b8a847STejun Heo struct pool_workqueue *pwq; 5663a0a1a5fdSTejun Heo 566468e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5665a0a1a5fdSTejun Heo 56666183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5667a0a1a5fdSTejun Heo 566824b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 566924b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 567024b8a847STejun Heo continue; 5671a0a1a5fdSTejun Heo /* 5672a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5673a0a1a5fdSTejun Heo * to peek without lock. 5674a0a1a5fdSTejun Heo */ 567524acfb71SThomas Gleixner rcu_read_lock(); 567624b8a847STejun Heo for_each_pwq(pwq, wq) { 56776183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5678112202d9STejun Heo if (pwq->nr_active) { 5679a0a1a5fdSTejun Heo busy = true; 568024acfb71SThomas Gleixner rcu_read_unlock(); 5681a0a1a5fdSTejun Heo goto out_unlock; 5682a0a1a5fdSTejun Heo } 5683a0a1a5fdSTejun Heo } 568424acfb71SThomas Gleixner rcu_read_unlock(); 5685a0a1a5fdSTejun Heo } 5686a0a1a5fdSTejun Heo out_unlock: 568768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5688a0a1a5fdSTejun Heo return busy; 5689a0a1a5fdSTejun Heo } 5690a0a1a5fdSTejun Heo 5691a0a1a5fdSTejun Heo /** 5692a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5693a0a1a5fdSTejun Heo * 5694a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5695706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5696a0a1a5fdSTejun Heo * 5697a0a1a5fdSTejun Heo * CONTEXT: 5698a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5699a0a1a5fdSTejun Heo */ 5700a0a1a5fdSTejun Heo void thaw_workqueues(void) 5701a0a1a5fdSTejun Heo { 570224b8a847STejun Heo struct workqueue_struct *wq; 570324b8a847STejun Heo struct pool_workqueue *pwq; 5704a0a1a5fdSTejun Heo 570568e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5706a0a1a5fdSTejun Heo 5707a0a1a5fdSTejun Heo if (!workqueue_freezing) 5708a0a1a5fdSTejun Heo goto out_unlock; 5709a0a1a5fdSTejun Heo 571074b414eaSLai Jiangshan workqueue_freezing = false; 571124b8a847STejun Heo 571224b8a847STejun Heo /* restore max_active and repopulate worklist */ 571324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5714a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5715699ce097STejun Heo for_each_pwq(pwq, wq) 5716699ce097STejun Heo pwq_adjust_max_active(pwq); 5717a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 571824b8a847STejun Heo } 571924b8a847STejun Heo 5720a0a1a5fdSTejun Heo out_unlock: 572168e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5722a0a1a5fdSTejun Heo } 5723a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5724a0a1a5fdSTejun Heo 572599c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5726042f7df1SLai Jiangshan { 5727042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5728042f7df1SLai Jiangshan int ret = 0; 5729042f7df1SLai Jiangshan struct workqueue_struct *wq; 5730042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5731042f7df1SLai Jiangshan 5732042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5733042f7df1SLai Jiangshan 5734042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5735042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5736042f7df1SLai Jiangshan continue; 5737042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5738042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5739042f7df1SLai Jiangshan continue; 5740042f7df1SLai Jiangshan 574199c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 574284193c07STejun Heo if (IS_ERR(ctx)) { 574384193c07STejun Heo ret = PTR_ERR(ctx); 5744042f7df1SLai Jiangshan break; 5745042f7df1SLai Jiangshan } 5746042f7df1SLai Jiangshan 5747042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5748042f7df1SLai Jiangshan } 5749042f7df1SLai Jiangshan 5750042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5751042f7df1SLai Jiangshan if (!ret) 5752042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5753042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5754042f7df1SLai Jiangshan } 5755042f7df1SLai Jiangshan 575699c621efSLai Jiangshan if (!ret) { 575799c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 575899c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 575999c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 576099c621efSLai Jiangshan } 5761042f7df1SLai Jiangshan return ret; 5762042f7df1SLai Jiangshan } 5763042f7df1SLai Jiangshan 5764042f7df1SLai Jiangshan /** 5765042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5766042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5767042f7df1SLai Jiangshan * 5768042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5769042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5770042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5771042f7df1SLai Jiangshan * 577267dc8325SCai Huoqing * Return: 0 - Success 5773042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5774042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5775042f7df1SLai Jiangshan */ 5776042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5777042f7df1SLai Jiangshan { 5778042f7df1SLai Jiangshan int ret = -EINVAL; 5779042f7df1SLai Jiangshan 5780c98a9805STal Shorer /* 5781c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5782c98a9805STal Shorer * If the user wishes to include them, we allow that. 5783c98a9805STal Shorer */ 5784042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5785042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5786a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5787d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5788d25302e4SMenglong Dong ret = 0; 5789d25302e4SMenglong Dong goto out_unlock; 5790d25302e4SMenglong Dong } 5791d25302e4SMenglong Dong 579299c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5793042f7df1SLai Jiangshan 5794d25302e4SMenglong Dong out_unlock: 5795a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5796042f7df1SLai Jiangshan } 5797042f7df1SLai Jiangshan 5798042f7df1SLai Jiangshan return ret; 5799042f7df1SLai Jiangshan } 5800042f7df1SLai Jiangshan 580163c5484eSTejun Heo static int parse_affn_scope(const char *val) 580263c5484eSTejun Heo { 580363c5484eSTejun Heo int i; 580463c5484eSTejun Heo 580563c5484eSTejun Heo for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 580663c5484eSTejun Heo if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 580763c5484eSTejun Heo return i; 580863c5484eSTejun Heo } 580963c5484eSTejun Heo return -EINVAL; 581063c5484eSTejun Heo } 581163c5484eSTejun Heo 581263c5484eSTejun Heo static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 581363c5484eSTejun Heo { 581463c5484eSTejun Heo int affn; 581563c5484eSTejun Heo 581663c5484eSTejun Heo affn = parse_affn_scope(val); 581763c5484eSTejun Heo if (affn < 0) 581863c5484eSTejun Heo return affn; 581963c5484eSTejun Heo 582063c5484eSTejun Heo wq_affn_dfl = affn; 582163c5484eSTejun Heo return 0; 582263c5484eSTejun Heo } 582363c5484eSTejun Heo 582463c5484eSTejun Heo static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 582563c5484eSTejun Heo { 582663c5484eSTejun Heo return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 582763c5484eSTejun Heo } 582863c5484eSTejun Heo 582963c5484eSTejun Heo static const struct kernel_param_ops wq_affn_dfl_ops = { 583063c5484eSTejun Heo .set = wq_affn_dfl_set, 583163c5484eSTejun Heo .get = wq_affn_dfl_get, 583263c5484eSTejun Heo }; 583363c5484eSTejun Heo 583463c5484eSTejun Heo module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 583563c5484eSTejun Heo 58366ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 58376ba94429SFrederic Weisbecker /* 58386ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 58396ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 58406ba94429SFrederic Weisbecker * following attributes. 58416ba94429SFrederic Weisbecker * 58426ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 58436ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 58446ba94429SFrederic Weisbecker * 58456ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 58466ba94429SFrederic Weisbecker * 58476ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 58486ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 584963c5484eSTejun Heo * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 58506ba94429SFrederic Weisbecker */ 58516ba94429SFrederic Weisbecker struct wq_device { 58526ba94429SFrederic Weisbecker struct workqueue_struct *wq; 58536ba94429SFrederic Weisbecker struct device dev; 58546ba94429SFrederic Weisbecker }; 58556ba94429SFrederic Weisbecker 58566ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 58576ba94429SFrederic Weisbecker { 58586ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 58596ba94429SFrederic Weisbecker 58606ba94429SFrederic Weisbecker return wq_dev->wq; 58616ba94429SFrederic Weisbecker } 58626ba94429SFrederic Weisbecker 58636ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 58646ba94429SFrederic Weisbecker char *buf) 58656ba94429SFrederic Weisbecker { 58666ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58676ba94429SFrederic Weisbecker 58686ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 58696ba94429SFrederic Weisbecker } 58706ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 58716ba94429SFrederic Weisbecker 58726ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 58736ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 58746ba94429SFrederic Weisbecker { 58756ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58766ba94429SFrederic Weisbecker 58776ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 58786ba94429SFrederic Weisbecker } 58796ba94429SFrederic Weisbecker 58806ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 58816ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 58826ba94429SFrederic Weisbecker size_t count) 58836ba94429SFrederic Weisbecker { 58846ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 58856ba94429SFrederic Weisbecker int val; 58866ba94429SFrederic Weisbecker 58876ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 58886ba94429SFrederic Weisbecker return -EINVAL; 58896ba94429SFrederic Weisbecker 58906ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 58916ba94429SFrederic Weisbecker return count; 58926ba94429SFrederic Weisbecker } 58936ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 58946ba94429SFrederic Weisbecker 58956ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 58966ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 58976ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 58986ba94429SFrederic Weisbecker NULL, 58996ba94429SFrederic Weisbecker }; 59006ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 59016ba94429SFrederic Weisbecker 59026ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 59036ba94429SFrederic Weisbecker char *buf) 59046ba94429SFrederic Weisbecker { 59056ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59066ba94429SFrederic Weisbecker int written; 59076ba94429SFrederic Weisbecker 59086ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59096ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 59106ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59116ba94429SFrederic Weisbecker 59126ba94429SFrederic Weisbecker return written; 59136ba94429SFrederic Weisbecker } 59146ba94429SFrederic Weisbecker 59156ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 59166ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 59176ba94429SFrederic Weisbecker { 59186ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 59196ba94429SFrederic Weisbecker 5920899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5921899a94feSLai Jiangshan 5922be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 59236ba94429SFrederic Weisbecker if (!attrs) 59246ba94429SFrederic Weisbecker return NULL; 59256ba94429SFrederic Weisbecker 59266ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 59276ba94429SFrederic Weisbecker return attrs; 59286ba94429SFrederic Weisbecker } 59296ba94429SFrederic Weisbecker 59306ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59316ba94429SFrederic Weisbecker const char *buf, size_t count) 59326ba94429SFrederic Weisbecker { 59336ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59346ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5935d4d3e257SLai Jiangshan int ret = -ENOMEM; 5936d4d3e257SLai Jiangshan 5937d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59386ba94429SFrederic Weisbecker 59396ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59406ba94429SFrederic Weisbecker if (!attrs) 5941d4d3e257SLai Jiangshan goto out_unlock; 59426ba94429SFrederic Weisbecker 59436ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 59446ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 5945d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59466ba94429SFrederic Weisbecker else 59476ba94429SFrederic Weisbecker ret = -EINVAL; 59486ba94429SFrederic Weisbecker 5949d4d3e257SLai Jiangshan out_unlock: 5950d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59516ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59526ba94429SFrederic Weisbecker return ret ?: count; 59536ba94429SFrederic Weisbecker } 59546ba94429SFrederic Weisbecker 59556ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 59566ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59576ba94429SFrederic Weisbecker { 59586ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59596ba94429SFrederic Weisbecker int written; 59606ba94429SFrederic Weisbecker 59616ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59626ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 59636ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 59646ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59656ba94429SFrederic Weisbecker return written; 59666ba94429SFrederic Weisbecker } 59676ba94429SFrederic Weisbecker 59686ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 59696ba94429SFrederic Weisbecker struct device_attribute *attr, 59706ba94429SFrederic Weisbecker const char *buf, size_t count) 59716ba94429SFrederic Weisbecker { 59726ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59736ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 5974d4d3e257SLai Jiangshan int ret = -ENOMEM; 5975d4d3e257SLai Jiangshan 5976d4d3e257SLai Jiangshan apply_wqattrs_lock(); 59776ba94429SFrederic Weisbecker 59786ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 59796ba94429SFrederic Weisbecker if (!attrs) 5980d4d3e257SLai Jiangshan goto out_unlock; 59816ba94429SFrederic Weisbecker 59826ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 59836ba94429SFrederic Weisbecker if (!ret) 5984d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 59856ba94429SFrederic Weisbecker 5986d4d3e257SLai Jiangshan out_unlock: 5987d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 59886ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 59896ba94429SFrederic Weisbecker return ret ?: count; 59906ba94429SFrederic Weisbecker } 59916ba94429SFrederic Weisbecker 599263c5484eSTejun Heo static ssize_t wq_affn_scope_show(struct device *dev, 599363c5484eSTejun Heo struct device_attribute *attr, char *buf) 599463c5484eSTejun Heo { 599563c5484eSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 599663c5484eSTejun Heo int written; 599763c5484eSTejun Heo 599863c5484eSTejun Heo mutex_lock(&wq->mutex); 599963c5484eSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%s\n", 600063c5484eSTejun Heo wq_affn_names[wq->unbound_attrs->affn_scope]); 600163c5484eSTejun Heo mutex_unlock(&wq->mutex); 600263c5484eSTejun Heo 600363c5484eSTejun Heo return written; 600463c5484eSTejun Heo } 600563c5484eSTejun Heo 600663c5484eSTejun Heo static ssize_t wq_affn_scope_store(struct device *dev, 600763c5484eSTejun Heo struct device_attribute *attr, 600863c5484eSTejun Heo const char *buf, size_t count) 600963c5484eSTejun Heo { 601063c5484eSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 601163c5484eSTejun Heo struct workqueue_attrs *attrs; 601263c5484eSTejun Heo int affn, ret = -ENOMEM; 601363c5484eSTejun Heo 601463c5484eSTejun Heo affn = parse_affn_scope(buf); 601563c5484eSTejun Heo if (affn < 0) 601663c5484eSTejun Heo return affn; 601763c5484eSTejun Heo 601863c5484eSTejun Heo apply_wqattrs_lock(); 601963c5484eSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 602063c5484eSTejun Heo if (attrs) { 602163c5484eSTejun Heo attrs->affn_scope = affn; 602263c5484eSTejun Heo ret = apply_workqueue_attrs_locked(wq, attrs); 602363c5484eSTejun Heo } 602463c5484eSTejun Heo apply_wqattrs_unlock(); 602563c5484eSTejun Heo free_workqueue_attrs(attrs); 602663c5484eSTejun Heo return ret ?: count; 602763c5484eSTejun Heo } 602863c5484eSTejun Heo 60296ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 60306ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 60316ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 603263c5484eSTejun Heo __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 60336ba94429SFrederic Weisbecker __ATTR_NULL, 60346ba94429SFrederic Weisbecker }; 60356ba94429SFrederic Weisbecker 60366ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 60376ba94429SFrederic Weisbecker .name = "workqueue", 60386ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 60396ba94429SFrederic Weisbecker }; 60406ba94429SFrederic Weisbecker 6041b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 6042b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 6043b05a7928SFrederic Weisbecker { 6044b05a7928SFrederic Weisbecker int written; 6045b05a7928SFrederic Weisbecker 6046042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 6047b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6048b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 6049042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 6050b05a7928SFrederic Weisbecker 6051b05a7928SFrederic Weisbecker return written; 6052b05a7928SFrederic Weisbecker } 6053b05a7928SFrederic Weisbecker 6054042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 6055042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 6056042f7df1SLai Jiangshan { 6057042f7df1SLai Jiangshan cpumask_var_t cpumask; 6058042f7df1SLai Jiangshan int ret; 6059042f7df1SLai Jiangshan 6060042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6061042f7df1SLai Jiangshan return -ENOMEM; 6062042f7df1SLai Jiangshan 6063042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 6064042f7df1SLai Jiangshan if (!ret) 6065042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 6066042f7df1SLai Jiangshan 6067042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6068042f7df1SLai Jiangshan return ret ? ret : count; 6069042f7df1SLai Jiangshan } 6070042f7df1SLai Jiangshan 6071b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6072042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6073042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6074b05a7928SFrederic Weisbecker 60756ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 60766ba94429SFrederic Weisbecker { 6077686f6697SGreg Kroah-Hartman struct device *dev_root; 6078b05a7928SFrederic Weisbecker int err; 6079b05a7928SFrederic Weisbecker 6080b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6081b05a7928SFrederic Weisbecker if (err) 6082b05a7928SFrederic Weisbecker return err; 6083b05a7928SFrederic Weisbecker 6084686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6085686f6697SGreg Kroah-Hartman if (dev_root) { 6086686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6087686f6697SGreg Kroah-Hartman put_device(dev_root); 6088686f6697SGreg Kroah-Hartman } 6089686f6697SGreg Kroah-Hartman return err; 60906ba94429SFrederic Weisbecker } 60916ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 60926ba94429SFrederic Weisbecker 60936ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 60946ba94429SFrederic Weisbecker { 60956ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 60966ba94429SFrederic Weisbecker 60976ba94429SFrederic Weisbecker kfree(wq_dev); 60986ba94429SFrederic Weisbecker } 60996ba94429SFrederic Weisbecker 61006ba94429SFrederic Weisbecker /** 61016ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 61026ba94429SFrederic Weisbecker * @wq: the workqueue to register 61036ba94429SFrederic Weisbecker * 61046ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 61056ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 61066ba94429SFrederic Weisbecker * which is the preferred method. 61076ba94429SFrederic Weisbecker * 61086ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 61096ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 61106ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 61116ba94429SFrederic Weisbecker * attributes. 61126ba94429SFrederic Weisbecker * 61136ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 61146ba94429SFrederic Weisbecker */ 61156ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 61166ba94429SFrederic Weisbecker { 61176ba94429SFrederic Weisbecker struct wq_device *wq_dev; 61186ba94429SFrederic Weisbecker int ret; 61196ba94429SFrederic Weisbecker 61206ba94429SFrederic Weisbecker /* 6121402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 61226ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 61236ba94429SFrederic Weisbecker * workqueues. 61246ba94429SFrederic Weisbecker */ 61250a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 61266ba94429SFrederic Weisbecker return -EINVAL; 61276ba94429SFrederic Weisbecker 61286ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 61296ba94429SFrederic Weisbecker if (!wq_dev) 61306ba94429SFrederic Weisbecker return -ENOMEM; 61316ba94429SFrederic Weisbecker 61326ba94429SFrederic Weisbecker wq_dev->wq = wq; 61336ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 61346ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 613523217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 61366ba94429SFrederic Weisbecker 61376ba94429SFrederic Weisbecker /* 61386ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 61396ba94429SFrederic Weisbecker * everything is ready. 61406ba94429SFrederic Weisbecker */ 61416ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 61426ba94429SFrederic Weisbecker 61436ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 61446ba94429SFrederic Weisbecker if (ret) { 6145537f4146SArvind Yadav put_device(&wq_dev->dev); 61466ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61476ba94429SFrederic Weisbecker return ret; 61486ba94429SFrederic Weisbecker } 61496ba94429SFrederic Weisbecker 61506ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 61516ba94429SFrederic Weisbecker struct device_attribute *attr; 61526ba94429SFrederic Weisbecker 61536ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 61546ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 61556ba94429SFrederic Weisbecker if (ret) { 61566ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61576ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61586ba94429SFrederic Weisbecker return ret; 61596ba94429SFrederic Weisbecker } 61606ba94429SFrederic Weisbecker } 61616ba94429SFrederic Weisbecker } 61626ba94429SFrederic Weisbecker 61636ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 61646ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 61656ba94429SFrederic Weisbecker return 0; 61666ba94429SFrederic Weisbecker } 61676ba94429SFrederic Weisbecker 61686ba94429SFrederic Weisbecker /** 61696ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 61706ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 61716ba94429SFrederic Weisbecker * 61726ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 61736ba94429SFrederic Weisbecker */ 61746ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 61756ba94429SFrederic Weisbecker { 61766ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 61776ba94429SFrederic Weisbecker 61786ba94429SFrederic Weisbecker if (!wq->wq_dev) 61796ba94429SFrederic Weisbecker return; 61806ba94429SFrederic Weisbecker 61816ba94429SFrederic Weisbecker wq->wq_dev = NULL; 61826ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 61836ba94429SFrederic Weisbecker } 61846ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 61856ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 61866ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 61876ba94429SFrederic Weisbecker 618882607adcSTejun Heo /* 618982607adcSTejun Heo * Workqueue watchdog. 619082607adcSTejun Heo * 619182607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 619282607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 619382607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 619482607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 619582607adcSTejun Heo * largely opaque. 619682607adcSTejun Heo * 619782607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 619882607adcSTejun Heo * state if some pools failed to make forward progress for a while where 619982607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 620082607adcSTejun Heo * 620182607adcSTejun Heo * This mechanism is controlled through the kernel parameter 620282607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 620382607adcSTejun Heo * corresponding sysfs parameter file. 620482607adcSTejun Heo */ 620582607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 620682607adcSTejun Heo 620782607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 62085cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 620982607adcSTejun Heo 621082607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 621182607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 621282607adcSTejun Heo 6213cd2440d6SPetr Mladek /* 6214cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6215cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6216cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6217cd2440d6SPetr Mladek * in all other situations. 6218cd2440d6SPetr Mladek */ 6219cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6220cd2440d6SPetr Mladek { 6221cd2440d6SPetr Mladek struct worker *worker; 6222cd2440d6SPetr Mladek unsigned long flags; 6223cd2440d6SPetr Mladek int bkt; 6224cd2440d6SPetr Mladek 6225cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6226cd2440d6SPetr Mladek 6227cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6228cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6229cd2440d6SPetr Mladek /* 6230cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6231cd2440d6SPetr Mladek * drivers that queue work while holding locks 6232cd2440d6SPetr Mladek * also taken in their write paths. 6233cd2440d6SPetr Mladek */ 6234cd2440d6SPetr Mladek printk_deferred_enter(); 6235cd2440d6SPetr Mladek 6236cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6237cd2440d6SPetr Mladek sched_show_task(worker->task); 6238cd2440d6SPetr Mladek 6239cd2440d6SPetr Mladek printk_deferred_exit(); 6240cd2440d6SPetr Mladek } 6241cd2440d6SPetr Mladek } 6242cd2440d6SPetr Mladek 6243cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6244cd2440d6SPetr Mladek } 6245cd2440d6SPetr Mladek 6246cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6247cd2440d6SPetr Mladek { 6248cd2440d6SPetr Mladek struct worker_pool *pool; 6249cd2440d6SPetr Mladek int pi; 6250cd2440d6SPetr Mladek 6251cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6252cd2440d6SPetr Mladek 6253cd2440d6SPetr Mladek rcu_read_lock(); 6254cd2440d6SPetr Mladek 6255cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6256cd2440d6SPetr Mladek if (pool->cpu_stall) 6257cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6258cd2440d6SPetr Mladek 6259cd2440d6SPetr Mladek } 6260cd2440d6SPetr Mladek 6261cd2440d6SPetr Mladek rcu_read_unlock(); 6262cd2440d6SPetr Mladek } 6263cd2440d6SPetr Mladek 626482607adcSTejun Heo static void wq_watchdog_reset_touched(void) 626582607adcSTejun Heo { 626682607adcSTejun Heo int cpu; 626782607adcSTejun Heo 626882607adcSTejun Heo wq_watchdog_touched = jiffies; 626982607adcSTejun Heo for_each_possible_cpu(cpu) 627082607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 627182607adcSTejun Heo } 627282607adcSTejun Heo 62735cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 627482607adcSTejun Heo { 627582607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 627682607adcSTejun Heo bool lockup_detected = false; 6277cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6278940d71c6SSergey Senozhatsky unsigned long now = jiffies; 627982607adcSTejun Heo struct worker_pool *pool; 628082607adcSTejun Heo int pi; 628182607adcSTejun Heo 628282607adcSTejun Heo if (!thresh) 628382607adcSTejun Heo return; 628482607adcSTejun Heo 628582607adcSTejun Heo rcu_read_lock(); 628682607adcSTejun Heo 628782607adcSTejun Heo for_each_pool(pool, pi) { 628882607adcSTejun Heo unsigned long pool_ts, touched, ts; 628982607adcSTejun Heo 6290cd2440d6SPetr Mladek pool->cpu_stall = false; 629182607adcSTejun Heo if (list_empty(&pool->worklist)) 629282607adcSTejun Heo continue; 629382607adcSTejun Heo 6294940d71c6SSergey Senozhatsky /* 6295940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6296940d71c6SSergey Senozhatsky * the watchdog like a stall. 6297940d71c6SSergey Senozhatsky */ 6298940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6299940d71c6SSergey Senozhatsky 630082607adcSTejun Heo /* get the latest of pool and touched timestamps */ 630189e28ce6SWang Qing if (pool->cpu >= 0) 630289e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 630389e28ce6SWang Qing else 630482607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 630589e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 630682607adcSTejun Heo 630782607adcSTejun Heo if (time_after(pool_ts, touched)) 630882607adcSTejun Heo ts = pool_ts; 630982607adcSTejun Heo else 631082607adcSTejun Heo ts = touched; 631182607adcSTejun Heo 631282607adcSTejun Heo /* did we stall? */ 6313940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 631482607adcSTejun Heo lockup_detected = true; 6315cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6316cd2440d6SPetr Mladek pool->cpu_stall = true; 6317cd2440d6SPetr Mladek cpu_pool_stall = true; 6318cd2440d6SPetr Mladek } 631982607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 632082607adcSTejun Heo pr_cont_pool_info(pool); 632182607adcSTejun Heo pr_cont(" stuck for %us!\n", 6322940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 632382607adcSTejun Heo } 6324cd2440d6SPetr Mladek 6325cd2440d6SPetr Mladek 632682607adcSTejun Heo } 632782607adcSTejun Heo 632882607adcSTejun Heo rcu_read_unlock(); 632982607adcSTejun Heo 633082607adcSTejun Heo if (lockup_detected) 633155df0933SImran Khan show_all_workqueues(); 633282607adcSTejun Heo 6333cd2440d6SPetr Mladek if (cpu_pool_stall) 6334cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6335cd2440d6SPetr Mladek 633682607adcSTejun Heo wq_watchdog_reset_touched(); 633782607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 633882607adcSTejun Heo } 633982607adcSTejun Heo 6340cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 634182607adcSTejun Heo { 634282607adcSTejun Heo if (cpu >= 0) 634382607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 634489e28ce6SWang Qing 634582607adcSTejun Heo wq_watchdog_touched = jiffies; 634682607adcSTejun Heo } 634782607adcSTejun Heo 634882607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 634982607adcSTejun Heo { 635082607adcSTejun Heo wq_watchdog_thresh = 0; 635182607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 635282607adcSTejun Heo 635382607adcSTejun Heo if (thresh) { 635482607adcSTejun Heo wq_watchdog_thresh = thresh; 635582607adcSTejun Heo wq_watchdog_reset_touched(); 635682607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 635782607adcSTejun Heo } 635882607adcSTejun Heo } 635982607adcSTejun Heo 636082607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 636182607adcSTejun Heo const struct kernel_param *kp) 636282607adcSTejun Heo { 636382607adcSTejun Heo unsigned long thresh; 636482607adcSTejun Heo int ret; 636582607adcSTejun Heo 636682607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 636782607adcSTejun Heo if (ret) 636882607adcSTejun Heo return ret; 636982607adcSTejun Heo 637082607adcSTejun Heo if (system_wq) 637182607adcSTejun Heo wq_watchdog_set_thresh(thresh); 637282607adcSTejun Heo else 637382607adcSTejun Heo wq_watchdog_thresh = thresh; 637482607adcSTejun Heo 637582607adcSTejun Heo return 0; 637682607adcSTejun Heo } 637782607adcSTejun Heo 637882607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 637982607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 638082607adcSTejun Heo .get = param_get_ulong, 638182607adcSTejun Heo }; 638282607adcSTejun Heo 638382607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 638482607adcSTejun Heo 0644); 638582607adcSTejun Heo 638682607adcSTejun Heo static void wq_watchdog_init(void) 638782607adcSTejun Heo { 63885cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 638982607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 639082607adcSTejun Heo } 639182607adcSTejun Heo 639282607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 639382607adcSTejun Heo 639482607adcSTejun Heo static inline void wq_watchdog_init(void) { } 639582607adcSTejun Heo 639682607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 639782607adcSTejun Heo 63983347fa09STejun Heo /** 63993347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 64003347fa09STejun Heo * 64012930155bSTejun Heo * This is the first step of three-staged workqueue subsystem initialization and 64022930155bSTejun Heo * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 64032930155bSTejun Heo * up. It sets up all the data structures and system workqueues and allows early 64042930155bSTejun Heo * boot code to create workqueues and queue/cancel work items. Actual work item 64052930155bSTejun Heo * execution starts only after kthreads can be created and scheduled right 64062930155bSTejun Heo * before early initcalls. 64073347fa09STejun Heo */ 64082333e829SYu Chen void __init workqueue_init_early(void) 64091da177e4SLinus Torvalds { 641084193c07STejun Heo struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 64117a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 64127a4e344cSTejun Heo int i, cpu; 6413c34056a3STejun Heo 641410cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6415e904e6c2STejun Heo 6416b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 641704d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 641804d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6419b05a7928SFrederic Weisbecker 6420ace3c549Stiozhang if (!cpumask_empty(&wq_cmdline_cpumask)) 6421ace3c549Stiozhang cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); 6422ace3c549Stiozhang 6423e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6424e904e6c2STejun Heo 64252930155bSTejun Heo wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 64262930155bSTejun Heo BUG_ON(!wq_update_pod_attrs_buf); 64272930155bSTejun Heo 642884193c07STejun Heo /* initialize WQ_AFFN_SYSTEM pods */ 642984193c07STejun Heo pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 643084193c07STejun Heo pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 643184193c07STejun Heo pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 643284193c07STejun Heo BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 643384193c07STejun Heo 643484193c07STejun Heo BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 643584193c07STejun Heo 643684193c07STejun Heo wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 643784193c07STejun Heo BUG_ON(!wq_update_pod_attrs_buf); 643884193c07STejun Heo 643984193c07STejun Heo pt->nr_pods = 1; 644084193c07STejun Heo cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 644184193c07STejun Heo pt->pod_node[0] = NUMA_NO_NODE; 644284193c07STejun Heo pt->cpu_pod[0] = 0; 644384193c07STejun Heo 6444706026c2STejun Heo /* initialize CPU pools */ 644529c91e99STejun Heo for_each_possible_cpu(cpu) { 64464ce62e9eSTejun Heo struct worker_pool *pool; 64478b03ae3cSTejun Heo 64487a4e344cSTejun Heo i = 0; 6449f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 64507a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6451ec22ca5eSTejun Heo pool->cpu = cpu; 64527a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 6453*9546b29eSTejun Heo cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 64547a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 6455f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 64567a4e344cSTejun Heo 64579daf9e67STejun Heo /* alloc pool ID */ 645868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 64599daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 646068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 64614ce62e9eSTejun Heo } 64628b03ae3cSTejun Heo } 64638b03ae3cSTejun Heo 64648a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 646529c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 646629c91e99STejun Heo struct workqueue_attrs *attrs; 646729c91e99STejun Heo 6468be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 646929c91e99STejun Heo attrs->nice = std_nice[i]; 647029c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 64718a2b7538STejun Heo 64728a2b7538STejun Heo /* 64738a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 64748a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 64758a2b7538STejun Heo */ 6476be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 64778a2b7538STejun Heo attrs->nice = std_nice[i]; 6478af73f5c9STejun Heo attrs->ordered = true; 64798a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 648029c91e99STejun Heo } 648129c91e99STejun Heo 6482d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 64831aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6484d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6485f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6486636b927eSTejun Heo WQ_MAX_ACTIVE); 648724d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 648824d51addSTejun Heo WQ_FREEZABLE, 0); 64890668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 64900668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 64910668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 64920668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 64930668106cSViresh Kumar 0); 64941aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 64950668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 64960668106cSViresh Kumar !system_power_efficient_wq || 64970668106cSViresh Kumar !system_freezable_power_efficient_wq); 64983347fa09STejun Heo } 64993347fa09STejun Heo 6500aa6fde93STejun Heo static void __init wq_cpu_intensive_thresh_init(void) 6501aa6fde93STejun Heo { 6502aa6fde93STejun Heo unsigned long thresh; 6503aa6fde93STejun Heo unsigned long bogo; 6504aa6fde93STejun Heo 6505aa6fde93STejun Heo /* if the user set it to a specific value, keep it */ 6506aa6fde93STejun Heo if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6507aa6fde93STejun Heo return; 6508aa6fde93STejun Heo 6509967b494eSTejun Heo pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6510967b494eSTejun Heo BUG_ON(IS_ERR(pwq_release_worker)); 6511967b494eSTejun Heo 6512aa6fde93STejun Heo /* 6513aa6fde93STejun Heo * The default of 10ms is derived from the fact that most modern (as of 6514aa6fde93STejun Heo * 2023) processors can do a lot in 10ms and that it's just below what 6515aa6fde93STejun Heo * most consider human-perceivable. However, the kernel also runs on a 6516aa6fde93STejun Heo * lot slower CPUs including microcontrollers where the threshold is way 6517aa6fde93STejun Heo * too low. 6518aa6fde93STejun Heo * 6519aa6fde93STejun Heo * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6520aa6fde93STejun Heo * This is by no means accurate but it doesn't have to be. The mechanism 6521aa6fde93STejun Heo * is still useful even when the threshold is fully scaled up. Also, as 6522aa6fde93STejun Heo * the reports would usually be applicable to everyone, some machines 6523aa6fde93STejun Heo * operating on longer thresholds won't significantly diminish their 6524aa6fde93STejun Heo * usefulness. 6525aa6fde93STejun Heo */ 6526aa6fde93STejun Heo thresh = 10 * USEC_PER_MSEC; 6527aa6fde93STejun Heo 6528aa6fde93STejun Heo /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6529aa6fde93STejun Heo bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6530aa6fde93STejun Heo if (bogo < 4000) 6531aa6fde93STejun Heo thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6532aa6fde93STejun Heo 6533aa6fde93STejun Heo pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6534aa6fde93STejun Heo loops_per_jiffy, bogo, thresh); 6535aa6fde93STejun Heo 6536aa6fde93STejun Heo wq_cpu_intensive_thresh_us = thresh; 6537aa6fde93STejun Heo } 6538aa6fde93STejun Heo 65393347fa09STejun Heo /** 65403347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 65413347fa09STejun Heo * 65422930155bSTejun Heo * This is the second step of three-staged workqueue subsystem initialization 65432930155bSTejun Heo * and invoked as soon as kthreads can be created and scheduled. Workqueues have 65442930155bSTejun Heo * been created and work items queued on them, but there are no kworkers 65452930155bSTejun Heo * executing the work items yet. Populate the worker pools with the initial 65462930155bSTejun Heo * workers and enable future kworker creations. 65473347fa09STejun Heo */ 65482333e829SYu Chen void __init workqueue_init(void) 65493347fa09STejun Heo { 65502186d9f9STejun Heo struct workqueue_struct *wq; 65513347fa09STejun Heo struct worker_pool *pool; 65523347fa09STejun Heo int cpu, bkt; 65533347fa09STejun Heo 6554aa6fde93STejun Heo wq_cpu_intensive_thresh_init(); 6555aa6fde93STejun Heo 65562186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 65572186d9f9STejun Heo 65582930155bSTejun Heo /* 65592930155bSTejun Heo * Per-cpu pools created earlier could be missing node hint. Fix them 65602930155bSTejun Heo * up. Also, create a rescuer for workqueues that requested it. 65612930155bSTejun Heo */ 65622186d9f9STejun Heo for_each_possible_cpu(cpu) { 65632186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65642186d9f9STejun Heo pool->node = cpu_to_node(cpu); 65652186d9f9STejun Heo } 65662186d9f9STejun Heo } 65672186d9f9STejun Heo 656840c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 656940c17f75STejun Heo WARN(init_rescuer(wq), 657040c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 657140c17f75STejun Heo wq->name); 657240c17f75STejun Heo } 65732186d9f9STejun Heo 65742186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 65752186d9f9STejun Heo 65763347fa09STejun Heo /* create the initial workers */ 65773347fa09STejun Heo for_each_online_cpu(cpu) { 65783347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 65793347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 65803347fa09STejun Heo BUG_ON(!create_worker(pool)); 65813347fa09STejun Heo } 65823347fa09STejun Heo } 65833347fa09STejun Heo 65843347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 65853347fa09STejun Heo BUG_ON(!create_worker(pool)); 65863347fa09STejun Heo 65873347fa09STejun Heo wq_online = true; 658882607adcSTejun Heo wq_watchdog_init(); 65891da177e4SLinus Torvalds } 6590c4f135d6STetsuo Handa 6591025e1684STejun Heo /* 6592025e1684STejun Heo * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 6593025e1684STejun Heo * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 6594025e1684STejun Heo * and consecutive pod ID. The rest of @pt is initialized accordingly. 6595025e1684STejun Heo */ 6596025e1684STejun Heo static void __init init_pod_type(struct wq_pod_type *pt, 6597025e1684STejun Heo bool (*cpus_share_pod)(int, int)) 6598025e1684STejun Heo { 6599025e1684STejun Heo int cur, pre, cpu, pod; 6600025e1684STejun Heo 6601025e1684STejun Heo pt->nr_pods = 0; 6602025e1684STejun Heo 6603025e1684STejun Heo /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 6604025e1684STejun Heo pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6605025e1684STejun Heo BUG_ON(!pt->cpu_pod); 6606025e1684STejun Heo 6607025e1684STejun Heo for_each_possible_cpu(cur) { 6608025e1684STejun Heo for_each_possible_cpu(pre) { 6609025e1684STejun Heo if (pre >= cur) { 6610025e1684STejun Heo pt->cpu_pod[cur] = pt->nr_pods++; 6611025e1684STejun Heo break; 6612025e1684STejun Heo } 6613025e1684STejun Heo if (cpus_share_pod(cur, pre)) { 6614025e1684STejun Heo pt->cpu_pod[cur] = pt->cpu_pod[pre]; 6615025e1684STejun Heo break; 6616025e1684STejun Heo } 6617025e1684STejun Heo } 6618025e1684STejun Heo } 6619025e1684STejun Heo 6620025e1684STejun Heo /* init the rest to match @pt->cpu_pod[] */ 6621025e1684STejun Heo pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6622025e1684STejun Heo pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 6623025e1684STejun Heo BUG_ON(!pt->pod_cpus || !pt->pod_node); 6624025e1684STejun Heo 6625025e1684STejun Heo for (pod = 0; pod < pt->nr_pods; pod++) 6626025e1684STejun Heo BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 6627025e1684STejun Heo 6628025e1684STejun Heo for_each_possible_cpu(cpu) { 6629025e1684STejun Heo cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 6630025e1684STejun Heo pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 6631025e1684STejun Heo } 6632025e1684STejun Heo } 6633025e1684STejun Heo 663463c5484eSTejun Heo static bool __init cpus_dont_share(int cpu0, int cpu1) 663563c5484eSTejun Heo { 663663c5484eSTejun Heo return false; 663763c5484eSTejun Heo } 663863c5484eSTejun Heo 663963c5484eSTejun Heo static bool __init cpus_share_smt(int cpu0, int cpu1) 664063c5484eSTejun Heo { 664163c5484eSTejun Heo #ifdef CONFIG_SCHED_SMT 664263c5484eSTejun Heo return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 664363c5484eSTejun Heo #else 664463c5484eSTejun Heo return false; 664563c5484eSTejun Heo #endif 664663c5484eSTejun Heo } 664763c5484eSTejun Heo 6648025e1684STejun Heo static bool __init cpus_share_numa(int cpu0, int cpu1) 6649025e1684STejun Heo { 6650025e1684STejun Heo return cpu_to_node(cpu0) == cpu_to_node(cpu1); 6651025e1684STejun Heo } 6652025e1684STejun Heo 66532930155bSTejun Heo /** 66542930155bSTejun Heo * workqueue_init_topology - initialize CPU pods for unbound workqueues 66552930155bSTejun Heo * 66562930155bSTejun Heo * This is the third step of there-staged workqueue subsystem initialization and 66572930155bSTejun Heo * invoked after SMP and topology information are fully initialized. It 66582930155bSTejun Heo * initializes the unbound CPU pods accordingly. 66592930155bSTejun Heo */ 66602930155bSTejun Heo void __init workqueue_init_topology(void) 6661a86feae6STejun Heo { 66622930155bSTejun Heo struct workqueue_struct *wq; 6663025e1684STejun Heo int cpu; 6664a86feae6STejun Heo 666563c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 666663c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 666763c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 6668025e1684STejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 6669a86feae6STejun Heo 66702930155bSTejun Heo mutex_lock(&wq_pool_mutex); 6671a86feae6STejun Heo 6672a86feae6STejun Heo /* 66732930155bSTejun Heo * Workqueues allocated earlier would have all CPUs sharing the default 66742930155bSTejun Heo * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU 66752930155bSTejun Heo * combinations to apply per-pod sharing. 66762930155bSTejun Heo */ 66772930155bSTejun Heo list_for_each_entry(wq, &workqueues, list) { 66782930155bSTejun Heo for_each_online_cpu(cpu) { 66792930155bSTejun Heo wq_update_pod(wq, cpu, cpu, true); 66802930155bSTejun Heo } 66812930155bSTejun Heo } 66822930155bSTejun Heo 66832930155bSTejun Heo mutex_unlock(&wq_pool_mutex); 6684a86feae6STejun Heo } 6685a86feae6STejun Heo 668620bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 668720bdedafSTetsuo Handa { 668820bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 668920bdedafSTetsuo Handa dump_stack(); 669020bdedafSTetsuo Handa } 6691c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6692ace3c549Stiozhang 6693ace3c549Stiozhang static int __init workqueue_unbound_cpus_setup(char *str) 6694ace3c549Stiozhang { 6695ace3c549Stiozhang if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6696ace3c549Stiozhang cpumask_clear(&wq_cmdline_cpumask); 6697ace3c549Stiozhang pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6698ace3c549Stiozhang } 6699ace3c549Stiozhang 6700ace3c549Stiozhang return 1; 6701ace3c549Stiozhang } 6702ace3c549Stiozhang __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6703