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 */ 2148639ecebSTejun Heo PWQ_STAT_REPATRIATED, /* unbound workers brought back into scope */ 215725e8ec5STejun Heo PWQ_STAT_MAYDAY, /* maydays to rescuer */ 216725e8ec5STejun Heo PWQ_STAT_RESCUED, /* linked work items executed by rescuer */ 217725e8ec5STejun Heo 218725e8ec5STejun Heo PWQ_NR_STATS, 219725e8ec5STejun Heo }; 220725e8ec5STejun Heo 221725e8ec5STejun Heo /* 222112202d9STejun Heo * The per-pool workqueue. While queued, the lower WORK_STRUCT_FLAG_BITS 223112202d9STejun Heo * of work_struct->data are used for flags and the remaining high bits 224112202d9STejun Heo * point to the pwq; thus, pwqs need to be aligned at two's power of the 225112202d9STejun Heo * number of flag bits. 2261da177e4SLinus Torvalds */ 227112202d9STejun Heo struct pool_workqueue { 228bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2294690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 23073f53c4aSTejun Heo int work_color; /* L: current color */ 23173f53c4aSTejun Heo int flush_color; /* L: flushing color */ 2328864b4e5STejun Heo int refcnt; /* L: reference count */ 23373f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 23473f53c4aSTejun Heo /* L: nr of in_flight works */ 235018f3a13SLai Jiangshan 236018f3a13SLai Jiangshan /* 237018f3a13SLai Jiangshan * nr_active management and WORK_STRUCT_INACTIVE: 238018f3a13SLai Jiangshan * 239018f3a13SLai Jiangshan * When pwq->nr_active >= max_active, new work item is queued to 240018f3a13SLai Jiangshan * pwq->inactive_works instead of pool->worklist and marked with 241018f3a13SLai Jiangshan * WORK_STRUCT_INACTIVE. 242018f3a13SLai Jiangshan * 243018f3a13SLai Jiangshan * All work items marked with WORK_STRUCT_INACTIVE do not participate 244018f3a13SLai Jiangshan * in pwq->nr_active and all work items in pwq->inactive_works are 245018f3a13SLai Jiangshan * marked with WORK_STRUCT_INACTIVE. But not all WORK_STRUCT_INACTIVE 246018f3a13SLai Jiangshan * work items are in pwq->inactive_works. Some of them are ready to 247018f3a13SLai Jiangshan * run in pool->worklist or worker->scheduled. Those work itmes are 248018f3a13SLai Jiangshan * only struct wq_barrier which is used for flush_work() and should 249018f3a13SLai Jiangshan * not participate in pwq->nr_active. For non-barrier work item, it 250018f3a13SLai Jiangshan * is marked with WORK_STRUCT_INACTIVE iff it is in pwq->inactive_works. 251018f3a13SLai Jiangshan */ 2521e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 253a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 254f97a4a1aSLai Jiangshan struct list_head inactive_works; /* L: inactive works */ 2553c25a55dSLai Jiangshan struct list_head pwqs_node; /* WR: node on wq->pwqs */ 2562e109a28STejun Heo struct list_head mayday_node; /* MD: node on wq->maydays */ 2578864b4e5STejun Heo 258725e8ec5STejun Heo u64 stats[PWQ_NR_STATS]; 259725e8ec5STejun Heo 2608864b4e5STejun Heo /* 261967b494eSTejun Heo * Release of unbound pwq is punted to a kthread_worker. See put_pwq() 262687a9aa5STejun Heo * and pwq_release_workfn() for details. pool_workqueue itself is also 263687a9aa5STejun Heo * RCU protected so that the first pwq can be determined without 264967b494eSTejun Heo * grabbing wq->mutex. 2658864b4e5STejun Heo */ 266687a9aa5STejun Heo struct kthread_work release_work; 2678864b4e5STejun Heo struct rcu_head rcu; 268e904e6c2STejun Heo } __aligned(1 << WORK_STRUCT_FLAG_BITS); 2691da177e4SLinus Torvalds 2701da177e4SLinus Torvalds /* 27173f53c4aSTejun Heo * Structure used to wait for workqueue flush. 27273f53c4aSTejun Heo */ 27373f53c4aSTejun Heo struct wq_flusher { 2743c25a55dSLai Jiangshan struct list_head list; /* WQ: list of flushers */ 2753c25a55dSLai Jiangshan int flush_color; /* WQ: flush color waiting for */ 27673f53c4aSTejun Heo struct completion done; /* flush completion */ 27773f53c4aSTejun Heo }; 2781da177e4SLinus Torvalds 279226223abSTejun Heo struct wq_device; 280226223abSTejun Heo 28173f53c4aSTejun Heo /* 282c5aa87bbSTejun Heo * The externally visible workqueue. It relays the issued work items to 283c5aa87bbSTejun Heo * the appropriate worker_pool through its pool_workqueues. 2841da177e4SLinus Torvalds */ 2851da177e4SLinus Torvalds struct workqueue_struct { 2863c25a55dSLai Jiangshan struct list_head pwqs; /* WR: all pwqs of this wq */ 287e2dca7adSTejun Heo struct list_head list; /* PR: list of all workqueues */ 28873f53c4aSTejun Heo 2893c25a55dSLai Jiangshan struct mutex mutex; /* protects this wq */ 2903c25a55dSLai Jiangshan int work_color; /* WQ: current work color */ 2913c25a55dSLai Jiangshan int flush_color; /* WQ: current flush color */ 292112202d9STejun Heo atomic_t nr_pwqs_to_flush; /* flush in progress */ 2933c25a55dSLai Jiangshan struct wq_flusher *first_flusher; /* WQ: first flusher */ 2943c25a55dSLai Jiangshan struct list_head flusher_queue; /* WQ: flush waiters */ 2953c25a55dSLai Jiangshan struct list_head flusher_overflow; /* WQ: flush overflow list */ 29673f53c4aSTejun Heo 2972e109a28STejun Heo struct list_head maydays; /* MD: pwqs requesting rescue */ 29830ae2fc0STejun Heo struct worker *rescuer; /* MD: rescue worker */ 299e22bee78STejun Heo 30087fc741eSLai Jiangshan int nr_drainers; /* WQ: drain in progress */ 301a357fc03SLai Jiangshan int saved_max_active; /* WQ: saved pwq max_active */ 302226223abSTejun Heo 3035b95e1afSLai Jiangshan struct workqueue_attrs *unbound_attrs; /* PW: only for unbound wqs */ 3045b95e1afSLai Jiangshan struct pool_workqueue *dfl_pwq; /* PW: only for unbound wqs */ 3056029a918STejun Heo 306226223abSTejun Heo #ifdef CONFIG_SYSFS 307226223abSTejun Heo struct wq_device *wq_dev; /* I: for sysfs interface */ 308226223abSTejun Heo #endif 3094e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 310669de8bdSBart Van Assche char *lock_name; 311669de8bdSBart Van Assche struct lock_class_key key; 3124e6045f1SJohannes Berg struct lockdep_map lockdep_map; 3134e6045f1SJohannes Berg #endif 314ecf6881fSTejun Heo char name[WQ_NAME_LEN]; /* I: workqueue name */ 3152728fd2fSTejun Heo 316e2dca7adSTejun Heo /* 31724acfb71SThomas Gleixner * Destruction of workqueue_struct is RCU protected to allow walking 31824acfb71SThomas Gleixner * the workqueues list without grabbing wq_pool_mutex. 319e2dca7adSTejun Heo * This is used to dump all workqueues from sysrq. 320e2dca7adSTejun Heo */ 321e2dca7adSTejun Heo struct rcu_head rcu; 322e2dca7adSTejun Heo 3232728fd2fSTejun Heo /* hot fields used during command issue, aligned to cacheline */ 3242728fd2fSTejun Heo unsigned int flags ____cacheline_aligned; /* WQ: WQ_* flags */ 325636b927eSTejun Heo struct pool_workqueue __percpu __rcu **cpu_pwq; /* I: per-cpu pwqs */ 3261da177e4SLinus Torvalds }; 3271da177e4SLinus Torvalds 328e904e6c2STejun Heo static struct kmem_cache *pwq_cache; 329e904e6c2STejun Heo 33084193c07STejun Heo /* 33184193c07STejun Heo * Each pod type describes how CPUs should be grouped for unbound workqueues. 33284193c07STejun Heo * See the comment above workqueue_attrs->affn_scope. 33384193c07STejun Heo */ 33484193c07STejun Heo struct wq_pod_type { 33584193c07STejun Heo int nr_pods; /* number of pods */ 33684193c07STejun Heo cpumask_var_t *pod_cpus; /* pod -> cpus */ 33784193c07STejun Heo int *pod_node; /* pod -> node */ 33884193c07STejun Heo int *cpu_pod; /* cpu -> pod */ 33984193c07STejun Heo }; 34084193c07STejun Heo 34184193c07STejun Heo static struct wq_pod_type wq_pod_types[WQ_AFFN_NR_TYPES]; 342*523a301eSTejun Heo static enum wq_affn_scope wq_affn_dfl = WQ_AFFN_CACHE; 34363c5484eSTejun Heo 34463c5484eSTejun Heo static const char *wq_affn_names[WQ_AFFN_NR_TYPES] = { 345*523a301eSTejun Heo [WQ_AFFN_DFL] = "default", 34663c5484eSTejun Heo [WQ_AFFN_CPU] = "cpu", 34763c5484eSTejun Heo [WQ_AFFN_SMT] = "smt", 34863c5484eSTejun Heo [WQ_AFFN_CACHE] = "cache", 34963c5484eSTejun Heo [WQ_AFFN_NUMA] = "numa", 35063c5484eSTejun Heo [WQ_AFFN_SYSTEM] = "system", 35163c5484eSTejun Heo }; 352bce90380STejun Heo 353616db877STejun Heo /* 354616db877STejun Heo * Per-cpu work items which run for longer than the following threshold are 355616db877STejun Heo * automatically considered CPU intensive and excluded from concurrency 356616db877STejun Heo * management to prevent them from noticeably delaying other per-cpu work items. 357aa6fde93STejun Heo * ULONG_MAX indicates that the user hasn't overridden it with a boot parameter. 358aa6fde93STejun Heo * The actual value is initialized in wq_cpu_intensive_thresh_init(). 359616db877STejun Heo */ 360aa6fde93STejun Heo static unsigned long wq_cpu_intensive_thresh_us = ULONG_MAX; 361616db877STejun Heo module_param_named(cpu_intensive_thresh_us, wq_cpu_intensive_thresh_us, ulong, 0644); 362616db877STejun Heo 363cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 364552f530cSLuis R. Rodriguez static bool wq_power_efficient = IS_ENABLED(CONFIG_WQ_POWER_EFFICIENT_DEFAULT); 365cee22a15SViresh Kumar module_param_named(power_efficient, wq_power_efficient, bool, 0444); 366cee22a15SViresh Kumar 367863b710bSTejun Heo static bool wq_online; /* can kworkers be created yet? */ 3683347fa09STejun Heo 369fef59c9cSTejun Heo /* buf for wq_update_unbound_pod_attrs(), protected by CPU hotplug exclusion */ 370fef59c9cSTejun Heo static struct workqueue_attrs *wq_update_pod_attrs_buf; 3714c16bd32STejun Heo 37268e13a67SLai Jiangshan static DEFINE_MUTEX(wq_pool_mutex); /* protects pools and workqueues list */ 3731258fae7STejun Heo static DEFINE_MUTEX(wq_pool_attach_mutex); /* protects worker attach/detach */ 374a9b8a985SSebastian Andrzej Siewior static DEFINE_RAW_SPINLOCK(wq_mayday_lock); /* protects wq->maydays list */ 375d8bb65abSSebastian Andrzej Siewior /* wait for manager to go away */ 376d8bb65abSSebastian Andrzej Siewior static struct rcuwait manager_wait = __RCUWAIT_INITIALIZER(manager_wait); 3775bcab335STejun Heo 378e2dca7adSTejun Heo static LIST_HEAD(workqueues); /* PR: list of all workqueues */ 37968e13a67SLai Jiangshan static bool workqueue_freezing; /* PL: have wqs started freezing? */ 3807d19c5ceSTejun Heo 38199c621efSLai Jiangshan /* PL&A: allowable cpus for unbound wqs and work items */ 382ef557180SMike Galbraith static cpumask_var_t wq_unbound_cpumask; 383ef557180SMike Galbraith 384ace3c549Stiozhang /* for further constrain wq_unbound_cpumask by cmdline parameter*/ 385ace3c549Stiozhang static struct cpumask wq_cmdline_cpumask __initdata; 386ace3c549Stiozhang 387ef557180SMike Galbraith /* CPU where unbound work was last round robin scheduled from this CPU */ 388ef557180SMike Galbraith static DEFINE_PER_CPU(int, wq_rr_cpu_last); 389b05a7928SFrederic Weisbecker 390f303fccbSTejun Heo /* 391f303fccbSTejun Heo * Local execution of unbound work items is no longer guaranteed. The 392f303fccbSTejun Heo * following always forces round-robin CPU selection on unbound work items 393f303fccbSTejun Heo * to uncover usages which depend on it. 394f303fccbSTejun Heo */ 395f303fccbSTejun Heo #ifdef CONFIG_DEBUG_WQ_FORCE_RR_CPU 396f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = true; 397f303fccbSTejun Heo #else 398f303fccbSTejun Heo static bool wq_debug_force_rr_cpu = false; 399f303fccbSTejun Heo #endif 400f303fccbSTejun Heo module_param_named(debug_force_rr_cpu, wq_debug_force_rr_cpu, bool, 0644); 401f303fccbSTejun Heo 4027d19c5ceSTejun Heo /* the per-cpu worker pools */ 40325528213SPeter Zijlstra static DEFINE_PER_CPU_SHARED_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], cpu_worker_pools); 4047d19c5ceSTejun Heo 40568e13a67SLai Jiangshan static DEFINE_IDR(worker_pool_idr); /* PR: idr of all pools */ 4067d19c5ceSTejun Heo 40768e13a67SLai Jiangshan /* PL: hash of all unbound pools keyed by pool->attrs */ 40829c91e99STejun Heo static DEFINE_HASHTABLE(unbound_pool_hash, UNBOUND_POOL_HASH_ORDER); 40929c91e99STejun Heo 410c5aa87bbSTejun Heo /* I: attributes used when instantiating standard unbound pools on demand */ 41129c91e99STejun Heo static struct workqueue_attrs *unbound_std_wq_attrs[NR_STD_WORKER_POOLS]; 41229c91e99STejun Heo 4138a2b7538STejun Heo /* I: attributes used when instantiating ordered pools on demand */ 4148a2b7538STejun Heo static struct workqueue_attrs *ordered_wq_attrs[NR_STD_WORKER_POOLS]; 4158a2b7538STejun Heo 416967b494eSTejun Heo /* 417967b494eSTejun Heo * I: kthread_worker to release pwq's. pwq release needs to be bounced to a 418967b494eSTejun Heo * process context while holding a pool lock. Bounce to a dedicated kthread 419967b494eSTejun Heo * worker to avoid A-A deadlocks. 420967b494eSTejun Heo */ 421967b494eSTejun Heo static struct kthread_worker *pwq_release_worker; 422967b494eSTejun Heo 423d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 424ad7b1f84SMarc Dionne EXPORT_SYMBOL(system_wq); 425044c782cSValentin Ilie struct workqueue_struct *system_highpri_wq __read_mostly; 4261aabe902SJoonsoo Kim EXPORT_SYMBOL_GPL(system_highpri_wq); 427044c782cSValentin Ilie struct workqueue_struct *system_long_wq __read_mostly; 428d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 429044c782cSValentin Ilie struct workqueue_struct *system_unbound_wq __read_mostly; 430f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 431044c782cSValentin Ilie struct workqueue_struct *system_freezable_wq __read_mostly; 43224d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 4330668106cSViresh Kumar struct workqueue_struct *system_power_efficient_wq __read_mostly; 4340668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_power_efficient_wq); 4350668106cSViresh Kumar struct workqueue_struct *system_freezable_power_efficient_wq __read_mostly; 4360668106cSViresh Kumar EXPORT_SYMBOL_GPL(system_freezable_power_efficient_wq); 437d320c038STejun Heo 4387d19c5ceSTejun Heo static int worker_thread(void *__worker); 4396ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq); 440c29eb853STejun Heo static void show_pwq(struct pool_workqueue *pwq); 44155df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool); 4427d19c5ceSTejun Heo 44397bd2347STejun Heo #define CREATE_TRACE_POINTS 44497bd2347STejun Heo #include <trace/events/workqueue.h> 44597bd2347STejun Heo 44668e13a67SLai Jiangshan #define assert_rcu_or_pool_mutex() \ 44724acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 448f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 44924acfb71SThomas Gleixner "RCU or wq_pool_mutex should be held") 4505bcab335STejun Heo 4515b95e1afSLai Jiangshan #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \ 45224acfb71SThomas Gleixner RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \ 453f78f5b90SPaul E. McKenney !lockdep_is_held(&wq->mutex) && \ 454f78f5b90SPaul E. McKenney !lockdep_is_held(&wq_pool_mutex), \ 45524acfb71SThomas Gleixner "RCU, wq->mutex or wq_pool_mutex should be held") 4565b95e1afSLai Jiangshan 457f02ae73aSTejun Heo #define for_each_cpu_worker_pool(pool, cpu) \ 458f02ae73aSTejun Heo for ((pool) = &per_cpu(cpu_worker_pools, cpu)[0]; \ 459f02ae73aSTejun Heo (pool) < &per_cpu(cpu_worker_pools, cpu)[NR_STD_WORKER_POOLS]; \ 4607a62c2c8STejun Heo (pool)++) 4614ce62e9eSTejun Heo 46249e3cf44STejun Heo /** 46317116969STejun Heo * for_each_pool - iterate through all worker_pools in the system 46417116969STejun Heo * @pool: iteration cursor 465611c92a0STejun Heo * @pi: integer used for iteration 466fa1b54e6STejun Heo * 46724acfb71SThomas Gleixner * This must be called either with wq_pool_mutex held or RCU read 46868e13a67SLai Jiangshan * locked. If the pool needs to be used beyond the locking in effect, the 46968e13a67SLai Jiangshan * caller is responsible for guaranteeing that the pool stays online. 470fa1b54e6STejun Heo * 471fa1b54e6STejun Heo * The if/else clause exists only for the lockdep assertion and can be 472fa1b54e6STejun Heo * ignored. 47317116969STejun Heo */ 474611c92a0STejun Heo #define for_each_pool(pool, pi) \ 475611c92a0STejun Heo idr_for_each_entry(&worker_pool_idr, pool, pi) \ 47668e13a67SLai Jiangshan if (({ assert_rcu_or_pool_mutex(); false; })) { } \ 477fa1b54e6STejun Heo else 47817116969STejun Heo 47917116969STejun Heo /** 480822d8405STejun Heo * for_each_pool_worker - iterate through all workers of a worker_pool 481822d8405STejun Heo * @worker: iteration cursor 482822d8405STejun Heo * @pool: worker_pool to iterate workers of 483822d8405STejun Heo * 4841258fae7STejun Heo * This must be called with wq_pool_attach_mutex. 485822d8405STejun Heo * 486822d8405STejun Heo * The if/else clause exists only for the lockdep assertion and can be 487822d8405STejun Heo * ignored. 488822d8405STejun Heo */ 489da028469SLai Jiangshan #define for_each_pool_worker(worker, pool) \ 490da028469SLai Jiangshan list_for_each_entry((worker), &(pool)->workers, node) \ 4911258fae7STejun Heo if (({ lockdep_assert_held(&wq_pool_attach_mutex); false; })) { } \ 492822d8405STejun Heo else 493822d8405STejun Heo 494822d8405STejun Heo /** 49549e3cf44STejun Heo * for_each_pwq - iterate through all pool_workqueues of the specified workqueue 49649e3cf44STejun Heo * @pwq: iteration cursor 49749e3cf44STejun Heo * @wq: the target workqueue 49876af4d93STejun Heo * 49924acfb71SThomas Gleixner * This must be called either with wq->mutex held or RCU read locked. 500794b18bcSTejun Heo * If the pwq needs to be used beyond the locking in effect, the caller is 501794b18bcSTejun Heo * responsible for guaranteeing that the pwq stays online. 50276af4d93STejun Heo * 50376af4d93STejun Heo * The if/else clause exists only for the lockdep assertion and can be 50476af4d93STejun Heo * ignored. 50549e3cf44STejun Heo */ 50649e3cf44STejun Heo #define for_each_pwq(pwq, wq) \ 50749e9d1a9SSebastian Andrzej Siewior list_for_each_entry_rcu((pwq), &(wq)->pwqs, pwqs_node, \ 5085a644662SJoel Fernandes (Google) lockdep_is_held(&(wq->mutex))) 509f3421797STejun Heo 510dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 511dc186ad7SThomas Gleixner 512f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr; 513dc186ad7SThomas Gleixner 51499777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 51599777288SStanislaw Gruszka { 51699777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 51799777288SStanislaw Gruszka } 51899777288SStanislaw Gruszka 519b9fdac7fSDu, Changbin static bool work_is_static_object(void *addr) 520b9fdac7fSDu, Changbin { 521b9fdac7fSDu, Changbin struct work_struct *work = addr; 522b9fdac7fSDu, Changbin 523b9fdac7fSDu, Changbin return test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work)); 524b9fdac7fSDu, Changbin } 525b9fdac7fSDu, Changbin 526dc186ad7SThomas Gleixner /* 527dc186ad7SThomas Gleixner * fixup_init is called when: 528dc186ad7SThomas Gleixner * - an active object is initialized 529dc186ad7SThomas Gleixner */ 53002a982a6SDu, Changbin static bool work_fixup_init(void *addr, enum debug_obj_state state) 531dc186ad7SThomas Gleixner { 532dc186ad7SThomas Gleixner struct work_struct *work = addr; 533dc186ad7SThomas Gleixner 534dc186ad7SThomas Gleixner switch (state) { 535dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 536dc186ad7SThomas Gleixner cancel_work_sync(work); 537dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 53802a982a6SDu, Changbin return true; 539dc186ad7SThomas Gleixner default: 54002a982a6SDu, Changbin return false; 541dc186ad7SThomas Gleixner } 542dc186ad7SThomas Gleixner } 543dc186ad7SThomas Gleixner 544dc186ad7SThomas Gleixner /* 545dc186ad7SThomas Gleixner * fixup_free is called when: 546dc186ad7SThomas Gleixner * - an active object is freed 547dc186ad7SThomas Gleixner */ 54802a982a6SDu, Changbin static bool work_fixup_free(void *addr, enum debug_obj_state state) 549dc186ad7SThomas Gleixner { 550dc186ad7SThomas Gleixner struct work_struct *work = addr; 551dc186ad7SThomas Gleixner 552dc186ad7SThomas Gleixner switch (state) { 553dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 554dc186ad7SThomas Gleixner cancel_work_sync(work); 555dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 55602a982a6SDu, Changbin return true; 557dc186ad7SThomas Gleixner default: 55802a982a6SDu, Changbin return false; 559dc186ad7SThomas Gleixner } 560dc186ad7SThomas Gleixner } 561dc186ad7SThomas Gleixner 562f9e62f31SStephen Boyd static const struct debug_obj_descr work_debug_descr = { 563dc186ad7SThomas Gleixner .name = "work_struct", 56499777288SStanislaw Gruszka .debug_hint = work_debug_hint, 565b9fdac7fSDu, Changbin .is_static_object = work_is_static_object, 566dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 567dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 568dc186ad7SThomas Gleixner }; 569dc186ad7SThomas Gleixner 570dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 571dc186ad7SThomas Gleixner { 572dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 573dc186ad7SThomas Gleixner } 574dc186ad7SThomas Gleixner 575dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 576dc186ad7SThomas Gleixner { 577dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 578dc186ad7SThomas Gleixner } 579dc186ad7SThomas Gleixner 580dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 581dc186ad7SThomas Gleixner { 582dc186ad7SThomas Gleixner if (onstack) 583dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 584dc186ad7SThomas Gleixner else 585dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 586dc186ad7SThomas Gleixner } 587dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 588dc186ad7SThomas Gleixner 589dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 590dc186ad7SThomas Gleixner { 591dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 592dc186ad7SThomas Gleixner } 593dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 594dc186ad7SThomas Gleixner 595ea2e64f2SThomas Gleixner void destroy_delayed_work_on_stack(struct delayed_work *work) 596ea2e64f2SThomas Gleixner { 597ea2e64f2SThomas Gleixner destroy_timer_on_stack(&work->timer); 598ea2e64f2SThomas Gleixner debug_object_free(&work->work, &work_debug_descr); 599ea2e64f2SThomas Gleixner } 600ea2e64f2SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_delayed_work_on_stack); 601ea2e64f2SThomas Gleixner 602dc186ad7SThomas Gleixner #else 603dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 604dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 605dc186ad7SThomas Gleixner #endif 606dc186ad7SThomas Gleixner 6074e8b22bdSLi Bin /** 60867dc8325SCai Huoqing * worker_pool_assign_id - allocate ID and assign it to @pool 6094e8b22bdSLi Bin * @pool: the pool pointer of interest 6104e8b22bdSLi Bin * 6114e8b22bdSLi Bin * Returns 0 if ID in [0, WORK_OFFQ_POOL_NONE) is allocated and assigned 6124e8b22bdSLi Bin * successfully, -errno on failure. 6134e8b22bdSLi Bin */ 6149daf9e67STejun Heo static int worker_pool_assign_id(struct worker_pool *pool) 6159daf9e67STejun Heo { 6169daf9e67STejun Heo int ret; 6179daf9e67STejun Heo 61868e13a67SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 6195bcab335STejun Heo 6204e8b22bdSLi Bin ret = idr_alloc(&worker_pool_idr, pool, 0, WORK_OFFQ_POOL_NONE, 6214e8b22bdSLi Bin GFP_KERNEL); 622229641a6STejun Heo if (ret >= 0) { 623e68035fbSTejun Heo pool->id = ret; 624229641a6STejun Heo return 0; 625229641a6STejun Heo } 6269daf9e67STejun Heo return ret; 6279daf9e67STejun Heo } 6289daf9e67STejun Heo 62973f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 63073f53c4aSTejun Heo { 63173f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 63273f53c4aSTejun Heo } 63373f53c4aSTejun Heo 634c4560c2cSLai Jiangshan static int get_work_color(unsigned long work_data) 63573f53c4aSTejun Heo { 636c4560c2cSLai Jiangshan return (work_data >> WORK_STRUCT_COLOR_SHIFT) & 63773f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 63873f53c4aSTejun Heo } 63973f53c4aSTejun Heo 64073f53c4aSTejun Heo static int work_next_color(int color) 64173f53c4aSTejun Heo { 64273f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 643a848e3b6SOleg Nesterov } 644a848e3b6SOleg Nesterov 6454594bf15SDavid Howells /* 646112202d9STejun Heo * While queued, %WORK_STRUCT_PWQ is set and non flag bits of a work's data 647112202d9STejun Heo * contain the pointer to the queued pwq. Once execution starts, the flag 6487c3eed5cSTejun Heo * is cleared and the high bits contain OFFQ flags and pool ID. 6497a22ad75STejun Heo * 650112202d9STejun Heo * set_work_pwq(), set_work_pool_and_clear_pending(), mark_work_canceling() 651112202d9STejun Heo * and clear_work_data() can be used to set the pwq, pool or clear 652bbb68dfaSTejun Heo * work->data. These functions should only be called while the work is 653bbb68dfaSTejun Heo * owned - ie. while the PENDING bit is set. 6547a22ad75STejun Heo * 655112202d9STejun Heo * get_work_pool() and get_work_pwq() can be used to obtain the pool or pwq 6567c3eed5cSTejun Heo * corresponding to a work. Pool is available once the work has been 657112202d9STejun Heo * queued anywhere after initialization until it is sync canceled. pwq is 6587c3eed5cSTejun Heo * available only while the work item is queued. 659bbb68dfaSTejun Heo * 660bbb68dfaSTejun Heo * %WORK_OFFQ_CANCELING is used to mark a work item which is being 661bbb68dfaSTejun Heo * canceled. While being canceled, a work item may have its PENDING set 662bbb68dfaSTejun Heo * but stay off timer and worklist for arbitrarily long and nobody should 663bbb68dfaSTejun Heo * try to steal the PENDING bit. 6644594bf15SDavid Howells */ 6657a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 6667a22ad75STejun Heo unsigned long flags) 6677a22ad75STejun Heo { 6686183c009STejun Heo WARN_ON_ONCE(!work_pending(work)); 6697a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 6707a22ad75STejun Heo } 6717a22ad75STejun Heo 672112202d9STejun Heo static void set_work_pwq(struct work_struct *work, struct pool_workqueue *pwq, 6734690c4abSTejun Heo unsigned long extra_flags) 674365970a1SDavid Howells { 675112202d9STejun Heo set_work_data(work, (unsigned long)pwq, 676112202d9STejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_PWQ | extra_flags); 677365970a1SDavid Howells } 678365970a1SDavid Howells 6794468a00fSLai Jiangshan static void set_work_pool_and_keep_pending(struct work_struct *work, 6804468a00fSLai Jiangshan int pool_id) 6814468a00fSLai Jiangshan { 6824468a00fSLai Jiangshan set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 6834468a00fSLai Jiangshan WORK_STRUCT_PENDING); 6844468a00fSLai Jiangshan } 6854468a00fSLai Jiangshan 6867c3eed5cSTejun Heo static void set_work_pool_and_clear_pending(struct work_struct *work, 6877c3eed5cSTejun Heo int pool_id) 6884d707b9fSOleg Nesterov { 68923657bb1STejun Heo /* 69023657bb1STejun Heo * The following wmb is paired with the implied mb in 69123657bb1STejun Heo * test_and_set_bit(PENDING) and ensures all updates to @work made 69223657bb1STejun Heo * here are visible to and precede any updates by the next PENDING 69323657bb1STejun Heo * owner. 69423657bb1STejun Heo */ 69523657bb1STejun Heo smp_wmb(); 6967c3eed5cSTejun Heo set_work_data(work, (unsigned long)pool_id << WORK_OFFQ_POOL_SHIFT, 0); 697346c09f8SRoman Pen /* 698346c09f8SRoman Pen * The following mb guarantees that previous clear of a PENDING bit 699346c09f8SRoman Pen * will not be reordered with any speculative LOADS or STORES from 700346c09f8SRoman Pen * work->current_func, which is executed afterwards. This possible 7018bdc6201SLiu Song * reordering can lead to a missed execution on attempt to queue 702346c09f8SRoman Pen * the same @work. E.g. consider this case: 703346c09f8SRoman Pen * 704346c09f8SRoman Pen * CPU#0 CPU#1 705346c09f8SRoman Pen * ---------------------------- -------------------------------- 706346c09f8SRoman Pen * 707346c09f8SRoman Pen * 1 STORE event_indicated 708346c09f8SRoman Pen * 2 queue_work_on() { 709346c09f8SRoman Pen * 3 test_and_set_bit(PENDING) 710346c09f8SRoman Pen * 4 } set_..._and_clear_pending() { 711346c09f8SRoman Pen * 5 set_work_data() # clear bit 712346c09f8SRoman Pen * 6 smp_mb() 713346c09f8SRoman Pen * 7 work->current_func() { 714346c09f8SRoman Pen * 8 LOAD event_indicated 715346c09f8SRoman Pen * } 716346c09f8SRoman Pen * 717346c09f8SRoman Pen * Without an explicit full barrier speculative LOAD on line 8 can 718346c09f8SRoman Pen * be executed before CPU#0 does STORE on line 1. If that happens, 719346c09f8SRoman Pen * CPU#0 observes the PENDING bit is still set and new execution of 720346c09f8SRoman Pen * a @work is not queued in a hope, that CPU#1 will eventually 721346c09f8SRoman Pen * finish the queued @work. Meanwhile CPU#1 does not see 722346c09f8SRoman Pen * event_indicated is set, because speculative LOAD was executed 723346c09f8SRoman Pen * before actual STORE. 724346c09f8SRoman Pen */ 725346c09f8SRoman Pen smp_mb(); 7264d707b9fSOleg Nesterov } 7274d707b9fSOleg Nesterov 7287a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 729365970a1SDavid Howells { 7307c3eed5cSTejun Heo smp_wmb(); /* see set_work_pool_and_clear_pending() */ 7317c3eed5cSTejun Heo set_work_data(work, WORK_STRUCT_NO_POOL, 0); 7327a22ad75STejun Heo } 7337a22ad75STejun Heo 734afa4bb77SLinus Torvalds static inline struct pool_workqueue *work_struct_pwq(unsigned long data) 735afa4bb77SLinus Torvalds { 736afa4bb77SLinus Torvalds return (struct pool_workqueue *)(data & WORK_STRUCT_WQ_DATA_MASK); 737afa4bb77SLinus Torvalds } 738afa4bb77SLinus Torvalds 739112202d9STejun Heo static struct pool_workqueue *get_work_pwq(struct work_struct *work) 7407a22ad75STejun Heo { 741e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7427a22ad75STejun Heo 743112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 744afa4bb77SLinus Torvalds return work_struct_pwq(data); 745e120153dSTejun Heo else 746e120153dSTejun Heo return NULL; 7477a22ad75STejun Heo } 7487a22ad75STejun Heo 7497c3eed5cSTejun Heo /** 7507c3eed5cSTejun Heo * get_work_pool - return the worker_pool a given work was associated with 7517c3eed5cSTejun Heo * @work: the work item of interest 7527c3eed5cSTejun Heo * 75368e13a67SLai Jiangshan * Pools are created and destroyed under wq_pool_mutex, and allows read 75424acfb71SThomas Gleixner * access under RCU read lock. As such, this function should be 75524acfb71SThomas Gleixner * called under wq_pool_mutex or inside of a rcu_read_lock() region. 756fa1b54e6STejun Heo * 757fa1b54e6STejun Heo * All fields of the returned pool are accessible as long as the above 758fa1b54e6STejun Heo * mentioned locking is in effect. If the returned pool needs to be used 759fa1b54e6STejun Heo * beyond the critical section, the caller is responsible for ensuring the 760fa1b54e6STejun Heo * returned pool is and stays online. 761d185af30SYacine Belkadi * 762d185af30SYacine Belkadi * Return: The worker_pool @work was last associated with. %NULL if none. 7637c3eed5cSTejun Heo */ 7647c3eed5cSTejun Heo static struct worker_pool *get_work_pool(struct work_struct *work) 7657a22ad75STejun Heo { 766e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 7677c3eed5cSTejun Heo int pool_id; 7687a22ad75STejun Heo 76968e13a67SLai Jiangshan assert_rcu_or_pool_mutex(); 770fa1b54e6STejun Heo 771112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 772afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool; 7737a22ad75STejun Heo 7747c3eed5cSTejun Heo pool_id = data >> WORK_OFFQ_POOL_SHIFT; 7757c3eed5cSTejun Heo if (pool_id == WORK_OFFQ_POOL_NONE) 7767a22ad75STejun Heo return NULL; 7777a22ad75STejun Heo 778fa1b54e6STejun Heo return idr_find(&worker_pool_idr, pool_id); 7797c3eed5cSTejun Heo } 7807c3eed5cSTejun Heo 7817c3eed5cSTejun Heo /** 7827c3eed5cSTejun Heo * get_work_pool_id - return the worker pool ID a given work is associated with 7837c3eed5cSTejun Heo * @work: the work item of interest 7847c3eed5cSTejun Heo * 785d185af30SYacine Belkadi * Return: The worker_pool ID @work was last associated with. 7867c3eed5cSTejun Heo * %WORK_OFFQ_POOL_NONE if none. 7877c3eed5cSTejun Heo */ 7887c3eed5cSTejun Heo static int get_work_pool_id(struct work_struct *work) 7897c3eed5cSTejun Heo { 79054d5b7d0SLai Jiangshan unsigned long data = atomic_long_read(&work->data); 7917c3eed5cSTejun Heo 792112202d9STejun Heo if (data & WORK_STRUCT_PWQ) 793afa4bb77SLinus Torvalds return work_struct_pwq(data)->pool->id; 79454d5b7d0SLai Jiangshan 79554d5b7d0SLai Jiangshan return data >> WORK_OFFQ_POOL_SHIFT; 7967c3eed5cSTejun Heo } 7977c3eed5cSTejun Heo 798bbb68dfaSTejun Heo static void mark_work_canceling(struct work_struct *work) 799bbb68dfaSTejun Heo { 8007c3eed5cSTejun Heo unsigned long pool_id = get_work_pool_id(work); 801bbb68dfaSTejun Heo 8027c3eed5cSTejun Heo pool_id <<= WORK_OFFQ_POOL_SHIFT; 8037c3eed5cSTejun Heo set_work_data(work, pool_id | WORK_OFFQ_CANCELING, WORK_STRUCT_PENDING); 804bbb68dfaSTejun Heo } 805bbb68dfaSTejun Heo 806bbb68dfaSTejun Heo static bool work_is_canceling(struct work_struct *work) 807bbb68dfaSTejun Heo { 808bbb68dfaSTejun Heo unsigned long data = atomic_long_read(&work->data); 809bbb68dfaSTejun Heo 810112202d9STejun Heo return !(data & WORK_STRUCT_PWQ) && (data & WORK_OFFQ_CANCELING); 811bbb68dfaSTejun Heo } 812bbb68dfaSTejun Heo 813e22bee78STejun Heo /* 8143270476aSTejun Heo * Policy functions. These define the policies on how the global worker 8153270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 816d565ed63STejun Heo * they're being called with pool->lock held. 817e22bee78STejun Heo */ 818e22bee78STejun Heo 819e22bee78STejun Heo /* 820e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 821e22bee78STejun Heo * running workers. 822974271c4STejun Heo * 823974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 824706026c2STejun Heo * function will always return %true for unbound pools as long as the 825974271c4STejun Heo * worklist isn't empty. 826e22bee78STejun Heo */ 82763d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 828e22bee78STejun Heo { 8290219a352STejun Heo return !list_empty(&pool->worklist) && !pool->nr_running; 830e22bee78STejun Heo } 831e22bee78STejun Heo 832e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 83363d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 834e22bee78STejun Heo { 83563d95a91STejun Heo return pool->nr_idle; 836e22bee78STejun Heo } 837e22bee78STejun Heo 838e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 83963d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 840e22bee78STejun Heo { 841bc35f7efSLai Jiangshan return !list_empty(&pool->worklist) && (pool->nr_running <= 1); 842e22bee78STejun Heo } 843e22bee78STejun Heo 844e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 84563d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 846e22bee78STejun Heo { 84763d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 848e22bee78STejun Heo } 849e22bee78STejun Heo 850e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 85163d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 852e22bee78STejun Heo { 853692b4825STejun Heo bool managing = pool->flags & POOL_MANAGER_ACTIVE; 85463d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 85563d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 856e22bee78STejun Heo 857e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 858e22bee78STejun Heo } 859e22bee78STejun Heo 8604690c4abSTejun Heo /** 861e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 862cb444766STejun Heo * @worker: self 863d302f017STejun Heo * @flags: flags to set 864d302f017STejun Heo * 865228f1d00SLai Jiangshan * Set @flags in @worker->flags and adjust nr_running accordingly. 866d302f017STejun Heo */ 867228f1d00SLai Jiangshan static inline void worker_set_flags(struct worker *worker, unsigned int flags) 868d302f017STejun Heo { 869bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 870e22bee78STejun Heo 871bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 872cb444766STejun Heo 873228f1d00SLai Jiangshan /* If transitioning into NOT_RUNNING, adjust nr_running. */ 874e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 875e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 876bc35f7efSLai Jiangshan pool->nr_running--; 877e22bee78STejun Heo } 878e22bee78STejun Heo 879d302f017STejun Heo worker->flags |= flags; 880d302f017STejun Heo } 881d302f017STejun Heo 882d302f017STejun Heo /** 883e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 884cb444766STejun Heo * @worker: self 885d302f017STejun Heo * @flags: flags to clear 886d302f017STejun Heo * 887e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 888d302f017STejun Heo */ 889d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 890d302f017STejun Heo { 89163d95a91STejun Heo struct worker_pool *pool = worker->pool; 892e22bee78STejun Heo unsigned int oflags = worker->flags; 893e22bee78STejun Heo 894bc8b50c2STejun Heo lockdep_assert_held(&pool->lock); 895cb444766STejun Heo 896d302f017STejun Heo worker->flags &= ~flags; 897e22bee78STejun Heo 89842c025f3STejun Heo /* 89942c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 90042c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 90142c025f3STejun Heo * of multiple flags, not a single flag. 90242c025f3STejun Heo */ 903e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 904e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 905bc35f7efSLai Jiangshan pool->nr_running++; 906d302f017STejun Heo } 907d302f017STejun Heo 908797e8345STejun Heo /* Return the first idle worker. Called with pool->lock held. */ 909797e8345STejun Heo static struct worker *first_idle_worker(struct worker_pool *pool) 910797e8345STejun Heo { 911797e8345STejun Heo if (unlikely(list_empty(&pool->idle_list))) 912797e8345STejun Heo return NULL; 913797e8345STejun Heo 914797e8345STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 915797e8345STejun Heo } 916797e8345STejun Heo 917797e8345STejun Heo /** 918797e8345STejun Heo * worker_enter_idle - enter idle state 919797e8345STejun Heo * @worker: worker which is entering idle state 920797e8345STejun Heo * 921797e8345STejun Heo * @worker is entering idle state. Update stats and idle timer if 922797e8345STejun Heo * necessary. 923797e8345STejun Heo * 924797e8345STejun Heo * LOCKING: 925797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 926797e8345STejun Heo */ 927797e8345STejun Heo static void worker_enter_idle(struct worker *worker) 928797e8345STejun Heo { 929797e8345STejun Heo struct worker_pool *pool = worker->pool; 930797e8345STejun Heo 931797e8345STejun Heo if (WARN_ON_ONCE(worker->flags & WORKER_IDLE) || 932797e8345STejun Heo WARN_ON_ONCE(!list_empty(&worker->entry) && 933797e8345STejun Heo (worker->hentry.next || worker->hentry.pprev))) 934797e8345STejun Heo return; 935797e8345STejun Heo 936797e8345STejun Heo /* can't use worker_set_flags(), also called from create_worker() */ 937797e8345STejun Heo worker->flags |= WORKER_IDLE; 938797e8345STejun Heo pool->nr_idle++; 939797e8345STejun Heo worker->last_active = jiffies; 940797e8345STejun Heo 941797e8345STejun Heo /* idle_list is LIFO */ 942797e8345STejun Heo list_add(&worker->entry, &pool->idle_list); 943797e8345STejun Heo 944797e8345STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 945797e8345STejun Heo mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT); 946797e8345STejun Heo 947797e8345STejun Heo /* Sanity check nr_running. */ 948797e8345STejun Heo WARN_ON_ONCE(pool->nr_workers == pool->nr_idle && pool->nr_running); 949797e8345STejun Heo } 950797e8345STejun Heo 951797e8345STejun Heo /** 952797e8345STejun Heo * worker_leave_idle - leave idle state 953797e8345STejun Heo * @worker: worker which is leaving idle state 954797e8345STejun Heo * 955797e8345STejun Heo * @worker is leaving idle state. Update stats. 956797e8345STejun Heo * 957797e8345STejun Heo * LOCKING: 958797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 959797e8345STejun Heo */ 960797e8345STejun Heo static void worker_leave_idle(struct worker *worker) 961797e8345STejun Heo { 962797e8345STejun Heo struct worker_pool *pool = worker->pool; 963797e8345STejun Heo 964797e8345STejun Heo if (WARN_ON_ONCE(!(worker->flags & WORKER_IDLE))) 965797e8345STejun Heo return; 966797e8345STejun Heo worker_clr_flags(worker, WORKER_IDLE); 967797e8345STejun Heo pool->nr_idle--; 968797e8345STejun Heo list_del_init(&worker->entry); 969797e8345STejun Heo } 970797e8345STejun Heo 971797e8345STejun Heo /** 972797e8345STejun Heo * find_worker_executing_work - find worker which is executing a work 973797e8345STejun Heo * @pool: pool of interest 974797e8345STejun Heo * @work: work to find worker for 975797e8345STejun Heo * 976797e8345STejun Heo * Find a worker which is executing @work on @pool by searching 977797e8345STejun Heo * @pool->busy_hash which is keyed by the address of @work. For a worker 978797e8345STejun Heo * to match, its current execution should match the address of @work and 979797e8345STejun Heo * its work function. This is to avoid unwanted dependency between 980797e8345STejun Heo * unrelated work executions through a work item being recycled while still 981797e8345STejun Heo * being executed. 982797e8345STejun Heo * 983797e8345STejun Heo * This is a bit tricky. A work item may be freed once its execution 984797e8345STejun Heo * starts and nothing prevents the freed area from being recycled for 985797e8345STejun Heo * another work item. If the same work item address ends up being reused 986797e8345STejun Heo * before the original execution finishes, workqueue will identify the 987797e8345STejun Heo * recycled work item as currently executing and make it wait until the 988797e8345STejun Heo * current execution finishes, introducing an unwanted dependency. 989797e8345STejun Heo * 990797e8345STejun Heo * This function checks the work item address and work function to avoid 991797e8345STejun Heo * false positives. Note that this isn't complete as one may construct a 992797e8345STejun Heo * work function which can introduce dependency onto itself through a 993797e8345STejun Heo * recycled work item. Well, if somebody wants to shoot oneself in the 994797e8345STejun Heo * foot that badly, there's only so much we can do, and if such deadlock 995797e8345STejun Heo * actually occurs, it should be easy to locate the culprit work function. 996797e8345STejun Heo * 997797e8345STejun Heo * CONTEXT: 998797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 999797e8345STejun Heo * 1000797e8345STejun Heo * Return: 1001797e8345STejun Heo * Pointer to worker which is executing @work if found, %NULL 1002797e8345STejun Heo * otherwise. 1003797e8345STejun Heo */ 1004797e8345STejun Heo static struct worker *find_worker_executing_work(struct worker_pool *pool, 1005797e8345STejun Heo struct work_struct *work) 1006797e8345STejun Heo { 1007797e8345STejun Heo struct worker *worker; 1008797e8345STejun Heo 1009797e8345STejun Heo hash_for_each_possible(pool->busy_hash, worker, hentry, 1010797e8345STejun Heo (unsigned long)work) 1011797e8345STejun Heo if (worker->current_work == work && 1012797e8345STejun Heo worker->current_func == work->func) 1013797e8345STejun Heo return worker; 1014797e8345STejun Heo 1015797e8345STejun Heo return NULL; 1016797e8345STejun Heo } 1017797e8345STejun Heo 1018797e8345STejun Heo /** 1019797e8345STejun Heo * move_linked_works - move linked works to a list 1020797e8345STejun Heo * @work: start of series of works to be scheduled 1021797e8345STejun Heo * @head: target list to append @work to 1022797e8345STejun Heo * @nextp: out parameter for nested worklist walking 1023797e8345STejun Heo * 1024873eaca6STejun Heo * Schedule linked works starting from @work to @head. Work series to be 1025873eaca6STejun Heo * scheduled starts at @work and includes any consecutive work with 1026873eaca6STejun Heo * WORK_STRUCT_LINKED set in its predecessor. See assign_work() for details on 1027873eaca6STejun Heo * @nextp. 1028797e8345STejun Heo * 1029797e8345STejun Heo * CONTEXT: 1030797e8345STejun Heo * raw_spin_lock_irq(pool->lock). 1031797e8345STejun Heo */ 1032797e8345STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1033797e8345STejun Heo struct work_struct **nextp) 1034797e8345STejun Heo { 1035797e8345STejun Heo struct work_struct *n; 1036797e8345STejun Heo 1037797e8345STejun Heo /* 1038797e8345STejun Heo * Linked worklist will always end before the end of the list, 1039797e8345STejun Heo * use NULL for list head. 1040797e8345STejun Heo */ 1041797e8345STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1042797e8345STejun Heo list_move_tail(&work->entry, head); 1043797e8345STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1044797e8345STejun Heo break; 1045797e8345STejun Heo } 1046797e8345STejun Heo 1047797e8345STejun Heo /* 1048797e8345STejun Heo * If we're already inside safe list traversal and have moved 1049797e8345STejun Heo * multiple works to the scheduled queue, the next position 1050797e8345STejun Heo * needs to be updated. 1051797e8345STejun Heo */ 1052797e8345STejun Heo if (nextp) 1053797e8345STejun Heo *nextp = n; 1054797e8345STejun Heo } 1055797e8345STejun Heo 1056797e8345STejun Heo /** 1057873eaca6STejun Heo * assign_work - assign a work item and its linked work items to a worker 1058873eaca6STejun Heo * @work: work to assign 1059873eaca6STejun Heo * @worker: worker to assign to 1060873eaca6STejun Heo * @nextp: out parameter for nested worklist walking 1061873eaca6STejun Heo * 1062873eaca6STejun Heo * Assign @work and its linked work items to @worker. If @work is already being 1063873eaca6STejun Heo * executed by another worker in the same pool, it'll be punted there. 1064873eaca6STejun Heo * 1065873eaca6STejun Heo * If @nextp is not NULL, it's updated to point to the next work of the last 1066873eaca6STejun Heo * scheduled work. This allows assign_work() to be nested inside 1067873eaca6STejun Heo * list_for_each_entry_safe(). 1068873eaca6STejun Heo * 1069873eaca6STejun Heo * Returns %true if @work was successfully assigned to @worker. %false if @work 1070873eaca6STejun Heo * was punted to another worker already executing it. 1071873eaca6STejun Heo */ 1072873eaca6STejun Heo static bool assign_work(struct work_struct *work, struct worker *worker, 1073873eaca6STejun Heo struct work_struct **nextp) 1074873eaca6STejun Heo { 1075873eaca6STejun Heo struct worker_pool *pool = worker->pool; 1076873eaca6STejun Heo struct worker *collision; 1077873eaca6STejun Heo 1078873eaca6STejun Heo lockdep_assert_held(&pool->lock); 1079873eaca6STejun Heo 1080873eaca6STejun Heo /* 1081873eaca6STejun Heo * A single work shouldn't be executed concurrently by multiple workers. 1082873eaca6STejun Heo * __queue_work() ensures that @work doesn't jump to a different pool 1083873eaca6STejun Heo * while still running in the previous pool. Here, we should ensure that 1084873eaca6STejun Heo * @work is not executed concurrently by multiple workers from the same 1085873eaca6STejun Heo * pool. Check whether anyone is already processing the work. If so, 1086873eaca6STejun Heo * defer the work to the currently executing one. 1087873eaca6STejun Heo */ 1088873eaca6STejun Heo collision = find_worker_executing_work(pool, work); 1089873eaca6STejun Heo if (unlikely(collision)) { 1090873eaca6STejun Heo move_linked_works(work, &collision->scheduled, nextp); 1091873eaca6STejun Heo return false; 1092873eaca6STejun Heo } 1093873eaca6STejun Heo 1094873eaca6STejun Heo move_linked_works(work, &worker->scheduled, nextp); 1095873eaca6STejun Heo return true; 1096873eaca6STejun Heo } 1097873eaca6STejun Heo 1098873eaca6STejun Heo /** 10990219a352STejun Heo * kick_pool - wake up an idle worker if necessary 11000219a352STejun Heo * @pool: pool to kick 1101797e8345STejun Heo * 11020219a352STejun Heo * @pool may have pending work items. Wake up worker if necessary. Returns 11030219a352STejun Heo * whether a worker was woken up. 1104797e8345STejun Heo */ 11050219a352STejun Heo static bool kick_pool(struct worker_pool *pool) 1106797e8345STejun Heo { 1107797e8345STejun Heo struct worker *worker = first_idle_worker(pool); 11088639ecebSTejun Heo struct task_struct *p; 1109797e8345STejun Heo 11100219a352STejun Heo lockdep_assert_held(&pool->lock); 11110219a352STejun Heo 11120219a352STejun Heo if (!need_more_worker(pool) || !worker) 11130219a352STejun Heo return false; 11140219a352STejun Heo 11158639ecebSTejun Heo p = worker->task; 11168639ecebSTejun Heo 11178639ecebSTejun Heo #ifdef CONFIG_SMP 11188639ecebSTejun Heo /* 11198639ecebSTejun Heo * Idle @worker is about to execute @work and waking up provides an 11208639ecebSTejun Heo * opportunity to migrate @worker at a lower cost by setting the task's 11218639ecebSTejun Heo * wake_cpu field. Let's see if we want to move @worker to improve 11228639ecebSTejun Heo * execution locality. 11238639ecebSTejun Heo * 11248639ecebSTejun Heo * We're waking the worker that went idle the latest and there's some 11258639ecebSTejun Heo * chance that @worker is marked idle but hasn't gone off CPU yet. If 11268639ecebSTejun Heo * so, setting the wake_cpu won't do anything. As this is a best-effort 11278639ecebSTejun Heo * optimization and the race window is narrow, let's leave as-is for 11288639ecebSTejun Heo * now. If this becomes pronounced, we can skip over workers which are 11298639ecebSTejun Heo * still on cpu when picking an idle worker. 11308639ecebSTejun Heo * 11318639ecebSTejun Heo * If @pool has non-strict affinity, @worker might have ended up outside 11328639ecebSTejun Heo * its affinity scope. Repatriate. 11338639ecebSTejun Heo */ 11348639ecebSTejun Heo if (!pool->attrs->affn_strict && 11358639ecebSTejun Heo !cpumask_test_cpu(p->wake_cpu, pool->attrs->__pod_cpumask)) { 11368639ecebSTejun Heo struct work_struct *work = list_first_entry(&pool->worklist, 11378639ecebSTejun Heo struct work_struct, entry); 11388639ecebSTejun Heo p->wake_cpu = cpumask_any_distribute(pool->attrs->__pod_cpumask); 11398639ecebSTejun Heo get_work_pwq(work)->stats[PWQ_STAT_REPATRIATED]++; 11408639ecebSTejun Heo } 11418639ecebSTejun Heo #endif 11428639ecebSTejun Heo wake_up_process(p); 11430219a352STejun Heo return true; 1144797e8345STejun Heo } 1145797e8345STejun Heo 114663638450STejun Heo #ifdef CONFIG_WQ_CPU_INTENSIVE_REPORT 114763638450STejun Heo 114863638450STejun Heo /* 114963638450STejun Heo * Concurrency-managed per-cpu work items that hog CPU for longer than 115063638450STejun Heo * wq_cpu_intensive_thresh_us trigger the automatic CPU_INTENSIVE mechanism, 115163638450STejun Heo * which prevents them from stalling other concurrency-managed work items. If a 115263638450STejun Heo * work function keeps triggering this mechanism, it's likely that the work item 115363638450STejun Heo * should be using an unbound workqueue instead. 115463638450STejun Heo * 115563638450STejun Heo * wq_cpu_intensive_report() tracks work functions which trigger such conditions 115663638450STejun Heo * and report them so that they can be examined and converted to use unbound 115763638450STejun Heo * workqueues as appropriate. To avoid flooding the console, each violating work 115863638450STejun Heo * function is tracked and reported with exponential backoff. 115963638450STejun Heo */ 116063638450STejun Heo #define WCI_MAX_ENTS 128 116163638450STejun Heo 116263638450STejun Heo struct wci_ent { 116363638450STejun Heo work_func_t func; 116463638450STejun Heo atomic64_t cnt; 116563638450STejun Heo struct hlist_node hash_node; 116663638450STejun Heo }; 116763638450STejun Heo 116863638450STejun Heo static struct wci_ent wci_ents[WCI_MAX_ENTS]; 116963638450STejun Heo static int wci_nr_ents; 117063638450STejun Heo static DEFINE_RAW_SPINLOCK(wci_lock); 117163638450STejun Heo static DEFINE_HASHTABLE(wci_hash, ilog2(WCI_MAX_ENTS)); 117263638450STejun Heo 117363638450STejun Heo static struct wci_ent *wci_find_ent(work_func_t func) 117463638450STejun Heo { 117563638450STejun Heo struct wci_ent *ent; 117663638450STejun Heo 117763638450STejun Heo hash_for_each_possible_rcu(wci_hash, ent, hash_node, 117863638450STejun Heo (unsigned long)func) { 117963638450STejun Heo if (ent->func == func) 118063638450STejun Heo return ent; 118163638450STejun Heo } 118263638450STejun Heo return NULL; 118363638450STejun Heo } 118463638450STejun Heo 118563638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) 118663638450STejun Heo { 118763638450STejun Heo struct wci_ent *ent; 118863638450STejun Heo 118963638450STejun Heo restart: 119063638450STejun Heo ent = wci_find_ent(func); 119163638450STejun Heo if (ent) { 119263638450STejun Heo u64 cnt; 119363638450STejun Heo 119463638450STejun Heo /* 119563638450STejun Heo * Start reporting from the fourth time and back off 119663638450STejun Heo * exponentially. 119763638450STejun Heo */ 119863638450STejun Heo cnt = atomic64_inc_return_relaxed(&ent->cnt); 119963638450STejun Heo if (cnt >= 4 && is_power_of_2(cnt)) 120063638450STejun Heo printk_deferred(KERN_WARNING "workqueue: %ps hogged CPU for >%luus %llu times, consider switching to WQ_UNBOUND\n", 120163638450STejun Heo ent->func, wq_cpu_intensive_thresh_us, 120263638450STejun Heo atomic64_read(&ent->cnt)); 120363638450STejun Heo return; 120463638450STejun Heo } 120563638450STejun Heo 120663638450STejun Heo /* 120763638450STejun Heo * @func is a new violation. Allocate a new entry for it. If wcn_ents[] 120863638450STejun Heo * is exhausted, something went really wrong and we probably made enough 120963638450STejun Heo * noise already. 121063638450STejun Heo */ 121163638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) 121263638450STejun Heo return; 121363638450STejun Heo 121463638450STejun Heo raw_spin_lock(&wci_lock); 121563638450STejun Heo 121663638450STejun Heo if (wci_nr_ents >= WCI_MAX_ENTS) { 121763638450STejun Heo raw_spin_unlock(&wci_lock); 121863638450STejun Heo return; 121963638450STejun Heo } 122063638450STejun Heo 122163638450STejun Heo if (wci_find_ent(func)) { 122263638450STejun Heo raw_spin_unlock(&wci_lock); 122363638450STejun Heo goto restart; 122463638450STejun Heo } 122563638450STejun Heo 122663638450STejun Heo ent = &wci_ents[wci_nr_ents++]; 122763638450STejun Heo ent->func = func; 122863638450STejun Heo atomic64_set(&ent->cnt, 1); 122963638450STejun Heo hash_add_rcu(wci_hash, &ent->hash_node, (unsigned long)func); 123063638450STejun Heo 123163638450STejun Heo raw_spin_unlock(&wci_lock); 123263638450STejun Heo } 123363638450STejun Heo 123463638450STejun Heo #else /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 123563638450STejun Heo static void wq_cpu_intensive_report(work_func_t func) {} 123663638450STejun Heo #endif /* CONFIG_WQ_CPU_INTENSIVE_REPORT */ 123763638450STejun Heo 1238c54d5046STejun Heo /** 12391da177e4SLinus Torvalds * wq_worker_running - a worker is running again 12401da177e4SLinus Torvalds * @task: task waking up 12411da177e4SLinus Torvalds * 12421da177e4SLinus Torvalds * This function is called when a worker returns from schedule() 12431da177e4SLinus Torvalds */ 12441da177e4SLinus Torvalds void wq_worker_running(struct task_struct *task) 12451da177e4SLinus Torvalds { 12461da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 12471da177e4SLinus Torvalds 1248c8f6219bSZqiang if (!READ_ONCE(worker->sleeping)) 12491da177e4SLinus Torvalds return; 12501da177e4SLinus Torvalds 12511da177e4SLinus Torvalds /* 12521da177e4SLinus Torvalds * If preempted by unbind_workers() between the WORKER_NOT_RUNNING check 12531da177e4SLinus Torvalds * and the nr_running increment below, we may ruin the nr_running reset 12541da177e4SLinus Torvalds * and leave with an unexpected pool->nr_running == 1 on the newly unbound 12551da177e4SLinus Torvalds * pool. Protect against such race. 12561da177e4SLinus Torvalds */ 12571da177e4SLinus Torvalds preempt_disable(); 12581da177e4SLinus Torvalds if (!(worker->flags & WORKER_NOT_RUNNING)) 12591da177e4SLinus Torvalds worker->pool->nr_running++; 12601da177e4SLinus Torvalds preempt_enable(); 1261616db877STejun Heo 1262616db877STejun Heo /* 1263616db877STejun Heo * CPU intensive auto-detection cares about how long a work item hogged 1264616db877STejun Heo * CPU without sleeping. Reset the starting timestamp on wakeup. 1265616db877STejun Heo */ 1266616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 1267616db877STejun Heo 1268c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 0); 12691da177e4SLinus Torvalds } 12701da177e4SLinus Torvalds 12711da177e4SLinus Torvalds /** 12721da177e4SLinus Torvalds * wq_worker_sleeping - a worker is going to sleep 12731da177e4SLinus Torvalds * @task: task going to sleep 12741da177e4SLinus Torvalds * 12751da177e4SLinus Torvalds * This function is called from schedule() when a busy worker is 12761da177e4SLinus Torvalds * going to sleep. 12771da177e4SLinus Torvalds */ 12781da177e4SLinus Torvalds void wq_worker_sleeping(struct task_struct *task) 12791da177e4SLinus Torvalds { 12801da177e4SLinus Torvalds struct worker *worker = kthread_data(task); 12811da177e4SLinus Torvalds struct worker_pool *pool; 12821da177e4SLinus Torvalds 12831da177e4SLinus Torvalds /* 12841da177e4SLinus Torvalds * Rescuers, which may not have all the fields set up like normal 12851da177e4SLinus Torvalds * workers, also reach here, let's not access anything before 12861da177e4SLinus Torvalds * checking NOT_RUNNING. 12871da177e4SLinus Torvalds */ 12881da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) 12891da177e4SLinus Torvalds return; 12901da177e4SLinus Torvalds 12911da177e4SLinus Torvalds pool = worker->pool; 12921da177e4SLinus Torvalds 12931da177e4SLinus Torvalds /* Return if preempted before wq_worker_running() was reached */ 1294c8f6219bSZqiang if (READ_ONCE(worker->sleeping)) 12951da177e4SLinus Torvalds return; 12961da177e4SLinus Torvalds 1297c8f6219bSZqiang WRITE_ONCE(worker->sleeping, 1); 12981da177e4SLinus Torvalds raw_spin_lock_irq(&pool->lock); 12991da177e4SLinus Torvalds 13001da177e4SLinus Torvalds /* 13011da177e4SLinus Torvalds * Recheck in case unbind_workers() preempted us. We don't 13021da177e4SLinus Torvalds * want to decrement nr_running after the worker is unbound 13031da177e4SLinus Torvalds * and nr_running has been reset. 13041da177e4SLinus Torvalds */ 13051da177e4SLinus Torvalds if (worker->flags & WORKER_NOT_RUNNING) { 13061da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 13071da177e4SLinus Torvalds return; 13081da177e4SLinus Torvalds } 13091da177e4SLinus Torvalds 13101da177e4SLinus Torvalds pool->nr_running--; 13110219a352STejun Heo if (kick_pool(pool)) 1312725e8ec5STejun Heo worker->current_pwq->stats[PWQ_STAT_CM_WAKEUP]++; 13130219a352STejun Heo 13141da177e4SLinus Torvalds raw_spin_unlock_irq(&pool->lock); 13151da177e4SLinus Torvalds } 13161da177e4SLinus Torvalds 13171da177e4SLinus Torvalds /** 1318616db877STejun Heo * wq_worker_tick - a scheduler tick occurred while a kworker is running 1319616db877STejun Heo * @task: task currently running 1320616db877STejun Heo * 1321616db877STejun Heo * Called from scheduler_tick(). We're in the IRQ context and the current 1322616db877STejun Heo * worker's fields which follow the 'K' locking rule can be accessed safely. 1323616db877STejun Heo */ 1324616db877STejun Heo void wq_worker_tick(struct task_struct *task) 1325616db877STejun Heo { 1326616db877STejun Heo struct worker *worker = kthread_data(task); 1327616db877STejun Heo struct pool_workqueue *pwq = worker->current_pwq; 1328616db877STejun Heo struct worker_pool *pool = worker->pool; 1329616db877STejun Heo 1330616db877STejun Heo if (!pwq) 1331616db877STejun Heo return; 1332616db877STejun Heo 13338a1dd1e5STejun Heo pwq->stats[PWQ_STAT_CPU_TIME] += TICK_USEC; 13348a1dd1e5STejun Heo 133518c8ae81SZqiang if (!wq_cpu_intensive_thresh_us) 133618c8ae81SZqiang return; 133718c8ae81SZqiang 1338616db877STejun Heo /* 1339616db877STejun Heo * If the current worker is concurrency managed and hogged the CPU for 1340616db877STejun Heo * longer than wq_cpu_intensive_thresh_us, it's automatically marked 1341616db877STejun Heo * CPU_INTENSIVE to avoid stalling other concurrency-managed work items. 1342c8f6219bSZqiang * 1343c8f6219bSZqiang * Set @worker->sleeping means that @worker is in the process of 1344c8f6219bSZqiang * switching out voluntarily and won't be contributing to 1345c8f6219bSZqiang * @pool->nr_running until it wakes up. As wq_worker_sleeping() also 1346c8f6219bSZqiang * decrements ->nr_running, setting CPU_INTENSIVE here can lead to 1347c8f6219bSZqiang * double decrements. The task is releasing the CPU anyway. Let's skip. 1348c8f6219bSZqiang * We probably want to make this prettier in the future. 1349616db877STejun Heo */ 1350c8f6219bSZqiang if ((worker->flags & WORKER_NOT_RUNNING) || READ_ONCE(worker->sleeping) || 1351616db877STejun Heo worker->task->se.sum_exec_runtime - worker->current_at < 1352616db877STejun Heo wq_cpu_intensive_thresh_us * NSEC_PER_USEC) 1353616db877STejun Heo return; 1354616db877STejun Heo 1355616db877STejun Heo raw_spin_lock(&pool->lock); 1356616db877STejun Heo 1357616db877STejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE); 135863638450STejun Heo wq_cpu_intensive_report(worker->current_func); 1359616db877STejun Heo pwq->stats[PWQ_STAT_CPU_INTENSIVE]++; 1360616db877STejun Heo 13610219a352STejun Heo if (kick_pool(pool)) 1362616db877STejun Heo pwq->stats[PWQ_STAT_CM_WAKEUP]++; 1363616db877STejun Heo 1364616db877STejun Heo raw_spin_unlock(&pool->lock); 1365616db877STejun Heo } 1366616db877STejun Heo 1367616db877STejun Heo /** 13681da177e4SLinus Torvalds * wq_worker_last_func - retrieve worker's last work function 13691da177e4SLinus Torvalds * @task: Task to retrieve last work function of. 13701da177e4SLinus Torvalds * 13711da177e4SLinus Torvalds * Determine the last function a worker executed. This is called from 13721da177e4SLinus Torvalds * the scheduler to get a worker's last known identity. 13731da177e4SLinus Torvalds * 13741da177e4SLinus Torvalds * CONTEXT: 13759b41ea72SAndrew Morton * raw_spin_lock_irq(rq->lock) 13761da177e4SLinus Torvalds * 13771da177e4SLinus Torvalds * This function is called during schedule() when a kworker is going 1378f756d5e2SNathan Lynch * to sleep. It's used by psi to identify aggregation workers during 1379f756d5e2SNathan Lynch * dequeuing, to allow periodic aggregation to shut-off when that 13801da177e4SLinus Torvalds * worker is the last task in the system or cgroup to go to sleep. 13811da177e4SLinus Torvalds * 13821da177e4SLinus Torvalds * As this function doesn't involve any workqueue-related locking, it 13831da177e4SLinus Torvalds * only returns stable values when called from inside the scheduler's 13841da177e4SLinus Torvalds * queuing and dequeuing paths, when @task, which must be a kworker, 13851da177e4SLinus Torvalds * is guaranteed to not be processing any works. 1386365970a1SDavid Howells * 1387365970a1SDavid Howells * Return: 1388365970a1SDavid Howells * The last work function %current executed as a worker, NULL if it 1389365970a1SDavid Howells * hasn't executed any work yet. 1390365970a1SDavid Howells */ 1391365970a1SDavid Howells work_func_t wq_worker_last_func(struct task_struct *task) 1392365970a1SDavid Howells { 1393365970a1SDavid Howells struct worker *worker = kthread_data(task); 1394365970a1SDavid Howells 1395365970a1SDavid Howells return worker->last_func; 1396365970a1SDavid Howells } 1397365970a1SDavid Howells 1398d302f017STejun Heo /** 13998864b4e5STejun Heo * get_pwq - get an extra reference on the specified pool_workqueue 14008864b4e5STejun Heo * @pwq: pool_workqueue to get 14018864b4e5STejun Heo * 14028864b4e5STejun Heo * Obtain an extra reference on @pwq. The caller should guarantee that 14038864b4e5STejun Heo * @pwq has positive refcnt and be holding the matching pool->lock. 14048864b4e5STejun Heo */ 14058864b4e5STejun Heo static void get_pwq(struct pool_workqueue *pwq) 14068864b4e5STejun Heo { 14078864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 14088864b4e5STejun Heo WARN_ON_ONCE(pwq->refcnt <= 0); 14098864b4e5STejun Heo pwq->refcnt++; 14108864b4e5STejun Heo } 14118864b4e5STejun Heo 14128864b4e5STejun Heo /** 14138864b4e5STejun Heo * put_pwq - put a pool_workqueue reference 14148864b4e5STejun Heo * @pwq: pool_workqueue to put 14158864b4e5STejun Heo * 14168864b4e5STejun Heo * Drop a reference of @pwq. If its refcnt reaches zero, schedule its 14178864b4e5STejun Heo * destruction. The caller should be holding the matching pool->lock. 14188864b4e5STejun Heo */ 14198864b4e5STejun Heo static void put_pwq(struct pool_workqueue *pwq) 14208864b4e5STejun Heo { 14218864b4e5STejun Heo lockdep_assert_held(&pwq->pool->lock); 14228864b4e5STejun Heo if (likely(--pwq->refcnt)) 14238864b4e5STejun Heo return; 14248864b4e5STejun Heo /* 1425967b494eSTejun Heo * @pwq can't be released under pool->lock, bounce to a dedicated 1426967b494eSTejun Heo * kthread_worker to avoid A-A deadlocks. 14278864b4e5STejun Heo */ 1428687a9aa5STejun Heo kthread_queue_work(pwq_release_worker, &pwq->release_work); 14298864b4e5STejun Heo } 14308864b4e5STejun Heo 1431dce90d47STejun Heo /** 1432dce90d47STejun Heo * put_pwq_unlocked - put_pwq() with surrounding pool lock/unlock 1433dce90d47STejun Heo * @pwq: pool_workqueue to put (can be %NULL) 1434dce90d47STejun Heo * 1435dce90d47STejun Heo * put_pwq() with locking. This function also allows %NULL @pwq. 1436dce90d47STejun Heo */ 1437dce90d47STejun Heo static void put_pwq_unlocked(struct pool_workqueue *pwq) 1438dce90d47STejun Heo { 1439dce90d47STejun Heo if (pwq) { 1440dce90d47STejun Heo /* 144124acfb71SThomas Gleixner * As both pwqs and pools are RCU protected, the 1442dce90d47STejun Heo * following lock operations are safe. 1443dce90d47STejun Heo */ 1444a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 1445dce90d47STejun Heo put_pwq(pwq); 1446a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 1447dce90d47STejun Heo } 1448dce90d47STejun Heo } 1449dce90d47STejun Heo 1450f97a4a1aSLai Jiangshan static void pwq_activate_inactive_work(struct work_struct *work) 1451bf4ede01STejun Heo { 1452112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 1453bf4ede01STejun Heo 1454bf4ede01STejun Heo trace_workqueue_activate_work(work); 145582607adcSTejun Heo if (list_empty(&pwq->pool->worklist)) 145682607adcSTejun Heo pwq->pool->watchdog_ts = jiffies; 1457112202d9STejun Heo move_linked_works(work, &pwq->pool->worklist, NULL); 1458f97a4a1aSLai Jiangshan __clear_bit(WORK_STRUCT_INACTIVE_BIT, work_data_bits(work)); 1459112202d9STejun Heo pwq->nr_active++; 1460bf4ede01STejun Heo } 1461bf4ede01STejun Heo 1462f97a4a1aSLai Jiangshan static void pwq_activate_first_inactive(struct pool_workqueue *pwq) 14633aa62497SLai Jiangshan { 1464f97a4a1aSLai Jiangshan struct work_struct *work = list_first_entry(&pwq->inactive_works, 14653aa62497SLai Jiangshan struct work_struct, entry); 14663aa62497SLai Jiangshan 1467f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 14683aa62497SLai Jiangshan } 14693aa62497SLai Jiangshan 1470bf4ede01STejun Heo /** 1471112202d9STejun Heo * pwq_dec_nr_in_flight - decrement pwq's nr_in_flight 1472112202d9STejun Heo * @pwq: pwq of interest 1473c4560c2cSLai Jiangshan * @work_data: work_data of work which left the queue 1474bf4ede01STejun Heo * 1475bf4ede01STejun Heo * A work either has completed or is removed from pending queue, 1476112202d9STejun Heo * decrement nr_in_flight of its pwq and handle workqueue flushing. 1477bf4ede01STejun Heo * 1478bf4ede01STejun Heo * CONTEXT: 1479a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1480bf4ede01STejun Heo */ 1481c4560c2cSLai Jiangshan static void pwq_dec_nr_in_flight(struct pool_workqueue *pwq, unsigned long work_data) 1482bf4ede01STejun Heo { 1483c4560c2cSLai Jiangshan int color = get_work_color(work_data); 1484c4560c2cSLai Jiangshan 1485018f3a13SLai Jiangshan if (!(work_data & WORK_STRUCT_INACTIVE)) { 1486112202d9STejun Heo pwq->nr_active--; 1487f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 1488f97a4a1aSLai Jiangshan /* one down, submit an inactive one */ 1489112202d9STejun Heo if (pwq->nr_active < pwq->max_active) 1490f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 1491bf4ede01STejun Heo } 1492018f3a13SLai Jiangshan } 1493018f3a13SLai Jiangshan 1494018f3a13SLai Jiangshan pwq->nr_in_flight[color]--; 1495bf4ede01STejun Heo 1496bf4ede01STejun Heo /* is flush in progress and are we at the flushing tip? */ 1497112202d9STejun Heo if (likely(pwq->flush_color != color)) 14988864b4e5STejun Heo goto out_put; 1499bf4ede01STejun Heo 1500bf4ede01STejun Heo /* are there still in-flight works? */ 1501112202d9STejun Heo if (pwq->nr_in_flight[color]) 15028864b4e5STejun Heo goto out_put; 1503bf4ede01STejun Heo 1504112202d9STejun Heo /* this pwq is done, clear flush_color */ 1505112202d9STejun Heo pwq->flush_color = -1; 1506bf4ede01STejun Heo 1507bf4ede01STejun Heo /* 1508112202d9STejun Heo * If this was the last pwq, wake up the first flusher. It 1509bf4ede01STejun Heo * will handle the rest. 1510bf4ede01STejun Heo */ 1511112202d9STejun Heo if (atomic_dec_and_test(&pwq->wq->nr_pwqs_to_flush)) 1512112202d9STejun Heo complete(&pwq->wq->first_flusher->done); 15138864b4e5STejun Heo out_put: 15148864b4e5STejun Heo put_pwq(pwq); 1515bf4ede01STejun Heo } 1516bf4ede01STejun Heo 151736e227d2STejun Heo /** 1518bbb68dfaSTejun Heo * try_to_grab_pending - steal work item from worklist and disable irq 151936e227d2STejun Heo * @work: work item to steal 152036e227d2STejun Heo * @is_dwork: @work is a delayed_work 1521bbb68dfaSTejun Heo * @flags: place to store irq state 152236e227d2STejun Heo * 152336e227d2STejun Heo * Try to grab PENDING bit of @work. This function can handle @work in any 1524d185af30SYacine Belkadi * stable state - idle, on timer or on worklist. 152536e227d2STejun Heo * 1526d185af30SYacine Belkadi * Return: 15273eb6b31bSMauro Carvalho Chehab * 15283eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 152936e227d2STejun Heo * 1 if @work was pending and we successfully stole PENDING 153036e227d2STejun Heo * 0 if @work was idle and we claimed PENDING 153136e227d2STejun Heo * -EAGAIN if PENDING couldn't be grabbed at the moment, safe to busy-retry 1532bbb68dfaSTejun Heo * -ENOENT if someone else is canceling @work, this state may persist 1533bbb68dfaSTejun Heo * for arbitrarily long 15343eb6b31bSMauro Carvalho Chehab * ======== ================================================================ 153536e227d2STejun Heo * 1536d185af30SYacine Belkadi * Note: 1537bbb68dfaSTejun Heo * On >= 0 return, the caller owns @work's PENDING bit. To avoid getting 1538e0aecdd8STejun Heo * interrupted while holding PENDING and @work off queue, irq must be 1539e0aecdd8STejun Heo * disabled on entry. This, combined with delayed_work->timer being 1540e0aecdd8STejun Heo * irqsafe, ensures that we return -EAGAIN for finite short period of time. 1541bbb68dfaSTejun Heo * 1542bbb68dfaSTejun Heo * On successful return, >= 0, irq is disabled and the caller is 1543bbb68dfaSTejun Heo * responsible for releasing it using local_irq_restore(*@flags). 1544bbb68dfaSTejun Heo * 1545e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 1546bf4ede01STejun Heo */ 1547bbb68dfaSTejun Heo static int try_to_grab_pending(struct work_struct *work, bool is_dwork, 1548bbb68dfaSTejun Heo unsigned long *flags) 1549bf4ede01STejun Heo { 1550d565ed63STejun Heo struct worker_pool *pool; 1551112202d9STejun Heo struct pool_workqueue *pwq; 1552bf4ede01STejun Heo 1553bbb68dfaSTejun Heo local_irq_save(*flags); 1554bbb68dfaSTejun Heo 155536e227d2STejun Heo /* try to steal the timer if it exists */ 155636e227d2STejun Heo if (is_dwork) { 155736e227d2STejun Heo struct delayed_work *dwork = to_delayed_work(work); 155836e227d2STejun Heo 1559e0aecdd8STejun Heo /* 1560e0aecdd8STejun Heo * dwork->timer is irqsafe. If del_timer() fails, it's 1561e0aecdd8STejun Heo * guaranteed that the timer is not queued anywhere and not 1562e0aecdd8STejun Heo * running on the local CPU. 1563e0aecdd8STejun Heo */ 156436e227d2STejun Heo if (likely(del_timer(&dwork->timer))) 156536e227d2STejun Heo return 1; 156636e227d2STejun Heo } 156736e227d2STejun Heo 156836e227d2STejun Heo /* try to claim PENDING the normal way */ 1569bf4ede01STejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 1570bf4ede01STejun Heo return 0; 1571bf4ede01STejun Heo 157224acfb71SThomas Gleixner rcu_read_lock(); 1573bf4ede01STejun Heo /* 1574bf4ede01STejun Heo * The queueing is in progress, or it is already queued. Try to 1575bf4ede01STejun Heo * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 1576bf4ede01STejun Heo */ 1577d565ed63STejun Heo pool = get_work_pool(work); 1578d565ed63STejun Heo if (!pool) 1579bbb68dfaSTejun Heo goto fail; 1580bf4ede01STejun Heo 1581a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&pool->lock); 1582bf4ede01STejun Heo /* 1583112202d9STejun Heo * work->data is guaranteed to point to pwq only while the work 1584112202d9STejun Heo * item is queued on pwq->wq, and both updating work->data to point 1585112202d9STejun Heo * to pwq on queueing and to pool on dequeueing are done under 1586112202d9STejun Heo * pwq->pool->lock. This in turn guarantees that, if work->data 1587112202d9STejun Heo * points to pwq which is associated with a locked pool, the work 15880b3dae68SLai Jiangshan * item is currently queued on that pool. 1589bf4ede01STejun Heo */ 1590112202d9STejun Heo pwq = get_work_pwq(work); 1591112202d9STejun Heo if (pwq && pwq->pool == pool) { 1592bf4ede01STejun Heo debug_work_deactivate(work); 15933aa62497SLai Jiangshan 15943aa62497SLai Jiangshan /* 1595018f3a13SLai Jiangshan * A cancelable inactive work item must be in the 1596018f3a13SLai Jiangshan * pwq->inactive_works since a queued barrier can't be 1597018f3a13SLai Jiangshan * canceled (see the comments in insert_wq_barrier()). 1598018f3a13SLai Jiangshan * 1599f97a4a1aSLai Jiangshan * An inactive work item cannot be grabbed directly because 1600d812796eSLai Jiangshan * it might have linked barrier work items which, if left 1601f97a4a1aSLai Jiangshan * on the inactive_works list, will confuse pwq->nr_active 160216062836STejun Heo * management later on and cause stall. Make sure the work 160316062836STejun Heo * item is activated before grabbing. 16043aa62497SLai Jiangshan */ 1605f97a4a1aSLai Jiangshan if (*work_data_bits(work) & WORK_STRUCT_INACTIVE) 1606f97a4a1aSLai Jiangshan pwq_activate_inactive_work(work); 16073aa62497SLai Jiangshan 1608bf4ede01STejun Heo list_del_init(&work->entry); 1609c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, *work_data_bits(work)); 161036e227d2STejun Heo 1611112202d9STejun Heo /* work->data points to pwq iff queued, point to pool */ 16124468a00fSLai Jiangshan set_work_pool_and_keep_pending(work, pool->id); 16134468a00fSLai Jiangshan 1614a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 161524acfb71SThomas Gleixner rcu_read_unlock(); 161636e227d2STejun Heo return 1; 1617bf4ede01STejun Heo } 1618a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&pool->lock); 1619bbb68dfaSTejun Heo fail: 162024acfb71SThomas Gleixner rcu_read_unlock(); 1621bbb68dfaSTejun Heo local_irq_restore(*flags); 1622bbb68dfaSTejun Heo if (work_is_canceling(work)) 1623bbb68dfaSTejun Heo return -ENOENT; 1624bbb68dfaSTejun Heo cpu_relax(); 162536e227d2STejun Heo return -EAGAIN; 1626bf4ede01STejun Heo } 1627bf4ede01STejun Heo 1628bf4ede01STejun Heo /** 1629706026c2STejun Heo * insert_work - insert a work into a pool 1630112202d9STejun Heo * @pwq: pwq @work belongs to 16314690c4abSTejun Heo * @work: work to insert 16324690c4abSTejun Heo * @head: insertion point 16334690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 16344690c4abSTejun Heo * 1635112202d9STejun Heo * Insert @work which belongs to @pwq after @head. @extra_flags is or'd to 1636706026c2STejun Heo * work_struct flags. 16374690c4abSTejun Heo * 16384690c4abSTejun Heo * CONTEXT: 1639a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 1640365970a1SDavid Howells */ 1641112202d9STejun Heo static void insert_work(struct pool_workqueue *pwq, struct work_struct *work, 1642112202d9STejun Heo struct list_head *head, unsigned int extra_flags) 1643b89deed3SOleg Nesterov { 1644fe089f87STejun Heo debug_work_activate(work); 1645e1d8aa9fSFrederic Weisbecker 1646e89a85d6SWalter Wu /* record the work call stack in order to print it in KASAN reports */ 1647f70da745SMarco Elver kasan_record_aux_stack_noalloc(work); 1648e89a85d6SWalter Wu 16494690c4abSTejun Heo /* we own @work, set data and link */ 1650112202d9STejun Heo set_work_pwq(work, pwq, extra_flags); 16511a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 16528864b4e5STejun Heo get_pwq(pwq); 1653b89deed3SOleg Nesterov } 1654b89deed3SOleg Nesterov 1655c8efcc25STejun Heo /* 1656c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 16578d03ecfeSTejun Heo * same workqueue. 1658c8efcc25STejun Heo */ 1659c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 1660c8efcc25STejun Heo { 1661c8efcc25STejun Heo struct worker *worker; 1662c8efcc25STejun Heo 16638d03ecfeSTejun Heo worker = current_wq_worker(); 1664c8efcc25STejun Heo /* 1665bf393fd4SBart Van Assche * Return %true iff I'm a worker executing a work item on @wq. If 16668d03ecfeSTejun Heo * I'm @worker, it's safe to dereference it without locking. 1667c8efcc25STejun Heo */ 1668112202d9STejun Heo return worker && worker->current_pwq->wq == wq; 1669c8efcc25STejun Heo } 1670c8efcc25STejun Heo 1671ef557180SMike Galbraith /* 1672ef557180SMike Galbraith * When queueing an unbound work item to a wq, prefer local CPU if allowed 1673ef557180SMike Galbraith * by wq_unbound_cpumask. Otherwise, round robin among the allowed ones to 1674ef557180SMike Galbraith * avoid perturbing sensitive tasks. 1675ef557180SMike Galbraith */ 1676ef557180SMike Galbraith static int wq_select_unbound_cpu(int cpu) 1677ef557180SMike Galbraith { 1678ef557180SMike Galbraith int new_cpu; 1679ef557180SMike Galbraith 1680f303fccbSTejun Heo if (likely(!wq_debug_force_rr_cpu)) { 1681ef557180SMike Galbraith if (cpumask_test_cpu(cpu, wq_unbound_cpumask)) 1682ef557180SMike Galbraith return cpu; 1683a8ec5880SAmmar Faizi } else { 1684a8ec5880SAmmar Faizi pr_warn_once("workqueue: round-robin CPU selection forced, expect performance impact\n"); 1685f303fccbSTejun Heo } 1686f303fccbSTejun Heo 1687ef557180SMike Galbraith if (cpumask_empty(wq_unbound_cpumask)) 1688ef557180SMike Galbraith return cpu; 1689ef557180SMike Galbraith 1690ef557180SMike Galbraith new_cpu = __this_cpu_read(wq_rr_cpu_last); 1691ef557180SMike Galbraith new_cpu = cpumask_next_and(new_cpu, wq_unbound_cpumask, cpu_online_mask); 1692ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) { 1693ef557180SMike Galbraith new_cpu = cpumask_first_and(wq_unbound_cpumask, cpu_online_mask); 1694ef557180SMike Galbraith if (unlikely(new_cpu >= nr_cpu_ids)) 1695ef557180SMike Galbraith return cpu; 1696ef557180SMike Galbraith } 1697ef557180SMike Galbraith __this_cpu_write(wq_rr_cpu_last, new_cpu); 1698ef557180SMike Galbraith 1699ef557180SMike Galbraith return new_cpu; 1700ef557180SMike Galbraith } 1701ef557180SMike Galbraith 1702d84ff051STejun Heo static void __queue_work(int cpu, struct workqueue_struct *wq, 17031da177e4SLinus Torvalds struct work_struct *work) 17041da177e4SLinus Torvalds { 1705112202d9STejun Heo struct pool_workqueue *pwq; 1706fe089f87STejun Heo struct worker_pool *last_pool, *pool; 17078a2e8e5dSTejun Heo unsigned int work_flags; 1708b75cac93SJoonsoo Kim unsigned int req_cpu = cpu; 17098930cabaSTejun Heo 17108930cabaSTejun Heo /* 17118930cabaSTejun Heo * While a work item is PENDING && off queue, a task trying to 17128930cabaSTejun Heo * steal the PENDING will busy-loop waiting for it to either get 17138930cabaSTejun Heo * queued or lose PENDING. Grabbing PENDING and queueing should 17148930cabaSTejun Heo * happen with IRQ disabled. 17158930cabaSTejun Heo */ 17168e8eb730SFrederic Weisbecker lockdep_assert_irqs_disabled(); 17171da177e4SLinus Torvalds 17181e19ffc6STejun Heo 171933e3f0a3SRichard Clark /* 172033e3f0a3SRichard Clark * For a draining wq, only works from the same workqueue are 172133e3f0a3SRichard Clark * allowed. The __WQ_DESTROYING helps to spot the issue that 172233e3f0a3SRichard Clark * queues a new work item to a wq after destroy_workqueue(wq). 172333e3f0a3SRichard Clark */ 172433e3f0a3SRichard Clark if (unlikely(wq->flags & (__WQ_DESTROYING | __WQ_DRAINING) && 172533e3f0a3SRichard Clark WARN_ON_ONCE(!is_chained_work(wq)))) 1726e41e704bSTejun Heo return; 172724acfb71SThomas Gleixner rcu_read_lock(); 17289e8cd2f5STejun Heo retry: 1729aa202f1fSHillf Danton /* pwq which will be used unless @work is executing elsewhere */ 1730636b927eSTejun Heo if (req_cpu == WORK_CPU_UNBOUND) { 1731636b927eSTejun Heo if (wq->flags & WQ_UNBOUND) 1732ef557180SMike Galbraith cpu = wq_select_unbound_cpu(raw_smp_processor_id()); 1733636b927eSTejun Heo else 1734aa202f1fSHillf Danton cpu = raw_smp_processor_id(); 1735aa202f1fSHillf Danton } 1736f3421797STejun Heo 1737636b927eSTejun Heo pwq = rcu_dereference(*per_cpu_ptr(wq->cpu_pwq, cpu)); 1738fe089f87STejun Heo pool = pwq->pool; 1739fe089f87STejun Heo 174018aa9effSTejun Heo /* 1741c9178087STejun Heo * If @work was previously on a different pool, it might still be 1742c9178087STejun Heo * running there, in which case the work needs to be queued on that 1743c9178087STejun Heo * pool to guarantee non-reentrancy. 174418aa9effSTejun Heo */ 1745c9e7cf27STejun Heo last_pool = get_work_pool(work); 1746fe089f87STejun Heo if (last_pool && last_pool != pool) { 174718aa9effSTejun Heo struct worker *worker; 174818aa9effSTejun Heo 1749a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&last_pool->lock); 175018aa9effSTejun Heo 1751c9e7cf27STejun Heo worker = find_worker_executing_work(last_pool, work); 175218aa9effSTejun Heo 1753112202d9STejun Heo if (worker && worker->current_pwq->wq == wq) { 1754c9178087STejun Heo pwq = worker->current_pwq; 1755fe089f87STejun Heo pool = pwq->pool; 1756fe089f87STejun Heo WARN_ON_ONCE(pool != last_pool); 17578594fadeSLai Jiangshan } else { 175818aa9effSTejun Heo /* meh... not running there, queue here */ 1759a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&last_pool->lock); 1760fe089f87STejun Heo raw_spin_lock(&pool->lock); 176118aa9effSTejun Heo } 17628930cabaSTejun Heo } else { 1763fe089f87STejun Heo raw_spin_lock(&pool->lock); 17648930cabaSTejun Heo } 1765502ca9d8STejun Heo 17669e8cd2f5STejun Heo /* 1767636b927eSTejun Heo * pwq is determined and locked. For unbound pools, we could have raced 1768636b927eSTejun Heo * with pwq release and it could already be dead. If its refcnt is zero, 1769636b927eSTejun Heo * repeat pwq selection. Note that unbound pwqs never die without 1770636b927eSTejun Heo * another pwq replacing it in cpu_pwq or while work items are executing 1771636b927eSTejun Heo * on it, so the retrying is guaranteed to make forward-progress. 17729e8cd2f5STejun Heo */ 17739e8cd2f5STejun Heo if (unlikely(!pwq->refcnt)) { 17749e8cd2f5STejun Heo if (wq->flags & WQ_UNBOUND) { 1775fe089f87STejun Heo raw_spin_unlock(&pool->lock); 17769e8cd2f5STejun Heo cpu_relax(); 17779e8cd2f5STejun Heo goto retry; 17789e8cd2f5STejun Heo } 17799e8cd2f5STejun Heo /* oops */ 17809e8cd2f5STejun Heo WARN_ONCE(true, "workqueue: per-cpu pwq for %s on cpu%d has 0 refcnt", 17819e8cd2f5STejun Heo wq->name, cpu); 17829e8cd2f5STejun Heo } 17839e8cd2f5STejun Heo 1784112202d9STejun Heo /* pwq determined, queue */ 1785112202d9STejun Heo trace_workqueue_queue_work(req_cpu, pwq, work); 1786502ca9d8STejun Heo 178724acfb71SThomas Gleixner if (WARN_ON(!list_empty(&work->entry))) 178824acfb71SThomas Gleixner goto out; 17891e19ffc6STejun Heo 1790112202d9STejun Heo pwq->nr_in_flight[pwq->work_color]++; 1791112202d9STejun Heo work_flags = work_color_to_flags(pwq->work_color); 17921e19ffc6STejun Heo 1793112202d9STejun Heo if (likely(pwq->nr_active < pwq->max_active)) { 1794fe089f87STejun Heo if (list_empty(&pool->worklist)) 1795fe089f87STejun Heo pool->watchdog_ts = jiffies; 1796fe089f87STejun Heo 1797cdadf009STejun Heo trace_workqueue_activate_work(work); 1798112202d9STejun Heo pwq->nr_active++; 1799fe089f87STejun Heo insert_work(pwq, work, &pool->worklist, work_flags); 18000219a352STejun Heo kick_pool(pool); 18018a2e8e5dSTejun Heo } else { 1802f97a4a1aSLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 1803fe089f87STejun Heo insert_work(pwq, work, &pwq->inactive_works, work_flags); 18048a2e8e5dSTejun Heo } 18051e19ffc6STejun Heo 180624acfb71SThomas Gleixner out: 1807fe089f87STejun Heo raw_spin_unlock(&pool->lock); 180824acfb71SThomas Gleixner rcu_read_unlock(); 18091da177e4SLinus Torvalds } 18101da177e4SLinus Torvalds 18110fcb78c2SRolf Eike Beer /** 1812c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1813c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1814c1a220e7SZhang Rui * @wq: workqueue to use 1815c1a220e7SZhang Rui * @work: work to queue 1816c1a220e7SZhang Rui * 1817c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1818443378f0SPaul E. McKenney * can't go away. Callers that fail to ensure that the specified 1819443378f0SPaul E. McKenney * CPU cannot go away will execute on a randomly chosen CPU. 1820854f5cc5SPaul E. McKenney * But note well that callers specifying a CPU that never has been 1821854f5cc5SPaul E. McKenney * online will get a splat. 1822d185af30SYacine Belkadi * 1823d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. 1824c1a220e7SZhang Rui */ 1825d4283e93STejun Heo bool queue_work_on(int cpu, struct workqueue_struct *wq, 1826d4283e93STejun Heo struct work_struct *work) 1827c1a220e7SZhang Rui { 1828d4283e93STejun Heo bool ret = false; 18298930cabaSTejun Heo unsigned long flags; 18308930cabaSTejun Heo 18318930cabaSTejun Heo local_irq_save(flags); 1832c1a220e7SZhang Rui 183322df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 18344690c4abSTejun Heo __queue_work(cpu, wq, work); 1835d4283e93STejun Heo ret = true; 1836c1a220e7SZhang Rui } 18378930cabaSTejun Heo 18388930cabaSTejun Heo local_irq_restore(flags); 1839c1a220e7SZhang Rui return ret; 1840c1a220e7SZhang Rui } 1841ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_work_on); 1842c1a220e7SZhang Rui 18438204e0c1SAlexander Duyck /** 1844fef59c9cSTejun Heo * select_numa_node_cpu - Select a CPU based on NUMA node 18458204e0c1SAlexander Duyck * @node: NUMA node ID that we want to select a CPU from 18468204e0c1SAlexander Duyck * 18478204e0c1SAlexander Duyck * This function will attempt to find a "random" cpu available on a given 18488204e0c1SAlexander Duyck * node. If there are no CPUs available on the given node it will return 18498204e0c1SAlexander Duyck * WORK_CPU_UNBOUND indicating that we should just schedule to any 18508204e0c1SAlexander Duyck * available CPU if we need to schedule this work. 18518204e0c1SAlexander Duyck */ 1852fef59c9cSTejun Heo static int select_numa_node_cpu(int node) 18538204e0c1SAlexander Duyck { 18548204e0c1SAlexander Duyck int cpu; 18558204e0c1SAlexander Duyck 18568204e0c1SAlexander Duyck /* Delay binding to CPU if node is not valid or online */ 18578204e0c1SAlexander Duyck if (node < 0 || node >= MAX_NUMNODES || !node_online(node)) 18588204e0c1SAlexander Duyck return WORK_CPU_UNBOUND; 18598204e0c1SAlexander Duyck 18608204e0c1SAlexander Duyck /* Use local node/cpu if we are already there */ 18618204e0c1SAlexander Duyck cpu = raw_smp_processor_id(); 18628204e0c1SAlexander Duyck if (node == cpu_to_node(cpu)) 18638204e0c1SAlexander Duyck return cpu; 18648204e0c1SAlexander Duyck 18658204e0c1SAlexander Duyck /* Use "random" otherwise know as "first" online CPU of node */ 18668204e0c1SAlexander Duyck cpu = cpumask_any_and(cpumask_of_node(node), cpu_online_mask); 18678204e0c1SAlexander Duyck 18688204e0c1SAlexander Duyck /* If CPU is valid return that, otherwise just defer */ 18698204e0c1SAlexander Duyck return cpu < nr_cpu_ids ? cpu : WORK_CPU_UNBOUND; 18708204e0c1SAlexander Duyck } 18718204e0c1SAlexander Duyck 18728204e0c1SAlexander Duyck /** 18738204e0c1SAlexander Duyck * queue_work_node - queue work on a "random" cpu for a given NUMA node 18748204e0c1SAlexander Duyck * @node: NUMA node that we are targeting the work for 18758204e0c1SAlexander Duyck * @wq: workqueue to use 18768204e0c1SAlexander Duyck * @work: work to queue 18778204e0c1SAlexander Duyck * 18788204e0c1SAlexander Duyck * We queue the work to a "random" CPU within a given NUMA node. The basic 18798204e0c1SAlexander Duyck * idea here is to provide a way to somehow associate work with a given 18808204e0c1SAlexander Duyck * NUMA node. 18818204e0c1SAlexander Duyck * 18828204e0c1SAlexander Duyck * This function will only make a best effort attempt at getting this onto 18838204e0c1SAlexander Duyck * the right NUMA node. If no node is requested or the requested node is 18848204e0c1SAlexander Duyck * offline then we just fall back to standard queue_work behavior. 18858204e0c1SAlexander Duyck * 18868204e0c1SAlexander Duyck * Currently the "random" CPU ends up being the first available CPU in the 18878204e0c1SAlexander Duyck * intersection of cpu_online_mask and the cpumask of the node, unless we 18888204e0c1SAlexander Duyck * are running on the node. In that case we just use the current CPU. 18898204e0c1SAlexander Duyck * 18908204e0c1SAlexander Duyck * Return: %false if @work was already on a queue, %true otherwise. 18918204e0c1SAlexander Duyck */ 18928204e0c1SAlexander Duyck bool queue_work_node(int node, struct workqueue_struct *wq, 18938204e0c1SAlexander Duyck struct work_struct *work) 18948204e0c1SAlexander Duyck { 18958204e0c1SAlexander Duyck unsigned long flags; 18968204e0c1SAlexander Duyck bool ret = false; 18978204e0c1SAlexander Duyck 18988204e0c1SAlexander Duyck /* 18998204e0c1SAlexander Duyck * This current implementation is specific to unbound workqueues. 19008204e0c1SAlexander Duyck * Specifically we only return the first available CPU for a given 19018204e0c1SAlexander Duyck * node instead of cycling through individual CPUs within the node. 19028204e0c1SAlexander Duyck * 19038204e0c1SAlexander Duyck * If this is used with a per-cpu workqueue then the logic in 19048204e0c1SAlexander Duyck * workqueue_select_cpu_near would need to be updated to allow for 19058204e0c1SAlexander Duyck * some round robin type logic. 19068204e0c1SAlexander Duyck */ 19078204e0c1SAlexander Duyck WARN_ON_ONCE(!(wq->flags & WQ_UNBOUND)); 19088204e0c1SAlexander Duyck 19098204e0c1SAlexander Duyck local_irq_save(flags); 19108204e0c1SAlexander Duyck 19118204e0c1SAlexander Duyck if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1912fef59c9cSTejun Heo int cpu = select_numa_node_cpu(node); 19138204e0c1SAlexander Duyck 19148204e0c1SAlexander Duyck __queue_work(cpu, wq, work); 19158204e0c1SAlexander Duyck ret = true; 19168204e0c1SAlexander Duyck } 19178204e0c1SAlexander Duyck 19188204e0c1SAlexander Duyck local_irq_restore(flags); 19198204e0c1SAlexander Duyck return ret; 19208204e0c1SAlexander Duyck } 19218204e0c1SAlexander Duyck EXPORT_SYMBOL_GPL(queue_work_node); 19228204e0c1SAlexander Duyck 19238c20feb6SKees Cook void delayed_work_timer_fn(struct timer_list *t) 19241da177e4SLinus Torvalds { 19258c20feb6SKees Cook struct delayed_work *dwork = from_timer(dwork, t, timer); 19261da177e4SLinus Torvalds 1927e0aecdd8STejun Heo /* should have been called from irqsafe timer with irq already off */ 192860c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 19291da177e4SLinus Torvalds } 19301438ade5SKonstantin Khlebnikov EXPORT_SYMBOL(delayed_work_timer_fn); 19311da177e4SLinus Torvalds 19327beb2edfSTejun Heo static void __queue_delayed_work(int cpu, struct workqueue_struct *wq, 193352bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19341da177e4SLinus Torvalds { 19357beb2edfSTejun Heo struct timer_list *timer = &dwork->timer; 19367beb2edfSTejun Heo struct work_struct *work = &dwork->work; 19371da177e4SLinus Torvalds 1938637fdbaeSTejun Heo WARN_ON_ONCE(!wq); 19394b243563SSami Tolvanen WARN_ON_ONCE(timer->function != delayed_work_timer_fn); 1940fc4b514fSTejun Heo WARN_ON_ONCE(timer_pending(timer)); 1941fc4b514fSTejun Heo WARN_ON_ONCE(!list_empty(&work->entry)); 19427beb2edfSTejun Heo 19438852aac2STejun Heo /* 19448852aac2STejun Heo * If @delay is 0, queue @dwork->work immediately. This is for 19458852aac2STejun Heo * both optimization and correctness. The earliest @timer can 19468852aac2STejun Heo * expire is on the closest next tick and delayed_work users depend 19478852aac2STejun Heo * on that there's no such delay when @delay is 0. 19488852aac2STejun Heo */ 19498852aac2STejun Heo if (!delay) { 19508852aac2STejun Heo __queue_work(cpu, wq, &dwork->work); 19518852aac2STejun Heo return; 19528852aac2STejun Heo } 19538852aac2STejun Heo 195460c057bcSLai Jiangshan dwork->wq = wq; 19551265057fSTejun Heo dwork->cpu = cpu; 19567beb2edfSTejun Heo timer->expires = jiffies + delay; 19577beb2edfSTejun Heo 1958041bd12eSTejun Heo if (unlikely(cpu != WORK_CPU_UNBOUND)) 19597beb2edfSTejun Heo add_timer_on(timer, cpu); 1960041bd12eSTejun Heo else 1961041bd12eSTejun Heo add_timer(timer); 19627beb2edfSTejun Heo } 19631da177e4SLinus Torvalds 19640fcb78c2SRolf Eike Beer /** 19650fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 19660fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 19670fcb78c2SRolf Eike Beer * @wq: workqueue to use 1968af9997e4SRandy Dunlap * @dwork: work to queue 19690fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 19700fcb78c2SRolf Eike Beer * 1971d185af30SYacine Belkadi * Return: %false if @work was already on a queue, %true otherwise. If 1972715f1300STejun Heo * @delay is zero and @dwork is idle, it will be scheduled for immediate 1973715f1300STejun Heo * execution. 19740fcb78c2SRolf Eike Beer */ 1975d4283e93STejun Heo bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 197652bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 19777a6bc1cdSVenkatesh Pallipadi { 197852bad64dSDavid Howells struct work_struct *work = &dwork->work; 1979d4283e93STejun Heo bool ret = false; 19808930cabaSTejun Heo unsigned long flags; 19818930cabaSTejun Heo 19828930cabaSTejun Heo /* read the comment in __queue_work() */ 19838930cabaSTejun Heo local_irq_save(flags); 19847a6bc1cdSVenkatesh Pallipadi 198522df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 19867beb2edfSTejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 1987d4283e93STejun Heo ret = true; 19887a6bc1cdSVenkatesh Pallipadi } 19898930cabaSTejun Heo 19908930cabaSTejun Heo local_irq_restore(flags); 19917a6bc1cdSVenkatesh Pallipadi return ret; 19927a6bc1cdSVenkatesh Pallipadi } 1993ad7b1f84SMarc Dionne EXPORT_SYMBOL(queue_delayed_work_on); 19941da177e4SLinus Torvalds 1995c8e55f36STejun Heo /** 19968376fe22STejun Heo * mod_delayed_work_on - modify delay of or queue a delayed work on specific CPU 19978376fe22STejun Heo * @cpu: CPU number to execute work on 19988376fe22STejun Heo * @wq: workqueue to use 19998376fe22STejun Heo * @dwork: work to queue 20008376fe22STejun Heo * @delay: number of jiffies to wait before queueing 20018376fe22STejun Heo * 20028376fe22STejun Heo * If @dwork is idle, equivalent to queue_delayed_work_on(); otherwise, 20038376fe22STejun Heo * modify @dwork's timer so that it expires after @delay. If @delay is 20048376fe22STejun Heo * zero, @work is guaranteed to be scheduled immediately regardless of its 20058376fe22STejun Heo * current state. 20068376fe22STejun Heo * 2007d185af30SYacine Belkadi * Return: %false if @dwork was idle and queued, %true if @dwork was 20088376fe22STejun Heo * pending and its timer was modified. 20098376fe22STejun Heo * 2010e0aecdd8STejun Heo * This function is safe to call from any context including IRQ handler. 20118376fe22STejun Heo * See try_to_grab_pending() for details. 20128376fe22STejun Heo */ 20138376fe22STejun Heo bool mod_delayed_work_on(int cpu, struct workqueue_struct *wq, 20148376fe22STejun Heo struct delayed_work *dwork, unsigned long delay) 20158376fe22STejun Heo { 20168376fe22STejun Heo unsigned long flags; 20178376fe22STejun Heo int ret; 20188376fe22STejun Heo 20198376fe22STejun Heo do { 20208376fe22STejun Heo ret = try_to_grab_pending(&dwork->work, true, &flags); 20218376fe22STejun Heo } while (unlikely(ret == -EAGAIN)); 20228376fe22STejun Heo 20238376fe22STejun Heo if (likely(ret >= 0)) { 20248376fe22STejun Heo __queue_delayed_work(cpu, wq, dwork, delay); 20258376fe22STejun Heo local_irq_restore(flags); 20268376fe22STejun Heo } 20278376fe22STejun Heo 20288376fe22STejun Heo /* -ENOENT from try_to_grab_pending() becomes %true */ 20298376fe22STejun Heo return ret; 20308376fe22STejun Heo } 20318376fe22STejun Heo EXPORT_SYMBOL_GPL(mod_delayed_work_on); 20328376fe22STejun Heo 203305f0fe6bSTejun Heo static void rcu_work_rcufn(struct rcu_head *rcu) 203405f0fe6bSTejun Heo { 203505f0fe6bSTejun Heo struct rcu_work *rwork = container_of(rcu, struct rcu_work, rcu); 203605f0fe6bSTejun Heo 203705f0fe6bSTejun Heo /* read the comment in __queue_work() */ 203805f0fe6bSTejun Heo local_irq_disable(); 203905f0fe6bSTejun Heo __queue_work(WORK_CPU_UNBOUND, rwork->wq, &rwork->work); 204005f0fe6bSTejun Heo local_irq_enable(); 204105f0fe6bSTejun Heo } 204205f0fe6bSTejun Heo 204305f0fe6bSTejun Heo /** 204405f0fe6bSTejun Heo * queue_rcu_work - queue work after a RCU grace period 204505f0fe6bSTejun Heo * @wq: workqueue to use 204605f0fe6bSTejun Heo * @rwork: work to queue 204705f0fe6bSTejun Heo * 204805f0fe6bSTejun Heo * Return: %false if @rwork was already pending, %true otherwise. Note 204905f0fe6bSTejun Heo * that a full RCU grace period is guaranteed only after a %true return. 2050bf393fd4SBart Van Assche * While @rwork is guaranteed to be executed after a %false return, the 205105f0fe6bSTejun Heo * execution may happen before a full RCU grace period has passed. 205205f0fe6bSTejun Heo */ 205305f0fe6bSTejun Heo bool queue_rcu_work(struct workqueue_struct *wq, struct rcu_work *rwork) 205405f0fe6bSTejun Heo { 205505f0fe6bSTejun Heo struct work_struct *work = &rwork->work; 205605f0fe6bSTejun Heo 205705f0fe6bSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 205805f0fe6bSTejun Heo rwork->wq = wq; 2059a7e30c0eSUladzislau Rezki call_rcu_hurry(&rwork->rcu, rcu_work_rcufn); 206005f0fe6bSTejun Heo return true; 206105f0fe6bSTejun Heo } 206205f0fe6bSTejun Heo 206305f0fe6bSTejun Heo return false; 206405f0fe6bSTejun Heo } 206505f0fe6bSTejun Heo EXPORT_SYMBOL(queue_rcu_work); 206605f0fe6bSTejun Heo 2067f7537df5SLai Jiangshan static struct worker *alloc_worker(int node) 2068c34056a3STejun Heo { 2069c34056a3STejun Heo struct worker *worker; 2070c34056a3STejun Heo 2071f7537df5SLai Jiangshan worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, node); 2072c8e55f36STejun Heo if (worker) { 2073c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 2074affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 2075da028469SLai Jiangshan INIT_LIST_HEAD(&worker->node); 2076e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 2077e22bee78STejun Heo worker->flags = WORKER_PREP; 2078c8e55f36STejun Heo } 2079c34056a3STejun Heo return worker; 2080c34056a3STejun Heo } 2081c34056a3STejun Heo 20829546b29eSTejun Heo static cpumask_t *pool_allowed_cpus(struct worker_pool *pool) 20839546b29eSTejun Heo { 20848639ecebSTejun Heo if (pool->cpu < 0 && pool->attrs->affn_strict) 20859546b29eSTejun Heo return pool->attrs->__pod_cpumask; 20868639ecebSTejun Heo else 20878639ecebSTejun Heo return pool->attrs->cpumask; 20889546b29eSTejun Heo } 20899546b29eSTejun Heo 2090c34056a3STejun Heo /** 20914736cbf7SLai Jiangshan * worker_attach_to_pool() - attach a worker to a pool 20924736cbf7SLai Jiangshan * @worker: worker to be attached 20934736cbf7SLai Jiangshan * @pool: the target pool 20944736cbf7SLai Jiangshan * 20954736cbf7SLai Jiangshan * Attach @worker to @pool. Once attached, the %WORKER_UNBOUND flag and 20964736cbf7SLai Jiangshan * cpu-binding of @worker are kept coordinated with the pool across 20974736cbf7SLai Jiangshan * cpu-[un]hotplugs. 20984736cbf7SLai Jiangshan */ 20994736cbf7SLai Jiangshan static void worker_attach_to_pool(struct worker *worker, 21004736cbf7SLai Jiangshan struct worker_pool *pool) 21014736cbf7SLai Jiangshan { 21021258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 21034736cbf7SLai Jiangshan 21044736cbf7SLai Jiangshan /* 21051258fae7STejun Heo * The wq_pool_attach_mutex ensures %POOL_DISASSOCIATED remains 21061258fae7STejun Heo * stable across this function. See the comments above the flag 21071258fae7STejun Heo * definition for details. 21084736cbf7SLai Jiangshan */ 21094736cbf7SLai Jiangshan if (pool->flags & POOL_DISASSOCIATED) 21104736cbf7SLai Jiangshan worker->flags |= WORKER_UNBOUND; 21115c25b5ffSPeter Zijlstra else 21125c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, pool->cpu); 21134736cbf7SLai Jiangshan 2114640f17c8SPeter Zijlstra if (worker->rescue_wq) 21159546b29eSTejun Heo set_cpus_allowed_ptr(worker->task, pool_allowed_cpus(pool)); 2116640f17c8SPeter Zijlstra 21174736cbf7SLai Jiangshan list_add_tail(&worker->node, &pool->workers); 2118a2d812a2STejun Heo worker->pool = pool; 21194736cbf7SLai Jiangshan 21201258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 21214736cbf7SLai Jiangshan } 21224736cbf7SLai Jiangshan 21234736cbf7SLai Jiangshan /** 212460f5a4bcSLai Jiangshan * worker_detach_from_pool() - detach a worker from its pool 212560f5a4bcSLai Jiangshan * @worker: worker which is attached to its pool 212660f5a4bcSLai Jiangshan * 21274736cbf7SLai Jiangshan * Undo the attaching which had been done in worker_attach_to_pool(). The 21284736cbf7SLai Jiangshan * caller worker shouldn't access to the pool after detached except it has 21294736cbf7SLai Jiangshan * other reference to the pool. 213060f5a4bcSLai Jiangshan */ 2131a2d812a2STejun Heo static void worker_detach_from_pool(struct worker *worker) 213260f5a4bcSLai Jiangshan { 2133a2d812a2STejun Heo struct worker_pool *pool = worker->pool; 213460f5a4bcSLai Jiangshan struct completion *detach_completion = NULL; 213560f5a4bcSLai Jiangshan 21361258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 2137a2d812a2STejun Heo 21385c25b5ffSPeter Zijlstra kthread_set_per_cpu(worker->task, -1); 2139da028469SLai Jiangshan list_del(&worker->node); 2140a2d812a2STejun Heo worker->pool = NULL; 2141a2d812a2STejun Heo 2142e02b9312SValentin Schneider if (list_empty(&pool->workers) && list_empty(&pool->dying_workers)) 214360f5a4bcSLai Jiangshan detach_completion = pool->detach_completion; 21441258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 214560f5a4bcSLai Jiangshan 2146b62c0751SLai Jiangshan /* clear leftover flags without pool->lock after it is detached */ 2147b62c0751SLai Jiangshan worker->flags &= ~(WORKER_UNBOUND | WORKER_REBOUND); 2148b62c0751SLai Jiangshan 214960f5a4bcSLai Jiangshan if (detach_completion) 215060f5a4bcSLai Jiangshan complete(detach_completion); 215160f5a4bcSLai Jiangshan } 215260f5a4bcSLai Jiangshan 215360f5a4bcSLai Jiangshan /** 2154c34056a3STejun Heo * create_worker - create a new workqueue worker 215563d95a91STejun Heo * @pool: pool the new worker will belong to 2156c34056a3STejun Heo * 2157051e1850SLai Jiangshan * Create and start a new worker which is attached to @pool. 2158c34056a3STejun Heo * 2159c34056a3STejun Heo * CONTEXT: 2160c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 2161c34056a3STejun Heo * 2162d185af30SYacine Belkadi * Return: 2163c34056a3STejun Heo * Pointer to the newly created worker. 2164c34056a3STejun Heo */ 2165bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 2166c34056a3STejun Heo { 2167e441b56fSZhen Lei struct worker *worker; 2168e441b56fSZhen Lei int id; 2169e3c916a4STejun Heo char id_buf[16]; 2170c34056a3STejun Heo 21717cda9aaeSLai Jiangshan /* ID is needed to determine kthread name */ 2172e441b56fSZhen Lei id = ida_alloc(&pool->worker_ida, GFP_KERNEL); 21733f0ea0b8SPetr Mladek if (id < 0) { 21743f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker ID: %pe\n", 21753f0ea0b8SPetr Mladek ERR_PTR(id)); 2176e441b56fSZhen Lei return NULL; 21773f0ea0b8SPetr Mladek } 2178c34056a3STejun Heo 2179f7537df5SLai Jiangshan worker = alloc_worker(pool->node); 21803f0ea0b8SPetr Mladek if (!worker) { 21813f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to allocate a worker\n"); 2182c34056a3STejun Heo goto fail; 21833f0ea0b8SPetr Mladek } 2184c34056a3STejun Heo 2185c34056a3STejun Heo worker->id = id; 2186c34056a3STejun Heo 218729c91e99STejun Heo if (pool->cpu >= 0) 2188e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "%d:%d%s", pool->cpu, id, 2189e3c916a4STejun Heo pool->attrs->nice < 0 ? "H" : ""); 2190f3421797STejun Heo else 2191e3c916a4STejun Heo snprintf(id_buf, sizeof(id_buf), "u%d:%d", pool->id, id); 2192e3c916a4STejun Heo 2193f3f90ad4STejun Heo worker->task = kthread_create_on_node(worker_thread, worker, pool->node, 2194e3c916a4STejun Heo "kworker/%s", id_buf); 21953f0ea0b8SPetr Mladek if (IS_ERR(worker->task)) { 219660f54038SPetr Mladek if (PTR_ERR(worker->task) == -EINTR) { 219760f54038SPetr Mladek pr_err("workqueue: Interrupted when creating a worker thread \"kworker/%s\"\n", 219860f54038SPetr Mladek id_buf); 219960f54038SPetr Mladek } else { 22003f0ea0b8SPetr Mladek pr_err_once("workqueue: Failed to create a worker thread: %pe", 22013f0ea0b8SPetr Mladek worker->task); 220260f54038SPetr Mladek } 2203c34056a3STejun Heo goto fail; 22043f0ea0b8SPetr Mladek } 2205c34056a3STejun Heo 220691151228SOleg Nesterov set_user_nice(worker->task, pool->attrs->nice); 22079546b29eSTejun Heo kthread_bind_mask(worker->task, pool_allowed_cpus(pool)); 220891151228SOleg Nesterov 2209da028469SLai Jiangshan /* successful, attach the worker to the pool */ 22104736cbf7SLai Jiangshan worker_attach_to_pool(worker, pool); 2211822d8405STejun Heo 2212051e1850SLai Jiangshan /* start the newly created worker */ 2213a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 22140219a352STejun Heo 2215051e1850SLai Jiangshan worker->pool->nr_workers++; 2216051e1850SLai Jiangshan worker_enter_idle(worker); 22170219a352STejun Heo kick_pool(pool); 22180219a352STejun Heo 22190219a352STejun Heo /* 22200219a352STejun Heo * @worker is waiting on a completion in kthread() and will trigger hung 22210219a352STejun Heo * check if not woken up soon. As kick_pool() might not have waken it 22220219a352STejun Heo * up, wake it up explicitly once more. 22230219a352STejun Heo */ 2224051e1850SLai Jiangshan wake_up_process(worker->task); 22250219a352STejun Heo 2226a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2227051e1850SLai Jiangshan 2228c34056a3STejun Heo return worker; 2229822d8405STejun Heo 2230c34056a3STejun Heo fail: 2231e441b56fSZhen Lei ida_free(&pool->worker_ida, id); 2232c34056a3STejun Heo kfree(worker); 2233c34056a3STejun Heo return NULL; 2234c34056a3STejun Heo } 2235c34056a3STejun Heo 2236793777bcSValentin Schneider static void unbind_worker(struct worker *worker) 2237793777bcSValentin Schneider { 2238793777bcSValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2239793777bcSValentin Schneider 2240793777bcSValentin Schneider kthread_set_per_cpu(worker->task, -1); 2241793777bcSValentin Schneider if (cpumask_intersects(wq_unbound_cpumask, cpu_active_mask)) 2242793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, wq_unbound_cpumask) < 0); 2243793777bcSValentin Schneider else 2244793777bcSValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, cpu_possible_mask) < 0); 2245793777bcSValentin Schneider } 2246793777bcSValentin Schneider 2247e02b9312SValentin Schneider static void wake_dying_workers(struct list_head *cull_list) 2248e02b9312SValentin Schneider { 2249e02b9312SValentin Schneider struct worker *worker, *tmp; 2250e02b9312SValentin Schneider 2251e02b9312SValentin Schneider list_for_each_entry_safe(worker, tmp, cull_list, entry) { 2252e02b9312SValentin Schneider list_del_init(&worker->entry); 2253e02b9312SValentin Schneider unbind_worker(worker); 2254e02b9312SValentin Schneider /* 2255e02b9312SValentin Schneider * If the worker was somehow already running, then it had to be 2256e02b9312SValentin Schneider * in pool->idle_list when set_worker_dying() happened or we 2257e02b9312SValentin Schneider * wouldn't have gotten here. 2258c34056a3STejun Heo * 2259e02b9312SValentin Schneider * Thus, the worker must either have observed the WORKER_DIE 2260e02b9312SValentin Schneider * flag, or have set its state to TASK_IDLE. Either way, the 2261e02b9312SValentin Schneider * below will be observed by the worker and is safe to do 2262e02b9312SValentin Schneider * outside of pool->lock. 2263e02b9312SValentin Schneider */ 2264e02b9312SValentin Schneider wake_up_process(worker->task); 2265e02b9312SValentin Schneider } 2266e02b9312SValentin Schneider } 2267e02b9312SValentin Schneider 2268e02b9312SValentin Schneider /** 2269e02b9312SValentin Schneider * set_worker_dying - Tag a worker for destruction 2270e02b9312SValentin Schneider * @worker: worker to be destroyed 2271e02b9312SValentin Schneider * @list: transfer worker away from its pool->idle_list and into list 2272e02b9312SValentin Schneider * 2273e02b9312SValentin Schneider * Tag @worker for destruction and adjust @pool stats accordingly. The worker 2274e02b9312SValentin Schneider * should be idle. 2275c8e55f36STejun Heo * 2276c8e55f36STejun Heo * CONTEXT: 2277a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 2278c34056a3STejun Heo */ 2279e02b9312SValentin Schneider static void set_worker_dying(struct worker *worker, struct list_head *list) 2280c34056a3STejun Heo { 2281bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2282c34056a3STejun Heo 2283cd549687STejun Heo lockdep_assert_held(&pool->lock); 2284e02b9312SValentin Schneider lockdep_assert_held(&wq_pool_attach_mutex); 2285cd549687STejun Heo 2286c34056a3STejun Heo /* sanity check frenzy */ 22876183c009STejun Heo if (WARN_ON(worker->current_work) || 228873eb7fe7SLai Jiangshan WARN_ON(!list_empty(&worker->scheduled)) || 228973eb7fe7SLai Jiangshan WARN_ON(!(worker->flags & WORKER_IDLE))) 22906183c009STejun Heo return; 2291c34056a3STejun Heo 2292bd7bdd43STejun Heo pool->nr_workers--; 2293bd7bdd43STejun Heo pool->nr_idle--; 2294c8e55f36STejun Heo 2295cb444766STejun Heo worker->flags |= WORKER_DIE; 2296e02b9312SValentin Schneider 2297e02b9312SValentin Schneider list_move(&worker->entry, list); 2298e02b9312SValentin Schneider list_move(&worker->node, &pool->dying_workers); 2299c34056a3STejun Heo } 2300c34056a3STejun Heo 23013f959aa3SValentin Schneider /** 23023f959aa3SValentin Schneider * idle_worker_timeout - check if some idle workers can now be deleted. 23033f959aa3SValentin Schneider * @t: The pool's idle_timer that just expired 23043f959aa3SValentin Schneider * 23053f959aa3SValentin Schneider * The timer is armed in worker_enter_idle(). Note that it isn't disarmed in 23063f959aa3SValentin Schneider * worker_leave_idle(), as a worker flicking between idle and active while its 23073f959aa3SValentin Schneider * pool is at the too_many_workers() tipping point would cause too much timer 23083f959aa3SValentin Schneider * housekeeping overhead. Since IDLE_WORKER_TIMEOUT is long enough, we just let 23093f959aa3SValentin Schneider * it expire and re-evaluate things from there. 23103f959aa3SValentin Schneider */ 231132a6c723SKees Cook static void idle_worker_timeout(struct timer_list *t) 2312e22bee78STejun Heo { 231332a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, idle_timer); 23143f959aa3SValentin Schneider bool do_cull = false; 23153f959aa3SValentin Schneider 23163f959aa3SValentin Schneider if (work_pending(&pool->idle_cull_work)) 23173f959aa3SValentin Schneider return; 23183f959aa3SValentin Schneider 23193f959aa3SValentin Schneider raw_spin_lock_irq(&pool->lock); 23203f959aa3SValentin Schneider 23213f959aa3SValentin Schneider if (too_many_workers(pool)) { 23223f959aa3SValentin Schneider struct worker *worker; 23233f959aa3SValentin Schneider unsigned long expires; 23243f959aa3SValentin Schneider 23253f959aa3SValentin Schneider /* idle_list is kept in LIFO order, check the last one */ 23263f959aa3SValentin Schneider worker = list_entry(pool->idle_list.prev, struct worker, entry); 23273f959aa3SValentin Schneider expires = worker->last_active + IDLE_WORKER_TIMEOUT; 23283f959aa3SValentin Schneider do_cull = !time_before(jiffies, expires); 23293f959aa3SValentin Schneider 23303f959aa3SValentin Schneider if (!do_cull) 23313f959aa3SValentin Schneider mod_timer(&pool->idle_timer, expires); 23323f959aa3SValentin Schneider } 23333f959aa3SValentin Schneider raw_spin_unlock_irq(&pool->lock); 23343f959aa3SValentin Schneider 23353f959aa3SValentin Schneider if (do_cull) 23363f959aa3SValentin Schneider queue_work(system_unbound_wq, &pool->idle_cull_work); 23373f959aa3SValentin Schneider } 23383f959aa3SValentin Schneider 23393f959aa3SValentin Schneider /** 23403f959aa3SValentin Schneider * idle_cull_fn - cull workers that have been idle for too long. 23413f959aa3SValentin Schneider * @work: the pool's work for handling these idle workers 23423f959aa3SValentin Schneider * 23433f959aa3SValentin Schneider * This goes through a pool's idle workers and gets rid of those that have been 23443f959aa3SValentin Schneider * idle for at least IDLE_WORKER_TIMEOUT seconds. 2345e02b9312SValentin Schneider * 2346e02b9312SValentin Schneider * We don't want to disturb isolated CPUs because of a pcpu kworker being 2347e02b9312SValentin Schneider * culled, so this also resets worker affinity. This requires a sleepable 2348e02b9312SValentin Schneider * context, hence the split between timer callback and work item. 23493f959aa3SValentin Schneider */ 23503f959aa3SValentin Schneider static void idle_cull_fn(struct work_struct *work) 23513f959aa3SValentin Schneider { 23523f959aa3SValentin Schneider struct worker_pool *pool = container_of(work, struct worker_pool, idle_cull_work); 23539680540cSYang Yingliang LIST_HEAD(cull_list); 2354e22bee78STejun Heo 2355e02b9312SValentin Schneider /* 2356e02b9312SValentin Schneider * Grabbing wq_pool_attach_mutex here ensures an already-running worker 2357e02b9312SValentin Schneider * cannot proceed beyong worker_detach_from_pool() in its self-destruct 2358e02b9312SValentin Schneider * path. This is required as a previously-preempted worker could run after 2359e02b9312SValentin Schneider * set_worker_dying() has happened but before wake_dying_workers() did. 2360e02b9312SValentin Schneider */ 2361e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 2362a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2363e22bee78STejun Heo 23643347fc9fSLai Jiangshan while (too_many_workers(pool)) { 2365e22bee78STejun Heo struct worker *worker; 2366e22bee78STejun Heo unsigned long expires; 2367e22bee78STejun Heo 236863d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 2369e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 2370e22bee78STejun Heo 23713347fc9fSLai Jiangshan if (time_before(jiffies, expires)) { 237263d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 23733347fc9fSLai Jiangshan break; 2374e22bee78STejun Heo } 23753347fc9fSLai Jiangshan 2376e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 2377e22bee78STejun Heo } 2378e22bee78STejun Heo 2379a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2380e02b9312SValentin Schneider wake_dying_workers(&cull_list); 2381e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 2382e22bee78STejun Heo } 2383e22bee78STejun Heo 2384493a1724STejun Heo static void send_mayday(struct work_struct *work) 2385e22bee78STejun Heo { 2386112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2387112202d9STejun Heo struct workqueue_struct *wq = pwq->wq; 2388493a1724STejun Heo 23892e109a28STejun Heo lockdep_assert_held(&wq_mayday_lock); 2390e22bee78STejun Heo 2391493008a8STejun Heo if (!wq->rescuer) 2392493a1724STejun Heo return; 2393e22bee78STejun Heo 2394e22bee78STejun Heo /* mayday mayday mayday */ 2395493a1724STejun Heo if (list_empty(&pwq->mayday_node)) { 239677668c8bSLai Jiangshan /* 239777668c8bSLai Jiangshan * If @pwq is for an unbound wq, its base ref may be put at 239877668c8bSLai Jiangshan * any time due to an attribute change. Pin @pwq until the 239977668c8bSLai Jiangshan * rescuer is done with it. 240077668c8bSLai Jiangshan */ 240177668c8bSLai Jiangshan get_pwq(pwq); 2402493a1724STejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2403e22bee78STejun Heo wake_up_process(wq->rescuer->task); 2404725e8ec5STejun Heo pwq->stats[PWQ_STAT_MAYDAY]++; 2405493a1724STejun Heo } 2406e22bee78STejun Heo } 2407e22bee78STejun Heo 240832a6c723SKees Cook static void pool_mayday_timeout(struct timer_list *t) 2409e22bee78STejun Heo { 241032a6c723SKees Cook struct worker_pool *pool = from_timer(pool, t, mayday_timer); 2411e22bee78STejun Heo struct work_struct *work; 2412e22bee78STejun Heo 2413a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2414a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); /* for wq->maydays */ 2415e22bee78STejun Heo 241663d95a91STejun Heo if (need_to_create_worker(pool)) { 2417e22bee78STejun Heo /* 2418e22bee78STejun Heo * We've been trying to create a new worker but 2419e22bee78STejun Heo * haven't been successful. We might be hitting an 2420e22bee78STejun Heo * allocation deadlock. Send distress signals to 2421e22bee78STejun Heo * rescuers. 2422e22bee78STejun Heo */ 242363d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 2424e22bee78STejun Heo send_mayday(work); 2425e22bee78STejun Heo } 2426e22bee78STejun Heo 2427a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2428a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2429e22bee78STejun Heo 243063d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 2431e22bee78STejun Heo } 2432e22bee78STejun Heo 2433e22bee78STejun Heo /** 2434e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 243563d95a91STejun Heo * @pool: pool to create a new worker for 2436e22bee78STejun Heo * 243763d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 2438e22bee78STejun Heo * have at least one idle worker on return from this function. If 2439e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 244063d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 2441e22bee78STejun Heo * possible allocation deadlock. 2442e22bee78STejun Heo * 2443c5aa87bbSTejun Heo * On return, need_to_create_worker() is guaranteed to be %false and 2444c5aa87bbSTejun Heo * may_start_working() %true. 2445e22bee78STejun Heo * 2446e22bee78STejun Heo * LOCKING: 2447a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2448e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 2449e22bee78STejun Heo * manager. 2450e22bee78STejun Heo */ 245129187a9eSTejun Heo static void maybe_create_worker(struct worker_pool *pool) 2452d565ed63STejun Heo __releases(&pool->lock) 2453d565ed63STejun Heo __acquires(&pool->lock) 2454e22bee78STejun Heo { 2455e22bee78STejun Heo restart: 2456a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 24579f9c2364STejun Heo 2458e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 245963d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 2460e22bee78STejun Heo 2461e22bee78STejun Heo while (true) { 2462051e1850SLai Jiangshan if (create_worker(pool) || !need_to_create_worker(pool)) 2463e22bee78STejun Heo break; 2464e22bee78STejun Heo 2465e212f361SLai Jiangshan schedule_timeout_interruptible(CREATE_COOLDOWN); 24669f9c2364STejun Heo 246763d95a91STejun Heo if (!need_to_create_worker(pool)) 2468e22bee78STejun Heo break; 2469e22bee78STejun Heo } 2470e22bee78STejun Heo 247163d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 2472a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2473051e1850SLai Jiangshan /* 2474051e1850SLai Jiangshan * This is necessary even after a new worker was just successfully 2475051e1850SLai Jiangshan * created as @pool->lock was dropped and the new worker might have 2476051e1850SLai Jiangshan * already become busy. 2477051e1850SLai Jiangshan */ 247863d95a91STejun Heo if (need_to_create_worker(pool)) 2479e22bee78STejun Heo goto restart; 2480e22bee78STejun Heo } 2481e22bee78STejun Heo 2482e22bee78STejun Heo /** 2483e22bee78STejun Heo * manage_workers - manage worker pool 2484e22bee78STejun Heo * @worker: self 2485e22bee78STejun Heo * 2486706026c2STejun Heo * Assume the manager role and manage the worker pool @worker belongs 2487e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 2488706026c2STejun Heo * pool. The exclusion is handled automatically by this function. 2489e22bee78STejun Heo * 2490e22bee78STejun Heo * The caller can safely start processing works on false return. On 2491e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 2492e22bee78STejun Heo * and may_start_working() is true. 2493e22bee78STejun Heo * 2494e22bee78STejun Heo * CONTEXT: 2495a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2496e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 2497e22bee78STejun Heo * 2498d185af30SYacine Belkadi * Return: 249929187a9eSTejun Heo * %false if the pool doesn't need management and the caller can safely 250029187a9eSTejun Heo * start processing works, %true if management function was performed and 250129187a9eSTejun Heo * the conditions that the caller verified before calling the function may 250229187a9eSTejun Heo * no longer be true. 2503e22bee78STejun Heo */ 2504e22bee78STejun Heo static bool manage_workers(struct worker *worker) 2505e22bee78STejun Heo { 250663d95a91STejun Heo struct worker_pool *pool = worker->pool; 2507e22bee78STejun Heo 2508692b4825STejun Heo if (pool->flags & POOL_MANAGER_ACTIVE) 250929187a9eSTejun Heo return false; 2510692b4825STejun Heo 2511692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 25122607d7a6STejun Heo pool->manager = worker; 2513e22bee78STejun Heo 251429187a9eSTejun Heo maybe_create_worker(pool); 2515e22bee78STejun Heo 25162607d7a6STejun Heo pool->manager = NULL; 2517692b4825STejun Heo pool->flags &= ~POOL_MANAGER_ACTIVE; 2518d8bb65abSSebastian Andrzej Siewior rcuwait_wake_up(&manager_wait); 251929187a9eSTejun Heo return true; 2520e22bee78STejun Heo } 2521e22bee78STejun Heo 2522a62428c0STejun Heo /** 2523a62428c0STejun Heo * process_one_work - process single work 2524c34056a3STejun Heo * @worker: self 2525a62428c0STejun Heo * @work: work to process 2526a62428c0STejun Heo * 2527a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 2528a62428c0STejun Heo * process a single work including synchronization against and 2529a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 2530a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 2531a62428c0STejun Heo * call this function to process a work. 2532a62428c0STejun Heo * 2533a62428c0STejun Heo * CONTEXT: 2534a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which is released and regrabbed. 2535a62428c0STejun Heo */ 2536c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 2537d565ed63STejun Heo __releases(&pool->lock) 2538d565ed63STejun Heo __acquires(&pool->lock) 25391da177e4SLinus Torvalds { 2540112202d9STejun Heo struct pool_workqueue *pwq = get_work_pwq(work); 2541bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2542c4560c2cSLai Jiangshan unsigned long work_data; 25434e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 25444e6045f1SJohannes Berg /* 2545a62428c0STejun Heo * It is permissible to free the struct work_struct from 2546a62428c0STejun Heo * inside the function that is called from it, this we need to 2547a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 2548a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 2549a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 25504e6045f1SJohannes Berg */ 25514d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 25524d82a1deSPeter Zijlstra 25534d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 25544e6045f1SJohannes Berg #endif 2555807407c0SLai Jiangshan /* ensure we're on the correct CPU */ 255685327af6SLai Jiangshan WARN_ON_ONCE(!(pool->flags & POOL_DISASSOCIATED) && 2557ec22ca5eSTejun Heo raw_smp_processor_id() != pool->cpu); 255825511a47STejun Heo 25598930cabaSTejun Heo /* claim and dequeue */ 2560dc186ad7SThomas Gleixner debug_work_deactivate(work); 2561c9e7cf27STejun Heo hash_add(pool->busy_hash, &worker->hentry, (unsigned long)work); 2562c34056a3STejun Heo worker->current_work = work; 2563a2c1c57bSTejun Heo worker->current_func = work->func; 2564112202d9STejun Heo worker->current_pwq = pwq; 2565616db877STejun Heo worker->current_at = worker->task->se.sum_exec_runtime; 2566c4560c2cSLai Jiangshan work_data = *work_data_bits(work); 2567d812796eSLai Jiangshan worker->current_color = get_work_color(work_data); 25687a22ad75STejun Heo 25698bf89593STejun Heo /* 25708bf89593STejun Heo * Record wq name for cmdline and debug reporting, may get 25718bf89593STejun Heo * overridden through set_worker_desc(). 25728bf89593STejun Heo */ 25738bf89593STejun Heo strscpy(worker->desc, pwq->wq->name, WORKER_DESC_LEN); 25748bf89593STejun Heo 2575a62428c0STejun Heo list_del_init(&work->entry); 2576a62428c0STejun Heo 2577649027d7STejun Heo /* 2578228f1d00SLai Jiangshan * CPU intensive works don't participate in concurrency management. 2579228f1d00SLai Jiangshan * They're the scheduler's responsibility. This takes @worker out 2580228f1d00SLai Jiangshan * of concurrency management and the next code block will chain 2581228f1d00SLai Jiangshan * execution of the pending work items. 2582fb0e7bebSTejun Heo */ 2583616db877STejun Heo if (unlikely(pwq->wq->flags & WQ_CPU_INTENSIVE)) 2584228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_CPU_INTENSIVE); 2585fb0e7bebSTejun Heo 2586974271c4STejun Heo /* 25870219a352STejun Heo * Kick @pool if necessary. It's always noop for per-cpu worker pools 25880219a352STejun Heo * since nr_running would always be >= 1 at this point. This is used to 25890219a352STejun Heo * chain execution of the pending work items for WORKER_NOT_RUNNING 25900219a352STejun Heo * workers such as the UNBOUND and CPU_INTENSIVE ones. 2591974271c4STejun Heo */ 25920219a352STejun Heo kick_pool(pool); 2593974271c4STejun Heo 25948930cabaSTejun Heo /* 25957c3eed5cSTejun Heo * Record the last pool and clear PENDING which should be the last 2596d565ed63STejun Heo * update to @work. Also, do this inside @pool->lock so that 259723657bb1STejun Heo * PENDING and queued state changes happen together while IRQ is 259823657bb1STejun Heo * disabled. 25998930cabaSTejun Heo */ 26007c3eed5cSTejun Heo set_work_pool_and_clear_pending(work, pool->id); 26011da177e4SLinus Torvalds 2602a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2603365970a1SDavid Howells 2604a1d14934SPeter Zijlstra lock_map_acquire(&pwq->wq->lockdep_map); 26053295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2606e6f3faa7SPeter Zijlstra /* 2607f52be570SPeter Zijlstra * Strictly speaking we should mark the invariant state without holding 2608f52be570SPeter Zijlstra * any locks, that is, before these two lock_map_acquire()'s. 2609e6f3faa7SPeter Zijlstra * 2610e6f3faa7SPeter Zijlstra * However, that would result in: 2611e6f3faa7SPeter Zijlstra * 2612e6f3faa7SPeter Zijlstra * A(W1) 2613e6f3faa7SPeter Zijlstra * WFC(C) 2614e6f3faa7SPeter Zijlstra * A(W1) 2615e6f3faa7SPeter Zijlstra * C(C) 2616e6f3faa7SPeter Zijlstra * 2617e6f3faa7SPeter Zijlstra * Which would create W1->C->W1 dependencies, even though there is no 2618e6f3faa7SPeter Zijlstra * actual deadlock possible. There are two solutions, using a 2619e6f3faa7SPeter Zijlstra * read-recursive acquire on the work(queue) 'locks', but this will then 2620f52be570SPeter Zijlstra * hit the lockdep limitation on recursive locks, or simply discard 2621e6f3faa7SPeter Zijlstra * these locks. 2622e6f3faa7SPeter Zijlstra * 2623e6f3faa7SPeter Zijlstra * AFAICT there is no possible deadlock scenario between the 2624e6f3faa7SPeter Zijlstra * flush_work() and complete() primitives (except for single-threaded 2625e6f3faa7SPeter Zijlstra * workqueues), so hiding them isn't a problem. 2626e6f3faa7SPeter Zijlstra */ 2627f52be570SPeter Zijlstra lockdep_invariant_state(true); 2628725e8ec5STejun Heo pwq->stats[PWQ_STAT_STARTED]++; 2629e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 2630a2c1c57bSTejun Heo worker->current_func(work); 2631e36c886aSArjan van de Ven /* 2632e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2633e36c886aSArjan van de Ven * point will only record its address. 2634e36c886aSArjan van de Ven */ 26351c5da0ecSDaniel Jordan trace_workqueue_execute_end(work, worker->current_func); 2636725e8ec5STejun Heo pwq->stats[PWQ_STAT_COMPLETED]++; 26373295f0efSIngo Molnar lock_map_release(&lockdep_map); 2638112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 26391da177e4SLinus Torvalds 2640d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2641044c782cSValentin Ilie pr_err("BUG: workqueue leaked lock or atomic: %s/0x%08x/%d\n" 2642d75f773cSSakari Ailus " last function: %ps\n", 2643a2c1c57bSTejun Heo current->comm, preempt_count(), task_pid_nr(current), 2644a2c1c57bSTejun Heo worker->current_func); 2645d5abe669SPeter Zijlstra debug_show_held_locks(current); 2646d5abe669SPeter Zijlstra dump_stack(); 2647d5abe669SPeter Zijlstra } 2648d5abe669SPeter Zijlstra 2649b22ce278STejun Heo /* 2650025f50f3SSebastian Andrzej Siewior * The following prevents a kworker from hogging CPU on !PREEMPTION 2651b22ce278STejun Heo * kernels, where a requeueing work item waiting for something to 2652b22ce278STejun Heo * happen could deadlock with stop_machine as such work item could 2653b22ce278STejun Heo * indefinitely requeue itself while all other CPUs are trapped in 2654789cbbecSJoe Lawrence * stop_machine. At the same time, report a quiescent RCU state so 2655789cbbecSJoe Lawrence * the same condition doesn't freeze RCU. 2656b22ce278STejun Heo */ 2657a7e6425eSPaul E. McKenney cond_resched(); 2658b22ce278STejun Heo 2659a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2660a62428c0STejun Heo 2661616db877STejun Heo /* 2662616db877STejun Heo * In addition to %WQ_CPU_INTENSIVE, @worker may also have been marked 2663616db877STejun Heo * CPU intensive by wq_worker_tick() if @work hogged CPU longer than 2664616db877STejun Heo * wq_cpu_intensive_thresh_us. Clear it. 2665616db877STejun Heo */ 2666fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2667fb0e7bebSTejun Heo 26681b69ac6bSJohannes Weiner /* tag the worker for identification in schedule() */ 26691b69ac6bSJohannes Weiner worker->last_func = worker->current_func; 26701b69ac6bSJohannes Weiner 2671a62428c0STejun Heo /* we're done with it, release */ 267242f8570fSSasha Levin hash_del(&worker->hentry); 2673c34056a3STejun Heo worker->current_work = NULL; 2674a2c1c57bSTejun Heo worker->current_func = NULL; 2675112202d9STejun Heo worker->current_pwq = NULL; 2676d812796eSLai Jiangshan worker->current_color = INT_MAX; 2677c4560c2cSLai Jiangshan pwq_dec_nr_in_flight(pwq, work_data); 26781da177e4SLinus Torvalds } 26791da177e4SLinus Torvalds 2680affee4b2STejun Heo /** 2681affee4b2STejun Heo * process_scheduled_works - process scheduled works 2682affee4b2STejun Heo * @worker: self 2683affee4b2STejun Heo * 2684affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2685affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2686affee4b2STejun Heo * fetches a work from the top and executes it. 2687affee4b2STejun Heo * 2688affee4b2STejun Heo * CONTEXT: 2689a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock) which may be released and regrabbed 2690affee4b2STejun Heo * multiple times. 2691affee4b2STejun Heo */ 2692affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 26931da177e4SLinus Torvalds { 2694c0ab017dSTejun Heo struct work_struct *work; 2695c0ab017dSTejun Heo bool first = true; 2696c0ab017dSTejun Heo 2697c0ab017dSTejun Heo while ((work = list_first_entry_or_null(&worker->scheduled, 2698c0ab017dSTejun Heo struct work_struct, entry))) { 2699c0ab017dSTejun Heo if (first) { 2700c0ab017dSTejun Heo worker->pool->watchdog_ts = jiffies; 2701c0ab017dSTejun Heo first = false; 2702c0ab017dSTejun Heo } 2703c34056a3STejun Heo process_one_work(worker, work); 2704a62428c0STejun Heo } 27051da177e4SLinus Torvalds } 27061da177e4SLinus Torvalds 2707197f6accSTejun Heo static void set_pf_worker(bool val) 2708197f6accSTejun Heo { 2709197f6accSTejun Heo mutex_lock(&wq_pool_attach_mutex); 2710197f6accSTejun Heo if (val) 2711197f6accSTejun Heo current->flags |= PF_WQ_WORKER; 2712197f6accSTejun Heo else 2713197f6accSTejun Heo current->flags &= ~PF_WQ_WORKER; 2714197f6accSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 2715197f6accSTejun Heo } 2716197f6accSTejun Heo 27174690c4abSTejun Heo /** 27184690c4abSTejun Heo * worker_thread - the worker thread function 2719c34056a3STejun Heo * @__worker: self 27204690c4abSTejun Heo * 2721c5aa87bbSTejun Heo * The worker thread function. All workers belong to a worker_pool - 2722c5aa87bbSTejun Heo * either a per-cpu one or dynamic unbound one. These workers process all 2723c5aa87bbSTejun Heo * work items regardless of their specific target workqueue. The only 2724c5aa87bbSTejun Heo * exception is work items which belong to workqueues with a rescuer which 2725c5aa87bbSTejun Heo * will be explained in rescuer_thread(). 2726d185af30SYacine Belkadi * 2727d185af30SYacine Belkadi * Return: 0 27284690c4abSTejun Heo */ 2729c34056a3STejun Heo static int worker_thread(void *__worker) 27301da177e4SLinus Torvalds { 2731c34056a3STejun Heo struct worker *worker = __worker; 2732bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 27331da177e4SLinus Torvalds 2734e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2735197f6accSTejun Heo set_pf_worker(true); 2736c8e55f36STejun Heo woke_up: 2737a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2738affee4b2STejun Heo 2739a9ab775bSTejun Heo /* am I supposed to die? */ 2740a9ab775bSTejun Heo if (unlikely(worker->flags & WORKER_DIE)) { 2741a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 2742197f6accSTejun Heo set_pf_worker(false); 274360f5a4bcSLai Jiangshan 274460f5a4bcSLai Jiangshan set_task_comm(worker->task, "kworker/dying"); 2745e441b56fSZhen Lei ida_free(&pool->worker_ida, worker->id); 2746a2d812a2STejun Heo worker_detach_from_pool(worker); 2747e02b9312SValentin Schneider WARN_ON_ONCE(!list_empty(&worker->entry)); 274860f5a4bcSLai Jiangshan kfree(worker); 2749c8e55f36STejun Heo return 0; 2750c8e55f36STejun Heo } 2751c8e55f36STejun Heo 2752c8e55f36STejun Heo worker_leave_idle(worker); 2753db7bccf4STejun Heo recheck: 2754e22bee78STejun Heo /* no more worker necessary? */ 275563d95a91STejun Heo if (!need_more_worker(pool)) 2756e22bee78STejun Heo goto sleep; 2757e22bee78STejun Heo 2758e22bee78STejun Heo /* do we need to manage? */ 275963d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2760e22bee78STejun Heo goto recheck; 2761e22bee78STejun Heo 2762c8e55f36STejun Heo /* 2763c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2764c8e55f36STejun Heo * preparing to process a work or actually processing it. 2765c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2766c8e55f36STejun Heo */ 27676183c009STejun Heo WARN_ON_ONCE(!list_empty(&worker->scheduled)); 2768c8e55f36STejun Heo 2769e22bee78STejun Heo /* 2770a9ab775bSTejun Heo * Finish PREP stage. We're guaranteed to have at least one idle 2771a9ab775bSTejun Heo * worker or that someone else has already assumed the manager 2772a9ab775bSTejun Heo * role. This is where @worker starts participating in concurrency 2773a9ab775bSTejun Heo * management if applicable and concurrency management is restored 2774a9ab775bSTejun Heo * after being rebound. See rebind_workers() for details. 2775e22bee78STejun Heo */ 2776a9ab775bSTejun Heo worker_clr_flags(worker, WORKER_PREP | WORKER_REBOUND); 2777e22bee78STejun Heo 2778e22bee78STejun Heo do { 2779affee4b2STejun Heo struct work_struct *work = 2780bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2781affee4b2STejun Heo struct work_struct, entry); 2782affee4b2STejun Heo 2783873eaca6STejun Heo if (assign_work(work, worker, NULL)) 2784affee4b2STejun Heo process_scheduled_works(worker); 278563d95a91STejun Heo } while (keep_working(pool)); 2786affee4b2STejun Heo 2787228f1d00SLai Jiangshan worker_set_flags(worker, WORKER_PREP); 2788d313dd85STejun Heo sleep: 2789c8e55f36STejun Heo /* 2790d565ed63STejun Heo * pool->lock is held and there's no work to process and no need to 2791d565ed63STejun Heo * manage, sleep. Workers are woken up only while holding 2792d565ed63STejun Heo * pool->lock or from local cpu, so setting the current state 2793d565ed63STejun Heo * before releasing pool->lock is enough to prevent losing any 2794d565ed63STejun Heo * event. 2795c8e55f36STejun Heo */ 2796c8e55f36STejun Heo worker_enter_idle(worker); 2797c5a94a61SPeter Zijlstra __set_current_state(TASK_IDLE); 2798a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 27991da177e4SLinus Torvalds schedule(); 2800c8e55f36STejun Heo goto woke_up; 28011da177e4SLinus Torvalds } 28021da177e4SLinus Torvalds 2803e22bee78STejun Heo /** 2804e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2805111c225aSTejun Heo * @__rescuer: self 2806e22bee78STejun Heo * 2807e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2808493008a8STejun Heo * workqueue which has WQ_MEM_RECLAIM set. 2809e22bee78STejun Heo * 2810706026c2STejun Heo * Regular work processing on a pool may block trying to create a new 2811e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2812e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2813e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2814e22bee78STejun Heo * the problem rescuer solves. 2815e22bee78STejun Heo * 2816706026c2STejun Heo * When such condition is possible, the pool summons rescuers of all 2817706026c2STejun Heo * workqueues which have works queued on the pool and let them process 2818e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2819e22bee78STejun Heo * 2820e22bee78STejun Heo * This should happen rarely. 2821d185af30SYacine Belkadi * 2822d185af30SYacine Belkadi * Return: 0 2823e22bee78STejun Heo */ 2824111c225aSTejun Heo static int rescuer_thread(void *__rescuer) 2825e22bee78STejun Heo { 2826111c225aSTejun Heo struct worker *rescuer = __rescuer; 2827111c225aSTejun Heo struct workqueue_struct *wq = rescuer->rescue_wq; 28284d595b86SLai Jiangshan bool should_stop; 2829e22bee78STejun Heo 2830e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2831111c225aSTejun Heo 2832111c225aSTejun Heo /* 2833111c225aSTejun Heo * Mark rescuer as worker too. As WORKER_PREP is never cleared, it 2834111c225aSTejun Heo * doesn't participate in concurrency management. 2835111c225aSTejun Heo */ 2836197f6accSTejun Heo set_pf_worker(true); 2837e22bee78STejun Heo repeat: 2838c5a94a61SPeter Zijlstra set_current_state(TASK_IDLE); 28391da177e4SLinus Torvalds 28404d595b86SLai Jiangshan /* 28414d595b86SLai Jiangshan * By the time the rescuer is requested to stop, the workqueue 28424d595b86SLai Jiangshan * shouldn't have any work pending, but @wq->maydays may still have 28434d595b86SLai Jiangshan * pwq(s) queued. This can happen by non-rescuer workers consuming 28444d595b86SLai Jiangshan * all the work items before the rescuer got to them. Go through 28454d595b86SLai Jiangshan * @wq->maydays processing before acting on should_stop so that the 28464d595b86SLai Jiangshan * list is always empty on exit. 28474d595b86SLai Jiangshan */ 28484d595b86SLai Jiangshan should_stop = kthread_should_stop(); 28491da177e4SLinus Torvalds 2850493a1724STejun Heo /* see whether any pwq is asking for help */ 2851a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 2852493a1724STejun Heo 2853493a1724STejun Heo while (!list_empty(&wq->maydays)) { 2854493a1724STejun Heo struct pool_workqueue *pwq = list_first_entry(&wq->maydays, 2855493a1724STejun Heo struct pool_workqueue, mayday_node); 2856112202d9STejun Heo struct worker_pool *pool = pwq->pool; 2857e22bee78STejun Heo struct work_struct *work, *n; 2858e22bee78STejun Heo 2859e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2860493a1724STejun Heo list_del_init(&pwq->mayday_node); 2861493a1724STejun Heo 2862a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2863e22bee78STejun Heo 286451697d39SLai Jiangshan worker_attach_to_pool(rescuer, pool); 286551697d39SLai Jiangshan 2866a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 2867e22bee78STejun Heo 2868e22bee78STejun Heo /* 2869e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2870e22bee78STejun Heo * process'em. 2871e22bee78STejun Heo */ 2872873eaca6STejun Heo WARN_ON_ONCE(!list_empty(&rescuer->scheduled)); 287382607adcSTejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) { 2874873eaca6STejun Heo if (get_work_pwq(work) == pwq && 2875873eaca6STejun Heo assign_work(work, rescuer, &n)) 2876725e8ec5STejun Heo pwq->stats[PWQ_STAT_RESCUED]++; 287782607adcSTejun Heo } 2878e22bee78STejun Heo 2879873eaca6STejun Heo if (!list_empty(&rescuer->scheduled)) { 2880e22bee78STejun Heo process_scheduled_works(rescuer); 28817576958aSTejun Heo 28827576958aSTejun Heo /* 2883008847f6SNeilBrown * The above execution of rescued work items could 2884008847f6SNeilBrown * have created more to rescue through 2885f97a4a1aSLai Jiangshan * pwq_activate_first_inactive() or chained 2886008847f6SNeilBrown * queueing. Let's put @pwq back on mayday list so 2887008847f6SNeilBrown * that such back-to-back work items, which may be 2888008847f6SNeilBrown * being used to relieve memory pressure, don't 2889008847f6SNeilBrown * incur MAYDAY_INTERVAL delay inbetween. 2890008847f6SNeilBrown */ 28914f3f4cf3SLai Jiangshan if (pwq->nr_active && need_to_create_worker(pool)) { 2892a9b8a985SSebastian Andrzej Siewior raw_spin_lock(&wq_mayday_lock); 2893e66b39afSTejun Heo /* 2894e66b39afSTejun Heo * Queue iff we aren't racing destruction 2895e66b39afSTejun Heo * and somebody else hasn't queued it already. 2896e66b39afSTejun Heo */ 2897e66b39afSTejun Heo if (wq->rescuer && list_empty(&pwq->mayday_node)) { 2898008847f6SNeilBrown get_pwq(pwq); 2899e66b39afSTejun Heo list_add_tail(&pwq->mayday_node, &wq->maydays); 2900e66b39afSTejun Heo } 2901a9b8a985SSebastian Andrzej Siewior raw_spin_unlock(&wq_mayday_lock); 2902008847f6SNeilBrown } 2903008847f6SNeilBrown } 2904008847f6SNeilBrown 2905008847f6SNeilBrown /* 290677668c8bSLai Jiangshan * Put the reference grabbed by send_mayday(). @pool won't 290713b1d625SLai Jiangshan * go away while we're still attached to it. 290877668c8bSLai Jiangshan */ 290977668c8bSLai Jiangshan put_pwq(pwq); 291077668c8bSLai Jiangshan 291177668c8bSLai Jiangshan /* 29120219a352STejun Heo * Leave this pool. Notify regular workers; otherwise, we end up 29130219a352STejun Heo * with 0 concurrency and stalling the execution. 29147576958aSTejun Heo */ 29150219a352STejun Heo kick_pool(pool); 29167576958aSTejun Heo 2917a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 291813b1d625SLai Jiangshan 2919a2d812a2STejun Heo worker_detach_from_pool(rescuer); 292013b1d625SLai Jiangshan 2921a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 29221da177e4SLinus Torvalds } 29231da177e4SLinus Torvalds 2924a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 2925493a1724STejun Heo 29264d595b86SLai Jiangshan if (should_stop) { 29274d595b86SLai Jiangshan __set_current_state(TASK_RUNNING); 2928197f6accSTejun Heo set_pf_worker(false); 29294d595b86SLai Jiangshan return 0; 29304d595b86SLai Jiangshan } 29314d595b86SLai Jiangshan 2932111c225aSTejun Heo /* rescuers should never participate in concurrency management */ 2933111c225aSTejun Heo WARN_ON_ONCE(!(rescuer->flags & WORKER_NOT_RUNNING)); 2934e22bee78STejun Heo schedule(); 2935e22bee78STejun Heo goto repeat; 29361da177e4SLinus Torvalds } 29371da177e4SLinus Torvalds 2938fca839c0STejun Heo /** 2939fca839c0STejun Heo * check_flush_dependency - check for flush dependency sanity 2940fca839c0STejun Heo * @target_wq: workqueue being flushed 2941fca839c0STejun Heo * @target_work: work item being flushed (NULL for workqueue flushes) 2942fca839c0STejun Heo * 2943fca839c0STejun Heo * %current is trying to flush the whole @target_wq or @target_work on it. 2944fca839c0STejun Heo * If @target_wq doesn't have %WQ_MEM_RECLAIM, verify that %current is not 2945fca839c0STejun Heo * reclaiming memory or running on a workqueue which doesn't have 2946fca839c0STejun Heo * %WQ_MEM_RECLAIM as that can break forward-progress guarantee leading to 2947fca839c0STejun Heo * a deadlock. 2948fca839c0STejun Heo */ 2949fca839c0STejun Heo static void check_flush_dependency(struct workqueue_struct *target_wq, 2950fca839c0STejun Heo struct work_struct *target_work) 2951fca839c0STejun Heo { 2952fca839c0STejun Heo work_func_t target_func = target_work ? target_work->func : NULL; 2953fca839c0STejun Heo struct worker *worker; 2954fca839c0STejun Heo 2955fca839c0STejun Heo if (target_wq->flags & WQ_MEM_RECLAIM) 2956fca839c0STejun Heo return; 2957fca839c0STejun Heo 2958fca839c0STejun Heo worker = current_wq_worker(); 2959fca839c0STejun Heo 2960fca839c0STejun Heo WARN_ONCE(current->flags & PF_MEMALLOC, 2961d75f773cSSakari Ailus "workqueue: PF_MEMALLOC task %d(%s) is flushing !WQ_MEM_RECLAIM %s:%ps", 2962fca839c0STejun Heo current->pid, current->comm, target_wq->name, target_func); 296323d11a58STejun Heo WARN_ONCE(worker && ((worker->current_pwq->wq->flags & 296423d11a58STejun Heo (WQ_MEM_RECLAIM | __WQ_LEGACY)) == WQ_MEM_RECLAIM), 2965d75f773cSSakari Ailus "workqueue: WQ_MEM_RECLAIM %s:%ps is flushing !WQ_MEM_RECLAIM %s:%ps", 2966fca839c0STejun Heo worker->current_pwq->wq->name, worker->current_func, 2967fca839c0STejun Heo target_wq->name, target_func); 2968fca839c0STejun Heo } 2969fca839c0STejun Heo 2970fc2e4d70SOleg Nesterov struct wq_barrier { 2971fc2e4d70SOleg Nesterov struct work_struct work; 2972fc2e4d70SOleg Nesterov struct completion done; 29732607d7a6STejun Heo struct task_struct *task; /* purely informational */ 2974fc2e4d70SOleg Nesterov }; 2975fc2e4d70SOleg Nesterov 2976fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2977fc2e4d70SOleg Nesterov { 2978fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2979fc2e4d70SOleg Nesterov complete(&barr->done); 2980fc2e4d70SOleg Nesterov } 2981fc2e4d70SOleg Nesterov 29824690c4abSTejun Heo /** 29834690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 2984112202d9STejun Heo * @pwq: pwq to insert barrier into 29854690c4abSTejun Heo * @barr: wq_barrier to insert 2986affee4b2STejun Heo * @target: target work to attach @barr to 2987affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 29884690c4abSTejun Heo * 2989affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2990affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2991affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2992affee4b2STejun Heo * cpu. 2993affee4b2STejun Heo * 2994affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2995affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2996affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2997affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2998affee4b2STejun Heo * after a work with LINKED flag set. 2999affee4b2STejun Heo * 3000affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 3001112202d9STejun Heo * underneath us, so we can't reliably determine pwq from @target. 30024690c4abSTejun Heo * 30034690c4abSTejun Heo * CONTEXT: 3004a9b8a985SSebastian Andrzej Siewior * raw_spin_lock_irq(pool->lock). 30054690c4abSTejun Heo */ 3006112202d9STejun Heo static void insert_wq_barrier(struct pool_workqueue *pwq, 3007affee4b2STejun Heo struct wq_barrier *barr, 3008affee4b2STejun Heo struct work_struct *target, struct worker *worker) 3009fc2e4d70SOleg Nesterov { 3010d812796eSLai Jiangshan unsigned int work_flags = 0; 3011d812796eSLai Jiangshan unsigned int work_color; 3012affee4b2STejun Heo struct list_head *head; 3013affee4b2STejun Heo 3014dc186ad7SThomas Gleixner /* 3015d565ed63STejun Heo * debugobject calls are safe here even with pool->lock locked 3016dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 3017dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 3018dc186ad7SThomas Gleixner * might deadlock. 3019dc186ad7SThomas Gleixner */ 3020ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 302122df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 302252fa5bc5SBoqun Feng 3023fd1a5b04SByungchul Park init_completion_map(&barr->done, &target->lockdep_map); 3024fd1a5b04SByungchul Park 30252607d7a6STejun Heo barr->task = current; 302683c22520SOleg Nesterov 3027018f3a13SLai Jiangshan /* The barrier work item does not participate in pwq->nr_active. */ 3028018f3a13SLai Jiangshan work_flags |= WORK_STRUCT_INACTIVE; 3029018f3a13SLai Jiangshan 3030affee4b2STejun Heo /* 3031affee4b2STejun Heo * If @target is currently being executed, schedule the 3032affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 3033affee4b2STejun Heo */ 3034d812796eSLai Jiangshan if (worker) { 3035affee4b2STejun Heo head = worker->scheduled.next; 3036d812796eSLai Jiangshan work_color = worker->current_color; 3037d812796eSLai Jiangshan } else { 3038affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 3039affee4b2STejun Heo 3040affee4b2STejun Heo head = target->entry.next; 3041affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 3042d21cece0SLai Jiangshan work_flags |= *bits & WORK_STRUCT_LINKED; 3043d812796eSLai Jiangshan work_color = get_work_color(*bits); 3044affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 3045affee4b2STejun Heo } 3046affee4b2STejun Heo 3047d812796eSLai Jiangshan pwq->nr_in_flight[work_color]++; 3048d812796eSLai Jiangshan work_flags |= work_color_to_flags(work_color); 3049d812796eSLai Jiangshan 3050d21cece0SLai Jiangshan insert_work(pwq, &barr->work, head, work_flags); 3051fc2e4d70SOleg Nesterov } 3052fc2e4d70SOleg Nesterov 305373f53c4aSTejun Heo /** 3054112202d9STejun Heo * flush_workqueue_prep_pwqs - prepare pwqs for workqueue flushing 305573f53c4aSTejun Heo * @wq: workqueue being flushed 305673f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 305773f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 305873f53c4aSTejun Heo * 3059112202d9STejun Heo * Prepare pwqs for workqueue flushing. 306073f53c4aSTejun Heo * 3061112202d9STejun Heo * If @flush_color is non-negative, flush_color on all pwqs should be 3062112202d9STejun Heo * -1. If no pwq has in-flight commands at the specified color, all 3063112202d9STejun Heo * pwq->flush_color's stay at -1 and %false is returned. If any pwq 3064112202d9STejun Heo * has in flight commands, its pwq->flush_color is set to 3065112202d9STejun Heo * @flush_color, @wq->nr_pwqs_to_flush is updated accordingly, pwq 306673f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 306773f53c4aSTejun Heo * 306873f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 306973f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 307073f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 307173f53c4aSTejun Heo * is returned. 307273f53c4aSTejun Heo * 3073112202d9STejun Heo * If @work_color is non-negative, all pwqs should have the same 307473f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 307573f53c4aSTejun Heo * advanced to @work_color. 307673f53c4aSTejun Heo * 307773f53c4aSTejun Heo * CONTEXT: 30783c25a55dSLai Jiangshan * mutex_lock(wq->mutex). 307973f53c4aSTejun Heo * 3080d185af30SYacine Belkadi * Return: 308173f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 308273f53c4aSTejun Heo * otherwise. 308373f53c4aSTejun Heo */ 3084112202d9STejun Heo static bool flush_workqueue_prep_pwqs(struct workqueue_struct *wq, 308573f53c4aSTejun Heo int flush_color, int work_color) 30861da177e4SLinus Torvalds { 308773f53c4aSTejun Heo bool wait = false; 308849e3cf44STejun Heo struct pool_workqueue *pwq; 30891da177e4SLinus Torvalds 309073f53c4aSTejun Heo if (flush_color >= 0) { 30916183c009STejun Heo WARN_ON_ONCE(atomic_read(&wq->nr_pwqs_to_flush)); 3092112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 1); 3093dc186ad7SThomas Gleixner } 309414441960SOleg Nesterov 309549e3cf44STejun Heo for_each_pwq(pwq, wq) { 3096112202d9STejun Heo struct worker_pool *pool = pwq->pool; 30971da177e4SLinus Torvalds 3098a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 309973f53c4aSTejun Heo 310073f53c4aSTejun Heo if (flush_color >= 0) { 31016183c009STejun Heo WARN_ON_ONCE(pwq->flush_color != -1); 310273f53c4aSTejun Heo 3103112202d9STejun Heo if (pwq->nr_in_flight[flush_color]) { 3104112202d9STejun Heo pwq->flush_color = flush_color; 3105112202d9STejun Heo atomic_inc(&wq->nr_pwqs_to_flush); 310673f53c4aSTejun Heo wait = true; 31071da177e4SLinus Torvalds } 310873f53c4aSTejun Heo } 310973f53c4aSTejun Heo 311073f53c4aSTejun Heo if (work_color >= 0) { 31116183c009STejun Heo WARN_ON_ONCE(work_color != work_next_color(pwq->work_color)); 3112112202d9STejun Heo pwq->work_color = work_color; 311373f53c4aSTejun Heo } 311473f53c4aSTejun Heo 3115a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 31161da177e4SLinus Torvalds } 31171da177e4SLinus Torvalds 3118112202d9STejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_pwqs_to_flush)) 311973f53c4aSTejun Heo complete(&wq->first_flusher->done); 312073f53c4aSTejun Heo 312173f53c4aSTejun Heo return wait; 312283c22520SOleg Nesterov } 31231da177e4SLinus Torvalds 31240fcb78c2SRolf Eike Beer /** 3125c4f135d6STetsuo Handa * __flush_workqueue - ensure that any scheduled work has run to completion. 31260fcb78c2SRolf Eike Beer * @wq: workqueue to flush 31271da177e4SLinus Torvalds * 3128c5aa87bbSTejun Heo * This function sleeps until all work items which were queued on entry 3129c5aa87bbSTejun Heo * have finished execution, but it is not livelocked by new incoming ones. 31301da177e4SLinus Torvalds */ 3131c4f135d6STetsuo Handa void __flush_workqueue(struct workqueue_struct *wq) 31321da177e4SLinus Torvalds { 313373f53c4aSTejun Heo struct wq_flusher this_flusher = { 313473f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 313573f53c4aSTejun Heo .flush_color = -1, 3136fd1a5b04SByungchul Park .done = COMPLETION_INITIALIZER_ONSTACK_MAP(this_flusher.done, wq->lockdep_map), 313773f53c4aSTejun Heo }; 313873f53c4aSTejun Heo int next_color; 3139b1f4ec17SOleg Nesterov 31403347fa09STejun Heo if (WARN_ON(!wq_online)) 31413347fa09STejun Heo return; 31423347fa09STejun Heo 314387915adcSJohannes Berg lock_map_acquire(&wq->lockdep_map); 314487915adcSJohannes Berg lock_map_release(&wq->lockdep_map); 314587915adcSJohannes Berg 31463c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 314773f53c4aSTejun Heo 314873f53c4aSTejun Heo /* 314973f53c4aSTejun Heo * Start-to-wait phase 315073f53c4aSTejun Heo */ 315173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 315273f53c4aSTejun Heo 315373f53c4aSTejun Heo if (next_color != wq->flush_color) { 315473f53c4aSTejun Heo /* 315573f53c4aSTejun Heo * Color space is not full. The current work_color 315673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 315773f53c4aSTejun Heo * by one. 315873f53c4aSTejun Heo */ 31596183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow)); 316073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 316173f53c4aSTejun Heo wq->work_color = next_color; 316273f53c4aSTejun Heo 316373f53c4aSTejun Heo if (!wq->first_flusher) { 316473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 31656183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 316673f53c4aSTejun Heo 316773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 316873f53c4aSTejun Heo 3169112202d9STejun Heo if (!flush_workqueue_prep_pwqs(wq, wq->flush_color, 317073f53c4aSTejun Heo wq->work_color)) { 317173f53c4aSTejun Heo /* nothing to flush, done */ 317273f53c4aSTejun Heo wq->flush_color = next_color; 317373f53c4aSTejun Heo wq->first_flusher = NULL; 317473f53c4aSTejun Heo goto out_unlock; 317573f53c4aSTejun Heo } 317673f53c4aSTejun Heo } else { 317773f53c4aSTejun Heo /* wait in queue */ 31786183c009STejun Heo WARN_ON_ONCE(wq->flush_color == this_flusher.flush_color); 317973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 3180112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 318173f53c4aSTejun Heo } 318273f53c4aSTejun Heo } else { 318373f53c4aSTejun Heo /* 318473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 318573f53c4aSTejun Heo * The next flush completion will assign us 318673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 318773f53c4aSTejun Heo */ 318873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 318973f53c4aSTejun Heo } 319073f53c4aSTejun Heo 3191fca839c0STejun Heo check_flush_dependency(wq, NULL); 3192fca839c0STejun Heo 31933c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 319473f53c4aSTejun Heo 319573f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 319673f53c4aSTejun Heo 319773f53c4aSTejun Heo /* 319873f53c4aSTejun Heo * Wake-up-and-cascade phase 319973f53c4aSTejun Heo * 320073f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 320173f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 320273f53c4aSTejun Heo */ 320300d5d15bSChris Wilson if (READ_ONCE(wq->first_flusher) != &this_flusher) 320473f53c4aSTejun Heo return; 320573f53c4aSTejun Heo 32063c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 320773f53c4aSTejun Heo 32084ce48b37STejun Heo /* we might have raced, check again with mutex held */ 32094ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 32104ce48b37STejun Heo goto out_unlock; 32114ce48b37STejun Heo 321200d5d15bSChris Wilson WRITE_ONCE(wq->first_flusher, NULL); 321373f53c4aSTejun Heo 32146183c009STejun Heo WARN_ON_ONCE(!list_empty(&this_flusher.list)); 32156183c009STejun Heo WARN_ON_ONCE(wq->flush_color != this_flusher.flush_color); 321673f53c4aSTejun Heo 321773f53c4aSTejun Heo while (true) { 321873f53c4aSTejun Heo struct wq_flusher *next, *tmp; 321973f53c4aSTejun Heo 322073f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 322173f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 322273f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 322373f53c4aSTejun Heo break; 322473f53c4aSTejun Heo list_del_init(&next->list); 322573f53c4aSTejun Heo complete(&next->done); 322673f53c4aSTejun Heo } 322773f53c4aSTejun Heo 32286183c009STejun Heo WARN_ON_ONCE(!list_empty(&wq->flusher_overflow) && 322973f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 323073f53c4aSTejun Heo 323173f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 323273f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 323373f53c4aSTejun Heo 323473f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 323573f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 323673f53c4aSTejun Heo /* 323773f53c4aSTejun Heo * Assign the same color to all overflowed 323873f53c4aSTejun Heo * flushers, advance work_color and append to 323973f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 324073f53c4aSTejun Heo * phase for these overflowed flushers. 324173f53c4aSTejun Heo */ 324273f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 324373f53c4aSTejun Heo tmp->flush_color = wq->work_color; 324473f53c4aSTejun Heo 324573f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 324673f53c4aSTejun Heo 324773f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 324873f53c4aSTejun Heo &wq->flusher_queue); 3249112202d9STejun Heo flush_workqueue_prep_pwqs(wq, -1, wq->work_color); 325073f53c4aSTejun Heo } 325173f53c4aSTejun Heo 325273f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 32536183c009STejun Heo WARN_ON_ONCE(wq->flush_color != wq->work_color); 325473f53c4aSTejun Heo break; 325573f53c4aSTejun Heo } 325673f53c4aSTejun Heo 325773f53c4aSTejun Heo /* 325873f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 3259112202d9STejun Heo * the new first flusher and arm pwqs. 326073f53c4aSTejun Heo */ 32616183c009STejun Heo WARN_ON_ONCE(wq->flush_color == wq->work_color); 32626183c009STejun Heo WARN_ON_ONCE(wq->flush_color != next->flush_color); 326373f53c4aSTejun Heo 326473f53c4aSTejun Heo list_del_init(&next->list); 326573f53c4aSTejun Heo wq->first_flusher = next; 326673f53c4aSTejun Heo 3267112202d9STejun Heo if (flush_workqueue_prep_pwqs(wq, wq->flush_color, -1)) 326873f53c4aSTejun Heo break; 326973f53c4aSTejun Heo 327073f53c4aSTejun Heo /* 327173f53c4aSTejun Heo * Meh... this color is already done, clear first 327273f53c4aSTejun Heo * flusher and repeat cascading. 327373f53c4aSTejun Heo */ 327473f53c4aSTejun Heo wq->first_flusher = NULL; 327573f53c4aSTejun Heo } 327673f53c4aSTejun Heo 327773f53c4aSTejun Heo out_unlock: 32783c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 32791da177e4SLinus Torvalds } 3280c4f135d6STetsuo Handa EXPORT_SYMBOL(__flush_workqueue); 32811da177e4SLinus Torvalds 32829c5a2ba7STejun Heo /** 32839c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 32849c5a2ba7STejun Heo * @wq: workqueue to drain 32859c5a2ba7STejun Heo * 32869c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 32879c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 32889c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 3289b749b1b6SChen Hanxiao * repeatedly until it becomes empty. The number of flushing is determined 32909c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 32919c5a2ba7STejun Heo * takes too long. 32929c5a2ba7STejun Heo */ 32939c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 32949c5a2ba7STejun Heo { 32959c5a2ba7STejun Heo unsigned int flush_cnt = 0; 329649e3cf44STejun Heo struct pool_workqueue *pwq; 32979c5a2ba7STejun Heo 32989c5a2ba7STejun Heo /* 32999c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 33009c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 3301618b01ebSTejun Heo * Use __WQ_DRAINING so that queue doesn't have to check nr_drainers. 33029c5a2ba7STejun Heo */ 330387fc741eSLai Jiangshan mutex_lock(&wq->mutex); 33049c5a2ba7STejun Heo if (!wq->nr_drainers++) 3305618b01ebSTejun Heo wq->flags |= __WQ_DRAINING; 330687fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33079c5a2ba7STejun Heo reflush: 3308c4f135d6STetsuo Handa __flush_workqueue(wq); 33099c5a2ba7STejun Heo 3310b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 331176af4d93STejun Heo 331249e3cf44STejun Heo for_each_pwq(pwq, wq) { 3313fa2563e4SThomas Tuttle bool drained; 33149c5a2ba7STejun Heo 3315a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 3316f97a4a1aSLai Jiangshan drained = !pwq->nr_active && list_empty(&pwq->inactive_works); 3317a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 3318fa2563e4SThomas Tuttle 3319fa2563e4SThomas Tuttle if (drained) 33209c5a2ba7STejun Heo continue; 33219c5a2ba7STejun Heo 33229c5a2ba7STejun Heo if (++flush_cnt == 10 || 33239c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 3324e9ad2eb3SStephen Zhang pr_warn("workqueue %s: %s() isn't complete after %u tries\n", 3325e9ad2eb3SStephen Zhang wq->name, __func__, flush_cnt); 332676af4d93STejun Heo 3327b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 33289c5a2ba7STejun Heo goto reflush; 33299c5a2ba7STejun Heo } 33309c5a2ba7STejun Heo 33319c5a2ba7STejun Heo if (!--wq->nr_drainers) 3332618b01ebSTejun Heo wq->flags &= ~__WQ_DRAINING; 333387fc741eSLai Jiangshan mutex_unlock(&wq->mutex); 33349c5a2ba7STejun Heo } 33359c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 33369c5a2ba7STejun Heo 3337d6e89786SJohannes Berg static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 3338d6e89786SJohannes Berg bool from_cancel) 3339baf59022STejun Heo { 3340baf59022STejun Heo struct worker *worker = NULL; 3341c9e7cf27STejun Heo struct worker_pool *pool; 3342112202d9STejun Heo struct pool_workqueue *pwq; 3343baf59022STejun Heo 3344baf59022STejun Heo might_sleep(); 3345baf59022STejun Heo 334624acfb71SThomas Gleixner rcu_read_lock(); 3347fa1b54e6STejun Heo pool = get_work_pool(work); 3348fa1b54e6STejun Heo if (!pool) { 334924acfb71SThomas Gleixner rcu_read_unlock(); 3350fa1b54e6STejun Heo return false; 3351fa1b54e6STejun Heo } 3352fa1b54e6STejun Heo 3353a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 33540b3dae68SLai Jiangshan /* see the comment in try_to_grab_pending() with the same code */ 3355112202d9STejun Heo pwq = get_work_pwq(work); 3356112202d9STejun Heo if (pwq) { 3357112202d9STejun Heo if (unlikely(pwq->pool != pool)) 3358baf59022STejun Heo goto already_gone; 3359606a5020STejun Heo } else { 3360c9e7cf27STejun Heo worker = find_worker_executing_work(pool, work); 3361baf59022STejun Heo if (!worker) 3362baf59022STejun Heo goto already_gone; 3363112202d9STejun Heo pwq = worker->current_pwq; 3364606a5020STejun Heo } 3365baf59022STejun Heo 3366fca839c0STejun Heo check_flush_dependency(pwq->wq, work); 3367fca839c0STejun Heo 3368112202d9STejun Heo insert_wq_barrier(pwq, barr, work, worker); 3369a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 3370baf59022STejun Heo 3371e159489bSTejun Heo /* 3372a1d14934SPeter Zijlstra * Force a lock recursion deadlock when using flush_work() inside a 3373a1d14934SPeter Zijlstra * single-threaded or rescuer equipped workqueue. 3374a1d14934SPeter Zijlstra * 3375a1d14934SPeter Zijlstra * For single threaded workqueues the deadlock happens when the work 3376a1d14934SPeter Zijlstra * is after the work issuing the flush_work(). For rescuer equipped 3377a1d14934SPeter Zijlstra * workqueues the deadlock happens when the rescuer stalls, blocking 3378a1d14934SPeter Zijlstra * forward progress. 3379e159489bSTejun Heo */ 3380d6e89786SJohannes Berg if (!from_cancel && 3381d6e89786SJohannes Berg (pwq->wq->saved_max_active == 1 || pwq->wq->rescuer)) { 3382112202d9STejun Heo lock_map_acquire(&pwq->wq->lockdep_map); 3383112202d9STejun Heo lock_map_release(&pwq->wq->lockdep_map); 3384a1d14934SPeter Zijlstra } 338524acfb71SThomas Gleixner rcu_read_unlock(); 3386baf59022STejun Heo return true; 3387baf59022STejun Heo already_gone: 3388a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 338924acfb71SThomas Gleixner rcu_read_unlock(); 3390baf59022STejun Heo return false; 3391baf59022STejun Heo } 3392baf59022STejun Heo 3393d6e89786SJohannes Berg static bool __flush_work(struct work_struct *work, bool from_cancel) 3394d6e89786SJohannes Berg { 3395d6e89786SJohannes Berg struct wq_barrier barr; 3396d6e89786SJohannes Berg 3397d6e89786SJohannes Berg if (WARN_ON(!wq_online)) 3398d6e89786SJohannes Berg return false; 3399d6e89786SJohannes Berg 34004d43d395STetsuo Handa if (WARN_ON(!work->func)) 34014d43d395STetsuo Handa return false; 34024d43d395STetsuo Handa 340387915adcSJohannes Berg lock_map_acquire(&work->lockdep_map); 340487915adcSJohannes Berg lock_map_release(&work->lockdep_map); 340587915adcSJohannes Berg 3406d6e89786SJohannes Berg if (start_flush_work(work, &barr, from_cancel)) { 3407d6e89786SJohannes Berg wait_for_completion(&barr.done); 3408d6e89786SJohannes Berg destroy_work_on_stack(&barr.work); 3409d6e89786SJohannes Berg return true; 3410d6e89786SJohannes Berg } else { 3411d6e89786SJohannes Berg return false; 3412d6e89786SJohannes Berg } 3413d6e89786SJohannes Berg } 3414d6e89786SJohannes Berg 3415db700897SOleg Nesterov /** 3416401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 3417401a8d04STejun Heo * @work: the work to flush 3418db700897SOleg Nesterov * 3419606a5020STejun Heo * Wait until @work has finished execution. @work is guaranteed to be idle 3420606a5020STejun Heo * on return if it hasn't been requeued since flush started. 3421401a8d04STejun Heo * 3422d185af30SYacine Belkadi * Return: 3423401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3424401a8d04STejun Heo * %false if it was already idle. 3425db700897SOleg Nesterov */ 3426401a8d04STejun Heo bool flush_work(struct work_struct *work) 3427db700897SOleg Nesterov { 3428d6e89786SJohannes Berg return __flush_work(work, false); 3429606a5020STejun Heo } 3430db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 3431db700897SOleg Nesterov 34328603e1b3STejun Heo struct cwt_wait { 3433ac6424b9SIngo Molnar wait_queue_entry_t wait; 34348603e1b3STejun Heo struct work_struct *work; 34358603e1b3STejun Heo }; 34368603e1b3STejun Heo 3437ac6424b9SIngo Molnar static int cwt_wakefn(wait_queue_entry_t *wait, unsigned mode, int sync, void *key) 34388603e1b3STejun Heo { 34398603e1b3STejun Heo struct cwt_wait *cwait = container_of(wait, struct cwt_wait, wait); 34408603e1b3STejun Heo 34418603e1b3STejun Heo if (cwait->work != key) 34428603e1b3STejun Heo return 0; 34438603e1b3STejun Heo return autoremove_wake_function(wait, mode, sync, key); 34448603e1b3STejun Heo } 34458603e1b3STejun Heo 344636e227d2STejun Heo static bool __cancel_work_timer(struct work_struct *work, bool is_dwork) 3447401a8d04STejun Heo { 34488603e1b3STejun Heo static DECLARE_WAIT_QUEUE_HEAD(cancel_waitq); 3449bbb68dfaSTejun Heo unsigned long flags; 34501f1f642eSOleg Nesterov int ret; 34511f1f642eSOleg Nesterov 34521f1f642eSOleg Nesterov do { 3453bbb68dfaSTejun Heo ret = try_to_grab_pending(work, is_dwork, &flags); 3454bbb68dfaSTejun Heo /* 34558603e1b3STejun Heo * If someone else is already canceling, wait for it to 34568603e1b3STejun Heo * finish. flush_work() doesn't work for PREEMPT_NONE 34578603e1b3STejun Heo * because we may get scheduled between @work's completion 34588603e1b3STejun Heo * and the other canceling task resuming and clearing 34598603e1b3STejun Heo * CANCELING - flush_work() will return false immediately 34608603e1b3STejun Heo * as @work is no longer busy, try_to_grab_pending() will 34618603e1b3STejun Heo * return -ENOENT as @work is still being canceled and the 34628603e1b3STejun Heo * other canceling task won't be able to clear CANCELING as 34638603e1b3STejun Heo * we're hogging the CPU. 34648603e1b3STejun Heo * 34658603e1b3STejun Heo * Let's wait for completion using a waitqueue. As this 34668603e1b3STejun Heo * may lead to the thundering herd problem, use a custom 34678603e1b3STejun Heo * wake function which matches @work along with exclusive 34688603e1b3STejun Heo * wait and wakeup. 3469bbb68dfaSTejun Heo */ 34708603e1b3STejun Heo if (unlikely(ret == -ENOENT)) { 34718603e1b3STejun Heo struct cwt_wait cwait; 34728603e1b3STejun Heo 34738603e1b3STejun Heo init_wait(&cwait.wait); 34748603e1b3STejun Heo cwait.wait.func = cwt_wakefn; 34758603e1b3STejun Heo cwait.work = work; 34768603e1b3STejun Heo 34778603e1b3STejun Heo prepare_to_wait_exclusive(&cancel_waitq, &cwait.wait, 34788603e1b3STejun Heo TASK_UNINTERRUPTIBLE); 34798603e1b3STejun Heo if (work_is_canceling(work)) 34808603e1b3STejun Heo schedule(); 34818603e1b3STejun Heo finish_wait(&cancel_waitq, &cwait.wait); 34828603e1b3STejun Heo } 34831f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 34841f1f642eSOleg Nesterov 3485bbb68dfaSTejun Heo /* tell other tasks trying to grab @work to back off */ 3486bbb68dfaSTejun Heo mark_work_canceling(work); 3487bbb68dfaSTejun Heo local_irq_restore(flags); 3488bbb68dfaSTejun Heo 34893347fa09STejun Heo /* 34903347fa09STejun Heo * This allows canceling during early boot. We know that @work 34913347fa09STejun Heo * isn't executing. 34923347fa09STejun Heo */ 34933347fa09STejun Heo if (wq_online) 3494d6e89786SJohannes Berg __flush_work(work, true); 34953347fa09STejun Heo 34967a22ad75STejun Heo clear_work_data(work); 34978603e1b3STejun Heo 34988603e1b3STejun Heo /* 34998603e1b3STejun Heo * Paired with prepare_to_wait() above so that either 35008603e1b3STejun Heo * waitqueue_active() is visible here or !work_is_canceling() is 35018603e1b3STejun Heo * visible there. 35028603e1b3STejun Heo */ 35038603e1b3STejun Heo smp_mb(); 35048603e1b3STejun Heo if (waitqueue_active(&cancel_waitq)) 35058603e1b3STejun Heo __wake_up(&cancel_waitq, TASK_NORMAL, 1, work); 35068603e1b3STejun Heo 35071f1f642eSOleg Nesterov return ret; 35081f1f642eSOleg Nesterov } 35091f1f642eSOleg Nesterov 35106e84d644SOleg Nesterov /** 3511401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 3512401a8d04STejun Heo * @work: the work to cancel 35136e84d644SOleg Nesterov * 3514401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 3515401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 3516401a8d04STejun Heo * another workqueue. On return from this function, @work is 3517401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 35181f1f642eSOleg Nesterov * 3519401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 3520401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 35216e84d644SOleg Nesterov * 3522401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 35236e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 3524401a8d04STejun Heo * 3525d185af30SYacine Belkadi * Return: 3526401a8d04STejun Heo * %true if @work was pending, %false otherwise. 35276e84d644SOleg Nesterov */ 3528401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 35296e84d644SOleg Nesterov { 353036e227d2STejun Heo return __cancel_work_timer(work, false); 3531b89deed3SOleg Nesterov } 353228e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 3533b89deed3SOleg Nesterov 35346e84d644SOleg Nesterov /** 3535401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 3536401a8d04STejun Heo * @dwork: the delayed work to flush 35376e84d644SOleg Nesterov * 3538401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 3539401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 3540401a8d04STejun Heo * considers the last queueing instance of @dwork. 35411f1f642eSOleg Nesterov * 3542d185af30SYacine Belkadi * Return: 3543401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 3544401a8d04STejun Heo * %false if it was already idle. 35456e84d644SOleg Nesterov */ 3546401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 3547401a8d04STejun Heo { 35488930cabaSTejun Heo local_irq_disable(); 3549401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 355060c057bcSLai Jiangshan __queue_work(dwork->cpu, dwork->wq, &dwork->work); 35518930cabaSTejun Heo local_irq_enable(); 3552401a8d04STejun Heo return flush_work(&dwork->work); 3553401a8d04STejun Heo } 3554401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 3555401a8d04STejun Heo 355605f0fe6bSTejun Heo /** 355705f0fe6bSTejun Heo * flush_rcu_work - wait for a rwork to finish executing the last queueing 355805f0fe6bSTejun Heo * @rwork: the rcu work to flush 355905f0fe6bSTejun Heo * 356005f0fe6bSTejun Heo * Return: 356105f0fe6bSTejun Heo * %true if flush_rcu_work() waited for the work to finish execution, 356205f0fe6bSTejun Heo * %false if it was already idle. 356305f0fe6bSTejun Heo */ 356405f0fe6bSTejun Heo bool flush_rcu_work(struct rcu_work *rwork) 356505f0fe6bSTejun Heo { 356605f0fe6bSTejun Heo if (test_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&rwork->work))) { 356705f0fe6bSTejun Heo rcu_barrier(); 356805f0fe6bSTejun Heo flush_work(&rwork->work); 356905f0fe6bSTejun Heo return true; 357005f0fe6bSTejun Heo } else { 357105f0fe6bSTejun Heo return flush_work(&rwork->work); 357205f0fe6bSTejun Heo } 357305f0fe6bSTejun Heo } 357405f0fe6bSTejun Heo EXPORT_SYMBOL(flush_rcu_work); 357505f0fe6bSTejun Heo 3576f72b8792SJens Axboe static bool __cancel_work(struct work_struct *work, bool is_dwork) 3577f72b8792SJens Axboe { 3578f72b8792SJens Axboe unsigned long flags; 3579f72b8792SJens Axboe int ret; 3580f72b8792SJens Axboe 3581f72b8792SJens Axboe do { 3582f72b8792SJens Axboe ret = try_to_grab_pending(work, is_dwork, &flags); 3583f72b8792SJens Axboe } while (unlikely(ret == -EAGAIN)); 3584f72b8792SJens Axboe 3585f72b8792SJens Axboe if (unlikely(ret < 0)) 3586f72b8792SJens Axboe return false; 3587f72b8792SJens Axboe 3588f72b8792SJens Axboe set_work_pool_and_clear_pending(work, get_work_pool_id(work)); 3589f72b8792SJens Axboe local_irq_restore(flags); 3590f72b8792SJens Axboe return ret; 3591f72b8792SJens Axboe } 3592f72b8792SJens Axboe 359373b4b532SAndrey Grodzovsky /* 359473b4b532SAndrey Grodzovsky * See cancel_delayed_work() 359573b4b532SAndrey Grodzovsky */ 359673b4b532SAndrey Grodzovsky bool cancel_work(struct work_struct *work) 359773b4b532SAndrey Grodzovsky { 359873b4b532SAndrey Grodzovsky return __cancel_work(work, false); 359973b4b532SAndrey Grodzovsky } 360073b4b532SAndrey Grodzovsky EXPORT_SYMBOL(cancel_work); 360173b4b532SAndrey Grodzovsky 3602401a8d04STejun Heo /** 360357b30ae7STejun Heo * cancel_delayed_work - cancel a delayed work 360457b30ae7STejun Heo * @dwork: delayed_work to cancel 360509383498STejun Heo * 3606d185af30SYacine Belkadi * Kill off a pending delayed_work. 3607d185af30SYacine Belkadi * 3608d185af30SYacine Belkadi * Return: %true if @dwork was pending and canceled; %false if it wasn't 3609d185af30SYacine Belkadi * pending. 3610d185af30SYacine Belkadi * 3611d185af30SYacine Belkadi * Note: 3612d185af30SYacine Belkadi * The work callback function may still be running on return, unless 3613d185af30SYacine Belkadi * it returns %true and the work doesn't re-arm itself. Explicitly flush or 3614d185af30SYacine Belkadi * use cancel_delayed_work_sync() to wait on it. 361509383498STejun Heo * 361657b30ae7STejun Heo * This function is safe to call from any context including IRQ handler. 361709383498STejun Heo */ 361857b30ae7STejun Heo bool cancel_delayed_work(struct delayed_work *dwork) 361909383498STejun Heo { 3620f72b8792SJens Axboe return __cancel_work(&dwork->work, true); 362109383498STejun Heo } 362257b30ae7STejun Heo EXPORT_SYMBOL(cancel_delayed_work); 362309383498STejun Heo 362409383498STejun Heo /** 3625401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 3626401a8d04STejun Heo * @dwork: the delayed work cancel 3627401a8d04STejun Heo * 3628401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 3629401a8d04STejun Heo * 3630d185af30SYacine Belkadi * Return: 3631401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 3632401a8d04STejun Heo */ 3633401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 36346e84d644SOleg Nesterov { 363536e227d2STejun Heo return __cancel_work_timer(&dwork->work, true); 36366e84d644SOleg Nesterov } 3637f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 36381da177e4SLinus Torvalds 36390fcb78c2SRolf Eike Beer /** 364031ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 3641b6136773SAndrew Morton * @func: the function to call 3642b6136773SAndrew Morton * 364331ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 364431ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 3645b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 364631ddd871STejun Heo * 3647d185af30SYacine Belkadi * Return: 364831ddd871STejun Heo * 0 on success, -errno on failure. 3649b6136773SAndrew Morton */ 365065f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 365115316ba8SChristoph Lameter { 365215316ba8SChristoph Lameter int cpu; 365338f51568SNamhyung Kim struct work_struct __percpu *works; 365415316ba8SChristoph Lameter 3655b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 3656b6136773SAndrew Morton if (!works) 365715316ba8SChristoph Lameter return -ENOMEM; 3658b6136773SAndrew Morton 3659ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 366093981800STejun Heo 366115316ba8SChristoph Lameter for_each_online_cpu(cpu) { 36629bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 36639bfb1839SIngo Molnar 36649bfb1839SIngo Molnar INIT_WORK(work, func); 36658de6d308SOleg Nesterov schedule_work_on(cpu, work); 366615316ba8SChristoph Lameter } 366793981800STejun Heo 366893981800STejun Heo for_each_online_cpu(cpu) 36698616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 367093981800STejun Heo 3671ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 3672b6136773SAndrew Morton free_percpu(works); 367315316ba8SChristoph Lameter return 0; 367415316ba8SChristoph Lameter } 367515316ba8SChristoph Lameter 3676eef6a7d5SAlan Stern /** 36771fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 36781fa44ecaSJames Bottomley * @fn: the function to execute 36791fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 36801fa44ecaSJames Bottomley * be available when the work executes) 36811fa44ecaSJames Bottomley * 36821fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 36831fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 36841fa44ecaSJames Bottomley * 3685d185af30SYacine Belkadi * Return: 0 - function was executed 36861fa44ecaSJames Bottomley * 1 - function was scheduled for execution 36871fa44ecaSJames Bottomley */ 368865f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 36891fa44ecaSJames Bottomley { 36901fa44ecaSJames Bottomley if (!in_interrupt()) { 369165f27f38SDavid Howells fn(&ew->work); 36921fa44ecaSJames Bottomley return 0; 36931fa44ecaSJames Bottomley } 36941fa44ecaSJames Bottomley 369565f27f38SDavid Howells INIT_WORK(&ew->work, fn); 36961fa44ecaSJames Bottomley schedule_work(&ew->work); 36971fa44ecaSJames Bottomley 36981fa44ecaSJames Bottomley return 1; 36991fa44ecaSJames Bottomley } 37001fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 37011fa44ecaSJames Bottomley 37027a4e344cSTejun Heo /** 37037a4e344cSTejun Heo * free_workqueue_attrs - free a workqueue_attrs 37047a4e344cSTejun Heo * @attrs: workqueue_attrs to free 37057a4e344cSTejun Heo * 37067a4e344cSTejun Heo * Undo alloc_workqueue_attrs(). 37077a4e344cSTejun Heo */ 3708513c98d0SDaniel Jordan void free_workqueue_attrs(struct workqueue_attrs *attrs) 37097a4e344cSTejun Heo { 37107a4e344cSTejun Heo if (attrs) { 37117a4e344cSTejun Heo free_cpumask_var(attrs->cpumask); 37129546b29eSTejun Heo free_cpumask_var(attrs->__pod_cpumask); 37137a4e344cSTejun Heo kfree(attrs); 37147a4e344cSTejun Heo } 37157a4e344cSTejun Heo } 37167a4e344cSTejun Heo 37177a4e344cSTejun Heo /** 37187a4e344cSTejun Heo * alloc_workqueue_attrs - allocate a workqueue_attrs 37197a4e344cSTejun Heo * 37207a4e344cSTejun Heo * Allocate a new workqueue_attrs, initialize with default settings and 3721d185af30SYacine Belkadi * return it. 3722d185af30SYacine Belkadi * 3723d185af30SYacine Belkadi * Return: The allocated new workqueue_attr on success. %NULL on failure. 37247a4e344cSTejun Heo */ 3725513c98d0SDaniel Jordan struct workqueue_attrs *alloc_workqueue_attrs(void) 37267a4e344cSTejun Heo { 37277a4e344cSTejun Heo struct workqueue_attrs *attrs; 37287a4e344cSTejun Heo 3729be69d00dSThomas Gleixner attrs = kzalloc(sizeof(*attrs), GFP_KERNEL); 37307a4e344cSTejun Heo if (!attrs) 37317a4e344cSTejun Heo goto fail; 3732be69d00dSThomas Gleixner if (!alloc_cpumask_var(&attrs->cpumask, GFP_KERNEL)) 37337a4e344cSTejun Heo goto fail; 37349546b29eSTejun Heo if (!alloc_cpumask_var(&attrs->__pod_cpumask, GFP_KERNEL)) 37359546b29eSTejun Heo goto fail; 37367a4e344cSTejun Heo 373713e2e556STejun Heo cpumask_copy(attrs->cpumask, cpu_possible_mask); 3738*523a301eSTejun Heo attrs->affn_scope = WQ_AFFN_DFL; 37397a4e344cSTejun Heo return attrs; 37407a4e344cSTejun Heo fail: 37417a4e344cSTejun Heo free_workqueue_attrs(attrs); 37427a4e344cSTejun Heo return NULL; 37437a4e344cSTejun Heo } 37447a4e344cSTejun Heo 374529c91e99STejun Heo static void copy_workqueue_attrs(struct workqueue_attrs *to, 374629c91e99STejun Heo const struct workqueue_attrs *from) 374729c91e99STejun Heo { 374829c91e99STejun Heo to->nice = from->nice; 374929c91e99STejun Heo cpumask_copy(to->cpumask, from->cpumask); 37509546b29eSTejun Heo cpumask_copy(to->__pod_cpumask, from->__pod_cpumask); 37518639ecebSTejun Heo to->affn_strict = from->affn_strict; 375284193c07STejun Heo 37532865a8fbSShaohua Li /* 375484193c07STejun Heo * Unlike hash and equality test, copying shouldn't ignore wq-only 375584193c07STejun Heo * fields as copying is used for both pool and wq attrs. Instead, 375684193c07STejun Heo * get_unbound_pool() explicitly clears the fields. 37572865a8fbSShaohua Li */ 375884193c07STejun Heo to->affn_scope = from->affn_scope; 3759af73f5c9STejun Heo to->ordered = from->ordered; 376029c91e99STejun Heo } 376129c91e99STejun Heo 37625de7a03cSTejun Heo /* 37635de7a03cSTejun Heo * Some attrs fields are workqueue-only. Clear them for worker_pool's. See the 37645de7a03cSTejun Heo * comments in 'struct workqueue_attrs' definition. 37655de7a03cSTejun Heo */ 37665de7a03cSTejun Heo static void wqattrs_clear_for_pool(struct workqueue_attrs *attrs) 37675de7a03cSTejun Heo { 376884193c07STejun Heo attrs->affn_scope = WQ_AFFN_NR_TYPES; 37695de7a03cSTejun Heo attrs->ordered = false; 37705de7a03cSTejun Heo } 37715de7a03cSTejun Heo 377229c91e99STejun Heo /* hash value of the content of @attr */ 377329c91e99STejun Heo static u32 wqattrs_hash(const struct workqueue_attrs *attrs) 377429c91e99STejun Heo { 377529c91e99STejun Heo u32 hash = 0; 377629c91e99STejun Heo 377729c91e99STejun Heo hash = jhash_1word(attrs->nice, hash); 377813e2e556STejun Heo hash = jhash(cpumask_bits(attrs->cpumask), 377913e2e556STejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 37809546b29eSTejun Heo hash = jhash(cpumask_bits(attrs->__pod_cpumask), 37819546b29eSTejun Heo BITS_TO_LONGS(nr_cpumask_bits) * sizeof(long), hash); 37828639ecebSTejun Heo hash = jhash_1word(attrs->affn_strict, hash); 378329c91e99STejun Heo return hash; 378429c91e99STejun Heo } 378529c91e99STejun Heo 378629c91e99STejun Heo /* content equality test */ 378729c91e99STejun Heo static bool wqattrs_equal(const struct workqueue_attrs *a, 378829c91e99STejun Heo const struct workqueue_attrs *b) 378929c91e99STejun Heo { 379029c91e99STejun Heo if (a->nice != b->nice) 379129c91e99STejun Heo return false; 379229c91e99STejun Heo if (!cpumask_equal(a->cpumask, b->cpumask)) 379329c91e99STejun Heo return false; 37949546b29eSTejun Heo if (!cpumask_equal(a->__pod_cpumask, b->__pod_cpumask)) 37959546b29eSTejun Heo return false; 37968639ecebSTejun Heo if (a->affn_strict != b->affn_strict) 37978639ecebSTejun Heo return false; 379829c91e99STejun Heo return true; 379929c91e99STejun Heo } 380029c91e99STejun Heo 38010f36ee24STejun Heo /* Update @attrs with actually available CPUs */ 38020f36ee24STejun Heo static void wqattrs_actualize_cpumask(struct workqueue_attrs *attrs, 38030f36ee24STejun Heo const cpumask_t *unbound_cpumask) 38040f36ee24STejun Heo { 38050f36ee24STejun Heo /* 38060f36ee24STejun Heo * Calculate the effective CPU mask of @attrs given @unbound_cpumask. If 38070f36ee24STejun Heo * @attrs->cpumask doesn't overlap with @unbound_cpumask, we fallback to 38080f36ee24STejun Heo * @unbound_cpumask. 38090f36ee24STejun Heo */ 38100f36ee24STejun Heo cpumask_and(attrs->cpumask, attrs->cpumask, unbound_cpumask); 38110f36ee24STejun Heo if (unlikely(cpumask_empty(attrs->cpumask))) 38120f36ee24STejun Heo cpumask_copy(attrs->cpumask, unbound_cpumask); 38130f36ee24STejun Heo } 38140f36ee24STejun Heo 381584193c07STejun Heo /* find wq_pod_type to use for @attrs */ 381684193c07STejun Heo static const struct wq_pod_type * 381784193c07STejun Heo wqattrs_pod_type(const struct workqueue_attrs *attrs) 381884193c07STejun Heo { 3819*523a301eSTejun Heo enum wq_affn_scope scope; 3820*523a301eSTejun Heo struct wq_pod_type *pt; 3821*523a301eSTejun Heo 3822*523a301eSTejun Heo /* to synchronize access to wq_affn_dfl */ 3823*523a301eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3824*523a301eSTejun Heo 3825*523a301eSTejun Heo if (attrs->affn_scope == WQ_AFFN_DFL) 3826*523a301eSTejun Heo scope = wq_affn_dfl; 3827*523a301eSTejun Heo else 3828*523a301eSTejun Heo scope = attrs->affn_scope; 3829*523a301eSTejun Heo 3830*523a301eSTejun Heo pt = &wq_pod_types[scope]; 383184193c07STejun Heo 383284193c07STejun Heo if (!WARN_ON_ONCE(attrs->affn_scope == WQ_AFFN_NR_TYPES) && 383384193c07STejun Heo likely(pt->nr_pods)) 383484193c07STejun Heo return pt; 383584193c07STejun Heo 383684193c07STejun Heo /* 383784193c07STejun Heo * Before workqueue_init_topology(), only SYSTEM is available which is 383884193c07STejun Heo * initialized in workqueue_init_early(). 383984193c07STejun Heo */ 384084193c07STejun Heo pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 384184193c07STejun Heo BUG_ON(!pt->nr_pods); 384284193c07STejun Heo return pt; 384384193c07STejun Heo } 384484193c07STejun Heo 38457a4e344cSTejun Heo /** 38467a4e344cSTejun Heo * init_worker_pool - initialize a newly zalloc'd worker_pool 38477a4e344cSTejun Heo * @pool: worker_pool to initialize 38487a4e344cSTejun Heo * 3849402dd89dSShailendra Verma * Initialize a newly zalloc'd @pool. It also allocates @pool->attrs. 3850d185af30SYacine Belkadi * 3851d185af30SYacine Belkadi * Return: 0 on success, -errno on failure. Even on failure, all fields 385229c91e99STejun Heo * inside @pool proper are initialized and put_unbound_pool() can be called 385329c91e99STejun Heo * on @pool safely to release it. 38547a4e344cSTejun Heo */ 38557a4e344cSTejun Heo static int init_worker_pool(struct worker_pool *pool) 38564e1a1f9aSTejun Heo { 3857a9b8a985SSebastian Andrzej Siewior raw_spin_lock_init(&pool->lock); 385829c91e99STejun Heo pool->id = -1; 385929c91e99STejun Heo pool->cpu = -1; 3860f3f90ad4STejun Heo pool->node = NUMA_NO_NODE; 38614e1a1f9aSTejun Heo pool->flags |= POOL_DISASSOCIATED; 386282607adcSTejun Heo pool->watchdog_ts = jiffies; 38634e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->worklist); 38644e1a1f9aSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 38654e1a1f9aSTejun Heo hash_init(pool->busy_hash); 38664e1a1f9aSTejun Heo 386732a6c723SKees Cook timer_setup(&pool->idle_timer, idle_worker_timeout, TIMER_DEFERRABLE); 38683f959aa3SValentin Schneider INIT_WORK(&pool->idle_cull_work, idle_cull_fn); 38694e1a1f9aSTejun Heo 387032a6c723SKees Cook timer_setup(&pool->mayday_timer, pool_mayday_timeout, 0); 38714e1a1f9aSTejun Heo 3872da028469SLai Jiangshan INIT_LIST_HEAD(&pool->workers); 3873e02b9312SValentin Schneider INIT_LIST_HEAD(&pool->dying_workers); 38747a4e344cSTejun Heo 38757cda9aaeSLai Jiangshan ida_init(&pool->worker_ida); 387629c91e99STejun Heo INIT_HLIST_NODE(&pool->hash_node); 387729c91e99STejun Heo pool->refcnt = 1; 387829c91e99STejun Heo 387929c91e99STejun Heo /* shouldn't fail above this point */ 3880be69d00dSThomas Gleixner pool->attrs = alloc_workqueue_attrs(); 38817a4e344cSTejun Heo if (!pool->attrs) 38827a4e344cSTejun Heo return -ENOMEM; 38835de7a03cSTejun Heo 38845de7a03cSTejun Heo wqattrs_clear_for_pool(pool->attrs); 38855de7a03cSTejun Heo 38867a4e344cSTejun Heo return 0; 38874e1a1f9aSTejun Heo } 38884e1a1f9aSTejun Heo 3889669de8bdSBart Van Assche #ifdef CONFIG_LOCKDEP 3890669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3891669de8bdSBart Van Assche { 3892669de8bdSBart Van Assche char *lock_name; 3893669de8bdSBart Van Assche 3894669de8bdSBart Van Assche lockdep_register_key(&wq->key); 3895669de8bdSBart Van Assche lock_name = kasprintf(GFP_KERNEL, "%s%s", "(wq_completion)", wq->name); 3896669de8bdSBart Van Assche if (!lock_name) 3897669de8bdSBart Van Assche lock_name = wq->name; 389869a106c0SQian Cai 389969a106c0SQian Cai wq->lock_name = lock_name; 3900669de8bdSBart Van Assche lockdep_init_map(&wq->lockdep_map, lock_name, &wq->key, 0); 3901669de8bdSBart Van Assche } 3902669de8bdSBart Van Assche 3903669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3904669de8bdSBart Van Assche { 3905669de8bdSBart Van Assche lockdep_unregister_key(&wq->key); 3906669de8bdSBart Van Assche } 3907669de8bdSBart Van Assche 3908669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3909669de8bdSBart Van Assche { 3910669de8bdSBart Van Assche if (wq->lock_name != wq->name) 3911669de8bdSBart Van Assche kfree(wq->lock_name); 3912669de8bdSBart Van Assche } 3913669de8bdSBart Van Assche #else 3914669de8bdSBart Van Assche static void wq_init_lockdep(struct workqueue_struct *wq) 3915669de8bdSBart Van Assche { 3916669de8bdSBart Van Assche } 3917669de8bdSBart Van Assche 3918669de8bdSBart Van Assche static void wq_unregister_lockdep(struct workqueue_struct *wq) 3919669de8bdSBart Van Assche { 3920669de8bdSBart Van Assche } 3921669de8bdSBart Van Assche 3922669de8bdSBart Van Assche static void wq_free_lockdep(struct workqueue_struct *wq) 3923669de8bdSBart Van Assche { 3924669de8bdSBart Van Assche } 3925669de8bdSBart Van Assche #endif 3926669de8bdSBart Van Assche 3927e2dca7adSTejun Heo static void rcu_free_wq(struct rcu_head *rcu) 3928e2dca7adSTejun Heo { 3929e2dca7adSTejun Heo struct workqueue_struct *wq = 3930e2dca7adSTejun Heo container_of(rcu, struct workqueue_struct, rcu); 3931e2dca7adSTejun Heo 3932669de8bdSBart Van Assche wq_free_lockdep(wq); 3933ee1ceef7STejun Heo free_percpu(wq->cpu_pwq); 3934e2dca7adSTejun Heo free_workqueue_attrs(wq->unbound_attrs); 3935e2dca7adSTejun Heo kfree(wq); 3936e2dca7adSTejun Heo } 3937e2dca7adSTejun Heo 393829c91e99STejun Heo static void rcu_free_pool(struct rcu_head *rcu) 393929c91e99STejun Heo { 394029c91e99STejun Heo struct worker_pool *pool = container_of(rcu, struct worker_pool, rcu); 394129c91e99STejun Heo 39427cda9aaeSLai Jiangshan ida_destroy(&pool->worker_ida); 394329c91e99STejun Heo free_workqueue_attrs(pool->attrs); 394429c91e99STejun Heo kfree(pool); 394529c91e99STejun Heo } 394629c91e99STejun Heo 394729c91e99STejun Heo /** 394829c91e99STejun Heo * put_unbound_pool - put a worker_pool 394929c91e99STejun Heo * @pool: worker_pool to put 395029c91e99STejun Heo * 395124acfb71SThomas Gleixner * Put @pool. If its refcnt reaches zero, it gets destroyed in RCU 3952c5aa87bbSTejun Heo * safe manner. get_unbound_pool() calls this function on its failure path 3953c5aa87bbSTejun Heo * and this function should be able to release pools which went through, 3954c5aa87bbSTejun Heo * successfully or not, init_worker_pool(). 3955a892caccSTejun Heo * 3956a892caccSTejun Heo * Should be called with wq_pool_mutex held. 395729c91e99STejun Heo */ 395829c91e99STejun Heo static void put_unbound_pool(struct worker_pool *pool) 395929c91e99STejun Heo { 396060f5a4bcSLai Jiangshan DECLARE_COMPLETION_ONSTACK(detach_completion); 396129c91e99STejun Heo struct worker *worker; 39629680540cSYang Yingliang LIST_HEAD(cull_list); 3963e02b9312SValentin Schneider 3964a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 3965a892caccSTejun Heo 3966a892caccSTejun Heo if (--pool->refcnt) 396729c91e99STejun Heo return; 396829c91e99STejun Heo 396929c91e99STejun Heo /* sanity checks */ 397061d0fbb4SLai Jiangshan if (WARN_ON(!(pool->cpu < 0)) || 3971a892caccSTejun Heo WARN_ON(!list_empty(&pool->worklist))) 397229c91e99STejun Heo return; 397329c91e99STejun Heo 397429c91e99STejun Heo /* release id and unhash */ 397529c91e99STejun Heo if (pool->id >= 0) 397629c91e99STejun Heo idr_remove(&worker_pool_idr, pool->id); 397729c91e99STejun Heo hash_del(&pool->hash_node); 397829c91e99STejun Heo 3979c5aa87bbSTejun Heo /* 3980692b4825STejun Heo * Become the manager and destroy all workers. This prevents 3981692b4825STejun Heo * @pool's workers from blocking on attach_mutex. We're the last 3982692b4825STejun Heo * manager and @pool gets freed with the flag set. 39839ab03be4SValentin Schneider * 39849ab03be4SValentin Schneider * Having a concurrent manager is quite unlikely to happen as we can 39859ab03be4SValentin Schneider * only get here with 39869ab03be4SValentin Schneider * pwq->refcnt == pool->refcnt == 0 39879ab03be4SValentin Schneider * which implies no work queued to the pool, which implies no worker can 39889ab03be4SValentin Schneider * become the manager. However a worker could have taken the role of 39899ab03be4SValentin Schneider * manager before the refcnts dropped to 0, since maybe_create_worker() 39909ab03be4SValentin Schneider * drops pool->lock 3991c5aa87bbSTejun Heo */ 39929ab03be4SValentin Schneider while (true) { 39939ab03be4SValentin Schneider rcuwait_wait_event(&manager_wait, 39949ab03be4SValentin Schneider !(pool->flags & POOL_MANAGER_ACTIVE), 3995d8bb65abSSebastian Andrzej Siewior TASK_UNINTERRUPTIBLE); 3996e02b9312SValentin Schneider 3997e02b9312SValentin Schneider mutex_lock(&wq_pool_attach_mutex); 39989ab03be4SValentin Schneider raw_spin_lock_irq(&pool->lock); 39999ab03be4SValentin Schneider if (!(pool->flags & POOL_MANAGER_ACTIVE)) { 4000692b4825STejun Heo pool->flags |= POOL_MANAGER_ACTIVE; 40019ab03be4SValentin Schneider break; 40029ab03be4SValentin Schneider } 40039ab03be4SValentin Schneider raw_spin_unlock_irq(&pool->lock); 4004e02b9312SValentin Schneider mutex_unlock(&wq_pool_attach_mutex); 40059ab03be4SValentin Schneider } 4006692b4825STejun Heo 40071037de36SLai Jiangshan while ((worker = first_idle_worker(pool))) 4008e02b9312SValentin Schneider set_worker_dying(worker, &cull_list); 400929c91e99STejun Heo WARN_ON(pool->nr_workers || pool->nr_idle); 4010a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 401160f5a4bcSLai Jiangshan 4012e02b9312SValentin Schneider wake_dying_workers(&cull_list); 4013e02b9312SValentin Schneider 4014e02b9312SValentin Schneider if (!list_empty(&pool->workers) || !list_empty(&pool->dying_workers)) 401560f5a4bcSLai Jiangshan pool->detach_completion = &detach_completion; 40161258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 401760f5a4bcSLai Jiangshan 401860f5a4bcSLai Jiangshan if (pool->detach_completion) 401960f5a4bcSLai Jiangshan wait_for_completion(pool->detach_completion); 402060f5a4bcSLai Jiangshan 402129c91e99STejun Heo /* shut down the timers */ 402229c91e99STejun Heo del_timer_sync(&pool->idle_timer); 40233f959aa3SValentin Schneider cancel_work_sync(&pool->idle_cull_work); 402429c91e99STejun Heo del_timer_sync(&pool->mayday_timer); 402529c91e99STejun Heo 402624acfb71SThomas Gleixner /* RCU protected to allow dereferences from get_work_pool() */ 402725b00775SPaul E. McKenney call_rcu(&pool->rcu, rcu_free_pool); 402829c91e99STejun Heo } 402929c91e99STejun Heo 403029c91e99STejun Heo /** 403129c91e99STejun Heo * get_unbound_pool - get a worker_pool with the specified attributes 403229c91e99STejun Heo * @attrs: the attributes of the worker_pool to get 403329c91e99STejun Heo * 403429c91e99STejun Heo * Obtain a worker_pool which has the same attributes as @attrs, bump the 403529c91e99STejun Heo * reference count and return it. If there already is a matching 403629c91e99STejun Heo * worker_pool, it will be used; otherwise, this function attempts to 4037d185af30SYacine Belkadi * create a new one. 4038a892caccSTejun Heo * 4039a892caccSTejun Heo * Should be called with wq_pool_mutex held. 4040d185af30SYacine Belkadi * 4041d185af30SYacine Belkadi * Return: On success, a worker_pool with the same attributes as @attrs. 4042d185af30SYacine Belkadi * On failure, %NULL. 404329c91e99STejun Heo */ 404429c91e99STejun Heo static struct worker_pool *get_unbound_pool(const struct workqueue_attrs *attrs) 404529c91e99STejun Heo { 404684193c07STejun Heo struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_NUMA]; 404729c91e99STejun Heo u32 hash = wqattrs_hash(attrs); 404829c91e99STejun Heo struct worker_pool *pool; 404984193c07STejun Heo int pod, node = NUMA_NO_NODE; 405029c91e99STejun Heo 4051a892caccSTejun Heo lockdep_assert_held(&wq_pool_mutex); 405229c91e99STejun Heo 405329c91e99STejun Heo /* do we already have a matching pool? */ 405429c91e99STejun Heo hash_for_each_possible(unbound_pool_hash, pool, hash_node, hash) { 405529c91e99STejun Heo if (wqattrs_equal(pool->attrs, attrs)) { 405629c91e99STejun Heo pool->refcnt++; 40573fb1823cSLai Jiangshan return pool; 405829c91e99STejun Heo } 405929c91e99STejun Heo } 406029c91e99STejun Heo 40619546b29eSTejun Heo /* If __pod_cpumask is contained inside a NUMA pod, that's our node */ 406284193c07STejun Heo for (pod = 0; pod < pt->nr_pods; pod++) { 40639546b29eSTejun Heo if (cpumask_subset(attrs->__pod_cpumask, pt->pod_cpus[pod])) { 406484193c07STejun Heo node = pt->pod_node[pod]; 4065e2273584SXunlei Pang break; 4066e2273584SXunlei Pang } 4067e2273584SXunlei Pang } 4068e2273584SXunlei Pang 406929c91e99STejun Heo /* nope, create a new one */ 407084193c07STejun Heo pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node); 407129c91e99STejun Heo if (!pool || init_worker_pool(pool) < 0) 407229c91e99STejun Heo goto fail; 407329c91e99STejun Heo 407484193c07STejun Heo pool->node = node; 40755de7a03cSTejun Heo copy_workqueue_attrs(pool->attrs, attrs); 40765de7a03cSTejun Heo wqattrs_clear_for_pool(pool->attrs); 40772865a8fbSShaohua Li 407829c91e99STejun Heo if (worker_pool_assign_id(pool) < 0) 407929c91e99STejun Heo goto fail; 408029c91e99STejun Heo 408129c91e99STejun Heo /* create and start the initial worker */ 40823347fa09STejun Heo if (wq_online && !create_worker(pool)) 408329c91e99STejun Heo goto fail; 408429c91e99STejun Heo 408529c91e99STejun Heo /* install */ 408629c91e99STejun Heo hash_add(unbound_pool_hash, &pool->hash_node, hash); 40873fb1823cSLai Jiangshan 408829c91e99STejun Heo return pool; 408929c91e99STejun Heo fail: 409029c91e99STejun Heo if (pool) 409129c91e99STejun Heo put_unbound_pool(pool); 409229c91e99STejun Heo return NULL; 409329c91e99STejun Heo } 409429c91e99STejun Heo 40958864b4e5STejun Heo static void rcu_free_pwq(struct rcu_head *rcu) 40968864b4e5STejun Heo { 40978864b4e5STejun Heo kmem_cache_free(pwq_cache, 40988864b4e5STejun Heo container_of(rcu, struct pool_workqueue, rcu)); 40998864b4e5STejun Heo } 41008864b4e5STejun Heo 41018864b4e5STejun Heo /* 4102967b494eSTejun Heo * Scheduled on pwq_release_worker by put_pwq() when an unbound pwq hits zero 4103967b494eSTejun Heo * refcnt and needs to be destroyed. 41048864b4e5STejun Heo */ 4105687a9aa5STejun Heo static void pwq_release_workfn(struct kthread_work *work) 41068864b4e5STejun Heo { 41078864b4e5STejun Heo struct pool_workqueue *pwq = container_of(work, struct pool_workqueue, 4108687a9aa5STejun Heo release_work); 41098864b4e5STejun Heo struct workqueue_struct *wq = pwq->wq; 41108864b4e5STejun Heo struct worker_pool *pool = pwq->pool; 4111b42b0bddSYang Yingliang bool is_last = false; 41128864b4e5STejun Heo 4113b42b0bddSYang Yingliang /* 4114687a9aa5STejun Heo * When @pwq is not linked, it doesn't hold any reference to the 4115b42b0bddSYang Yingliang * @wq, and @wq is invalid to access. 4116b42b0bddSYang Yingliang */ 4117b42b0bddSYang Yingliang if (!list_empty(&pwq->pwqs_node)) { 41183c25a55dSLai Jiangshan mutex_lock(&wq->mutex); 41198864b4e5STejun Heo list_del_rcu(&pwq->pwqs_node); 4120bc0caf09STejun Heo is_last = list_empty(&wq->pwqs); 41213c25a55dSLai Jiangshan mutex_unlock(&wq->mutex); 4122b42b0bddSYang Yingliang } 41238864b4e5STejun Heo 4124687a9aa5STejun Heo if (wq->flags & WQ_UNBOUND) { 4125a892caccSTejun Heo mutex_lock(&wq_pool_mutex); 41268864b4e5STejun Heo put_unbound_pool(pool); 4127a892caccSTejun Heo mutex_unlock(&wq_pool_mutex); 4128687a9aa5STejun Heo } 4129a892caccSTejun Heo 413025b00775SPaul E. McKenney call_rcu(&pwq->rcu, rcu_free_pwq); 41318864b4e5STejun Heo 41328864b4e5STejun Heo /* 41338864b4e5STejun Heo * If we're the last pwq going away, @wq is already dead and no one 4134e2dca7adSTejun Heo * is gonna access it anymore. Schedule RCU free. 41358864b4e5STejun Heo */ 4136669de8bdSBart Van Assche if (is_last) { 4137669de8bdSBart Van Assche wq_unregister_lockdep(wq); 413825b00775SPaul E. McKenney call_rcu(&wq->rcu, rcu_free_wq); 41396029a918STejun Heo } 4140669de8bdSBart Van Assche } 41418864b4e5STejun Heo 41420fbd95aaSTejun Heo /** 4143699ce097STejun Heo * pwq_adjust_max_active - update a pwq's max_active to the current setting 41440fbd95aaSTejun Heo * @pwq: target pool_workqueue 41450fbd95aaSTejun Heo * 4146699ce097STejun Heo * If @pwq isn't freezing, set @pwq->max_active to the associated 4147f97a4a1aSLai Jiangshan * workqueue's saved_max_active and activate inactive work items 4148699ce097STejun Heo * accordingly. If @pwq is freezing, clear @pwq->max_active to zero. 41490fbd95aaSTejun Heo */ 4150699ce097STejun Heo static void pwq_adjust_max_active(struct pool_workqueue *pwq) 41510fbd95aaSTejun Heo { 4152699ce097STejun Heo struct workqueue_struct *wq = pwq->wq; 4153699ce097STejun Heo bool freezable = wq->flags & WQ_FREEZABLE; 41543347fa09STejun Heo unsigned long flags; 4155699ce097STejun Heo 4156699ce097STejun Heo /* for @wq->saved_max_active */ 4157a357fc03SLai Jiangshan lockdep_assert_held(&wq->mutex); 4158699ce097STejun Heo 4159699ce097STejun Heo /* fast exit for non-freezable wqs */ 4160699ce097STejun Heo if (!freezable && pwq->max_active == wq->saved_max_active) 4161699ce097STejun Heo return; 4162699ce097STejun Heo 41633347fa09STejun Heo /* this function can be called during early boot w/ irq disabled */ 4164a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 4165699ce097STejun Heo 416674b414eaSLai Jiangshan /* 416774b414eaSLai Jiangshan * During [un]freezing, the caller is responsible for ensuring that 416874b414eaSLai Jiangshan * this function is called at least once after @workqueue_freezing 416974b414eaSLai Jiangshan * is updated and visible. 417074b414eaSLai Jiangshan */ 417174b414eaSLai Jiangshan if (!freezable || !workqueue_freezing) { 4172699ce097STejun Heo pwq->max_active = wq->saved_max_active; 41730fbd95aaSTejun Heo 4174f97a4a1aSLai Jiangshan while (!list_empty(&pwq->inactive_works) && 41750219a352STejun Heo pwq->nr_active < pwq->max_active) 4176f97a4a1aSLai Jiangshan pwq_activate_first_inactive(pwq); 4177951a078aSLai Jiangshan 41780219a352STejun Heo kick_pool(pwq->pool); 4179699ce097STejun Heo } else { 4180699ce097STejun Heo pwq->max_active = 0; 4181699ce097STejun Heo } 4182699ce097STejun Heo 4183a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 41840fbd95aaSTejun Heo } 41850fbd95aaSTejun Heo 418667dc8325SCai Huoqing /* initialize newly allocated @pwq which is associated with @wq and @pool */ 4187f147f29eSTejun Heo static void init_pwq(struct pool_workqueue *pwq, struct workqueue_struct *wq, 4188f147f29eSTejun Heo struct worker_pool *pool) 4189d2c1d404STejun Heo { 4190d2c1d404STejun Heo BUG_ON((unsigned long)pwq & WORK_STRUCT_FLAG_MASK); 4191d2c1d404STejun Heo 4192e50aba9aSTejun Heo memset(pwq, 0, sizeof(*pwq)); 4193e50aba9aSTejun Heo 4194d2c1d404STejun Heo pwq->pool = pool; 4195d2c1d404STejun Heo pwq->wq = wq; 4196d2c1d404STejun Heo pwq->flush_color = -1; 41978864b4e5STejun Heo pwq->refcnt = 1; 4198f97a4a1aSLai Jiangshan INIT_LIST_HEAD(&pwq->inactive_works); 41991befcf30STejun Heo INIT_LIST_HEAD(&pwq->pwqs_node); 4200d2c1d404STejun Heo INIT_LIST_HEAD(&pwq->mayday_node); 4201687a9aa5STejun Heo kthread_init_work(&pwq->release_work, pwq_release_workfn); 4202f147f29eSTejun Heo } 4203d2c1d404STejun Heo 4204f147f29eSTejun Heo /* sync @pwq with the current state of its associated wq and link it */ 42051befcf30STejun Heo static void link_pwq(struct pool_workqueue *pwq) 4206f147f29eSTejun Heo { 4207f147f29eSTejun Heo struct workqueue_struct *wq = pwq->wq; 4208f147f29eSTejun Heo 4209f147f29eSTejun Heo lockdep_assert_held(&wq->mutex); 421075ccf595STejun Heo 42111befcf30STejun Heo /* may be called multiple times, ignore if already linked */ 42121befcf30STejun Heo if (!list_empty(&pwq->pwqs_node)) 42131befcf30STejun Heo return; 42141befcf30STejun Heo 421529b1cb41SLai Jiangshan /* set the matching work_color */ 421675ccf595STejun Heo pwq->work_color = wq->work_color; 4217983ca25eSTejun Heo 4218983ca25eSTejun Heo /* sync max_active to the current setting */ 4219983ca25eSTejun Heo pwq_adjust_max_active(pwq); 4220983ca25eSTejun Heo 4221983ca25eSTejun Heo /* link in @pwq */ 42229e8cd2f5STejun Heo list_add_rcu(&pwq->pwqs_node, &wq->pwqs); 4223df2d5ae4STejun Heo } 42246029a918STejun Heo 4225f147f29eSTejun Heo /* obtain a pool matching @attr and create a pwq associating the pool and @wq */ 4226f147f29eSTejun Heo static struct pool_workqueue *alloc_unbound_pwq(struct workqueue_struct *wq, 4227f147f29eSTejun Heo const struct workqueue_attrs *attrs) 4228f147f29eSTejun Heo { 4229f147f29eSTejun Heo struct worker_pool *pool; 4230f147f29eSTejun Heo struct pool_workqueue *pwq; 4231f147f29eSTejun Heo 4232f147f29eSTejun Heo lockdep_assert_held(&wq_pool_mutex); 4233f147f29eSTejun Heo 4234f147f29eSTejun Heo pool = get_unbound_pool(attrs); 4235f147f29eSTejun Heo if (!pool) 4236f147f29eSTejun Heo return NULL; 4237f147f29eSTejun Heo 4238e50aba9aSTejun Heo pwq = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, pool->node); 4239f147f29eSTejun Heo if (!pwq) { 4240f147f29eSTejun Heo put_unbound_pool(pool); 4241f147f29eSTejun Heo return NULL; 4242f147f29eSTejun Heo } 4243f147f29eSTejun Heo 4244f147f29eSTejun Heo init_pwq(pwq, wq, pool); 4245f147f29eSTejun Heo return pwq; 4246d2c1d404STejun Heo } 4247d2c1d404STejun Heo 42484c16bd32STejun Heo /** 4249fef59c9cSTejun Heo * wq_calc_pod_cpumask - calculate a wq_attrs' cpumask for a pod 4250042f7df1SLai Jiangshan * @attrs: the wq_attrs of the default pwq of the target workqueue 425184193c07STejun Heo * @cpu: the target CPU 42524c16bd32STejun Heo * @cpu_going_down: if >= 0, the CPU to consider as offline 42534c16bd32STejun Heo * 4254fef59c9cSTejun Heo * Calculate the cpumask a workqueue with @attrs should use on @pod. If 4255fef59c9cSTejun Heo * @cpu_going_down is >= 0, that cpu is considered offline during calculation. 42569546b29eSTejun Heo * The result is stored in @attrs->__pod_cpumask. 42574c16bd32STejun Heo * 4258fef59c9cSTejun Heo * If pod affinity is not enabled, @attrs->cpumask is always used. If enabled 4259fef59c9cSTejun Heo * and @pod has online CPUs requested by @attrs, the returned cpumask is the 4260fef59c9cSTejun Heo * intersection of the possible CPUs of @pod and @attrs->cpumask. 42614c16bd32STejun Heo * 4262fef59c9cSTejun Heo * The caller is responsible for ensuring that the cpumask of @pod stays stable. 42634c16bd32STejun Heo */ 42649546b29eSTejun Heo static void wq_calc_pod_cpumask(struct workqueue_attrs *attrs, int cpu, 42659546b29eSTejun Heo int cpu_going_down) 42664c16bd32STejun Heo { 426784193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 426884193c07STejun Heo int pod = pt->cpu_pod[cpu]; 42694c16bd32STejun Heo 4270fef59c9cSTejun Heo /* does @pod have any online CPUs @attrs wants? */ 42719546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, pt->pod_cpus[pod], attrs->cpumask); 42729546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, attrs->__pod_cpumask, cpu_online_mask); 42734c16bd32STejun Heo if (cpu_going_down >= 0) 42749546b29eSTejun Heo cpumask_clear_cpu(cpu_going_down, attrs->__pod_cpumask); 42754c16bd32STejun Heo 42769546b29eSTejun Heo if (cpumask_empty(attrs->__pod_cpumask)) { 42779546b29eSTejun Heo cpumask_copy(attrs->__pod_cpumask, attrs->cpumask); 427884193c07STejun Heo return; 427984193c07STejun Heo } 42804c16bd32STejun Heo 4281fef59c9cSTejun Heo /* yeap, return possible CPUs in @pod that @attrs wants */ 42829546b29eSTejun Heo cpumask_and(attrs->__pod_cpumask, attrs->cpumask, pt->pod_cpus[pod]); 42831ad0f0a7SMichael Bringmann 42849546b29eSTejun Heo if (cpumask_empty(attrs->__pod_cpumask)) 42851ad0f0a7SMichael Bringmann pr_warn_once("WARNING: workqueue cpumask: online intersect > " 42861ad0f0a7SMichael Bringmann "possible intersect\n"); 42874c16bd32STejun Heo } 42884c16bd32STejun Heo 4289636b927eSTejun Heo /* install @pwq into @wq's cpu_pwq and return the old pwq */ 4290636b927eSTejun Heo static struct pool_workqueue *install_unbound_pwq(struct workqueue_struct *wq, 4291636b927eSTejun Heo int cpu, struct pool_workqueue *pwq) 42921befcf30STejun Heo { 42931befcf30STejun Heo struct pool_workqueue *old_pwq; 42941befcf30STejun Heo 42955b95e1afSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 42961befcf30STejun Heo lockdep_assert_held(&wq->mutex); 42971befcf30STejun Heo 42981befcf30STejun Heo /* link_pwq() can handle duplicate calls */ 42991befcf30STejun Heo link_pwq(pwq); 43001befcf30STejun Heo 4301636b927eSTejun Heo old_pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4302636b927eSTejun Heo rcu_assign_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu), pwq); 43031befcf30STejun Heo return old_pwq; 43041befcf30STejun Heo } 43051befcf30STejun Heo 43062d5f0764SLai Jiangshan /* context to store the prepared attrs & pwqs before applying */ 43072d5f0764SLai Jiangshan struct apply_wqattrs_ctx { 43082d5f0764SLai Jiangshan struct workqueue_struct *wq; /* target workqueue */ 43092d5f0764SLai Jiangshan struct workqueue_attrs *attrs; /* attrs to apply */ 4310042f7df1SLai Jiangshan struct list_head list; /* queued for batching commit */ 43112d5f0764SLai Jiangshan struct pool_workqueue *dfl_pwq; 43122d5f0764SLai Jiangshan struct pool_workqueue *pwq_tbl[]; 43132d5f0764SLai Jiangshan }; 43142d5f0764SLai Jiangshan 43152d5f0764SLai Jiangshan /* free the resources after success or abort */ 43162d5f0764SLai Jiangshan static void apply_wqattrs_cleanup(struct apply_wqattrs_ctx *ctx) 43172d5f0764SLai Jiangshan { 43182d5f0764SLai Jiangshan if (ctx) { 4319636b927eSTejun Heo int cpu; 43202d5f0764SLai Jiangshan 4321636b927eSTejun Heo for_each_possible_cpu(cpu) 4322636b927eSTejun Heo put_pwq_unlocked(ctx->pwq_tbl[cpu]); 43232d5f0764SLai Jiangshan put_pwq_unlocked(ctx->dfl_pwq); 43242d5f0764SLai Jiangshan 43252d5f0764SLai Jiangshan free_workqueue_attrs(ctx->attrs); 43262d5f0764SLai Jiangshan 43272d5f0764SLai Jiangshan kfree(ctx); 43282d5f0764SLai Jiangshan } 43292d5f0764SLai Jiangshan } 43302d5f0764SLai Jiangshan 43312d5f0764SLai Jiangshan /* allocate the attrs and pwqs for later installation */ 43322d5f0764SLai Jiangshan static struct apply_wqattrs_ctx * 43332d5f0764SLai Jiangshan apply_wqattrs_prepare(struct workqueue_struct *wq, 433499c621efSLai Jiangshan const struct workqueue_attrs *attrs, 433599c621efSLai Jiangshan const cpumask_var_t unbound_cpumask) 43362d5f0764SLai Jiangshan { 43372d5f0764SLai Jiangshan struct apply_wqattrs_ctx *ctx; 43389546b29eSTejun Heo struct workqueue_attrs *new_attrs; 4339636b927eSTejun Heo int cpu; 43402d5f0764SLai Jiangshan 43412d5f0764SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 43422d5f0764SLai Jiangshan 434384193c07STejun Heo if (WARN_ON(attrs->affn_scope < 0 || 434484193c07STejun Heo attrs->affn_scope >= WQ_AFFN_NR_TYPES)) 434584193c07STejun Heo return ERR_PTR(-EINVAL); 434684193c07STejun Heo 4347636b927eSTejun Heo ctx = kzalloc(struct_size(ctx, pwq_tbl, nr_cpu_ids), GFP_KERNEL); 43482d5f0764SLai Jiangshan 4349be69d00dSThomas Gleixner new_attrs = alloc_workqueue_attrs(); 43509546b29eSTejun Heo if (!ctx || !new_attrs) 43512d5f0764SLai Jiangshan goto out_free; 43522d5f0764SLai Jiangshan 4353042f7df1SLai Jiangshan /* 43542d5f0764SLai Jiangshan * If something goes wrong during CPU up/down, we'll fall back to 43552d5f0764SLai Jiangshan * the default pwq covering whole @attrs->cpumask. Always create 43562d5f0764SLai Jiangshan * it even if we don't use it immediately. 43572d5f0764SLai Jiangshan */ 43580f36ee24STejun Heo copy_workqueue_attrs(new_attrs, attrs); 43590f36ee24STejun Heo wqattrs_actualize_cpumask(new_attrs, unbound_cpumask); 43609546b29eSTejun Heo cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 43612d5f0764SLai Jiangshan ctx->dfl_pwq = alloc_unbound_pwq(wq, new_attrs); 43622d5f0764SLai Jiangshan if (!ctx->dfl_pwq) 43632d5f0764SLai Jiangshan goto out_free; 43642d5f0764SLai Jiangshan 4365636b927eSTejun Heo for_each_possible_cpu(cpu) { 4366af73f5c9STejun Heo if (new_attrs->ordered) { 43672d5f0764SLai Jiangshan ctx->dfl_pwq->refcnt++; 4368636b927eSTejun Heo ctx->pwq_tbl[cpu] = ctx->dfl_pwq; 4369636b927eSTejun Heo } else { 43709546b29eSTejun Heo wq_calc_pod_cpumask(new_attrs, cpu, -1); 43719546b29eSTejun Heo ctx->pwq_tbl[cpu] = alloc_unbound_pwq(wq, new_attrs); 4372636b927eSTejun Heo if (!ctx->pwq_tbl[cpu]) 4373636b927eSTejun Heo goto out_free; 43742d5f0764SLai Jiangshan } 43752d5f0764SLai Jiangshan } 43762d5f0764SLai Jiangshan 4377042f7df1SLai Jiangshan /* save the user configured attrs and sanitize it. */ 4378042f7df1SLai Jiangshan copy_workqueue_attrs(new_attrs, attrs); 4379042f7df1SLai Jiangshan cpumask_and(new_attrs->cpumask, new_attrs->cpumask, cpu_possible_mask); 43809546b29eSTejun Heo cpumask_copy(new_attrs->__pod_cpumask, new_attrs->cpumask); 43812d5f0764SLai Jiangshan ctx->attrs = new_attrs; 4382042f7df1SLai Jiangshan 43832d5f0764SLai Jiangshan ctx->wq = wq; 43842d5f0764SLai Jiangshan return ctx; 43852d5f0764SLai Jiangshan 43862d5f0764SLai Jiangshan out_free: 43872d5f0764SLai Jiangshan free_workqueue_attrs(new_attrs); 43882d5f0764SLai Jiangshan apply_wqattrs_cleanup(ctx); 438984193c07STejun Heo return ERR_PTR(-ENOMEM); 43902d5f0764SLai Jiangshan } 43912d5f0764SLai Jiangshan 43922d5f0764SLai Jiangshan /* set attrs and install prepared pwqs, @ctx points to old pwqs on return */ 43932d5f0764SLai Jiangshan static void apply_wqattrs_commit(struct apply_wqattrs_ctx *ctx) 43942d5f0764SLai Jiangshan { 4395636b927eSTejun Heo int cpu; 43962d5f0764SLai Jiangshan 43972d5f0764SLai Jiangshan /* all pwqs have been created successfully, let's install'em */ 43982d5f0764SLai Jiangshan mutex_lock(&ctx->wq->mutex); 43992d5f0764SLai Jiangshan 44002d5f0764SLai Jiangshan copy_workqueue_attrs(ctx->wq->unbound_attrs, ctx->attrs); 44012d5f0764SLai Jiangshan 44022d5f0764SLai Jiangshan /* save the previous pwq and install the new one */ 4403636b927eSTejun Heo for_each_possible_cpu(cpu) 4404636b927eSTejun Heo ctx->pwq_tbl[cpu] = install_unbound_pwq(ctx->wq, cpu, 4405636b927eSTejun Heo ctx->pwq_tbl[cpu]); 44062d5f0764SLai Jiangshan 44072d5f0764SLai Jiangshan /* @dfl_pwq might not have been used, ensure it's linked */ 44082d5f0764SLai Jiangshan link_pwq(ctx->dfl_pwq); 44092d5f0764SLai Jiangshan swap(ctx->wq->dfl_pwq, ctx->dfl_pwq); 44102d5f0764SLai Jiangshan 44112d5f0764SLai Jiangshan mutex_unlock(&ctx->wq->mutex); 44122d5f0764SLai Jiangshan } 44132d5f0764SLai Jiangshan 4414a0111cf6SLai Jiangshan static void apply_wqattrs_lock(void) 4415a0111cf6SLai Jiangshan { 4416a0111cf6SLai Jiangshan /* CPUs should stay stable across pwq creations and installations */ 4417ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4418a0111cf6SLai Jiangshan mutex_lock(&wq_pool_mutex); 4419a0111cf6SLai Jiangshan } 4420a0111cf6SLai Jiangshan 4421a0111cf6SLai Jiangshan static void apply_wqattrs_unlock(void) 4422a0111cf6SLai Jiangshan { 4423a0111cf6SLai Jiangshan mutex_unlock(&wq_pool_mutex); 4424ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4425a0111cf6SLai Jiangshan } 4426a0111cf6SLai Jiangshan 4427a0111cf6SLai Jiangshan static int apply_workqueue_attrs_locked(struct workqueue_struct *wq, 4428a0111cf6SLai Jiangshan const struct workqueue_attrs *attrs) 4429a0111cf6SLai Jiangshan { 4430a0111cf6SLai Jiangshan struct apply_wqattrs_ctx *ctx; 4431a0111cf6SLai Jiangshan 4432a0111cf6SLai Jiangshan /* only unbound workqueues can change attributes */ 4433a0111cf6SLai Jiangshan if (WARN_ON(!(wq->flags & WQ_UNBOUND))) 4434a0111cf6SLai Jiangshan return -EINVAL; 4435a0111cf6SLai Jiangshan 4436a0111cf6SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 44370a94efb5STejun Heo if (!list_empty(&wq->pwqs)) { 44380a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 4439a0111cf6SLai Jiangshan return -EINVAL; 4440a0111cf6SLai Jiangshan 44410a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 44420a94efb5STejun Heo } 44430a94efb5STejun Heo 444499c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, attrs, wq_unbound_cpumask); 444584193c07STejun Heo if (IS_ERR(ctx)) 444684193c07STejun Heo return PTR_ERR(ctx); 4447a0111cf6SLai Jiangshan 4448a0111cf6SLai Jiangshan /* the ctx has been prepared successfully, let's commit it */ 4449a0111cf6SLai Jiangshan apply_wqattrs_commit(ctx); 4450a0111cf6SLai Jiangshan apply_wqattrs_cleanup(ctx); 4451a0111cf6SLai Jiangshan 44526201171eSwanghaibin return 0; 4453a0111cf6SLai Jiangshan } 4454a0111cf6SLai Jiangshan 44559e8cd2f5STejun Heo /** 44569e8cd2f5STejun Heo * apply_workqueue_attrs - apply new workqueue_attrs to an unbound workqueue 44579e8cd2f5STejun Heo * @wq: the target workqueue 44589e8cd2f5STejun Heo * @attrs: the workqueue_attrs to apply, allocated with alloc_workqueue_attrs() 44599e8cd2f5STejun Heo * 4460fef59c9cSTejun Heo * Apply @attrs to an unbound workqueue @wq. Unless disabled, this function maps 4461fef59c9cSTejun Heo * a separate pwq to each CPU pod with possibles CPUs in @attrs->cpumask so that 4462fef59c9cSTejun Heo * work items are affine to the pod it was issued on. Older pwqs are released as 4463fef59c9cSTejun Heo * in-flight work items finish. Note that a work item which repeatedly requeues 4464fef59c9cSTejun Heo * itself back-to-back will stay on its current pwq. 44659e8cd2f5STejun Heo * 4466d185af30SYacine Belkadi * Performs GFP_KERNEL allocations. 4467d185af30SYacine Belkadi * 4468ffd8bea8SSebastian Andrzej Siewior * Assumes caller has CPU hotplug read exclusion, i.e. cpus_read_lock(). 4469509b3204SDaniel Jordan * 4470d185af30SYacine Belkadi * Return: 0 on success and -errno on failure. 44719e8cd2f5STejun Heo */ 4472513c98d0SDaniel Jordan int apply_workqueue_attrs(struct workqueue_struct *wq, 44739e8cd2f5STejun Heo const struct workqueue_attrs *attrs) 44749e8cd2f5STejun Heo { 4475a0111cf6SLai Jiangshan int ret; 44769e8cd2f5STejun Heo 4477509b3204SDaniel Jordan lockdep_assert_cpus_held(); 4478509b3204SDaniel Jordan 4479509b3204SDaniel Jordan mutex_lock(&wq_pool_mutex); 4480a0111cf6SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 4481509b3204SDaniel Jordan mutex_unlock(&wq_pool_mutex); 44822d5f0764SLai Jiangshan 44832d5f0764SLai Jiangshan return ret; 44849e8cd2f5STejun Heo } 44859e8cd2f5STejun Heo 44864c16bd32STejun Heo /** 4487fef59c9cSTejun Heo * wq_update_pod - update pod affinity of a wq for CPU hot[un]plug 44884c16bd32STejun Heo * @wq: the target workqueue 44894cbfd3deSTejun Heo * @cpu: the CPU to update pool association for 44904cbfd3deSTejun Heo * @hotplug_cpu: the CPU coming up or going down 44914c16bd32STejun Heo * @online: whether @cpu is coming up or going down 44924c16bd32STejun Heo * 44934c16bd32STejun Heo * This function is to be called from %CPU_DOWN_PREPARE, %CPU_ONLINE and 4494fef59c9cSTejun Heo * %CPU_DOWN_FAILED. @cpu is being hot[un]plugged, update pod affinity of 44954c16bd32STejun Heo * @wq accordingly. 44964c16bd32STejun Heo * 44974c16bd32STejun Heo * 4498fef59c9cSTejun Heo * If pod affinity can't be adjusted due to memory allocation failure, it falls 4499fef59c9cSTejun Heo * back to @wq->dfl_pwq which may not be optimal but is always correct. 4500fef59c9cSTejun Heo * 4501fef59c9cSTejun Heo * Note that when the last allowed CPU of a pod goes offline for a workqueue 4502fef59c9cSTejun Heo * with a cpumask spanning multiple pods, the workers which were already 4503fef59c9cSTejun Heo * executing the work items for the workqueue will lose their CPU affinity and 4504fef59c9cSTejun Heo * may execute on any CPU. This is similar to how per-cpu workqueues behave on 4505fef59c9cSTejun Heo * CPU_DOWN. If a workqueue user wants strict affinity, it's the user's 4506fef59c9cSTejun Heo * responsibility to flush the work item from CPU_DOWN_PREPARE. 45074c16bd32STejun Heo */ 4508fef59c9cSTejun Heo static void wq_update_pod(struct workqueue_struct *wq, int cpu, 45094cbfd3deSTejun Heo int hotplug_cpu, bool online) 45104c16bd32STejun Heo { 45114cbfd3deSTejun Heo int off_cpu = online ? -1 : hotplug_cpu; 45124c16bd32STejun Heo struct pool_workqueue *old_pwq = NULL, *pwq; 45134c16bd32STejun Heo struct workqueue_attrs *target_attrs; 45144c16bd32STejun Heo 45154c16bd32STejun Heo lockdep_assert_held(&wq_pool_mutex); 45164c16bd32STejun Heo 451784193c07STejun Heo if (!(wq->flags & WQ_UNBOUND) || wq->unbound_attrs->ordered) 45184c16bd32STejun Heo return; 45194c16bd32STejun Heo 45204c16bd32STejun Heo /* 45214c16bd32STejun Heo * We don't wanna alloc/free wq_attrs for each wq for each CPU. 45224c16bd32STejun Heo * Let's use a preallocated one. The following buf is protected by 45234c16bd32STejun Heo * CPU hotplug exclusion. 45244c16bd32STejun Heo */ 4525fef59c9cSTejun Heo target_attrs = wq_update_pod_attrs_buf; 45264c16bd32STejun Heo 45274c16bd32STejun Heo copy_workqueue_attrs(target_attrs, wq->unbound_attrs); 45280f36ee24STejun Heo wqattrs_actualize_cpumask(target_attrs, wq_unbound_cpumask); 45294c16bd32STejun Heo 4530636b927eSTejun Heo /* nothing to do if the target cpumask matches the current pwq */ 45319546b29eSTejun Heo wq_calc_pod_cpumask(target_attrs, cpu, off_cpu); 4532636b927eSTejun Heo pwq = rcu_dereference_protected(*per_cpu_ptr(wq->cpu_pwq, cpu), 4533636b927eSTejun Heo lockdep_is_held(&wq_pool_mutex)); 45349546b29eSTejun Heo if (wqattrs_equal(target_attrs, pwq->pool->attrs)) 4535f7142ed4SLai Jiangshan return; 45364c16bd32STejun Heo 45374c16bd32STejun Heo /* create a new pwq */ 45384c16bd32STejun Heo pwq = alloc_unbound_pwq(wq, target_attrs); 45394c16bd32STejun Heo if (!pwq) { 4540fef59c9cSTejun Heo pr_warn("workqueue: allocation failed while updating CPU pod affinity of \"%s\"\n", 45414c16bd32STejun Heo wq->name); 454277f300b1SDaeseok Youn goto use_dfl_pwq; 45434c16bd32STejun Heo } 45444c16bd32STejun Heo 4545f7142ed4SLai Jiangshan /* Install the new pwq. */ 45464c16bd32STejun Heo mutex_lock(&wq->mutex); 4547636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, pwq); 45484c16bd32STejun Heo goto out_unlock; 45494c16bd32STejun Heo 45504c16bd32STejun Heo use_dfl_pwq: 4551f7142ed4SLai Jiangshan mutex_lock(&wq->mutex); 4552a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq->dfl_pwq->pool->lock); 45534c16bd32STejun Heo get_pwq(wq->dfl_pwq); 4554a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq->dfl_pwq->pool->lock); 4555636b927eSTejun Heo old_pwq = install_unbound_pwq(wq, cpu, wq->dfl_pwq); 45564c16bd32STejun Heo out_unlock: 45574c16bd32STejun Heo mutex_unlock(&wq->mutex); 45584c16bd32STejun Heo put_pwq_unlocked(old_pwq); 45594c16bd32STejun Heo } 45604c16bd32STejun Heo 456130cdf249STejun Heo static int alloc_and_link_pwqs(struct workqueue_struct *wq) 45621da177e4SLinus Torvalds { 456349e3cf44STejun Heo bool highpri = wq->flags & WQ_HIGHPRI; 45648a2b7538STejun Heo int cpu, ret; 4565e1d8aa9fSFrederic Weisbecker 4566687a9aa5STejun Heo wq->cpu_pwq = alloc_percpu(struct pool_workqueue *); 4567ee1ceef7STejun Heo if (!wq->cpu_pwq) 4568687a9aa5STejun Heo goto enomem; 456930cdf249STejun Heo 4570636b927eSTejun Heo if (!(wq->flags & WQ_UNBOUND)) { 457130cdf249STejun Heo for_each_possible_cpu(cpu) { 4572687a9aa5STejun Heo struct pool_workqueue **pwq_p = 4573ee1ceef7STejun Heo per_cpu_ptr(wq->cpu_pwq, cpu); 4574687a9aa5STejun Heo struct worker_pool *pool = 4575687a9aa5STejun Heo &(per_cpu_ptr(cpu_worker_pools, cpu)[highpri]); 457630cdf249STejun Heo 4577687a9aa5STejun Heo *pwq_p = kmem_cache_alloc_node(pwq_cache, GFP_KERNEL, 4578687a9aa5STejun Heo pool->node); 4579687a9aa5STejun Heo if (!*pwq_p) 4580687a9aa5STejun Heo goto enomem; 4581687a9aa5STejun Heo 4582687a9aa5STejun Heo init_pwq(*pwq_p, wq, pool); 4583f147f29eSTejun Heo 4584f147f29eSTejun Heo mutex_lock(&wq->mutex); 4585687a9aa5STejun Heo link_pwq(*pwq_p); 4586f147f29eSTejun Heo mutex_unlock(&wq->mutex); 458730cdf249STejun Heo } 458830cdf249STejun Heo return 0; 4589509b3204SDaniel Jordan } 4590509b3204SDaniel Jordan 4591ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 4592509b3204SDaniel Jordan if (wq->flags & __WQ_ORDERED) { 45938a2b7538STejun Heo ret = apply_workqueue_attrs(wq, ordered_wq_attrs[highpri]); 45948a2b7538STejun Heo /* there should only be single pwq for ordering guarantee */ 45958a2b7538STejun Heo WARN(!ret && (wq->pwqs.next != &wq->dfl_pwq->pwqs_node || 45968a2b7538STejun Heo wq->pwqs.prev != &wq->dfl_pwq->pwqs_node), 45978a2b7538STejun Heo "ordering guarantee broken for workqueue %s\n", wq->name); 45989e8cd2f5STejun Heo } else { 4599509b3204SDaniel Jordan ret = apply_workqueue_attrs(wq, unbound_std_wq_attrs[highpri]); 46009e8cd2f5STejun Heo } 4601ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 4602509b3204SDaniel Jordan 4603509b3204SDaniel Jordan return ret; 4604687a9aa5STejun Heo 4605687a9aa5STejun Heo enomem: 4606687a9aa5STejun Heo if (wq->cpu_pwq) { 4607687a9aa5STejun Heo for_each_possible_cpu(cpu) 4608687a9aa5STejun Heo kfree(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4609687a9aa5STejun Heo free_percpu(wq->cpu_pwq); 4610687a9aa5STejun Heo wq->cpu_pwq = NULL; 4611687a9aa5STejun Heo } 4612687a9aa5STejun Heo return -ENOMEM; 46130f900049STejun Heo } 46140f900049STejun Heo 4615f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 4616f3421797STejun Heo const char *name) 4617b71ab8c2STejun Heo { 4618636b927eSTejun Heo if (max_active < 1 || max_active > WQ_MAX_ACTIVE) 4619044c782cSValentin Ilie pr_warn("workqueue: max_active %d requested for %s is out of range, clamping between %d and %d\n", 4620636b927eSTejun Heo max_active, name, 1, WQ_MAX_ACTIVE); 4621b71ab8c2STejun Heo 4622636b927eSTejun Heo return clamp_val(max_active, 1, WQ_MAX_ACTIVE); 4623b71ab8c2STejun Heo } 4624b71ab8c2STejun Heo 4625983c7515STejun Heo /* 4626983c7515STejun Heo * Workqueues which may be used during memory reclaim should have a rescuer 4627983c7515STejun Heo * to guarantee forward progress. 4628983c7515STejun Heo */ 4629983c7515STejun Heo static int init_rescuer(struct workqueue_struct *wq) 4630983c7515STejun Heo { 4631983c7515STejun Heo struct worker *rescuer; 4632b92b36eaSDan Carpenter int ret; 4633983c7515STejun Heo 4634983c7515STejun Heo if (!(wq->flags & WQ_MEM_RECLAIM)) 4635983c7515STejun Heo return 0; 4636983c7515STejun Heo 4637983c7515STejun Heo rescuer = alloc_worker(NUMA_NO_NODE); 46384c0736a7SPetr Mladek if (!rescuer) { 46394c0736a7SPetr Mladek pr_err("workqueue: Failed to allocate a rescuer for wq \"%s\"\n", 46404c0736a7SPetr Mladek wq->name); 4641983c7515STejun Heo return -ENOMEM; 46424c0736a7SPetr Mladek } 4643983c7515STejun Heo 4644983c7515STejun Heo rescuer->rescue_wq = wq; 4645983c7515STejun Heo rescuer->task = kthread_create(rescuer_thread, rescuer, "%s", wq->name); 4646f187b697SSean Fu if (IS_ERR(rescuer->task)) { 4647b92b36eaSDan Carpenter ret = PTR_ERR(rescuer->task); 46484c0736a7SPetr Mladek pr_err("workqueue: Failed to create a rescuer kthread for wq \"%s\": %pe", 46494c0736a7SPetr Mladek wq->name, ERR_PTR(ret)); 4650983c7515STejun Heo kfree(rescuer); 4651b92b36eaSDan Carpenter return ret; 4652983c7515STejun Heo } 4653983c7515STejun Heo 4654983c7515STejun Heo wq->rescuer = rescuer; 4655983c7515STejun Heo kthread_bind_mask(rescuer->task, cpu_possible_mask); 4656983c7515STejun Heo wake_up_process(rescuer->task); 4657983c7515STejun Heo 4658983c7515STejun Heo return 0; 4659983c7515STejun Heo } 4660983c7515STejun Heo 4661a2775bbcSMathieu Malaterre __printf(1, 4) 4662669de8bdSBart Van Assche struct workqueue_struct *alloc_workqueue(const char *fmt, 466397e37d7bSTejun Heo unsigned int flags, 4664669de8bdSBart Van Assche int max_active, ...) 46653af24433SOleg Nesterov { 4666ecf6881fSTejun Heo va_list args; 46673af24433SOleg Nesterov struct workqueue_struct *wq; 466849e3cf44STejun Heo struct pool_workqueue *pwq; 4669b196be89STejun Heo 46705c0338c6STejun Heo /* 4671fef59c9cSTejun Heo * Unbound && max_active == 1 used to imply ordered, which is no longer 4672fef59c9cSTejun Heo * the case on many machines due to per-pod pools. While 46735c0338c6STejun Heo * alloc_ordered_workqueue() is the right way to create an ordered 4674fef59c9cSTejun Heo * workqueue, keep the previous behavior to avoid subtle breakages. 46755c0338c6STejun Heo */ 46765c0338c6STejun Heo if ((flags & WQ_UNBOUND) && max_active == 1) 46775c0338c6STejun Heo flags |= __WQ_ORDERED; 46785c0338c6STejun Heo 4679cee22a15SViresh Kumar /* see the comment above the definition of WQ_POWER_EFFICIENT */ 4680cee22a15SViresh Kumar if ((flags & WQ_POWER_EFFICIENT) && wq_power_efficient) 4681cee22a15SViresh Kumar flags |= WQ_UNBOUND; 4682cee22a15SViresh Kumar 4683ecf6881fSTejun Heo /* allocate wq and format name */ 4684636b927eSTejun Heo wq = kzalloc(sizeof(*wq), GFP_KERNEL); 4685b196be89STejun Heo if (!wq) 4686d2c1d404STejun Heo return NULL; 4687b196be89STejun Heo 46886029a918STejun Heo if (flags & WQ_UNBOUND) { 4689be69d00dSThomas Gleixner wq->unbound_attrs = alloc_workqueue_attrs(); 46906029a918STejun Heo if (!wq->unbound_attrs) 46916029a918STejun Heo goto err_free_wq; 46926029a918STejun Heo } 46936029a918STejun Heo 4694669de8bdSBart Van Assche va_start(args, max_active); 4695ecf6881fSTejun Heo vsnprintf(wq->name, sizeof(wq->name), fmt, args); 4696b196be89STejun Heo va_end(args); 46973af24433SOleg Nesterov 4698d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 4699b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 47003af24433SOleg Nesterov 4701b196be89STejun Heo /* init wq */ 470297e37d7bSTejun Heo wq->flags = flags; 4703a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 47043c25a55dSLai Jiangshan mutex_init(&wq->mutex); 4705112202d9STejun Heo atomic_set(&wq->nr_pwqs_to_flush, 0); 470630cdf249STejun Heo INIT_LIST_HEAD(&wq->pwqs); 470773f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 470873f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 4709493a1724STejun Heo INIT_LIST_HEAD(&wq->maydays); 47103af24433SOleg Nesterov 4711669de8bdSBart Van Assche wq_init_lockdep(wq); 4712cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 47133af24433SOleg Nesterov 471430cdf249STejun Heo if (alloc_and_link_pwqs(wq) < 0) 471582efcab3SBart Van Assche goto err_unreg_lockdep; 47161537663fSTejun Heo 471740c17f75STejun Heo if (wq_online && init_rescuer(wq) < 0) 4718d2c1d404STejun Heo goto err_destroy; 4719e22bee78STejun Heo 4720226223abSTejun Heo if ((wq->flags & WQ_SYSFS) && workqueue_sysfs_register(wq)) 4721226223abSTejun Heo goto err_destroy; 4722226223abSTejun Heo 47236af8bf3dSOleg Nesterov /* 472468e13a67SLai Jiangshan * wq_pool_mutex protects global freeze state and workqueues list. 472568e13a67SLai Jiangshan * Grab it, adjust max_active and add the new @wq to workqueues 472668e13a67SLai Jiangshan * list. 47276af8bf3dSOleg Nesterov */ 472868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 4729a0a1a5fdSTejun Heo 4730a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 473149e3cf44STejun Heo for_each_pwq(pwq, wq) 4732699ce097STejun Heo pwq_adjust_max_active(pwq); 4733a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4734a0a1a5fdSTejun Heo 4735e2dca7adSTejun Heo list_add_tail_rcu(&wq->list, &workqueues); 4736a0a1a5fdSTejun Heo 473768e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 47383af24433SOleg Nesterov 47393af24433SOleg Nesterov return wq; 4740d2c1d404STejun Heo 474182efcab3SBart Van Assche err_unreg_lockdep: 4742009bb421SBart Van Assche wq_unregister_lockdep(wq); 4743009bb421SBart Van Assche wq_free_lockdep(wq); 474482efcab3SBart Van Assche err_free_wq: 47456029a918STejun Heo free_workqueue_attrs(wq->unbound_attrs); 47464690c4abSTejun Heo kfree(wq); 4747d2c1d404STejun Heo return NULL; 4748d2c1d404STejun Heo err_destroy: 4749d2c1d404STejun Heo destroy_workqueue(wq); 47504690c4abSTejun Heo return NULL; 47511da177e4SLinus Torvalds } 4752669de8bdSBart Van Assche EXPORT_SYMBOL_GPL(alloc_workqueue); 47531da177e4SLinus Torvalds 4754c29eb853STejun Heo static bool pwq_busy(struct pool_workqueue *pwq) 4755c29eb853STejun Heo { 4756c29eb853STejun Heo int i; 4757c29eb853STejun Heo 4758c29eb853STejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 4759c29eb853STejun Heo if (pwq->nr_in_flight[i]) 4760c29eb853STejun Heo return true; 4761c29eb853STejun Heo 4762c29eb853STejun Heo if ((pwq != pwq->wq->dfl_pwq) && (pwq->refcnt > 1)) 4763c29eb853STejun Heo return true; 4764f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) 4765c29eb853STejun Heo return true; 4766c29eb853STejun Heo 4767c29eb853STejun Heo return false; 4768c29eb853STejun Heo } 4769c29eb853STejun Heo 47703af24433SOleg Nesterov /** 47713af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 47723af24433SOleg Nesterov * @wq: target workqueue 47733af24433SOleg Nesterov * 47743af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 47753af24433SOleg Nesterov */ 47763af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 47773af24433SOleg Nesterov { 477849e3cf44STejun Heo struct pool_workqueue *pwq; 4779636b927eSTejun Heo int cpu; 47803af24433SOleg Nesterov 4781def98c84STejun Heo /* 4782def98c84STejun Heo * Remove it from sysfs first so that sanity check failure doesn't 4783def98c84STejun Heo * lead to sysfs name conflicts. 4784def98c84STejun Heo */ 4785def98c84STejun Heo workqueue_sysfs_unregister(wq); 4786def98c84STejun Heo 478733e3f0a3SRichard Clark /* mark the workqueue destruction is in progress */ 478833e3f0a3SRichard Clark mutex_lock(&wq->mutex); 478933e3f0a3SRichard Clark wq->flags |= __WQ_DESTROYING; 479033e3f0a3SRichard Clark mutex_unlock(&wq->mutex); 479133e3f0a3SRichard Clark 47929c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 47939c5a2ba7STejun Heo drain_workqueue(wq); 4794c8efcc25STejun Heo 4795def98c84STejun Heo /* kill rescuer, if sanity checks fail, leave it w/o rescuer */ 4796def98c84STejun Heo if (wq->rescuer) { 4797def98c84STejun Heo struct worker *rescuer = wq->rescuer; 4798def98c84STejun Heo 4799def98c84STejun Heo /* this prevents new queueing */ 4800a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&wq_mayday_lock); 4801def98c84STejun Heo wq->rescuer = NULL; 4802a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&wq_mayday_lock); 4803def98c84STejun Heo 4804def98c84STejun Heo /* rescuer will empty maydays list before exiting */ 4805def98c84STejun Heo kthread_stop(rescuer->task); 48068efe1223STejun Heo kfree(rescuer); 4807def98c84STejun Heo } 4808def98c84STejun Heo 4809c29eb853STejun Heo /* 4810c29eb853STejun Heo * Sanity checks - grab all the locks so that we wait for all 4811c29eb853STejun Heo * in-flight operations which may do put_pwq(). 4812c29eb853STejun Heo */ 4813c29eb853STejun Heo mutex_lock(&wq_pool_mutex); 4814b09f4fd3SLai Jiangshan mutex_lock(&wq->mutex); 481549e3cf44STejun Heo for_each_pwq(pwq, wq) { 4816a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pwq->pool->lock); 4817c29eb853STejun Heo if (WARN_ON(pwq_busy(pwq))) { 48181d9a6159SKefeng Wang pr_warn("%s: %s has the following busy pwq\n", 4819e66b39afSTejun Heo __func__, wq->name); 4820c29eb853STejun Heo show_pwq(pwq); 4821a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 4822b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 4823c29eb853STejun Heo mutex_unlock(&wq_pool_mutex); 482455df0933SImran Khan show_one_workqueue(wq); 48256183c009STejun Heo return; 482676af4d93STejun Heo } 4827a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pwq->pool->lock); 482876af4d93STejun Heo } 4829b09f4fd3SLai Jiangshan mutex_unlock(&wq->mutex); 48306183c009STejun Heo 4831a0a1a5fdSTejun Heo /* 4832a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 4833a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 4834a0a1a5fdSTejun Heo */ 4835e2dca7adSTejun Heo list_del_rcu(&wq->list); 483668e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 48373af24433SOleg Nesterov 48388864b4e5STejun Heo /* 4839636b927eSTejun Heo * We're the sole accessor of @wq. Directly access cpu_pwq and dfl_pwq 4840636b927eSTejun Heo * to put the base refs. @wq will be auto-destroyed from the last 4841636b927eSTejun Heo * pwq_put. RCU read lock prevents @wq from going away from under us. 48428864b4e5STejun Heo */ 4843636b927eSTejun Heo rcu_read_lock(); 4844636b927eSTejun Heo 4845636b927eSTejun Heo for_each_possible_cpu(cpu) { 4846636b927eSTejun Heo pwq = rcu_access_pointer(*per_cpu_ptr(wq->cpu_pwq, cpu)); 4847636b927eSTejun Heo RCU_INIT_POINTER(*per_cpu_ptr(wq->cpu_pwq, cpu), NULL); 48484c16bd32STejun Heo put_pwq_unlocked(pwq); 48494c16bd32STejun Heo } 48504c16bd32STejun Heo 4851636b927eSTejun Heo put_pwq_unlocked(wq->dfl_pwq); 48524c16bd32STejun Heo wq->dfl_pwq = NULL; 4853636b927eSTejun Heo 4854636b927eSTejun Heo rcu_read_unlock(); 48553af24433SOleg Nesterov } 48563af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 48573af24433SOleg Nesterov 4858dcd989cbSTejun Heo /** 4859dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 4860dcd989cbSTejun Heo * @wq: target workqueue 4861dcd989cbSTejun Heo * @max_active: new max_active value. 4862dcd989cbSTejun Heo * 4863dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 4864dcd989cbSTejun Heo * 4865dcd989cbSTejun Heo * CONTEXT: 4866dcd989cbSTejun Heo * Don't call from IRQ context. 4867dcd989cbSTejun Heo */ 4868dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 4869dcd989cbSTejun Heo { 487049e3cf44STejun Heo struct pool_workqueue *pwq; 4871dcd989cbSTejun Heo 48728719dceaSTejun Heo /* disallow meddling with max_active for ordered workqueues */ 48730a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 48748719dceaSTejun Heo return; 48758719dceaSTejun Heo 4876f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 4877dcd989cbSTejun Heo 4878a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 4879dcd989cbSTejun Heo 48800a94efb5STejun Heo wq->flags &= ~__WQ_ORDERED; 4881dcd989cbSTejun Heo wq->saved_max_active = max_active; 4882dcd989cbSTejun Heo 4883699ce097STejun Heo for_each_pwq(pwq, wq) 4884699ce097STejun Heo pwq_adjust_max_active(pwq); 4885dcd989cbSTejun Heo 4886a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 4887dcd989cbSTejun Heo } 4888dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 4889dcd989cbSTejun Heo 4890dcd989cbSTejun Heo /** 489127d4ee03SLukas Wunner * current_work - retrieve %current task's work struct 489227d4ee03SLukas Wunner * 489327d4ee03SLukas Wunner * Determine if %current task is a workqueue worker and what it's working on. 489427d4ee03SLukas Wunner * Useful to find out the context that the %current task is running in. 489527d4ee03SLukas Wunner * 489627d4ee03SLukas Wunner * Return: work struct if %current task is a workqueue worker, %NULL otherwise. 489727d4ee03SLukas Wunner */ 489827d4ee03SLukas Wunner struct work_struct *current_work(void) 489927d4ee03SLukas Wunner { 490027d4ee03SLukas Wunner struct worker *worker = current_wq_worker(); 490127d4ee03SLukas Wunner 490227d4ee03SLukas Wunner return worker ? worker->current_work : NULL; 490327d4ee03SLukas Wunner } 490427d4ee03SLukas Wunner EXPORT_SYMBOL(current_work); 490527d4ee03SLukas Wunner 490627d4ee03SLukas Wunner /** 4907e6267616STejun Heo * current_is_workqueue_rescuer - is %current workqueue rescuer? 4908e6267616STejun Heo * 4909e6267616STejun Heo * Determine whether %current is a workqueue rescuer. Can be used from 4910e6267616STejun Heo * work functions to determine whether it's being run off the rescuer task. 4911d185af30SYacine Belkadi * 4912d185af30SYacine Belkadi * Return: %true if %current is a workqueue rescuer. %false otherwise. 4913e6267616STejun Heo */ 4914e6267616STejun Heo bool current_is_workqueue_rescuer(void) 4915e6267616STejun Heo { 4916e6267616STejun Heo struct worker *worker = current_wq_worker(); 4917e6267616STejun Heo 49186a092dfdSLai Jiangshan return worker && worker->rescue_wq; 4919e6267616STejun Heo } 4920e6267616STejun Heo 4921e6267616STejun Heo /** 4922dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 4923dcd989cbSTejun Heo * @cpu: CPU in question 4924dcd989cbSTejun Heo * @wq: target workqueue 4925dcd989cbSTejun Heo * 4926dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 4927dcd989cbSTejun Heo * no synchronization around this function and the test result is 4928dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4929dcd989cbSTejun Heo * 4930d3251859STejun Heo * If @cpu is WORK_CPU_UNBOUND, the test is performed on the local CPU. 4931636b927eSTejun Heo * 4932636b927eSTejun Heo * With the exception of ordered workqueues, all workqueues have per-cpu 4933636b927eSTejun Heo * pool_workqueues, each with its own congested state. A workqueue being 4934636b927eSTejun Heo * congested on one CPU doesn't mean that the workqueue is contested on any 4935636b927eSTejun Heo * other CPUs. 4936d3251859STejun Heo * 4937d185af30SYacine Belkadi * Return: 4938dcd989cbSTejun Heo * %true if congested, %false otherwise. 4939dcd989cbSTejun Heo */ 4940d84ff051STejun Heo bool workqueue_congested(int cpu, struct workqueue_struct *wq) 4941dcd989cbSTejun Heo { 49427fb98ea7STejun Heo struct pool_workqueue *pwq; 494376af4d93STejun Heo bool ret; 494476af4d93STejun Heo 494524acfb71SThomas Gleixner rcu_read_lock(); 494624acfb71SThomas Gleixner preempt_disable(); 49477fb98ea7STejun Heo 4948d3251859STejun Heo if (cpu == WORK_CPU_UNBOUND) 4949d3251859STejun Heo cpu = smp_processor_id(); 4950d3251859STejun Heo 4951687a9aa5STejun Heo pwq = *per_cpu_ptr(wq->cpu_pwq, cpu); 4952f97a4a1aSLai Jiangshan ret = !list_empty(&pwq->inactive_works); 4953636b927eSTejun Heo 495424acfb71SThomas Gleixner preempt_enable(); 495524acfb71SThomas Gleixner rcu_read_unlock(); 495676af4d93STejun Heo 495776af4d93STejun Heo return ret; 4958dcd989cbSTejun Heo } 4959dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 4960dcd989cbSTejun Heo 4961dcd989cbSTejun Heo /** 4962dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 4963dcd989cbSTejun Heo * @work: the work to be tested 4964dcd989cbSTejun Heo * 4965dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 4966dcd989cbSTejun Heo * synchronization around this function and the test result is 4967dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 4968dcd989cbSTejun Heo * 4969d185af30SYacine Belkadi * Return: 4970dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 4971dcd989cbSTejun Heo */ 4972dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 4973dcd989cbSTejun Heo { 4974fa1b54e6STejun Heo struct worker_pool *pool; 4975dcd989cbSTejun Heo unsigned long flags; 4976dcd989cbSTejun Heo unsigned int ret = 0; 4977dcd989cbSTejun Heo 4978dcd989cbSTejun Heo if (work_pending(work)) 4979dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 4980038366c5SLai Jiangshan 498124acfb71SThomas Gleixner rcu_read_lock(); 4982fa1b54e6STejun Heo pool = get_work_pool(work); 4983038366c5SLai Jiangshan if (pool) { 4984a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 4985c9e7cf27STejun Heo if (find_worker_executing_work(pool, work)) 4986dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 4987a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 4988038366c5SLai Jiangshan } 498924acfb71SThomas Gleixner rcu_read_unlock(); 4990dcd989cbSTejun Heo 4991dcd989cbSTejun Heo return ret; 4992dcd989cbSTejun Heo } 4993dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 4994dcd989cbSTejun Heo 49953d1cb205STejun Heo /** 49963d1cb205STejun Heo * set_worker_desc - set description for the current work item 49973d1cb205STejun Heo * @fmt: printf-style format string 49983d1cb205STejun Heo * @...: arguments for the format string 49993d1cb205STejun Heo * 50003d1cb205STejun Heo * This function can be called by a running work function to describe what 50013d1cb205STejun Heo * the work item is about. If the worker task gets dumped, this 50023d1cb205STejun Heo * information will be printed out together to help debugging. The 50033d1cb205STejun Heo * description can be at most WORKER_DESC_LEN including the trailing '\0'. 50043d1cb205STejun Heo */ 50053d1cb205STejun Heo void set_worker_desc(const char *fmt, ...) 50063d1cb205STejun Heo { 50073d1cb205STejun Heo struct worker *worker = current_wq_worker(); 50083d1cb205STejun Heo va_list args; 50093d1cb205STejun Heo 50103d1cb205STejun Heo if (worker) { 50113d1cb205STejun Heo va_start(args, fmt); 50123d1cb205STejun Heo vsnprintf(worker->desc, sizeof(worker->desc), fmt, args); 50133d1cb205STejun Heo va_end(args); 50143d1cb205STejun Heo } 50153d1cb205STejun Heo } 50165c750d58SSteffen Maier EXPORT_SYMBOL_GPL(set_worker_desc); 50173d1cb205STejun Heo 50183d1cb205STejun Heo /** 50193d1cb205STejun Heo * print_worker_info - print out worker information and description 50203d1cb205STejun Heo * @log_lvl: the log level to use when printing 50213d1cb205STejun Heo * @task: target task 50223d1cb205STejun Heo * 50233d1cb205STejun Heo * If @task is a worker and currently executing a work item, print out the 50243d1cb205STejun Heo * name of the workqueue being serviced and worker description set with 50253d1cb205STejun Heo * set_worker_desc() by the currently executing work item. 50263d1cb205STejun Heo * 50273d1cb205STejun Heo * This function can be safely called on any task as long as the 50283d1cb205STejun Heo * task_struct itself is accessible. While safe, this function isn't 50293d1cb205STejun Heo * synchronized and may print out mixups or garbages of limited length. 50303d1cb205STejun Heo */ 50313d1cb205STejun Heo void print_worker_info(const char *log_lvl, struct task_struct *task) 50323d1cb205STejun Heo { 50333d1cb205STejun Heo work_func_t *fn = NULL; 50343d1cb205STejun Heo char name[WQ_NAME_LEN] = { }; 50353d1cb205STejun Heo char desc[WORKER_DESC_LEN] = { }; 50363d1cb205STejun Heo struct pool_workqueue *pwq = NULL; 50373d1cb205STejun Heo struct workqueue_struct *wq = NULL; 50383d1cb205STejun Heo struct worker *worker; 50393d1cb205STejun Heo 50403d1cb205STejun Heo if (!(task->flags & PF_WQ_WORKER)) 50413d1cb205STejun Heo return; 50423d1cb205STejun Heo 50433d1cb205STejun Heo /* 50443d1cb205STejun Heo * This function is called without any synchronization and @task 50453d1cb205STejun Heo * could be in any state. Be careful with dereferences. 50463d1cb205STejun Heo */ 5047e700591aSPetr Mladek worker = kthread_probe_data(task); 50483d1cb205STejun Heo 50493d1cb205STejun Heo /* 50508bf89593STejun Heo * Carefully copy the associated workqueue's workfn, name and desc. 50518bf89593STejun Heo * Keep the original last '\0' in case the original is garbage. 50523d1cb205STejun Heo */ 5053fe557319SChristoph Hellwig copy_from_kernel_nofault(&fn, &worker->current_func, sizeof(fn)); 5054fe557319SChristoph Hellwig copy_from_kernel_nofault(&pwq, &worker->current_pwq, sizeof(pwq)); 5055fe557319SChristoph Hellwig copy_from_kernel_nofault(&wq, &pwq->wq, sizeof(wq)); 5056fe557319SChristoph Hellwig copy_from_kernel_nofault(name, wq->name, sizeof(name) - 1); 5057fe557319SChristoph Hellwig copy_from_kernel_nofault(desc, worker->desc, sizeof(desc) - 1); 50583d1cb205STejun Heo 50593d1cb205STejun Heo if (fn || name[0] || desc[0]) { 5060d75f773cSSakari Ailus printk("%sWorkqueue: %s %ps", log_lvl, name, fn); 50618bf89593STejun Heo if (strcmp(name, desc)) 50623d1cb205STejun Heo pr_cont(" (%s)", desc); 50633d1cb205STejun Heo pr_cont("\n"); 50643d1cb205STejun Heo } 50653d1cb205STejun Heo } 50663d1cb205STejun Heo 50673494fc30STejun Heo static void pr_cont_pool_info(struct worker_pool *pool) 50683494fc30STejun Heo { 50693494fc30STejun Heo pr_cont(" cpus=%*pbl", nr_cpumask_bits, pool->attrs->cpumask); 50703494fc30STejun Heo if (pool->node != NUMA_NO_NODE) 50713494fc30STejun Heo pr_cont(" node=%d", pool->node); 50723494fc30STejun Heo pr_cont(" flags=0x%x nice=%d", pool->flags, pool->attrs->nice); 50733494fc30STejun Heo } 50743494fc30STejun Heo 5075c76feb0dSPaul E. McKenney struct pr_cont_work_struct { 5076c76feb0dSPaul E. McKenney bool comma; 5077c76feb0dSPaul E. McKenney work_func_t func; 5078c76feb0dSPaul E. McKenney long ctr; 5079c76feb0dSPaul E. McKenney }; 5080c76feb0dSPaul E. McKenney 5081c76feb0dSPaul E. McKenney static void pr_cont_work_flush(bool comma, work_func_t func, struct pr_cont_work_struct *pcwsp) 5082c76feb0dSPaul E. McKenney { 5083c76feb0dSPaul E. McKenney if (!pcwsp->ctr) 5084c76feb0dSPaul E. McKenney goto out_record; 5085c76feb0dSPaul E. McKenney if (func == pcwsp->func) { 5086c76feb0dSPaul E. McKenney pcwsp->ctr++; 5087c76feb0dSPaul E. McKenney return; 5088c76feb0dSPaul E. McKenney } 5089c76feb0dSPaul E. McKenney if (pcwsp->ctr == 1) 5090c76feb0dSPaul E. McKenney pr_cont("%s %ps", pcwsp->comma ? "," : "", pcwsp->func); 5091c76feb0dSPaul E. McKenney else 5092c76feb0dSPaul E. McKenney pr_cont("%s %ld*%ps", pcwsp->comma ? "," : "", pcwsp->ctr, pcwsp->func); 5093c76feb0dSPaul E. McKenney pcwsp->ctr = 0; 5094c76feb0dSPaul E. McKenney out_record: 5095c76feb0dSPaul E. McKenney if ((long)func == -1L) 5096c76feb0dSPaul E. McKenney return; 5097c76feb0dSPaul E. McKenney pcwsp->comma = comma; 5098c76feb0dSPaul E. McKenney pcwsp->func = func; 5099c76feb0dSPaul E. McKenney pcwsp->ctr = 1; 5100c76feb0dSPaul E. McKenney } 5101c76feb0dSPaul E. McKenney 5102c76feb0dSPaul E. McKenney static void pr_cont_work(bool comma, struct work_struct *work, struct pr_cont_work_struct *pcwsp) 51033494fc30STejun Heo { 51043494fc30STejun Heo if (work->func == wq_barrier_func) { 51053494fc30STejun Heo struct wq_barrier *barr; 51063494fc30STejun Heo 51073494fc30STejun Heo barr = container_of(work, struct wq_barrier, work); 51083494fc30STejun Heo 5109c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 51103494fc30STejun Heo pr_cont("%s BAR(%d)", comma ? "," : "", 51113494fc30STejun Heo task_pid_nr(barr->task)); 51123494fc30STejun Heo } else { 5113c76feb0dSPaul E. McKenney if (!comma) 5114c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1, pcwsp); 5115c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, work->func, pcwsp); 51163494fc30STejun Heo } 51173494fc30STejun Heo } 51183494fc30STejun Heo 51193494fc30STejun Heo static void show_pwq(struct pool_workqueue *pwq) 51203494fc30STejun Heo { 5121c76feb0dSPaul E. McKenney struct pr_cont_work_struct pcws = { .ctr = 0, }; 51223494fc30STejun Heo struct worker_pool *pool = pwq->pool; 51233494fc30STejun Heo struct work_struct *work; 51243494fc30STejun Heo struct worker *worker; 51253494fc30STejun Heo bool has_in_flight = false, has_pending = false; 51263494fc30STejun Heo int bkt; 51273494fc30STejun Heo 51283494fc30STejun Heo pr_info(" pwq %d:", pool->id); 51293494fc30STejun Heo pr_cont_pool_info(pool); 51303494fc30STejun Heo 5131e66b39afSTejun Heo pr_cont(" active=%d/%d refcnt=%d%s\n", 5132e66b39afSTejun Heo pwq->nr_active, pwq->max_active, pwq->refcnt, 51333494fc30STejun Heo !list_empty(&pwq->mayday_node) ? " MAYDAY" : ""); 51343494fc30STejun Heo 51353494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51363494fc30STejun Heo if (worker->current_pwq == pwq) { 51373494fc30STejun Heo has_in_flight = true; 51383494fc30STejun Heo break; 51393494fc30STejun Heo } 51403494fc30STejun Heo } 51413494fc30STejun Heo if (has_in_flight) { 51423494fc30STejun Heo bool comma = false; 51433494fc30STejun Heo 51443494fc30STejun Heo pr_info(" in-flight:"); 51453494fc30STejun Heo hash_for_each(pool->busy_hash, bkt, worker, hentry) { 51463494fc30STejun Heo if (worker->current_pwq != pwq) 51473494fc30STejun Heo continue; 51483494fc30STejun Heo 5149d75f773cSSakari Ailus pr_cont("%s %d%s:%ps", comma ? "," : "", 51503494fc30STejun Heo task_pid_nr(worker->task), 515130ae2fc0STejun Heo worker->rescue_wq ? "(RESCUER)" : "", 51523494fc30STejun Heo worker->current_func); 51533494fc30STejun Heo list_for_each_entry(work, &worker->scheduled, entry) 5154c76feb0dSPaul E. McKenney pr_cont_work(false, work, &pcws); 5155c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51563494fc30STejun Heo comma = true; 51573494fc30STejun Heo } 51583494fc30STejun Heo pr_cont("\n"); 51593494fc30STejun Heo } 51603494fc30STejun Heo 51613494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51623494fc30STejun Heo if (get_work_pwq(work) == pwq) { 51633494fc30STejun Heo has_pending = true; 51643494fc30STejun Heo break; 51653494fc30STejun Heo } 51663494fc30STejun Heo } 51673494fc30STejun Heo if (has_pending) { 51683494fc30STejun Heo bool comma = false; 51693494fc30STejun Heo 51703494fc30STejun Heo pr_info(" pending:"); 51713494fc30STejun Heo list_for_each_entry(work, &pool->worklist, entry) { 51723494fc30STejun Heo if (get_work_pwq(work) != pwq) 51733494fc30STejun Heo continue; 51743494fc30STejun Heo 5175c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51763494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51773494fc30STejun Heo } 5178c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51793494fc30STejun Heo pr_cont("\n"); 51803494fc30STejun Heo } 51813494fc30STejun Heo 5182f97a4a1aSLai Jiangshan if (!list_empty(&pwq->inactive_works)) { 51833494fc30STejun Heo bool comma = false; 51843494fc30STejun Heo 5185f97a4a1aSLai Jiangshan pr_info(" inactive:"); 5186f97a4a1aSLai Jiangshan list_for_each_entry(work, &pwq->inactive_works, entry) { 5187c76feb0dSPaul E. McKenney pr_cont_work(comma, work, &pcws); 51883494fc30STejun Heo comma = !(*work_data_bits(work) & WORK_STRUCT_LINKED); 51893494fc30STejun Heo } 5190c76feb0dSPaul E. McKenney pr_cont_work_flush(comma, (work_func_t)-1L, &pcws); 51913494fc30STejun Heo pr_cont("\n"); 51923494fc30STejun Heo } 51933494fc30STejun Heo } 51943494fc30STejun Heo 51953494fc30STejun Heo /** 519655df0933SImran Khan * show_one_workqueue - dump state of specified workqueue 519755df0933SImran Khan * @wq: workqueue whose state will be printed 51983494fc30STejun Heo */ 519955df0933SImran Khan void show_one_workqueue(struct workqueue_struct *wq) 52003494fc30STejun Heo { 52013494fc30STejun Heo struct pool_workqueue *pwq; 52023494fc30STejun Heo bool idle = true; 520355df0933SImran Khan unsigned long flags; 52043494fc30STejun Heo 52053494fc30STejun Heo for_each_pwq(pwq, wq) { 5206f97a4a1aSLai Jiangshan if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 52073494fc30STejun Heo idle = false; 52083494fc30STejun Heo break; 52093494fc30STejun Heo } 52103494fc30STejun Heo } 521155df0933SImran Khan if (idle) /* Nothing to print for idle workqueue */ 521255df0933SImran Khan return; 52133494fc30STejun Heo 52143494fc30STejun Heo pr_info("workqueue %s: flags=0x%x\n", wq->name, wq->flags); 52153494fc30STejun Heo 52163494fc30STejun Heo for_each_pwq(pwq, wq) { 5217a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pwq->pool->lock, flags); 521857116ce1SJohan Hovold if (pwq->nr_active || !list_empty(&pwq->inactive_works)) { 521957116ce1SJohan Hovold /* 522057116ce1SJohan Hovold * Defer printing to avoid deadlocks in console 522157116ce1SJohan Hovold * drivers that queue work while holding locks 522257116ce1SJohan Hovold * also taken in their write paths. 522357116ce1SJohan Hovold */ 522457116ce1SJohan Hovold printk_deferred_enter(); 52253494fc30STejun Heo show_pwq(pwq); 522657116ce1SJohan Hovold printk_deferred_exit(); 522757116ce1SJohan Hovold } 5228a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pwq->pool->lock, flags); 522962635ea8SSergey Senozhatsky /* 523062635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 523155df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 523262635ea8SSergey Senozhatsky * hard lockup. 523362635ea8SSergey Senozhatsky */ 523462635ea8SSergey Senozhatsky touch_nmi_watchdog(); 52353494fc30STejun Heo } 523655df0933SImran Khan 52373494fc30STejun Heo } 52383494fc30STejun Heo 523955df0933SImran Khan /** 524055df0933SImran Khan * show_one_worker_pool - dump state of specified worker pool 524155df0933SImran Khan * @pool: worker pool whose state will be printed 524255df0933SImran Khan */ 524355df0933SImran Khan static void show_one_worker_pool(struct worker_pool *pool) 524455df0933SImran Khan { 52453494fc30STejun Heo struct worker *worker; 52463494fc30STejun Heo bool first = true; 524755df0933SImran Khan unsigned long flags; 5248335a42ebSPetr Mladek unsigned long hung = 0; 52493494fc30STejun Heo 5250a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irqsave(&pool->lock, flags); 52513494fc30STejun Heo if (pool->nr_workers == pool->nr_idle) 52523494fc30STejun Heo goto next_pool; 5253335a42ebSPetr Mladek 5254335a42ebSPetr Mladek /* How long the first pending work is waiting for a worker. */ 5255335a42ebSPetr Mladek if (!list_empty(&pool->worklist)) 5256335a42ebSPetr Mladek hung = jiffies_to_msecs(jiffies - pool->watchdog_ts) / 1000; 5257335a42ebSPetr Mladek 525857116ce1SJohan Hovold /* 525957116ce1SJohan Hovold * Defer printing to avoid deadlocks in console drivers that 526057116ce1SJohan Hovold * queue work while holding locks also taken in their write 526157116ce1SJohan Hovold * paths. 526257116ce1SJohan Hovold */ 526357116ce1SJohan Hovold printk_deferred_enter(); 52643494fc30STejun Heo pr_info("pool %d:", pool->id); 52653494fc30STejun Heo pr_cont_pool_info(pool); 5266335a42ebSPetr Mladek pr_cont(" hung=%lus workers=%d", hung, pool->nr_workers); 52673494fc30STejun Heo if (pool->manager) 52683494fc30STejun Heo pr_cont(" manager: %d", 52693494fc30STejun Heo task_pid_nr(pool->manager->task)); 52703494fc30STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 52713494fc30STejun Heo pr_cont(" %s%d", first ? "idle: " : "", 52723494fc30STejun Heo task_pid_nr(worker->task)); 52733494fc30STejun Heo first = false; 52743494fc30STejun Heo } 52753494fc30STejun Heo pr_cont("\n"); 527657116ce1SJohan Hovold printk_deferred_exit(); 52773494fc30STejun Heo next_pool: 5278a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irqrestore(&pool->lock, flags); 527962635ea8SSergey Senozhatsky /* 528062635ea8SSergey Senozhatsky * We could be printing a lot from atomic context, e.g. 528155df0933SImran Khan * sysrq-t -> show_all_workqueues(). Avoid triggering 528262635ea8SSergey Senozhatsky * hard lockup. 528362635ea8SSergey Senozhatsky */ 528462635ea8SSergey Senozhatsky touch_nmi_watchdog(); 528555df0933SImran Khan 52863494fc30STejun Heo } 52873494fc30STejun Heo 528855df0933SImran Khan /** 528955df0933SImran Khan * show_all_workqueues - dump workqueue state 529055df0933SImran Khan * 5291704bc669SJungseung Lee * Called from a sysrq handler and prints out all busy workqueues and pools. 529255df0933SImran Khan */ 529355df0933SImran Khan void show_all_workqueues(void) 529455df0933SImran Khan { 529555df0933SImran Khan struct workqueue_struct *wq; 529655df0933SImran Khan struct worker_pool *pool; 529755df0933SImran Khan int pi; 529855df0933SImran Khan 529955df0933SImran Khan rcu_read_lock(); 530055df0933SImran Khan 530155df0933SImran Khan pr_info("Showing busy workqueues and worker pools:\n"); 530255df0933SImran Khan 530355df0933SImran Khan list_for_each_entry_rcu(wq, &workqueues, list) 530455df0933SImran Khan show_one_workqueue(wq); 530555df0933SImran Khan 530655df0933SImran Khan for_each_pool(pool, pi) 530755df0933SImran Khan show_one_worker_pool(pool); 530855df0933SImran Khan 530924acfb71SThomas Gleixner rcu_read_unlock(); 53103494fc30STejun Heo } 53113494fc30STejun Heo 5312704bc669SJungseung Lee /** 5313704bc669SJungseung Lee * show_freezable_workqueues - dump freezable workqueue state 5314704bc669SJungseung Lee * 5315704bc669SJungseung Lee * Called from try_to_freeze_tasks() and prints out all freezable workqueues 5316704bc669SJungseung Lee * still busy. 5317704bc669SJungseung Lee */ 5318704bc669SJungseung Lee void show_freezable_workqueues(void) 5319704bc669SJungseung Lee { 5320704bc669SJungseung Lee struct workqueue_struct *wq; 5321704bc669SJungseung Lee 5322704bc669SJungseung Lee rcu_read_lock(); 5323704bc669SJungseung Lee 5324704bc669SJungseung Lee pr_info("Showing freezable workqueues that are still busy:\n"); 5325704bc669SJungseung Lee 5326704bc669SJungseung Lee list_for_each_entry_rcu(wq, &workqueues, list) { 5327704bc669SJungseung Lee if (!(wq->flags & WQ_FREEZABLE)) 5328704bc669SJungseung Lee continue; 5329704bc669SJungseung Lee show_one_workqueue(wq); 5330704bc669SJungseung Lee } 5331704bc669SJungseung Lee 5332704bc669SJungseung Lee rcu_read_unlock(); 5333704bc669SJungseung Lee } 5334704bc669SJungseung Lee 53356b59808bSTejun Heo /* used to show worker information through /proc/PID/{comm,stat,status} */ 53366b59808bSTejun Heo void wq_worker_comm(char *buf, size_t size, struct task_struct *task) 53376b59808bSTejun Heo { 53386b59808bSTejun Heo int off; 53396b59808bSTejun Heo 53406b59808bSTejun Heo /* always show the actual comm */ 53416b59808bSTejun Heo off = strscpy(buf, task->comm, size); 53426b59808bSTejun Heo if (off < 0) 53436b59808bSTejun Heo return; 53446b59808bSTejun Heo 5345197f6accSTejun Heo /* stabilize PF_WQ_WORKER and worker pool association */ 53466b59808bSTejun Heo mutex_lock(&wq_pool_attach_mutex); 53476b59808bSTejun Heo 5348197f6accSTejun Heo if (task->flags & PF_WQ_WORKER) { 5349197f6accSTejun Heo struct worker *worker = kthread_data(task); 5350197f6accSTejun Heo struct worker_pool *pool = worker->pool; 53516b59808bSTejun Heo 53526b59808bSTejun Heo if (pool) { 5353a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 53546b59808bSTejun Heo /* 5355197f6accSTejun Heo * ->desc tracks information (wq name or 5356197f6accSTejun Heo * set_worker_desc()) for the latest execution. If 5357197f6accSTejun Heo * current, prepend '+', otherwise '-'. 53586b59808bSTejun Heo */ 53596b59808bSTejun Heo if (worker->desc[0] != '\0') { 53606b59808bSTejun Heo if (worker->current_work) 53616b59808bSTejun Heo scnprintf(buf + off, size - off, "+%s", 53626b59808bSTejun Heo worker->desc); 53636b59808bSTejun Heo else 53646b59808bSTejun Heo scnprintf(buf + off, size - off, "-%s", 53656b59808bSTejun Heo worker->desc); 53666b59808bSTejun Heo } 5367a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 53686b59808bSTejun Heo } 5369197f6accSTejun Heo } 53706b59808bSTejun Heo 53716b59808bSTejun Heo mutex_unlock(&wq_pool_attach_mutex); 53726b59808bSTejun Heo } 53736b59808bSTejun Heo 537466448bc2SMathieu Malaterre #ifdef CONFIG_SMP 537566448bc2SMathieu Malaterre 5376db7bccf4STejun Heo /* 5377db7bccf4STejun Heo * CPU hotplug. 5378db7bccf4STejun Heo * 5379e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 5380112202d9STejun Heo * are a lot of assumptions on strong associations among work, pwq and 5381706026c2STejun Heo * pool which make migrating pending and scheduled works very 5382e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 538394cf58bbSTejun Heo * worker pools serve mix of short, long and very long running works making 5384e22bee78STejun Heo * blocked draining impractical. 5385e22bee78STejun Heo * 538624647570STejun Heo * This is solved by allowing the pools to be disassociated from the CPU 5387628c78e7STejun Heo * running as an unbound one and allowing it to be reattached later if the 5388628c78e7STejun Heo * cpu comes back online. 5389db7bccf4STejun Heo */ 5390db7bccf4STejun Heo 5391e8b3f8dbSLai Jiangshan static void unbind_workers(int cpu) 5392db7bccf4STejun Heo { 53934ce62e9eSTejun Heo struct worker_pool *pool; 5394db7bccf4STejun Heo struct worker *worker; 5395db7bccf4STejun Heo 5396f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 53971258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 5398a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5399e22bee78STejun Heo 5400f2d5a0eeSTejun Heo /* 540192f9c5c4SLai Jiangshan * We've blocked all attach/detach operations. Make all workers 540294cf58bbSTejun Heo * unbound and set DISASSOCIATED. Before this, all workers 540311b45b0bSLai Jiangshan * must be on the cpu. After this, they may become diasporas. 5404b4ac9384SLai Jiangshan * And the preemption disabled section in their sched callbacks 5405b4ac9384SLai Jiangshan * are guaranteed to see WORKER_UNBOUND since the code here 5406b4ac9384SLai Jiangshan * is on the same cpu. 5407f2d5a0eeSTejun Heo */ 5408da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5409403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 5410db7bccf4STejun Heo 541124647570STejun Heo pool->flags |= POOL_DISASSOCIATED; 5412f2d5a0eeSTejun Heo 5413e22bee78STejun Heo /* 5414989442d7SLai Jiangshan * The handling of nr_running in sched callbacks are disabled 5415989442d7SLai Jiangshan * now. Zap nr_running. After this, nr_running stays zero and 5416989442d7SLai Jiangshan * need_more_worker() and keep_working() are always true as 5417989442d7SLai Jiangshan * long as the worklist is not empty. This pool now behaves as 5418989442d7SLai Jiangshan * an unbound (in terms of concurrency management) pool which 5419eb283428SLai Jiangshan * are served by workers tied to the pool. 5420e22bee78STejun Heo */ 5421bc35f7efSLai Jiangshan pool->nr_running = 0; 5422eb283428SLai Jiangshan 5423eb283428SLai Jiangshan /* 5424eb283428SLai Jiangshan * With concurrency management just turned off, a busy 5425eb283428SLai Jiangshan * worker blocking could lead to lengthy stalls. Kick off 5426eb283428SLai Jiangshan * unbound chain execution of currently pending work items. 5427eb283428SLai Jiangshan */ 54280219a352STejun Heo kick_pool(pool); 5429989442d7SLai Jiangshan 5430a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5431989442d7SLai Jiangshan 5432793777bcSValentin Schneider for_each_pool_worker(worker, pool) 5433793777bcSValentin Schneider unbind_worker(worker); 5434989442d7SLai Jiangshan 5435989442d7SLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 5436eb283428SLai Jiangshan } 5437db7bccf4STejun Heo } 5438db7bccf4STejun Heo 5439bd7c089eSTejun Heo /** 5440bd7c089eSTejun Heo * rebind_workers - rebind all workers of a pool to the associated CPU 5441bd7c089eSTejun Heo * @pool: pool of interest 5442bd7c089eSTejun Heo * 5443a9ab775bSTejun Heo * @pool->cpu is coming online. Rebind all workers to the CPU. 5444bd7c089eSTejun Heo */ 5445bd7c089eSTejun Heo static void rebind_workers(struct worker_pool *pool) 5446bd7c089eSTejun Heo { 5447a9ab775bSTejun Heo struct worker *worker; 5448bd7c089eSTejun Heo 54491258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 5450bd7c089eSTejun Heo 5451bd7c089eSTejun Heo /* 5452a9ab775bSTejun Heo * Restore CPU affinity of all workers. As all idle workers should 5453a9ab775bSTejun Heo * be on the run-queue of the associated CPU before any local 5454402dd89dSShailendra Verma * wake-ups for concurrency management happen, restore CPU affinity 5455a9ab775bSTejun Heo * of all workers first and then clear UNBOUND. As we're called 5456a9ab775bSTejun Heo * from CPU_ONLINE, the following shouldn't fail. 5457bd7c089eSTejun Heo */ 5458c63a2e52SValentin Schneider for_each_pool_worker(worker, pool) { 5459c63a2e52SValentin Schneider kthread_set_per_cpu(worker->task, pool->cpu); 5460c63a2e52SValentin Schneider WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, 54619546b29eSTejun Heo pool_allowed_cpus(pool)) < 0); 5462c63a2e52SValentin Schneider } 5463a9ab775bSTejun Heo 5464a9b8a985SSebastian Andrzej Siewior raw_spin_lock_irq(&pool->lock); 5465f7c17d26SWanpeng Li 54663de5e884SLai Jiangshan pool->flags &= ~POOL_DISASSOCIATED; 5467a9ab775bSTejun Heo 5468da028469SLai Jiangshan for_each_pool_worker(worker, pool) { 5469a9ab775bSTejun Heo unsigned int worker_flags = worker->flags; 5470a9ab775bSTejun Heo 5471a9ab775bSTejun Heo /* 5472a9ab775bSTejun Heo * We want to clear UNBOUND but can't directly call 5473a9ab775bSTejun Heo * worker_clr_flags() or adjust nr_running. Atomically 5474a9ab775bSTejun Heo * replace UNBOUND with another NOT_RUNNING flag REBOUND. 5475a9ab775bSTejun Heo * @worker will clear REBOUND using worker_clr_flags() when 5476a9ab775bSTejun Heo * it initiates the next execution cycle thus restoring 5477a9ab775bSTejun Heo * concurrency management. Note that when or whether 5478a9ab775bSTejun Heo * @worker clears REBOUND doesn't affect correctness. 5479a9ab775bSTejun Heo * 5480c95491edSMark Rutland * WRITE_ONCE() is necessary because @worker->flags may be 5481a9ab775bSTejun Heo * tested without holding any lock in 54826d25be57SThomas Gleixner * wq_worker_running(). Without it, NOT_RUNNING test may 5483a9ab775bSTejun Heo * fail incorrectly leading to premature concurrency 5484a9ab775bSTejun Heo * management operations. 5485bd7c089eSTejun Heo */ 5486a9ab775bSTejun Heo WARN_ON_ONCE(!(worker_flags & WORKER_UNBOUND)); 5487a9ab775bSTejun Heo worker_flags |= WORKER_REBOUND; 5488a9ab775bSTejun Heo worker_flags &= ~WORKER_UNBOUND; 5489c95491edSMark Rutland WRITE_ONCE(worker->flags, worker_flags); 5490bd7c089eSTejun Heo } 5491a9ab775bSTejun Heo 5492a9b8a985SSebastian Andrzej Siewior raw_spin_unlock_irq(&pool->lock); 5493bd7c089eSTejun Heo } 5494bd7c089eSTejun Heo 54957dbc725eSTejun Heo /** 54967dbc725eSTejun Heo * restore_unbound_workers_cpumask - restore cpumask of unbound workers 54977dbc725eSTejun Heo * @pool: unbound pool of interest 54987dbc725eSTejun Heo * @cpu: the CPU which is coming up 54997dbc725eSTejun Heo * 55007dbc725eSTejun Heo * An unbound pool may end up with a cpumask which doesn't have any online 55017dbc725eSTejun Heo * CPUs. When a worker of such pool get scheduled, the scheduler resets 55027dbc725eSTejun Heo * its cpus_allowed. If @cpu is in @pool's cpumask which didn't have any 55037dbc725eSTejun Heo * online CPU before, cpus_allowed of all its workers should be restored. 55047dbc725eSTejun Heo */ 55057dbc725eSTejun Heo static void restore_unbound_workers_cpumask(struct worker_pool *pool, int cpu) 55067dbc725eSTejun Heo { 55077dbc725eSTejun Heo static cpumask_t cpumask; 55087dbc725eSTejun Heo struct worker *worker; 55097dbc725eSTejun Heo 55101258fae7STejun Heo lockdep_assert_held(&wq_pool_attach_mutex); 55117dbc725eSTejun Heo 55127dbc725eSTejun Heo /* is @cpu allowed for @pool? */ 55137dbc725eSTejun Heo if (!cpumask_test_cpu(cpu, pool->attrs->cpumask)) 55147dbc725eSTejun Heo return; 55157dbc725eSTejun Heo 55167dbc725eSTejun Heo cpumask_and(&cpumask, pool->attrs->cpumask, cpu_online_mask); 55177dbc725eSTejun Heo 55187dbc725eSTejun Heo /* as we're called from CPU_ONLINE, the following shouldn't fail */ 5519da028469SLai Jiangshan for_each_pool_worker(worker, pool) 5520d945b5e9SPeter Zijlstra WARN_ON_ONCE(set_cpus_allowed_ptr(worker->task, &cpumask) < 0); 55217dbc725eSTejun Heo } 55227dbc725eSTejun Heo 55237ee681b2SThomas Gleixner int workqueue_prepare_cpu(unsigned int cpu) 55241da177e4SLinus Torvalds { 55254ce62e9eSTejun Heo struct worker_pool *pool; 55261da177e4SLinus Torvalds 5527f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 55283ce63377STejun Heo if (pool->nr_workers) 55293ce63377STejun Heo continue; 5530051e1850SLai Jiangshan if (!create_worker(pool)) 55317ee681b2SThomas Gleixner return -ENOMEM; 55323af24433SOleg Nesterov } 55337ee681b2SThomas Gleixner return 0; 55347ee681b2SThomas Gleixner } 55351da177e4SLinus Torvalds 55367ee681b2SThomas Gleixner int workqueue_online_cpu(unsigned int cpu) 55377ee681b2SThomas Gleixner { 55387ee681b2SThomas Gleixner struct worker_pool *pool; 55397ee681b2SThomas Gleixner struct workqueue_struct *wq; 55407ee681b2SThomas Gleixner int pi; 55417ee681b2SThomas Gleixner 554268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 55437dbc725eSTejun Heo 55447dbc725eSTejun Heo for_each_pool(pool, pi) { 55451258fae7STejun Heo mutex_lock(&wq_pool_attach_mutex); 554694cf58bbSTejun Heo 5547f05b558dSLai Jiangshan if (pool->cpu == cpu) 554894cf58bbSTejun Heo rebind_workers(pool); 5549f05b558dSLai Jiangshan else if (pool->cpu < 0) 55507dbc725eSTejun Heo restore_unbound_workers_cpumask(pool, cpu); 555194cf58bbSTejun Heo 55521258fae7STejun Heo mutex_unlock(&wq_pool_attach_mutex); 555394cf58bbSTejun Heo } 55547dbc725eSTejun Heo 5555fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 55564cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 555784193c07STejun Heo struct workqueue_attrs *attrs = wq->unbound_attrs; 555884193c07STejun Heo 555984193c07STejun Heo if (attrs) { 556084193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 55614cbfd3deSTejun Heo int tcpu; 55624cbfd3deSTejun Heo 556384193c07STejun Heo for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5564fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, true); 55654cbfd3deSTejun Heo } 55664cbfd3deSTejun Heo } 55674c16bd32STejun Heo 556868e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 55697ee681b2SThomas Gleixner return 0; 557065758202STejun Heo } 557165758202STejun Heo 55727ee681b2SThomas Gleixner int workqueue_offline_cpu(unsigned int cpu) 557365758202STejun Heo { 55744c16bd32STejun Heo struct workqueue_struct *wq; 55758db25e78STejun Heo 55764c16bd32STejun Heo /* unbinding per-cpu workers should happen on the local CPU */ 5577e8b3f8dbSLai Jiangshan if (WARN_ON(cpu != smp_processor_id())) 5578e8b3f8dbSLai Jiangshan return -1; 5579e8b3f8dbSLai Jiangshan 5580e8b3f8dbSLai Jiangshan unbind_workers(cpu); 55814c16bd32STejun Heo 5582fef59c9cSTejun Heo /* update pod affinity of unbound workqueues */ 55834c16bd32STejun Heo mutex_lock(&wq_pool_mutex); 55844cbfd3deSTejun Heo list_for_each_entry(wq, &workqueues, list) { 558584193c07STejun Heo struct workqueue_attrs *attrs = wq->unbound_attrs; 558684193c07STejun Heo 558784193c07STejun Heo if (attrs) { 558884193c07STejun Heo const struct wq_pod_type *pt = wqattrs_pod_type(attrs); 55894cbfd3deSTejun Heo int tcpu; 55904cbfd3deSTejun Heo 559184193c07STejun Heo for_each_cpu(tcpu, pt->pod_cpus[pt->cpu_pod[cpu]]) 5592fef59c9cSTejun Heo wq_update_pod(wq, tcpu, cpu, false); 55934cbfd3deSTejun Heo } 55944cbfd3deSTejun Heo } 55954c16bd32STejun Heo mutex_unlock(&wq_pool_mutex); 55964c16bd32STejun Heo 55977ee681b2SThomas Gleixner return 0; 559865758202STejun Heo } 559965758202STejun Heo 56002d3854a3SRusty Russell struct work_for_cpu { 5601ed48ece2STejun Heo struct work_struct work; 56022d3854a3SRusty Russell long (*fn)(void *); 56032d3854a3SRusty Russell void *arg; 56042d3854a3SRusty Russell long ret; 56052d3854a3SRusty Russell }; 56062d3854a3SRusty Russell 5607ed48ece2STejun Heo static void work_for_cpu_fn(struct work_struct *work) 56082d3854a3SRusty Russell { 5609ed48ece2STejun Heo struct work_for_cpu *wfc = container_of(work, struct work_for_cpu, work); 5610ed48ece2STejun Heo 56112d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 56122d3854a3SRusty Russell } 56132d3854a3SRusty Russell 56142d3854a3SRusty Russell /** 561522aceb31SAnna-Maria Gleixner * work_on_cpu - run a function in thread context on a particular cpu 56162d3854a3SRusty Russell * @cpu: the cpu to run on 56172d3854a3SRusty Russell * @fn: the function to run 56182d3854a3SRusty Russell * @arg: the function arg 56192d3854a3SRusty Russell * 562031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 56216b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 5622d185af30SYacine Belkadi * 5623d185af30SYacine Belkadi * Return: The value @fn returns. 56242d3854a3SRusty Russell */ 5625d84ff051STejun Heo long work_on_cpu(int cpu, long (*fn)(void *), void *arg) 56262d3854a3SRusty Russell { 5627ed48ece2STejun Heo struct work_for_cpu wfc = { .fn = fn, .arg = arg }; 56282d3854a3SRusty Russell 5629ed48ece2STejun Heo INIT_WORK_ONSTACK(&wfc.work, work_for_cpu_fn); 5630ed48ece2STejun Heo schedule_work_on(cpu, &wfc.work); 563112997d1aSBjorn Helgaas flush_work(&wfc.work); 5632440a1136SChuansheng Liu destroy_work_on_stack(&wfc.work); 56332d3854a3SRusty Russell return wfc.ret; 56342d3854a3SRusty Russell } 56352d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 56360e8d6a93SThomas Gleixner 56370e8d6a93SThomas Gleixner /** 56380e8d6a93SThomas Gleixner * work_on_cpu_safe - run a function in thread context on a particular cpu 56390e8d6a93SThomas Gleixner * @cpu: the cpu to run on 56400e8d6a93SThomas Gleixner * @fn: the function to run 56410e8d6a93SThomas Gleixner * @arg: the function argument 56420e8d6a93SThomas Gleixner * 56430e8d6a93SThomas Gleixner * Disables CPU hotplug and calls work_on_cpu(). The caller must not hold 56440e8d6a93SThomas Gleixner * any locks which would prevent @fn from completing. 56450e8d6a93SThomas Gleixner * 56460e8d6a93SThomas Gleixner * Return: The value @fn returns. 56470e8d6a93SThomas Gleixner */ 56480e8d6a93SThomas Gleixner long work_on_cpu_safe(int cpu, long (*fn)(void *), void *arg) 56490e8d6a93SThomas Gleixner { 56500e8d6a93SThomas Gleixner long ret = -ENODEV; 56510e8d6a93SThomas Gleixner 5652ffd8bea8SSebastian Andrzej Siewior cpus_read_lock(); 56530e8d6a93SThomas Gleixner if (cpu_online(cpu)) 56540e8d6a93SThomas Gleixner ret = work_on_cpu(cpu, fn, arg); 5655ffd8bea8SSebastian Andrzej Siewior cpus_read_unlock(); 56560e8d6a93SThomas Gleixner return ret; 56570e8d6a93SThomas Gleixner } 56580e8d6a93SThomas Gleixner EXPORT_SYMBOL_GPL(work_on_cpu_safe); 56592d3854a3SRusty Russell #endif /* CONFIG_SMP */ 56602d3854a3SRusty Russell 5661a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 5662e7577c50SRusty Russell 5663a0a1a5fdSTejun Heo /** 5664a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 5665a0a1a5fdSTejun Heo * 566658a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 5667f97a4a1aSLai Jiangshan * workqueues will queue new works to their inactive_works list instead of 5668706026c2STejun Heo * pool->worklist. 5669a0a1a5fdSTejun Heo * 5670a0a1a5fdSTejun Heo * CONTEXT: 5671a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5672a0a1a5fdSTejun Heo */ 5673a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 5674a0a1a5fdSTejun Heo { 567524b8a847STejun Heo struct workqueue_struct *wq; 567624b8a847STejun Heo struct pool_workqueue *pwq; 5677a0a1a5fdSTejun Heo 567868e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5679a0a1a5fdSTejun Heo 56806183c009STejun Heo WARN_ON_ONCE(workqueue_freezing); 5681a0a1a5fdSTejun Heo workqueue_freezing = true; 5682a0a1a5fdSTejun Heo 568324b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5684a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5685699ce097STejun Heo for_each_pwq(pwq, wq) 5686699ce097STejun Heo pwq_adjust_max_active(pwq); 5687a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 5688a1056305STejun Heo } 56895bcab335STejun Heo 569068e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5691a0a1a5fdSTejun Heo } 5692a0a1a5fdSTejun Heo 5693a0a1a5fdSTejun Heo /** 569458a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 5695a0a1a5fdSTejun Heo * 5696a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 5697a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 5698a0a1a5fdSTejun Heo * 5699a0a1a5fdSTejun Heo * CONTEXT: 570068e13a67SLai Jiangshan * Grabs and releases wq_pool_mutex. 5701a0a1a5fdSTejun Heo * 5702d185af30SYacine Belkadi * Return: 570358a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 570458a69cb4STejun Heo * is complete. 5705a0a1a5fdSTejun Heo */ 5706a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 5707a0a1a5fdSTejun Heo { 5708a0a1a5fdSTejun Heo bool busy = false; 570924b8a847STejun Heo struct workqueue_struct *wq; 571024b8a847STejun Heo struct pool_workqueue *pwq; 5711a0a1a5fdSTejun Heo 571268e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5713a0a1a5fdSTejun Heo 57146183c009STejun Heo WARN_ON_ONCE(!workqueue_freezing); 5715a0a1a5fdSTejun Heo 571624b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 571724b8a847STejun Heo if (!(wq->flags & WQ_FREEZABLE)) 571824b8a847STejun Heo continue; 5719a0a1a5fdSTejun Heo /* 5720a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 5721a0a1a5fdSTejun Heo * to peek without lock. 5722a0a1a5fdSTejun Heo */ 572324acfb71SThomas Gleixner rcu_read_lock(); 572424b8a847STejun Heo for_each_pwq(pwq, wq) { 57256183c009STejun Heo WARN_ON_ONCE(pwq->nr_active < 0); 5726112202d9STejun Heo if (pwq->nr_active) { 5727a0a1a5fdSTejun Heo busy = true; 572824acfb71SThomas Gleixner rcu_read_unlock(); 5729a0a1a5fdSTejun Heo goto out_unlock; 5730a0a1a5fdSTejun Heo } 5731a0a1a5fdSTejun Heo } 573224acfb71SThomas Gleixner rcu_read_unlock(); 5733a0a1a5fdSTejun Heo } 5734a0a1a5fdSTejun Heo out_unlock: 573568e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5736a0a1a5fdSTejun Heo return busy; 5737a0a1a5fdSTejun Heo } 5738a0a1a5fdSTejun Heo 5739a0a1a5fdSTejun Heo /** 5740a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 5741a0a1a5fdSTejun Heo * 5742a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 5743706026c2STejun Heo * frozen works are transferred to their respective pool worklists. 5744a0a1a5fdSTejun Heo * 5745a0a1a5fdSTejun Heo * CONTEXT: 5746a357fc03SLai Jiangshan * Grabs and releases wq_pool_mutex, wq->mutex and pool->lock's. 5747a0a1a5fdSTejun Heo */ 5748a0a1a5fdSTejun Heo void thaw_workqueues(void) 5749a0a1a5fdSTejun Heo { 575024b8a847STejun Heo struct workqueue_struct *wq; 575124b8a847STejun Heo struct pool_workqueue *pwq; 5752a0a1a5fdSTejun Heo 575368e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 5754a0a1a5fdSTejun Heo 5755a0a1a5fdSTejun Heo if (!workqueue_freezing) 5756a0a1a5fdSTejun Heo goto out_unlock; 5757a0a1a5fdSTejun Heo 575874b414eaSLai Jiangshan workqueue_freezing = false; 575924b8a847STejun Heo 576024b8a847STejun Heo /* restore max_active and repopulate worklist */ 576124b8a847STejun Heo list_for_each_entry(wq, &workqueues, list) { 5762a357fc03SLai Jiangshan mutex_lock(&wq->mutex); 5763699ce097STejun Heo for_each_pwq(pwq, wq) 5764699ce097STejun Heo pwq_adjust_max_active(pwq); 5765a357fc03SLai Jiangshan mutex_unlock(&wq->mutex); 576624b8a847STejun Heo } 576724b8a847STejun Heo 5768a0a1a5fdSTejun Heo out_unlock: 576968e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 5770a0a1a5fdSTejun Heo } 5771a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 5772a0a1a5fdSTejun Heo 577399c621efSLai Jiangshan static int workqueue_apply_unbound_cpumask(const cpumask_var_t unbound_cpumask) 5774042f7df1SLai Jiangshan { 5775042f7df1SLai Jiangshan LIST_HEAD(ctxs); 5776042f7df1SLai Jiangshan int ret = 0; 5777042f7df1SLai Jiangshan struct workqueue_struct *wq; 5778042f7df1SLai Jiangshan struct apply_wqattrs_ctx *ctx, *n; 5779042f7df1SLai Jiangshan 5780042f7df1SLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5781042f7df1SLai Jiangshan 5782042f7df1SLai Jiangshan list_for_each_entry(wq, &workqueues, list) { 5783042f7df1SLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 5784042f7df1SLai Jiangshan continue; 5785042f7df1SLai Jiangshan /* creating multiple pwqs breaks ordering guarantee */ 5786042f7df1SLai Jiangshan if (wq->flags & __WQ_ORDERED) 5787042f7df1SLai Jiangshan continue; 5788042f7df1SLai Jiangshan 578999c621efSLai Jiangshan ctx = apply_wqattrs_prepare(wq, wq->unbound_attrs, unbound_cpumask); 579084193c07STejun Heo if (IS_ERR(ctx)) { 579184193c07STejun Heo ret = PTR_ERR(ctx); 5792042f7df1SLai Jiangshan break; 5793042f7df1SLai Jiangshan } 5794042f7df1SLai Jiangshan 5795042f7df1SLai Jiangshan list_add_tail(&ctx->list, &ctxs); 5796042f7df1SLai Jiangshan } 5797042f7df1SLai Jiangshan 5798042f7df1SLai Jiangshan list_for_each_entry_safe(ctx, n, &ctxs, list) { 5799042f7df1SLai Jiangshan if (!ret) 5800042f7df1SLai Jiangshan apply_wqattrs_commit(ctx); 5801042f7df1SLai Jiangshan apply_wqattrs_cleanup(ctx); 5802042f7df1SLai Jiangshan } 5803042f7df1SLai Jiangshan 580499c621efSLai Jiangshan if (!ret) { 580599c621efSLai Jiangshan mutex_lock(&wq_pool_attach_mutex); 580699c621efSLai Jiangshan cpumask_copy(wq_unbound_cpumask, unbound_cpumask); 580799c621efSLai Jiangshan mutex_unlock(&wq_pool_attach_mutex); 580899c621efSLai Jiangshan } 5809042f7df1SLai Jiangshan return ret; 5810042f7df1SLai Jiangshan } 5811042f7df1SLai Jiangshan 5812042f7df1SLai Jiangshan /** 5813042f7df1SLai Jiangshan * workqueue_set_unbound_cpumask - Set the low-level unbound cpumask 5814042f7df1SLai Jiangshan * @cpumask: the cpumask to set 5815042f7df1SLai Jiangshan * 5816042f7df1SLai Jiangshan * The low-level workqueues cpumask is a global cpumask that limits 5817042f7df1SLai Jiangshan * the affinity of all unbound workqueues. This function check the @cpumask 5818042f7df1SLai Jiangshan * and apply it to all unbound workqueues and updates all pwqs of them. 5819042f7df1SLai Jiangshan * 582067dc8325SCai Huoqing * Return: 0 - Success 5821042f7df1SLai Jiangshan * -EINVAL - Invalid @cpumask 5822042f7df1SLai Jiangshan * -ENOMEM - Failed to allocate memory for attrs or pwqs. 5823042f7df1SLai Jiangshan */ 5824042f7df1SLai Jiangshan int workqueue_set_unbound_cpumask(cpumask_var_t cpumask) 5825042f7df1SLai Jiangshan { 5826042f7df1SLai Jiangshan int ret = -EINVAL; 5827042f7df1SLai Jiangshan 5828c98a9805STal Shorer /* 5829c98a9805STal Shorer * Not excluding isolated cpus on purpose. 5830c98a9805STal Shorer * If the user wishes to include them, we allow that. 5831c98a9805STal Shorer */ 5832042f7df1SLai Jiangshan cpumask_and(cpumask, cpumask, cpu_possible_mask); 5833042f7df1SLai Jiangshan if (!cpumask_empty(cpumask)) { 5834a0111cf6SLai Jiangshan apply_wqattrs_lock(); 5835d25302e4SMenglong Dong if (cpumask_equal(cpumask, wq_unbound_cpumask)) { 5836d25302e4SMenglong Dong ret = 0; 5837d25302e4SMenglong Dong goto out_unlock; 5838d25302e4SMenglong Dong } 5839d25302e4SMenglong Dong 584099c621efSLai Jiangshan ret = workqueue_apply_unbound_cpumask(cpumask); 5841042f7df1SLai Jiangshan 5842d25302e4SMenglong Dong out_unlock: 5843a0111cf6SLai Jiangshan apply_wqattrs_unlock(); 5844042f7df1SLai Jiangshan } 5845042f7df1SLai Jiangshan 5846042f7df1SLai Jiangshan return ret; 5847042f7df1SLai Jiangshan } 5848042f7df1SLai Jiangshan 584963c5484eSTejun Heo static int parse_affn_scope(const char *val) 585063c5484eSTejun Heo { 585163c5484eSTejun Heo int i; 585263c5484eSTejun Heo 585363c5484eSTejun Heo for (i = 0; i < ARRAY_SIZE(wq_affn_names); i++) { 585463c5484eSTejun Heo if (!strncasecmp(val, wq_affn_names[i], strlen(wq_affn_names[i]))) 585563c5484eSTejun Heo return i; 585663c5484eSTejun Heo } 585763c5484eSTejun Heo return -EINVAL; 585863c5484eSTejun Heo } 585963c5484eSTejun Heo 586063c5484eSTejun Heo static int wq_affn_dfl_set(const char *val, const struct kernel_param *kp) 586163c5484eSTejun Heo { 5862*523a301eSTejun Heo struct workqueue_struct *wq; 5863*523a301eSTejun Heo int affn, cpu; 586463c5484eSTejun Heo 586563c5484eSTejun Heo affn = parse_affn_scope(val); 586663c5484eSTejun Heo if (affn < 0) 586763c5484eSTejun Heo return affn; 5868*523a301eSTejun Heo if (affn == WQ_AFFN_DFL) 5869*523a301eSTejun Heo return -EINVAL; 5870*523a301eSTejun Heo 5871*523a301eSTejun Heo cpus_read_lock(); 5872*523a301eSTejun Heo mutex_lock(&wq_pool_mutex); 587363c5484eSTejun Heo 587463c5484eSTejun Heo wq_affn_dfl = affn; 5875*523a301eSTejun Heo 5876*523a301eSTejun Heo list_for_each_entry(wq, &workqueues, list) { 5877*523a301eSTejun Heo for_each_online_cpu(cpu) { 5878*523a301eSTejun Heo wq_update_pod(wq, cpu, cpu, true); 5879*523a301eSTejun Heo } 5880*523a301eSTejun Heo } 5881*523a301eSTejun Heo 5882*523a301eSTejun Heo mutex_unlock(&wq_pool_mutex); 5883*523a301eSTejun Heo cpus_read_unlock(); 5884*523a301eSTejun Heo 588563c5484eSTejun Heo return 0; 588663c5484eSTejun Heo } 588763c5484eSTejun Heo 588863c5484eSTejun Heo static int wq_affn_dfl_get(char *buffer, const struct kernel_param *kp) 588963c5484eSTejun Heo { 589063c5484eSTejun Heo return scnprintf(buffer, PAGE_SIZE, "%s\n", wq_affn_names[wq_affn_dfl]); 589163c5484eSTejun Heo } 589263c5484eSTejun Heo 589363c5484eSTejun Heo static const struct kernel_param_ops wq_affn_dfl_ops = { 589463c5484eSTejun Heo .set = wq_affn_dfl_set, 589563c5484eSTejun Heo .get = wq_affn_dfl_get, 589663c5484eSTejun Heo }; 589763c5484eSTejun Heo 589863c5484eSTejun Heo module_param_cb(default_affinity_scope, &wq_affn_dfl_ops, NULL, 0644); 589963c5484eSTejun Heo 59006ba94429SFrederic Weisbecker #ifdef CONFIG_SYSFS 59016ba94429SFrederic Weisbecker /* 59026ba94429SFrederic Weisbecker * Workqueues with WQ_SYSFS flag set is visible to userland via 59036ba94429SFrederic Weisbecker * /sys/bus/workqueue/devices/WQ_NAME. All visible workqueues have the 59046ba94429SFrederic Weisbecker * following attributes. 59056ba94429SFrederic Weisbecker * 59066ba94429SFrederic Weisbecker * per_cpu RO bool : whether the workqueue is per-cpu or unbound 59076ba94429SFrederic Weisbecker * max_active RW int : maximum number of in-flight work items 59086ba94429SFrederic Weisbecker * 59096ba94429SFrederic Weisbecker * Unbound workqueues have the following extra attributes. 59106ba94429SFrederic Weisbecker * 59116ba94429SFrederic Weisbecker * nice RW int : nice value of the workers 59126ba94429SFrederic Weisbecker * cpumask RW mask : bitmask of allowed CPUs for the workers 591363c5484eSTejun Heo * affinity_scope RW str : worker CPU affinity scope (cache, numa, none) 59148639ecebSTejun Heo * affinity_strict RW bool : worker CPU affinity is strict 59156ba94429SFrederic Weisbecker */ 59166ba94429SFrederic Weisbecker struct wq_device { 59176ba94429SFrederic Weisbecker struct workqueue_struct *wq; 59186ba94429SFrederic Weisbecker struct device dev; 59196ba94429SFrederic Weisbecker }; 59206ba94429SFrederic Weisbecker 59216ba94429SFrederic Weisbecker static struct workqueue_struct *dev_to_wq(struct device *dev) 59226ba94429SFrederic Weisbecker { 59236ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 59246ba94429SFrederic Weisbecker 59256ba94429SFrederic Weisbecker return wq_dev->wq; 59266ba94429SFrederic Weisbecker } 59276ba94429SFrederic Weisbecker 59286ba94429SFrederic Weisbecker static ssize_t per_cpu_show(struct device *dev, struct device_attribute *attr, 59296ba94429SFrederic Weisbecker char *buf) 59306ba94429SFrederic Weisbecker { 59316ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59326ba94429SFrederic Weisbecker 59336ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", (bool)!(wq->flags & WQ_UNBOUND)); 59346ba94429SFrederic Weisbecker } 59356ba94429SFrederic Weisbecker static DEVICE_ATTR_RO(per_cpu); 59366ba94429SFrederic Weisbecker 59376ba94429SFrederic Weisbecker static ssize_t max_active_show(struct device *dev, 59386ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 59396ba94429SFrederic Weisbecker { 59406ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59416ba94429SFrederic Weisbecker 59426ba94429SFrederic Weisbecker return scnprintf(buf, PAGE_SIZE, "%d\n", wq->saved_max_active); 59436ba94429SFrederic Weisbecker } 59446ba94429SFrederic Weisbecker 59456ba94429SFrederic Weisbecker static ssize_t max_active_store(struct device *dev, 59466ba94429SFrederic Weisbecker struct device_attribute *attr, const char *buf, 59476ba94429SFrederic Weisbecker size_t count) 59486ba94429SFrederic Weisbecker { 59496ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59506ba94429SFrederic Weisbecker int val; 59516ba94429SFrederic Weisbecker 59526ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &val) != 1 || val <= 0) 59536ba94429SFrederic Weisbecker return -EINVAL; 59546ba94429SFrederic Weisbecker 59556ba94429SFrederic Weisbecker workqueue_set_max_active(wq, val); 59566ba94429SFrederic Weisbecker return count; 59576ba94429SFrederic Weisbecker } 59586ba94429SFrederic Weisbecker static DEVICE_ATTR_RW(max_active); 59596ba94429SFrederic Weisbecker 59606ba94429SFrederic Weisbecker static struct attribute *wq_sysfs_attrs[] = { 59616ba94429SFrederic Weisbecker &dev_attr_per_cpu.attr, 59626ba94429SFrederic Weisbecker &dev_attr_max_active.attr, 59636ba94429SFrederic Weisbecker NULL, 59646ba94429SFrederic Weisbecker }; 59656ba94429SFrederic Weisbecker ATTRIBUTE_GROUPS(wq_sysfs); 59666ba94429SFrederic Weisbecker 59676ba94429SFrederic Weisbecker static ssize_t wq_nice_show(struct device *dev, struct device_attribute *attr, 59686ba94429SFrederic Weisbecker char *buf) 59696ba94429SFrederic Weisbecker { 59706ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59716ba94429SFrederic Weisbecker int written; 59726ba94429SFrederic Weisbecker 59736ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 59746ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%d\n", wq->unbound_attrs->nice); 59756ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 59766ba94429SFrederic Weisbecker 59776ba94429SFrederic Weisbecker return written; 59786ba94429SFrederic Weisbecker } 59796ba94429SFrederic Weisbecker 59806ba94429SFrederic Weisbecker /* prepare workqueue_attrs for sysfs store operations */ 59816ba94429SFrederic Weisbecker static struct workqueue_attrs *wq_sysfs_prep_attrs(struct workqueue_struct *wq) 59826ba94429SFrederic Weisbecker { 59836ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 59846ba94429SFrederic Weisbecker 5985899a94feSLai Jiangshan lockdep_assert_held(&wq_pool_mutex); 5986899a94feSLai Jiangshan 5987be69d00dSThomas Gleixner attrs = alloc_workqueue_attrs(); 59886ba94429SFrederic Weisbecker if (!attrs) 59896ba94429SFrederic Weisbecker return NULL; 59906ba94429SFrederic Weisbecker 59916ba94429SFrederic Weisbecker copy_workqueue_attrs(attrs, wq->unbound_attrs); 59926ba94429SFrederic Weisbecker return attrs; 59936ba94429SFrederic Weisbecker } 59946ba94429SFrederic Weisbecker 59956ba94429SFrederic Weisbecker static ssize_t wq_nice_store(struct device *dev, struct device_attribute *attr, 59966ba94429SFrederic Weisbecker const char *buf, size_t count) 59976ba94429SFrederic Weisbecker { 59986ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 59996ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 6000d4d3e257SLai Jiangshan int ret = -ENOMEM; 6001d4d3e257SLai Jiangshan 6002d4d3e257SLai Jiangshan apply_wqattrs_lock(); 60036ba94429SFrederic Weisbecker 60046ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 60056ba94429SFrederic Weisbecker if (!attrs) 6006d4d3e257SLai Jiangshan goto out_unlock; 60076ba94429SFrederic Weisbecker 60086ba94429SFrederic Weisbecker if (sscanf(buf, "%d", &attrs->nice) == 1 && 60096ba94429SFrederic Weisbecker attrs->nice >= MIN_NICE && attrs->nice <= MAX_NICE) 6010d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 60116ba94429SFrederic Weisbecker else 60126ba94429SFrederic Weisbecker ret = -EINVAL; 60136ba94429SFrederic Weisbecker 6014d4d3e257SLai Jiangshan out_unlock: 6015d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 60166ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 60176ba94429SFrederic Weisbecker return ret ?: count; 60186ba94429SFrederic Weisbecker } 60196ba94429SFrederic Weisbecker 60206ba94429SFrederic Weisbecker static ssize_t wq_cpumask_show(struct device *dev, 60216ba94429SFrederic Weisbecker struct device_attribute *attr, char *buf) 60226ba94429SFrederic Weisbecker { 60236ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 60246ba94429SFrederic Weisbecker int written; 60256ba94429SFrederic Weisbecker 60266ba94429SFrederic Weisbecker mutex_lock(&wq->mutex); 60276ba94429SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 60286ba94429SFrederic Weisbecker cpumask_pr_args(wq->unbound_attrs->cpumask)); 60296ba94429SFrederic Weisbecker mutex_unlock(&wq->mutex); 60306ba94429SFrederic Weisbecker return written; 60316ba94429SFrederic Weisbecker } 60326ba94429SFrederic Weisbecker 60336ba94429SFrederic Weisbecker static ssize_t wq_cpumask_store(struct device *dev, 60346ba94429SFrederic Weisbecker struct device_attribute *attr, 60356ba94429SFrederic Weisbecker const char *buf, size_t count) 60366ba94429SFrederic Weisbecker { 60376ba94429SFrederic Weisbecker struct workqueue_struct *wq = dev_to_wq(dev); 60386ba94429SFrederic Weisbecker struct workqueue_attrs *attrs; 6039d4d3e257SLai Jiangshan int ret = -ENOMEM; 6040d4d3e257SLai Jiangshan 6041d4d3e257SLai Jiangshan apply_wqattrs_lock(); 60426ba94429SFrederic Weisbecker 60436ba94429SFrederic Weisbecker attrs = wq_sysfs_prep_attrs(wq); 60446ba94429SFrederic Weisbecker if (!attrs) 6045d4d3e257SLai Jiangshan goto out_unlock; 60466ba94429SFrederic Weisbecker 60476ba94429SFrederic Weisbecker ret = cpumask_parse(buf, attrs->cpumask); 60486ba94429SFrederic Weisbecker if (!ret) 6049d4d3e257SLai Jiangshan ret = apply_workqueue_attrs_locked(wq, attrs); 60506ba94429SFrederic Weisbecker 6051d4d3e257SLai Jiangshan out_unlock: 6052d4d3e257SLai Jiangshan apply_wqattrs_unlock(); 60536ba94429SFrederic Weisbecker free_workqueue_attrs(attrs); 60546ba94429SFrederic Weisbecker return ret ?: count; 60556ba94429SFrederic Weisbecker } 60566ba94429SFrederic Weisbecker 605763c5484eSTejun Heo static ssize_t wq_affn_scope_show(struct device *dev, 605863c5484eSTejun Heo struct device_attribute *attr, char *buf) 605963c5484eSTejun Heo { 606063c5484eSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 606163c5484eSTejun Heo int written; 606263c5484eSTejun Heo 606363c5484eSTejun Heo mutex_lock(&wq->mutex); 6064*523a301eSTejun Heo if (wq->unbound_attrs->affn_scope == WQ_AFFN_DFL) 6065*523a301eSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%s (%s)\n", 6066*523a301eSTejun Heo wq_affn_names[WQ_AFFN_DFL], 6067*523a301eSTejun Heo wq_affn_names[wq_affn_dfl]); 6068*523a301eSTejun Heo else 606963c5484eSTejun Heo written = scnprintf(buf, PAGE_SIZE, "%s\n", 607063c5484eSTejun Heo wq_affn_names[wq->unbound_attrs->affn_scope]); 607163c5484eSTejun Heo mutex_unlock(&wq->mutex); 607263c5484eSTejun Heo 607363c5484eSTejun Heo return written; 607463c5484eSTejun Heo } 607563c5484eSTejun Heo 607663c5484eSTejun Heo static ssize_t wq_affn_scope_store(struct device *dev, 607763c5484eSTejun Heo struct device_attribute *attr, 607863c5484eSTejun Heo const char *buf, size_t count) 607963c5484eSTejun Heo { 608063c5484eSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 608163c5484eSTejun Heo struct workqueue_attrs *attrs; 608263c5484eSTejun Heo int affn, ret = -ENOMEM; 608363c5484eSTejun Heo 608463c5484eSTejun Heo affn = parse_affn_scope(buf); 608563c5484eSTejun Heo if (affn < 0) 608663c5484eSTejun Heo return affn; 608763c5484eSTejun Heo 608863c5484eSTejun Heo apply_wqattrs_lock(); 608963c5484eSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 609063c5484eSTejun Heo if (attrs) { 609163c5484eSTejun Heo attrs->affn_scope = affn; 609263c5484eSTejun Heo ret = apply_workqueue_attrs_locked(wq, attrs); 609363c5484eSTejun Heo } 609463c5484eSTejun Heo apply_wqattrs_unlock(); 609563c5484eSTejun Heo free_workqueue_attrs(attrs); 609663c5484eSTejun Heo return ret ?: count; 609763c5484eSTejun Heo } 609863c5484eSTejun Heo 60998639ecebSTejun Heo static ssize_t wq_affinity_strict_show(struct device *dev, 61008639ecebSTejun Heo struct device_attribute *attr, char *buf) 61018639ecebSTejun Heo { 61028639ecebSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 61038639ecebSTejun Heo 61048639ecebSTejun Heo return scnprintf(buf, PAGE_SIZE, "%d\n", 61058639ecebSTejun Heo wq->unbound_attrs->affn_strict); 61068639ecebSTejun Heo } 61078639ecebSTejun Heo 61088639ecebSTejun Heo static ssize_t wq_affinity_strict_store(struct device *dev, 61098639ecebSTejun Heo struct device_attribute *attr, 61108639ecebSTejun Heo const char *buf, size_t count) 61118639ecebSTejun Heo { 61128639ecebSTejun Heo struct workqueue_struct *wq = dev_to_wq(dev); 61138639ecebSTejun Heo struct workqueue_attrs *attrs; 61148639ecebSTejun Heo int v, ret = -ENOMEM; 61158639ecebSTejun Heo 61168639ecebSTejun Heo if (sscanf(buf, "%d", &v) != 1) 61178639ecebSTejun Heo return -EINVAL; 61188639ecebSTejun Heo 61198639ecebSTejun Heo apply_wqattrs_lock(); 61208639ecebSTejun Heo attrs = wq_sysfs_prep_attrs(wq); 61218639ecebSTejun Heo if (attrs) { 61228639ecebSTejun Heo attrs->affn_strict = (bool)v; 61238639ecebSTejun Heo ret = apply_workqueue_attrs_locked(wq, attrs); 61248639ecebSTejun Heo } 61258639ecebSTejun Heo apply_wqattrs_unlock(); 61268639ecebSTejun Heo free_workqueue_attrs(attrs); 61278639ecebSTejun Heo return ret ?: count; 61288639ecebSTejun Heo } 61298639ecebSTejun Heo 61306ba94429SFrederic Weisbecker static struct device_attribute wq_sysfs_unbound_attrs[] = { 61316ba94429SFrederic Weisbecker __ATTR(nice, 0644, wq_nice_show, wq_nice_store), 61326ba94429SFrederic Weisbecker __ATTR(cpumask, 0644, wq_cpumask_show, wq_cpumask_store), 613363c5484eSTejun Heo __ATTR(affinity_scope, 0644, wq_affn_scope_show, wq_affn_scope_store), 61348639ecebSTejun Heo __ATTR(affinity_strict, 0644, wq_affinity_strict_show, wq_affinity_strict_store), 61356ba94429SFrederic Weisbecker __ATTR_NULL, 61366ba94429SFrederic Weisbecker }; 61376ba94429SFrederic Weisbecker 61386ba94429SFrederic Weisbecker static struct bus_type wq_subsys = { 61396ba94429SFrederic Weisbecker .name = "workqueue", 61406ba94429SFrederic Weisbecker .dev_groups = wq_sysfs_groups, 61416ba94429SFrederic Weisbecker }; 61426ba94429SFrederic Weisbecker 6143b05a7928SFrederic Weisbecker static ssize_t wq_unbound_cpumask_show(struct device *dev, 6144b05a7928SFrederic Weisbecker struct device_attribute *attr, char *buf) 6145b05a7928SFrederic Weisbecker { 6146b05a7928SFrederic Weisbecker int written; 6147b05a7928SFrederic Weisbecker 6148042f7df1SLai Jiangshan mutex_lock(&wq_pool_mutex); 6149b05a7928SFrederic Weisbecker written = scnprintf(buf, PAGE_SIZE, "%*pb\n", 6150b05a7928SFrederic Weisbecker cpumask_pr_args(wq_unbound_cpumask)); 6151042f7df1SLai Jiangshan mutex_unlock(&wq_pool_mutex); 6152b05a7928SFrederic Weisbecker 6153b05a7928SFrederic Weisbecker return written; 6154b05a7928SFrederic Weisbecker } 6155b05a7928SFrederic Weisbecker 6156042f7df1SLai Jiangshan static ssize_t wq_unbound_cpumask_store(struct device *dev, 6157042f7df1SLai Jiangshan struct device_attribute *attr, const char *buf, size_t count) 6158042f7df1SLai Jiangshan { 6159042f7df1SLai Jiangshan cpumask_var_t cpumask; 6160042f7df1SLai Jiangshan int ret; 6161042f7df1SLai Jiangshan 6162042f7df1SLai Jiangshan if (!zalloc_cpumask_var(&cpumask, GFP_KERNEL)) 6163042f7df1SLai Jiangshan return -ENOMEM; 6164042f7df1SLai Jiangshan 6165042f7df1SLai Jiangshan ret = cpumask_parse(buf, cpumask); 6166042f7df1SLai Jiangshan if (!ret) 6167042f7df1SLai Jiangshan ret = workqueue_set_unbound_cpumask(cpumask); 6168042f7df1SLai Jiangshan 6169042f7df1SLai Jiangshan free_cpumask_var(cpumask); 6170042f7df1SLai Jiangshan return ret ? ret : count; 6171042f7df1SLai Jiangshan } 6172042f7df1SLai Jiangshan 6173b05a7928SFrederic Weisbecker static struct device_attribute wq_sysfs_cpumask_attr = 6174042f7df1SLai Jiangshan __ATTR(cpumask, 0644, wq_unbound_cpumask_show, 6175042f7df1SLai Jiangshan wq_unbound_cpumask_store); 6176b05a7928SFrederic Weisbecker 61776ba94429SFrederic Weisbecker static int __init wq_sysfs_init(void) 61786ba94429SFrederic Weisbecker { 6179686f6697SGreg Kroah-Hartman struct device *dev_root; 6180b05a7928SFrederic Weisbecker int err; 6181b05a7928SFrederic Weisbecker 6182b05a7928SFrederic Weisbecker err = subsys_virtual_register(&wq_subsys, NULL); 6183b05a7928SFrederic Weisbecker if (err) 6184b05a7928SFrederic Weisbecker return err; 6185b05a7928SFrederic Weisbecker 6186686f6697SGreg Kroah-Hartman dev_root = bus_get_dev_root(&wq_subsys); 6187686f6697SGreg Kroah-Hartman if (dev_root) { 6188686f6697SGreg Kroah-Hartman err = device_create_file(dev_root, &wq_sysfs_cpumask_attr); 6189686f6697SGreg Kroah-Hartman put_device(dev_root); 6190686f6697SGreg Kroah-Hartman } 6191686f6697SGreg Kroah-Hartman return err; 61926ba94429SFrederic Weisbecker } 61936ba94429SFrederic Weisbecker core_initcall(wq_sysfs_init); 61946ba94429SFrederic Weisbecker 61956ba94429SFrederic Weisbecker static void wq_device_release(struct device *dev) 61966ba94429SFrederic Weisbecker { 61976ba94429SFrederic Weisbecker struct wq_device *wq_dev = container_of(dev, struct wq_device, dev); 61986ba94429SFrederic Weisbecker 61996ba94429SFrederic Weisbecker kfree(wq_dev); 62006ba94429SFrederic Weisbecker } 62016ba94429SFrederic Weisbecker 62026ba94429SFrederic Weisbecker /** 62036ba94429SFrederic Weisbecker * workqueue_sysfs_register - make a workqueue visible in sysfs 62046ba94429SFrederic Weisbecker * @wq: the workqueue to register 62056ba94429SFrederic Weisbecker * 62066ba94429SFrederic Weisbecker * Expose @wq in sysfs under /sys/bus/workqueue/devices. 62076ba94429SFrederic Weisbecker * alloc_workqueue*() automatically calls this function if WQ_SYSFS is set 62086ba94429SFrederic Weisbecker * which is the preferred method. 62096ba94429SFrederic Weisbecker * 62106ba94429SFrederic Weisbecker * Workqueue user should use this function directly iff it wants to apply 62116ba94429SFrederic Weisbecker * workqueue_attrs before making the workqueue visible in sysfs; otherwise, 62126ba94429SFrederic Weisbecker * apply_workqueue_attrs() may race against userland updating the 62136ba94429SFrederic Weisbecker * attributes. 62146ba94429SFrederic Weisbecker * 62156ba94429SFrederic Weisbecker * Return: 0 on success, -errno on failure. 62166ba94429SFrederic Weisbecker */ 62176ba94429SFrederic Weisbecker int workqueue_sysfs_register(struct workqueue_struct *wq) 62186ba94429SFrederic Weisbecker { 62196ba94429SFrederic Weisbecker struct wq_device *wq_dev; 62206ba94429SFrederic Weisbecker int ret; 62216ba94429SFrederic Weisbecker 62226ba94429SFrederic Weisbecker /* 6223402dd89dSShailendra Verma * Adjusting max_active or creating new pwqs by applying 62246ba94429SFrederic Weisbecker * attributes breaks ordering guarantee. Disallow exposing ordered 62256ba94429SFrederic Weisbecker * workqueues. 62266ba94429SFrederic Weisbecker */ 62270a94efb5STejun Heo if (WARN_ON(wq->flags & __WQ_ORDERED_EXPLICIT)) 62286ba94429SFrederic Weisbecker return -EINVAL; 62296ba94429SFrederic Weisbecker 62306ba94429SFrederic Weisbecker wq->wq_dev = wq_dev = kzalloc(sizeof(*wq_dev), GFP_KERNEL); 62316ba94429SFrederic Weisbecker if (!wq_dev) 62326ba94429SFrederic Weisbecker return -ENOMEM; 62336ba94429SFrederic Weisbecker 62346ba94429SFrederic Weisbecker wq_dev->wq = wq; 62356ba94429SFrederic Weisbecker wq_dev->dev.bus = &wq_subsys; 62366ba94429SFrederic Weisbecker wq_dev->dev.release = wq_device_release; 623723217b44SLars-Peter Clausen dev_set_name(&wq_dev->dev, "%s", wq->name); 62386ba94429SFrederic Weisbecker 62396ba94429SFrederic Weisbecker /* 62406ba94429SFrederic Weisbecker * unbound_attrs are created separately. Suppress uevent until 62416ba94429SFrederic Weisbecker * everything is ready. 62426ba94429SFrederic Weisbecker */ 62436ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, true); 62446ba94429SFrederic Weisbecker 62456ba94429SFrederic Weisbecker ret = device_register(&wq_dev->dev); 62466ba94429SFrederic Weisbecker if (ret) { 6247537f4146SArvind Yadav put_device(&wq_dev->dev); 62486ba94429SFrederic Weisbecker wq->wq_dev = NULL; 62496ba94429SFrederic Weisbecker return ret; 62506ba94429SFrederic Weisbecker } 62516ba94429SFrederic Weisbecker 62526ba94429SFrederic Weisbecker if (wq->flags & WQ_UNBOUND) { 62536ba94429SFrederic Weisbecker struct device_attribute *attr; 62546ba94429SFrederic Weisbecker 62556ba94429SFrederic Weisbecker for (attr = wq_sysfs_unbound_attrs; attr->attr.name; attr++) { 62566ba94429SFrederic Weisbecker ret = device_create_file(&wq_dev->dev, attr); 62576ba94429SFrederic Weisbecker if (ret) { 62586ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 62596ba94429SFrederic Weisbecker wq->wq_dev = NULL; 62606ba94429SFrederic Weisbecker return ret; 62616ba94429SFrederic Weisbecker } 62626ba94429SFrederic Weisbecker } 62636ba94429SFrederic Weisbecker } 62646ba94429SFrederic Weisbecker 62656ba94429SFrederic Weisbecker dev_set_uevent_suppress(&wq_dev->dev, false); 62666ba94429SFrederic Weisbecker kobject_uevent(&wq_dev->dev.kobj, KOBJ_ADD); 62676ba94429SFrederic Weisbecker return 0; 62686ba94429SFrederic Weisbecker } 62696ba94429SFrederic Weisbecker 62706ba94429SFrederic Weisbecker /** 62716ba94429SFrederic Weisbecker * workqueue_sysfs_unregister - undo workqueue_sysfs_register() 62726ba94429SFrederic Weisbecker * @wq: the workqueue to unregister 62736ba94429SFrederic Weisbecker * 62746ba94429SFrederic Weisbecker * If @wq is registered to sysfs by workqueue_sysfs_register(), unregister. 62756ba94429SFrederic Weisbecker */ 62766ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) 62776ba94429SFrederic Weisbecker { 62786ba94429SFrederic Weisbecker struct wq_device *wq_dev = wq->wq_dev; 62796ba94429SFrederic Weisbecker 62806ba94429SFrederic Weisbecker if (!wq->wq_dev) 62816ba94429SFrederic Weisbecker return; 62826ba94429SFrederic Weisbecker 62836ba94429SFrederic Weisbecker wq->wq_dev = NULL; 62846ba94429SFrederic Weisbecker device_unregister(&wq_dev->dev); 62856ba94429SFrederic Weisbecker } 62866ba94429SFrederic Weisbecker #else /* CONFIG_SYSFS */ 62876ba94429SFrederic Weisbecker static void workqueue_sysfs_unregister(struct workqueue_struct *wq) { } 62886ba94429SFrederic Weisbecker #endif /* CONFIG_SYSFS */ 62896ba94429SFrederic Weisbecker 629082607adcSTejun Heo /* 629182607adcSTejun Heo * Workqueue watchdog. 629282607adcSTejun Heo * 629382607adcSTejun Heo * Stall may be caused by various bugs - missing WQ_MEM_RECLAIM, illegal 629482607adcSTejun Heo * flush dependency, a concurrency managed work item which stays RUNNING 629582607adcSTejun Heo * indefinitely. Workqueue stalls can be very difficult to debug as the 629682607adcSTejun Heo * usual warning mechanisms don't trigger and internal workqueue state is 629782607adcSTejun Heo * largely opaque. 629882607adcSTejun Heo * 629982607adcSTejun Heo * Workqueue watchdog monitors all worker pools periodically and dumps 630082607adcSTejun Heo * state if some pools failed to make forward progress for a while where 630182607adcSTejun Heo * forward progress is defined as the first item on ->worklist changing. 630282607adcSTejun Heo * 630382607adcSTejun Heo * This mechanism is controlled through the kernel parameter 630482607adcSTejun Heo * "workqueue.watchdog_thresh" which can be updated at runtime through the 630582607adcSTejun Heo * corresponding sysfs parameter file. 630682607adcSTejun Heo */ 630782607adcSTejun Heo #ifdef CONFIG_WQ_WATCHDOG 630882607adcSTejun Heo 630982607adcSTejun Heo static unsigned long wq_watchdog_thresh = 30; 63105cd79d6aSKees Cook static struct timer_list wq_watchdog_timer; 631182607adcSTejun Heo 631282607adcSTejun Heo static unsigned long wq_watchdog_touched = INITIAL_JIFFIES; 631382607adcSTejun Heo static DEFINE_PER_CPU(unsigned long, wq_watchdog_touched_cpu) = INITIAL_JIFFIES; 631482607adcSTejun Heo 6315cd2440d6SPetr Mladek /* 6316cd2440d6SPetr Mladek * Show workers that might prevent the processing of pending work items. 6317cd2440d6SPetr Mladek * The only candidates are CPU-bound workers in the running state. 6318cd2440d6SPetr Mladek * Pending work items should be handled by another idle worker 6319cd2440d6SPetr Mladek * in all other situations. 6320cd2440d6SPetr Mladek */ 6321cd2440d6SPetr Mladek static void show_cpu_pool_hog(struct worker_pool *pool) 6322cd2440d6SPetr Mladek { 6323cd2440d6SPetr Mladek struct worker *worker; 6324cd2440d6SPetr Mladek unsigned long flags; 6325cd2440d6SPetr Mladek int bkt; 6326cd2440d6SPetr Mladek 6327cd2440d6SPetr Mladek raw_spin_lock_irqsave(&pool->lock, flags); 6328cd2440d6SPetr Mladek 6329cd2440d6SPetr Mladek hash_for_each(pool->busy_hash, bkt, worker, hentry) { 6330cd2440d6SPetr Mladek if (task_is_running(worker->task)) { 6331cd2440d6SPetr Mladek /* 6332cd2440d6SPetr Mladek * Defer printing to avoid deadlocks in console 6333cd2440d6SPetr Mladek * drivers that queue work while holding locks 6334cd2440d6SPetr Mladek * also taken in their write paths. 6335cd2440d6SPetr Mladek */ 6336cd2440d6SPetr Mladek printk_deferred_enter(); 6337cd2440d6SPetr Mladek 6338cd2440d6SPetr Mladek pr_info("pool %d:\n", pool->id); 6339cd2440d6SPetr Mladek sched_show_task(worker->task); 6340cd2440d6SPetr Mladek 6341cd2440d6SPetr Mladek printk_deferred_exit(); 6342cd2440d6SPetr Mladek } 6343cd2440d6SPetr Mladek } 6344cd2440d6SPetr Mladek 6345cd2440d6SPetr Mladek raw_spin_unlock_irqrestore(&pool->lock, flags); 6346cd2440d6SPetr Mladek } 6347cd2440d6SPetr Mladek 6348cd2440d6SPetr Mladek static void show_cpu_pools_hogs(void) 6349cd2440d6SPetr Mladek { 6350cd2440d6SPetr Mladek struct worker_pool *pool; 6351cd2440d6SPetr Mladek int pi; 6352cd2440d6SPetr Mladek 6353cd2440d6SPetr Mladek pr_info("Showing backtraces of running workers in stalled CPU-bound worker pools:\n"); 6354cd2440d6SPetr Mladek 6355cd2440d6SPetr Mladek rcu_read_lock(); 6356cd2440d6SPetr Mladek 6357cd2440d6SPetr Mladek for_each_pool(pool, pi) { 6358cd2440d6SPetr Mladek if (pool->cpu_stall) 6359cd2440d6SPetr Mladek show_cpu_pool_hog(pool); 6360cd2440d6SPetr Mladek 6361cd2440d6SPetr Mladek } 6362cd2440d6SPetr Mladek 6363cd2440d6SPetr Mladek rcu_read_unlock(); 6364cd2440d6SPetr Mladek } 6365cd2440d6SPetr Mladek 636682607adcSTejun Heo static void wq_watchdog_reset_touched(void) 636782607adcSTejun Heo { 636882607adcSTejun Heo int cpu; 636982607adcSTejun Heo 637082607adcSTejun Heo wq_watchdog_touched = jiffies; 637182607adcSTejun Heo for_each_possible_cpu(cpu) 637282607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 637382607adcSTejun Heo } 637482607adcSTejun Heo 63755cd79d6aSKees Cook static void wq_watchdog_timer_fn(struct timer_list *unused) 637682607adcSTejun Heo { 637782607adcSTejun Heo unsigned long thresh = READ_ONCE(wq_watchdog_thresh) * HZ; 637882607adcSTejun Heo bool lockup_detected = false; 6379cd2440d6SPetr Mladek bool cpu_pool_stall = false; 6380940d71c6SSergey Senozhatsky unsigned long now = jiffies; 638182607adcSTejun Heo struct worker_pool *pool; 638282607adcSTejun Heo int pi; 638382607adcSTejun Heo 638482607adcSTejun Heo if (!thresh) 638582607adcSTejun Heo return; 638682607adcSTejun Heo 638782607adcSTejun Heo rcu_read_lock(); 638882607adcSTejun Heo 638982607adcSTejun Heo for_each_pool(pool, pi) { 639082607adcSTejun Heo unsigned long pool_ts, touched, ts; 639182607adcSTejun Heo 6392cd2440d6SPetr Mladek pool->cpu_stall = false; 639382607adcSTejun Heo if (list_empty(&pool->worklist)) 639482607adcSTejun Heo continue; 639582607adcSTejun Heo 6396940d71c6SSergey Senozhatsky /* 6397940d71c6SSergey Senozhatsky * If a virtual machine is stopped by the host it can look to 6398940d71c6SSergey Senozhatsky * the watchdog like a stall. 6399940d71c6SSergey Senozhatsky */ 6400940d71c6SSergey Senozhatsky kvm_check_and_clear_guest_paused(); 6401940d71c6SSergey Senozhatsky 640282607adcSTejun Heo /* get the latest of pool and touched timestamps */ 640389e28ce6SWang Qing if (pool->cpu >= 0) 640489e28ce6SWang Qing touched = READ_ONCE(per_cpu(wq_watchdog_touched_cpu, pool->cpu)); 640589e28ce6SWang Qing else 640682607adcSTejun Heo touched = READ_ONCE(wq_watchdog_touched); 640789e28ce6SWang Qing pool_ts = READ_ONCE(pool->watchdog_ts); 640882607adcSTejun Heo 640982607adcSTejun Heo if (time_after(pool_ts, touched)) 641082607adcSTejun Heo ts = pool_ts; 641182607adcSTejun Heo else 641282607adcSTejun Heo ts = touched; 641382607adcSTejun Heo 641482607adcSTejun Heo /* did we stall? */ 6415940d71c6SSergey Senozhatsky if (time_after(now, ts + thresh)) { 641682607adcSTejun Heo lockup_detected = true; 6417cd2440d6SPetr Mladek if (pool->cpu >= 0) { 6418cd2440d6SPetr Mladek pool->cpu_stall = true; 6419cd2440d6SPetr Mladek cpu_pool_stall = true; 6420cd2440d6SPetr Mladek } 642182607adcSTejun Heo pr_emerg("BUG: workqueue lockup - pool"); 642282607adcSTejun Heo pr_cont_pool_info(pool); 642382607adcSTejun Heo pr_cont(" stuck for %us!\n", 6424940d71c6SSergey Senozhatsky jiffies_to_msecs(now - pool_ts) / 1000); 642582607adcSTejun Heo } 6426cd2440d6SPetr Mladek 6427cd2440d6SPetr Mladek 642882607adcSTejun Heo } 642982607adcSTejun Heo 643082607adcSTejun Heo rcu_read_unlock(); 643182607adcSTejun Heo 643282607adcSTejun Heo if (lockup_detected) 643355df0933SImran Khan show_all_workqueues(); 643482607adcSTejun Heo 6435cd2440d6SPetr Mladek if (cpu_pool_stall) 6436cd2440d6SPetr Mladek show_cpu_pools_hogs(); 6437cd2440d6SPetr Mladek 643882607adcSTejun Heo wq_watchdog_reset_touched(); 643982607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh); 644082607adcSTejun Heo } 644182607adcSTejun Heo 6442cb9d7fd5SVincent Whitchurch notrace void wq_watchdog_touch(int cpu) 644382607adcSTejun Heo { 644482607adcSTejun Heo if (cpu >= 0) 644582607adcSTejun Heo per_cpu(wq_watchdog_touched_cpu, cpu) = jiffies; 644689e28ce6SWang Qing 644782607adcSTejun Heo wq_watchdog_touched = jiffies; 644882607adcSTejun Heo } 644982607adcSTejun Heo 645082607adcSTejun Heo static void wq_watchdog_set_thresh(unsigned long thresh) 645182607adcSTejun Heo { 645282607adcSTejun Heo wq_watchdog_thresh = 0; 645382607adcSTejun Heo del_timer_sync(&wq_watchdog_timer); 645482607adcSTejun Heo 645582607adcSTejun Heo if (thresh) { 645682607adcSTejun Heo wq_watchdog_thresh = thresh; 645782607adcSTejun Heo wq_watchdog_reset_touched(); 645882607adcSTejun Heo mod_timer(&wq_watchdog_timer, jiffies + thresh * HZ); 645982607adcSTejun Heo } 646082607adcSTejun Heo } 646182607adcSTejun Heo 646282607adcSTejun Heo static int wq_watchdog_param_set_thresh(const char *val, 646382607adcSTejun Heo const struct kernel_param *kp) 646482607adcSTejun Heo { 646582607adcSTejun Heo unsigned long thresh; 646682607adcSTejun Heo int ret; 646782607adcSTejun Heo 646882607adcSTejun Heo ret = kstrtoul(val, 0, &thresh); 646982607adcSTejun Heo if (ret) 647082607adcSTejun Heo return ret; 647182607adcSTejun Heo 647282607adcSTejun Heo if (system_wq) 647382607adcSTejun Heo wq_watchdog_set_thresh(thresh); 647482607adcSTejun Heo else 647582607adcSTejun Heo wq_watchdog_thresh = thresh; 647682607adcSTejun Heo 647782607adcSTejun Heo return 0; 647882607adcSTejun Heo } 647982607adcSTejun Heo 648082607adcSTejun Heo static const struct kernel_param_ops wq_watchdog_thresh_ops = { 648182607adcSTejun Heo .set = wq_watchdog_param_set_thresh, 648282607adcSTejun Heo .get = param_get_ulong, 648382607adcSTejun Heo }; 648482607adcSTejun Heo 648582607adcSTejun Heo module_param_cb(watchdog_thresh, &wq_watchdog_thresh_ops, &wq_watchdog_thresh, 648682607adcSTejun Heo 0644); 648782607adcSTejun Heo 648882607adcSTejun Heo static void wq_watchdog_init(void) 648982607adcSTejun Heo { 64905cd79d6aSKees Cook timer_setup(&wq_watchdog_timer, wq_watchdog_timer_fn, TIMER_DEFERRABLE); 649182607adcSTejun Heo wq_watchdog_set_thresh(wq_watchdog_thresh); 649282607adcSTejun Heo } 649382607adcSTejun Heo 649482607adcSTejun Heo #else /* CONFIG_WQ_WATCHDOG */ 649582607adcSTejun Heo 649682607adcSTejun Heo static inline void wq_watchdog_init(void) { } 649782607adcSTejun Heo 649882607adcSTejun Heo #endif /* CONFIG_WQ_WATCHDOG */ 649982607adcSTejun Heo 65003347fa09STejun Heo /** 65013347fa09STejun Heo * workqueue_init_early - early init for workqueue subsystem 65023347fa09STejun Heo * 65032930155bSTejun Heo * This is the first step of three-staged workqueue subsystem initialization and 65042930155bSTejun Heo * invoked as soon as the bare basics - memory allocation, cpumasks and idr are 65052930155bSTejun Heo * up. It sets up all the data structures and system workqueues and allows early 65062930155bSTejun Heo * boot code to create workqueues and queue/cancel work items. Actual work item 65072930155bSTejun Heo * execution starts only after kthreads can be created and scheduled right 65082930155bSTejun Heo * before early initcalls. 65093347fa09STejun Heo */ 65102333e829SYu Chen void __init workqueue_init_early(void) 65111da177e4SLinus Torvalds { 651284193c07STejun Heo struct wq_pod_type *pt = &wq_pod_types[WQ_AFFN_SYSTEM]; 65137a4e344cSTejun Heo int std_nice[NR_STD_WORKER_POOLS] = { 0, HIGHPRI_NICE_LEVEL }; 65147a4e344cSTejun Heo int i, cpu; 6515c34056a3STejun Heo 651610cdb157SLai Jiangshan BUILD_BUG_ON(__alignof__(struct pool_workqueue) < __alignof__(long long)); 6517e904e6c2STejun Heo 6518b05a7928SFrederic Weisbecker BUG_ON(!alloc_cpumask_var(&wq_unbound_cpumask, GFP_KERNEL)); 651904d4e665SFrederic Weisbecker cpumask_copy(wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_WQ)); 652004d4e665SFrederic Weisbecker cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, housekeeping_cpumask(HK_TYPE_DOMAIN)); 6521b05a7928SFrederic Weisbecker 6522ace3c549Stiozhang if (!cpumask_empty(&wq_cmdline_cpumask)) 6523ace3c549Stiozhang cpumask_and(wq_unbound_cpumask, wq_unbound_cpumask, &wq_cmdline_cpumask); 6524ace3c549Stiozhang 6525e904e6c2STejun Heo pwq_cache = KMEM_CACHE(pool_workqueue, SLAB_PANIC); 6526e904e6c2STejun Heo 65272930155bSTejun Heo wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 65282930155bSTejun Heo BUG_ON(!wq_update_pod_attrs_buf); 65292930155bSTejun Heo 653084193c07STejun Heo /* initialize WQ_AFFN_SYSTEM pods */ 653184193c07STejun Heo pt->pod_cpus = kcalloc(1, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 653284193c07STejun Heo pt->pod_node = kcalloc(1, sizeof(pt->pod_node[0]), GFP_KERNEL); 653384193c07STejun Heo pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 653484193c07STejun Heo BUG_ON(!pt->pod_cpus || !pt->pod_node || !pt->cpu_pod); 653584193c07STejun Heo 653684193c07STejun Heo BUG_ON(!zalloc_cpumask_var_node(&pt->pod_cpus[0], GFP_KERNEL, NUMA_NO_NODE)); 653784193c07STejun Heo 653884193c07STejun Heo wq_update_pod_attrs_buf = alloc_workqueue_attrs(); 653984193c07STejun Heo BUG_ON(!wq_update_pod_attrs_buf); 654084193c07STejun Heo 654184193c07STejun Heo pt->nr_pods = 1; 654284193c07STejun Heo cpumask_copy(pt->pod_cpus[0], cpu_possible_mask); 654384193c07STejun Heo pt->pod_node[0] = NUMA_NO_NODE; 654484193c07STejun Heo pt->cpu_pod[0] = 0; 654584193c07STejun Heo 6546706026c2STejun Heo /* initialize CPU pools */ 654729c91e99STejun Heo for_each_possible_cpu(cpu) { 65484ce62e9eSTejun Heo struct worker_pool *pool; 65498b03ae3cSTejun Heo 65507a4e344cSTejun Heo i = 0; 6551f02ae73aSTejun Heo for_each_cpu_worker_pool(pool, cpu) { 65527a4e344cSTejun Heo BUG_ON(init_worker_pool(pool)); 6553ec22ca5eSTejun Heo pool->cpu = cpu; 65547a4e344cSTejun Heo cpumask_copy(pool->attrs->cpumask, cpumask_of(cpu)); 65559546b29eSTejun Heo cpumask_copy(pool->attrs->__pod_cpumask, cpumask_of(cpu)); 65567a4e344cSTejun Heo pool->attrs->nice = std_nice[i++]; 65578639ecebSTejun Heo pool->attrs->affn_strict = true; 6558f3f90ad4STejun Heo pool->node = cpu_to_node(cpu); 65597a4e344cSTejun Heo 65609daf9e67STejun Heo /* alloc pool ID */ 656168e13a67SLai Jiangshan mutex_lock(&wq_pool_mutex); 65629daf9e67STejun Heo BUG_ON(worker_pool_assign_id(pool)); 656368e13a67SLai Jiangshan mutex_unlock(&wq_pool_mutex); 65644ce62e9eSTejun Heo } 65658b03ae3cSTejun Heo } 65668b03ae3cSTejun Heo 65678a2b7538STejun Heo /* create default unbound and ordered wq attrs */ 656829c91e99STejun Heo for (i = 0; i < NR_STD_WORKER_POOLS; i++) { 656929c91e99STejun Heo struct workqueue_attrs *attrs; 657029c91e99STejun Heo 6571be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 657229c91e99STejun Heo attrs->nice = std_nice[i]; 657329c91e99STejun Heo unbound_std_wq_attrs[i] = attrs; 65748a2b7538STejun Heo 65758a2b7538STejun Heo /* 65768a2b7538STejun Heo * An ordered wq should have only one pwq as ordering is 65778a2b7538STejun Heo * guaranteed by max_active which is enforced by pwqs. 65788a2b7538STejun Heo */ 6579be69d00dSThomas Gleixner BUG_ON(!(attrs = alloc_workqueue_attrs())); 65808a2b7538STejun Heo attrs->nice = std_nice[i]; 6581af73f5c9STejun Heo attrs->ordered = true; 65828a2b7538STejun Heo ordered_wq_attrs[i] = attrs; 658329c91e99STejun Heo } 658429c91e99STejun Heo 6585d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 65861aabe902SJoonsoo Kim system_highpri_wq = alloc_workqueue("events_highpri", WQ_HIGHPRI, 0); 6587d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 6588f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 6589636b927eSTejun Heo WQ_MAX_ACTIVE); 659024d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 659124d51addSTejun Heo WQ_FREEZABLE, 0); 65920668106cSViresh Kumar system_power_efficient_wq = alloc_workqueue("events_power_efficient", 65930668106cSViresh Kumar WQ_POWER_EFFICIENT, 0); 65940668106cSViresh Kumar system_freezable_power_efficient_wq = alloc_workqueue("events_freezable_power_efficient", 65950668106cSViresh Kumar WQ_FREEZABLE | WQ_POWER_EFFICIENT, 65960668106cSViresh Kumar 0); 65971aabe902SJoonsoo Kim BUG_ON(!system_wq || !system_highpri_wq || !system_long_wq || 65980668106cSViresh Kumar !system_unbound_wq || !system_freezable_wq || 65990668106cSViresh Kumar !system_power_efficient_wq || 66000668106cSViresh Kumar !system_freezable_power_efficient_wq); 66013347fa09STejun Heo } 66023347fa09STejun Heo 6603aa6fde93STejun Heo static void __init wq_cpu_intensive_thresh_init(void) 6604aa6fde93STejun Heo { 6605aa6fde93STejun Heo unsigned long thresh; 6606aa6fde93STejun Heo unsigned long bogo; 6607aa6fde93STejun Heo 6608aa6fde93STejun Heo /* if the user set it to a specific value, keep it */ 6609aa6fde93STejun Heo if (wq_cpu_intensive_thresh_us != ULONG_MAX) 6610aa6fde93STejun Heo return; 6611aa6fde93STejun Heo 6612967b494eSTejun Heo pwq_release_worker = kthread_create_worker(0, "pool_workqueue_release"); 6613967b494eSTejun Heo BUG_ON(IS_ERR(pwq_release_worker)); 6614967b494eSTejun Heo 6615aa6fde93STejun Heo /* 6616aa6fde93STejun Heo * The default of 10ms is derived from the fact that most modern (as of 6617aa6fde93STejun Heo * 2023) processors can do a lot in 10ms and that it's just below what 6618aa6fde93STejun Heo * most consider human-perceivable. However, the kernel also runs on a 6619aa6fde93STejun Heo * lot slower CPUs including microcontrollers where the threshold is way 6620aa6fde93STejun Heo * too low. 6621aa6fde93STejun Heo * 6622aa6fde93STejun Heo * Let's scale up the threshold upto 1 second if BogoMips is below 4000. 6623aa6fde93STejun Heo * This is by no means accurate but it doesn't have to be. The mechanism 6624aa6fde93STejun Heo * is still useful even when the threshold is fully scaled up. Also, as 6625aa6fde93STejun Heo * the reports would usually be applicable to everyone, some machines 6626aa6fde93STejun Heo * operating on longer thresholds won't significantly diminish their 6627aa6fde93STejun Heo * usefulness. 6628aa6fde93STejun Heo */ 6629aa6fde93STejun Heo thresh = 10 * USEC_PER_MSEC; 6630aa6fde93STejun Heo 6631aa6fde93STejun Heo /* see init/calibrate.c for lpj -> BogoMIPS calculation */ 6632aa6fde93STejun Heo bogo = max_t(unsigned long, loops_per_jiffy / 500000 * HZ, 1); 6633aa6fde93STejun Heo if (bogo < 4000) 6634aa6fde93STejun Heo thresh = min_t(unsigned long, thresh * 4000 / bogo, USEC_PER_SEC); 6635aa6fde93STejun Heo 6636aa6fde93STejun Heo pr_debug("wq_cpu_intensive_thresh: lpj=%lu BogoMIPS=%lu thresh_us=%lu\n", 6637aa6fde93STejun Heo loops_per_jiffy, bogo, thresh); 6638aa6fde93STejun Heo 6639aa6fde93STejun Heo wq_cpu_intensive_thresh_us = thresh; 6640aa6fde93STejun Heo } 6641aa6fde93STejun Heo 66423347fa09STejun Heo /** 66433347fa09STejun Heo * workqueue_init - bring workqueue subsystem fully online 66443347fa09STejun Heo * 66452930155bSTejun Heo * This is the second step of three-staged workqueue subsystem initialization 66462930155bSTejun Heo * and invoked as soon as kthreads can be created and scheduled. Workqueues have 66472930155bSTejun Heo * been created and work items queued on them, but there are no kworkers 66482930155bSTejun Heo * executing the work items yet. Populate the worker pools with the initial 66492930155bSTejun Heo * workers and enable future kworker creations. 66503347fa09STejun Heo */ 66512333e829SYu Chen void __init workqueue_init(void) 66523347fa09STejun Heo { 66532186d9f9STejun Heo struct workqueue_struct *wq; 66543347fa09STejun Heo struct worker_pool *pool; 66553347fa09STejun Heo int cpu, bkt; 66563347fa09STejun Heo 6657aa6fde93STejun Heo wq_cpu_intensive_thresh_init(); 6658aa6fde93STejun Heo 66592186d9f9STejun Heo mutex_lock(&wq_pool_mutex); 66602186d9f9STejun Heo 66612930155bSTejun Heo /* 66622930155bSTejun Heo * Per-cpu pools created earlier could be missing node hint. Fix them 66632930155bSTejun Heo * up. Also, create a rescuer for workqueues that requested it. 66642930155bSTejun Heo */ 66652186d9f9STejun Heo for_each_possible_cpu(cpu) { 66662186d9f9STejun Heo for_each_cpu_worker_pool(pool, cpu) { 66672186d9f9STejun Heo pool->node = cpu_to_node(cpu); 66682186d9f9STejun Heo } 66692186d9f9STejun Heo } 66702186d9f9STejun Heo 667140c17f75STejun Heo list_for_each_entry(wq, &workqueues, list) { 667240c17f75STejun Heo WARN(init_rescuer(wq), 667340c17f75STejun Heo "workqueue: failed to create early rescuer for %s", 667440c17f75STejun Heo wq->name); 667540c17f75STejun Heo } 66762186d9f9STejun Heo 66772186d9f9STejun Heo mutex_unlock(&wq_pool_mutex); 66782186d9f9STejun Heo 66793347fa09STejun Heo /* create the initial workers */ 66803347fa09STejun Heo for_each_online_cpu(cpu) { 66813347fa09STejun Heo for_each_cpu_worker_pool(pool, cpu) { 66823347fa09STejun Heo pool->flags &= ~POOL_DISASSOCIATED; 66833347fa09STejun Heo BUG_ON(!create_worker(pool)); 66843347fa09STejun Heo } 66853347fa09STejun Heo } 66863347fa09STejun Heo 66873347fa09STejun Heo hash_for_each(unbound_pool_hash, bkt, pool, hash_node) 66883347fa09STejun Heo BUG_ON(!create_worker(pool)); 66893347fa09STejun Heo 66903347fa09STejun Heo wq_online = true; 669182607adcSTejun Heo wq_watchdog_init(); 66921da177e4SLinus Torvalds } 6693c4f135d6STetsuo Handa 6694025e1684STejun Heo /* 6695025e1684STejun Heo * Initialize @pt by first initializing @pt->cpu_pod[] with pod IDs according to 6696025e1684STejun Heo * @cpu_shares_pod(). Each subset of CPUs that share a pod is assigned a unique 6697025e1684STejun Heo * and consecutive pod ID. The rest of @pt is initialized accordingly. 6698025e1684STejun Heo */ 6699025e1684STejun Heo static void __init init_pod_type(struct wq_pod_type *pt, 6700025e1684STejun Heo bool (*cpus_share_pod)(int, int)) 6701025e1684STejun Heo { 6702025e1684STejun Heo int cur, pre, cpu, pod; 6703025e1684STejun Heo 6704025e1684STejun Heo pt->nr_pods = 0; 6705025e1684STejun Heo 6706025e1684STejun Heo /* init @pt->cpu_pod[] according to @cpus_share_pod() */ 6707025e1684STejun Heo pt->cpu_pod = kcalloc(nr_cpu_ids, sizeof(pt->cpu_pod[0]), GFP_KERNEL); 6708025e1684STejun Heo BUG_ON(!pt->cpu_pod); 6709025e1684STejun Heo 6710025e1684STejun Heo for_each_possible_cpu(cur) { 6711025e1684STejun Heo for_each_possible_cpu(pre) { 6712025e1684STejun Heo if (pre >= cur) { 6713025e1684STejun Heo pt->cpu_pod[cur] = pt->nr_pods++; 6714025e1684STejun Heo break; 6715025e1684STejun Heo } 6716025e1684STejun Heo if (cpus_share_pod(cur, pre)) { 6717025e1684STejun Heo pt->cpu_pod[cur] = pt->cpu_pod[pre]; 6718025e1684STejun Heo break; 6719025e1684STejun Heo } 6720025e1684STejun Heo } 6721025e1684STejun Heo } 6722025e1684STejun Heo 6723025e1684STejun Heo /* init the rest to match @pt->cpu_pod[] */ 6724025e1684STejun Heo pt->pod_cpus = kcalloc(pt->nr_pods, sizeof(pt->pod_cpus[0]), GFP_KERNEL); 6725025e1684STejun Heo pt->pod_node = kcalloc(pt->nr_pods, sizeof(pt->pod_node[0]), GFP_KERNEL); 6726025e1684STejun Heo BUG_ON(!pt->pod_cpus || !pt->pod_node); 6727025e1684STejun Heo 6728025e1684STejun Heo for (pod = 0; pod < pt->nr_pods; pod++) 6729025e1684STejun Heo BUG_ON(!zalloc_cpumask_var(&pt->pod_cpus[pod], GFP_KERNEL)); 6730025e1684STejun Heo 6731025e1684STejun Heo for_each_possible_cpu(cpu) { 6732025e1684STejun Heo cpumask_set_cpu(cpu, pt->pod_cpus[pt->cpu_pod[cpu]]); 6733025e1684STejun Heo pt->pod_node[pt->cpu_pod[cpu]] = cpu_to_node(cpu); 6734025e1684STejun Heo } 6735025e1684STejun Heo } 6736025e1684STejun Heo 673763c5484eSTejun Heo static bool __init cpus_dont_share(int cpu0, int cpu1) 673863c5484eSTejun Heo { 673963c5484eSTejun Heo return false; 674063c5484eSTejun Heo } 674163c5484eSTejun Heo 674263c5484eSTejun Heo static bool __init cpus_share_smt(int cpu0, int cpu1) 674363c5484eSTejun Heo { 674463c5484eSTejun Heo #ifdef CONFIG_SCHED_SMT 674563c5484eSTejun Heo return cpumask_test_cpu(cpu0, cpu_smt_mask(cpu1)); 674663c5484eSTejun Heo #else 674763c5484eSTejun Heo return false; 674863c5484eSTejun Heo #endif 674963c5484eSTejun Heo } 675063c5484eSTejun Heo 6751025e1684STejun Heo static bool __init cpus_share_numa(int cpu0, int cpu1) 6752025e1684STejun Heo { 6753025e1684STejun Heo return cpu_to_node(cpu0) == cpu_to_node(cpu1); 6754025e1684STejun Heo } 6755025e1684STejun Heo 67562930155bSTejun Heo /** 67572930155bSTejun Heo * workqueue_init_topology - initialize CPU pods for unbound workqueues 67582930155bSTejun Heo * 67592930155bSTejun Heo * This is the third step of there-staged workqueue subsystem initialization and 67602930155bSTejun Heo * invoked after SMP and topology information are fully initialized. It 67612930155bSTejun Heo * initializes the unbound CPU pods accordingly. 67622930155bSTejun Heo */ 67632930155bSTejun Heo void __init workqueue_init_topology(void) 6764a86feae6STejun Heo { 67652930155bSTejun Heo struct workqueue_struct *wq; 6766025e1684STejun Heo int cpu; 6767a86feae6STejun Heo 676863c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_CPU], cpus_dont_share); 676963c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_SMT], cpus_share_smt); 677063c5484eSTejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_CACHE], cpus_share_cache); 6771025e1684STejun Heo init_pod_type(&wq_pod_types[WQ_AFFN_NUMA], cpus_share_numa); 6772a86feae6STejun Heo 67732930155bSTejun Heo mutex_lock(&wq_pool_mutex); 6774a86feae6STejun Heo 6775a86feae6STejun Heo /* 67762930155bSTejun Heo * Workqueues allocated earlier would have all CPUs sharing the default 67772930155bSTejun Heo * worker pool. Explicitly call wq_update_pod() on all workqueue and CPU 67782930155bSTejun Heo * combinations to apply per-pod sharing. 67792930155bSTejun Heo */ 67802930155bSTejun Heo list_for_each_entry(wq, &workqueues, list) { 67812930155bSTejun Heo for_each_online_cpu(cpu) { 67822930155bSTejun Heo wq_update_pod(wq, cpu, cpu, true); 67832930155bSTejun Heo } 67842930155bSTejun Heo } 67852930155bSTejun Heo 67862930155bSTejun Heo mutex_unlock(&wq_pool_mutex); 6787a86feae6STejun Heo } 6788a86feae6STejun Heo 678920bdedafSTetsuo Handa void __warn_flushing_systemwide_wq(void) 679020bdedafSTetsuo Handa { 679120bdedafSTetsuo Handa pr_warn("WARNING: Flushing system-wide workqueues will be prohibited in near future.\n"); 679220bdedafSTetsuo Handa dump_stack(); 679320bdedafSTetsuo Handa } 6794c4f135d6STetsuo Handa EXPORT_SYMBOL(__warn_flushing_systemwide_wq); 6795ace3c549Stiozhang 6796ace3c549Stiozhang static int __init workqueue_unbound_cpus_setup(char *str) 6797ace3c549Stiozhang { 6798ace3c549Stiozhang if (cpulist_parse(str, &wq_cmdline_cpumask) < 0) { 6799ace3c549Stiozhang cpumask_clear(&wq_cmdline_cpumask); 6800ace3c549Stiozhang pr_warn("workqueue.unbound_cpus: incorrect CPU range, using default\n"); 6801ace3c549Stiozhang } 6802ace3c549Stiozhang 6803ace3c549Stiozhang return 1; 6804ace3c549Stiozhang } 6805ace3c549Stiozhang __setup("workqueue.unbound_cpus=", workqueue_unbound_cpus_setup); 6806