11da177e4SLinus Torvalds /* 2c54fce6eSTejun Heo * kernel/workqueue.c - generic async execution with shared worker pool 31da177e4SLinus Torvalds * 4c54fce6eSTejun Heo * Copyright (C) 2002 Ingo Molnar 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Derived from the taskqueue/keventd code by: 71da177e4SLinus Torvalds * David Woodhouse <dwmw2@infradead.org> 8e1f8e874SFrancois Cami * Andrew Morton 91da177e4SLinus Torvalds * Kai Petzke <wpp@marie.physik.tu-berlin.de> 101da177e4SLinus Torvalds * Theodore Ts'o <tytso@mit.edu> 1189ada679SChristoph Lameter * 12cde53535SChristoph Lameter * Made to use alloc_percpu by Christoph Lameter. 13c54fce6eSTejun Heo * 14c54fce6eSTejun Heo * Copyright (C) 2010 SUSE Linux Products GmbH 15c54fce6eSTejun Heo * Copyright (C) 2010 Tejun Heo <tj@kernel.org> 16c54fce6eSTejun Heo * 17c54fce6eSTejun Heo * This is the generic async execution mechanism. Work items as are 18c54fce6eSTejun Heo * executed in process context. The worker pool is shared and 19c54fce6eSTejun Heo * automatically managed. There is one worker pool for each CPU and 20c54fce6eSTejun Heo * one extra for works which are better served by workers which are 21c54fce6eSTejun Heo * not bound to any specific CPU. 22c54fce6eSTejun Heo * 23c54fce6eSTejun Heo * Please read Documentation/workqueue.txt for details. 241da177e4SLinus Torvalds */ 251da177e4SLinus Torvalds 269984de1aSPaul Gortmaker #include <linux/export.h> 271da177e4SLinus Torvalds #include <linux/kernel.h> 281da177e4SLinus Torvalds #include <linux/sched.h> 291da177e4SLinus Torvalds #include <linux/init.h> 301da177e4SLinus Torvalds #include <linux/signal.h> 311da177e4SLinus Torvalds #include <linux/completion.h> 321da177e4SLinus Torvalds #include <linux/workqueue.h> 331da177e4SLinus Torvalds #include <linux/slab.h> 341da177e4SLinus Torvalds #include <linux/cpu.h> 351da177e4SLinus Torvalds #include <linux/notifier.h> 361da177e4SLinus Torvalds #include <linux/kthread.h> 371fa44ecaSJames Bottomley #include <linux/hardirq.h> 3846934023SChristoph Lameter #include <linux/mempolicy.h> 39341a5958SRafael J. Wysocki #include <linux/freezer.h> 40d5abe669SPeter Zijlstra #include <linux/kallsyms.h> 41d5abe669SPeter Zijlstra #include <linux/debug_locks.h> 424e6045f1SJohannes Berg #include <linux/lockdep.h> 43c34056a3STejun Heo #include <linux/idr.h> 44e22bee78STejun Heo 45e22bee78STejun Heo #include "workqueue_sched.h" 461da177e4SLinus Torvalds 47c8e55f36STejun Heo enum { 48bc2ae0f5STejun Heo /* 49bc2ae0f5STejun Heo * global_cwq flags 50bc2ae0f5STejun Heo * 51bc2ae0f5STejun Heo * A bound gcwq is either associated or disassociated with its CPU. 52bc2ae0f5STejun Heo * While associated (!DISASSOCIATED), all workers are bound to the 53bc2ae0f5STejun Heo * CPU and none has %WORKER_UNBOUND set and concurrency management 54bc2ae0f5STejun Heo * is in effect. 55bc2ae0f5STejun Heo * 56bc2ae0f5STejun Heo * While DISASSOCIATED, the cpu may be offline and all workers have 57bc2ae0f5STejun Heo * %WORKER_UNBOUND set and concurrency management disabled, and may 58bc2ae0f5STejun Heo * be executing on any CPU. The gcwq behaves as an unbound one. 59bc2ae0f5STejun Heo * 60bc2ae0f5STejun Heo * Note that DISASSOCIATED can be flipped only while holding 61bc2ae0f5STejun Heo * managership of all pools on the gcwq to avoid changing binding 62bc2ae0f5STejun Heo * state while create_worker() is in progress. 63bc2ae0f5STejun Heo */ 6411ebea50STejun Heo GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */ 6511ebea50STejun Heo GCWQ_FREEZING = 1 << 1, /* freeze in progress */ 6611ebea50STejun Heo 6711ebea50STejun Heo /* pool flags */ 6811ebea50STejun Heo POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ 69db7bccf4STejun Heo 70c8e55f36STejun Heo /* worker flags */ 71c8e55f36STejun Heo WORKER_STARTED = 1 << 0, /* started */ 72c8e55f36STejun Heo WORKER_DIE = 1 << 1, /* die die die */ 73c8e55f36STejun Heo WORKER_IDLE = 1 << 2, /* is idle */ 74e22bee78STejun Heo WORKER_PREP = 1 << 3, /* preparing to run works */ 75e22bee78STejun Heo WORKER_REBIND = 1 << 5, /* mom is home, come back */ 76fb0e7bebSTejun Heo WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ 77f3421797STejun Heo WORKER_UNBOUND = 1 << 7, /* worker is unbound */ 78e22bee78STejun Heo 79403c821dSTejun Heo WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | WORKER_UNBOUND | 80403c821dSTejun Heo WORKER_CPU_INTENSIVE, 81db7bccf4STejun Heo 82db7bccf4STejun Heo /* gcwq->trustee_state */ 83db7bccf4STejun Heo TRUSTEE_START = 0, /* start */ 84db7bccf4STejun Heo TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */ 85db7bccf4STejun Heo TRUSTEE_BUTCHER = 2, /* butcher workers */ 86db7bccf4STejun Heo TRUSTEE_RELEASE = 3, /* release workers */ 87db7bccf4STejun Heo TRUSTEE_DONE = 4, /* trustee is done */ 88c8e55f36STejun Heo 893270476aSTejun Heo NR_WORKER_POOLS = 2, /* # worker pools per gcwq */ 904ce62e9eSTejun Heo 91c8e55f36STejun Heo BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ 92c8e55f36STejun Heo BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, 93c8e55f36STejun Heo BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, 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 */ 103db7bccf4STejun Heo TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ 1041da177e4SLinus Torvalds 1051da177e4SLinus Torvalds /* 106e22bee78STejun Heo * Rescue workers are used only on emergencies and shared by 107e22bee78STejun Heo * all cpus. Give -20. 108e22bee78STejun Heo */ 109e22bee78STejun Heo RESCUER_NICE_LEVEL = -20, 1103270476aSTejun Heo HIGHPRI_NICE_LEVEL = -20, 111c8e55f36STejun Heo }; 112c8e55f36STejun Heo 1131da177e4SLinus Torvalds /* 1144690c4abSTejun Heo * Structure fields follow one of the following exclusion rules. 1154690c4abSTejun Heo * 116e41e704bSTejun Heo * I: Modifiable by initialization/destruction paths and read-only for 117e41e704bSTejun Heo * everyone else. 1184690c4abSTejun Heo * 119e22bee78STejun Heo * P: Preemption protected. Disabling preemption is enough and should 120e22bee78STejun Heo * only be modified and accessed from the local cpu. 121e22bee78STejun Heo * 1228b03ae3cSTejun Heo * L: gcwq->lock protected. Access with gcwq->lock held. 1234690c4abSTejun Heo * 124e22bee78STejun Heo * X: During normal operation, modification requires gcwq->lock and 125e22bee78STejun Heo * should be done only from local cpu. Either disabling preemption 126e22bee78STejun Heo * on local cpu or grabbing gcwq->lock is enough for read access. 127f3421797STejun Heo * If GCWQ_DISASSOCIATED is set, it's identical to L. 128e22bee78STejun Heo * 12973f53c4aSTejun Heo * F: wq->flush_mutex protected. 13073f53c4aSTejun Heo * 1314690c4abSTejun Heo * W: workqueue_lock protected. 1324690c4abSTejun Heo */ 1334690c4abSTejun Heo 1348b03ae3cSTejun Heo struct global_cwq; 135bd7bdd43STejun Heo struct worker_pool; 13625511a47STejun Heo struct idle_rebind; 137c34056a3STejun Heo 138e22bee78STejun Heo /* 139e22bee78STejun Heo * The poor guys doing the actual heavy lifting. All on-duty workers 140e22bee78STejun Heo * are either serving the manager role, on idle list or on busy hash. 141e22bee78STejun Heo */ 142c34056a3STejun Heo struct worker { 143c8e55f36STejun Heo /* on idle list while idle, on busy hash table while busy */ 144c8e55f36STejun Heo union { 145c8e55f36STejun Heo struct list_head entry; /* L: while idle */ 146c8e55f36STejun Heo struct hlist_node hentry; /* L: while busy */ 147c8e55f36STejun Heo }; 148c8e55f36STejun Heo 149c34056a3STejun Heo struct work_struct *current_work; /* L: work being processed */ 1508cca0eeaSTejun Heo struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ 151affee4b2STejun Heo struct list_head scheduled; /* L: scheduled works */ 152c34056a3STejun Heo struct task_struct *task; /* I: worker task */ 153bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 154e22bee78STejun Heo /* 64 bytes boundary on 64bit, 32 on 32bit */ 155e22bee78STejun Heo unsigned long last_active; /* L: last active timestamp */ 156e22bee78STejun Heo unsigned int flags; /* X: flags */ 157c34056a3STejun Heo int id; /* I: worker id */ 15825511a47STejun Heo 15925511a47STejun Heo /* for rebinding worker to CPU */ 16025511a47STejun Heo struct idle_rebind *idle_rebind; /* L: for idle worker */ 16125511a47STejun Heo struct work_struct rebind_work; /* L: for busy worker */ 162c34056a3STejun Heo }; 163c34056a3STejun Heo 164bd7bdd43STejun Heo struct worker_pool { 165bd7bdd43STejun Heo struct global_cwq *gcwq; /* I: the owning gcwq */ 16611ebea50STejun Heo unsigned int flags; /* X: flags */ 167bd7bdd43STejun Heo 168bd7bdd43STejun Heo struct list_head worklist; /* L: list of pending works */ 169bd7bdd43STejun Heo int nr_workers; /* L: total number of workers */ 170bd7bdd43STejun Heo int nr_idle; /* L: currently idle ones */ 171bd7bdd43STejun Heo 172bd7bdd43STejun Heo struct list_head idle_list; /* X: list of idle workers */ 173bd7bdd43STejun Heo struct timer_list idle_timer; /* L: worker idle timeout */ 174bd7bdd43STejun Heo struct timer_list mayday_timer; /* L: SOS timer for workers */ 175bd7bdd43STejun Heo 17660373152STejun Heo struct mutex manager_mutex; /* mutex manager should hold */ 177bd7bdd43STejun Heo struct ida worker_ida; /* L: for worker IDs */ 178bd7bdd43STejun Heo }; 179bd7bdd43STejun Heo 1804690c4abSTejun Heo /* 181e22bee78STejun Heo * Global per-cpu workqueue. There's one and only one for each cpu 182e22bee78STejun Heo * and all works are queued and processed here regardless of their 183e22bee78STejun Heo * target workqueues. 1848b03ae3cSTejun Heo */ 1858b03ae3cSTejun Heo struct global_cwq { 1868b03ae3cSTejun Heo spinlock_t lock; /* the gcwq lock */ 1878b03ae3cSTejun Heo unsigned int cpu; /* I: the associated cpu */ 188db7bccf4STejun Heo unsigned int flags; /* L: GCWQ_* flags */ 189c8e55f36STejun Heo 190bd7bdd43STejun Heo /* workers are chained either in busy_hash or pool idle_list */ 191c8e55f36STejun Heo struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; 192c8e55f36STejun Heo /* L: hash of busy workers */ 193c8e55f36STejun Heo 1943270476aSTejun Heo struct worker_pool pools[2]; /* normal and highpri pools */ 195db7bccf4STejun Heo 19625511a47STejun Heo wait_queue_head_t rebind_hold; /* rebind hold wait */ 19725511a47STejun Heo 198db7bccf4STejun Heo struct task_struct *trustee; /* L: for gcwq shutdown */ 199db7bccf4STejun Heo unsigned int trustee_state; /* L: trustee state */ 200db7bccf4STejun Heo wait_queue_head_t trustee_wait; /* trustee wait */ 2018b03ae3cSTejun Heo } ____cacheline_aligned_in_smp; 2028b03ae3cSTejun Heo 2038b03ae3cSTejun Heo /* 204502ca9d8STejun Heo * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of 2050f900049STejun Heo * work_struct->data are used for flags and thus cwqs need to be 2060f900049STejun Heo * aligned at two's power of the number of flag bits. 2071da177e4SLinus Torvalds */ 2081da177e4SLinus Torvalds struct cpu_workqueue_struct { 209bd7bdd43STejun Heo struct worker_pool *pool; /* I: the associated pool */ 2104690c4abSTejun Heo struct workqueue_struct *wq; /* I: the owning workqueue */ 21173f53c4aSTejun Heo int work_color; /* L: current color */ 21273f53c4aSTejun Heo int flush_color; /* L: flushing color */ 21373f53c4aSTejun Heo int nr_in_flight[WORK_NR_COLORS]; 21473f53c4aSTejun Heo /* L: nr of in_flight works */ 2151e19ffc6STejun Heo int nr_active; /* L: nr of active works */ 216a0a1a5fdSTejun Heo int max_active; /* L: max active works */ 2171e19ffc6STejun Heo struct list_head delayed_works; /* L: delayed works */ 2180f900049STejun Heo }; 2191da177e4SLinus Torvalds 2201da177e4SLinus Torvalds /* 22173f53c4aSTejun Heo * Structure used to wait for workqueue flush. 22273f53c4aSTejun Heo */ 22373f53c4aSTejun Heo struct wq_flusher { 22473f53c4aSTejun Heo struct list_head list; /* F: list of flushers */ 22573f53c4aSTejun Heo int flush_color; /* F: flush color waiting for */ 22673f53c4aSTejun Heo struct completion done; /* flush completion */ 22773f53c4aSTejun Heo }; 2281da177e4SLinus Torvalds 22973f53c4aSTejun Heo /* 230f2e005aaSTejun Heo * All cpumasks are assumed to be always set on UP and thus can't be 231f2e005aaSTejun Heo * used to determine whether there's something to be done. 232f2e005aaSTejun Heo */ 233f2e005aaSTejun Heo #ifdef CONFIG_SMP 234f2e005aaSTejun Heo typedef cpumask_var_t mayday_mask_t; 235f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) \ 236f2e005aaSTejun Heo cpumask_test_and_set_cpu((cpu), (mask)) 237f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) 238f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) 2399c37547aSTejun Heo #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) 240f2e005aaSTejun Heo #define free_mayday_mask(mask) free_cpumask_var((mask)) 241f2e005aaSTejun Heo #else 242f2e005aaSTejun Heo typedef unsigned long mayday_mask_t; 243f2e005aaSTejun Heo #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) 244f2e005aaSTejun Heo #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) 245f2e005aaSTejun Heo #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) 246f2e005aaSTejun Heo #define alloc_mayday_mask(maskp, gfp) true 247f2e005aaSTejun Heo #define free_mayday_mask(mask) do { } while (0) 248f2e005aaSTejun Heo #endif 2491da177e4SLinus Torvalds 2501da177e4SLinus Torvalds /* 2511da177e4SLinus Torvalds * The externally visible workqueue abstraction is an array of 2521da177e4SLinus Torvalds * per-CPU workqueues: 2531da177e4SLinus Torvalds */ 2541da177e4SLinus Torvalds struct workqueue_struct { 2559c5a2ba7STejun Heo unsigned int flags; /* W: WQ_* flags */ 256bdbc5dd7STejun Heo union { 257bdbc5dd7STejun Heo struct cpu_workqueue_struct __percpu *pcpu; 258bdbc5dd7STejun Heo struct cpu_workqueue_struct *single; 259bdbc5dd7STejun Heo unsigned long v; 260bdbc5dd7STejun Heo } cpu_wq; /* I: cwq's */ 2614690c4abSTejun Heo struct list_head list; /* W: list of all workqueues */ 26273f53c4aSTejun Heo 26373f53c4aSTejun Heo struct mutex flush_mutex; /* protects wq flushing */ 26473f53c4aSTejun Heo int work_color; /* F: current work color */ 26573f53c4aSTejun Heo int flush_color; /* F: current flush color */ 26673f53c4aSTejun Heo atomic_t nr_cwqs_to_flush; /* flush in progress */ 26773f53c4aSTejun Heo struct wq_flusher *first_flusher; /* F: first flusher */ 26873f53c4aSTejun Heo struct list_head flusher_queue; /* F: flush waiters */ 26973f53c4aSTejun Heo struct list_head flusher_overflow; /* F: flush overflow list */ 27073f53c4aSTejun Heo 271f2e005aaSTejun Heo mayday_mask_t mayday_mask; /* cpus requesting rescue */ 272e22bee78STejun Heo struct worker *rescuer; /* I: rescue worker */ 273e22bee78STejun Heo 2749c5a2ba7STejun Heo int nr_drainers; /* W: drain in progress */ 275dcd989cbSTejun Heo int saved_max_active; /* W: saved cwq max_active */ 2764e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 2774e6045f1SJohannes Berg struct lockdep_map lockdep_map; 2784e6045f1SJohannes Berg #endif 279b196be89STejun Heo char name[]; /* I: workqueue name */ 2801da177e4SLinus Torvalds }; 2811da177e4SLinus Torvalds 282d320c038STejun Heo struct workqueue_struct *system_wq __read_mostly; 283d320c038STejun Heo struct workqueue_struct *system_long_wq __read_mostly; 284d320c038STejun Heo struct workqueue_struct *system_nrt_wq __read_mostly; 285f3421797STejun Heo struct workqueue_struct *system_unbound_wq __read_mostly; 28624d51addSTejun Heo struct workqueue_struct *system_freezable_wq __read_mostly; 28762d3c543SAlan Stern struct workqueue_struct *system_nrt_freezable_wq __read_mostly; 288d320c038STejun Heo EXPORT_SYMBOL_GPL(system_wq); 289d320c038STejun Heo EXPORT_SYMBOL_GPL(system_long_wq); 290d320c038STejun Heo EXPORT_SYMBOL_GPL(system_nrt_wq); 291f3421797STejun Heo EXPORT_SYMBOL_GPL(system_unbound_wq); 29224d51addSTejun Heo EXPORT_SYMBOL_GPL(system_freezable_wq); 29362d3c543SAlan Stern EXPORT_SYMBOL_GPL(system_nrt_freezable_wq); 294d320c038STejun Heo 29597bd2347STejun Heo #define CREATE_TRACE_POINTS 29697bd2347STejun Heo #include <trace/events/workqueue.h> 29797bd2347STejun Heo 2984ce62e9eSTejun Heo #define for_each_worker_pool(pool, gcwq) \ 2993270476aSTejun Heo for ((pool) = &(gcwq)->pools[0]; \ 3003270476aSTejun Heo (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++) 3014ce62e9eSTejun Heo 302db7bccf4STejun Heo #define for_each_busy_worker(worker, i, pos, gcwq) \ 303db7bccf4STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ 304db7bccf4STejun Heo hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) 305db7bccf4STejun Heo 306f3421797STejun Heo static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, 307f3421797STejun Heo unsigned int sw) 308f3421797STejun Heo { 309f3421797STejun Heo if (cpu < nr_cpu_ids) { 310f3421797STejun Heo if (sw & 1) { 311f3421797STejun Heo cpu = cpumask_next(cpu, mask); 312f3421797STejun Heo if (cpu < nr_cpu_ids) 313f3421797STejun Heo return cpu; 314f3421797STejun Heo } 315f3421797STejun Heo if (sw & 2) 316f3421797STejun Heo return WORK_CPU_UNBOUND; 317f3421797STejun Heo } 318f3421797STejun Heo return WORK_CPU_NONE; 319f3421797STejun Heo } 320f3421797STejun Heo 321f3421797STejun Heo static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, 322f3421797STejun Heo struct workqueue_struct *wq) 323f3421797STejun Heo { 324f3421797STejun Heo return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); 325f3421797STejun Heo } 326f3421797STejun Heo 32709884951STejun Heo /* 32809884951STejun Heo * CPU iterators 32909884951STejun Heo * 33009884951STejun Heo * An extra gcwq is defined for an invalid cpu number 33109884951STejun Heo * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any 33209884951STejun Heo * specific CPU. The following iterators are similar to 33309884951STejun Heo * for_each_*_cpu() iterators but also considers the unbound gcwq. 33409884951STejun Heo * 33509884951STejun Heo * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND 33609884951STejun Heo * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND 33709884951STejun Heo * for_each_cwq_cpu() : possible CPUs for bound workqueues, 33809884951STejun Heo * WORK_CPU_UNBOUND for unbound workqueues 33909884951STejun Heo */ 340f3421797STejun Heo #define for_each_gcwq_cpu(cpu) \ 341f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ 342f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 343f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) 344f3421797STejun Heo 345f3421797STejun Heo #define for_each_online_gcwq_cpu(cpu) \ 346f3421797STejun Heo for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ 347f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 348f3421797STejun Heo (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) 349f3421797STejun Heo 350f3421797STejun Heo #define for_each_cwq_cpu(cpu, wq) \ 351f3421797STejun Heo for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ 352f3421797STejun Heo (cpu) < WORK_CPU_NONE; \ 353f3421797STejun Heo (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) 354f3421797STejun Heo 355dc186ad7SThomas Gleixner #ifdef CONFIG_DEBUG_OBJECTS_WORK 356dc186ad7SThomas Gleixner 357dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr; 358dc186ad7SThomas Gleixner 35999777288SStanislaw Gruszka static void *work_debug_hint(void *addr) 36099777288SStanislaw Gruszka { 36199777288SStanislaw Gruszka return ((struct work_struct *) addr)->func; 36299777288SStanislaw Gruszka } 36399777288SStanislaw Gruszka 364dc186ad7SThomas Gleixner /* 365dc186ad7SThomas Gleixner * fixup_init is called when: 366dc186ad7SThomas Gleixner * - an active object is initialized 367dc186ad7SThomas Gleixner */ 368dc186ad7SThomas Gleixner static int work_fixup_init(void *addr, enum debug_obj_state state) 369dc186ad7SThomas Gleixner { 370dc186ad7SThomas Gleixner struct work_struct *work = addr; 371dc186ad7SThomas Gleixner 372dc186ad7SThomas Gleixner switch (state) { 373dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 374dc186ad7SThomas Gleixner cancel_work_sync(work); 375dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 376dc186ad7SThomas Gleixner return 1; 377dc186ad7SThomas Gleixner default: 378dc186ad7SThomas Gleixner return 0; 379dc186ad7SThomas Gleixner } 380dc186ad7SThomas Gleixner } 381dc186ad7SThomas Gleixner 382dc186ad7SThomas Gleixner /* 383dc186ad7SThomas Gleixner * fixup_activate is called when: 384dc186ad7SThomas Gleixner * - an active object is activated 385dc186ad7SThomas Gleixner * - an unknown object is activated (might be a statically initialized object) 386dc186ad7SThomas Gleixner */ 387dc186ad7SThomas Gleixner static int work_fixup_activate(void *addr, enum debug_obj_state state) 388dc186ad7SThomas Gleixner { 389dc186ad7SThomas Gleixner struct work_struct *work = addr; 390dc186ad7SThomas Gleixner 391dc186ad7SThomas Gleixner switch (state) { 392dc186ad7SThomas Gleixner 393dc186ad7SThomas Gleixner case ODEBUG_STATE_NOTAVAILABLE: 394dc186ad7SThomas Gleixner /* 395dc186ad7SThomas Gleixner * This is not really a fixup. The work struct was 396dc186ad7SThomas Gleixner * statically initialized. We just make sure that it 397dc186ad7SThomas Gleixner * is tracked in the object tracker. 398dc186ad7SThomas Gleixner */ 39922df02bbSTejun Heo if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { 400dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 401dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 402dc186ad7SThomas Gleixner return 0; 403dc186ad7SThomas Gleixner } 404dc186ad7SThomas Gleixner WARN_ON_ONCE(1); 405dc186ad7SThomas Gleixner return 0; 406dc186ad7SThomas Gleixner 407dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 408dc186ad7SThomas Gleixner WARN_ON(1); 409dc186ad7SThomas Gleixner 410dc186ad7SThomas Gleixner default: 411dc186ad7SThomas Gleixner return 0; 412dc186ad7SThomas Gleixner } 413dc186ad7SThomas Gleixner } 414dc186ad7SThomas Gleixner 415dc186ad7SThomas Gleixner /* 416dc186ad7SThomas Gleixner * fixup_free is called when: 417dc186ad7SThomas Gleixner * - an active object is freed 418dc186ad7SThomas Gleixner */ 419dc186ad7SThomas Gleixner static int work_fixup_free(void *addr, enum debug_obj_state state) 420dc186ad7SThomas Gleixner { 421dc186ad7SThomas Gleixner struct work_struct *work = addr; 422dc186ad7SThomas Gleixner 423dc186ad7SThomas Gleixner switch (state) { 424dc186ad7SThomas Gleixner case ODEBUG_STATE_ACTIVE: 425dc186ad7SThomas Gleixner cancel_work_sync(work); 426dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 427dc186ad7SThomas Gleixner return 1; 428dc186ad7SThomas Gleixner default: 429dc186ad7SThomas Gleixner return 0; 430dc186ad7SThomas Gleixner } 431dc186ad7SThomas Gleixner } 432dc186ad7SThomas Gleixner 433dc186ad7SThomas Gleixner static struct debug_obj_descr work_debug_descr = { 434dc186ad7SThomas Gleixner .name = "work_struct", 43599777288SStanislaw Gruszka .debug_hint = work_debug_hint, 436dc186ad7SThomas Gleixner .fixup_init = work_fixup_init, 437dc186ad7SThomas Gleixner .fixup_activate = work_fixup_activate, 438dc186ad7SThomas Gleixner .fixup_free = work_fixup_free, 439dc186ad7SThomas Gleixner }; 440dc186ad7SThomas Gleixner 441dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) 442dc186ad7SThomas Gleixner { 443dc186ad7SThomas Gleixner debug_object_activate(work, &work_debug_descr); 444dc186ad7SThomas Gleixner } 445dc186ad7SThomas Gleixner 446dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) 447dc186ad7SThomas Gleixner { 448dc186ad7SThomas Gleixner debug_object_deactivate(work, &work_debug_descr); 449dc186ad7SThomas Gleixner } 450dc186ad7SThomas Gleixner 451dc186ad7SThomas Gleixner void __init_work(struct work_struct *work, int onstack) 452dc186ad7SThomas Gleixner { 453dc186ad7SThomas Gleixner if (onstack) 454dc186ad7SThomas Gleixner debug_object_init_on_stack(work, &work_debug_descr); 455dc186ad7SThomas Gleixner else 456dc186ad7SThomas Gleixner debug_object_init(work, &work_debug_descr); 457dc186ad7SThomas Gleixner } 458dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(__init_work); 459dc186ad7SThomas Gleixner 460dc186ad7SThomas Gleixner void destroy_work_on_stack(struct work_struct *work) 461dc186ad7SThomas Gleixner { 462dc186ad7SThomas Gleixner debug_object_free(work, &work_debug_descr); 463dc186ad7SThomas Gleixner } 464dc186ad7SThomas Gleixner EXPORT_SYMBOL_GPL(destroy_work_on_stack); 465dc186ad7SThomas Gleixner 466dc186ad7SThomas Gleixner #else 467dc186ad7SThomas Gleixner static inline void debug_work_activate(struct work_struct *work) { } 468dc186ad7SThomas Gleixner static inline void debug_work_deactivate(struct work_struct *work) { } 469dc186ad7SThomas Gleixner #endif 470dc186ad7SThomas Gleixner 47195402b38SGautham R Shenoy /* Serializes the accesses to the list of workqueues. */ 47295402b38SGautham R Shenoy static DEFINE_SPINLOCK(workqueue_lock); 4731da177e4SLinus Torvalds static LIST_HEAD(workqueues); 474a0a1a5fdSTejun Heo static bool workqueue_freezing; /* W: have wqs started freezing? */ 4751da177e4SLinus Torvalds 47614441960SOleg Nesterov /* 477e22bee78STejun Heo * The almighty global cpu workqueues. nr_running is the only field 478e22bee78STejun Heo * which is expected to be used frequently by other cpus via 479e22bee78STejun Heo * try_to_wake_up(). Put it in a separate cacheline. 48014441960SOleg Nesterov */ 4818b03ae3cSTejun Heo static DEFINE_PER_CPU(struct global_cwq, global_cwq); 4824ce62e9eSTejun Heo static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]); 483f756d5e2SNathan Lynch 484f3421797STejun Heo /* 485f3421797STejun Heo * Global cpu workqueue and nr_running counter for unbound gcwq. The 486f3421797STejun Heo * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its 487f3421797STejun Heo * workers have WORKER_UNBOUND set. 488f3421797STejun Heo */ 489f3421797STejun Heo static struct global_cwq unbound_global_cwq; 4904ce62e9eSTejun Heo static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = { 4914ce62e9eSTejun Heo [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */ 4924ce62e9eSTejun Heo }; 493f3421797STejun Heo 494c34056a3STejun Heo static int worker_thread(void *__worker); 4951da177e4SLinus Torvalds 4963270476aSTejun Heo static int worker_pool_pri(struct worker_pool *pool) 4973270476aSTejun Heo { 4983270476aSTejun Heo return pool - pool->gcwq->pools; 4993270476aSTejun Heo } 5003270476aSTejun Heo 5018b03ae3cSTejun Heo static struct global_cwq *get_gcwq(unsigned int cpu) 5021da177e4SLinus Torvalds { 503f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 5048b03ae3cSTejun Heo return &per_cpu(global_cwq, cpu); 505f3421797STejun Heo else 506f3421797STejun Heo return &unbound_global_cwq; 5071da177e4SLinus Torvalds } 5081da177e4SLinus Torvalds 50963d95a91STejun Heo static atomic_t *get_pool_nr_running(struct worker_pool *pool) 510b1f4ec17SOleg Nesterov { 51163d95a91STejun Heo int cpu = pool->gcwq->cpu; 5123270476aSTejun Heo int idx = worker_pool_pri(pool); 51363d95a91STejun Heo 514f3421797STejun Heo if (cpu != WORK_CPU_UNBOUND) 5154ce62e9eSTejun Heo return &per_cpu(pool_nr_running, cpu)[idx]; 516f3421797STejun Heo else 5174ce62e9eSTejun Heo return &unbound_pool_nr_running[idx]; 518b1f4ec17SOleg Nesterov } 519b1f4ec17SOleg Nesterov 5204690c4abSTejun Heo static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, 5214690c4abSTejun Heo struct workqueue_struct *wq) 522a848e3b6SOleg Nesterov { 523f3421797STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 524e06ffa1eSLai Jiangshan if (likely(cpu < nr_cpu_ids)) 525bdbc5dd7STejun Heo return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); 526f3421797STejun Heo } else if (likely(cpu == WORK_CPU_UNBOUND)) 527f3421797STejun Heo return wq->cpu_wq.single; 528f3421797STejun Heo return NULL; 529f3421797STejun Heo } 530a848e3b6SOleg Nesterov 53173f53c4aSTejun Heo static unsigned int work_color_to_flags(int color) 53273f53c4aSTejun Heo { 53373f53c4aSTejun Heo return color << WORK_STRUCT_COLOR_SHIFT; 53473f53c4aSTejun Heo } 53573f53c4aSTejun Heo 53673f53c4aSTejun Heo static int get_work_color(struct work_struct *work) 53773f53c4aSTejun Heo { 53873f53c4aSTejun Heo return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & 53973f53c4aSTejun Heo ((1 << WORK_STRUCT_COLOR_BITS) - 1); 54073f53c4aSTejun Heo } 54173f53c4aSTejun Heo 54273f53c4aSTejun Heo static int work_next_color(int color) 54373f53c4aSTejun Heo { 54473f53c4aSTejun Heo return (color + 1) % WORK_NR_COLORS; 5451da177e4SLinus Torvalds } 5461da177e4SLinus Torvalds 5474594bf15SDavid Howells /* 548e120153dSTejun Heo * A work's data points to the cwq with WORK_STRUCT_CWQ set while the 549e120153dSTejun Heo * work is on queue. Once execution starts, WORK_STRUCT_CWQ is 550e120153dSTejun Heo * cleared and the work data contains the cpu number it was last on. 5517a22ad75STejun Heo * 5527a22ad75STejun Heo * set_work_{cwq|cpu}() and clear_work_data() can be used to set the 5537a22ad75STejun Heo * cwq, cpu or clear work->data. These functions should only be 5547a22ad75STejun Heo * called while the work is owned - ie. while the PENDING bit is set. 5557a22ad75STejun Heo * 5567a22ad75STejun Heo * get_work_[g]cwq() can be used to obtain the gcwq or cwq 5577a22ad75STejun Heo * corresponding to a work. gcwq is available once the work has been 5587a22ad75STejun Heo * queued anywhere after initialization. cwq is available only from 5597a22ad75STejun Heo * queueing until execution starts. 5604594bf15SDavid Howells */ 5617a22ad75STejun Heo static inline void set_work_data(struct work_struct *work, unsigned long data, 5627a22ad75STejun Heo unsigned long flags) 5637a22ad75STejun Heo { 5647a22ad75STejun Heo BUG_ON(!work_pending(work)); 5657a22ad75STejun Heo atomic_long_set(&work->data, data | flags | work_static(work)); 5667a22ad75STejun Heo } 5677a22ad75STejun Heo 5687a22ad75STejun Heo static void set_work_cwq(struct work_struct *work, 5694690c4abSTejun Heo struct cpu_workqueue_struct *cwq, 5704690c4abSTejun Heo unsigned long extra_flags) 571365970a1SDavid Howells { 5727a22ad75STejun Heo set_work_data(work, (unsigned long)cwq, 573e120153dSTejun Heo WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); 574365970a1SDavid Howells } 575365970a1SDavid Howells 5767a22ad75STejun Heo static void set_work_cpu(struct work_struct *work, unsigned int cpu) 5774d707b9fSOleg Nesterov { 5787a22ad75STejun Heo set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); 5794d707b9fSOleg Nesterov } 5804d707b9fSOleg Nesterov 5817a22ad75STejun Heo static void clear_work_data(struct work_struct *work) 582365970a1SDavid Howells { 5837a22ad75STejun Heo set_work_data(work, WORK_STRUCT_NO_CPU, 0); 5847a22ad75STejun Heo } 5857a22ad75STejun Heo 5867a22ad75STejun Heo static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) 5877a22ad75STejun Heo { 588e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5897a22ad75STejun Heo 590e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 591e120153dSTejun Heo return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); 592e120153dSTejun Heo else 593e120153dSTejun Heo return NULL; 5947a22ad75STejun Heo } 5957a22ad75STejun Heo 5967a22ad75STejun Heo static struct global_cwq *get_work_gcwq(struct work_struct *work) 5977a22ad75STejun Heo { 598e120153dSTejun Heo unsigned long data = atomic_long_read(&work->data); 5997a22ad75STejun Heo unsigned int cpu; 6007a22ad75STejun Heo 601e120153dSTejun Heo if (data & WORK_STRUCT_CWQ) 602e120153dSTejun Heo return ((struct cpu_workqueue_struct *) 603bd7bdd43STejun Heo (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq; 6047a22ad75STejun Heo 6057a22ad75STejun Heo cpu = data >> WORK_STRUCT_FLAG_BITS; 606bdbc5dd7STejun Heo if (cpu == WORK_CPU_NONE) 6077a22ad75STejun Heo return NULL; 6087a22ad75STejun Heo 609f3421797STejun Heo BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); 6107a22ad75STejun Heo return get_gcwq(cpu); 611365970a1SDavid Howells } 612365970a1SDavid Howells 613e22bee78STejun Heo /* 6143270476aSTejun Heo * Policy functions. These define the policies on how the global worker 6153270476aSTejun Heo * pools are managed. Unless noted otherwise, these functions assume that 6163270476aSTejun Heo * they're being called with gcwq->lock held. 617e22bee78STejun Heo */ 618e22bee78STejun Heo 61963d95a91STejun Heo static bool __need_more_worker(struct worker_pool *pool) 620649027d7STejun Heo { 6213270476aSTejun Heo return !atomic_read(get_pool_nr_running(pool)); 622649027d7STejun Heo } 623649027d7STejun Heo 624e22bee78STejun Heo /* 625e22bee78STejun Heo * Need to wake up a worker? Called from anything but currently 626e22bee78STejun Heo * running workers. 627974271c4STejun Heo * 628974271c4STejun Heo * Note that, because unbound workers never contribute to nr_running, this 629974271c4STejun Heo * function will always return %true for unbound gcwq as long as the 630974271c4STejun Heo * worklist isn't empty. 631e22bee78STejun Heo */ 63263d95a91STejun Heo static bool need_more_worker(struct worker_pool *pool) 633e22bee78STejun Heo { 63463d95a91STejun Heo return !list_empty(&pool->worklist) && __need_more_worker(pool); 635e22bee78STejun Heo } 636e22bee78STejun Heo 637e22bee78STejun Heo /* Can I start working? Called from busy but !running workers. */ 63863d95a91STejun Heo static bool may_start_working(struct worker_pool *pool) 639e22bee78STejun Heo { 64063d95a91STejun Heo return pool->nr_idle; 641e22bee78STejun Heo } 642e22bee78STejun Heo 643e22bee78STejun Heo /* Do I need to keep working? Called from currently running workers. */ 64463d95a91STejun Heo static bool keep_working(struct worker_pool *pool) 645e22bee78STejun Heo { 64663d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 647e22bee78STejun Heo 6483270476aSTejun Heo return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1; 649e22bee78STejun Heo } 650e22bee78STejun Heo 651e22bee78STejun Heo /* Do we need a new worker? Called from manager. */ 65263d95a91STejun Heo static bool need_to_create_worker(struct worker_pool *pool) 653e22bee78STejun Heo { 65463d95a91STejun Heo return need_more_worker(pool) && !may_start_working(pool); 655e22bee78STejun Heo } 656e22bee78STejun Heo 657e22bee78STejun Heo /* Do I need to be the manager? */ 65863d95a91STejun Heo static bool need_to_manage_workers(struct worker_pool *pool) 659e22bee78STejun Heo { 66063d95a91STejun Heo return need_to_create_worker(pool) || 66111ebea50STejun Heo (pool->flags & POOL_MANAGE_WORKERS); 662e22bee78STejun Heo } 663e22bee78STejun Heo 664e22bee78STejun Heo /* Do we have too many workers and should some go away? */ 66563d95a91STejun Heo static bool too_many_workers(struct worker_pool *pool) 666e22bee78STejun Heo { 66760373152STejun Heo bool managing = mutex_is_locked(&pool->manager_mutex); 66863d95a91STejun Heo int nr_idle = pool->nr_idle + managing; /* manager is considered idle */ 66963d95a91STejun Heo int nr_busy = pool->nr_workers - nr_idle; 670e22bee78STejun Heo 671e22bee78STejun Heo return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; 672e22bee78STejun Heo } 673e22bee78STejun Heo 674e22bee78STejun Heo /* 675e22bee78STejun Heo * Wake up functions. 676e22bee78STejun Heo */ 677e22bee78STejun Heo 6787e11629dSTejun Heo /* Return the first worker. Safe with preemption disabled */ 67963d95a91STejun Heo static struct worker *first_worker(struct worker_pool *pool) 6807e11629dSTejun Heo { 68163d95a91STejun Heo if (unlikely(list_empty(&pool->idle_list))) 6827e11629dSTejun Heo return NULL; 6837e11629dSTejun Heo 68463d95a91STejun Heo return list_first_entry(&pool->idle_list, struct worker, entry); 6857e11629dSTejun Heo } 6867e11629dSTejun Heo 6877e11629dSTejun Heo /** 6887e11629dSTejun Heo * wake_up_worker - wake up an idle worker 68963d95a91STejun Heo * @pool: worker pool to wake worker from 6907e11629dSTejun Heo * 69163d95a91STejun Heo * Wake up the first idle worker of @pool. 6927e11629dSTejun Heo * 6937e11629dSTejun Heo * CONTEXT: 6947e11629dSTejun Heo * spin_lock_irq(gcwq->lock). 6957e11629dSTejun Heo */ 69663d95a91STejun Heo static void wake_up_worker(struct worker_pool *pool) 6977e11629dSTejun Heo { 69863d95a91STejun Heo struct worker *worker = first_worker(pool); 6997e11629dSTejun Heo 7007e11629dSTejun Heo if (likely(worker)) 7017e11629dSTejun Heo wake_up_process(worker->task); 7027e11629dSTejun Heo } 7037e11629dSTejun Heo 7044690c4abSTejun Heo /** 705e22bee78STejun Heo * wq_worker_waking_up - a worker is waking up 706e22bee78STejun Heo * @task: task waking up 707e22bee78STejun Heo * @cpu: CPU @task is waking up to 708e22bee78STejun Heo * 709e22bee78STejun Heo * This function is called during try_to_wake_up() when a worker is 710e22bee78STejun Heo * being awoken. 711e22bee78STejun Heo * 712e22bee78STejun Heo * CONTEXT: 713e22bee78STejun Heo * spin_lock_irq(rq->lock) 714e22bee78STejun Heo */ 715e22bee78STejun Heo void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) 716e22bee78STejun Heo { 717e22bee78STejun Heo struct worker *worker = kthread_data(task); 718e22bee78STejun Heo 7192d64672eSSteven Rostedt if (!(worker->flags & WORKER_NOT_RUNNING)) 72063d95a91STejun Heo atomic_inc(get_pool_nr_running(worker->pool)); 721e22bee78STejun Heo } 722e22bee78STejun Heo 723e22bee78STejun Heo /** 724e22bee78STejun Heo * wq_worker_sleeping - a worker is going to sleep 725e22bee78STejun Heo * @task: task going to sleep 726e22bee78STejun Heo * @cpu: CPU in question, must be the current CPU number 727e22bee78STejun Heo * 728e22bee78STejun Heo * This function is called during schedule() when a busy worker is 729e22bee78STejun Heo * going to sleep. Worker on the same cpu can be woken up by 730e22bee78STejun Heo * returning pointer to its task. 731e22bee78STejun Heo * 732e22bee78STejun Heo * CONTEXT: 733e22bee78STejun Heo * spin_lock_irq(rq->lock) 734e22bee78STejun Heo * 735e22bee78STejun Heo * RETURNS: 736e22bee78STejun Heo * Worker task on @cpu to wake up, %NULL if none. 737e22bee78STejun Heo */ 738e22bee78STejun Heo struct task_struct *wq_worker_sleeping(struct task_struct *task, 739e22bee78STejun Heo unsigned int cpu) 740e22bee78STejun Heo { 741e22bee78STejun Heo struct worker *worker = kthread_data(task), *to_wakeup = NULL; 742bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 74363d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 744e22bee78STejun Heo 7452d64672eSSteven Rostedt if (worker->flags & WORKER_NOT_RUNNING) 746e22bee78STejun Heo return NULL; 747e22bee78STejun Heo 748e22bee78STejun Heo /* this can only happen on the local cpu */ 749e22bee78STejun Heo BUG_ON(cpu != raw_smp_processor_id()); 750e22bee78STejun Heo 751e22bee78STejun Heo /* 752e22bee78STejun Heo * The counterpart of the following dec_and_test, implied mb, 753e22bee78STejun Heo * worklist not empty test sequence is in insert_work(). 754e22bee78STejun Heo * Please read comment there. 755e22bee78STejun Heo * 756e22bee78STejun Heo * NOT_RUNNING is clear. This means that trustee is not in 757e22bee78STejun Heo * charge and we're running on the local cpu w/ rq lock held 758e22bee78STejun Heo * and preemption disabled, which in turn means that none else 759e22bee78STejun Heo * could be manipulating idle_list, so dereferencing idle_list 760e22bee78STejun Heo * without gcwq lock is safe. 761e22bee78STejun Heo */ 762bd7bdd43STejun Heo if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist)) 76363d95a91STejun Heo to_wakeup = first_worker(pool); 764e22bee78STejun Heo return to_wakeup ? to_wakeup->task : NULL; 765e22bee78STejun Heo } 766e22bee78STejun Heo 767e22bee78STejun Heo /** 768e22bee78STejun Heo * worker_set_flags - set worker flags and adjust nr_running accordingly 769cb444766STejun Heo * @worker: self 770d302f017STejun Heo * @flags: flags to set 771d302f017STejun Heo * @wakeup: wakeup an idle worker if necessary 772d302f017STejun Heo * 773e22bee78STejun Heo * Set @flags in @worker->flags and adjust nr_running accordingly. If 774e22bee78STejun Heo * nr_running becomes zero and @wakeup is %true, an idle worker is 775e22bee78STejun Heo * woken up. 776d302f017STejun Heo * 777cb444766STejun Heo * CONTEXT: 778cb444766STejun Heo * spin_lock_irq(gcwq->lock) 779d302f017STejun Heo */ 780d302f017STejun Heo static inline void worker_set_flags(struct worker *worker, unsigned int flags, 781d302f017STejun Heo bool wakeup) 782d302f017STejun Heo { 783bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 784e22bee78STejun Heo 785cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 786cb444766STejun Heo 787e22bee78STejun Heo /* 788e22bee78STejun Heo * If transitioning into NOT_RUNNING, adjust nr_running and 789e22bee78STejun Heo * wake up an idle worker as necessary if requested by 790e22bee78STejun Heo * @wakeup. 791e22bee78STejun Heo */ 792e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && 793e22bee78STejun Heo !(worker->flags & WORKER_NOT_RUNNING)) { 79463d95a91STejun Heo atomic_t *nr_running = get_pool_nr_running(pool); 795e22bee78STejun Heo 796e22bee78STejun Heo if (wakeup) { 797e22bee78STejun Heo if (atomic_dec_and_test(nr_running) && 798bd7bdd43STejun Heo !list_empty(&pool->worklist)) 79963d95a91STejun Heo wake_up_worker(pool); 800e22bee78STejun Heo } else 801e22bee78STejun Heo atomic_dec(nr_running); 802e22bee78STejun Heo } 803e22bee78STejun Heo 804d302f017STejun Heo worker->flags |= flags; 805d302f017STejun Heo } 806d302f017STejun Heo 807d302f017STejun Heo /** 808e22bee78STejun Heo * worker_clr_flags - clear worker flags and adjust nr_running accordingly 809cb444766STejun Heo * @worker: self 810d302f017STejun Heo * @flags: flags to clear 811d302f017STejun Heo * 812e22bee78STejun Heo * Clear @flags in @worker->flags and adjust nr_running accordingly. 813d302f017STejun Heo * 814cb444766STejun Heo * CONTEXT: 815cb444766STejun Heo * spin_lock_irq(gcwq->lock) 816d302f017STejun Heo */ 817d302f017STejun Heo static inline void worker_clr_flags(struct worker *worker, unsigned int flags) 818d302f017STejun Heo { 81963d95a91STejun Heo struct worker_pool *pool = worker->pool; 820e22bee78STejun Heo unsigned int oflags = worker->flags; 821e22bee78STejun Heo 822cb444766STejun Heo WARN_ON_ONCE(worker->task != current); 823cb444766STejun Heo 824d302f017STejun Heo worker->flags &= ~flags; 825e22bee78STejun Heo 82642c025f3STejun Heo /* 82742c025f3STejun Heo * If transitioning out of NOT_RUNNING, increment nr_running. Note 82842c025f3STejun Heo * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask 82942c025f3STejun Heo * of multiple flags, not a single flag. 83042c025f3STejun Heo */ 831e22bee78STejun Heo if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) 832e22bee78STejun Heo if (!(worker->flags & WORKER_NOT_RUNNING)) 83363d95a91STejun Heo atomic_inc(get_pool_nr_running(pool)); 834d302f017STejun Heo } 835d302f017STejun Heo 836d302f017STejun Heo /** 837c8e55f36STejun Heo * busy_worker_head - return the busy hash head for a work 838c8e55f36STejun Heo * @gcwq: gcwq of interest 839c8e55f36STejun Heo * @work: work to be hashed 840c8e55f36STejun Heo * 841c8e55f36STejun Heo * Return hash head of @gcwq for @work. 842c8e55f36STejun Heo * 843c8e55f36STejun Heo * CONTEXT: 844c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 845c8e55f36STejun Heo * 846c8e55f36STejun Heo * RETURNS: 847c8e55f36STejun Heo * Pointer to the hash head. 848c8e55f36STejun Heo */ 849c8e55f36STejun Heo static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, 850c8e55f36STejun Heo struct work_struct *work) 851c8e55f36STejun Heo { 852c8e55f36STejun Heo const int base_shift = ilog2(sizeof(struct work_struct)); 853c8e55f36STejun Heo unsigned long v = (unsigned long)work; 854c8e55f36STejun Heo 855c8e55f36STejun Heo /* simple shift and fold hash, do we need something better? */ 856c8e55f36STejun Heo v >>= base_shift; 857c8e55f36STejun Heo v += v >> BUSY_WORKER_HASH_ORDER; 858c8e55f36STejun Heo v &= BUSY_WORKER_HASH_MASK; 859c8e55f36STejun Heo 860c8e55f36STejun Heo return &gcwq->busy_hash[v]; 861c8e55f36STejun Heo } 862c8e55f36STejun Heo 863c8e55f36STejun Heo /** 8648cca0eeaSTejun Heo * __find_worker_executing_work - find worker which is executing a work 8658cca0eeaSTejun Heo * @gcwq: gcwq of interest 8668cca0eeaSTejun Heo * @bwh: hash head as returned by busy_worker_head() 8678cca0eeaSTejun Heo * @work: work to find worker for 8688cca0eeaSTejun Heo * 8698cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. @bwh should be 8708cca0eeaSTejun Heo * the hash head obtained by calling busy_worker_head() with the same 8718cca0eeaSTejun Heo * work. 8728cca0eeaSTejun Heo * 8738cca0eeaSTejun Heo * CONTEXT: 8748cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 8758cca0eeaSTejun Heo * 8768cca0eeaSTejun Heo * RETURNS: 8778cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 8788cca0eeaSTejun Heo * otherwise. 8798cca0eeaSTejun Heo */ 8808cca0eeaSTejun Heo static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, 8818cca0eeaSTejun Heo struct hlist_head *bwh, 8828cca0eeaSTejun Heo struct work_struct *work) 8838cca0eeaSTejun Heo { 8848cca0eeaSTejun Heo struct worker *worker; 8858cca0eeaSTejun Heo struct hlist_node *tmp; 8868cca0eeaSTejun Heo 8878cca0eeaSTejun Heo hlist_for_each_entry(worker, tmp, bwh, hentry) 8888cca0eeaSTejun Heo if (worker->current_work == work) 8898cca0eeaSTejun Heo return worker; 8908cca0eeaSTejun Heo return NULL; 8918cca0eeaSTejun Heo } 8928cca0eeaSTejun Heo 8938cca0eeaSTejun Heo /** 8948cca0eeaSTejun Heo * find_worker_executing_work - find worker which is executing a work 8958cca0eeaSTejun Heo * @gcwq: gcwq of interest 8968cca0eeaSTejun Heo * @work: work to find worker for 8978cca0eeaSTejun Heo * 8988cca0eeaSTejun Heo * Find a worker which is executing @work on @gcwq. This function is 8998cca0eeaSTejun Heo * identical to __find_worker_executing_work() except that this 9008cca0eeaSTejun Heo * function calculates @bwh itself. 9018cca0eeaSTejun Heo * 9028cca0eeaSTejun Heo * CONTEXT: 9038cca0eeaSTejun Heo * spin_lock_irq(gcwq->lock). 9048cca0eeaSTejun Heo * 9058cca0eeaSTejun Heo * RETURNS: 9068cca0eeaSTejun Heo * Pointer to worker which is executing @work if found, NULL 9078cca0eeaSTejun Heo * otherwise. 9088cca0eeaSTejun Heo */ 9098cca0eeaSTejun Heo static struct worker *find_worker_executing_work(struct global_cwq *gcwq, 9108cca0eeaSTejun Heo struct work_struct *work) 9118cca0eeaSTejun Heo { 9128cca0eeaSTejun Heo return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), 9138cca0eeaSTejun Heo work); 9148cca0eeaSTejun Heo } 9158cca0eeaSTejun Heo 9168cca0eeaSTejun Heo /** 9177e11629dSTejun Heo * insert_work - insert a work into gcwq 9184690c4abSTejun Heo * @cwq: cwq @work belongs to 9194690c4abSTejun Heo * @work: work to insert 9204690c4abSTejun Heo * @head: insertion point 9214690c4abSTejun Heo * @extra_flags: extra WORK_STRUCT_* flags to set 9224690c4abSTejun Heo * 9237e11629dSTejun Heo * Insert @work which belongs to @cwq into @gcwq after @head. 9247e11629dSTejun Heo * @extra_flags is or'd to work_struct flags. 9254690c4abSTejun Heo * 9264690c4abSTejun Heo * CONTEXT: 9278b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 9281da177e4SLinus Torvalds */ 929b89deed3SOleg Nesterov static void insert_work(struct cpu_workqueue_struct *cwq, 9304690c4abSTejun Heo struct work_struct *work, struct list_head *head, 9314690c4abSTejun Heo unsigned int extra_flags) 932b89deed3SOleg Nesterov { 93363d95a91STejun Heo struct worker_pool *pool = cwq->pool; 934e1d8aa9fSFrederic Weisbecker 9354690c4abSTejun Heo /* we own @work, set data and link */ 9367a22ad75STejun Heo set_work_cwq(work, cwq, extra_flags); 9374690c4abSTejun Heo 9386e84d644SOleg Nesterov /* 9396e84d644SOleg Nesterov * Ensure that we get the right work->data if we see the 9406e84d644SOleg Nesterov * result of list_add() below, see try_to_grab_pending(). 9416e84d644SOleg Nesterov */ 9426e84d644SOleg Nesterov smp_wmb(); 9434690c4abSTejun Heo 9441a4d9b0aSOleg Nesterov list_add_tail(&work->entry, head); 945e22bee78STejun Heo 946e22bee78STejun Heo /* 947e22bee78STejun Heo * Ensure either worker_sched_deactivated() sees the above 948e22bee78STejun Heo * list_add_tail() or we see zero nr_running to avoid workers 949e22bee78STejun Heo * lying around lazily while there are works to be processed. 950e22bee78STejun Heo */ 951e22bee78STejun Heo smp_mb(); 952e22bee78STejun Heo 95363d95a91STejun Heo if (__need_more_worker(pool)) 95463d95a91STejun Heo wake_up_worker(pool); 955b89deed3SOleg Nesterov } 956b89deed3SOleg Nesterov 957c8efcc25STejun Heo /* 958c8efcc25STejun Heo * Test whether @work is being queued from another work executing on the 959c8efcc25STejun Heo * same workqueue. This is rather expensive and should only be used from 960c8efcc25STejun Heo * cold paths. 961c8efcc25STejun Heo */ 962c8efcc25STejun Heo static bool is_chained_work(struct workqueue_struct *wq) 963c8efcc25STejun Heo { 964c8efcc25STejun Heo unsigned long flags; 965c8efcc25STejun Heo unsigned int cpu; 966c8efcc25STejun Heo 967c8efcc25STejun Heo for_each_gcwq_cpu(cpu) { 968c8efcc25STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 969c8efcc25STejun Heo struct worker *worker; 970c8efcc25STejun Heo struct hlist_node *pos; 971c8efcc25STejun Heo int i; 972c8efcc25STejun Heo 973c8efcc25STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 974c8efcc25STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 975c8efcc25STejun Heo if (worker->task != current) 976c8efcc25STejun Heo continue; 977c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 978c8efcc25STejun Heo /* 979c8efcc25STejun Heo * I'm @worker, no locking necessary. See if @work 980c8efcc25STejun Heo * is headed to the same workqueue. 981c8efcc25STejun Heo */ 982c8efcc25STejun Heo return worker->current_cwq->wq == wq; 983c8efcc25STejun Heo } 984c8efcc25STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 985c8efcc25STejun Heo } 986c8efcc25STejun Heo return false; 987c8efcc25STejun Heo } 988c8efcc25STejun Heo 9894690c4abSTejun Heo static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, 9901da177e4SLinus Torvalds struct work_struct *work) 9911da177e4SLinus Torvalds { 992502ca9d8STejun Heo struct global_cwq *gcwq; 993502ca9d8STejun Heo struct cpu_workqueue_struct *cwq; 9941e19ffc6STejun Heo struct list_head *worklist; 9958a2e8e5dSTejun Heo unsigned int work_flags; 9961da177e4SLinus Torvalds unsigned long flags; 9971da177e4SLinus Torvalds 998dc186ad7SThomas Gleixner debug_work_activate(work); 9991e19ffc6STejun Heo 1000c8efcc25STejun Heo /* if dying, only works from the same workqueue are allowed */ 10019c5a2ba7STejun Heo if (unlikely(wq->flags & WQ_DRAINING) && 1002c8efcc25STejun Heo WARN_ON_ONCE(!is_chained_work(wq))) 1003e41e704bSTejun Heo return; 1004e41e704bSTejun Heo 1005c7fc77f7STejun Heo /* determine gcwq to use */ 1006c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1007c7fc77f7STejun Heo struct global_cwq *last_gcwq; 1008c7fc77f7STejun Heo 1009f3421797STejun Heo if (unlikely(cpu == WORK_CPU_UNBOUND)) 1010f3421797STejun Heo cpu = raw_smp_processor_id(); 1011f3421797STejun Heo 101218aa9effSTejun Heo /* 101318aa9effSTejun Heo * It's multi cpu. If @wq is non-reentrant and @work 101418aa9effSTejun Heo * was previously on a different cpu, it might still 101518aa9effSTejun Heo * be running there, in which case the work needs to 101618aa9effSTejun Heo * be queued on that cpu to guarantee non-reentrance. 101718aa9effSTejun Heo */ 1018502ca9d8STejun Heo gcwq = get_gcwq(cpu); 101918aa9effSTejun Heo if (wq->flags & WQ_NON_REENTRANT && 102018aa9effSTejun Heo (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { 102118aa9effSTejun Heo struct worker *worker; 102218aa9effSTejun Heo 102318aa9effSTejun Heo spin_lock_irqsave(&last_gcwq->lock, flags); 102418aa9effSTejun Heo 102518aa9effSTejun Heo worker = find_worker_executing_work(last_gcwq, work); 102618aa9effSTejun Heo 102718aa9effSTejun Heo if (worker && worker->current_cwq->wq == wq) 102818aa9effSTejun Heo gcwq = last_gcwq; 102918aa9effSTejun Heo else { 103018aa9effSTejun Heo /* meh... not running there, queue here */ 103118aa9effSTejun Heo spin_unlock_irqrestore(&last_gcwq->lock, flags); 103218aa9effSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 103318aa9effSTejun Heo } 103418aa9effSTejun Heo } else 10358b03ae3cSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1036f3421797STejun Heo } else { 1037f3421797STejun Heo gcwq = get_gcwq(WORK_CPU_UNBOUND); 1038f3421797STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 1039502ca9d8STejun Heo } 1040502ca9d8STejun Heo 1041502ca9d8STejun Heo /* gcwq determined, get cwq and queue */ 1042502ca9d8STejun Heo cwq = get_cwq(gcwq->cpu, wq); 1043cdadf009STejun Heo trace_workqueue_queue_work(cpu, cwq, work); 1044502ca9d8STejun Heo 1045f5b2552bSDan Carpenter if (WARN_ON(!list_empty(&work->entry))) { 1046f5b2552bSDan Carpenter spin_unlock_irqrestore(&gcwq->lock, flags); 1047f5b2552bSDan Carpenter return; 1048f5b2552bSDan Carpenter } 10491e19ffc6STejun Heo 105073f53c4aSTejun Heo cwq->nr_in_flight[cwq->work_color]++; 10518a2e8e5dSTejun Heo work_flags = work_color_to_flags(cwq->work_color); 10521e19ffc6STejun Heo 10531e19ffc6STejun Heo if (likely(cwq->nr_active < cwq->max_active)) { 1054cdadf009STejun Heo trace_workqueue_activate_work(work); 10551e19ffc6STejun Heo cwq->nr_active++; 10563270476aSTejun Heo worklist = &cwq->pool->worklist; 10578a2e8e5dSTejun Heo } else { 10588a2e8e5dSTejun Heo work_flags |= WORK_STRUCT_DELAYED; 10591e19ffc6STejun Heo worklist = &cwq->delayed_works; 10608a2e8e5dSTejun Heo } 10611e19ffc6STejun Heo 10628a2e8e5dSTejun Heo insert_work(cwq, work, worklist, work_flags); 10631e19ffc6STejun Heo 10648b03ae3cSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 10651da177e4SLinus Torvalds } 10661da177e4SLinus Torvalds 10670fcb78c2SRolf Eike Beer /** 10680fcb78c2SRolf Eike Beer * queue_work - queue work on a workqueue 10690fcb78c2SRolf Eike Beer * @wq: workqueue to use 10700fcb78c2SRolf Eike Beer * @work: work to queue 10710fcb78c2SRolf Eike Beer * 1072057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 10731da177e4SLinus Torvalds * 107400dfcaf7SOleg Nesterov * We queue the work to the CPU on which it was submitted, but if the CPU dies 107500dfcaf7SOleg Nesterov * it can be processed by another CPU. 10761da177e4SLinus Torvalds */ 10777ad5b3a5SHarvey Harrison int queue_work(struct workqueue_struct *wq, struct work_struct *work) 10781da177e4SLinus Torvalds { 1079ef1ca236SOleg Nesterov int ret; 10801da177e4SLinus Torvalds 1081ef1ca236SOleg Nesterov ret = queue_work_on(get_cpu(), wq, work); 1082a848e3b6SOleg Nesterov put_cpu(); 1083ef1ca236SOleg Nesterov 10841da177e4SLinus Torvalds return ret; 10851da177e4SLinus Torvalds } 1086ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_work); 10871da177e4SLinus Torvalds 1088c1a220e7SZhang Rui /** 1089c1a220e7SZhang Rui * queue_work_on - queue work on specific cpu 1090c1a220e7SZhang Rui * @cpu: CPU number to execute work on 1091c1a220e7SZhang Rui * @wq: workqueue to use 1092c1a220e7SZhang Rui * @work: work to queue 1093c1a220e7SZhang Rui * 1094c1a220e7SZhang Rui * Returns 0 if @work was already on a queue, non-zero otherwise. 1095c1a220e7SZhang Rui * 1096c1a220e7SZhang Rui * We queue the work to a specific CPU, the caller must ensure it 1097c1a220e7SZhang Rui * can't go away. 1098c1a220e7SZhang Rui */ 1099c1a220e7SZhang Rui int 1100c1a220e7SZhang Rui queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) 1101c1a220e7SZhang Rui { 1102c1a220e7SZhang Rui int ret = 0; 1103c1a220e7SZhang Rui 110422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 11054690c4abSTejun Heo __queue_work(cpu, wq, work); 1106c1a220e7SZhang Rui ret = 1; 1107c1a220e7SZhang Rui } 1108c1a220e7SZhang Rui return ret; 1109c1a220e7SZhang Rui } 1110c1a220e7SZhang Rui EXPORT_SYMBOL_GPL(queue_work_on); 1111c1a220e7SZhang Rui 11126d141c3fSLi Zefan static void delayed_work_timer_fn(unsigned long __data) 11131da177e4SLinus Torvalds { 111452bad64dSDavid Howells struct delayed_work *dwork = (struct delayed_work *)__data; 11157a22ad75STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); 11161da177e4SLinus Torvalds 11174690c4abSTejun Heo __queue_work(smp_processor_id(), cwq->wq, &dwork->work); 11181da177e4SLinus Torvalds } 11191da177e4SLinus Torvalds 11200fcb78c2SRolf Eike Beer /** 11210fcb78c2SRolf Eike Beer * queue_delayed_work - queue work on a workqueue after delay 11220fcb78c2SRolf Eike Beer * @wq: workqueue to use 1123af9997e4SRandy Dunlap * @dwork: delayable work to queue 11240fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11250fcb78c2SRolf Eike Beer * 1126057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11270fcb78c2SRolf Eike Beer */ 11287ad5b3a5SHarvey Harrison int queue_delayed_work(struct workqueue_struct *wq, 112952bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11301da177e4SLinus Torvalds { 113152bad64dSDavid Howells if (delay == 0) 113263bc0362SOleg Nesterov return queue_work(wq, &dwork->work); 11331da177e4SLinus Torvalds 113463bc0362SOleg Nesterov return queue_delayed_work_on(-1, wq, dwork, delay); 11351da177e4SLinus Torvalds } 1136ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work); 11371da177e4SLinus Torvalds 11380fcb78c2SRolf Eike Beer /** 11390fcb78c2SRolf Eike Beer * queue_delayed_work_on - queue work on specific CPU after delay 11400fcb78c2SRolf Eike Beer * @cpu: CPU number to execute work on 11410fcb78c2SRolf Eike Beer * @wq: workqueue to use 1142af9997e4SRandy Dunlap * @dwork: work to queue 11430fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait before queueing 11440fcb78c2SRolf Eike Beer * 1145057647fcSAlan Stern * Returns 0 if @work was already on a queue, non-zero otherwise. 11460fcb78c2SRolf Eike Beer */ 11477a6bc1cdSVenkatesh Pallipadi int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, 114852bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 11497a6bc1cdSVenkatesh Pallipadi { 11507a6bc1cdSVenkatesh Pallipadi int ret = 0; 115152bad64dSDavid Howells struct timer_list *timer = &dwork->timer; 115252bad64dSDavid Howells struct work_struct *work = &dwork->work; 11537a6bc1cdSVenkatesh Pallipadi 115422df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { 1155c7fc77f7STejun Heo unsigned int lcpu; 11567a22ad75STejun Heo 11577a6bc1cdSVenkatesh Pallipadi BUG_ON(timer_pending(timer)); 11587a6bc1cdSVenkatesh Pallipadi BUG_ON(!list_empty(&work->entry)); 11597a6bc1cdSVenkatesh Pallipadi 11608a3e77ccSAndrew Liu timer_stats_timer_set_start_info(&dwork->timer); 11618a3e77ccSAndrew Liu 11627a22ad75STejun Heo /* 11637a22ad75STejun Heo * This stores cwq for the moment, for the timer_fn. 11647a22ad75STejun Heo * Note that the work's gcwq is preserved to allow 11657a22ad75STejun Heo * reentrance detection for delayed works. 11667a22ad75STejun Heo */ 1167c7fc77f7STejun Heo if (!(wq->flags & WQ_UNBOUND)) { 1168c7fc77f7STejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 1169c7fc77f7STejun Heo 1170c7fc77f7STejun Heo if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND) 1171c7fc77f7STejun Heo lcpu = gcwq->cpu; 1172c7fc77f7STejun Heo else 1173c7fc77f7STejun Heo lcpu = raw_smp_processor_id(); 1174c7fc77f7STejun Heo } else 1175c7fc77f7STejun Heo lcpu = WORK_CPU_UNBOUND; 1176c7fc77f7STejun Heo 11777a22ad75STejun Heo set_work_cwq(work, get_cwq(lcpu, wq), 0); 1178c7fc77f7STejun Heo 11797a6bc1cdSVenkatesh Pallipadi timer->expires = jiffies + delay; 118052bad64dSDavid Howells timer->data = (unsigned long)dwork; 11817a6bc1cdSVenkatesh Pallipadi timer->function = delayed_work_timer_fn; 118263bc0362SOleg Nesterov 118363bc0362SOleg Nesterov if (unlikely(cpu >= 0)) 11847a6bc1cdSVenkatesh Pallipadi add_timer_on(timer, cpu); 118563bc0362SOleg Nesterov else 118663bc0362SOleg Nesterov add_timer(timer); 11877a6bc1cdSVenkatesh Pallipadi ret = 1; 11887a6bc1cdSVenkatesh Pallipadi } 11897a6bc1cdSVenkatesh Pallipadi return ret; 11907a6bc1cdSVenkatesh Pallipadi } 1191ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(queue_delayed_work_on); 11921da177e4SLinus Torvalds 1193c8e55f36STejun Heo /** 1194c8e55f36STejun Heo * worker_enter_idle - enter idle state 1195c8e55f36STejun Heo * @worker: worker which is entering idle state 1196c8e55f36STejun Heo * 1197c8e55f36STejun Heo * @worker is entering idle state. Update stats and idle timer if 1198c8e55f36STejun Heo * necessary. 1199c8e55f36STejun Heo * 1200c8e55f36STejun Heo * LOCKING: 1201c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1202c8e55f36STejun Heo */ 1203c8e55f36STejun Heo static void worker_enter_idle(struct worker *worker) 12041da177e4SLinus Torvalds { 1205bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1206bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1207c8e55f36STejun Heo 1208c8e55f36STejun Heo BUG_ON(worker->flags & WORKER_IDLE); 1209c8e55f36STejun Heo BUG_ON(!list_empty(&worker->entry) && 1210c8e55f36STejun Heo (worker->hentry.next || worker->hentry.pprev)); 1211c8e55f36STejun Heo 1212cb444766STejun Heo /* can't use worker_set_flags(), also called from start_worker() */ 1213cb444766STejun Heo worker->flags |= WORKER_IDLE; 1214bd7bdd43STejun Heo pool->nr_idle++; 1215e22bee78STejun Heo worker->last_active = jiffies; 1216c8e55f36STejun Heo 1217c8e55f36STejun Heo /* idle_list is LIFO */ 1218bd7bdd43STejun Heo list_add(&worker->entry, &pool->idle_list); 1219db7bccf4STejun Heo 1220403c821dSTejun Heo if (likely(gcwq->trustee_state != TRUSTEE_DONE)) { 122163d95a91STejun Heo if (too_many_workers(pool) && !timer_pending(&pool->idle_timer)) 1222bd7bdd43STejun Heo mod_timer(&pool->idle_timer, 1223e22bee78STejun Heo jiffies + IDLE_WORKER_TIMEOUT); 1224e22bee78STejun Heo } else 1225db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 1226cb444766STejun Heo 1227544ecf31STejun Heo /* 1228544ecf31STejun Heo * Sanity check nr_running. Because trustee releases gcwq->lock 1229403c821dSTejun Heo * between setting %WORKER_UNBOUND and zapping nr_running, the 1230544ecf31STejun Heo * warning may trigger spuriously. Check iff trustee is idle. 1231544ecf31STejun Heo */ 1232544ecf31STejun Heo WARN_ON_ONCE(gcwq->trustee_state == TRUSTEE_DONE && 1233bd7bdd43STejun Heo pool->nr_workers == pool->nr_idle && 123463d95a91STejun Heo atomic_read(get_pool_nr_running(pool))); 1235c8e55f36STejun Heo } 1236c8e55f36STejun Heo 1237c8e55f36STejun Heo /** 1238c8e55f36STejun Heo * worker_leave_idle - leave idle state 1239c8e55f36STejun Heo * @worker: worker which is leaving idle state 1240c8e55f36STejun Heo * 1241c8e55f36STejun Heo * @worker is leaving idle state. Update stats. 1242c8e55f36STejun Heo * 1243c8e55f36STejun Heo * LOCKING: 1244c8e55f36STejun Heo * spin_lock_irq(gcwq->lock). 1245c8e55f36STejun Heo */ 1246c8e55f36STejun Heo static void worker_leave_idle(struct worker *worker) 1247c8e55f36STejun Heo { 1248bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1249c8e55f36STejun Heo 1250c8e55f36STejun Heo BUG_ON(!(worker->flags & WORKER_IDLE)); 1251d302f017STejun Heo worker_clr_flags(worker, WORKER_IDLE); 1252bd7bdd43STejun Heo pool->nr_idle--; 1253c8e55f36STejun Heo list_del_init(&worker->entry); 1254c8e55f36STejun Heo } 1255c8e55f36STejun Heo 1256e22bee78STejun Heo /** 1257e22bee78STejun Heo * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq 1258e22bee78STejun Heo * @worker: self 1259e22bee78STejun Heo * 1260e22bee78STejun Heo * Works which are scheduled while the cpu is online must at least be 1261e22bee78STejun Heo * scheduled to a worker which is bound to the cpu so that if they are 1262e22bee78STejun Heo * flushed from cpu callbacks while cpu is going down, they are 1263e22bee78STejun Heo * guaranteed to execute on the cpu. 1264e22bee78STejun Heo * 1265e22bee78STejun Heo * This function is to be used by rogue workers and rescuers to bind 1266e22bee78STejun Heo * themselves to the target cpu and may race with cpu going down or 1267e22bee78STejun Heo * coming online. kthread_bind() can't be used because it may put the 1268e22bee78STejun Heo * worker to already dead cpu and set_cpus_allowed_ptr() can't be used 1269e22bee78STejun Heo * verbatim as it's best effort and blocking and gcwq may be 1270e22bee78STejun Heo * [dis]associated in the meantime. 1271e22bee78STejun Heo * 1272f2d5a0eeSTejun Heo * This function tries set_cpus_allowed() and locks gcwq and verifies the 1273f2d5a0eeSTejun Heo * binding against %GCWQ_DISASSOCIATED which is set during 1274f2d5a0eeSTejun Heo * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker 1275f2d5a0eeSTejun Heo * enters idle state or fetches works without dropping lock, it can 1276f2d5a0eeSTejun Heo * guarantee the scheduling requirement described in the first paragraph. 1277e22bee78STejun Heo * 1278e22bee78STejun Heo * CONTEXT: 1279e22bee78STejun Heo * Might sleep. Called without any lock but returns with gcwq->lock 1280e22bee78STejun Heo * held. 1281e22bee78STejun Heo * 1282e22bee78STejun Heo * RETURNS: 1283e22bee78STejun Heo * %true if the associated gcwq is online (@worker is successfully 1284e22bee78STejun Heo * bound), %false if offline. 1285e22bee78STejun Heo */ 1286e22bee78STejun Heo static bool worker_maybe_bind_and_lock(struct worker *worker) 1287972fa1c5SNamhyung Kim __acquires(&gcwq->lock) 1288e22bee78STejun Heo { 1289bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1290e22bee78STejun Heo struct task_struct *task = worker->task; 1291e22bee78STejun Heo 1292e22bee78STejun Heo while (true) { 1293e22bee78STejun Heo /* 1294e22bee78STejun Heo * The following call may fail, succeed or succeed 1295e22bee78STejun Heo * without actually migrating the task to the cpu if 1296e22bee78STejun Heo * it races with cpu hotunplug operation. Verify 1297e22bee78STejun Heo * against GCWQ_DISASSOCIATED. 1298e22bee78STejun Heo */ 1299f3421797STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) 1300e22bee78STejun Heo set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); 1301e22bee78STejun Heo 1302e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1303e22bee78STejun Heo if (gcwq->flags & GCWQ_DISASSOCIATED) 1304e22bee78STejun Heo return false; 1305e22bee78STejun Heo if (task_cpu(task) == gcwq->cpu && 1306e22bee78STejun Heo cpumask_equal(¤t->cpus_allowed, 1307e22bee78STejun Heo get_cpu_mask(gcwq->cpu))) 1308e22bee78STejun Heo return true; 1309e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1310e22bee78STejun Heo 13115035b20fSTejun Heo /* 13125035b20fSTejun Heo * We've raced with CPU hot[un]plug. Give it a breather 13135035b20fSTejun Heo * and retry migration. cond_resched() is required here; 13145035b20fSTejun Heo * otherwise, we might deadlock against cpu_stop trying to 13155035b20fSTejun Heo * bring down the CPU on non-preemptive kernel. 13165035b20fSTejun Heo */ 1317e22bee78STejun Heo cpu_relax(); 13185035b20fSTejun Heo cond_resched(); 1319e22bee78STejun Heo } 1320e22bee78STejun Heo } 1321e22bee78STejun Heo 132225511a47STejun Heo struct idle_rebind { 132325511a47STejun Heo int cnt; /* # workers to be rebound */ 132425511a47STejun Heo struct completion done; /* all workers rebound */ 132525511a47STejun Heo }; 132625511a47STejun Heo 1327e22bee78STejun Heo /* 132825511a47STejun Heo * Rebind an idle @worker to its CPU. During CPU onlining, this has to 132925511a47STejun Heo * happen synchronously for idle workers. worker_thread() will test 133025511a47STejun Heo * %WORKER_REBIND before leaving idle and call this function. 133125511a47STejun Heo */ 133225511a47STejun Heo static void idle_worker_rebind(struct worker *worker) 133325511a47STejun Heo { 133425511a47STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 133525511a47STejun Heo 133625511a47STejun Heo /* CPU must be online at this point */ 133725511a47STejun Heo WARN_ON(!worker_maybe_bind_and_lock(worker)); 133825511a47STejun Heo if (!--worker->idle_rebind->cnt) 133925511a47STejun Heo complete(&worker->idle_rebind->done); 134025511a47STejun Heo spin_unlock_irq(&worker->pool->gcwq->lock); 134125511a47STejun Heo 134225511a47STejun Heo /* we did our part, wait for rebind_workers() to finish up */ 134325511a47STejun Heo wait_event(gcwq->rebind_hold, !(worker->flags & WORKER_REBIND)); 134425511a47STejun Heo } 134525511a47STejun Heo 134625511a47STejun Heo /* 134725511a47STejun Heo * Function for @worker->rebind.work used to rebind unbound busy workers to 1348403c821dSTejun Heo * the associated cpu which is coming back online. This is scheduled by 1349403c821dSTejun Heo * cpu up but can race with other cpu hotplug operations and may be 1350403c821dSTejun Heo * executed twice without intervening cpu down. 1351e22bee78STejun Heo */ 135225511a47STejun Heo static void busy_worker_rebind_fn(struct work_struct *work) 1353e22bee78STejun Heo { 1354e22bee78STejun Heo struct worker *worker = container_of(work, struct worker, rebind_work); 1355bd7bdd43STejun Heo struct global_cwq *gcwq = worker->pool->gcwq; 1356e22bee78STejun Heo 1357e22bee78STejun Heo if (worker_maybe_bind_and_lock(worker)) 1358e22bee78STejun Heo worker_clr_flags(worker, WORKER_REBIND); 1359e22bee78STejun Heo 1360e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1361e22bee78STejun Heo } 1362e22bee78STejun Heo 136325511a47STejun Heo /** 136425511a47STejun Heo * rebind_workers - rebind all workers of a gcwq to the associated CPU 136525511a47STejun Heo * @gcwq: gcwq of interest 136625511a47STejun Heo * 136725511a47STejun Heo * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding 136825511a47STejun Heo * is different for idle and busy ones. 136925511a47STejun Heo * 137025511a47STejun Heo * The idle ones should be rebound synchronously and idle rebinding should 137125511a47STejun Heo * be complete before any worker starts executing work items with 137225511a47STejun Heo * concurrency management enabled; otherwise, scheduler may oops trying to 137325511a47STejun Heo * wake up non-local idle worker from wq_worker_sleeping(). 137425511a47STejun Heo * 137525511a47STejun Heo * This is achieved by repeatedly requesting rebinding until all idle 137625511a47STejun Heo * workers are known to have been rebound under @gcwq->lock and holding all 137725511a47STejun Heo * idle workers from becoming busy until idle rebinding is complete. 137825511a47STejun Heo * 137925511a47STejun Heo * Once idle workers are rebound, busy workers can be rebound as they 138025511a47STejun Heo * finish executing their current work items. Queueing the rebind work at 138125511a47STejun Heo * the head of their scheduled lists is enough. Note that nr_running will 138225511a47STejun Heo * be properbly bumped as busy workers rebind. 138325511a47STejun Heo * 138425511a47STejun Heo * On return, all workers are guaranteed to either be bound or have rebind 138525511a47STejun Heo * work item scheduled. 138625511a47STejun Heo */ 138725511a47STejun Heo static void rebind_workers(struct global_cwq *gcwq) 138825511a47STejun Heo __releases(&gcwq->lock) __acquires(&gcwq->lock) 138925511a47STejun Heo { 139025511a47STejun Heo struct idle_rebind idle_rebind; 139125511a47STejun Heo struct worker_pool *pool; 139225511a47STejun Heo struct worker *worker; 139325511a47STejun Heo struct hlist_node *pos; 139425511a47STejun Heo int i; 139525511a47STejun Heo 139625511a47STejun Heo lockdep_assert_held(&gcwq->lock); 139725511a47STejun Heo 139825511a47STejun Heo for_each_worker_pool(pool, gcwq) 139925511a47STejun Heo lockdep_assert_held(&pool->manager_mutex); 140025511a47STejun Heo 140125511a47STejun Heo /* 140225511a47STejun Heo * Rebind idle workers. Interlocked both ways. We wait for 140325511a47STejun Heo * workers to rebind via @idle_rebind.done. Workers will wait for 140425511a47STejun Heo * us to finish up by watching %WORKER_REBIND. 140525511a47STejun Heo */ 140625511a47STejun Heo init_completion(&idle_rebind.done); 140725511a47STejun Heo retry: 140825511a47STejun Heo idle_rebind.cnt = 1; 140925511a47STejun Heo INIT_COMPLETION(idle_rebind.done); 141025511a47STejun Heo 141125511a47STejun Heo /* set REBIND and kick idle ones, we'll wait for these later */ 141225511a47STejun Heo for_each_worker_pool(pool, gcwq) { 141325511a47STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 141425511a47STejun Heo if (worker->flags & WORKER_REBIND) 141525511a47STejun Heo continue; 141625511a47STejun Heo 141725511a47STejun Heo /* morph UNBOUND to REBIND */ 141825511a47STejun Heo worker->flags &= ~WORKER_UNBOUND; 141925511a47STejun Heo worker->flags |= WORKER_REBIND; 142025511a47STejun Heo 142125511a47STejun Heo idle_rebind.cnt++; 142225511a47STejun Heo worker->idle_rebind = &idle_rebind; 142325511a47STejun Heo 142425511a47STejun Heo /* worker_thread() will call idle_worker_rebind() */ 142525511a47STejun Heo wake_up_process(worker->task); 142625511a47STejun Heo } 142725511a47STejun Heo } 142825511a47STejun Heo 142925511a47STejun Heo if (--idle_rebind.cnt) { 143025511a47STejun Heo spin_unlock_irq(&gcwq->lock); 143125511a47STejun Heo wait_for_completion(&idle_rebind.done); 143225511a47STejun Heo spin_lock_irq(&gcwq->lock); 143325511a47STejun Heo /* busy ones might have become idle while waiting, retry */ 143425511a47STejun Heo goto retry; 143525511a47STejun Heo } 143625511a47STejun Heo 143725511a47STejun Heo /* 143825511a47STejun Heo * All idle workers are rebound and waiting for %WORKER_REBIND to 143925511a47STejun Heo * be cleared inside idle_worker_rebind(). Clear and release. 144025511a47STejun Heo * Clearing %WORKER_REBIND from this foreign context is safe 144125511a47STejun Heo * because these workers are still guaranteed to be idle. 144225511a47STejun Heo */ 144325511a47STejun Heo for_each_worker_pool(pool, gcwq) 144425511a47STejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 144525511a47STejun Heo worker->flags &= ~WORKER_REBIND; 144625511a47STejun Heo 144725511a47STejun Heo wake_up_all(&gcwq->rebind_hold); 144825511a47STejun Heo 144925511a47STejun Heo /* rebind busy workers */ 145025511a47STejun Heo for_each_busy_worker(worker, i, pos, gcwq) { 145125511a47STejun Heo struct work_struct *rebind_work = &worker->rebind_work; 145225511a47STejun Heo 145325511a47STejun Heo /* morph UNBOUND to REBIND */ 145425511a47STejun Heo worker->flags &= ~WORKER_UNBOUND; 145525511a47STejun Heo worker->flags |= WORKER_REBIND; 145625511a47STejun Heo 145725511a47STejun Heo if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, 145825511a47STejun Heo work_data_bits(rebind_work))) 145925511a47STejun Heo continue; 146025511a47STejun Heo 146125511a47STejun Heo /* wq doesn't matter, use the default one */ 146225511a47STejun Heo debug_work_activate(rebind_work); 146325511a47STejun Heo insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, 146425511a47STejun Heo worker->scheduled.next, 146525511a47STejun Heo work_color_to_flags(WORK_NO_COLOR)); 146625511a47STejun Heo } 146725511a47STejun Heo } 146825511a47STejun Heo 1469c34056a3STejun Heo static struct worker *alloc_worker(void) 1470c34056a3STejun Heo { 1471c34056a3STejun Heo struct worker *worker; 1472c34056a3STejun Heo 1473c34056a3STejun Heo worker = kzalloc(sizeof(*worker), GFP_KERNEL); 1474c8e55f36STejun Heo if (worker) { 1475c8e55f36STejun Heo INIT_LIST_HEAD(&worker->entry); 1476affee4b2STejun Heo INIT_LIST_HEAD(&worker->scheduled); 147725511a47STejun Heo INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn); 1478e22bee78STejun Heo /* on creation a worker is in !idle && prep state */ 1479e22bee78STejun Heo worker->flags = WORKER_PREP; 1480c8e55f36STejun Heo } 1481c34056a3STejun Heo return worker; 1482c34056a3STejun Heo } 1483c34056a3STejun Heo 1484c34056a3STejun Heo /** 1485c34056a3STejun Heo * create_worker - create a new workqueue worker 148663d95a91STejun Heo * @pool: pool the new worker will belong to 1487c34056a3STejun Heo * 148863d95a91STejun Heo * Create a new worker which is bound to @pool. The returned worker 1489c34056a3STejun Heo * can be started by calling start_worker() or destroyed using 1490c34056a3STejun Heo * destroy_worker(). 1491c34056a3STejun Heo * 1492c34056a3STejun Heo * CONTEXT: 1493c34056a3STejun Heo * Might sleep. Does GFP_KERNEL allocations. 1494c34056a3STejun Heo * 1495c34056a3STejun Heo * RETURNS: 1496c34056a3STejun Heo * Pointer to the newly created worker. 1497c34056a3STejun Heo */ 1498bc2ae0f5STejun Heo static struct worker *create_worker(struct worker_pool *pool) 1499c34056a3STejun Heo { 150063d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 15013270476aSTejun Heo const char *pri = worker_pool_pri(pool) ? "H" : ""; 1502c34056a3STejun Heo struct worker *worker = NULL; 1503f3421797STejun Heo int id = -1; 1504c34056a3STejun Heo 15058b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1506bd7bdd43STejun Heo while (ida_get_new(&pool->worker_ida, &id)) { 15078b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1508bd7bdd43STejun Heo if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL)) 1509c34056a3STejun Heo goto fail; 15108b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1511c34056a3STejun Heo } 15128b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1513c34056a3STejun Heo 1514c34056a3STejun Heo worker = alloc_worker(); 1515c34056a3STejun Heo if (!worker) 1516c34056a3STejun Heo goto fail; 1517c34056a3STejun Heo 1518bd7bdd43STejun Heo worker->pool = pool; 1519c34056a3STejun Heo worker->id = id; 1520c34056a3STejun Heo 1521bc2ae0f5STejun Heo if (gcwq->cpu != WORK_CPU_UNBOUND) 152294dcf29aSEric Dumazet worker->task = kthread_create_on_node(worker_thread, 15233270476aSTejun Heo worker, cpu_to_node(gcwq->cpu), 15243270476aSTejun Heo "kworker/%u:%d%s", gcwq->cpu, id, pri); 1525f3421797STejun Heo else 1526f3421797STejun Heo worker->task = kthread_create(worker_thread, worker, 15273270476aSTejun Heo "kworker/u:%d%s", id, pri); 1528c34056a3STejun Heo if (IS_ERR(worker->task)) 1529c34056a3STejun Heo goto fail; 1530c34056a3STejun Heo 15313270476aSTejun Heo if (worker_pool_pri(pool)) 15323270476aSTejun Heo set_user_nice(worker->task, HIGHPRI_NICE_LEVEL); 15333270476aSTejun Heo 1534db7bccf4STejun Heo /* 1535bc2ae0f5STejun Heo * Determine CPU binding of the new worker depending on 1536bc2ae0f5STejun Heo * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the 1537bc2ae0f5STejun Heo * flag remains stable across this function. See the comments 1538bc2ae0f5STejun Heo * above the flag definition for details. 1539bc2ae0f5STejun Heo * 1540bc2ae0f5STejun Heo * As an unbound worker may later become a regular one if CPU comes 1541bc2ae0f5STejun Heo * online, make sure every worker has %PF_THREAD_BOUND set. 1542db7bccf4STejun Heo */ 1543bc2ae0f5STejun Heo if (!(gcwq->flags & GCWQ_DISASSOCIATED)) { 15448b03ae3cSTejun Heo kthread_bind(worker->task, gcwq->cpu); 1545bc2ae0f5STejun Heo } else { 1546db7bccf4STejun Heo worker->task->flags |= PF_THREAD_BOUND; 1547f3421797STejun Heo worker->flags |= WORKER_UNBOUND; 1548f3421797STejun Heo } 1549c34056a3STejun Heo 1550c34056a3STejun Heo return worker; 1551c34056a3STejun Heo fail: 1552c34056a3STejun Heo if (id >= 0) { 15538b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1554bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 15558b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 1556c34056a3STejun Heo } 1557c34056a3STejun Heo kfree(worker); 1558c34056a3STejun Heo return NULL; 1559c34056a3STejun Heo } 1560c34056a3STejun Heo 1561c34056a3STejun Heo /** 1562c34056a3STejun Heo * start_worker - start a newly created worker 1563c34056a3STejun Heo * @worker: worker to start 1564c34056a3STejun Heo * 1565c8e55f36STejun Heo * Make the gcwq aware of @worker and start it. 1566c34056a3STejun Heo * 1567c34056a3STejun Heo * CONTEXT: 15688b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1569c34056a3STejun Heo */ 1570c34056a3STejun Heo static void start_worker(struct worker *worker) 1571c34056a3STejun Heo { 1572cb444766STejun Heo worker->flags |= WORKER_STARTED; 1573bd7bdd43STejun Heo worker->pool->nr_workers++; 1574c8e55f36STejun Heo worker_enter_idle(worker); 1575c34056a3STejun Heo wake_up_process(worker->task); 1576c34056a3STejun Heo } 1577c34056a3STejun Heo 1578c34056a3STejun Heo /** 1579c34056a3STejun Heo * destroy_worker - destroy a workqueue worker 1580c34056a3STejun Heo * @worker: worker to be destroyed 1581c34056a3STejun Heo * 1582c8e55f36STejun Heo * Destroy @worker and adjust @gcwq stats accordingly. 1583c8e55f36STejun Heo * 1584c8e55f36STejun Heo * CONTEXT: 1585c8e55f36STejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1586c34056a3STejun Heo */ 1587c34056a3STejun Heo static void destroy_worker(struct worker *worker) 1588c34056a3STejun Heo { 1589bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1590bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1591c34056a3STejun Heo int id = worker->id; 1592c34056a3STejun Heo 1593c34056a3STejun Heo /* sanity check frenzy */ 1594c34056a3STejun Heo BUG_ON(worker->current_work); 1595affee4b2STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 1596c34056a3STejun Heo 1597c8e55f36STejun Heo if (worker->flags & WORKER_STARTED) 1598bd7bdd43STejun Heo pool->nr_workers--; 1599c8e55f36STejun Heo if (worker->flags & WORKER_IDLE) 1600bd7bdd43STejun Heo pool->nr_idle--; 1601c8e55f36STejun Heo 1602c8e55f36STejun Heo list_del_init(&worker->entry); 1603cb444766STejun Heo worker->flags |= WORKER_DIE; 1604c8e55f36STejun Heo 1605c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 1606c8e55f36STejun Heo 1607c34056a3STejun Heo kthread_stop(worker->task); 1608c34056a3STejun Heo kfree(worker); 1609c34056a3STejun Heo 16108b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 1611bd7bdd43STejun Heo ida_remove(&pool->worker_ida, id); 1612c34056a3STejun Heo } 1613c34056a3STejun Heo 161463d95a91STejun Heo static void idle_worker_timeout(unsigned long __pool) 1615e22bee78STejun Heo { 161663d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 161763d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1618e22bee78STejun Heo 1619e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1620e22bee78STejun Heo 162163d95a91STejun Heo if (too_many_workers(pool)) { 1622e22bee78STejun Heo struct worker *worker; 1623e22bee78STejun Heo unsigned long expires; 1624e22bee78STejun Heo 1625e22bee78STejun Heo /* idle_list is kept in LIFO order, check the last one */ 162663d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1627e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1628e22bee78STejun Heo 1629e22bee78STejun Heo if (time_before(jiffies, expires)) 163063d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1631e22bee78STejun Heo else { 1632e22bee78STejun Heo /* it's been idle for too long, wake up manager */ 163311ebea50STejun Heo pool->flags |= POOL_MANAGE_WORKERS; 163463d95a91STejun Heo wake_up_worker(pool); 1635e22bee78STejun Heo } 1636e22bee78STejun Heo } 1637e22bee78STejun Heo 1638e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1639e22bee78STejun Heo } 1640e22bee78STejun Heo 1641e22bee78STejun Heo static bool send_mayday(struct work_struct *work) 1642e22bee78STejun Heo { 1643e22bee78STejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1644e22bee78STejun Heo struct workqueue_struct *wq = cwq->wq; 1645f3421797STejun Heo unsigned int cpu; 1646e22bee78STejun Heo 1647e22bee78STejun Heo if (!(wq->flags & WQ_RESCUER)) 1648e22bee78STejun Heo return false; 1649e22bee78STejun Heo 1650e22bee78STejun Heo /* mayday mayday mayday */ 1651bd7bdd43STejun Heo cpu = cwq->pool->gcwq->cpu; 1652f3421797STejun Heo /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ 1653f3421797STejun Heo if (cpu == WORK_CPU_UNBOUND) 1654f3421797STejun Heo cpu = 0; 1655f2e005aaSTejun Heo if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) 1656e22bee78STejun Heo wake_up_process(wq->rescuer->task); 1657e22bee78STejun Heo return true; 1658e22bee78STejun Heo } 1659e22bee78STejun Heo 166063d95a91STejun Heo static void gcwq_mayday_timeout(unsigned long __pool) 1661e22bee78STejun Heo { 166263d95a91STejun Heo struct worker_pool *pool = (void *)__pool; 166363d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 1664e22bee78STejun Heo struct work_struct *work; 1665e22bee78STejun Heo 1666e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1667e22bee78STejun Heo 166863d95a91STejun Heo if (need_to_create_worker(pool)) { 1669e22bee78STejun Heo /* 1670e22bee78STejun Heo * We've been trying to create a new worker but 1671e22bee78STejun Heo * haven't been successful. We might be hitting an 1672e22bee78STejun Heo * allocation deadlock. Send distress signals to 1673e22bee78STejun Heo * rescuers. 1674e22bee78STejun Heo */ 167563d95a91STejun Heo list_for_each_entry(work, &pool->worklist, entry) 1676e22bee78STejun Heo send_mayday(work); 1677e22bee78STejun Heo } 1678e22bee78STejun Heo 1679e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 1680e22bee78STejun Heo 168163d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL); 1682e22bee78STejun Heo } 1683e22bee78STejun Heo 1684e22bee78STejun Heo /** 1685e22bee78STejun Heo * maybe_create_worker - create a new worker if necessary 168663d95a91STejun Heo * @pool: pool to create a new worker for 1687e22bee78STejun Heo * 168863d95a91STejun Heo * Create a new worker for @pool if necessary. @pool is guaranteed to 1689e22bee78STejun Heo * have at least one idle worker on return from this function. If 1690e22bee78STejun Heo * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is 169163d95a91STejun Heo * sent to all rescuers with works scheduled on @pool to resolve 1692e22bee78STejun Heo * possible allocation deadlock. 1693e22bee78STejun Heo * 1694e22bee78STejun Heo * On return, need_to_create_worker() is guaranteed to be false and 1695e22bee78STejun Heo * may_start_working() true. 1696e22bee78STejun Heo * 1697e22bee78STejun Heo * LOCKING: 1698e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1699e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. Called only from 1700e22bee78STejun Heo * manager. 1701e22bee78STejun Heo * 1702e22bee78STejun Heo * RETURNS: 1703e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1704e22bee78STejun Heo * otherwise. 1705e22bee78STejun Heo */ 170663d95a91STejun Heo static bool maybe_create_worker(struct worker_pool *pool) 170706bd6ebfSNamhyung Kim __releases(&gcwq->lock) 170806bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 1709e22bee78STejun Heo { 171063d95a91STejun Heo struct global_cwq *gcwq = pool->gcwq; 171163d95a91STejun Heo 171263d95a91STejun Heo if (!need_to_create_worker(pool)) 1713e22bee78STejun Heo return false; 1714e22bee78STejun Heo restart: 17159f9c2364STejun Heo spin_unlock_irq(&gcwq->lock); 17169f9c2364STejun Heo 1717e22bee78STejun Heo /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ 171863d95a91STejun Heo mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); 1719e22bee78STejun Heo 1720e22bee78STejun Heo while (true) { 1721e22bee78STejun Heo struct worker *worker; 1722e22bee78STejun Heo 1723bc2ae0f5STejun Heo worker = create_worker(pool); 1724e22bee78STejun Heo if (worker) { 172563d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1726e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 1727e22bee78STejun Heo start_worker(worker); 172863d95a91STejun Heo BUG_ON(need_to_create_worker(pool)); 1729e22bee78STejun Heo return true; 1730e22bee78STejun Heo } 1731e22bee78STejun Heo 173263d95a91STejun Heo if (!need_to_create_worker(pool)) 1733e22bee78STejun Heo break; 1734e22bee78STejun Heo 1735e22bee78STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 1736e22bee78STejun Heo schedule_timeout(CREATE_COOLDOWN); 17379f9c2364STejun Heo 173863d95a91STejun Heo if (!need_to_create_worker(pool)) 1739e22bee78STejun Heo break; 1740e22bee78STejun Heo } 1741e22bee78STejun Heo 174263d95a91STejun Heo del_timer_sync(&pool->mayday_timer); 1743e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 174463d95a91STejun Heo if (need_to_create_worker(pool)) 1745e22bee78STejun Heo goto restart; 1746e22bee78STejun Heo return true; 1747e22bee78STejun Heo } 1748e22bee78STejun Heo 1749e22bee78STejun Heo /** 1750e22bee78STejun Heo * maybe_destroy_worker - destroy workers which have been idle for a while 175163d95a91STejun Heo * @pool: pool to destroy workers for 1752e22bee78STejun Heo * 175363d95a91STejun Heo * Destroy @pool workers which have been idle for longer than 1754e22bee78STejun Heo * IDLE_WORKER_TIMEOUT. 1755e22bee78STejun Heo * 1756e22bee78STejun Heo * LOCKING: 1757e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1758e22bee78STejun Heo * multiple times. Called only from manager. 1759e22bee78STejun Heo * 1760e22bee78STejun Heo * RETURNS: 1761e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true 1762e22bee78STejun Heo * otherwise. 1763e22bee78STejun Heo */ 176463d95a91STejun Heo static bool maybe_destroy_workers(struct worker_pool *pool) 1765e22bee78STejun Heo { 1766e22bee78STejun Heo bool ret = false; 1767e22bee78STejun Heo 176863d95a91STejun Heo while (too_many_workers(pool)) { 1769e22bee78STejun Heo struct worker *worker; 1770e22bee78STejun Heo unsigned long expires; 1771e22bee78STejun Heo 177263d95a91STejun Heo worker = list_entry(pool->idle_list.prev, struct worker, entry); 1773e22bee78STejun Heo expires = worker->last_active + IDLE_WORKER_TIMEOUT; 1774e22bee78STejun Heo 1775e22bee78STejun Heo if (time_before(jiffies, expires)) { 177663d95a91STejun Heo mod_timer(&pool->idle_timer, expires); 1777e22bee78STejun Heo break; 1778e22bee78STejun Heo } 1779e22bee78STejun Heo 1780e22bee78STejun Heo destroy_worker(worker); 1781e22bee78STejun Heo ret = true; 1782e22bee78STejun Heo } 1783e22bee78STejun Heo 1784e22bee78STejun Heo return ret; 1785e22bee78STejun Heo } 1786e22bee78STejun Heo 1787e22bee78STejun Heo /** 1788e22bee78STejun Heo * manage_workers - manage worker pool 1789e22bee78STejun Heo * @worker: self 1790e22bee78STejun Heo * 1791e22bee78STejun Heo * Assume the manager role and manage gcwq worker pool @worker belongs 1792e22bee78STejun Heo * to. At any given time, there can be only zero or one manager per 1793e22bee78STejun Heo * gcwq. The exclusion is handled automatically by this function. 1794e22bee78STejun Heo * 1795e22bee78STejun Heo * The caller can safely start processing works on false return. On 1796e22bee78STejun Heo * true return, it's guaranteed that need_to_create_worker() is false 1797e22bee78STejun Heo * and may_start_working() is true. 1798e22bee78STejun Heo * 1799e22bee78STejun Heo * CONTEXT: 1800e22bee78STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 1801e22bee78STejun Heo * multiple times. Does GFP_KERNEL allocations. 1802e22bee78STejun Heo * 1803e22bee78STejun Heo * RETURNS: 1804e22bee78STejun Heo * false if no action was taken and gcwq->lock stayed locked, true if 1805e22bee78STejun Heo * some action was taken. 1806e22bee78STejun Heo */ 1807e22bee78STejun Heo static bool manage_workers(struct worker *worker) 1808e22bee78STejun Heo { 180963d95a91STejun Heo struct worker_pool *pool = worker->pool; 1810e22bee78STejun Heo bool ret = false; 1811e22bee78STejun Heo 181260373152STejun Heo if (!mutex_trylock(&pool->manager_mutex)) 1813e22bee78STejun Heo return ret; 1814e22bee78STejun Heo 181511ebea50STejun Heo pool->flags &= ~POOL_MANAGE_WORKERS; 1816e22bee78STejun Heo 1817e22bee78STejun Heo /* 1818e22bee78STejun Heo * Destroy and then create so that may_start_working() is true 1819e22bee78STejun Heo * on return. 1820e22bee78STejun Heo */ 182163d95a91STejun Heo ret |= maybe_destroy_workers(pool); 182263d95a91STejun Heo ret |= maybe_create_worker(pool); 1823e22bee78STejun Heo 182460373152STejun Heo mutex_unlock(&pool->manager_mutex); 1825e22bee78STejun Heo return ret; 1826e22bee78STejun Heo } 1827e22bee78STejun Heo 1828a62428c0STejun Heo /** 1829affee4b2STejun Heo * move_linked_works - move linked works to a list 1830affee4b2STejun Heo * @work: start of series of works to be scheduled 1831affee4b2STejun Heo * @head: target list to append @work to 1832affee4b2STejun Heo * @nextp: out paramter for nested worklist walking 1833affee4b2STejun Heo * 1834affee4b2STejun Heo * Schedule linked works starting from @work to @head. Work series to 1835affee4b2STejun Heo * be scheduled starts at @work and includes any consecutive work with 1836affee4b2STejun Heo * WORK_STRUCT_LINKED set in its predecessor. 1837affee4b2STejun Heo * 1838affee4b2STejun Heo * If @nextp is not NULL, it's updated to point to the next work of 1839affee4b2STejun Heo * the last scheduled work. This allows move_linked_works() to be 1840affee4b2STejun Heo * nested inside outer list_for_each_entry_safe(). 1841affee4b2STejun Heo * 1842affee4b2STejun Heo * CONTEXT: 18438b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 1844affee4b2STejun Heo */ 1845affee4b2STejun Heo static void move_linked_works(struct work_struct *work, struct list_head *head, 1846affee4b2STejun Heo struct work_struct **nextp) 1847affee4b2STejun Heo { 1848affee4b2STejun Heo struct work_struct *n; 1849affee4b2STejun Heo 1850affee4b2STejun Heo /* 1851affee4b2STejun Heo * Linked worklist will always end before the end of the list, 1852affee4b2STejun Heo * use NULL for list head. 1853affee4b2STejun Heo */ 1854affee4b2STejun Heo list_for_each_entry_safe_from(work, n, NULL, entry) { 1855affee4b2STejun Heo list_move_tail(&work->entry, head); 1856affee4b2STejun Heo if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) 1857affee4b2STejun Heo break; 1858affee4b2STejun Heo } 1859affee4b2STejun Heo 1860affee4b2STejun Heo /* 1861affee4b2STejun Heo * If we're already inside safe list traversal and have moved 1862affee4b2STejun Heo * multiple works to the scheduled queue, the next position 1863affee4b2STejun Heo * needs to be updated. 1864affee4b2STejun Heo */ 1865affee4b2STejun Heo if (nextp) 1866affee4b2STejun Heo *nextp = n; 1867affee4b2STejun Heo } 1868affee4b2STejun Heo 18691e19ffc6STejun Heo static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) 18701e19ffc6STejun Heo { 18711e19ffc6STejun Heo struct work_struct *work = list_first_entry(&cwq->delayed_works, 18721da177e4SLinus Torvalds struct work_struct, entry); 18731e19ffc6STejun Heo 1874cdadf009STejun Heo trace_workqueue_activate_work(work); 18753270476aSTejun Heo move_linked_works(work, &cwq->pool->worklist, NULL); 18768a2e8e5dSTejun Heo __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); 18771e19ffc6STejun Heo cwq->nr_active++; 18781e19ffc6STejun Heo } 18791e19ffc6STejun Heo 1880affee4b2STejun Heo /** 188173f53c4aSTejun Heo * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight 188273f53c4aSTejun Heo * @cwq: cwq of interest 188373f53c4aSTejun Heo * @color: color of work which left the queue 18848a2e8e5dSTejun Heo * @delayed: for a delayed work 188573f53c4aSTejun Heo * 188673f53c4aSTejun Heo * A work either has completed or is removed from pending queue, 188773f53c4aSTejun Heo * decrement nr_in_flight of its cwq and handle workqueue flushing. 188873f53c4aSTejun Heo * 188973f53c4aSTejun Heo * CONTEXT: 18908b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 189173f53c4aSTejun Heo */ 18928a2e8e5dSTejun Heo static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color, 18938a2e8e5dSTejun Heo bool delayed) 189473f53c4aSTejun Heo { 189573f53c4aSTejun Heo /* ignore uncolored works */ 189673f53c4aSTejun Heo if (color == WORK_NO_COLOR) 189773f53c4aSTejun Heo return; 189873f53c4aSTejun Heo 189973f53c4aSTejun Heo cwq->nr_in_flight[color]--; 19001e19ffc6STejun Heo 19018a2e8e5dSTejun Heo if (!delayed) { 19028a2e8e5dSTejun Heo cwq->nr_active--; 1903502ca9d8STejun Heo if (!list_empty(&cwq->delayed_works)) { 19041e19ffc6STejun Heo /* one down, submit a delayed one */ 1905502ca9d8STejun Heo if (cwq->nr_active < cwq->max_active) 19061e19ffc6STejun Heo cwq_activate_first_delayed(cwq); 1907502ca9d8STejun Heo } 19088a2e8e5dSTejun Heo } 190973f53c4aSTejun Heo 191073f53c4aSTejun Heo /* is flush in progress and are we at the flushing tip? */ 191173f53c4aSTejun Heo if (likely(cwq->flush_color != color)) 191273f53c4aSTejun Heo return; 191373f53c4aSTejun Heo 191473f53c4aSTejun Heo /* are there still in-flight works? */ 191573f53c4aSTejun Heo if (cwq->nr_in_flight[color]) 191673f53c4aSTejun Heo return; 191773f53c4aSTejun Heo 191873f53c4aSTejun Heo /* this cwq is done, clear flush_color */ 191973f53c4aSTejun Heo cwq->flush_color = -1; 192073f53c4aSTejun Heo 192173f53c4aSTejun Heo /* 192273f53c4aSTejun Heo * If this was the last cwq, wake up the first flusher. It 192373f53c4aSTejun Heo * will handle the rest. 192473f53c4aSTejun Heo */ 192573f53c4aSTejun Heo if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) 192673f53c4aSTejun Heo complete(&cwq->wq->first_flusher->done); 192773f53c4aSTejun Heo } 192873f53c4aSTejun Heo 192973f53c4aSTejun Heo /** 1930a62428c0STejun Heo * process_one_work - process single work 1931c34056a3STejun Heo * @worker: self 1932a62428c0STejun Heo * @work: work to process 1933a62428c0STejun Heo * 1934a62428c0STejun Heo * Process @work. This function contains all the logics necessary to 1935a62428c0STejun Heo * process a single work including synchronization against and 1936a62428c0STejun Heo * interaction with other workers on the same cpu, queueing and 1937a62428c0STejun Heo * flushing. As long as context requirement is met, any worker can 1938a62428c0STejun Heo * call this function to process a work. 1939a62428c0STejun Heo * 1940a62428c0STejun Heo * CONTEXT: 19418b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which is released and regrabbed. 1942a62428c0STejun Heo */ 1943c34056a3STejun Heo static void process_one_work(struct worker *worker, struct work_struct *work) 194406bd6ebfSNamhyung Kim __releases(&gcwq->lock) 194506bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 19461da177e4SLinus Torvalds { 19477e11629dSTejun Heo struct cpu_workqueue_struct *cwq = get_work_cwq(work); 1948bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 1949bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 1950c8e55f36STejun Heo struct hlist_head *bwh = busy_worker_head(gcwq, work); 1951fb0e7bebSTejun Heo bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; 19526bb49e59SDavid Howells work_func_t f = work->func; 195373f53c4aSTejun Heo int work_color; 19547e11629dSTejun Heo struct worker *collision; 19554e6045f1SJohannes Berg #ifdef CONFIG_LOCKDEP 19564e6045f1SJohannes Berg /* 1957a62428c0STejun Heo * It is permissible to free the struct work_struct from 1958a62428c0STejun Heo * inside the function that is called from it, this we need to 1959a62428c0STejun Heo * take into account for lockdep too. To avoid bogus "held 1960a62428c0STejun Heo * lock freed" warnings as well as problems when looking into 1961a62428c0STejun Heo * work->lockdep_map, make a copy and use that here. 19624e6045f1SJohannes Berg */ 19634d82a1deSPeter Zijlstra struct lockdep_map lockdep_map; 19644d82a1deSPeter Zijlstra 19654d82a1deSPeter Zijlstra lockdep_copy_map(&lockdep_map, &work->lockdep_map); 19664e6045f1SJohannes Berg #endif 196725511a47STejun Heo WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) && 196825511a47STejun Heo raw_smp_processor_id() != gcwq->cpu); 196925511a47STejun Heo 19707e11629dSTejun Heo /* 19717e11629dSTejun Heo * A single work shouldn't be executed concurrently by 19727e11629dSTejun Heo * multiple workers on a single cpu. Check whether anyone is 19737e11629dSTejun Heo * already processing the work. If so, defer the work to the 19747e11629dSTejun Heo * currently executing one. 19757e11629dSTejun Heo */ 19767e11629dSTejun Heo collision = __find_worker_executing_work(gcwq, bwh, work); 19777e11629dSTejun Heo if (unlikely(collision)) { 19787e11629dSTejun Heo move_linked_works(work, &collision->scheduled, NULL); 19797e11629dSTejun Heo return; 19807e11629dSTejun Heo } 19811da177e4SLinus Torvalds 1982a62428c0STejun Heo /* claim and process */ 19831da177e4SLinus Torvalds debug_work_deactivate(work); 1984c8e55f36STejun Heo hlist_add_head(&worker->hentry, bwh); 1985c34056a3STejun Heo worker->current_work = work; 19868cca0eeaSTejun Heo worker->current_cwq = cwq; 198773f53c4aSTejun Heo work_color = get_work_color(work); 19887a22ad75STejun Heo 19897a22ad75STejun Heo /* record the current cpu number in the work data and dequeue */ 19907a22ad75STejun Heo set_work_cpu(work, gcwq->cpu); 1991a62428c0STejun Heo list_del_init(&work->entry); 1992a62428c0STejun Heo 1993649027d7STejun Heo /* 1994fb0e7bebSTejun Heo * CPU intensive works don't participate in concurrency 1995fb0e7bebSTejun Heo * management. They're the scheduler's responsibility. 1996fb0e7bebSTejun Heo */ 1997fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 1998fb0e7bebSTejun Heo worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); 1999fb0e7bebSTejun Heo 2000974271c4STejun Heo /* 2001974271c4STejun Heo * Unbound gcwq isn't concurrency managed and work items should be 2002974271c4STejun Heo * executed ASAP. Wake up another worker if necessary. 2003974271c4STejun Heo */ 200463d95a91STejun Heo if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool)) 200563d95a91STejun Heo wake_up_worker(pool); 2006974271c4STejun Heo 20078b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 20081da177e4SLinus Torvalds 200923b2e599SOleg Nesterov work_clear_pending(work); 2010e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 20113295f0efSIngo Molnar lock_map_acquire(&lockdep_map); 2012e36c886aSArjan van de Ven trace_workqueue_execute_start(work); 201365f27f38SDavid Howells f(work); 2014e36c886aSArjan van de Ven /* 2015e36c886aSArjan van de Ven * While we must be careful to not use "work" after this, the trace 2016e36c886aSArjan van de Ven * point will only record its address. 2017e36c886aSArjan van de Ven */ 2018e36c886aSArjan van de Ven trace_workqueue_execute_end(work); 20193295f0efSIngo Molnar lock_map_release(&lockdep_map); 20203295f0efSIngo Molnar lock_map_release(&cwq->wq->lockdep_map); 20211da177e4SLinus Torvalds 2022d5abe669SPeter Zijlstra if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { 2023d5abe669SPeter Zijlstra printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " 2024d5abe669SPeter Zijlstra "%s/0x%08x/%d\n", 2025a62428c0STejun Heo current->comm, preempt_count(), task_pid_nr(current)); 2026d5abe669SPeter Zijlstra printk(KERN_ERR " last function: "); 2027d5abe669SPeter Zijlstra print_symbol("%s\n", (unsigned long)f); 2028d5abe669SPeter Zijlstra debug_show_held_locks(current); 2029d5abe669SPeter Zijlstra dump_stack(); 2030d5abe669SPeter Zijlstra } 2031d5abe669SPeter Zijlstra 20328b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2033a62428c0STejun Heo 2034fb0e7bebSTejun Heo /* clear cpu intensive status */ 2035fb0e7bebSTejun Heo if (unlikely(cpu_intensive)) 2036fb0e7bebSTejun Heo worker_clr_flags(worker, WORKER_CPU_INTENSIVE); 2037fb0e7bebSTejun Heo 2038a62428c0STejun Heo /* we're done with it, release */ 2039c8e55f36STejun Heo hlist_del_init(&worker->hentry); 2040c34056a3STejun Heo worker->current_work = NULL; 20418cca0eeaSTejun Heo worker->current_cwq = NULL; 20428a2e8e5dSTejun Heo cwq_dec_nr_in_flight(cwq, work_color, false); 20431da177e4SLinus Torvalds } 20441da177e4SLinus Torvalds 2045affee4b2STejun Heo /** 2046affee4b2STejun Heo * process_scheduled_works - process scheduled works 2047affee4b2STejun Heo * @worker: self 2048affee4b2STejun Heo * 2049affee4b2STejun Heo * Process all scheduled works. Please note that the scheduled list 2050affee4b2STejun Heo * may change while processing a work, so this function repeatedly 2051affee4b2STejun Heo * fetches a work from the top and executes it. 2052affee4b2STejun Heo * 2053affee4b2STejun Heo * CONTEXT: 20548b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 2055affee4b2STejun Heo * multiple times. 2056affee4b2STejun Heo */ 2057affee4b2STejun Heo static void process_scheduled_works(struct worker *worker) 20581da177e4SLinus Torvalds { 2059affee4b2STejun Heo while (!list_empty(&worker->scheduled)) { 2060affee4b2STejun Heo struct work_struct *work = list_first_entry(&worker->scheduled, 2061a62428c0STejun Heo struct work_struct, entry); 2062c34056a3STejun Heo process_one_work(worker, work); 2063a62428c0STejun Heo } 20641da177e4SLinus Torvalds } 20651da177e4SLinus Torvalds 20664690c4abSTejun Heo /** 20674690c4abSTejun Heo * worker_thread - the worker thread function 2068c34056a3STejun Heo * @__worker: self 20694690c4abSTejun Heo * 2070e22bee78STejun Heo * The gcwq worker thread function. There's a single dynamic pool of 2071e22bee78STejun Heo * these per each cpu. These workers process all works regardless of 2072e22bee78STejun Heo * their specific target workqueue. The only exception is works which 2073e22bee78STejun Heo * belong to workqueues with a rescuer which will be explained in 2074e22bee78STejun Heo * rescuer_thread(). 20754690c4abSTejun Heo */ 2076c34056a3STejun Heo static int worker_thread(void *__worker) 20771da177e4SLinus Torvalds { 2078c34056a3STejun Heo struct worker *worker = __worker; 2079bd7bdd43STejun Heo struct worker_pool *pool = worker->pool; 2080bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 20811da177e4SLinus Torvalds 2082e22bee78STejun Heo /* tell the scheduler that this is a workqueue worker */ 2083e22bee78STejun Heo worker->task->flags |= PF_WQ_WORKER; 2084c8e55f36STejun Heo woke_up: 20858b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 2086affee4b2STejun Heo 208725511a47STejun Heo /* 208825511a47STejun Heo * DIE can be set only while idle and REBIND set while busy has 208925511a47STejun Heo * @worker->rebind_work scheduled. Checking here is enough. 209025511a47STejun Heo */ 209125511a47STejun Heo if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) { 2092c8e55f36STejun Heo spin_unlock_irq(&gcwq->lock); 209325511a47STejun Heo 209425511a47STejun Heo if (worker->flags & WORKER_DIE) { 2095e22bee78STejun Heo worker->task->flags &= ~PF_WQ_WORKER; 2096c8e55f36STejun Heo return 0; 2097c8e55f36STejun Heo } 2098c8e55f36STejun Heo 209925511a47STejun Heo idle_worker_rebind(worker); 210025511a47STejun Heo goto woke_up; 210125511a47STejun Heo } 210225511a47STejun Heo 2103c8e55f36STejun Heo worker_leave_idle(worker); 2104db7bccf4STejun Heo recheck: 2105e22bee78STejun Heo /* no more worker necessary? */ 210663d95a91STejun Heo if (!need_more_worker(pool)) 2107e22bee78STejun Heo goto sleep; 2108e22bee78STejun Heo 2109e22bee78STejun Heo /* do we need to manage? */ 211063d95a91STejun Heo if (unlikely(!may_start_working(pool)) && manage_workers(worker)) 2111e22bee78STejun Heo goto recheck; 2112e22bee78STejun Heo 2113c8e55f36STejun Heo /* 2114c8e55f36STejun Heo * ->scheduled list can only be filled while a worker is 2115c8e55f36STejun Heo * preparing to process a work or actually processing it. 2116c8e55f36STejun Heo * Make sure nobody diddled with it while I was sleeping. 2117c8e55f36STejun Heo */ 2118c8e55f36STejun Heo BUG_ON(!list_empty(&worker->scheduled)); 2119c8e55f36STejun Heo 2120e22bee78STejun Heo /* 2121e22bee78STejun Heo * When control reaches this point, we're guaranteed to have 2122e22bee78STejun Heo * at least one idle worker or that someone else has already 2123e22bee78STejun Heo * assumed the manager role. 2124e22bee78STejun Heo */ 2125e22bee78STejun Heo worker_clr_flags(worker, WORKER_PREP); 2126e22bee78STejun Heo 2127e22bee78STejun Heo do { 2128affee4b2STejun Heo struct work_struct *work = 2129bd7bdd43STejun Heo list_first_entry(&pool->worklist, 2130affee4b2STejun Heo struct work_struct, entry); 2131affee4b2STejun Heo 2132c8e55f36STejun Heo if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { 2133affee4b2STejun Heo /* optimization path, not strictly necessary */ 2134affee4b2STejun Heo process_one_work(worker, work); 2135affee4b2STejun Heo if (unlikely(!list_empty(&worker->scheduled))) 2136affee4b2STejun Heo process_scheduled_works(worker); 2137affee4b2STejun Heo } else { 2138c8e55f36STejun Heo move_linked_works(work, &worker->scheduled, NULL); 2139affee4b2STejun Heo process_scheduled_works(worker); 2140affee4b2STejun Heo } 214163d95a91STejun Heo } while (keep_working(pool)); 2142affee4b2STejun Heo 2143e22bee78STejun Heo worker_set_flags(worker, WORKER_PREP, false); 2144d313dd85STejun Heo sleep: 214563d95a91STejun Heo if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker)) 2146e22bee78STejun Heo goto recheck; 2147d313dd85STejun Heo 2148c8e55f36STejun Heo /* 2149e22bee78STejun Heo * gcwq->lock is held and there's no work to process and no 2150e22bee78STejun Heo * need to manage, sleep. Workers are woken up only while 2151e22bee78STejun Heo * holding gcwq->lock or from local cpu, so setting the 2152e22bee78STejun Heo * current state before releasing gcwq->lock is enough to 2153e22bee78STejun Heo * prevent losing any event. 2154c8e55f36STejun Heo */ 2155c8e55f36STejun Heo worker_enter_idle(worker); 2156c8e55f36STejun Heo __set_current_state(TASK_INTERRUPTIBLE); 21578b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 21581da177e4SLinus Torvalds schedule(); 2159c8e55f36STejun Heo goto woke_up; 21601da177e4SLinus Torvalds } 21611da177e4SLinus Torvalds 2162e22bee78STejun Heo /** 2163e22bee78STejun Heo * rescuer_thread - the rescuer thread function 2164e22bee78STejun Heo * @__wq: the associated workqueue 2165e22bee78STejun Heo * 2166e22bee78STejun Heo * Workqueue rescuer thread function. There's one rescuer for each 2167e22bee78STejun Heo * workqueue which has WQ_RESCUER set. 2168e22bee78STejun Heo * 2169e22bee78STejun Heo * Regular work processing on a gcwq may block trying to create a new 2170e22bee78STejun Heo * worker which uses GFP_KERNEL allocation which has slight chance of 2171e22bee78STejun Heo * developing into deadlock if some works currently on the same queue 2172e22bee78STejun Heo * need to be processed to satisfy the GFP_KERNEL allocation. This is 2173e22bee78STejun Heo * the problem rescuer solves. 2174e22bee78STejun Heo * 2175e22bee78STejun Heo * When such condition is possible, the gcwq summons rescuers of all 2176e22bee78STejun Heo * workqueues which have works queued on the gcwq and let them process 2177e22bee78STejun Heo * those works so that forward progress can be guaranteed. 2178e22bee78STejun Heo * 2179e22bee78STejun Heo * This should happen rarely. 2180e22bee78STejun Heo */ 2181e22bee78STejun Heo static int rescuer_thread(void *__wq) 2182e22bee78STejun Heo { 2183e22bee78STejun Heo struct workqueue_struct *wq = __wq; 2184e22bee78STejun Heo struct worker *rescuer = wq->rescuer; 2185e22bee78STejun Heo struct list_head *scheduled = &rescuer->scheduled; 2186f3421797STejun Heo bool is_unbound = wq->flags & WQ_UNBOUND; 2187e22bee78STejun Heo unsigned int cpu; 2188e22bee78STejun Heo 2189e22bee78STejun Heo set_user_nice(current, RESCUER_NICE_LEVEL); 2190e22bee78STejun Heo repeat: 2191e22bee78STejun Heo set_current_state(TASK_INTERRUPTIBLE); 21921da177e4SLinus Torvalds 21931da177e4SLinus Torvalds if (kthread_should_stop()) 2194e22bee78STejun Heo return 0; 21951da177e4SLinus Torvalds 2196f3421797STejun Heo /* 2197f3421797STejun Heo * See whether any cpu is asking for help. Unbounded 2198f3421797STejun Heo * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. 2199f3421797STejun Heo */ 2200f2e005aaSTejun Heo for_each_mayday_cpu(cpu, wq->mayday_mask) { 2201f3421797STejun Heo unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; 2202f3421797STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); 2203bd7bdd43STejun Heo struct worker_pool *pool = cwq->pool; 2204bd7bdd43STejun Heo struct global_cwq *gcwq = pool->gcwq; 2205e22bee78STejun Heo struct work_struct *work, *n; 2206e22bee78STejun Heo 2207e22bee78STejun Heo __set_current_state(TASK_RUNNING); 2208f2e005aaSTejun Heo mayday_clear_cpu(cpu, wq->mayday_mask); 2209e22bee78STejun Heo 2210e22bee78STejun Heo /* migrate to the target cpu if possible */ 2211bd7bdd43STejun Heo rescuer->pool = pool; 2212e22bee78STejun Heo worker_maybe_bind_and_lock(rescuer); 2213e22bee78STejun Heo 2214e22bee78STejun Heo /* 2215e22bee78STejun Heo * Slurp in all works issued via this workqueue and 2216e22bee78STejun Heo * process'em. 2217e22bee78STejun Heo */ 2218e22bee78STejun Heo BUG_ON(!list_empty(&rescuer->scheduled)); 2219bd7bdd43STejun Heo list_for_each_entry_safe(work, n, &pool->worklist, entry) 2220e22bee78STejun Heo if (get_work_cwq(work) == cwq) 2221e22bee78STejun Heo move_linked_works(work, scheduled, &n); 2222e22bee78STejun Heo 2223e22bee78STejun Heo process_scheduled_works(rescuer); 22247576958aSTejun Heo 22257576958aSTejun Heo /* 22267576958aSTejun Heo * Leave this gcwq. If keep_working() is %true, notify a 22277576958aSTejun Heo * regular worker; otherwise, we end up with 0 concurrency 22287576958aSTejun Heo * and stalling the execution. 22297576958aSTejun Heo */ 223063d95a91STejun Heo if (keep_working(pool)) 223163d95a91STejun Heo wake_up_worker(pool); 22327576958aSTejun Heo 2233e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 22341da177e4SLinus Torvalds } 22351da177e4SLinus Torvalds 2236e22bee78STejun Heo schedule(); 2237e22bee78STejun Heo goto repeat; 22381da177e4SLinus Torvalds } 22391da177e4SLinus Torvalds 2240fc2e4d70SOleg Nesterov struct wq_barrier { 2241fc2e4d70SOleg Nesterov struct work_struct work; 2242fc2e4d70SOleg Nesterov struct completion done; 2243fc2e4d70SOleg Nesterov }; 2244fc2e4d70SOleg Nesterov 2245fc2e4d70SOleg Nesterov static void wq_barrier_func(struct work_struct *work) 2246fc2e4d70SOleg Nesterov { 2247fc2e4d70SOleg Nesterov struct wq_barrier *barr = container_of(work, struct wq_barrier, work); 2248fc2e4d70SOleg Nesterov complete(&barr->done); 2249fc2e4d70SOleg Nesterov } 2250fc2e4d70SOleg Nesterov 22514690c4abSTejun Heo /** 22524690c4abSTejun Heo * insert_wq_barrier - insert a barrier work 22534690c4abSTejun Heo * @cwq: cwq to insert barrier into 22544690c4abSTejun Heo * @barr: wq_barrier to insert 2255affee4b2STejun Heo * @target: target work to attach @barr to 2256affee4b2STejun Heo * @worker: worker currently executing @target, NULL if @target is not executing 22574690c4abSTejun Heo * 2258affee4b2STejun Heo * @barr is linked to @target such that @barr is completed only after 2259affee4b2STejun Heo * @target finishes execution. Please note that the ordering 2260affee4b2STejun Heo * guarantee is observed only with respect to @target and on the local 2261affee4b2STejun Heo * cpu. 2262affee4b2STejun Heo * 2263affee4b2STejun Heo * Currently, a queued barrier can't be canceled. This is because 2264affee4b2STejun Heo * try_to_grab_pending() can't determine whether the work to be 2265affee4b2STejun Heo * grabbed is at the head of the queue and thus can't clear LINKED 2266affee4b2STejun Heo * flag of the previous work while there must be a valid next work 2267affee4b2STejun Heo * after a work with LINKED flag set. 2268affee4b2STejun Heo * 2269affee4b2STejun Heo * Note that when @worker is non-NULL, @target may be modified 2270affee4b2STejun Heo * underneath us, so we can't reliably determine cwq from @target. 22714690c4abSTejun Heo * 22724690c4abSTejun Heo * CONTEXT: 22738b03ae3cSTejun Heo * spin_lock_irq(gcwq->lock). 22744690c4abSTejun Heo */ 227583c22520SOleg Nesterov static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, 2276affee4b2STejun Heo struct wq_barrier *barr, 2277affee4b2STejun Heo struct work_struct *target, struct worker *worker) 2278fc2e4d70SOleg Nesterov { 2279affee4b2STejun Heo struct list_head *head; 2280affee4b2STejun Heo unsigned int linked = 0; 2281affee4b2STejun Heo 2282dc186ad7SThomas Gleixner /* 22838b03ae3cSTejun Heo * debugobject calls are safe here even with gcwq->lock locked 2284dc186ad7SThomas Gleixner * as we know for sure that this will not trigger any of the 2285dc186ad7SThomas Gleixner * checks and call back into the fixup functions where we 2286dc186ad7SThomas Gleixner * might deadlock. 2287dc186ad7SThomas Gleixner */ 2288ca1cab37SAndrew Morton INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); 228922df02bbSTejun Heo __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); 2290fc2e4d70SOleg Nesterov init_completion(&barr->done); 229183c22520SOleg Nesterov 2292affee4b2STejun Heo /* 2293affee4b2STejun Heo * If @target is currently being executed, schedule the 2294affee4b2STejun Heo * barrier to the worker; otherwise, put it after @target. 2295affee4b2STejun Heo */ 2296affee4b2STejun Heo if (worker) 2297affee4b2STejun Heo head = worker->scheduled.next; 2298affee4b2STejun Heo else { 2299affee4b2STejun Heo unsigned long *bits = work_data_bits(target); 2300affee4b2STejun Heo 2301affee4b2STejun Heo head = target->entry.next; 2302affee4b2STejun Heo /* there can already be other linked works, inherit and set */ 2303affee4b2STejun Heo linked = *bits & WORK_STRUCT_LINKED; 2304affee4b2STejun Heo __set_bit(WORK_STRUCT_LINKED_BIT, bits); 2305affee4b2STejun Heo } 2306affee4b2STejun Heo 2307dc186ad7SThomas Gleixner debug_work_activate(&barr->work); 2308affee4b2STejun Heo insert_work(cwq, &barr->work, head, 2309affee4b2STejun Heo work_color_to_flags(WORK_NO_COLOR) | linked); 2310fc2e4d70SOleg Nesterov } 2311fc2e4d70SOleg Nesterov 231273f53c4aSTejun Heo /** 231373f53c4aSTejun Heo * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing 231473f53c4aSTejun Heo * @wq: workqueue being flushed 231573f53c4aSTejun Heo * @flush_color: new flush color, < 0 for no-op 231673f53c4aSTejun Heo * @work_color: new work color, < 0 for no-op 231773f53c4aSTejun Heo * 231873f53c4aSTejun Heo * Prepare cwqs for workqueue flushing. 231973f53c4aSTejun Heo * 232073f53c4aSTejun Heo * If @flush_color is non-negative, flush_color on all cwqs should be 232173f53c4aSTejun Heo * -1. If no cwq has in-flight commands at the specified color, all 232273f53c4aSTejun Heo * cwq->flush_color's stay at -1 and %false is returned. If any cwq 232373f53c4aSTejun Heo * has in flight commands, its cwq->flush_color is set to 232473f53c4aSTejun Heo * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq 232573f53c4aSTejun Heo * wakeup logic is armed and %true is returned. 232673f53c4aSTejun Heo * 232773f53c4aSTejun Heo * The caller should have initialized @wq->first_flusher prior to 232873f53c4aSTejun Heo * calling this function with non-negative @flush_color. If 232973f53c4aSTejun Heo * @flush_color is negative, no flush color update is done and %false 233073f53c4aSTejun Heo * is returned. 233173f53c4aSTejun Heo * 233273f53c4aSTejun Heo * If @work_color is non-negative, all cwqs should have the same 233373f53c4aSTejun Heo * work_color which is previous to @work_color and all will be 233473f53c4aSTejun Heo * advanced to @work_color. 233573f53c4aSTejun Heo * 233673f53c4aSTejun Heo * CONTEXT: 233773f53c4aSTejun Heo * mutex_lock(wq->flush_mutex). 233873f53c4aSTejun Heo * 233973f53c4aSTejun Heo * RETURNS: 234073f53c4aSTejun Heo * %true if @flush_color >= 0 and there's something to flush. %false 234173f53c4aSTejun Heo * otherwise. 234273f53c4aSTejun Heo */ 234373f53c4aSTejun Heo static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, 234473f53c4aSTejun Heo int flush_color, int work_color) 23451da177e4SLinus Torvalds { 234673f53c4aSTejun Heo bool wait = false; 234773f53c4aSTejun Heo unsigned int cpu; 23481da177e4SLinus Torvalds 234973f53c4aSTejun Heo if (flush_color >= 0) { 235073f53c4aSTejun Heo BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); 235173f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 1); 2352dc186ad7SThomas Gleixner } 235314441960SOleg Nesterov 2354f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 235573f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2356bd7bdd43STejun Heo struct global_cwq *gcwq = cwq->pool->gcwq; 23571da177e4SLinus Torvalds 23588b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 235973f53c4aSTejun Heo 236073f53c4aSTejun Heo if (flush_color >= 0) { 236173f53c4aSTejun Heo BUG_ON(cwq->flush_color != -1); 236273f53c4aSTejun Heo 236373f53c4aSTejun Heo if (cwq->nr_in_flight[flush_color]) { 236473f53c4aSTejun Heo cwq->flush_color = flush_color; 236573f53c4aSTejun Heo atomic_inc(&wq->nr_cwqs_to_flush); 236673f53c4aSTejun Heo wait = true; 23671da177e4SLinus Torvalds } 236873f53c4aSTejun Heo } 236973f53c4aSTejun Heo 237073f53c4aSTejun Heo if (work_color >= 0) { 237173f53c4aSTejun Heo BUG_ON(work_color != work_next_color(cwq->work_color)); 237273f53c4aSTejun Heo cwq->work_color = work_color; 237373f53c4aSTejun Heo } 237473f53c4aSTejun Heo 23758b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 23761da177e4SLinus Torvalds } 23771da177e4SLinus Torvalds 237873f53c4aSTejun Heo if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) 237973f53c4aSTejun Heo complete(&wq->first_flusher->done); 238073f53c4aSTejun Heo 238173f53c4aSTejun Heo return wait; 238283c22520SOleg Nesterov } 23831da177e4SLinus Torvalds 23840fcb78c2SRolf Eike Beer /** 23851da177e4SLinus Torvalds * flush_workqueue - ensure that any scheduled work has run to completion. 23860fcb78c2SRolf Eike Beer * @wq: workqueue to flush 23871da177e4SLinus Torvalds * 23881da177e4SLinus Torvalds * Forces execution of the workqueue and blocks until its completion. 23891da177e4SLinus Torvalds * This is typically used in driver shutdown handlers. 23901da177e4SLinus Torvalds * 2391fc2e4d70SOleg Nesterov * We sleep until all works which were queued on entry have been handled, 2392fc2e4d70SOleg Nesterov * but we are not livelocked by new incoming ones. 23931da177e4SLinus Torvalds */ 23947ad5b3a5SHarvey Harrison void flush_workqueue(struct workqueue_struct *wq) 23951da177e4SLinus Torvalds { 239673f53c4aSTejun Heo struct wq_flusher this_flusher = { 239773f53c4aSTejun Heo .list = LIST_HEAD_INIT(this_flusher.list), 239873f53c4aSTejun Heo .flush_color = -1, 239973f53c4aSTejun Heo .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), 240073f53c4aSTejun Heo }; 240173f53c4aSTejun Heo int next_color; 2402b1f4ec17SOleg Nesterov 24033295f0efSIngo Molnar lock_map_acquire(&wq->lockdep_map); 24043295f0efSIngo Molnar lock_map_release(&wq->lockdep_map); 240573f53c4aSTejun Heo 240673f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 240773f53c4aSTejun Heo 240873f53c4aSTejun Heo /* 240973f53c4aSTejun Heo * Start-to-wait phase 241073f53c4aSTejun Heo */ 241173f53c4aSTejun Heo next_color = work_next_color(wq->work_color); 241273f53c4aSTejun Heo 241373f53c4aSTejun Heo if (next_color != wq->flush_color) { 241473f53c4aSTejun Heo /* 241573f53c4aSTejun Heo * Color space is not full. The current work_color 241673f53c4aSTejun Heo * becomes our flush_color and work_color is advanced 241773f53c4aSTejun Heo * by one. 241873f53c4aSTejun Heo */ 241973f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow)); 242073f53c4aSTejun Heo this_flusher.flush_color = wq->work_color; 242173f53c4aSTejun Heo wq->work_color = next_color; 242273f53c4aSTejun Heo 242373f53c4aSTejun Heo if (!wq->first_flusher) { 242473f53c4aSTejun Heo /* no flush in progress, become the first flusher */ 242573f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 242673f53c4aSTejun Heo 242773f53c4aSTejun Heo wq->first_flusher = &this_flusher; 242873f53c4aSTejun Heo 242973f53c4aSTejun Heo if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, 243073f53c4aSTejun Heo wq->work_color)) { 243173f53c4aSTejun Heo /* nothing to flush, done */ 243273f53c4aSTejun Heo wq->flush_color = next_color; 243373f53c4aSTejun Heo wq->first_flusher = NULL; 243473f53c4aSTejun Heo goto out_unlock; 243573f53c4aSTejun Heo } 243673f53c4aSTejun Heo } else { 243773f53c4aSTejun Heo /* wait in queue */ 243873f53c4aSTejun Heo BUG_ON(wq->flush_color == this_flusher.flush_color); 243973f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_queue); 244073f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 244173f53c4aSTejun Heo } 244273f53c4aSTejun Heo } else { 244373f53c4aSTejun Heo /* 244473f53c4aSTejun Heo * Oops, color space is full, wait on overflow queue. 244573f53c4aSTejun Heo * The next flush completion will assign us 244673f53c4aSTejun Heo * flush_color and transfer to flusher_queue. 244773f53c4aSTejun Heo */ 244873f53c4aSTejun Heo list_add_tail(&this_flusher.list, &wq->flusher_overflow); 244973f53c4aSTejun Heo } 245073f53c4aSTejun Heo 245173f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 245273f53c4aSTejun Heo 245373f53c4aSTejun Heo wait_for_completion(&this_flusher.done); 245473f53c4aSTejun Heo 245573f53c4aSTejun Heo /* 245673f53c4aSTejun Heo * Wake-up-and-cascade phase 245773f53c4aSTejun Heo * 245873f53c4aSTejun Heo * First flushers are responsible for cascading flushes and 245973f53c4aSTejun Heo * handling overflow. Non-first flushers can simply return. 246073f53c4aSTejun Heo */ 246173f53c4aSTejun Heo if (wq->first_flusher != &this_flusher) 246273f53c4aSTejun Heo return; 246373f53c4aSTejun Heo 246473f53c4aSTejun Heo mutex_lock(&wq->flush_mutex); 246573f53c4aSTejun Heo 24664ce48b37STejun Heo /* we might have raced, check again with mutex held */ 24674ce48b37STejun Heo if (wq->first_flusher != &this_flusher) 24684ce48b37STejun Heo goto out_unlock; 24694ce48b37STejun Heo 247073f53c4aSTejun Heo wq->first_flusher = NULL; 247173f53c4aSTejun Heo 247273f53c4aSTejun Heo BUG_ON(!list_empty(&this_flusher.list)); 247373f53c4aSTejun Heo BUG_ON(wq->flush_color != this_flusher.flush_color); 247473f53c4aSTejun Heo 247573f53c4aSTejun Heo while (true) { 247673f53c4aSTejun Heo struct wq_flusher *next, *tmp; 247773f53c4aSTejun Heo 247873f53c4aSTejun Heo /* complete all the flushers sharing the current flush color */ 247973f53c4aSTejun Heo list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { 248073f53c4aSTejun Heo if (next->flush_color != wq->flush_color) 248173f53c4aSTejun Heo break; 248273f53c4aSTejun Heo list_del_init(&next->list); 248373f53c4aSTejun Heo complete(&next->done); 248473f53c4aSTejun Heo } 248573f53c4aSTejun Heo 248673f53c4aSTejun Heo BUG_ON(!list_empty(&wq->flusher_overflow) && 248773f53c4aSTejun Heo wq->flush_color != work_next_color(wq->work_color)); 248873f53c4aSTejun Heo 248973f53c4aSTejun Heo /* this flush_color is finished, advance by one */ 249073f53c4aSTejun Heo wq->flush_color = work_next_color(wq->flush_color); 249173f53c4aSTejun Heo 249273f53c4aSTejun Heo /* one color has been freed, handle overflow queue */ 249373f53c4aSTejun Heo if (!list_empty(&wq->flusher_overflow)) { 249473f53c4aSTejun Heo /* 249573f53c4aSTejun Heo * Assign the same color to all overflowed 249673f53c4aSTejun Heo * flushers, advance work_color and append to 249773f53c4aSTejun Heo * flusher_queue. This is the start-to-wait 249873f53c4aSTejun Heo * phase for these overflowed flushers. 249973f53c4aSTejun Heo */ 250073f53c4aSTejun Heo list_for_each_entry(tmp, &wq->flusher_overflow, list) 250173f53c4aSTejun Heo tmp->flush_color = wq->work_color; 250273f53c4aSTejun Heo 250373f53c4aSTejun Heo wq->work_color = work_next_color(wq->work_color); 250473f53c4aSTejun Heo 250573f53c4aSTejun Heo list_splice_tail_init(&wq->flusher_overflow, 250673f53c4aSTejun Heo &wq->flusher_queue); 250773f53c4aSTejun Heo flush_workqueue_prep_cwqs(wq, -1, wq->work_color); 250873f53c4aSTejun Heo } 250973f53c4aSTejun Heo 251073f53c4aSTejun Heo if (list_empty(&wq->flusher_queue)) { 251173f53c4aSTejun Heo BUG_ON(wq->flush_color != wq->work_color); 251273f53c4aSTejun Heo break; 251373f53c4aSTejun Heo } 251473f53c4aSTejun Heo 251573f53c4aSTejun Heo /* 251673f53c4aSTejun Heo * Need to flush more colors. Make the next flusher 251773f53c4aSTejun Heo * the new first flusher and arm cwqs. 251873f53c4aSTejun Heo */ 251973f53c4aSTejun Heo BUG_ON(wq->flush_color == wq->work_color); 252073f53c4aSTejun Heo BUG_ON(wq->flush_color != next->flush_color); 252173f53c4aSTejun Heo 252273f53c4aSTejun Heo list_del_init(&next->list); 252373f53c4aSTejun Heo wq->first_flusher = next; 252473f53c4aSTejun Heo 252573f53c4aSTejun Heo if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) 252673f53c4aSTejun Heo break; 252773f53c4aSTejun Heo 252873f53c4aSTejun Heo /* 252973f53c4aSTejun Heo * Meh... this color is already done, clear first 253073f53c4aSTejun Heo * flusher and repeat cascading. 253173f53c4aSTejun Heo */ 253273f53c4aSTejun Heo wq->first_flusher = NULL; 253373f53c4aSTejun Heo } 253473f53c4aSTejun Heo 253573f53c4aSTejun Heo out_unlock: 253673f53c4aSTejun Heo mutex_unlock(&wq->flush_mutex); 25371da177e4SLinus Torvalds } 2538ae90dd5dSDave Jones EXPORT_SYMBOL_GPL(flush_workqueue); 25391da177e4SLinus Torvalds 25409c5a2ba7STejun Heo /** 25419c5a2ba7STejun Heo * drain_workqueue - drain a workqueue 25429c5a2ba7STejun Heo * @wq: workqueue to drain 25439c5a2ba7STejun Heo * 25449c5a2ba7STejun Heo * Wait until the workqueue becomes empty. While draining is in progress, 25459c5a2ba7STejun Heo * only chain queueing is allowed. IOW, only currently pending or running 25469c5a2ba7STejun Heo * work items on @wq can queue further work items on it. @wq is flushed 25479c5a2ba7STejun Heo * repeatedly until it becomes empty. The number of flushing is detemined 25489c5a2ba7STejun Heo * by the depth of chaining and should be relatively short. Whine if it 25499c5a2ba7STejun Heo * takes too long. 25509c5a2ba7STejun Heo */ 25519c5a2ba7STejun Heo void drain_workqueue(struct workqueue_struct *wq) 25529c5a2ba7STejun Heo { 25539c5a2ba7STejun Heo unsigned int flush_cnt = 0; 25549c5a2ba7STejun Heo unsigned int cpu; 25559c5a2ba7STejun Heo 25569c5a2ba7STejun Heo /* 25579c5a2ba7STejun Heo * __queue_work() needs to test whether there are drainers, is much 25589c5a2ba7STejun Heo * hotter than drain_workqueue() and already looks at @wq->flags. 25599c5a2ba7STejun Heo * Use WQ_DRAINING so that queue doesn't have to check nr_drainers. 25609c5a2ba7STejun Heo */ 25619c5a2ba7STejun Heo spin_lock(&workqueue_lock); 25629c5a2ba7STejun Heo if (!wq->nr_drainers++) 25639c5a2ba7STejun Heo wq->flags |= WQ_DRAINING; 25649c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 25659c5a2ba7STejun Heo reflush: 25669c5a2ba7STejun Heo flush_workqueue(wq); 25679c5a2ba7STejun Heo 25689c5a2ba7STejun Heo for_each_cwq_cpu(cpu, wq) { 25699c5a2ba7STejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 2570fa2563e4SThomas Tuttle bool drained; 25719c5a2ba7STejun Heo 2572bd7bdd43STejun Heo spin_lock_irq(&cwq->pool->gcwq->lock); 2573fa2563e4SThomas Tuttle drained = !cwq->nr_active && list_empty(&cwq->delayed_works); 2574bd7bdd43STejun Heo spin_unlock_irq(&cwq->pool->gcwq->lock); 2575fa2563e4SThomas Tuttle 2576fa2563e4SThomas Tuttle if (drained) 25779c5a2ba7STejun Heo continue; 25789c5a2ba7STejun Heo 25799c5a2ba7STejun Heo if (++flush_cnt == 10 || 25809c5a2ba7STejun Heo (flush_cnt % 100 == 0 && flush_cnt <= 1000)) 25819c5a2ba7STejun Heo pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n", 25829c5a2ba7STejun Heo wq->name, flush_cnt); 25839c5a2ba7STejun Heo goto reflush; 25849c5a2ba7STejun Heo } 25859c5a2ba7STejun Heo 25869c5a2ba7STejun Heo spin_lock(&workqueue_lock); 25879c5a2ba7STejun Heo if (!--wq->nr_drainers) 25889c5a2ba7STejun Heo wq->flags &= ~WQ_DRAINING; 25899c5a2ba7STejun Heo spin_unlock(&workqueue_lock); 25909c5a2ba7STejun Heo } 25919c5a2ba7STejun Heo EXPORT_SYMBOL_GPL(drain_workqueue); 25929c5a2ba7STejun Heo 2593baf59022STejun Heo static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, 2594baf59022STejun Heo bool wait_executing) 2595baf59022STejun Heo { 2596baf59022STejun Heo struct worker *worker = NULL; 2597baf59022STejun Heo struct global_cwq *gcwq; 2598baf59022STejun Heo struct cpu_workqueue_struct *cwq; 2599baf59022STejun Heo 2600baf59022STejun Heo might_sleep(); 2601baf59022STejun Heo gcwq = get_work_gcwq(work); 2602baf59022STejun Heo if (!gcwq) 2603baf59022STejun Heo return false; 2604baf59022STejun Heo 2605baf59022STejun Heo spin_lock_irq(&gcwq->lock); 2606baf59022STejun Heo if (!list_empty(&work->entry)) { 2607baf59022STejun Heo /* 2608baf59022STejun Heo * See the comment near try_to_grab_pending()->smp_rmb(). 2609baf59022STejun Heo * If it was re-queued to a different gcwq under us, we 2610baf59022STejun Heo * are not going to wait. 2611baf59022STejun Heo */ 2612baf59022STejun Heo smp_rmb(); 2613baf59022STejun Heo cwq = get_work_cwq(work); 2614bd7bdd43STejun Heo if (unlikely(!cwq || gcwq != cwq->pool->gcwq)) 2615baf59022STejun Heo goto already_gone; 2616baf59022STejun Heo } else if (wait_executing) { 2617baf59022STejun Heo worker = find_worker_executing_work(gcwq, work); 2618baf59022STejun Heo if (!worker) 2619baf59022STejun Heo goto already_gone; 2620baf59022STejun Heo cwq = worker->current_cwq; 2621baf59022STejun Heo } else 2622baf59022STejun Heo goto already_gone; 2623baf59022STejun Heo 2624baf59022STejun Heo insert_wq_barrier(cwq, barr, work, worker); 2625baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2626baf59022STejun Heo 2627e159489bSTejun Heo /* 2628e159489bSTejun Heo * If @max_active is 1 or rescuer is in use, flushing another work 2629e159489bSTejun Heo * item on the same workqueue may lead to deadlock. Make sure the 2630e159489bSTejun Heo * flusher is not running on the same workqueue by verifying write 2631e159489bSTejun Heo * access. 2632e159489bSTejun Heo */ 2633e159489bSTejun Heo if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) 2634baf59022STejun Heo lock_map_acquire(&cwq->wq->lockdep_map); 2635e159489bSTejun Heo else 2636e159489bSTejun Heo lock_map_acquire_read(&cwq->wq->lockdep_map); 2637baf59022STejun Heo lock_map_release(&cwq->wq->lockdep_map); 2638e159489bSTejun Heo 2639baf59022STejun Heo return true; 2640baf59022STejun Heo already_gone: 2641baf59022STejun Heo spin_unlock_irq(&gcwq->lock); 2642baf59022STejun Heo return false; 2643baf59022STejun Heo } 2644baf59022STejun Heo 2645db700897SOleg Nesterov /** 2646401a8d04STejun Heo * flush_work - wait for a work to finish executing the last queueing instance 2647401a8d04STejun Heo * @work: the work to flush 2648db700897SOleg Nesterov * 2649401a8d04STejun Heo * Wait until @work has finished execution. This function considers 2650401a8d04STejun Heo * only the last queueing instance of @work. If @work has been 2651401a8d04STejun Heo * enqueued across different CPUs on a non-reentrant workqueue or on 2652401a8d04STejun Heo * multiple workqueues, @work might still be executing on return on 2653401a8d04STejun Heo * some of the CPUs from earlier queueing. 2654a67da70dSOleg Nesterov * 2655401a8d04STejun Heo * If @work was queued only on a non-reentrant, ordered or unbound 2656401a8d04STejun Heo * workqueue, @work is guaranteed to be idle on return if it hasn't 2657401a8d04STejun Heo * been requeued since flush started. 2658401a8d04STejun Heo * 2659401a8d04STejun Heo * RETURNS: 2660401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2661401a8d04STejun Heo * %false if it was already idle. 2662db700897SOleg Nesterov */ 2663401a8d04STejun Heo bool flush_work(struct work_struct *work) 2664db700897SOleg Nesterov { 2665db700897SOleg Nesterov struct wq_barrier barr; 2666db700897SOleg Nesterov 26670976dfc1SStephen Boyd lock_map_acquire(&work->lockdep_map); 26680976dfc1SStephen Boyd lock_map_release(&work->lockdep_map); 26690976dfc1SStephen Boyd 2670baf59022STejun Heo if (start_flush_work(work, &barr, true)) { 2671db700897SOleg Nesterov wait_for_completion(&barr.done); 2672dc186ad7SThomas Gleixner destroy_work_on_stack(&barr.work); 2673401a8d04STejun Heo return true; 2674baf59022STejun Heo } else 2675401a8d04STejun Heo return false; 2676db700897SOleg Nesterov } 2677db700897SOleg Nesterov EXPORT_SYMBOL_GPL(flush_work); 2678db700897SOleg Nesterov 2679401a8d04STejun Heo static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) 2680401a8d04STejun Heo { 2681401a8d04STejun Heo struct wq_barrier barr; 2682401a8d04STejun Heo struct worker *worker; 2683401a8d04STejun Heo 2684401a8d04STejun Heo spin_lock_irq(&gcwq->lock); 2685401a8d04STejun Heo 2686401a8d04STejun Heo worker = find_worker_executing_work(gcwq, work); 2687401a8d04STejun Heo if (unlikely(worker)) 2688401a8d04STejun Heo insert_wq_barrier(worker->current_cwq, &barr, work, worker); 2689401a8d04STejun Heo 2690401a8d04STejun Heo spin_unlock_irq(&gcwq->lock); 2691401a8d04STejun Heo 2692401a8d04STejun Heo if (unlikely(worker)) { 2693401a8d04STejun Heo wait_for_completion(&barr.done); 2694401a8d04STejun Heo destroy_work_on_stack(&barr.work); 2695401a8d04STejun Heo return true; 2696401a8d04STejun Heo } else 2697401a8d04STejun Heo return false; 2698401a8d04STejun Heo } 2699401a8d04STejun Heo 2700401a8d04STejun Heo static bool wait_on_work(struct work_struct *work) 2701401a8d04STejun Heo { 2702401a8d04STejun Heo bool ret = false; 2703401a8d04STejun Heo int cpu; 2704401a8d04STejun Heo 2705401a8d04STejun Heo might_sleep(); 2706401a8d04STejun Heo 2707401a8d04STejun Heo lock_map_acquire(&work->lockdep_map); 2708401a8d04STejun Heo lock_map_release(&work->lockdep_map); 2709401a8d04STejun Heo 2710401a8d04STejun Heo for_each_gcwq_cpu(cpu) 2711401a8d04STejun Heo ret |= wait_on_cpu_work(get_gcwq(cpu), work); 2712401a8d04STejun Heo return ret; 2713401a8d04STejun Heo } 2714401a8d04STejun Heo 271509383498STejun Heo /** 271609383498STejun Heo * flush_work_sync - wait until a work has finished execution 271709383498STejun Heo * @work: the work to flush 271809383498STejun Heo * 271909383498STejun Heo * Wait until @work has finished execution. On return, it's 272009383498STejun Heo * guaranteed that all queueing instances of @work which happened 272109383498STejun Heo * before this function is called are finished. In other words, if 272209383498STejun Heo * @work hasn't been requeued since this function was called, @work is 272309383498STejun Heo * guaranteed to be idle on return. 272409383498STejun Heo * 272509383498STejun Heo * RETURNS: 272609383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 272709383498STejun Heo * %false if it was already idle. 272809383498STejun Heo */ 272909383498STejun Heo bool flush_work_sync(struct work_struct *work) 273009383498STejun Heo { 273109383498STejun Heo struct wq_barrier barr; 273209383498STejun Heo bool pending, waited; 273309383498STejun Heo 273409383498STejun Heo /* we'll wait for executions separately, queue barr only if pending */ 273509383498STejun Heo pending = start_flush_work(work, &barr, false); 273609383498STejun Heo 273709383498STejun Heo /* wait for executions to finish */ 273809383498STejun Heo waited = wait_on_work(work); 273909383498STejun Heo 274009383498STejun Heo /* wait for the pending one */ 274109383498STejun Heo if (pending) { 274209383498STejun Heo wait_for_completion(&barr.done); 274309383498STejun Heo destroy_work_on_stack(&barr.work); 274409383498STejun Heo } 274509383498STejun Heo 274609383498STejun Heo return pending || waited; 274709383498STejun Heo } 274809383498STejun Heo EXPORT_SYMBOL_GPL(flush_work_sync); 274909383498STejun Heo 27506e84d644SOleg Nesterov /* 27511f1f642eSOleg Nesterov * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, 27526e84d644SOleg Nesterov * so this work can't be re-armed in any way. 27536e84d644SOleg Nesterov */ 27546e84d644SOleg Nesterov static int try_to_grab_pending(struct work_struct *work) 27556e84d644SOleg Nesterov { 27568b03ae3cSTejun Heo struct global_cwq *gcwq; 27571f1f642eSOleg Nesterov int ret = -1; 27586e84d644SOleg Nesterov 275922df02bbSTejun Heo if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) 27601f1f642eSOleg Nesterov return 0; 27616e84d644SOleg Nesterov 27626e84d644SOleg Nesterov /* 27636e84d644SOleg Nesterov * The queueing is in progress, or it is already queued. Try to 27646e84d644SOleg Nesterov * steal it from ->worklist without clearing WORK_STRUCT_PENDING. 27656e84d644SOleg Nesterov */ 27667a22ad75STejun Heo gcwq = get_work_gcwq(work); 27677a22ad75STejun Heo if (!gcwq) 27686e84d644SOleg Nesterov return ret; 27696e84d644SOleg Nesterov 27708b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 27716e84d644SOleg Nesterov if (!list_empty(&work->entry)) { 27726e84d644SOleg Nesterov /* 27737a22ad75STejun Heo * This work is queued, but perhaps we locked the wrong gcwq. 27746e84d644SOleg Nesterov * In that case we must see the new value after rmb(), see 27756e84d644SOleg Nesterov * insert_work()->wmb(). 27766e84d644SOleg Nesterov */ 27776e84d644SOleg Nesterov smp_rmb(); 27787a22ad75STejun Heo if (gcwq == get_work_gcwq(work)) { 2779dc186ad7SThomas Gleixner debug_work_deactivate(work); 27806e84d644SOleg Nesterov list_del_init(&work->entry); 27817a22ad75STejun Heo cwq_dec_nr_in_flight(get_work_cwq(work), 27828a2e8e5dSTejun Heo get_work_color(work), 27838a2e8e5dSTejun Heo *work_data_bits(work) & WORK_STRUCT_DELAYED); 27846e84d644SOleg Nesterov ret = 1; 27856e84d644SOleg Nesterov } 27866e84d644SOleg Nesterov } 27878b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 27886e84d644SOleg Nesterov 27896e84d644SOleg Nesterov return ret; 27906e84d644SOleg Nesterov } 27916e84d644SOleg Nesterov 2792401a8d04STejun Heo static bool __cancel_work_timer(struct work_struct *work, 27931f1f642eSOleg Nesterov struct timer_list* timer) 27941f1f642eSOleg Nesterov { 27951f1f642eSOleg Nesterov int ret; 27961f1f642eSOleg Nesterov 27971f1f642eSOleg Nesterov do { 27981f1f642eSOleg Nesterov ret = (timer && likely(del_timer(timer))); 27991f1f642eSOleg Nesterov if (!ret) 28001f1f642eSOleg Nesterov ret = try_to_grab_pending(work); 28011f1f642eSOleg Nesterov wait_on_work(work); 28021f1f642eSOleg Nesterov } while (unlikely(ret < 0)); 28031f1f642eSOleg Nesterov 28047a22ad75STejun Heo clear_work_data(work); 28051f1f642eSOleg Nesterov return ret; 28061f1f642eSOleg Nesterov } 28071f1f642eSOleg Nesterov 28086e84d644SOleg Nesterov /** 2809401a8d04STejun Heo * cancel_work_sync - cancel a work and wait for it to finish 2810401a8d04STejun Heo * @work: the work to cancel 28116e84d644SOleg Nesterov * 2812401a8d04STejun Heo * Cancel @work and wait for its execution to finish. This function 2813401a8d04STejun Heo * can be used even if the work re-queues itself or migrates to 2814401a8d04STejun Heo * another workqueue. On return from this function, @work is 2815401a8d04STejun Heo * guaranteed to be not pending or executing on any CPU. 28161f1f642eSOleg Nesterov * 2817401a8d04STejun Heo * cancel_work_sync(&delayed_work->work) must not be used for 2818401a8d04STejun Heo * delayed_work's. Use cancel_delayed_work_sync() instead. 28196e84d644SOleg Nesterov * 2820401a8d04STejun Heo * The caller must ensure that the workqueue on which @work was last 28216e84d644SOleg Nesterov * queued can't be destroyed before this function returns. 2822401a8d04STejun Heo * 2823401a8d04STejun Heo * RETURNS: 2824401a8d04STejun Heo * %true if @work was pending, %false otherwise. 28256e84d644SOleg Nesterov */ 2826401a8d04STejun Heo bool cancel_work_sync(struct work_struct *work) 28276e84d644SOleg Nesterov { 28281f1f642eSOleg Nesterov return __cancel_work_timer(work, NULL); 2829b89deed3SOleg Nesterov } 283028e53bddSOleg Nesterov EXPORT_SYMBOL_GPL(cancel_work_sync); 2831b89deed3SOleg Nesterov 28326e84d644SOleg Nesterov /** 2833401a8d04STejun Heo * flush_delayed_work - wait for a dwork to finish executing the last queueing 2834401a8d04STejun Heo * @dwork: the delayed work to flush 28356e84d644SOleg Nesterov * 2836401a8d04STejun Heo * Delayed timer is cancelled and the pending work is queued for 2837401a8d04STejun Heo * immediate execution. Like flush_work(), this function only 2838401a8d04STejun Heo * considers the last queueing instance of @dwork. 28391f1f642eSOleg Nesterov * 2840401a8d04STejun Heo * RETURNS: 2841401a8d04STejun Heo * %true if flush_work() waited for the work to finish execution, 2842401a8d04STejun Heo * %false if it was already idle. 28436e84d644SOleg Nesterov */ 2844401a8d04STejun Heo bool flush_delayed_work(struct delayed_work *dwork) 2845401a8d04STejun Heo { 2846401a8d04STejun Heo if (del_timer_sync(&dwork->timer)) 2847401a8d04STejun Heo __queue_work(raw_smp_processor_id(), 2848401a8d04STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 2849401a8d04STejun Heo return flush_work(&dwork->work); 2850401a8d04STejun Heo } 2851401a8d04STejun Heo EXPORT_SYMBOL(flush_delayed_work); 2852401a8d04STejun Heo 2853401a8d04STejun Heo /** 285409383498STejun Heo * flush_delayed_work_sync - wait for a dwork to finish 285509383498STejun Heo * @dwork: the delayed work to flush 285609383498STejun Heo * 285709383498STejun Heo * Delayed timer is cancelled and the pending work is queued for 285809383498STejun Heo * execution immediately. Other than timer handling, its behavior 285909383498STejun Heo * is identical to flush_work_sync(). 286009383498STejun Heo * 286109383498STejun Heo * RETURNS: 286209383498STejun Heo * %true if flush_work_sync() waited for the work to finish execution, 286309383498STejun Heo * %false if it was already idle. 286409383498STejun Heo */ 286509383498STejun Heo bool flush_delayed_work_sync(struct delayed_work *dwork) 286609383498STejun Heo { 286709383498STejun Heo if (del_timer_sync(&dwork->timer)) 286809383498STejun Heo __queue_work(raw_smp_processor_id(), 286909383498STejun Heo get_work_cwq(&dwork->work)->wq, &dwork->work); 287009383498STejun Heo return flush_work_sync(&dwork->work); 287109383498STejun Heo } 287209383498STejun Heo EXPORT_SYMBOL(flush_delayed_work_sync); 287309383498STejun Heo 287409383498STejun Heo /** 2875401a8d04STejun Heo * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish 2876401a8d04STejun Heo * @dwork: the delayed work cancel 2877401a8d04STejun Heo * 2878401a8d04STejun Heo * This is cancel_work_sync() for delayed works. 2879401a8d04STejun Heo * 2880401a8d04STejun Heo * RETURNS: 2881401a8d04STejun Heo * %true if @dwork was pending, %false otherwise. 2882401a8d04STejun Heo */ 2883401a8d04STejun Heo bool cancel_delayed_work_sync(struct delayed_work *dwork) 28846e84d644SOleg Nesterov { 28851f1f642eSOleg Nesterov return __cancel_work_timer(&dwork->work, &dwork->timer); 28866e84d644SOleg Nesterov } 2887f5a421a4SOleg Nesterov EXPORT_SYMBOL(cancel_delayed_work_sync); 28881da177e4SLinus Torvalds 28890fcb78c2SRolf Eike Beer /** 28900fcb78c2SRolf Eike Beer * schedule_work - put work task in global workqueue 28910fcb78c2SRolf Eike Beer * @work: job to be done 28920fcb78c2SRolf Eike Beer * 28935b0f437dSBart Van Assche * Returns zero if @work was already on the kernel-global workqueue and 28945b0f437dSBart Van Assche * non-zero otherwise. 28955b0f437dSBart Van Assche * 28965b0f437dSBart Van Assche * This puts a job in the kernel-global workqueue if it was not already 28975b0f437dSBart Van Assche * queued and leaves it in the same position on the kernel-global 28985b0f437dSBart Van Assche * workqueue otherwise. 28990fcb78c2SRolf Eike Beer */ 29007ad5b3a5SHarvey Harrison int schedule_work(struct work_struct *work) 29011da177e4SLinus Torvalds { 2902d320c038STejun Heo return queue_work(system_wq, work); 29031da177e4SLinus Torvalds } 2904ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_work); 29051da177e4SLinus Torvalds 2906c1a220e7SZhang Rui /* 2907c1a220e7SZhang Rui * schedule_work_on - put work task on a specific cpu 2908c1a220e7SZhang Rui * @cpu: cpu to put the work task on 2909c1a220e7SZhang Rui * @work: job to be done 2910c1a220e7SZhang Rui * 2911c1a220e7SZhang Rui * This puts a job on a specific cpu 2912c1a220e7SZhang Rui */ 2913c1a220e7SZhang Rui int schedule_work_on(int cpu, struct work_struct *work) 2914c1a220e7SZhang Rui { 2915d320c038STejun Heo return queue_work_on(cpu, system_wq, work); 2916c1a220e7SZhang Rui } 2917c1a220e7SZhang Rui EXPORT_SYMBOL(schedule_work_on); 2918c1a220e7SZhang Rui 29190fcb78c2SRolf Eike Beer /** 29200fcb78c2SRolf Eike Beer * schedule_delayed_work - put work task in global workqueue after delay 292152bad64dSDavid Howells * @dwork: job to be done 292252bad64dSDavid Howells * @delay: number of jiffies to wait or 0 for immediate execution 29230fcb78c2SRolf Eike Beer * 29240fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 29250fcb78c2SRolf Eike Beer * workqueue. 29260fcb78c2SRolf Eike Beer */ 29277ad5b3a5SHarvey Harrison int schedule_delayed_work(struct delayed_work *dwork, 292882f67cd9SIngo Molnar unsigned long delay) 29291da177e4SLinus Torvalds { 2930d320c038STejun Heo return queue_delayed_work(system_wq, dwork, delay); 29311da177e4SLinus Torvalds } 2932ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work); 29331da177e4SLinus Torvalds 29340fcb78c2SRolf Eike Beer /** 29350fcb78c2SRolf Eike Beer * schedule_delayed_work_on - queue work in global workqueue on CPU after delay 29360fcb78c2SRolf Eike Beer * @cpu: cpu to use 293752bad64dSDavid Howells * @dwork: job to be done 29380fcb78c2SRolf Eike Beer * @delay: number of jiffies to wait 29390fcb78c2SRolf Eike Beer * 29400fcb78c2SRolf Eike Beer * After waiting for a given time this puts a job in the kernel-global 29410fcb78c2SRolf Eike Beer * workqueue on the specified CPU. 29420fcb78c2SRolf Eike Beer */ 29431da177e4SLinus Torvalds int schedule_delayed_work_on(int cpu, 294452bad64dSDavid Howells struct delayed_work *dwork, unsigned long delay) 29451da177e4SLinus Torvalds { 2946d320c038STejun Heo return queue_delayed_work_on(cpu, system_wq, dwork, delay); 29471da177e4SLinus Torvalds } 2948ae90dd5dSDave Jones EXPORT_SYMBOL(schedule_delayed_work_on); 29491da177e4SLinus Torvalds 2950b6136773SAndrew Morton /** 295131ddd871STejun Heo * schedule_on_each_cpu - execute a function synchronously on each online CPU 2952b6136773SAndrew Morton * @func: the function to call 2953b6136773SAndrew Morton * 295431ddd871STejun Heo * schedule_on_each_cpu() executes @func on each online CPU using the 295531ddd871STejun Heo * system workqueue and blocks until all CPUs have completed. 2956b6136773SAndrew Morton * schedule_on_each_cpu() is very slow. 295731ddd871STejun Heo * 295831ddd871STejun Heo * RETURNS: 295931ddd871STejun Heo * 0 on success, -errno on failure. 2960b6136773SAndrew Morton */ 296165f27f38SDavid Howells int schedule_on_each_cpu(work_func_t func) 296215316ba8SChristoph Lameter { 296315316ba8SChristoph Lameter int cpu; 296438f51568SNamhyung Kim struct work_struct __percpu *works; 296515316ba8SChristoph Lameter 2966b6136773SAndrew Morton works = alloc_percpu(struct work_struct); 2967b6136773SAndrew Morton if (!works) 296815316ba8SChristoph Lameter return -ENOMEM; 2969b6136773SAndrew Morton 297095402b38SGautham R Shenoy get_online_cpus(); 297193981800STejun Heo 297215316ba8SChristoph Lameter for_each_online_cpu(cpu) { 29739bfb1839SIngo Molnar struct work_struct *work = per_cpu_ptr(works, cpu); 29749bfb1839SIngo Molnar 29759bfb1839SIngo Molnar INIT_WORK(work, func); 29768de6d308SOleg Nesterov schedule_work_on(cpu, work); 297715316ba8SChristoph Lameter } 297893981800STejun Heo 297993981800STejun Heo for_each_online_cpu(cpu) 29808616a89aSOleg Nesterov flush_work(per_cpu_ptr(works, cpu)); 298193981800STejun Heo 298295402b38SGautham R Shenoy put_online_cpus(); 2983b6136773SAndrew Morton free_percpu(works); 298415316ba8SChristoph Lameter return 0; 298515316ba8SChristoph Lameter } 298615316ba8SChristoph Lameter 2987eef6a7d5SAlan Stern /** 2988eef6a7d5SAlan Stern * flush_scheduled_work - ensure that any scheduled work has run to completion. 2989eef6a7d5SAlan Stern * 2990eef6a7d5SAlan Stern * Forces execution of the kernel-global workqueue and blocks until its 2991eef6a7d5SAlan Stern * completion. 2992eef6a7d5SAlan Stern * 2993eef6a7d5SAlan Stern * Think twice before calling this function! It's very easy to get into 2994eef6a7d5SAlan Stern * trouble if you don't take great care. Either of the following situations 2995eef6a7d5SAlan Stern * will lead to deadlock: 2996eef6a7d5SAlan Stern * 2997eef6a7d5SAlan Stern * One of the work items currently on the workqueue needs to acquire 2998eef6a7d5SAlan Stern * a lock held by your code or its caller. 2999eef6a7d5SAlan Stern * 3000eef6a7d5SAlan Stern * Your code is running in the context of a work routine. 3001eef6a7d5SAlan Stern * 3002eef6a7d5SAlan Stern * They will be detected by lockdep when they occur, but the first might not 3003eef6a7d5SAlan Stern * occur very often. It depends on what work items are on the workqueue and 3004eef6a7d5SAlan Stern * what locks they need, which you have no control over. 3005eef6a7d5SAlan Stern * 3006eef6a7d5SAlan Stern * In most situations flushing the entire workqueue is overkill; you merely 3007eef6a7d5SAlan Stern * need to know that a particular work item isn't queued and isn't running. 3008eef6a7d5SAlan Stern * In such cases you should use cancel_delayed_work_sync() or 3009eef6a7d5SAlan Stern * cancel_work_sync() instead. 3010eef6a7d5SAlan Stern */ 30111da177e4SLinus Torvalds void flush_scheduled_work(void) 30121da177e4SLinus Torvalds { 3013d320c038STejun Heo flush_workqueue(system_wq); 30141da177e4SLinus Torvalds } 3015ae90dd5dSDave Jones EXPORT_SYMBOL(flush_scheduled_work); 30161da177e4SLinus Torvalds 30171da177e4SLinus Torvalds /** 30181fa44ecaSJames Bottomley * execute_in_process_context - reliably execute the routine with user context 30191fa44ecaSJames Bottomley * @fn: the function to execute 30201fa44ecaSJames Bottomley * @ew: guaranteed storage for the execute work structure (must 30211fa44ecaSJames Bottomley * be available when the work executes) 30221fa44ecaSJames Bottomley * 30231fa44ecaSJames Bottomley * Executes the function immediately if process context is available, 30241fa44ecaSJames Bottomley * otherwise schedules the function for delayed execution. 30251fa44ecaSJames Bottomley * 30261fa44ecaSJames Bottomley * Returns: 0 - function was executed 30271fa44ecaSJames Bottomley * 1 - function was scheduled for execution 30281fa44ecaSJames Bottomley */ 302965f27f38SDavid Howells int execute_in_process_context(work_func_t fn, struct execute_work *ew) 30301fa44ecaSJames Bottomley { 30311fa44ecaSJames Bottomley if (!in_interrupt()) { 303265f27f38SDavid Howells fn(&ew->work); 30331fa44ecaSJames Bottomley return 0; 30341fa44ecaSJames Bottomley } 30351fa44ecaSJames Bottomley 303665f27f38SDavid Howells INIT_WORK(&ew->work, fn); 30371fa44ecaSJames Bottomley schedule_work(&ew->work); 30381fa44ecaSJames Bottomley 30391fa44ecaSJames Bottomley return 1; 30401fa44ecaSJames Bottomley } 30411fa44ecaSJames Bottomley EXPORT_SYMBOL_GPL(execute_in_process_context); 30421fa44ecaSJames Bottomley 30431da177e4SLinus Torvalds int keventd_up(void) 30441da177e4SLinus Torvalds { 3045d320c038STejun Heo return system_wq != NULL; 30461da177e4SLinus Torvalds } 30471da177e4SLinus Torvalds 3048bdbc5dd7STejun Heo static int alloc_cwqs(struct workqueue_struct *wq) 30491da177e4SLinus Torvalds { 30503af24433SOleg Nesterov /* 30510f900049STejun Heo * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. 30520f900049STejun Heo * Make sure that the alignment isn't lower than that of 30530f900049STejun Heo * unsigned long long. 30543af24433SOleg Nesterov */ 30550f900049STejun Heo const size_t size = sizeof(struct cpu_workqueue_struct); 30560f900049STejun Heo const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, 30570f900049STejun Heo __alignof__(unsigned long long)); 30583af24433SOleg Nesterov 3059e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3060f3421797STejun Heo wq->cpu_wq.pcpu = __alloc_percpu(size, align); 3061931ac77eSTejun Heo else { 30620f900049STejun Heo void *ptr; 3063e1d8aa9fSFrederic Weisbecker 30640f900049STejun Heo /* 3065f3421797STejun Heo * Allocate enough room to align cwq and put an extra 3066f3421797STejun Heo * pointer at the end pointing back to the originally 3067f3421797STejun Heo * allocated pointer which will be used for free. 30680f900049STejun Heo */ 3069bdbc5dd7STejun Heo ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); 3070bdbc5dd7STejun Heo if (ptr) { 3071bdbc5dd7STejun Heo wq->cpu_wq.single = PTR_ALIGN(ptr, align); 3072bdbc5dd7STejun Heo *(void **)(wq->cpu_wq.single + 1) = ptr; 3073bdbc5dd7STejun Heo } 30743af24433SOleg Nesterov } 30753af24433SOleg Nesterov 30760415b00dSTejun Heo /* just in case, make sure it's actually aligned */ 3077bdbc5dd7STejun Heo BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); 3078bdbc5dd7STejun Heo return wq->cpu_wq.v ? 0 : -ENOMEM; 30790f900049STejun Heo } 30800f900049STejun Heo 3081bdbc5dd7STejun Heo static void free_cwqs(struct workqueue_struct *wq) 308206ba38a9SOleg Nesterov { 3083e06ffa1eSLai Jiangshan if (!(wq->flags & WQ_UNBOUND)) 3084bdbc5dd7STejun Heo free_percpu(wq->cpu_wq.pcpu); 3085f3421797STejun Heo else if (wq->cpu_wq.single) { 3086f3421797STejun Heo /* the pointer to free is stored right after the cwq */ 3087f3421797STejun Heo kfree(*(void **)(wq->cpu_wq.single + 1)); 308806ba38a9SOleg Nesterov } 308906ba38a9SOleg Nesterov } 309006ba38a9SOleg Nesterov 3091f3421797STejun Heo static int wq_clamp_max_active(int max_active, unsigned int flags, 3092f3421797STejun Heo const char *name) 3093b71ab8c2STejun Heo { 3094f3421797STejun Heo int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; 3095f3421797STejun Heo 3096f3421797STejun Heo if (max_active < 1 || max_active > lim) 3097b71ab8c2STejun Heo printk(KERN_WARNING "workqueue: max_active %d requested for %s " 3098b71ab8c2STejun Heo "is out of range, clamping between %d and %d\n", 3099f3421797STejun Heo max_active, name, 1, lim); 3100b71ab8c2STejun Heo 3101f3421797STejun Heo return clamp_val(max_active, 1, lim); 3102b71ab8c2STejun Heo } 3103b71ab8c2STejun Heo 3104b196be89STejun Heo struct workqueue_struct *__alloc_workqueue_key(const char *fmt, 310597e37d7bSTejun Heo unsigned int flags, 31061e19ffc6STejun Heo int max_active, 3107eb13ba87SJohannes Berg struct lock_class_key *key, 3108b196be89STejun Heo const char *lock_name, ...) 31093af24433SOleg Nesterov { 3110b196be89STejun Heo va_list args, args1; 31113af24433SOleg Nesterov struct workqueue_struct *wq; 3112c34056a3STejun Heo unsigned int cpu; 3113b196be89STejun Heo size_t namelen; 3114b196be89STejun Heo 3115b196be89STejun Heo /* determine namelen, allocate wq and format name */ 3116b196be89STejun Heo va_start(args, lock_name); 3117b196be89STejun Heo va_copy(args1, args); 3118b196be89STejun Heo namelen = vsnprintf(NULL, 0, fmt, args) + 1; 3119b196be89STejun Heo 3120b196be89STejun Heo wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL); 3121b196be89STejun Heo if (!wq) 3122b196be89STejun Heo goto err; 3123b196be89STejun Heo 3124b196be89STejun Heo vsnprintf(wq->name, namelen, fmt, args1); 3125b196be89STejun Heo va_end(args); 3126b196be89STejun Heo va_end(args1); 31273af24433SOleg Nesterov 3128f3421797STejun Heo /* 31296370a6adSTejun Heo * Workqueues which may be used during memory reclaim should 31306370a6adSTejun Heo * have a rescuer to guarantee forward progress. 31316370a6adSTejun Heo */ 31326370a6adSTejun Heo if (flags & WQ_MEM_RECLAIM) 31336370a6adSTejun Heo flags |= WQ_RESCUER; 31346370a6adSTejun Heo 3135d320c038STejun Heo max_active = max_active ?: WQ_DFL_ACTIVE; 3136b196be89STejun Heo max_active = wq_clamp_max_active(max_active, flags, wq->name); 31373af24433SOleg Nesterov 3138b196be89STejun Heo /* init wq */ 313997e37d7bSTejun Heo wq->flags = flags; 3140a0a1a5fdSTejun Heo wq->saved_max_active = max_active; 314173f53c4aSTejun Heo mutex_init(&wq->flush_mutex); 314273f53c4aSTejun Heo atomic_set(&wq->nr_cwqs_to_flush, 0); 314373f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_queue); 314473f53c4aSTejun Heo INIT_LIST_HEAD(&wq->flusher_overflow); 31453af24433SOleg Nesterov 3146eb13ba87SJohannes Berg lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); 3147cce1a165SOleg Nesterov INIT_LIST_HEAD(&wq->list); 31483af24433SOleg Nesterov 3149bdbc5dd7STejun Heo if (alloc_cwqs(wq) < 0) 3150bdbc5dd7STejun Heo goto err; 3151bdbc5dd7STejun Heo 3152f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 31531537663fSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 31548b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 31553270476aSTejun Heo int pool_idx = (bool)(flags & WQ_HIGHPRI); 31561537663fSTejun Heo 31570f900049STejun Heo BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); 31583270476aSTejun Heo cwq->pool = &gcwq->pools[pool_idx]; 3159c34056a3STejun Heo cwq->wq = wq; 316073f53c4aSTejun Heo cwq->flush_color = -1; 31611e19ffc6STejun Heo cwq->max_active = max_active; 31621e19ffc6STejun Heo INIT_LIST_HEAD(&cwq->delayed_works); 3163e22bee78STejun Heo } 31641537663fSTejun Heo 3165e22bee78STejun Heo if (flags & WQ_RESCUER) { 3166e22bee78STejun Heo struct worker *rescuer; 3167e22bee78STejun Heo 3168f2e005aaSTejun Heo if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) 3169e22bee78STejun Heo goto err; 3170e22bee78STejun Heo 3171e22bee78STejun Heo wq->rescuer = rescuer = alloc_worker(); 3172e22bee78STejun Heo if (!rescuer) 3173e22bee78STejun Heo goto err; 3174e22bee78STejun Heo 3175b196be89STejun Heo rescuer->task = kthread_create(rescuer_thread, wq, "%s", 3176b196be89STejun Heo wq->name); 3177e22bee78STejun Heo if (IS_ERR(rescuer->task)) 3178e22bee78STejun Heo goto err; 3179e22bee78STejun Heo 3180e22bee78STejun Heo rescuer->task->flags |= PF_THREAD_BOUND; 3181e22bee78STejun Heo wake_up_process(rescuer->task); 31823af24433SOleg Nesterov } 31831537663fSTejun Heo 31843af24433SOleg Nesterov /* 3185a0a1a5fdSTejun Heo * workqueue_lock protects global freeze state and workqueues 3186a0a1a5fdSTejun Heo * list. Grab it, set max_active accordingly and add the new 3187a0a1a5fdSTejun Heo * workqueue to workqueues list. 31883af24433SOleg Nesterov */ 31893af24433SOleg Nesterov spin_lock(&workqueue_lock); 3190a0a1a5fdSTejun Heo 319158a69cb4STejun Heo if (workqueue_freezing && wq->flags & WQ_FREEZABLE) 3192f3421797STejun Heo for_each_cwq_cpu(cpu, wq) 3193a0a1a5fdSTejun Heo get_cwq(cpu, wq)->max_active = 0; 3194a0a1a5fdSTejun Heo 31953af24433SOleg Nesterov list_add(&wq->list, &workqueues); 3196a0a1a5fdSTejun Heo 31973af24433SOleg Nesterov spin_unlock(&workqueue_lock); 31983af24433SOleg Nesterov 31993af24433SOleg Nesterov return wq; 32004690c4abSTejun Heo err: 32014690c4abSTejun Heo if (wq) { 3202bdbc5dd7STejun Heo free_cwqs(wq); 3203f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 3204e22bee78STejun Heo kfree(wq->rescuer); 32054690c4abSTejun Heo kfree(wq); 32063af24433SOleg Nesterov } 32074690c4abSTejun Heo return NULL; 32081da177e4SLinus Torvalds } 3209d320c038STejun Heo EXPORT_SYMBOL_GPL(__alloc_workqueue_key); 32101da177e4SLinus Torvalds 32113af24433SOleg Nesterov /** 32123af24433SOleg Nesterov * destroy_workqueue - safely terminate a workqueue 32133af24433SOleg Nesterov * @wq: target workqueue 32143af24433SOleg Nesterov * 32153af24433SOleg Nesterov * Safely destroy a workqueue. All work currently pending will be done first. 32163af24433SOleg Nesterov */ 32173af24433SOleg Nesterov void destroy_workqueue(struct workqueue_struct *wq) 32183af24433SOleg Nesterov { 3219c8e55f36STejun Heo unsigned int cpu; 32203af24433SOleg Nesterov 32219c5a2ba7STejun Heo /* drain it before proceeding with destruction */ 32229c5a2ba7STejun Heo drain_workqueue(wq); 3223c8efcc25STejun Heo 3224a0a1a5fdSTejun Heo /* 3225a0a1a5fdSTejun Heo * wq list is used to freeze wq, remove from list after 3226a0a1a5fdSTejun Heo * flushing is complete in case freeze races us. 3227a0a1a5fdSTejun Heo */ 322895402b38SGautham R Shenoy spin_lock(&workqueue_lock); 32293af24433SOleg Nesterov list_del(&wq->list); 323095402b38SGautham R Shenoy spin_unlock(&workqueue_lock); 32313af24433SOleg Nesterov 3232e22bee78STejun Heo /* sanity check */ 3233f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 323473f53c4aSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 323573f53c4aSTejun Heo int i; 32363af24433SOleg Nesterov 323773f53c4aSTejun Heo for (i = 0; i < WORK_NR_COLORS; i++) 323873f53c4aSTejun Heo BUG_ON(cwq->nr_in_flight[i]); 32391e19ffc6STejun Heo BUG_ON(cwq->nr_active); 32401e19ffc6STejun Heo BUG_ON(!list_empty(&cwq->delayed_works)); 324173f53c4aSTejun Heo } 32421537663fSTejun Heo 3243e22bee78STejun Heo if (wq->flags & WQ_RESCUER) { 3244e22bee78STejun Heo kthread_stop(wq->rescuer->task); 3245f2e005aaSTejun Heo free_mayday_mask(wq->mayday_mask); 32468d9df9f0SXiaotian Feng kfree(wq->rescuer); 3247e22bee78STejun Heo } 3248e22bee78STejun Heo 3249bdbc5dd7STejun Heo free_cwqs(wq); 32503af24433SOleg Nesterov kfree(wq); 32513af24433SOleg Nesterov } 32523af24433SOleg Nesterov EXPORT_SYMBOL_GPL(destroy_workqueue); 32533af24433SOleg Nesterov 3254dcd989cbSTejun Heo /** 3255dcd989cbSTejun Heo * workqueue_set_max_active - adjust max_active of a workqueue 3256dcd989cbSTejun Heo * @wq: target workqueue 3257dcd989cbSTejun Heo * @max_active: new max_active value. 3258dcd989cbSTejun Heo * 3259dcd989cbSTejun Heo * Set max_active of @wq to @max_active. 3260dcd989cbSTejun Heo * 3261dcd989cbSTejun Heo * CONTEXT: 3262dcd989cbSTejun Heo * Don't call from IRQ context. 3263dcd989cbSTejun Heo */ 3264dcd989cbSTejun Heo void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) 3265dcd989cbSTejun Heo { 3266dcd989cbSTejun Heo unsigned int cpu; 3267dcd989cbSTejun Heo 3268f3421797STejun Heo max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); 3269dcd989cbSTejun Heo 3270dcd989cbSTejun Heo spin_lock(&workqueue_lock); 3271dcd989cbSTejun Heo 3272dcd989cbSTejun Heo wq->saved_max_active = max_active; 3273dcd989cbSTejun Heo 3274f3421797STejun Heo for_each_cwq_cpu(cpu, wq) { 3275dcd989cbSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3276dcd989cbSTejun Heo 3277dcd989cbSTejun Heo spin_lock_irq(&gcwq->lock); 3278dcd989cbSTejun Heo 327958a69cb4STejun Heo if (!(wq->flags & WQ_FREEZABLE) || 3280dcd989cbSTejun Heo !(gcwq->flags & GCWQ_FREEZING)) 3281dcd989cbSTejun Heo get_cwq(gcwq->cpu, wq)->max_active = max_active; 3282dcd989cbSTejun Heo 3283dcd989cbSTejun Heo spin_unlock_irq(&gcwq->lock); 3284dcd989cbSTejun Heo } 3285dcd989cbSTejun Heo 3286dcd989cbSTejun Heo spin_unlock(&workqueue_lock); 3287dcd989cbSTejun Heo } 3288dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_set_max_active); 3289dcd989cbSTejun Heo 3290dcd989cbSTejun Heo /** 3291dcd989cbSTejun Heo * workqueue_congested - test whether a workqueue is congested 3292dcd989cbSTejun Heo * @cpu: CPU in question 3293dcd989cbSTejun Heo * @wq: target workqueue 3294dcd989cbSTejun Heo * 3295dcd989cbSTejun Heo * Test whether @wq's cpu workqueue for @cpu is congested. There is 3296dcd989cbSTejun Heo * no synchronization around this function and the test result is 3297dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3298dcd989cbSTejun Heo * 3299dcd989cbSTejun Heo * RETURNS: 3300dcd989cbSTejun Heo * %true if congested, %false otherwise. 3301dcd989cbSTejun Heo */ 3302dcd989cbSTejun Heo bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) 3303dcd989cbSTejun Heo { 3304dcd989cbSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3305dcd989cbSTejun Heo 3306dcd989cbSTejun Heo return !list_empty(&cwq->delayed_works); 3307dcd989cbSTejun Heo } 3308dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(workqueue_congested); 3309dcd989cbSTejun Heo 3310dcd989cbSTejun Heo /** 3311dcd989cbSTejun Heo * work_cpu - return the last known associated cpu for @work 3312dcd989cbSTejun Heo * @work: the work of interest 3313dcd989cbSTejun Heo * 3314dcd989cbSTejun Heo * RETURNS: 3315bdbc5dd7STejun Heo * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. 3316dcd989cbSTejun Heo */ 3317dcd989cbSTejun Heo unsigned int work_cpu(struct work_struct *work) 3318dcd989cbSTejun Heo { 3319dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3320dcd989cbSTejun Heo 3321bdbc5dd7STejun Heo return gcwq ? gcwq->cpu : WORK_CPU_NONE; 3322dcd989cbSTejun Heo } 3323dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_cpu); 3324dcd989cbSTejun Heo 3325dcd989cbSTejun Heo /** 3326dcd989cbSTejun Heo * work_busy - test whether a work is currently pending or running 3327dcd989cbSTejun Heo * @work: the work to be tested 3328dcd989cbSTejun Heo * 3329dcd989cbSTejun Heo * Test whether @work is currently pending or running. There is no 3330dcd989cbSTejun Heo * synchronization around this function and the test result is 3331dcd989cbSTejun Heo * unreliable and only useful as advisory hints or for debugging. 3332dcd989cbSTejun Heo * Especially for reentrant wqs, the pending state might hide the 3333dcd989cbSTejun Heo * running state. 3334dcd989cbSTejun Heo * 3335dcd989cbSTejun Heo * RETURNS: 3336dcd989cbSTejun Heo * OR'd bitmask of WORK_BUSY_* bits. 3337dcd989cbSTejun Heo */ 3338dcd989cbSTejun Heo unsigned int work_busy(struct work_struct *work) 3339dcd989cbSTejun Heo { 3340dcd989cbSTejun Heo struct global_cwq *gcwq = get_work_gcwq(work); 3341dcd989cbSTejun Heo unsigned long flags; 3342dcd989cbSTejun Heo unsigned int ret = 0; 3343dcd989cbSTejun Heo 3344dcd989cbSTejun Heo if (!gcwq) 3345dcd989cbSTejun Heo return false; 3346dcd989cbSTejun Heo 3347dcd989cbSTejun Heo spin_lock_irqsave(&gcwq->lock, flags); 3348dcd989cbSTejun Heo 3349dcd989cbSTejun Heo if (work_pending(work)) 3350dcd989cbSTejun Heo ret |= WORK_BUSY_PENDING; 3351dcd989cbSTejun Heo if (find_worker_executing_work(gcwq, work)) 3352dcd989cbSTejun Heo ret |= WORK_BUSY_RUNNING; 3353dcd989cbSTejun Heo 3354dcd989cbSTejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 3355dcd989cbSTejun Heo 3356dcd989cbSTejun Heo return ret; 3357dcd989cbSTejun Heo } 3358dcd989cbSTejun Heo EXPORT_SYMBOL_GPL(work_busy); 3359dcd989cbSTejun Heo 3360db7bccf4STejun Heo /* 3361db7bccf4STejun Heo * CPU hotplug. 3362db7bccf4STejun Heo * 3363e22bee78STejun Heo * There are two challenges in supporting CPU hotplug. Firstly, there 3364e22bee78STejun Heo * are a lot of assumptions on strong associations among work, cwq and 3365e22bee78STejun Heo * gcwq which make migrating pending and scheduled works very 3366e22bee78STejun Heo * difficult to implement without impacting hot paths. Secondly, 3367e22bee78STejun Heo * gcwqs serve mix of short, long and very long running works making 3368e22bee78STejun Heo * blocked draining impractical. 3369e22bee78STejun Heo * 3370403c821dSTejun Heo * This is solved by allowing a gcwq to be detached from CPU, running it 3371403c821dSTejun Heo * with unbound workers and allowing it to be reattached later if the cpu 3372403c821dSTejun Heo * comes back online. A separate thread is created to govern a gcwq in 3373403c821dSTejun Heo * such state and is called the trustee of the gcwq. 3374db7bccf4STejun Heo * 3375db7bccf4STejun Heo * Trustee states and their descriptions. 3376db7bccf4STejun Heo * 3377db7bccf4STejun Heo * START Command state used on startup. On CPU_DOWN_PREPARE, a 3378db7bccf4STejun Heo * new trustee is started with this state. 3379db7bccf4STejun Heo * 3380db7bccf4STejun Heo * IN_CHARGE Once started, trustee will enter this state after 3381e22bee78STejun Heo * assuming the manager role and making all existing 3382e22bee78STejun Heo * workers rogue. DOWN_PREPARE waits for trustee to 3383e22bee78STejun Heo * enter this state. After reaching IN_CHARGE, trustee 3384e22bee78STejun Heo * tries to execute the pending worklist until it's empty 3385e22bee78STejun Heo * and the state is set to BUTCHER, or the state is set 3386e22bee78STejun Heo * to RELEASE. 3387db7bccf4STejun Heo * 3388db7bccf4STejun Heo * BUTCHER Command state which is set by the cpu callback after 3389db7bccf4STejun Heo * the cpu has went down. Once this state is set trustee 3390db7bccf4STejun Heo * knows that there will be no new works on the worklist 3391db7bccf4STejun Heo * and once the worklist is empty it can proceed to 3392db7bccf4STejun Heo * killing idle workers. 3393db7bccf4STejun Heo * 3394db7bccf4STejun Heo * RELEASE Command state which is set by the cpu callback if the 3395db7bccf4STejun Heo * cpu down has been canceled or it has come online 3396db7bccf4STejun Heo * again. After recognizing this state, trustee stops 3397e22bee78STejun Heo * trying to drain or butcher and clears ROGUE, rebinds 3398e22bee78STejun Heo * all remaining workers back to the cpu and releases 3399e22bee78STejun Heo * manager role. 3400db7bccf4STejun Heo * 3401db7bccf4STejun Heo * DONE Trustee will enter this state after BUTCHER or RELEASE 3402db7bccf4STejun Heo * is complete. 3403db7bccf4STejun Heo * 3404db7bccf4STejun Heo * trustee CPU draining 3405db7bccf4STejun Heo * took over down complete 3406db7bccf4STejun Heo * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE 3407db7bccf4STejun Heo * | | ^ 3408db7bccf4STejun Heo * | CPU is back online v return workers | 3409db7bccf4STejun Heo * ----------------> RELEASE -------------- 3410db7bccf4STejun Heo */ 3411db7bccf4STejun Heo 341260373152STejun Heo /* claim manager positions of all pools */ 341360373152STejun Heo static void gcwq_claim_management(struct global_cwq *gcwq) 341460373152STejun Heo { 341560373152STejun Heo struct worker_pool *pool; 341660373152STejun Heo 341760373152STejun Heo for_each_worker_pool(pool, gcwq) 341860373152STejun Heo mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools); 341960373152STejun Heo } 342060373152STejun Heo 342160373152STejun Heo /* release manager positions */ 342260373152STejun Heo static void gcwq_release_management(struct global_cwq *gcwq) 342360373152STejun Heo { 342460373152STejun Heo struct worker_pool *pool; 342560373152STejun Heo 342660373152STejun Heo for_each_worker_pool(pool, gcwq) 342760373152STejun Heo mutex_unlock(&pool->manager_mutex); 342860373152STejun Heo } 342960373152STejun Heo 3430db7bccf4STejun Heo /** 3431db7bccf4STejun Heo * trustee_wait_event_timeout - timed event wait for trustee 3432db7bccf4STejun Heo * @cond: condition to wait for 3433db7bccf4STejun Heo * @timeout: timeout in jiffies 3434db7bccf4STejun Heo * 3435db7bccf4STejun Heo * wait_event_timeout() for trustee to use. Handles locking and 3436db7bccf4STejun Heo * checks for RELEASE request. 3437db7bccf4STejun Heo * 3438db7bccf4STejun Heo * CONTEXT: 3439db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3440db7bccf4STejun Heo * multiple times. To be used by trustee. 3441db7bccf4STejun Heo * 3442db7bccf4STejun Heo * RETURNS: 3443db7bccf4STejun Heo * Positive indicating left time if @cond is satisfied, 0 if timed 3444db7bccf4STejun Heo * out, -1 if canceled. 3445db7bccf4STejun Heo */ 3446db7bccf4STejun Heo #define trustee_wait_event_timeout(cond, timeout) ({ \ 3447db7bccf4STejun Heo long __ret = (timeout); \ 3448db7bccf4STejun Heo while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ 3449db7bccf4STejun Heo __ret) { \ 3450db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); \ 3451db7bccf4STejun Heo __wait_event_timeout(gcwq->trustee_wait, (cond) || \ 3452db7bccf4STejun Heo (gcwq->trustee_state == TRUSTEE_RELEASE), \ 3453db7bccf4STejun Heo __ret); \ 3454db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); \ 3455db7bccf4STejun Heo } \ 3456db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ 3457db7bccf4STejun Heo }) 3458db7bccf4STejun Heo 3459db7bccf4STejun Heo /** 3460db7bccf4STejun Heo * trustee_wait_event - event wait for trustee 3461db7bccf4STejun Heo * @cond: condition to wait for 3462db7bccf4STejun Heo * 3463db7bccf4STejun Heo * wait_event() for trustee to use. Automatically handles locking and 3464db7bccf4STejun Heo * checks for CANCEL request. 3465db7bccf4STejun Heo * 3466db7bccf4STejun Heo * CONTEXT: 3467db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3468db7bccf4STejun Heo * multiple times. To be used by trustee. 3469db7bccf4STejun Heo * 3470db7bccf4STejun Heo * RETURNS: 3471db7bccf4STejun Heo * 0 if @cond is satisfied, -1 if canceled. 3472db7bccf4STejun Heo */ 3473db7bccf4STejun Heo #define trustee_wait_event(cond) ({ \ 3474db7bccf4STejun Heo long __ret1; \ 3475db7bccf4STejun Heo __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ 3476db7bccf4STejun Heo __ret1 < 0 ? -1 : 0; \ 3477db7bccf4STejun Heo }) 3478db7bccf4STejun Heo 3479db7bccf4STejun Heo static int __cpuinit trustee_thread(void *__gcwq) 3480db7bccf4STejun Heo { 3481db7bccf4STejun Heo struct global_cwq *gcwq = __gcwq; 34824ce62e9eSTejun Heo struct worker_pool *pool; 3483db7bccf4STejun Heo struct worker *worker; 3484e22bee78STejun Heo struct work_struct *work; 3485db7bccf4STejun Heo struct hlist_node *pos; 3486db7bccf4STejun Heo int i; 3487db7bccf4STejun Heo 3488db7bccf4STejun Heo BUG_ON(gcwq->cpu != smp_processor_id()); 3489db7bccf4STejun Heo 349060373152STejun Heo gcwq_claim_management(gcwq); 3491db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3492e22bee78STejun Heo 3493f2d5a0eeSTejun Heo /* 3494f2d5a0eeSTejun Heo * We've claimed all manager positions. Make all workers unbound 3495f2d5a0eeSTejun Heo * and set DISASSOCIATED. Before this, all workers except for the 3496f2d5a0eeSTejun Heo * ones which are still executing works from before the last CPU 3497f2d5a0eeSTejun Heo * down must be on the cpu. After this, they may become diasporas. 3498f2d5a0eeSTejun Heo */ 349960373152STejun Heo for_each_worker_pool(pool, gcwq) 35004ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) 3501403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3502db7bccf4STejun Heo 3503db7bccf4STejun Heo for_each_busy_worker(worker, i, pos, gcwq) 3504403c821dSTejun Heo worker->flags |= WORKER_UNBOUND; 3505db7bccf4STejun Heo 3506f2d5a0eeSTejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 3507f2d5a0eeSTejun Heo 3508db7bccf4STejun Heo /* 3509403c821dSTejun Heo * Call schedule() so that we cross rq->lock and thus can guarantee 3510403c821dSTejun Heo * sched callbacks see the unbound flag. This is necessary as 3511403c821dSTejun Heo * scheduler callbacks may be invoked from other cpus. 3512e22bee78STejun Heo */ 3513e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3514e22bee78STejun Heo schedule(); 3515e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3516e22bee78STejun Heo 3517e22bee78STejun Heo /* 3518cb444766STejun Heo * Sched callbacks are disabled now. Zap nr_running. After 3519cb444766STejun Heo * this, nr_running stays zero and need_more_worker() and 3520cb444766STejun Heo * keep_working() are always true as long as the worklist is 3521cb444766STejun Heo * not empty. 3522e22bee78STejun Heo */ 35234ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35244ce62e9eSTejun Heo atomic_set(get_pool_nr_running(pool), 0); 3525e22bee78STejun Heo 3526e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 35274ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35284ce62e9eSTejun Heo del_timer_sync(&pool->idle_timer); 3529e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3530e22bee78STejun Heo 3531e22bee78STejun Heo /* 3532db7bccf4STejun Heo * We're now in charge. Notify and proceed to drain. We need 3533db7bccf4STejun Heo * to keep the gcwq running during the whole CPU down 3534db7bccf4STejun Heo * procedure as other cpu hotunplug callbacks may need to 3535db7bccf4STejun Heo * flush currently running tasks. 3536db7bccf4STejun Heo */ 3537db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_IN_CHARGE; 3538db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3539db7bccf4STejun Heo 3540db7bccf4STejun Heo /* 3541db7bccf4STejun Heo * The original cpu is in the process of dying and may go away 3542db7bccf4STejun Heo * anytime now. When that happens, we and all workers would 3543e22bee78STejun Heo * be migrated to other cpus. Try draining any left work. We 3544e22bee78STejun Heo * want to get it over with ASAP - spam rescuers, wake up as 3545e22bee78STejun Heo * many idlers as necessary and create new ones till the 3546e22bee78STejun Heo * worklist is empty. Note that if the gcwq is frozen, there 354758a69cb4STejun Heo * may be frozen works in freezable cwqs. Don't declare 3548e22bee78STejun Heo * completion while frozen. 3549db7bccf4STejun Heo */ 35504ce62e9eSTejun Heo while (true) { 35514ce62e9eSTejun Heo bool busy = false; 35524ce62e9eSTejun Heo 35534ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 35544ce62e9eSTejun Heo busy |= pool->nr_workers != pool->nr_idle; 35554ce62e9eSTejun Heo 35564ce62e9eSTejun Heo if (!busy && !(gcwq->flags & GCWQ_FREEZING) && 35574ce62e9eSTejun Heo gcwq->trustee_state != TRUSTEE_IN_CHARGE) 35584ce62e9eSTejun Heo break; 35594ce62e9eSTejun Heo 35604ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 3561e22bee78STejun Heo int nr_works = 0; 3562e22bee78STejun Heo 35634ce62e9eSTejun Heo list_for_each_entry(work, &pool->worklist, entry) { 3564e22bee78STejun Heo send_mayday(work); 3565e22bee78STejun Heo nr_works++; 3566e22bee78STejun Heo } 3567e22bee78STejun Heo 35684ce62e9eSTejun Heo list_for_each_entry(worker, &pool->idle_list, entry) { 3569e22bee78STejun Heo if (!nr_works--) 3570e22bee78STejun Heo break; 3571e22bee78STejun Heo wake_up_process(worker->task); 3572e22bee78STejun Heo } 3573e22bee78STejun Heo 35744ce62e9eSTejun Heo if (need_to_create_worker(pool)) { 3575e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3576bc2ae0f5STejun Heo worker = create_worker(pool); 3577e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3578bc2ae0f5STejun Heo if (worker) 3579e22bee78STejun Heo start_worker(worker); 3580e22bee78STejun Heo } 3581e22bee78STejun Heo } 3582e22bee78STejun Heo 3583db7bccf4STejun Heo /* give a breather */ 3584db7bccf4STejun Heo if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) 3585db7bccf4STejun Heo break; 3586db7bccf4STejun Heo } 3587db7bccf4STejun Heo 358860373152STejun Heo gcwq_release_management(gcwq); 3589e22bee78STejun Heo 3590db7bccf4STejun Heo /* notify completion */ 3591db7bccf4STejun Heo gcwq->trustee = NULL; 3592db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3593db7bccf4STejun Heo wake_up_all(&gcwq->trustee_wait); 3594db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3595db7bccf4STejun Heo return 0; 3596db7bccf4STejun Heo } 3597db7bccf4STejun Heo 3598db7bccf4STejun Heo /** 3599db7bccf4STejun Heo * wait_trustee_state - wait for trustee to enter the specified state 3600db7bccf4STejun Heo * @gcwq: gcwq the trustee of interest belongs to 3601db7bccf4STejun Heo * @state: target state to wait for 3602db7bccf4STejun Heo * 3603db7bccf4STejun Heo * Wait for the trustee to reach @state. DONE is already matched. 3604db7bccf4STejun Heo * 3605db7bccf4STejun Heo * CONTEXT: 3606db7bccf4STejun Heo * spin_lock_irq(gcwq->lock) which may be released and regrabbed 3607db7bccf4STejun Heo * multiple times. To be used by cpu_callback. 3608db7bccf4STejun Heo */ 3609db7bccf4STejun Heo static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) 361006bd6ebfSNamhyung Kim __releases(&gcwq->lock) 361106bd6ebfSNamhyung Kim __acquires(&gcwq->lock) 3612db7bccf4STejun Heo { 3613db7bccf4STejun Heo if (!(gcwq->trustee_state == state || 3614db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE)) { 3615db7bccf4STejun Heo spin_unlock_irq(&gcwq->lock); 3616db7bccf4STejun Heo __wait_event(gcwq->trustee_wait, 3617db7bccf4STejun Heo gcwq->trustee_state == state || 3618db7bccf4STejun Heo gcwq->trustee_state == TRUSTEE_DONE); 3619db7bccf4STejun Heo spin_lock_irq(&gcwq->lock); 3620db7bccf4STejun Heo } 3621db7bccf4STejun Heo } 3622db7bccf4STejun Heo 36239c7b216dSChandra Seetharaman static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, 36241da177e4SLinus Torvalds unsigned long action, 36251da177e4SLinus Torvalds void *hcpu) 36261da177e4SLinus Torvalds { 36273af24433SOleg Nesterov unsigned int cpu = (unsigned long)hcpu; 3628db7bccf4STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3629db7bccf4STejun Heo struct task_struct *new_trustee = NULL; 36304ce62e9eSTejun Heo struct worker_pool *pool; 3631db7bccf4STejun Heo unsigned long flags; 36321da177e4SLinus Torvalds 36338bb78442SRafael J. Wysocki action &= ~CPU_TASKS_FROZEN; 36348bb78442SRafael J. Wysocki 36351da177e4SLinus Torvalds switch (action) { 3636db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3637db7bccf4STejun Heo new_trustee = kthread_create(trustee_thread, gcwq, 3638db7bccf4STejun Heo "workqueue_trustee/%d\n", cpu); 3639db7bccf4STejun Heo if (IS_ERR(new_trustee)) 3640db7bccf4STejun Heo return notifier_from_errno(PTR_ERR(new_trustee)); 3641db7bccf4STejun Heo kthread_bind(new_trustee, cpu); 36423ce63377STejun Heo break; 36433ce63377STejun Heo 36443af24433SOleg Nesterov case CPU_UP_PREPARE: 36454ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 36463ce63377STejun Heo struct worker *worker; 36473ce63377STejun Heo 36483ce63377STejun Heo if (pool->nr_workers) 36493ce63377STejun Heo continue; 36503ce63377STejun Heo 36513ce63377STejun Heo worker = create_worker(pool); 36523ce63377STejun Heo if (!worker) 36533ce63377STejun Heo return NOTIFY_BAD; 36543ce63377STejun Heo 36553ce63377STejun Heo spin_lock_irq(&gcwq->lock); 36563ce63377STejun Heo start_worker(worker); 36573ce63377STejun Heo spin_unlock_irq(&gcwq->lock); 36583af24433SOleg Nesterov } 3659db7bccf4STejun Heo } 36601537663fSTejun Heo 3661db7bccf4STejun Heo /* some are called w/ irq disabled, don't disturb irq status */ 3662db7bccf4STejun Heo spin_lock_irqsave(&gcwq->lock, flags); 36633af24433SOleg Nesterov 36643af24433SOleg Nesterov switch (action) { 3665db7bccf4STejun Heo case CPU_DOWN_PREPARE: 3666db7bccf4STejun Heo /* initialize trustee and tell it to acquire the gcwq */ 3667db7bccf4STejun Heo BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); 3668db7bccf4STejun Heo gcwq->trustee = new_trustee; 3669db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_START; 3670db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3671db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); 36721da177e4SLinus Torvalds break; 36731da177e4SLinus Torvalds 36743da1c84cSOleg Nesterov case CPU_POST_DEAD: 3675db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_BUTCHER; 3676db7bccf4STejun Heo break; 3677db7bccf4STejun Heo 3678db7bccf4STejun Heo case CPU_DOWN_FAILED: 36791da177e4SLinus Torvalds case CPU_ONLINE: 3680db7bccf4STejun Heo if (gcwq->trustee_state != TRUSTEE_DONE) { 3681db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_RELEASE; 3682db7bccf4STejun Heo wake_up_process(gcwq->trustee); 3683db7bccf4STejun Heo wait_trustee_state(gcwq, TRUSTEE_DONE); 3684db7bccf4STejun Heo } 36851da177e4SLinus Torvalds 368625511a47STejun Heo spin_unlock_irq(&gcwq->lock); 368725511a47STejun Heo gcwq_claim_management(gcwq); 368825511a47STejun Heo spin_lock_irq(&gcwq->lock); 368925511a47STejun Heo 3690bc2ae0f5STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 3691bc2ae0f5STejun Heo 369225511a47STejun Heo rebind_workers(gcwq); 369325511a47STejun Heo 369425511a47STejun Heo gcwq_release_management(gcwq); 36951da177e4SLinus Torvalds break; 36961da177e4SLinus Torvalds } 36971da177e4SLinus Torvalds 3698db7bccf4STejun Heo spin_unlock_irqrestore(&gcwq->lock, flags); 369900dfcaf7SOleg Nesterov 37001537663fSTejun Heo return notifier_from_errno(0); 37011da177e4SLinus Torvalds } 37021da177e4SLinus Torvalds 370365758202STejun Heo /* 370465758202STejun Heo * Workqueues should be brought up before normal priority CPU notifiers. 370565758202STejun Heo * This will be registered high priority CPU notifier. 370665758202STejun Heo */ 370765758202STejun Heo static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb, 370865758202STejun Heo unsigned long action, 370965758202STejun Heo void *hcpu) 371065758202STejun Heo { 371165758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 371265758202STejun Heo case CPU_UP_PREPARE: 371365758202STejun Heo case CPU_DOWN_FAILED: 371465758202STejun Heo case CPU_ONLINE: 371565758202STejun Heo return workqueue_cpu_callback(nfb, action, hcpu); 371665758202STejun Heo } 371765758202STejun Heo return NOTIFY_OK; 371865758202STejun Heo } 371965758202STejun Heo 372065758202STejun Heo /* 372165758202STejun Heo * Workqueues should be brought down after normal priority CPU notifiers. 372265758202STejun Heo * This will be registered as low priority CPU notifier. 372365758202STejun Heo */ 372465758202STejun Heo static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb, 372565758202STejun Heo unsigned long action, 372665758202STejun Heo void *hcpu) 372765758202STejun Heo { 372865758202STejun Heo switch (action & ~CPU_TASKS_FROZEN) { 372965758202STejun Heo case CPU_DOWN_PREPARE: 373065758202STejun Heo case CPU_POST_DEAD: 373165758202STejun Heo return workqueue_cpu_callback(nfb, action, hcpu); 373265758202STejun Heo } 373365758202STejun Heo return NOTIFY_OK; 373465758202STejun Heo } 373565758202STejun Heo 37362d3854a3SRusty Russell #ifdef CONFIG_SMP 37378ccad40dSRusty Russell 37382d3854a3SRusty Russell struct work_for_cpu { 37396b44003eSAndrew Morton struct completion completion; 37402d3854a3SRusty Russell long (*fn)(void *); 37412d3854a3SRusty Russell void *arg; 37422d3854a3SRusty Russell long ret; 37432d3854a3SRusty Russell }; 37442d3854a3SRusty Russell 37456b44003eSAndrew Morton static int do_work_for_cpu(void *_wfc) 37462d3854a3SRusty Russell { 37476b44003eSAndrew Morton struct work_for_cpu *wfc = _wfc; 37482d3854a3SRusty Russell wfc->ret = wfc->fn(wfc->arg); 37496b44003eSAndrew Morton complete(&wfc->completion); 37506b44003eSAndrew Morton return 0; 37512d3854a3SRusty Russell } 37522d3854a3SRusty Russell 37532d3854a3SRusty Russell /** 37542d3854a3SRusty Russell * work_on_cpu - run a function in user context on a particular cpu 37552d3854a3SRusty Russell * @cpu: the cpu to run on 37562d3854a3SRusty Russell * @fn: the function to run 37572d3854a3SRusty Russell * @arg: the function arg 37582d3854a3SRusty Russell * 375931ad9081SRusty Russell * This will return the value @fn returns. 376031ad9081SRusty Russell * It is up to the caller to ensure that the cpu doesn't go offline. 37616b44003eSAndrew Morton * The caller must not hold any locks which would prevent @fn from completing. 37622d3854a3SRusty Russell */ 37632d3854a3SRusty Russell long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) 37642d3854a3SRusty Russell { 37656b44003eSAndrew Morton struct task_struct *sub_thread; 37666b44003eSAndrew Morton struct work_for_cpu wfc = { 37676b44003eSAndrew Morton .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), 37686b44003eSAndrew Morton .fn = fn, 37696b44003eSAndrew Morton .arg = arg, 37706b44003eSAndrew Morton }; 37712d3854a3SRusty Russell 37726b44003eSAndrew Morton sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); 37736b44003eSAndrew Morton if (IS_ERR(sub_thread)) 37746b44003eSAndrew Morton return PTR_ERR(sub_thread); 37756b44003eSAndrew Morton kthread_bind(sub_thread, cpu); 37766b44003eSAndrew Morton wake_up_process(sub_thread); 37776b44003eSAndrew Morton wait_for_completion(&wfc.completion); 37782d3854a3SRusty Russell return wfc.ret; 37792d3854a3SRusty Russell } 37802d3854a3SRusty Russell EXPORT_SYMBOL_GPL(work_on_cpu); 37812d3854a3SRusty Russell #endif /* CONFIG_SMP */ 37822d3854a3SRusty Russell 3783a0a1a5fdSTejun Heo #ifdef CONFIG_FREEZER 3784e7577c50SRusty Russell 3785a0a1a5fdSTejun Heo /** 3786a0a1a5fdSTejun Heo * freeze_workqueues_begin - begin freezing workqueues 3787a0a1a5fdSTejun Heo * 378858a69cb4STejun Heo * Start freezing workqueues. After this function returns, all freezable 378958a69cb4STejun Heo * workqueues will queue new works to their frozen_works list instead of 379058a69cb4STejun Heo * gcwq->worklist. 3791a0a1a5fdSTejun Heo * 3792a0a1a5fdSTejun Heo * CONTEXT: 37938b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3794a0a1a5fdSTejun Heo */ 3795a0a1a5fdSTejun Heo void freeze_workqueues_begin(void) 3796a0a1a5fdSTejun Heo { 3797a0a1a5fdSTejun Heo unsigned int cpu; 3798a0a1a5fdSTejun Heo 3799a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3800a0a1a5fdSTejun Heo 3801a0a1a5fdSTejun Heo BUG_ON(workqueue_freezing); 3802a0a1a5fdSTejun Heo workqueue_freezing = true; 3803a0a1a5fdSTejun Heo 3804f3421797STejun Heo for_each_gcwq_cpu(cpu) { 38058b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 3806bdbc5dd7STejun Heo struct workqueue_struct *wq; 38078b03ae3cSTejun Heo 38088b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 38098b03ae3cSTejun Heo 3810db7bccf4STejun Heo BUG_ON(gcwq->flags & GCWQ_FREEZING); 3811db7bccf4STejun Heo gcwq->flags |= GCWQ_FREEZING; 3812db7bccf4STejun Heo 3813a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3814a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3815a0a1a5fdSTejun Heo 381658a69cb4STejun Heo if (cwq && wq->flags & WQ_FREEZABLE) 3817a0a1a5fdSTejun Heo cwq->max_active = 0; 38181da177e4SLinus Torvalds } 38198b03ae3cSTejun Heo 38208b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3821a0a1a5fdSTejun Heo } 3822a0a1a5fdSTejun Heo 3823a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3824a0a1a5fdSTejun Heo } 3825a0a1a5fdSTejun Heo 3826a0a1a5fdSTejun Heo /** 382758a69cb4STejun Heo * freeze_workqueues_busy - are freezable workqueues still busy? 3828a0a1a5fdSTejun Heo * 3829a0a1a5fdSTejun Heo * Check whether freezing is complete. This function must be called 3830a0a1a5fdSTejun Heo * between freeze_workqueues_begin() and thaw_workqueues(). 3831a0a1a5fdSTejun Heo * 3832a0a1a5fdSTejun Heo * CONTEXT: 3833a0a1a5fdSTejun Heo * Grabs and releases workqueue_lock. 3834a0a1a5fdSTejun Heo * 3835a0a1a5fdSTejun Heo * RETURNS: 383658a69cb4STejun Heo * %true if some freezable workqueues are still busy. %false if freezing 383758a69cb4STejun Heo * is complete. 3838a0a1a5fdSTejun Heo */ 3839a0a1a5fdSTejun Heo bool freeze_workqueues_busy(void) 3840a0a1a5fdSTejun Heo { 3841a0a1a5fdSTejun Heo unsigned int cpu; 3842a0a1a5fdSTejun Heo bool busy = false; 3843a0a1a5fdSTejun Heo 3844a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3845a0a1a5fdSTejun Heo 3846a0a1a5fdSTejun Heo BUG_ON(!workqueue_freezing); 3847a0a1a5fdSTejun Heo 3848f3421797STejun Heo for_each_gcwq_cpu(cpu) { 3849bdbc5dd7STejun Heo struct workqueue_struct *wq; 3850a0a1a5fdSTejun Heo /* 3851a0a1a5fdSTejun Heo * nr_active is monotonically decreasing. It's safe 3852a0a1a5fdSTejun Heo * to peek without lock. 3853a0a1a5fdSTejun Heo */ 3854a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3855a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3856a0a1a5fdSTejun Heo 385758a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3858a0a1a5fdSTejun Heo continue; 3859a0a1a5fdSTejun Heo 3860a0a1a5fdSTejun Heo BUG_ON(cwq->nr_active < 0); 3861a0a1a5fdSTejun Heo if (cwq->nr_active) { 3862a0a1a5fdSTejun Heo busy = true; 3863a0a1a5fdSTejun Heo goto out_unlock; 3864a0a1a5fdSTejun Heo } 3865a0a1a5fdSTejun Heo } 3866a0a1a5fdSTejun Heo } 3867a0a1a5fdSTejun Heo out_unlock: 3868a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3869a0a1a5fdSTejun Heo return busy; 3870a0a1a5fdSTejun Heo } 3871a0a1a5fdSTejun Heo 3872a0a1a5fdSTejun Heo /** 3873a0a1a5fdSTejun Heo * thaw_workqueues - thaw workqueues 3874a0a1a5fdSTejun Heo * 3875a0a1a5fdSTejun Heo * Thaw workqueues. Normal queueing is restored and all collected 38767e11629dSTejun Heo * frozen works are transferred to their respective gcwq worklists. 3877a0a1a5fdSTejun Heo * 3878a0a1a5fdSTejun Heo * CONTEXT: 38798b03ae3cSTejun Heo * Grabs and releases workqueue_lock and gcwq->lock's. 3880a0a1a5fdSTejun Heo */ 3881a0a1a5fdSTejun Heo void thaw_workqueues(void) 3882a0a1a5fdSTejun Heo { 3883a0a1a5fdSTejun Heo unsigned int cpu; 3884a0a1a5fdSTejun Heo 3885a0a1a5fdSTejun Heo spin_lock(&workqueue_lock); 3886a0a1a5fdSTejun Heo 3887a0a1a5fdSTejun Heo if (!workqueue_freezing) 3888a0a1a5fdSTejun Heo goto out_unlock; 3889a0a1a5fdSTejun Heo 3890f3421797STejun Heo for_each_gcwq_cpu(cpu) { 38918b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 38924ce62e9eSTejun Heo struct worker_pool *pool; 3893bdbc5dd7STejun Heo struct workqueue_struct *wq; 38948b03ae3cSTejun Heo 38958b03ae3cSTejun Heo spin_lock_irq(&gcwq->lock); 38968b03ae3cSTejun Heo 3897db7bccf4STejun Heo BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); 3898db7bccf4STejun Heo gcwq->flags &= ~GCWQ_FREEZING; 3899db7bccf4STejun Heo 3900a0a1a5fdSTejun Heo list_for_each_entry(wq, &workqueues, list) { 3901a0a1a5fdSTejun Heo struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); 3902a0a1a5fdSTejun Heo 390358a69cb4STejun Heo if (!cwq || !(wq->flags & WQ_FREEZABLE)) 3904a0a1a5fdSTejun Heo continue; 3905a0a1a5fdSTejun Heo 3906a0a1a5fdSTejun Heo /* restore max_active and repopulate worklist */ 3907a0a1a5fdSTejun Heo cwq->max_active = wq->saved_max_active; 3908a0a1a5fdSTejun Heo 3909a0a1a5fdSTejun Heo while (!list_empty(&cwq->delayed_works) && 3910a0a1a5fdSTejun Heo cwq->nr_active < cwq->max_active) 3911a0a1a5fdSTejun Heo cwq_activate_first_delayed(cwq); 3912a0a1a5fdSTejun Heo } 39138b03ae3cSTejun Heo 39144ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) 39154ce62e9eSTejun Heo wake_up_worker(pool); 3916e22bee78STejun Heo 39178b03ae3cSTejun Heo spin_unlock_irq(&gcwq->lock); 3918a0a1a5fdSTejun Heo } 3919a0a1a5fdSTejun Heo 3920a0a1a5fdSTejun Heo workqueue_freezing = false; 3921a0a1a5fdSTejun Heo out_unlock: 3922a0a1a5fdSTejun Heo spin_unlock(&workqueue_lock); 3923a0a1a5fdSTejun Heo } 3924a0a1a5fdSTejun Heo #endif /* CONFIG_FREEZER */ 3925a0a1a5fdSTejun Heo 39266ee0578bSSuresh Siddha static int __init init_workqueues(void) 39271da177e4SLinus Torvalds { 3928c34056a3STejun Heo unsigned int cpu; 3929c8e55f36STejun Heo int i; 3930c34056a3STejun Heo 393165758202STejun Heo cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP); 393265758202STejun Heo cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN); 39338b03ae3cSTejun Heo 39348b03ae3cSTejun Heo /* initialize gcwqs */ 3935f3421797STejun Heo for_each_gcwq_cpu(cpu) { 39368b03ae3cSTejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 39374ce62e9eSTejun Heo struct worker_pool *pool; 39388b03ae3cSTejun Heo 39398b03ae3cSTejun Heo spin_lock_init(&gcwq->lock); 39408b03ae3cSTejun Heo gcwq->cpu = cpu; 3941f3421797STejun Heo gcwq->flags |= GCWQ_DISASSOCIATED; 39428b03ae3cSTejun Heo 3943c8e55f36STejun Heo for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) 3944c8e55f36STejun Heo INIT_HLIST_HEAD(&gcwq->busy_hash[i]); 3945c8e55f36STejun Heo 39464ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 39474ce62e9eSTejun Heo pool->gcwq = gcwq; 39484ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->worklist); 39494ce62e9eSTejun Heo INIT_LIST_HEAD(&pool->idle_list); 3950e22bee78STejun Heo 39514ce62e9eSTejun Heo init_timer_deferrable(&pool->idle_timer); 39524ce62e9eSTejun Heo pool->idle_timer.function = idle_worker_timeout; 39534ce62e9eSTejun Heo pool->idle_timer.data = (unsigned long)pool; 3954e22bee78STejun Heo 39554ce62e9eSTejun Heo setup_timer(&pool->mayday_timer, gcwq_mayday_timeout, 39564ce62e9eSTejun Heo (unsigned long)pool); 39574ce62e9eSTejun Heo 395860373152STejun Heo mutex_init(&pool->manager_mutex); 39594ce62e9eSTejun Heo ida_init(&pool->worker_ida); 39604ce62e9eSTejun Heo } 3961db7bccf4STejun Heo 396225511a47STejun Heo init_waitqueue_head(&gcwq->rebind_hold); 396325511a47STejun Heo 3964db7bccf4STejun Heo gcwq->trustee_state = TRUSTEE_DONE; 3965db7bccf4STejun Heo init_waitqueue_head(&gcwq->trustee_wait); 39668b03ae3cSTejun Heo } 39678b03ae3cSTejun Heo 3968e22bee78STejun Heo /* create the initial worker */ 3969f3421797STejun Heo for_each_online_gcwq_cpu(cpu) { 3970e22bee78STejun Heo struct global_cwq *gcwq = get_gcwq(cpu); 39714ce62e9eSTejun Heo struct worker_pool *pool; 3972e22bee78STejun Heo 3973477a3c33STejun Heo if (cpu != WORK_CPU_UNBOUND) 3974477a3c33STejun Heo gcwq->flags &= ~GCWQ_DISASSOCIATED; 39754ce62e9eSTejun Heo 39764ce62e9eSTejun Heo for_each_worker_pool(pool, gcwq) { 39774ce62e9eSTejun Heo struct worker *worker; 39784ce62e9eSTejun Heo 3979bc2ae0f5STejun Heo worker = create_worker(pool); 3980e22bee78STejun Heo BUG_ON(!worker); 3981e22bee78STejun Heo spin_lock_irq(&gcwq->lock); 3982e22bee78STejun Heo start_worker(worker); 3983e22bee78STejun Heo spin_unlock_irq(&gcwq->lock); 3984e22bee78STejun Heo } 39854ce62e9eSTejun Heo } 3986e22bee78STejun Heo 3987d320c038STejun Heo system_wq = alloc_workqueue("events", 0, 0); 3988d320c038STejun Heo system_long_wq = alloc_workqueue("events_long", 0, 0); 3989d320c038STejun Heo system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); 3990f3421797STejun Heo system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, 3991f3421797STejun Heo WQ_UNBOUND_MAX_ACTIVE); 399224d51addSTejun Heo system_freezable_wq = alloc_workqueue("events_freezable", 399324d51addSTejun Heo WQ_FREEZABLE, 0); 399462d3c543SAlan Stern system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable", 399562d3c543SAlan Stern WQ_NON_REENTRANT | WQ_FREEZABLE, 0); 3996e5cba24eSHitoshi Mitake BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq || 399762d3c543SAlan Stern !system_unbound_wq || !system_freezable_wq || 399862d3c543SAlan Stern !system_nrt_freezable_wq); 39996ee0578bSSuresh Siddha return 0; 40001da177e4SLinus Torvalds } 40016ee0578bSSuresh Siddha early_initcall(init_workqueues); 4002