blob: 04ef830690eceea974d90add3de55b57bc4048f6 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Tejun Heoc54fce62010-09-10 16:51:36 +02002 * kernel/workqueue.c - generic async execution with shared worker pool
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
Tejun Heoc54fce62010-09-10 16:51:36 +02004 * Copyright (C) 2002 Ingo Molnar
Linus Torvalds1da177e2005-04-16 15:20:36 -07005 *
Tejun Heoc54fce62010-09-10 16:51:36 +02006 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
Christoph Lameter89ada672005-10-30 15:01:59 -080011 *
Christoph Lametercde53532008-07-04 09:59:22 -070012 * Made to use alloc_percpu by Christoph Lameter.
Tejun Heoc54fce62010-09-10 16:51:36 +020013 *
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
16 *
17 * This is the generic async execution mechanism. Work items as are
18 * executed in process context. The worker pool is shared and
19 * automatically managed. There is one worker pool for each CPU and
20 * one extra for works which are better served by workers which are
21 * not bound to any specific CPU.
22 *
23 * Please read Documentation/workqueue.txt for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -070024 */
25
26#include <linux/module.h>
27#include <linux/kernel.h>
28#include <linux/sched.h>
29#include <linux/init.h>
30#include <linux/signal.h>
31#include <linux/completion.h>
32#include <linux/workqueue.h>
33#include <linux/slab.h>
34#include <linux/cpu.h>
35#include <linux/notifier.h>
36#include <linux/kthread.h>
James Bottomley1fa44ec2006-02-23 12:43:43 -060037#include <linux/hardirq.h>
Christoph Lameter46934022006-10-11 01:21:26 -070038#include <linux/mempolicy.h>
Rafael J. Wysocki341a5952006-12-06 20:34:49 -080039#include <linux/freezer.h>
Peter Zijlstrad5abe662006-12-06 20:37:26 -080040#include <linux/kallsyms.h>
41#include <linux/debug_locks.h>
Johannes Berg4e6045f2007-10-18 23:39:55 -070042#include <linux/lockdep.h>
Tejun Heoc34056a2010-06-29 10:07:11 +020043#include <linux/idr.h>
Tejun Heoe22bee72010-06-29 10:07:14 +020044
45#include "workqueue_sched.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
Tejun Heoc8e55f32010-06-29 10:07:12 +020047enum {
Tejun Heodb7bccf2010-06-29 10:07:12 +020048 /* global_cwq flags */
Tejun Heoe22bee72010-06-29 10:07:14 +020049 GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
50 GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */
51 GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */
Tejun Heodb7bccf2010-06-29 10:07:12 +020052 GCWQ_FREEZING = 1 << 3, /* freeze in progress */
Tejun Heo649027d2010-06-29 10:07:14 +020053 GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */
Tejun Heodb7bccf2010-06-29 10:07:12 +020054
Tejun Heoc8e55f32010-06-29 10:07:12 +020055 /* worker flags */
56 WORKER_STARTED = 1 << 0, /* started */
57 WORKER_DIE = 1 << 1, /* die die die */
58 WORKER_IDLE = 1 << 2, /* is idle */
Tejun Heoe22bee72010-06-29 10:07:14 +020059 WORKER_PREP = 1 << 3, /* preparing to run works */
Tejun Heodb7bccf2010-06-29 10:07:12 +020060 WORKER_ROGUE = 1 << 4, /* not bound to any cpu */
Tejun Heoe22bee72010-06-29 10:07:14 +020061 WORKER_REBIND = 1 << 5, /* mom is home, come back */
Tejun Heofb0e7be2010-06-29 10:07:15 +020062 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
Tejun Heof3421792010-07-02 10:03:51 +020063 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
Tejun Heoe22bee72010-06-29 10:07:14 +020064
Tejun Heofb0e7be2010-06-29 10:07:15 +020065 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND |
Tejun Heof3421792010-07-02 10:03:51 +020066 WORKER_CPU_INTENSIVE | WORKER_UNBOUND,
Tejun Heodb7bccf2010-06-29 10:07:12 +020067
68 /* gcwq->trustee_state */
69 TRUSTEE_START = 0, /* start */
70 TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */
71 TRUSTEE_BUTCHER = 2, /* butcher workers */
72 TRUSTEE_RELEASE = 3, /* release workers */
73 TRUSTEE_DONE = 4, /* trustee is done */
Tejun Heoc8e55f32010-06-29 10:07:12 +020074
75 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
76 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
77 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
Tejun Heodb7bccf2010-06-29 10:07:12 +020078
Tejun Heoe22bee72010-06-29 10:07:14 +020079 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
80 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
81
Tejun Heo3233cdb2011-02-16 18:10:19 +010082 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
83 /* call for help after 10ms
84 (min two ticks) */
Tejun Heoe22bee72010-06-29 10:07:14 +020085 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
86 CREATE_COOLDOWN = HZ, /* time to breath after fail */
Tejun Heodb7bccf2010-06-29 10:07:12 +020087 TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */
Tejun Heoe22bee72010-06-29 10:07:14 +020088
89 /*
90 * Rescue workers are used only on emergencies and shared by
91 * all cpus. Give -20.
92 */
93 RESCUER_NICE_LEVEL = -20,
Tejun Heoc8e55f32010-06-29 10:07:12 +020094};
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
96/*
Tejun Heo4690c4a2010-06-29 10:07:10 +020097 * Structure fields follow one of the following exclusion rules.
98 *
Tejun Heoe41e7042010-08-24 14:22:47 +020099 * I: Modifiable by initialization/destruction paths and read-only for
100 * everyone else.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200101 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200102 * P: Preemption protected. Disabling preemption is enough and should
103 * only be modified and accessed from the local cpu.
104 *
Tejun Heo8b03ae32010-06-29 10:07:12 +0200105 * L: gcwq->lock protected. Access with gcwq->lock held.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200106 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200107 * X: During normal operation, modification requires gcwq->lock and
108 * should be done only from local cpu. Either disabling preemption
109 * on local cpu or grabbing gcwq->lock is enough for read access.
Tejun Heof3421792010-07-02 10:03:51 +0200110 * If GCWQ_DISASSOCIATED is set, it's identical to L.
Tejun Heoe22bee72010-06-29 10:07:14 +0200111 *
Tejun Heo73f53c42010-06-29 10:07:11 +0200112 * F: wq->flush_mutex protected.
113 *
Tejun Heo4690c4a2010-06-29 10:07:10 +0200114 * W: workqueue_lock protected.
115 */
116
Tejun Heo8b03ae32010-06-29 10:07:12 +0200117struct global_cwq;
Tejun Heoc34056a2010-06-29 10:07:11 +0200118
Tejun Heoe22bee72010-06-29 10:07:14 +0200119/*
120 * The poor guys doing the actual heavy lifting. All on-duty workers
121 * are either serving the manager role, on idle list or on busy hash.
122 */
Tejun Heoc34056a2010-06-29 10:07:11 +0200123struct worker {
Tejun Heoc8e55f32010-06-29 10:07:12 +0200124 /* on idle list while idle, on busy hash table while busy */
125 union {
126 struct list_head entry; /* L: while idle */
127 struct hlist_node hentry; /* L: while busy */
128 };
129
Tejun Heoc34056a2010-06-29 10:07:11 +0200130 struct work_struct *current_work; /* L: work being processed */
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200131 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
Tejun Heoaffee4b2010-06-29 10:07:12 +0200132 struct list_head scheduled; /* L: scheduled works */
Tejun Heoc34056a2010-06-29 10:07:11 +0200133 struct task_struct *task; /* I: worker task */
Tejun Heo8b03ae32010-06-29 10:07:12 +0200134 struct global_cwq *gcwq; /* I: the associated gcwq */
Tejun Heoe22bee72010-06-29 10:07:14 +0200135 /* 64 bytes boundary on 64bit, 32 on 32bit */
136 unsigned long last_active; /* L: last active timestamp */
137 unsigned int flags; /* X: flags */
Tejun Heoc34056a2010-06-29 10:07:11 +0200138 int id; /* I: worker id */
Tejun Heoe22bee72010-06-29 10:07:14 +0200139 struct work_struct rebind_work; /* L: rebind worker to cpu */
Tejun Heoc34056a2010-06-29 10:07:11 +0200140};
141
Tejun Heo4690c4a2010-06-29 10:07:10 +0200142/*
Tejun Heoe22bee72010-06-29 10:07:14 +0200143 * Global per-cpu workqueue. There's one and only one for each cpu
144 * and all works are queued and processed here regardless of their
145 * target workqueues.
Tejun Heo8b03ae32010-06-29 10:07:12 +0200146 */
147struct global_cwq {
148 spinlock_t lock; /* the gcwq lock */
Tejun Heo7e116292010-06-29 10:07:13 +0200149 struct list_head worklist; /* L: list of pending works */
Tejun Heo8b03ae32010-06-29 10:07:12 +0200150 unsigned int cpu; /* I: the associated cpu */
Tejun Heodb7bccf2010-06-29 10:07:12 +0200151 unsigned int flags; /* L: GCWQ_* flags */
Tejun Heoc8e55f32010-06-29 10:07:12 +0200152
153 int nr_workers; /* L: total number of workers */
154 int nr_idle; /* L: currently idle ones */
155
156 /* workers are chained either in the idle_list or busy_hash */
Tejun Heoe22bee72010-06-29 10:07:14 +0200157 struct list_head idle_list; /* X: list of idle workers */
Tejun Heoc8e55f32010-06-29 10:07:12 +0200158 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
159 /* L: hash of busy workers */
160
Tejun Heoe22bee72010-06-29 10:07:14 +0200161 struct timer_list idle_timer; /* L: worker idle timeout */
162 struct timer_list mayday_timer; /* L: SOS timer for dworkers */
163
Tejun Heo8b03ae32010-06-29 10:07:12 +0200164 struct ida worker_ida; /* L: for worker IDs */
Tejun Heodb7bccf2010-06-29 10:07:12 +0200165
166 struct task_struct *trustee; /* L: for gcwq shutdown */
167 unsigned int trustee_state; /* L: trustee state */
168 wait_queue_head_t trustee_wait; /* trustee wait */
Tejun Heoe22bee72010-06-29 10:07:14 +0200169 struct worker *first_idle; /* L: first idle worker */
Tejun Heo8b03ae32010-06-29 10:07:12 +0200170} ____cacheline_aligned_in_smp;
171
172/*
Tejun Heo502ca9d2010-06-29 10:07:13 +0200173 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
Tejun Heo0f900042010-06-29 10:07:11 +0200174 * work_struct->data are used for flags and thus cwqs need to be
175 * aligned at two's power of the number of flag bits.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 */
177struct cpu_workqueue_struct {
Tejun Heo8b03ae32010-06-29 10:07:12 +0200178 struct global_cwq *gcwq; /* I: the associated gcwq */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200179 struct workqueue_struct *wq; /* I: the owning workqueue */
Tejun Heo73f53c42010-06-29 10:07:11 +0200180 int work_color; /* L: current color */
181 int flush_color; /* L: flushing color */
182 int nr_in_flight[WORK_NR_COLORS];
183 /* L: nr of in_flight works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200184 int nr_active; /* L: nr of active works */
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200185 int max_active; /* L: max active works */
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200186 struct list_head delayed_works; /* L: delayed works */
Tejun Heo0f900042010-06-29 10:07:11 +0200187};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189/*
Tejun Heo73f53c42010-06-29 10:07:11 +0200190 * Structure used to wait for workqueue flush.
191 */
192struct wq_flusher {
193 struct list_head list; /* F: list of flushers */
194 int flush_color; /* F: flush color waiting for */
195 struct completion done; /* flush completion */
196};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197
Tejun Heo73f53c42010-06-29 10:07:11 +0200198/*
Tejun Heof2e005a2010-07-20 15:59:09 +0200199 * All cpumasks are assumed to be always set on UP and thus can't be
200 * used to determine whether there's something to be done.
201 */
202#ifdef CONFIG_SMP
203typedef cpumask_var_t mayday_mask_t;
204#define mayday_test_and_set_cpu(cpu, mask) \
205 cpumask_test_and_set_cpu((cpu), (mask))
206#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
207#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
Tejun Heo9c375472010-08-31 11:18:34 +0200208#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
Tejun Heof2e005a2010-07-20 15:59:09 +0200209#define free_mayday_mask(mask) free_cpumask_var((mask))
210#else
211typedef unsigned long mayday_mask_t;
212#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
213#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
214#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
215#define alloc_mayday_mask(maskp, gfp) true
216#define free_mayday_mask(mask) do { } while (0)
217#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218
219/*
220 * The externally visible workqueue abstraction is an array of
221 * per-CPU workqueues:
222 */
223struct workqueue_struct {
Tejun Heo97e37d72010-06-29 10:07:10 +0200224 unsigned int flags; /* I: WQ_* flags */
Tejun Heobdbc5dd2010-07-02 10:03:51 +0200225 union {
226 struct cpu_workqueue_struct __percpu *pcpu;
227 struct cpu_workqueue_struct *single;
228 unsigned long v;
229 } cpu_wq; /* I: cwq's */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200230 struct list_head list; /* W: list of all workqueues */
Tejun Heo73f53c42010-06-29 10:07:11 +0200231
232 struct mutex flush_mutex; /* protects wq flushing */
233 int work_color; /* F: current work color */
234 int flush_color; /* F: current flush color */
235 atomic_t nr_cwqs_to_flush; /* flush in progress */
236 struct wq_flusher *first_flusher; /* F: first flusher */
237 struct list_head flusher_queue; /* F: flush waiters */
238 struct list_head flusher_overflow; /* F: flush overflow list */
239
Tejun Heof2e005a2010-07-20 15:59:09 +0200240 mayday_mask_t mayday_mask; /* cpus requesting rescue */
Tejun Heoe22bee72010-06-29 10:07:14 +0200241 struct worker *rescuer; /* I: rescue worker */
242
Tejun Heodcd989c2010-06-29 10:07:14 +0200243 int saved_max_active; /* W: saved cwq max_active */
Tejun Heo4690c4a2010-06-29 10:07:10 +0200244 const char *name; /* I: workqueue name */
Johannes Berg4e6045f2007-10-18 23:39:55 -0700245#ifdef CONFIG_LOCKDEP
Tejun Heo4690c4a2010-06-29 10:07:10 +0200246 struct lockdep_map lockdep_map;
Johannes Berg4e6045f2007-10-18 23:39:55 -0700247#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248};
249
Tejun Heod320c032010-06-29 10:07:14 +0200250struct workqueue_struct *system_wq __read_mostly;
251struct workqueue_struct *system_long_wq __read_mostly;
252struct workqueue_struct *system_nrt_wq __read_mostly;
Tejun Heof3421792010-07-02 10:03:51 +0200253struct workqueue_struct *system_unbound_wq __read_mostly;
Tejun Heo24d51ad2011-02-21 09:52:50 +0100254struct workqueue_struct *system_freezable_wq __read_mostly;
Tejun Heod320c032010-06-29 10:07:14 +0200255EXPORT_SYMBOL_GPL(system_wq);
256EXPORT_SYMBOL_GPL(system_long_wq);
257EXPORT_SYMBOL_GPL(system_nrt_wq);
Tejun Heof3421792010-07-02 10:03:51 +0200258EXPORT_SYMBOL_GPL(system_unbound_wq);
Tejun Heo24d51ad2011-02-21 09:52:50 +0100259EXPORT_SYMBOL_GPL(system_freezable_wq);
Tejun Heod320c032010-06-29 10:07:14 +0200260
Tejun Heo97bd2342010-10-05 10:41:14 +0200261#define CREATE_TRACE_POINTS
262#include <trace/events/workqueue.h>
263
Tejun Heodb7bccf2010-06-29 10:07:12 +0200264#define for_each_busy_worker(worker, i, pos, gcwq) \
265 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
266 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
267
Tejun Heof3421792010-07-02 10:03:51 +0200268static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
269 unsigned int sw)
270{
271 if (cpu < nr_cpu_ids) {
272 if (sw & 1) {
273 cpu = cpumask_next(cpu, mask);
274 if (cpu < nr_cpu_ids)
275 return cpu;
276 }
277 if (sw & 2)
278 return WORK_CPU_UNBOUND;
279 }
280 return WORK_CPU_NONE;
281}
282
283static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
284 struct workqueue_struct *wq)
285{
286 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
287}
288
Tejun Heo09884952010-08-01 11:50:12 +0200289/*
290 * CPU iterators
291 *
292 * An extra gcwq is defined for an invalid cpu number
293 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
294 * specific CPU. The following iterators are similar to
295 * for_each_*_cpu() iterators but also considers the unbound gcwq.
296 *
297 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
298 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
299 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
300 * WORK_CPU_UNBOUND for unbound workqueues
301 */
Tejun Heof3421792010-07-02 10:03:51 +0200302#define for_each_gcwq_cpu(cpu) \
303 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
304 (cpu) < WORK_CPU_NONE; \
305 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
306
307#define for_each_online_gcwq_cpu(cpu) \
308 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
309 (cpu) < WORK_CPU_NONE; \
310 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
311
312#define for_each_cwq_cpu(cpu, wq) \
313 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
314 (cpu) < WORK_CPU_NONE; \
315 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
316
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900317#ifdef CONFIG_DEBUG_OBJECTS_WORK
318
319static struct debug_obj_descr work_debug_descr;
320
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100321static void *work_debug_hint(void *addr)
322{
323 return ((struct work_struct *) addr)->func;
324}
325
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900326/*
327 * fixup_init is called when:
328 * - an active object is initialized
329 */
330static int work_fixup_init(void *addr, enum debug_obj_state state)
331{
332 struct work_struct *work = addr;
333
334 switch (state) {
335 case ODEBUG_STATE_ACTIVE:
336 cancel_work_sync(work);
337 debug_object_init(work, &work_debug_descr);
338 return 1;
339 default:
340 return 0;
341 }
342}
343
344/*
345 * fixup_activate is called when:
346 * - an active object is activated
347 * - an unknown object is activated (might be a statically initialized object)
348 */
349static int work_fixup_activate(void *addr, enum debug_obj_state state)
350{
351 struct work_struct *work = addr;
352
353 switch (state) {
354
355 case ODEBUG_STATE_NOTAVAILABLE:
356 /*
357 * This is not really a fixup. The work struct was
358 * statically initialized. We just make sure that it
359 * is tracked in the object tracker.
360 */
Tejun Heo22df02b2010-06-29 10:07:10 +0200361 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900362 debug_object_init(work, &work_debug_descr);
363 debug_object_activate(work, &work_debug_descr);
364 return 0;
365 }
366 WARN_ON_ONCE(1);
367 return 0;
368
369 case ODEBUG_STATE_ACTIVE:
370 WARN_ON(1);
371
372 default:
373 return 0;
374 }
375}
376
377/*
378 * fixup_free is called when:
379 * - an active object is freed
380 */
381static int work_fixup_free(void *addr, enum debug_obj_state state)
382{
383 struct work_struct *work = addr;
384
385 switch (state) {
386 case ODEBUG_STATE_ACTIVE:
387 cancel_work_sync(work);
388 debug_object_free(work, &work_debug_descr);
389 return 1;
390 default:
391 return 0;
392 }
393}
394
395static struct debug_obj_descr work_debug_descr = {
396 .name = "work_struct",
Stanislaw Gruszka99777282011-03-07 09:58:33 +0100397 .debug_hint = work_debug_hint,
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900398 .fixup_init = work_fixup_init,
399 .fixup_activate = work_fixup_activate,
400 .fixup_free = work_fixup_free,
401};
402
403static inline void debug_work_activate(struct work_struct *work)
404{
405 debug_object_activate(work, &work_debug_descr);
406}
407
408static inline void debug_work_deactivate(struct work_struct *work)
409{
410 debug_object_deactivate(work, &work_debug_descr);
411}
412
413void __init_work(struct work_struct *work, int onstack)
414{
415 if (onstack)
416 debug_object_init_on_stack(work, &work_debug_descr);
417 else
418 debug_object_init(work, &work_debug_descr);
419}
420EXPORT_SYMBOL_GPL(__init_work);
421
422void destroy_work_on_stack(struct work_struct *work)
423{
424 debug_object_free(work, &work_debug_descr);
425}
426EXPORT_SYMBOL_GPL(destroy_work_on_stack);
427
428#else
429static inline void debug_work_activate(struct work_struct *work) { }
430static inline void debug_work_deactivate(struct work_struct *work) { }
431#endif
432
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100433/* Serializes the accesses to the list of workqueues. */
434static DEFINE_SPINLOCK(workqueue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435static LIST_HEAD(workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +0200436static bool workqueue_freezing; /* W: have wqs started freezing? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437
Oleg Nesterov14441962007-05-23 13:57:57 -0700438/*
Tejun Heoe22bee72010-06-29 10:07:14 +0200439 * The almighty global cpu workqueues. nr_running is the only field
440 * which is expected to be used frequently by other cpus via
441 * try_to_wake_up(). Put it in a separate cacheline.
Oleg Nesterov14441962007-05-23 13:57:57 -0700442 */
Tejun Heo8b03ae32010-06-29 10:07:12 +0200443static DEFINE_PER_CPU(struct global_cwq, global_cwq);
Tejun Heoe22bee72010-06-29 10:07:14 +0200444static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running);
Nathan Lynchf756d5e2006-01-08 01:05:12 -0800445
Tejun Heof3421792010-07-02 10:03:51 +0200446/*
447 * Global cpu workqueue and nr_running counter for unbound gcwq. The
448 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
449 * workers have WORKER_UNBOUND set.
450 */
451static struct global_cwq unbound_global_cwq;
452static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */
453
Tejun Heoc34056a2010-06-29 10:07:11 +0200454static int worker_thread(void *__worker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455
Tejun Heo8b03ae32010-06-29 10:07:12 +0200456static struct global_cwq *get_gcwq(unsigned int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457{
Tejun Heof3421792010-07-02 10:03:51 +0200458 if (cpu != WORK_CPU_UNBOUND)
459 return &per_cpu(global_cwq, cpu);
460 else
461 return &unbound_global_cwq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462}
463
Tejun Heoe22bee72010-06-29 10:07:14 +0200464static atomic_t *get_gcwq_nr_running(unsigned int cpu)
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -0700465{
Tejun Heof3421792010-07-02 10:03:51 +0200466 if (cpu != WORK_CPU_UNBOUND)
467 return &per_cpu(gcwq_nr_running, cpu);
468 else
469 return &unbound_gcwq_nr_running;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -0700470}
471
Tejun Heo4690c4a2010-06-29 10:07:10 +0200472static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
473 struct workqueue_struct *wq)
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700474{
Tejun Heof3421792010-07-02 10:03:51 +0200475 if (!(wq->flags & WQ_UNBOUND)) {
476 if (likely(cpu < nr_cpu_ids)) {
477#ifdef CONFIG_SMP
478 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
Tejun Heobdbc5dd2010-07-02 10:03:51 +0200479#else
Tejun Heof3421792010-07-02 10:03:51 +0200480 return wq->cpu_wq.single;
Tejun Heobdbc5dd2010-07-02 10:03:51 +0200481#endif
Tejun Heof3421792010-07-02 10:03:51 +0200482 }
483 } else if (likely(cpu == WORK_CPU_UNBOUND))
484 return wq->cpu_wq.single;
485 return NULL;
Oleg Nesterova848e3b2007-05-09 02:34:17 -0700486}
487
Tejun Heo73f53c42010-06-29 10:07:11 +0200488static unsigned int work_color_to_flags(int color)
489{
490 return color << WORK_STRUCT_COLOR_SHIFT;
491}
492
493static int get_work_color(struct work_struct *work)
494{
495 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
496 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
497}
498
499static int work_next_color(int color)
500{
501 return (color + 1) % WORK_NR_COLORS;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502}
503
David Howells4594bf12006-12-07 11:33:26 +0000504/*
Tejun Heoe1201532010-07-22 14:14:25 +0200505 * A work's data points to the cwq with WORK_STRUCT_CWQ set while the
506 * work is on queue. Once execution starts, WORK_STRUCT_CWQ is
507 * cleared and the work data contains the cpu number it was last on.
Tejun Heo7a22ad72010-06-29 10:07:13 +0200508 *
509 * set_work_{cwq|cpu}() and clear_work_data() can be used to set the
510 * cwq, cpu or clear work->data. These functions should only be
511 * called while the work is owned - ie. while the PENDING bit is set.
512 *
513 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
514 * corresponding to a work. gcwq is available once the work has been
515 * queued anywhere after initialization. cwq is available only from
516 * queueing until execution starts.
David Howells4594bf12006-12-07 11:33:26 +0000517 */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200518static inline void set_work_data(struct work_struct *work, unsigned long data,
519 unsigned long flags)
David Howells365970a2006-11-22 14:54:49 +0000520{
David Howells4594bf12006-12-07 11:33:26 +0000521 BUG_ON(!work_pending(work));
Tejun Heo7a22ad72010-06-29 10:07:13 +0200522 atomic_long_set(&work->data, data | flags | work_static(work));
David Howells365970a2006-11-22 14:54:49 +0000523}
David Howells365970a2006-11-22 14:54:49 +0000524
Tejun Heo7a22ad72010-06-29 10:07:13 +0200525static void set_work_cwq(struct work_struct *work,
526 struct cpu_workqueue_struct *cwq,
527 unsigned long extra_flags)
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200528{
Tejun Heo7a22ad72010-06-29 10:07:13 +0200529 set_work_data(work, (unsigned long)cwq,
Tejun Heoe1201532010-07-22 14:14:25 +0200530 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
Oleg Nesterov4d707b92010-04-23 17:40:40 +0200531}
532
Tejun Heo7a22ad72010-06-29 10:07:13 +0200533static void set_work_cpu(struct work_struct *work, unsigned int cpu)
David Howells365970a2006-11-22 14:54:49 +0000534{
Tejun Heo7a22ad72010-06-29 10:07:13 +0200535 set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING);
536}
537
538static void clear_work_data(struct work_struct *work)
539{
540 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
541}
542
Tejun Heo7a22ad72010-06-29 10:07:13 +0200543static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
544{
Tejun Heoe1201532010-07-22 14:14:25 +0200545 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200546
Tejun Heoe1201532010-07-22 14:14:25 +0200547 if (data & WORK_STRUCT_CWQ)
548 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
549 else
550 return NULL;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200551}
552
553static struct global_cwq *get_work_gcwq(struct work_struct *work)
554{
Tejun Heoe1201532010-07-22 14:14:25 +0200555 unsigned long data = atomic_long_read(&work->data);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200556 unsigned int cpu;
557
Tejun Heoe1201532010-07-22 14:14:25 +0200558 if (data & WORK_STRUCT_CWQ)
559 return ((struct cpu_workqueue_struct *)
560 (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq;
Tejun Heo7a22ad72010-06-29 10:07:13 +0200561
562 cpu = data >> WORK_STRUCT_FLAG_BITS;
Tejun Heobdbc5dd2010-07-02 10:03:51 +0200563 if (cpu == WORK_CPU_NONE)
Tejun Heo7a22ad72010-06-29 10:07:13 +0200564 return NULL;
565
Tejun Heof3421792010-07-02 10:03:51 +0200566 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
Tejun Heo7a22ad72010-06-29 10:07:13 +0200567 return get_gcwq(cpu);
David Howells365970a2006-11-22 14:54:49 +0000568}
569
570/*
Tejun Heoe22bee72010-06-29 10:07:14 +0200571 * Policy functions. These define the policies on how the global
572 * worker pool is managed. Unless noted otherwise, these functions
573 * assume that they're being called with gcwq->lock held.
David Howells365970a2006-11-22 14:54:49 +0000574 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200575
Tejun Heo649027d2010-06-29 10:07:14 +0200576static bool __need_more_worker(struct global_cwq *gcwq)
David Howells365970a2006-11-22 14:54:49 +0000577{
Tejun Heo649027d2010-06-29 10:07:14 +0200578 return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) ||
579 gcwq->flags & GCWQ_HIGHPRI_PENDING;
David Howells365970a2006-11-22 14:54:49 +0000580}
581
Tejun Heoe22bee72010-06-29 10:07:14 +0200582/*
583 * Need to wake up a worker? Called from anything but currently
584 * running workers.
585 */
586static bool need_more_worker(struct global_cwq *gcwq)
David Howells365970a2006-11-22 14:54:49 +0000587{
Tejun Heo649027d2010-06-29 10:07:14 +0200588 return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq);
David Howells365970a2006-11-22 14:54:49 +0000589}
590
Tejun Heoe22bee72010-06-29 10:07:14 +0200591/* Can I start working? Called from busy but !running workers. */
592static bool may_start_working(struct global_cwq *gcwq)
593{
594 return gcwq->nr_idle;
595}
596
597/* Do I need to keep working? Called from currently running workers. */
598static bool keep_working(struct global_cwq *gcwq)
599{
600 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
601
Tejun Heo30310042010-10-11 11:51:57 +0200602 return !list_empty(&gcwq->worklist) &&
603 (atomic_read(nr_running) <= 1 ||
604 gcwq->flags & GCWQ_HIGHPRI_PENDING);
Tejun Heoe22bee72010-06-29 10:07:14 +0200605}
606
607/* Do we need a new worker? Called from manager. */
608static bool need_to_create_worker(struct global_cwq *gcwq)
609{
610 return need_more_worker(gcwq) && !may_start_working(gcwq);
611}
612
613/* Do I need to be the manager? */
614static bool need_to_manage_workers(struct global_cwq *gcwq)
615{
616 return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS;
617}
618
619/* Do we have too many workers and should some go away? */
620static bool too_many_workers(struct global_cwq *gcwq)
621{
622 bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS;
623 int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */
624 int nr_busy = gcwq->nr_workers - nr_idle;
625
626 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
627}
628
629/*
630 * Wake up functions.
631 */
632
Tejun Heo7e116292010-06-29 10:07:13 +0200633/* Return the first worker. Safe with preemption disabled */
634static struct worker *first_worker(struct global_cwq *gcwq)
635{
636 if (unlikely(list_empty(&gcwq->idle_list)))
637 return NULL;
638
639 return list_first_entry(&gcwq->idle_list, struct worker, entry);
640}
641
642/**
643 * wake_up_worker - wake up an idle worker
644 * @gcwq: gcwq to wake worker for
645 *
646 * Wake up the first idle worker of @gcwq.
647 *
648 * CONTEXT:
649 * spin_lock_irq(gcwq->lock).
650 */
651static void wake_up_worker(struct global_cwq *gcwq)
652{
653 struct worker *worker = first_worker(gcwq);
654
655 if (likely(worker))
656 wake_up_process(worker->task);
657}
658
Tejun Heo4690c4a2010-06-29 10:07:10 +0200659/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200660 * wq_worker_waking_up - a worker is waking up
661 * @task: task waking up
662 * @cpu: CPU @task is waking up to
663 *
664 * This function is called during try_to_wake_up() when a worker is
665 * being awoken.
666 *
667 * CONTEXT:
668 * spin_lock_irq(rq->lock)
669 */
670void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
671{
672 struct worker *worker = kthread_data(task);
673
Steven Rostedt2d646722010-12-03 23:12:33 -0500674 if (!(worker->flags & WORKER_NOT_RUNNING))
Tejun Heoe22bee72010-06-29 10:07:14 +0200675 atomic_inc(get_gcwq_nr_running(cpu));
676}
677
678/**
679 * wq_worker_sleeping - a worker is going to sleep
680 * @task: task going to sleep
681 * @cpu: CPU in question, must be the current CPU number
682 *
683 * This function is called during schedule() when a busy worker is
684 * going to sleep. Worker on the same cpu can be woken up by
685 * returning pointer to its task.
686 *
687 * CONTEXT:
688 * spin_lock_irq(rq->lock)
689 *
690 * RETURNS:
691 * Worker task on @cpu to wake up, %NULL if none.
692 */
693struct task_struct *wq_worker_sleeping(struct task_struct *task,
694 unsigned int cpu)
695{
696 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
697 struct global_cwq *gcwq = get_gcwq(cpu);
698 atomic_t *nr_running = get_gcwq_nr_running(cpu);
699
Steven Rostedt2d646722010-12-03 23:12:33 -0500700 if (worker->flags & WORKER_NOT_RUNNING)
Tejun Heoe22bee72010-06-29 10:07:14 +0200701 return NULL;
702
703 /* this can only happen on the local cpu */
704 BUG_ON(cpu != raw_smp_processor_id());
705
706 /*
707 * The counterpart of the following dec_and_test, implied mb,
708 * worklist not empty test sequence is in insert_work().
709 * Please read comment there.
710 *
711 * NOT_RUNNING is clear. This means that trustee is not in
712 * charge and we're running on the local cpu w/ rq lock held
713 * and preemption disabled, which in turn means that none else
714 * could be manipulating idle_list, so dereferencing idle_list
715 * without gcwq lock is safe.
716 */
717 if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist))
718 to_wakeup = first_worker(gcwq);
719 return to_wakeup ? to_wakeup->task : NULL;
720}
721
722/**
723 * worker_set_flags - set worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200724 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200725 * @flags: flags to set
726 * @wakeup: wakeup an idle worker if necessary
727 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200728 * Set @flags in @worker->flags and adjust nr_running accordingly. If
729 * nr_running becomes zero and @wakeup is %true, an idle worker is
730 * woken up.
Tejun Heod302f012010-06-29 10:07:13 +0200731 *
Tejun Heocb444762010-07-02 10:03:50 +0200732 * CONTEXT:
733 * spin_lock_irq(gcwq->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200734 */
735static inline void worker_set_flags(struct worker *worker, unsigned int flags,
736 bool wakeup)
737{
Tejun Heoe22bee72010-06-29 10:07:14 +0200738 struct global_cwq *gcwq = worker->gcwq;
739
Tejun Heocb444762010-07-02 10:03:50 +0200740 WARN_ON_ONCE(worker->task != current);
741
Tejun Heoe22bee72010-06-29 10:07:14 +0200742 /*
743 * If transitioning into NOT_RUNNING, adjust nr_running and
744 * wake up an idle worker as necessary if requested by
745 * @wakeup.
746 */
747 if ((flags & WORKER_NOT_RUNNING) &&
748 !(worker->flags & WORKER_NOT_RUNNING)) {
749 atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu);
750
751 if (wakeup) {
752 if (atomic_dec_and_test(nr_running) &&
753 !list_empty(&gcwq->worklist))
754 wake_up_worker(gcwq);
755 } else
756 atomic_dec(nr_running);
757 }
758
Tejun Heod302f012010-06-29 10:07:13 +0200759 worker->flags |= flags;
760}
761
762/**
Tejun Heoe22bee72010-06-29 10:07:14 +0200763 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
Tejun Heocb444762010-07-02 10:03:50 +0200764 * @worker: self
Tejun Heod302f012010-06-29 10:07:13 +0200765 * @flags: flags to clear
766 *
Tejun Heoe22bee72010-06-29 10:07:14 +0200767 * Clear @flags in @worker->flags and adjust nr_running accordingly.
Tejun Heod302f012010-06-29 10:07:13 +0200768 *
Tejun Heocb444762010-07-02 10:03:50 +0200769 * CONTEXT:
770 * spin_lock_irq(gcwq->lock)
Tejun Heod302f012010-06-29 10:07:13 +0200771 */
772static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
773{
Tejun Heoe22bee72010-06-29 10:07:14 +0200774 struct global_cwq *gcwq = worker->gcwq;
775 unsigned int oflags = worker->flags;
776
Tejun Heocb444762010-07-02 10:03:50 +0200777 WARN_ON_ONCE(worker->task != current);
778
Tejun Heod302f012010-06-29 10:07:13 +0200779 worker->flags &= ~flags;
Tejun Heoe22bee72010-06-29 10:07:14 +0200780
Tejun Heo42c025f2011-01-11 15:58:49 +0100781 /*
782 * If transitioning out of NOT_RUNNING, increment nr_running. Note
783 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
784 * of multiple flags, not a single flag.
785 */
Tejun Heoe22bee72010-06-29 10:07:14 +0200786 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
787 if (!(worker->flags & WORKER_NOT_RUNNING))
788 atomic_inc(get_gcwq_nr_running(gcwq->cpu));
Tejun Heod302f012010-06-29 10:07:13 +0200789}
790
791/**
Tejun Heoc8e55f32010-06-29 10:07:12 +0200792 * busy_worker_head - return the busy hash head for a work
793 * @gcwq: gcwq of interest
794 * @work: work to be hashed
795 *
796 * Return hash head of @gcwq for @work.
797 *
798 * CONTEXT:
799 * spin_lock_irq(gcwq->lock).
800 *
801 * RETURNS:
802 * Pointer to the hash head.
803 */
804static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
805 struct work_struct *work)
806{
807 const int base_shift = ilog2(sizeof(struct work_struct));
808 unsigned long v = (unsigned long)work;
809
810 /* simple shift and fold hash, do we need something better? */
811 v >>= base_shift;
812 v += v >> BUSY_WORKER_HASH_ORDER;
813 v &= BUSY_WORKER_HASH_MASK;
814
815 return &gcwq->busy_hash[v];
816}
817
818/**
Tejun Heo8cca0ee2010-06-29 10:07:13 +0200819 * __find_worker_executing_work - find worker which is executing a work
820 * @gcwq: gcwq of interest
821 * @bwh: hash head as returned by busy_worker_head()
822 * @work: work to find worker for
823 *
824 * Find a worker which is executing @work on @gcwq. @bwh should be
825 * the hash head obtained by calling busy_worker_head() with the same
826 * work.
827 *
828 * CONTEXT:
829 * spin_lock_irq(gcwq->lock).
830 *
831 * RETURNS:
832 * Pointer to worker which is executing @work if found, NULL
833 * otherwise.
834 */
835static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
836 struct hlist_head *bwh,
837 struct work_struct *work)
838{
839 struct worker *worker;
840 struct hlist_node *tmp;
841
842 hlist_for_each_entry(worker, tmp, bwh, hentry)
843 if (worker->current_work == work)
844 return worker;
845 return NULL;
846}
847
848/**
849 * find_worker_executing_work - find worker which is executing a work
850 * @gcwq: gcwq of interest
851 * @work: work to find worker for
852 *
853 * Find a worker which is executing @work on @gcwq. This function is
854 * identical to __find_worker_executing_work() except that this
855 * function calculates @bwh itself.
856 *
857 * CONTEXT:
858 * spin_lock_irq(gcwq->lock).
859 *
860 * RETURNS:
861 * Pointer to worker which is executing @work if found, NULL
862 * otherwise.
863 */
864static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
865 struct work_struct *work)
866{
867 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
868 work);
869}
870
871/**
Tejun Heo649027d2010-06-29 10:07:14 +0200872 * gcwq_determine_ins_pos - find insertion position
873 * @gcwq: gcwq of interest
874 * @cwq: cwq a work is being queued for
875 *
876 * A work for @cwq is about to be queued on @gcwq, determine insertion
877 * position for the work. If @cwq is for HIGHPRI wq, the work is
878 * queued at the head of the queue but in FIFO order with respect to
879 * other HIGHPRI works; otherwise, at the end of the queue. This
880 * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that
881 * there are HIGHPRI works pending.
882 *
883 * CONTEXT:
884 * spin_lock_irq(gcwq->lock).
885 *
886 * RETURNS:
887 * Pointer to inserstion position.
888 */
889static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq,
890 struct cpu_workqueue_struct *cwq)
891{
892 struct work_struct *twork;
893
894 if (likely(!(cwq->wq->flags & WQ_HIGHPRI)))
895 return &gcwq->worklist;
896
897 list_for_each_entry(twork, &gcwq->worklist, entry) {
898 struct cpu_workqueue_struct *tcwq = get_work_cwq(twork);
899
900 if (!(tcwq->wq->flags & WQ_HIGHPRI))
901 break;
902 }
903
904 gcwq->flags |= GCWQ_HIGHPRI_PENDING;
905 return &twork->entry;
906}
907
908/**
Tejun Heo7e116292010-06-29 10:07:13 +0200909 * insert_work - insert a work into gcwq
Tejun Heo4690c4a2010-06-29 10:07:10 +0200910 * @cwq: cwq @work belongs to
911 * @work: work to insert
912 * @head: insertion point
913 * @extra_flags: extra WORK_STRUCT_* flags to set
914 *
Tejun Heo7e116292010-06-29 10:07:13 +0200915 * Insert @work which belongs to @cwq into @gcwq after @head.
916 * @extra_flags is or'd to work_struct flags.
Tejun Heo4690c4a2010-06-29 10:07:10 +0200917 *
918 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +0200919 * spin_lock_irq(gcwq->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +0200920 */
Oleg Nesterovb89deed2007-05-09 02:33:52 -0700921static void insert_work(struct cpu_workqueue_struct *cwq,
Tejun Heo4690c4a2010-06-29 10:07:10 +0200922 struct work_struct *work, struct list_head *head,
923 unsigned int extra_flags)
Oleg Nesterovb89deed2007-05-09 02:33:52 -0700924{
Tejun Heoe22bee72010-06-29 10:07:14 +0200925 struct global_cwq *gcwq = cwq->gcwq;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +0100926
Tejun Heo4690c4a2010-06-29 10:07:10 +0200927 /* we own @work, set data and link */
Tejun Heo7a22ad72010-06-29 10:07:13 +0200928 set_work_cwq(work, cwq, extra_flags);
Tejun Heo4690c4a2010-06-29 10:07:10 +0200929
Oleg Nesterov6e84d642007-05-09 02:34:46 -0700930 /*
931 * Ensure that we get the right work->data if we see the
932 * result of list_add() below, see try_to_grab_pending().
933 */
934 smp_wmb();
Tejun Heo4690c4a2010-06-29 10:07:10 +0200935
Oleg Nesterov1a4d9b02008-07-25 01:47:47 -0700936 list_add_tail(&work->entry, head);
Tejun Heoe22bee72010-06-29 10:07:14 +0200937
938 /*
939 * Ensure either worker_sched_deactivated() sees the above
940 * list_add_tail() or we see zero nr_running to avoid workers
941 * lying around lazily while there are works to be processed.
942 */
943 smp_mb();
944
Tejun Heo649027d2010-06-29 10:07:14 +0200945 if (__need_more_worker(gcwq))
Tejun Heoe22bee72010-06-29 10:07:14 +0200946 wake_up_worker(gcwq);
Oleg Nesterovb89deed2007-05-09 02:33:52 -0700947}
948
Tejun Heoc8efcc22010-12-20 19:32:04 +0100949/*
950 * Test whether @work is being queued from another work executing on the
951 * same workqueue. This is rather expensive and should only be used from
952 * cold paths.
953 */
954static bool is_chained_work(struct workqueue_struct *wq)
955{
956 unsigned long flags;
957 unsigned int cpu;
958
959 for_each_gcwq_cpu(cpu) {
960 struct global_cwq *gcwq = get_gcwq(cpu);
961 struct worker *worker;
962 struct hlist_node *pos;
963 int i;
964
965 spin_lock_irqsave(&gcwq->lock, flags);
966 for_each_busy_worker(worker, i, pos, gcwq) {
967 if (worker->task != current)
968 continue;
969 spin_unlock_irqrestore(&gcwq->lock, flags);
970 /*
971 * I'm @worker, no locking necessary. See if @work
972 * is headed to the same workqueue.
973 */
974 return worker->current_cwq->wq == wq;
975 }
976 spin_unlock_irqrestore(&gcwq->lock, flags);
977 }
978 return false;
979}
980
Tejun Heo4690c4a2010-06-29 10:07:10 +0200981static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 struct work_struct *work)
983{
Tejun Heo502ca9d2010-06-29 10:07:13 +0200984 struct global_cwq *gcwq;
985 struct cpu_workqueue_struct *cwq;
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200986 struct list_head *worklist;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +0200987 unsigned int work_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 unsigned long flags;
989
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +0900990 debug_work_activate(work);
Tejun Heo1e19ffc2010-06-29 10:07:12 +0200991
Tejun Heoc8efcc22010-12-20 19:32:04 +0100992 /* if dying, only works from the same workqueue are allowed */
993 if (unlikely(wq->flags & WQ_DYING) &&
994 WARN_ON_ONCE(!is_chained_work(wq)))
Tejun Heoe41e7042010-08-24 14:22:47 +0200995 return;
996
Tejun Heoc7fc77f2010-07-02 10:03:51 +0200997 /* determine gcwq to use */
998 if (!(wq->flags & WQ_UNBOUND)) {
Tejun Heo18aa9ef2010-06-29 10:07:13 +0200999 struct global_cwq *last_gcwq;
1000
Tejun Heoc7fc77f2010-07-02 10:03:51 +02001001 if (unlikely(cpu == WORK_CPU_UNBOUND))
1002 cpu = raw_smp_processor_id();
1003
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001004 /*
1005 * It's multi cpu. If @wq is non-reentrant and @work
1006 * was previously on a different cpu, it might still
1007 * be running there, in which case the work needs to
1008 * be queued on that cpu to guarantee non-reentrance.
1009 */
Tejun Heo502ca9d2010-06-29 10:07:13 +02001010 gcwq = get_gcwq(cpu);
Tejun Heo18aa9ef2010-06-29 10:07:13 +02001011 if (wq->flags & WQ_NON_REENTRANT &&
1012 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1013 struct worker *worker;
1014
1015 spin_lock_irqsave(&last_gcwq->lock, flags);
1016
1017 worker = find_worker_executing_work(last_gcwq, work);
1018
1019 if (worker && worker->current_cwq->wq == wq)
1020 gcwq = last_gcwq;
1021 else {
1022 /* meh... not running there, queue here */
1023 spin_unlock_irqrestore(&last_gcwq->lock, flags);
1024 spin_lock_irqsave(&gcwq->lock, flags);
1025 }
1026 } else
1027 spin_lock_irqsave(&gcwq->lock, flags);
Tejun Heof3421792010-07-02 10:03:51 +02001028 } else {
1029 gcwq = get_gcwq(WORK_CPU_UNBOUND);
1030 spin_lock_irqsave(&gcwq->lock, flags);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001031 }
1032
1033 /* gcwq determined, get cwq and queue */
1034 cwq = get_cwq(gcwq->cpu, wq);
Tejun Heocdadf002010-10-05 10:49:55 +02001035 trace_workqueue_queue_work(cpu, cwq, work);
Tejun Heo502ca9d2010-06-29 10:07:13 +02001036
Tejun Heo4690c4a2010-06-29 10:07:10 +02001037 BUG_ON(!list_empty(&work->entry));
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001038
Tejun Heo73f53c42010-06-29 10:07:11 +02001039 cwq->nr_in_flight[cwq->work_color]++;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001040 work_flags = work_color_to_flags(cwq->work_color);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001041
1042 if (likely(cwq->nr_active < cwq->max_active)) {
Tejun Heocdadf002010-10-05 10:49:55 +02001043 trace_workqueue_activate_work(work);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001044 cwq->nr_active++;
Tejun Heo649027d2010-06-29 10:07:14 +02001045 worklist = gcwq_determine_ins_pos(gcwq, cwq);
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001046 } else {
1047 work_flags |= WORK_STRUCT_DELAYED;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001048 worklist = &cwq->delayed_works;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001049 }
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001050
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001051 insert_work(cwq, work, worklist, work_flags);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001052
Tejun Heo8b03ae32010-06-29 10:07:12 +02001053 spin_unlock_irqrestore(&gcwq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054}
1055
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001056/**
1057 * queue_work - queue work on a workqueue
1058 * @wq: workqueue to use
1059 * @work: work to queue
1060 *
Alan Stern057647f2006-10-28 10:38:58 -07001061 * Returns 0 if @work was already on a queue, non-zero otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062 *
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07001063 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1064 * it can be processed by another CPU.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08001066int queue_work(struct workqueue_struct *wq, struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067{
Oleg Nesterovef1ca232008-07-25 01:47:53 -07001068 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069
Oleg Nesterovef1ca232008-07-25 01:47:53 -07001070 ret = queue_work_on(get_cpu(), wq, work);
1071 put_cpu();
1072
Linus Torvalds1da177e2005-04-16 15:20:36 -07001073 return ret;
1074}
Dave Jonesae90dd52006-06-30 01:40:45 -04001075EXPORT_SYMBOL_GPL(queue_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076
Zhang Ruic1a220e2008-07-23 21:28:39 -07001077/**
1078 * queue_work_on - queue work on specific cpu
1079 * @cpu: CPU number to execute work on
1080 * @wq: workqueue to use
1081 * @work: work to queue
1082 *
1083 * Returns 0 if @work was already on a queue, non-zero otherwise.
1084 *
1085 * We queue the work to a specific CPU, the caller must ensure it
1086 * can't go away.
1087 */
1088int
1089queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work)
1090{
1091 int ret = 0;
1092
Tejun Heo22df02b2010-06-29 10:07:10 +02001093 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heo4690c4a2010-06-29 10:07:10 +02001094 __queue_work(cpu, wq, work);
Zhang Ruic1a220e2008-07-23 21:28:39 -07001095 ret = 1;
1096 }
1097 return ret;
1098}
1099EXPORT_SYMBOL_GPL(queue_work_on);
1100
Li Zefan6d141c32008-02-08 04:21:09 -08001101static void delayed_work_timer_fn(unsigned long __data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102{
David Howells52bad642006-11-22 14:54:01 +00001103 struct delayed_work *dwork = (struct delayed_work *)__data;
Tejun Heo7a22ad72010-06-29 10:07:13 +02001104 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
Tejun Heo4690c4a2010-06-29 10:07:10 +02001106 __queue_work(smp_processor_id(), cwq->wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107}
1108
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001109/**
1110 * queue_delayed_work - queue work on a workqueue after delay
1111 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001112 * @dwork: delayable work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001113 * @delay: number of jiffies to wait before queueing
1114 *
Alan Stern057647f2006-10-28 10:38:58 -07001115 * Returns 0 if @work was already on a queue, non-zero otherwise.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001116 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08001117int queue_delayed_work(struct workqueue_struct *wq,
David Howells52bad642006-11-22 14:54:01 +00001118 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119{
David Howells52bad642006-11-22 14:54:01 +00001120 if (delay == 0)
Oleg Nesterov63bc0362007-05-09 02:34:16 -07001121 return queue_work(wq, &dwork->work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122
Oleg Nesterov63bc0362007-05-09 02:34:16 -07001123 return queue_delayed_work_on(-1, wq, dwork, delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124}
Dave Jonesae90dd52006-06-30 01:40:45 -04001125EXPORT_SYMBOL_GPL(queue_delayed_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001127/**
1128 * queue_delayed_work_on - queue work on specific CPU after delay
1129 * @cpu: CPU number to execute work on
1130 * @wq: workqueue to use
Randy Dunlapaf9997e2006-12-22 01:06:52 -08001131 * @dwork: work to queue
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001132 * @delay: number of jiffies to wait before queueing
1133 *
Alan Stern057647f2006-10-28 10:38:58 -07001134 * Returns 0 if @work was already on a queue, non-zero otherwise.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07001135 */
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001136int queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
David Howells52bad642006-11-22 14:54:01 +00001137 struct delayed_work *dwork, unsigned long delay)
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001138{
1139 int ret = 0;
David Howells52bad642006-11-22 14:54:01 +00001140 struct timer_list *timer = &dwork->timer;
1141 struct work_struct *work = &dwork->work;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001142
Tejun Heo22df02b2010-06-29 10:07:10 +02001143 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
Tejun Heoc7fc77f2010-07-02 10:03:51 +02001144 unsigned int lcpu;
Tejun Heo7a22ad72010-06-29 10:07:13 +02001145
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001146 BUG_ON(timer_pending(timer));
1147 BUG_ON(!list_empty(&work->entry));
1148
Andrew Liu8a3e77c2008-05-01 04:35:14 -07001149 timer_stats_timer_set_start_info(&dwork->timer);
1150
Tejun Heo7a22ad72010-06-29 10:07:13 +02001151 /*
1152 * This stores cwq for the moment, for the timer_fn.
1153 * Note that the work's gcwq is preserved to allow
1154 * reentrance detection for delayed works.
1155 */
Tejun Heoc7fc77f2010-07-02 10:03:51 +02001156 if (!(wq->flags & WQ_UNBOUND)) {
1157 struct global_cwq *gcwq = get_work_gcwq(work);
1158
1159 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1160 lcpu = gcwq->cpu;
1161 else
1162 lcpu = raw_smp_processor_id();
1163 } else
1164 lcpu = WORK_CPU_UNBOUND;
1165
Tejun Heo7a22ad72010-06-29 10:07:13 +02001166 set_work_cwq(work, get_cwq(lcpu, wq), 0);
Tejun Heoc7fc77f2010-07-02 10:03:51 +02001167
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001168 timer->expires = jiffies + delay;
David Howells52bad642006-11-22 14:54:01 +00001169 timer->data = (unsigned long)dwork;
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001170 timer->function = delayed_work_timer_fn;
Oleg Nesterov63bc0362007-05-09 02:34:16 -07001171
1172 if (unlikely(cpu >= 0))
1173 add_timer_on(timer, cpu);
1174 else
1175 add_timer(timer);
Venkatesh Pallipadi7a6bc1c2006-06-28 13:50:33 -07001176 ret = 1;
1177 }
1178 return ret;
1179}
Dave Jonesae90dd52006-06-30 01:40:45 -04001180EXPORT_SYMBOL_GPL(queue_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181
Tejun Heoc8e55f32010-06-29 10:07:12 +02001182/**
1183 * worker_enter_idle - enter idle state
1184 * @worker: worker which is entering idle state
1185 *
1186 * @worker is entering idle state. Update stats and idle timer if
1187 * necessary.
1188 *
1189 * LOCKING:
1190 * spin_lock_irq(gcwq->lock).
1191 */
1192static void worker_enter_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193{
Tejun Heoc8e55f32010-06-29 10:07:12 +02001194 struct global_cwq *gcwq = worker->gcwq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195
Tejun Heoc8e55f32010-06-29 10:07:12 +02001196 BUG_ON(worker->flags & WORKER_IDLE);
1197 BUG_ON(!list_empty(&worker->entry) &&
1198 (worker->hentry.next || worker->hentry.pprev));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
Tejun Heocb444762010-07-02 10:03:50 +02001200 /* can't use worker_set_flags(), also called from start_worker() */
1201 worker->flags |= WORKER_IDLE;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001202 gcwq->nr_idle++;
Tejun Heoe22bee72010-06-29 10:07:14 +02001203 worker->last_active = jiffies;
Peter Zijlstrad5abe662006-12-06 20:37:26 -08001204
Tejun Heoc8e55f32010-06-29 10:07:12 +02001205 /* idle_list is LIFO */
1206 list_add(&worker->entry, &gcwq->idle_list);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001207
Tejun Heoe22bee72010-06-29 10:07:14 +02001208 if (likely(!(worker->flags & WORKER_ROGUE))) {
1209 if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer))
1210 mod_timer(&gcwq->idle_timer,
1211 jiffies + IDLE_WORKER_TIMEOUT);
1212 } else
Tejun Heodb7bccf2010-06-29 10:07:12 +02001213 wake_up_all(&gcwq->trustee_wait);
Tejun Heocb444762010-07-02 10:03:50 +02001214
1215 /* sanity check nr_running */
1216 WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle &&
1217 atomic_read(get_gcwq_nr_running(gcwq->cpu)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218}
1219
Tejun Heoc8e55f32010-06-29 10:07:12 +02001220/**
1221 * worker_leave_idle - leave idle state
1222 * @worker: worker which is leaving idle state
1223 *
1224 * @worker is leaving idle state. Update stats.
1225 *
1226 * LOCKING:
1227 * spin_lock_irq(gcwq->lock).
1228 */
1229static void worker_leave_idle(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230{
Tejun Heoc8e55f32010-06-29 10:07:12 +02001231 struct global_cwq *gcwq = worker->gcwq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232
Tejun Heoc8e55f32010-06-29 10:07:12 +02001233 BUG_ON(!(worker->flags & WORKER_IDLE));
Tejun Heod302f012010-06-29 10:07:13 +02001234 worker_clr_flags(worker, WORKER_IDLE);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001235 gcwq->nr_idle--;
1236 list_del_init(&worker->entry);
1237}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238
Tejun Heoe22bee72010-06-29 10:07:14 +02001239/**
1240 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1241 * @worker: self
1242 *
1243 * Works which are scheduled while the cpu is online must at least be
1244 * scheduled to a worker which is bound to the cpu so that if they are
1245 * flushed from cpu callbacks while cpu is going down, they are
1246 * guaranteed to execute on the cpu.
1247 *
1248 * This function is to be used by rogue workers and rescuers to bind
1249 * themselves to the target cpu and may race with cpu going down or
1250 * coming online. kthread_bind() can't be used because it may put the
1251 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1252 * verbatim as it's best effort and blocking and gcwq may be
1253 * [dis]associated in the meantime.
1254 *
1255 * This function tries set_cpus_allowed() and locks gcwq and verifies
1256 * the binding against GCWQ_DISASSOCIATED which is set during
1257 * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters
1258 * idle state or fetches works without dropping lock, it can guarantee
1259 * the scheduling requirement described in the first paragraph.
1260 *
1261 * CONTEXT:
1262 * Might sleep. Called without any lock but returns with gcwq->lock
1263 * held.
1264 *
1265 * RETURNS:
1266 * %true if the associated gcwq is online (@worker is successfully
1267 * bound), %false if offline.
1268 */
1269static bool worker_maybe_bind_and_lock(struct worker *worker)
Namhyung Kim972fa1c2010-08-22 23:19:43 +09001270__acquires(&gcwq->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001271{
1272 struct global_cwq *gcwq = worker->gcwq;
1273 struct task_struct *task = worker->task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001274
Tejun Heoe22bee72010-06-29 10:07:14 +02001275 while (true) {
1276 /*
1277 * The following call may fail, succeed or succeed
1278 * without actually migrating the task to the cpu if
1279 * it races with cpu hotunplug operation. Verify
1280 * against GCWQ_DISASSOCIATED.
1281 */
Tejun Heof3421792010-07-02 10:03:51 +02001282 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1283 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
Oleg Nesterov85f41862007-05-09 02:34:20 -07001284
Tejun Heoe22bee72010-06-29 10:07:14 +02001285 spin_lock_irq(&gcwq->lock);
1286 if (gcwq->flags & GCWQ_DISASSOCIATED)
1287 return false;
1288 if (task_cpu(task) == gcwq->cpu &&
1289 cpumask_equal(&current->cpus_allowed,
1290 get_cpu_mask(gcwq->cpu)))
1291 return true;
1292 spin_unlock_irq(&gcwq->lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07001293
Tejun Heoe22bee72010-06-29 10:07:14 +02001294 /* CPU has come up inbetween, retry migration */
1295 cpu_relax();
1296 }
1297}
1298
1299/*
1300 * Function for worker->rebind_work used to rebind rogue busy workers
1301 * to the associated cpu which is coming back online. This is
1302 * scheduled by cpu up but can race with other cpu hotplug operations
1303 * and may be executed twice without intervening cpu down.
1304 */
1305static void worker_rebind_fn(struct work_struct *work)
1306{
1307 struct worker *worker = container_of(work, struct worker, rebind_work);
1308 struct global_cwq *gcwq = worker->gcwq;
1309
1310 if (worker_maybe_bind_and_lock(worker))
1311 worker_clr_flags(worker, WORKER_REBIND);
1312
1313 spin_unlock_irq(&gcwq->lock);
1314}
1315
Tejun Heoc34056a2010-06-29 10:07:11 +02001316static struct worker *alloc_worker(void)
1317{
1318 struct worker *worker;
1319
1320 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001321 if (worker) {
1322 INIT_LIST_HEAD(&worker->entry);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001323 INIT_LIST_HEAD(&worker->scheduled);
Tejun Heoe22bee72010-06-29 10:07:14 +02001324 INIT_WORK(&worker->rebind_work, worker_rebind_fn);
1325 /* on creation a worker is in !idle && prep state */
1326 worker->flags = WORKER_PREP;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001327 }
Tejun Heoc34056a2010-06-29 10:07:11 +02001328 return worker;
1329}
1330
1331/**
1332 * create_worker - create a new workqueue worker
Tejun Heo7e116292010-06-29 10:07:13 +02001333 * @gcwq: gcwq the new worker will belong to
Tejun Heoc34056a2010-06-29 10:07:11 +02001334 * @bind: whether to set affinity to @cpu or not
1335 *
Tejun Heo7e116292010-06-29 10:07:13 +02001336 * Create a new worker which is bound to @gcwq. The returned worker
Tejun Heoc34056a2010-06-29 10:07:11 +02001337 * can be started by calling start_worker() or destroyed using
1338 * destroy_worker().
1339 *
1340 * CONTEXT:
1341 * Might sleep. Does GFP_KERNEL allocations.
1342 *
1343 * RETURNS:
1344 * Pointer to the newly created worker.
1345 */
Tejun Heo7e116292010-06-29 10:07:13 +02001346static struct worker *create_worker(struct global_cwq *gcwq, bool bind)
Tejun Heoc34056a2010-06-29 10:07:11 +02001347{
Tejun Heof3421792010-07-02 10:03:51 +02001348 bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND;
Tejun Heoc34056a2010-06-29 10:07:11 +02001349 struct worker *worker = NULL;
Tejun Heof3421792010-07-02 10:03:51 +02001350 int id = -1;
Tejun Heoc34056a2010-06-29 10:07:11 +02001351
Tejun Heo8b03ae32010-06-29 10:07:12 +02001352 spin_lock_irq(&gcwq->lock);
1353 while (ida_get_new(&gcwq->worker_ida, &id)) {
1354 spin_unlock_irq(&gcwq->lock);
1355 if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL))
Tejun Heoc34056a2010-06-29 10:07:11 +02001356 goto fail;
Tejun Heo8b03ae32010-06-29 10:07:12 +02001357 spin_lock_irq(&gcwq->lock);
Tejun Heoc34056a2010-06-29 10:07:11 +02001358 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02001359 spin_unlock_irq(&gcwq->lock);
Tejun Heoc34056a2010-06-29 10:07:11 +02001360
1361 worker = alloc_worker();
1362 if (!worker)
1363 goto fail;
1364
Tejun Heo8b03ae32010-06-29 10:07:12 +02001365 worker->gcwq = gcwq;
Tejun Heoc34056a2010-06-29 10:07:11 +02001366 worker->id = id;
1367
Tejun Heof3421792010-07-02 10:03:51 +02001368 if (!on_unbound_cpu)
Eric Dumazet94dcf292011-03-22 16:30:45 -07001369 worker->task = kthread_create_on_node(worker_thread,
1370 worker,
1371 cpu_to_node(gcwq->cpu),
1372 "kworker/%u:%d", gcwq->cpu, id);
Tejun Heof3421792010-07-02 10:03:51 +02001373 else
1374 worker->task = kthread_create(worker_thread, worker,
1375 "kworker/u:%d", id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001376 if (IS_ERR(worker->task))
1377 goto fail;
1378
Tejun Heodb7bccf2010-06-29 10:07:12 +02001379 /*
1380 * A rogue worker will become a regular one if CPU comes
1381 * online later on. Make sure every worker has
1382 * PF_THREAD_BOUND set.
1383 */
Tejun Heof3421792010-07-02 10:03:51 +02001384 if (bind && !on_unbound_cpu)
Tejun Heo8b03ae32010-06-29 10:07:12 +02001385 kthread_bind(worker->task, gcwq->cpu);
Tejun Heof3421792010-07-02 10:03:51 +02001386 else {
Tejun Heodb7bccf2010-06-29 10:07:12 +02001387 worker->task->flags |= PF_THREAD_BOUND;
Tejun Heof3421792010-07-02 10:03:51 +02001388 if (on_unbound_cpu)
1389 worker->flags |= WORKER_UNBOUND;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07001391
Tejun Heoc34056a2010-06-29 10:07:11 +02001392 return worker;
1393fail:
1394 if (id >= 0) {
Tejun Heo8b03ae32010-06-29 10:07:12 +02001395 spin_lock_irq(&gcwq->lock);
1396 ida_remove(&gcwq->worker_ida, id);
1397 spin_unlock_irq(&gcwq->lock);
Tejun Heoc34056a2010-06-29 10:07:11 +02001398 }
1399 kfree(worker);
1400 return NULL;
1401}
1402
1403/**
1404 * start_worker - start a newly created worker
1405 * @worker: worker to start
1406 *
Tejun Heoc8e55f32010-06-29 10:07:12 +02001407 * Make the gcwq aware of @worker and start it.
Tejun Heoc34056a2010-06-29 10:07:11 +02001408 *
1409 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02001410 * spin_lock_irq(gcwq->lock).
Tejun Heoc34056a2010-06-29 10:07:11 +02001411 */
1412static void start_worker(struct worker *worker)
1413{
Tejun Heocb444762010-07-02 10:03:50 +02001414 worker->flags |= WORKER_STARTED;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001415 worker->gcwq->nr_workers++;
1416 worker_enter_idle(worker);
Tejun Heoc34056a2010-06-29 10:07:11 +02001417 wake_up_process(worker->task);
1418}
1419
1420/**
1421 * destroy_worker - destroy a workqueue worker
1422 * @worker: worker to be destroyed
1423 *
Tejun Heoc8e55f32010-06-29 10:07:12 +02001424 * Destroy @worker and adjust @gcwq stats accordingly.
1425 *
1426 * CONTEXT:
1427 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
Tejun Heoc34056a2010-06-29 10:07:11 +02001428 */
1429static void destroy_worker(struct worker *worker)
1430{
Tejun Heo8b03ae32010-06-29 10:07:12 +02001431 struct global_cwq *gcwq = worker->gcwq;
Tejun Heoc34056a2010-06-29 10:07:11 +02001432 int id = worker->id;
1433
1434 /* sanity check frenzy */
1435 BUG_ON(worker->current_work);
Tejun Heoaffee4b2010-06-29 10:07:12 +02001436 BUG_ON(!list_empty(&worker->scheduled));
Tejun Heoc34056a2010-06-29 10:07:11 +02001437
Tejun Heoc8e55f32010-06-29 10:07:12 +02001438 if (worker->flags & WORKER_STARTED)
1439 gcwq->nr_workers--;
1440 if (worker->flags & WORKER_IDLE)
1441 gcwq->nr_idle--;
1442
1443 list_del_init(&worker->entry);
Tejun Heocb444762010-07-02 10:03:50 +02001444 worker->flags |= WORKER_DIE;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001445
1446 spin_unlock_irq(&gcwq->lock);
1447
Tejun Heoc34056a2010-06-29 10:07:11 +02001448 kthread_stop(worker->task);
1449 kfree(worker);
1450
Tejun Heo8b03ae32010-06-29 10:07:12 +02001451 spin_lock_irq(&gcwq->lock);
1452 ida_remove(&gcwq->worker_ida, id);
Tejun Heoc34056a2010-06-29 10:07:11 +02001453}
1454
Tejun Heoe22bee72010-06-29 10:07:14 +02001455static void idle_worker_timeout(unsigned long __gcwq)
1456{
1457 struct global_cwq *gcwq = (void *)__gcwq;
1458
1459 spin_lock_irq(&gcwq->lock);
1460
1461 if (too_many_workers(gcwq)) {
1462 struct worker *worker;
1463 unsigned long expires;
1464
1465 /* idle_list is kept in LIFO order, check the last one */
1466 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1467 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1468
1469 if (time_before(jiffies, expires))
1470 mod_timer(&gcwq->idle_timer, expires);
1471 else {
1472 /* it's been idle for too long, wake up manager */
1473 gcwq->flags |= GCWQ_MANAGE_WORKERS;
1474 wake_up_worker(gcwq);
1475 }
1476 }
1477
1478 spin_unlock_irq(&gcwq->lock);
1479}
1480
1481static bool send_mayday(struct work_struct *work)
1482{
1483 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1484 struct workqueue_struct *wq = cwq->wq;
Tejun Heof3421792010-07-02 10:03:51 +02001485 unsigned int cpu;
Tejun Heoe22bee72010-06-29 10:07:14 +02001486
1487 if (!(wq->flags & WQ_RESCUER))
1488 return false;
1489
1490 /* mayday mayday mayday */
Tejun Heof3421792010-07-02 10:03:51 +02001491 cpu = cwq->gcwq->cpu;
1492 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1493 if (cpu == WORK_CPU_UNBOUND)
1494 cpu = 0;
Tejun Heof2e005a2010-07-20 15:59:09 +02001495 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
Tejun Heoe22bee72010-06-29 10:07:14 +02001496 wake_up_process(wq->rescuer->task);
1497 return true;
1498}
1499
1500static void gcwq_mayday_timeout(unsigned long __gcwq)
1501{
1502 struct global_cwq *gcwq = (void *)__gcwq;
1503 struct work_struct *work;
1504
1505 spin_lock_irq(&gcwq->lock);
1506
1507 if (need_to_create_worker(gcwq)) {
1508 /*
1509 * We've been trying to create a new worker but
1510 * haven't been successful. We might be hitting an
1511 * allocation deadlock. Send distress signals to
1512 * rescuers.
1513 */
1514 list_for_each_entry(work, &gcwq->worklist, entry)
1515 send_mayday(work);
1516 }
1517
1518 spin_unlock_irq(&gcwq->lock);
1519
1520 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL);
1521}
1522
1523/**
1524 * maybe_create_worker - create a new worker if necessary
1525 * @gcwq: gcwq to create a new worker for
1526 *
1527 * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to
1528 * have at least one idle worker on return from this function. If
1529 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
1530 * sent to all rescuers with works scheduled on @gcwq to resolve
1531 * possible allocation deadlock.
1532 *
1533 * On return, need_to_create_worker() is guaranteed to be false and
1534 * may_start_working() true.
1535 *
1536 * LOCKING:
1537 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1538 * multiple times. Does GFP_KERNEL allocations. Called only from
1539 * manager.
1540 *
1541 * RETURNS:
1542 * false if no action was taken and gcwq->lock stayed locked, true
1543 * otherwise.
1544 */
1545static bool maybe_create_worker(struct global_cwq *gcwq)
Namhyung Kim06bd6eb2010-08-22 23:19:42 +09001546__releases(&gcwq->lock)
1547__acquires(&gcwq->lock)
Tejun Heoe22bee72010-06-29 10:07:14 +02001548{
1549 if (!need_to_create_worker(gcwq))
1550 return false;
1551restart:
Tejun Heo9f9c2362010-07-14 11:31:20 +02001552 spin_unlock_irq(&gcwq->lock);
1553
Tejun Heoe22bee72010-06-29 10:07:14 +02001554 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
1555 mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
1556
1557 while (true) {
1558 struct worker *worker;
1559
Tejun Heoe22bee72010-06-29 10:07:14 +02001560 worker = create_worker(gcwq, true);
1561 if (worker) {
1562 del_timer_sync(&gcwq->mayday_timer);
1563 spin_lock_irq(&gcwq->lock);
1564 start_worker(worker);
1565 BUG_ON(need_to_create_worker(gcwq));
1566 return true;
1567 }
1568
1569 if (!need_to_create_worker(gcwq))
1570 break;
1571
Tejun Heoe22bee72010-06-29 10:07:14 +02001572 __set_current_state(TASK_INTERRUPTIBLE);
1573 schedule_timeout(CREATE_COOLDOWN);
Tejun Heo9f9c2362010-07-14 11:31:20 +02001574
Tejun Heoe22bee72010-06-29 10:07:14 +02001575 if (!need_to_create_worker(gcwq))
1576 break;
1577 }
1578
Tejun Heoe22bee72010-06-29 10:07:14 +02001579 del_timer_sync(&gcwq->mayday_timer);
1580 spin_lock_irq(&gcwq->lock);
1581 if (need_to_create_worker(gcwq))
1582 goto restart;
1583 return true;
1584}
1585
1586/**
1587 * maybe_destroy_worker - destroy workers which have been idle for a while
1588 * @gcwq: gcwq to destroy workers for
1589 *
1590 * Destroy @gcwq workers which have been idle for longer than
1591 * IDLE_WORKER_TIMEOUT.
1592 *
1593 * LOCKING:
1594 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1595 * multiple times. Called only from manager.
1596 *
1597 * RETURNS:
1598 * false if no action was taken and gcwq->lock stayed locked, true
1599 * otherwise.
1600 */
1601static bool maybe_destroy_workers(struct global_cwq *gcwq)
1602{
1603 bool ret = false;
1604
1605 while (too_many_workers(gcwq)) {
1606 struct worker *worker;
1607 unsigned long expires;
1608
1609 worker = list_entry(gcwq->idle_list.prev, struct worker, entry);
1610 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1611
1612 if (time_before(jiffies, expires)) {
1613 mod_timer(&gcwq->idle_timer, expires);
1614 break;
1615 }
1616
1617 destroy_worker(worker);
1618 ret = true;
1619 }
1620
1621 return ret;
1622}
1623
1624/**
1625 * manage_workers - manage worker pool
1626 * @worker: self
1627 *
1628 * Assume the manager role and manage gcwq worker pool @worker belongs
1629 * to. At any given time, there can be only zero or one manager per
1630 * gcwq. The exclusion is handled automatically by this function.
1631 *
1632 * The caller can safely start processing works on false return. On
1633 * true return, it's guaranteed that need_to_create_worker() is false
1634 * and may_start_working() is true.
1635 *
1636 * CONTEXT:
1637 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1638 * multiple times. Does GFP_KERNEL allocations.
1639 *
1640 * RETURNS:
1641 * false if no action was taken and gcwq->lock stayed locked, true if
1642 * some action was taken.
1643 */
1644static bool manage_workers(struct worker *worker)
1645{
1646 struct global_cwq *gcwq = worker->gcwq;
1647 bool ret = false;
1648
1649 if (gcwq->flags & GCWQ_MANAGING_WORKERS)
1650 return ret;
1651
1652 gcwq->flags &= ~GCWQ_MANAGE_WORKERS;
1653 gcwq->flags |= GCWQ_MANAGING_WORKERS;
1654
1655 /*
1656 * Destroy and then create so that may_start_working() is true
1657 * on return.
1658 */
1659 ret |= maybe_destroy_workers(gcwq);
1660 ret |= maybe_create_worker(gcwq);
1661
1662 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
1663
1664 /*
1665 * The trustee might be waiting to take over the manager
1666 * position, tell it we're done.
1667 */
1668 if (unlikely(gcwq->trustee))
1669 wake_up_all(&gcwq->trustee_wait);
1670
1671 return ret;
1672}
1673
Tejun Heoa62428c2010-06-29 10:07:10 +02001674/**
Tejun Heoaffee4b2010-06-29 10:07:12 +02001675 * move_linked_works - move linked works to a list
1676 * @work: start of series of works to be scheduled
1677 * @head: target list to append @work to
1678 * @nextp: out paramter for nested worklist walking
1679 *
1680 * Schedule linked works starting from @work to @head. Work series to
1681 * be scheduled starts at @work and includes any consecutive work with
1682 * WORK_STRUCT_LINKED set in its predecessor.
1683 *
1684 * If @nextp is not NULL, it's updated to point to the next work of
1685 * the last scheduled work. This allows move_linked_works() to be
1686 * nested inside outer list_for_each_entry_safe().
1687 *
1688 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02001689 * spin_lock_irq(gcwq->lock).
Tejun Heoaffee4b2010-06-29 10:07:12 +02001690 */
1691static void move_linked_works(struct work_struct *work, struct list_head *head,
1692 struct work_struct **nextp)
1693{
1694 struct work_struct *n;
1695
1696 /*
1697 * Linked worklist will always end before the end of the list,
1698 * use NULL for list head.
1699 */
1700 list_for_each_entry_safe_from(work, n, NULL, entry) {
1701 list_move_tail(&work->entry, head);
1702 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
1703 break;
1704 }
1705
1706 /*
1707 * If we're already inside safe list traversal and have moved
1708 * multiple works to the scheduled queue, the next position
1709 * needs to be updated.
1710 */
1711 if (nextp)
1712 *nextp = n;
1713}
1714
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001715static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
1716{
1717 struct work_struct *work = list_first_entry(&cwq->delayed_works,
1718 struct work_struct, entry);
Tejun Heo649027d2010-06-29 10:07:14 +02001719 struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001720
Tejun Heocdadf002010-10-05 10:49:55 +02001721 trace_workqueue_activate_work(work);
Tejun Heo649027d2010-06-29 10:07:14 +02001722 move_linked_works(work, pos, NULL);
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001723 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001724 cwq->nr_active++;
1725}
1726
Tejun Heoaffee4b2010-06-29 10:07:12 +02001727/**
Tejun Heo73f53c42010-06-29 10:07:11 +02001728 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
1729 * @cwq: cwq of interest
1730 * @color: color of work which left the queue
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001731 * @delayed: for a delayed work
Tejun Heo73f53c42010-06-29 10:07:11 +02001732 *
1733 * A work either has completed or is removed from pending queue,
1734 * decrement nr_in_flight of its cwq and handle workqueue flushing.
1735 *
1736 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02001737 * spin_lock_irq(gcwq->lock).
Tejun Heo73f53c42010-06-29 10:07:11 +02001738 */
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001739static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
1740 bool delayed)
Tejun Heo73f53c42010-06-29 10:07:11 +02001741{
1742 /* ignore uncolored works */
1743 if (color == WORK_NO_COLOR)
1744 return;
1745
1746 cwq->nr_in_flight[color]--;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02001747
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001748 if (!delayed) {
1749 cwq->nr_active--;
1750 if (!list_empty(&cwq->delayed_works)) {
1751 /* one down, submit a delayed one */
1752 if (cwq->nr_active < cwq->max_active)
1753 cwq_activate_first_delayed(cwq);
1754 }
Tejun Heo502ca9d2010-06-29 10:07:13 +02001755 }
Tejun Heo73f53c42010-06-29 10:07:11 +02001756
1757 /* is flush in progress and are we at the flushing tip? */
1758 if (likely(cwq->flush_color != color))
1759 return;
1760
1761 /* are there still in-flight works? */
1762 if (cwq->nr_in_flight[color])
1763 return;
1764
1765 /* this cwq is done, clear flush_color */
1766 cwq->flush_color = -1;
1767
1768 /*
1769 * If this was the last cwq, wake up the first flusher. It
1770 * will handle the rest.
1771 */
1772 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1773 complete(&cwq->wq->first_flusher->done);
1774}
1775
1776/**
Tejun Heoa62428c2010-06-29 10:07:10 +02001777 * process_one_work - process single work
Tejun Heoc34056a2010-06-29 10:07:11 +02001778 * @worker: self
Tejun Heoa62428c2010-06-29 10:07:10 +02001779 * @work: work to process
1780 *
1781 * Process @work. This function contains all the logics necessary to
1782 * process a single work including synchronization against and
1783 * interaction with other workers on the same cpu, queueing and
1784 * flushing. As long as context requirement is met, any worker can
1785 * call this function to process a work.
1786 *
1787 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02001788 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
Tejun Heoa62428c2010-06-29 10:07:10 +02001789 */
Tejun Heoc34056a2010-06-29 10:07:11 +02001790static void process_one_work(struct worker *worker, struct work_struct *work)
Namhyung Kim06bd6eb2010-08-22 23:19:42 +09001791__releases(&gcwq->lock)
1792__acquires(&gcwq->lock)
Tejun Heoa62428c2010-06-29 10:07:10 +02001793{
Tejun Heo7e116292010-06-29 10:07:13 +02001794 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
Tejun Heo8b03ae32010-06-29 10:07:12 +02001795 struct global_cwq *gcwq = cwq->gcwq;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001796 struct hlist_head *bwh = busy_worker_head(gcwq, work);
Tejun Heofb0e7be2010-06-29 10:07:15 +02001797 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
Tejun Heoa62428c2010-06-29 10:07:10 +02001798 work_func_t f = work->func;
Tejun Heo73f53c42010-06-29 10:07:11 +02001799 int work_color;
Tejun Heo7e116292010-06-29 10:07:13 +02001800 struct worker *collision;
Tejun Heoa62428c2010-06-29 10:07:10 +02001801#ifdef CONFIG_LOCKDEP
1802 /*
1803 * It is permissible to free the struct work_struct from
1804 * inside the function that is called from it, this we need to
1805 * take into account for lockdep too. To avoid bogus "held
1806 * lock freed" warnings as well as problems when looking into
1807 * work->lockdep_map, make a copy and use that here.
1808 */
1809 struct lockdep_map lockdep_map = work->lockdep_map;
1810#endif
Tejun Heo7e116292010-06-29 10:07:13 +02001811 /*
1812 * A single work shouldn't be executed concurrently by
1813 * multiple workers on a single cpu. Check whether anyone is
1814 * already processing the work. If so, defer the work to the
1815 * currently executing one.
1816 */
1817 collision = __find_worker_executing_work(gcwq, bwh, work);
1818 if (unlikely(collision)) {
1819 move_linked_works(work, &collision->scheduled, NULL);
1820 return;
1821 }
1822
Tejun Heoa62428c2010-06-29 10:07:10 +02001823 /* claim and process */
Tejun Heoa62428c2010-06-29 10:07:10 +02001824 debug_work_deactivate(work);
Tejun Heoc8e55f32010-06-29 10:07:12 +02001825 hlist_add_head(&worker->hentry, bwh);
Tejun Heoc34056a2010-06-29 10:07:11 +02001826 worker->current_work = work;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001827 worker->current_cwq = cwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02001828 work_color = get_work_color(work);
Tejun Heo7a22ad72010-06-29 10:07:13 +02001829
Tejun Heo7a22ad72010-06-29 10:07:13 +02001830 /* record the current cpu number in the work data and dequeue */
1831 set_work_cpu(work, gcwq->cpu);
Tejun Heoa62428c2010-06-29 10:07:10 +02001832 list_del_init(&work->entry);
1833
Tejun Heo649027d2010-06-29 10:07:14 +02001834 /*
1835 * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI,
1836 * wake up another worker; otherwise, clear HIGHPRI_PENDING.
1837 */
1838 if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) {
1839 struct work_struct *nwork = list_first_entry(&gcwq->worklist,
1840 struct work_struct, entry);
1841
1842 if (!list_empty(&gcwq->worklist) &&
1843 get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI)
1844 wake_up_worker(gcwq);
1845 else
1846 gcwq->flags &= ~GCWQ_HIGHPRI_PENDING;
1847 }
1848
Tejun Heofb0e7be2010-06-29 10:07:15 +02001849 /*
1850 * CPU intensive works don't participate in concurrency
1851 * management. They're the scheduler's responsibility.
1852 */
1853 if (unlikely(cpu_intensive))
1854 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
1855
Tejun Heo8b03ae32010-06-29 10:07:12 +02001856 spin_unlock_irq(&gcwq->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02001857
Tejun Heoa62428c2010-06-29 10:07:10 +02001858 work_clear_pending(work);
Tejun Heoe1594892011-01-09 23:32:15 +01001859 lock_map_acquire_read(&cwq->wq->lockdep_map);
Tejun Heoa62428c2010-06-29 10:07:10 +02001860 lock_map_acquire(&lockdep_map);
Arjan van de Vene36c8862010-08-21 13:07:26 -07001861 trace_workqueue_execute_start(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02001862 f(work);
Arjan van de Vene36c8862010-08-21 13:07:26 -07001863 /*
1864 * While we must be careful to not use "work" after this, the trace
1865 * point will only record its address.
1866 */
1867 trace_workqueue_execute_end(work);
Tejun Heoa62428c2010-06-29 10:07:10 +02001868 lock_map_release(&lockdep_map);
1869 lock_map_release(&cwq->wq->lockdep_map);
1870
1871 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
1872 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
1873 "%s/0x%08x/%d\n",
1874 current->comm, preempt_count(), task_pid_nr(current));
1875 printk(KERN_ERR " last function: ");
1876 print_symbol("%s\n", (unsigned long)f);
1877 debug_show_held_locks(current);
1878 dump_stack();
1879 }
1880
Tejun Heo8b03ae32010-06-29 10:07:12 +02001881 spin_lock_irq(&gcwq->lock);
Tejun Heoa62428c2010-06-29 10:07:10 +02001882
Tejun Heofb0e7be2010-06-29 10:07:15 +02001883 /* clear cpu intensive status */
1884 if (unlikely(cpu_intensive))
1885 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
1886
Tejun Heoa62428c2010-06-29 10:07:10 +02001887 /* we're done with it, release */
Tejun Heoc8e55f32010-06-29 10:07:12 +02001888 hlist_del_init(&worker->hentry);
Tejun Heoc34056a2010-06-29 10:07:11 +02001889 worker->current_work = NULL;
Tejun Heo8cca0ee2010-06-29 10:07:13 +02001890 worker->current_cwq = NULL;
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02001891 cwq_dec_nr_in_flight(cwq, work_color, false);
Tejun Heoa62428c2010-06-29 10:07:10 +02001892}
1893
Tejun Heoaffee4b2010-06-29 10:07:12 +02001894/**
1895 * process_scheduled_works - process scheduled works
1896 * @worker: self
1897 *
1898 * Process all scheduled works. Please note that the scheduled list
1899 * may change while processing a work, so this function repeatedly
1900 * fetches a work from the top and executes it.
1901 *
1902 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02001903 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
Tejun Heoaffee4b2010-06-29 10:07:12 +02001904 * multiple times.
1905 */
1906static void process_scheduled_works(struct worker *worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907{
Tejun Heoaffee4b2010-06-29 10:07:12 +02001908 while (!list_empty(&worker->scheduled)) {
1909 struct work_struct *work = list_first_entry(&worker->scheduled,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 struct work_struct, entry);
Tejun Heoc34056a2010-06-29 10:07:11 +02001911 process_one_work(worker, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913}
1914
Tejun Heo4690c4a2010-06-29 10:07:10 +02001915/**
1916 * worker_thread - the worker thread function
Tejun Heoc34056a2010-06-29 10:07:11 +02001917 * @__worker: self
Tejun Heo4690c4a2010-06-29 10:07:10 +02001918 *
Tejun Heoe22bee72010-06-29 10:07:14 +02001919 * The gcwq worker thread function. There's a single dynamic pool of
1920 * these per each cpu. These workers process all works regardless of
1921 * their specific target workqueue. The only exception is works which
1922 * belong to workqueues with a rescuer which will be explained in
1923 * rescuer_thread().
Tejun Heo4690c4a2010-06-29 10:07:10 +02001924 */
Tejun Heoc34056a2010-06-29 10:07:11 +02001925static int worker_thread(void *__worker)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926{
Tejun Heoc34056a2010-06-29 10:07:11 +02001927 struct worker *worker = __worker;
Tejun Heo8b03ae32010-06-29 10:07:12 +02001928 struct global_cwq *gcwq = worker->gcwq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929
Tejun Heoe22bee72010-06-29 10:07:14 +02001930 /* tell the scheduler that this is a workqueue worker */
1931 worker->task->flags |= PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001932woke_up:
Tejun Heoc8e55f32010-06-29 10:07:12 +02001933 spin_lock_irq(&gcwq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934
Tejun Heoc8e55f32010-06-29 10:07:12 +02001935 /* DIE can be set only while we're idle, checking here is enough */
1936 if (worker->flags & WORKER_DIE) {
Tejun Heo8b03ae32010-06-29 10:07:12 +02001937 spin_unlock_irq(&gcwq->lock);
Tejun Heoe22bee72010-06-29 10:07:14 +02001938 worker->task->flags &= ~PF_WQ_WORKER;
Tejun Heoc8e55f32010-06-29 10:07:12 +02001939 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 }
1941
Tejun Heoc8e55f32010-06-29 10:07:12 +02001942 worker_leave_idle(worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02001943recheck:
Tejun Heoe22bee72010-06-29 10:07:14 +02001944 /* no more worker necessary? */
1945 if (!need_more_worker(gcwq))
1946 goto sleep;
1947
1948 /* do we need to manage? */
1949 if (unlikely(!may_start_working(gcwq)) && manage_workers(worker))
1950 goto recheck;
1951
Tejun Heoc8e55f32010-06-29 10:07:12 +02001952 /*
1953 * ->scheduled list can only be filled while a worker is
1954 * preparing to process a work or actually processing it.
1955 * Make sure nobody diddled with it while I was sleeping.
1956 */
1957 BUG_ON(!list_empty(&worker->scheduled));
1958
Tejun Heoe22bee72010-06-29 10:07:14 +02001959 /*
1960 * When control reaches this point, we're guaranteed to have
1961 * at least one idle worker or that someone else has already
1962 * assumed the manager role.
1963 */
1964 worker_clr_flags(worker, WORKER_PREP);
1965
1966 do {
Tejun Heoc8e55f32010-06-29 10:07:12 +02001967 struct work_struct *work =
Tejun Heo7e116292010-06-29 10:07:13 +02001968 list_first_entry(&gcwq->worklist,
Tejun Heoc8e55f32010-06-29 10:07:12 +02001969 struct work_struct, entry);
1970
1971 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
1972 /* optimization path, not strictly necessary */
1973 process_one_work(worker, work);
1974 if (unlikely(!list_empty(&worker->scheduled)))
1975 process_scheduled_works(worker);
1976 } else {
1977 move_linked_works(work, &worker->scheduled, NULL);
1978 process_scheduled_works(worker);
1979 }
Tejun Heoe22bee72010-06-29 10:07:14 +02001980 } while (keep_working(gcwq));
Tejun Heoc8e55f32010-06-29 10:07:12 +02001981
Tejun Heoe22bee72010-06-29 10:07:14 +02001982 worker_set_flags(worker, WORKER_PREP, false);
Tejun Heod313dd82010-07-02 10:03:51 +02001983sleep:
Tejun Heoe22bee72010-06-29 10:07:14 +02001984 if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker))
1985 goto recheck;
Tejun Heod313dd82010-07-02 10:03:51 +02001986
Tejun Heoc8e55f32010-06-29 10:07:12 +02001987 /*
Tejun Heoe22bee72010-06-29 10:07:14 +02001988 * gcwq->lock is held and there's no work to process and no
1989 * need to manage, sleep. Workers are woken up only while
1990 * holding gcwq->lock or from local cpu, so setting the
1991 * current state before releasing gcwq->lock is enough to
1992 * prevent losing any event.
Tejun Heoc8e55f32010-06-29 10:07:12 +02001993 */
1994 worker_enter_idle(worker);
1995 __set_current_state(TASK_INTERRUPTIBLE);
1996 spin_unlock_irq(&gcwq->lock);
1997 schedule();
1998 goto woke_up;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999}
2000
Tejun Heoe22bee72010-06-29 10:07:14 +02002001/**
2002 * rescuer_thread - the rescuer thread function
2003 * @__wq: the associated workqueue
2004 *
2005 * Workqueue rescuer thread function. There's one rescuer for each
2006 * workqueue which has WQ_RESCUER set.
2007 *
2008 * Regular work processing on a gcwq may block trying to create a new
2009 * worker which uses GFP_KERNEL allocation which has slight chance of
2010 * developing into deadlock if some works currently on the same queue
2011 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2012 * the problem rescuer solves.
2013 *
2014 * When such condition is possible, the gcwq summons rescuers of all
2015 * workqueues which have works queued on the gcwq and let them process
2016 * those works so that forward progress can be guaranteed.
2017 *
2018 * This should happen rarely.
2019 */
2020static int rescuer_thread(void *__wq)
2021{
2022 struct workqueue_struct *wq = __wq;
2023 struct worker *rescuer = wq->rescuer;
2024 struct list_head *scheduled = &rescuer->scheduled;
Tejun Heof3421792010-07-02 10:03:51 +02002025 bool is_unbound = wq->flags & WQ_UNBOUND;
Tejun Heoe22bee72010-06-29 10:07:14 +02002026 unsigned int cpu;
2027
2028 set_user_nice(current, RESCUER_NICE_LEVEL);
2029repeat:
2030 set_current_state(TASK_INTERRUPTIBLE);
2031
2032 if (kthread_should_stop())
2033 return 0;
2034
Tejun Heof3421792010-07-02 10:03:51 +02002035 /*
2036 * See whether any cpu is asking for help. Unbounded
2037 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2038 */
Tejun Heof2e005a2010-07-20 15:59:09 +02002039 for_each_mayday_cpu(cpu, wq->mayday_mask) {
Tejun Heof3421792010-07-02 10:03:51 +02002040 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2041 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
Tejun Heoe22bee72010-06-29 10:07:14 +02002042 struct global_cwq *gcwq = cwq->gcwq;
2043 struct work_struct *work, *n;
2044
2045 __set_current_state(TASK_RUNNING);
Tejun Heof2e005a2010-07-20 15:59:09 +02002046 mayday_clear_cpu(cpu, wq->mayday_mask);
Tejun Heoe22bee72010-06-29 10:07:14 +02002047
2048 /* migrate to the target cpu if possible */
2049 rescuer->gcwq = gcwq;
2050 worker_maybe_bind_and_lock(rescuer);
2051
2052 /*
2053 * Slurp in all works issued via this workqueue and
2054 * process'em.
2055 */
2056 BUG_ON(!list_empty(&rescuer->scheduled));
2057 list_for_each_entry_safe(work, n, &gcwq->worklist, entry)
2058 if (get_work_cwq(work) == cwq)
2059 move_linked_works(work, scheduled, &n);
2060
2061 process_scheduled_works(rescuer);
Tejun Heo75769582011-02-14 14:04:46 +01002062
2063 /*
2064 * Leave this gcwq. If keep_working() is %true, notify a
2065 * regular worker; otherwise, we end up with 0 concurrency
2066 * and stalling the execution.
2067 */
2068 if (keep_working(gcwq))
2069 wake_up_worker(gcwq);
2070
Tejun Heoe22bee72010-06-29 10:07:14 +02002071 spin_unlock_irq(&gcwq->lock);
2072 }
2073
2074 schedule();
2075 goto repeat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076}
2077
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002078struct wq_barrier {
2079 struct work_struct work;
2080 struct completion done;
2081};
2082
2083static void wq_barrier_func(struct work_struct *work)
2084{
2085 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2086 complete(&barr->done);
2087}
2088
Tejun Heo4690c4a2010-06-29 10:07:10 +02002089/**
2090 * insert_wq_barrier - insert a barrier work
2091 * @cwq: cwq to insert barrier into
2092 * @barr: wq_barrier to insert
Tejun Heoaffee4b2010-06-29 10:07:12 +02002093 * @target: target work to attach @barr to
2094 * @worker: worker currently executing @target, NULL if @target is not executing
Tejun Heo4690c4a2010-06-29 10:07:10 +02002095 *
Tejun Heoaffee4b2010-06-29 10:07:12 +02002096 * @barr is linked to @target such that @barr is completed only after
2097 * @target finishes execution. Please note that the ordering
2098 * guarantee is observed only with respect to @target and on the local
2099 * cpu.
2100 *
2101 * Currently, a queued barrier can't be canceled. This is because
2102 * try_to_grab_pending() can't determine whether the work to be
2103 * grabbed is at the head of the queue and thus can't clear LINKED
2104 * flag of the previous work while there must be a valid next work
2105 * after a work with LINKED flag set.
2106 *
2107 * Note that when @worker is non-NULL, @target may be modified
2108 * underneath us, so we can't reliably determine cwq from @target.
Tejun Heo4690c4a2010-06-29 10:07:10 +02002109 *
2110 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02002111 * spin_lock_irq(gcwq->lock).
Tejun Heo4690c4a2010-06-29 10:07:10 +02002112 */
Oleg Nesterov83c22522007-05-09 02:33:54 -07002113static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
Tejun Heoaffee4b2010-06-29 10:07:12 +02002114 struct wq_barrier *barr,
2115 struct work_struct *target, struct worker *worker)
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002116{
Tejun Heoaffee4b2010-06-29 10:07:12 +02002117 struct list_head *head;
2118 unsigned int linked = 0;
2119
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002120 /*
Tejun Heo8b03ae32010-06-29 10:07:12 +02002121 * debugobject calls are safe here even with gcwq->lock locked
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002122 * as we know for sure that this will not trigger any of the
2123 * checks and call back into the fixup functions where we
2124 * might deadlock.
2125 */
Andrew Mortonca1cab32010-10-26 14:22:34 -07002126 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
Tejun Heo22df02b2010-06-29 10:07:10 +02002127 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002128 init_completion(&barr->done);
Oleg Nesterov83c22522007-05-09 02:33:54 -07002129
Tejun Heoaffee4b2010-06-29 10:07:12 +02002130 /*
2131 * If @target is currently being executed, schedule the
2132 * barrier to the worker; otherwise, put it after @target.
2133 */
2134 if (worker)
2135 head = worker->scheduled.next;
2136 else {
2137 unsigned long *bits = work_data_bits(target);
2138
2139 head = target->entry.next;
2140 /* there can already be other linked works, inherit and set */
2141 linked = *bits & WORK_STRUCT_LINKED;
2142 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2143 }
2144
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002145 debug_work_activate(&barr->work);
Tejun Heoaffee4b2010-06-29 10:07:12 +02002146 insert_work(cwq, &barr->work, head,
2147 work_color_to_flags(WORK_NO_COLOR) | linked);
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002148}
2149
Tejun Heo73f53c42010-06-29 10:07:11 +02002150/**
2151 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2152 * @wq: workqueue being flushed
2153 * @flush_color: new flush color, < 0 for no-op
2154 * @work_color: new work color, < 0 for no-op
2155 *
2156 * Prepare cwqs for workqueue flushing.
2157 *
2158 * If @flush_color is non-negative, flush_color on all cwqs should be
2159 * -1. If no cwq has in-flight commands at the specified color, all
2160 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2161 * has in flight commands, its cwq->flush_color is set to
2162 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2163 * wakeup logic is armed and %true is returned.
2164 *
2165 * The caller should have initialized @wq->first_flusher prior to
2166 * calling this function with non-negative @flush_color. If
2167 * @flush_color is negative, no flush color update is done and %false
2168 * is returned.
2169 *
2170 * If @work_color is non-negative, all cwqs should have the same
2171 * work_color which is previous to @work_color and all will be
2172 * advanced to @work_color.
2173 *
2174 * CONTEXT:
2175 * mutex_lock(wq->flush_mutex).
2176 *
2177 * RETURNS:
2178 * %true if @flush_color >= 0 and there's something to flush. %false
2179 * otherwise.
2180 */
2181static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2182 int flush_color, int work_color)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183{
Tejun Heo73f53c42010-06-29 10:07:11 +02002184 bool wait = false;
2185 unsigned int cpu;
Oleg Nesterov14441962007-05-23 13:57:57 -07002186
Tejun Heo73f53c42010-06-29 10:07:11 +02002187 if (flush_color >= 0) {
2188 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2189 atomic_set(&wq->nr_cwqs_to_flush, 1);
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002190 }
Oleg Nesterov14441962007-05-23 13:57:57 -07002191
Tejun Heof3421792010-07-02 10:03:51 +02002192 for_each_cwq_cpu(cpu, wq) {
Tejun Heo73f53c42010-06-29 10:07:11 +02002193 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
Tejun Heo8b03ae32010-06-29 10:07:12 +02002194 struct global_cwq *gcwq = cwq->gcwq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002195
Tejun Heo8b03ae32010-06-29 10:07:12 +02002196 spin_lock_irq(&gcwq->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002197
2198 if (flush_color >= 0) {
2199 BUG_ON(cwq->flush_color != -1);
2200
2201 if (cwq->nr_in_flight[flush_color]) {
2202 cwq->flush_color = flush_color;
2203 atomic_inc(&wq->nr_cwqs_to_flush);
2204 wait = true;
2205 }
2206 }
2207
2208 if (work_color >= 0) {
2209 BUG_ON(work_color != work_next_color(cwq->work_color));
2210 cwq->work_color = work_color;
2211 }
2212
Tejun Heo8b03ae32010-06-29 10:07:12 +02002213 spin_unlock_irq(&gcwq->lock);
Tejun Heo73f53c42010-06-29 10:07:11 +02002214 }
2215
2216 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2217 complete(&wq->first_flusher->done);
2218
2219 return wait;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220}
2221
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002222/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 * flush_workqueue - ensure that any scheduled work has run to completion.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002224 * @wq: workqueue to flush
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 *
2226 * Forces execution of the workqueue and blocks until its completion.
2227 * This is typically used in driver shutdown handlers.
2228 *
Oleg Nesterovfc2e4d72007-05-09 02:33:51 -07002229 * We sleep until all works which were queued on entry have been handled,
2230 * but we are not livelocked by new incoming ones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002232void flush_workqueue(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233{
Tejun Heo73f53c42010-06-29 10:07:11 +02002234 struct wq_flusher this_flusher = {
2235 .list = LIST_HEAD_INIT(this_flusher.list),
2236 .flush_color = -1,
2237 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2238 };
2239 int next_color;
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07002240
Ingo Molnar3295f0e2008-08-11 10:30:30 +02002241 lock_map_acquire(&wq->lockdep_map);
2242 lock_map_release(&wq->lockdep_map);
Tejun Heo73f53c42010-06-29 10:07:11 +02002243
2244 mutex_lock(&wq->flush_mutex);
2245
2246 /*
2247 * Start-to-wait phase
2248 */
2249 next_color = work_next_color(wq->work_color);
2250
2251 if (next_color != wq->flush_color) {
2252 /*
2253 * Color space is not full. The current work_color
2254 * becomes our flush_color and work_color is advanced
2255 * by one.
2256 */
2257 BUG_ON(!list_empty(&wq->flusher_overflow));
2258 this_flusher.flush_color = wq->work_color;
2259 wq->work_color = next_color;
2260
2261 if (!wq->first_flusher) {
2262 /* no flush in progress, become the first flusher */
2263 BUG_ON(wq->flush_color != this_flusher.flush_color);
2264
2265 wq->first_flusher = &this_flusher;
2266
2267 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2268 wq->work_color)) {
2269 /* nothing to flush, done */
2270 wq->flush_color = next_color;
2271 wq->first_flusher = NULL;
2272 goto out_unlock;
2273 }
2274 } else {
2275 /* wait in queue */
2276 BUG_ON(wq->flush_color == this_flusher.flush_color);
2277 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2278 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2279 }
2280 } else {
2281 /*
2282 * Oops, color space is full, wait on overflow queue.
2283 * The next flush completion will assign us
2284 * flush_color and transfer to flusher_queue.
2285 */
2286 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2287 }
2288
2289 mutex_unlock(&wq->flush_mutex);
2290
2291 wait_for_completion(&this_flusher.done);
2292
2293 /*
2294 * Wake-up-and-cascade phase
2295 *
2296 * First flushers are responsible for cascading flushes and
2297 * handling overflow. Non-first flushers can simply return.
2298 */
2299 if (wq->first_flusher != &this_flusher)
2300 return;
2301
2302 mutex_lock(&wq->flush_mutex);
2303
Tejun Heo4ce48b32010-07-02 10:03:51 +02002304 /* we might have raced, check again with mutex held */
2305 if (wq->first_flusher != &this_flusher)
2306 goto out_unlock;
2307
Tejun Heo73f53c42010-06-29 10:07:11 +02002308 wq->first_flusher = NULL;
2309
2310 BUG_ON(!list_empty(&this_flusher.list));
2311 BUG_ON(wq->flush_color != this_flusher.flush_color);
2312
2313 while (true) {
2314 struct wq_flusher *next, *tmp;
2315
2316 /* complete all the flushers sharing the current flush color */
2317 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2318 if (next->flush_color != wq->flush_color)
2319 break;
2320 list_del_init(&next->list);
2321 complete(&next->done);
2322 }
2323
2324 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2325 wq->flush_color != work_next_color(wq->work_color));
2326
2327 /* this flush_color is finished, advance by one */
2328 wq->flush_color = work_next_color(wq->flush_color);
2329
2330 /* one color has been freed, handle overflow queue */
2331 if (!list_empty(&wq->flusher_overflow)) {
2332 /*
2333 * Assign the same color to all overflowed
2334 * flushers, advance work_color and append to
2335 * flusher_queue. This is the start-to-wait
2336 * phase for these overflowed flushers.
2337 */
2338 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2339 tmp->flush_color = wq->work_color;
2340
2341 wq->work_color = work_next_color(wq->work_color);
2342
2343 list_splice_tail_init(&wq->flusher_overflow,
2344 &wq->flusher_queue);
2345 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2346 }
2347
2348 if (list_empty(&wq->flusher_queue)) {
2349 BUG_ON(wq->flush_color != wq->work_color);
2350 break;
2351 }
2352
2353 /*
2354 * Need to flush more colors. Make the next flusher
2355 * the new first flusher and arm cwqs.
2356 */
2357 BUG_ON(wq->flush_color == wq->work_color);
2358 BUG_ON(wq->flush_color != next->flush_color);
2359
2360 list_del_init(&next->list);
2361 wq->first_flusher = next;
2362
2363 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2364 break;
2365
2366 /*
2367 * Meh... this color is already done, clear first
2368 * flusher and repeat cascading.
2369 */
2370 wq->first_flusher = NULL;
2371 }
2372
2373out_unlock:
2374 mutex_unlock(&wq->flush_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375}
Dave Jonesae90dd52006-06-30 01:40:45 -04002376EXPORT_SYMBOL_GPL(flush_workqueue);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377
Tejun Heobaf59022010-09-16 10:42:16 +02002378static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2379 bool wait_executing)
2380{
2381 struct worker *worker = NULL;
2382 struct global_cwq *gcwq;
2383 struct cpu_workqueue_struct *cwq;
2384
2385 might_sleep();
2386 gcwq = get_work_gcwq(work);
2387 if (!gcwq)
2388 return false;
2389
2390 spin_lock_irq(&gcwq->lock);
2391 if (!list_empty(&work->entry)) {
2392 /*
2393 * See the comment near try_to_grab_pending()->smp_rmb().
2394 * If it was re-queued to a different gcwq under us, we
2395 * are not going to wait.
2396 */
2397 smp_rmb();
2398 cwq = get_work_cwq(work);
2399 if (unlikely(!cwq || gcwq != cwq->gcwq))
2400 goto already_gone;
2401 } else if (wait_executing) {
2402 worker = find_worker_executing_work(gcwq, work);
2403 if (!worker)
2404 goto already_gone;
2405 cwq = worker->current_cwq;
2406 } else
2407 goto already_gone;
2408
2409 insert_wq_barrier(cwq, barr, work, worker);
2410 spin_unlock_irq(&gcwq->lock);
2411
Tejun Heoe1594892011-01-09 23:32:15 +01002412 /*
2413 * If @max_active is 1 or rescuer is in use, flushing another work
2414 * item on the same workqueue may lead to deadlock. Make sure the
2415 * flusher is not running on the same workqueue by verifying write
2416 * access.
2417 */
2418 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2419 lock_map_acquire(&cwq->wq->lockdep_map);
2420 else
2421 lock_map_acquire_read(&cwq->wq->lockdep_map);
Tejun Heobaf59022010-09-16 10:42:16 +02002422 lock_map_release(&cwq->wq->lockdep_map);
Tejun Heoe1594892011-01-09 23:32:15 +01002423
Tejun Heobaf59022010-09-16 10:42:16 +02002424 return true;
2425already_gone:
2426 spin_unlock_irq(&gcwq->lock);
2427 return false;
2428}
2429
Oleg Nesterovdb700892008-07-25 01:47:49 -07002430/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002431 * flush_work - wait for a work to finish executing the last queueing instance
2432 * @work: the work to flush
Oleg Nesterovdb700892008-07-25 01:47:49 -07002433 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002434 * Wait until @work has finished execution. This function considers
2435 * only the last queueing instance of @work. If @work has been
2436 * enqueued across different CPUs on a non-reentrant workqueue or on
2437 * multiple workqueues, @work might still be executing on return on
2438 * some of the CPUs from earlier queueing.
Oleg Nesterova67da702008-07-25 01:47:52 -07002439 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002440 * If @work was queued only on a non-reentrant, ordered or unbound
2441 * workqueue, @work is guaranteed to be idle on return if it hasn't
2442 * been requeued since flush started.
2443 *
2444 * RETURNS:
2445 * %true if flush_work() waited for the work to finish execution,
2446 * %false if it was already idle.
Oleg Nesterovdb700892008-07-25 01:47:49 -07002447 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002448bool flush_work(struct work_struct *work)
Oleg Nesterovdb700892008-07-25 01:47:49 -07002449{
Oleg Nesterovdb700892008-07-25 01:47:49 -07002450 struct wq_barrier barr;
2451
Tejun Heobaf59022010-09-16 10:42:16 +02002452 if (start_flush_work(work, &barr, true)) {
2453 wait_for_completion(&barr.done);
2454 destroy_work_on_stack(&barr.work);
2455 return true;
2456 } else
2457 return false;
Oleg Nesterovdb700892008-07-25 01:47:49 -07002458}
2459EXPORT_SYMBOL_GPL(flush_work);
2460
Tejun Heo401a8d02010-09-16 10:36:00 +02002461static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2462{
2463 struct wq_barrier barr;
2464 struct worker *worker;
2465
2466 spin_lock_irq(&gcwq->lock);
2467
2468 worker = find_worker_executing_work(gcwq, work);
2469 if (unlikely(worker))
2470 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2471
2472 spin_unlock_irq(&gcwq->lock);
2473
2474 if (unlikely(worker)) {
2475 wait_for_completion(&barr.done);
2476 destroy_work_on_stack(&barr.work);
2477 return true;
2478 } else
2479 return false;
2480}
2481
2482static bool wait_on_work(struct work_struct *work)
2483{
2484 bool ret = false;
2485 int cpu;
2486
2487 might_sleep();
2488
2489 lock_map_acquire(&work->lockdep_map);
2490 lock_map_release(&work->lockdep_map);
2491
2492 for_each_gcwq_cpu(cpu)
2493 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2494 return ret;
2495}
2496
Tejun Heo09383492010-09-16 10:48:29 +02002497/**
2498 * flush_work_sync - wait until a work has finished execution
2499 * @work: the work to flush
2500 *
2501 * Wait until @work has finished execution. On return, it's
2502 * guaranteed that all queueing instances of @work which happened
2503 * before this function is called are finished. In other words, if
2504 * @work hasn't been requeued since this function was called, @work is
2505 * guaranteed to be idle on return.
2506 *
2507 * RETURNS:
2508 * %true if flush_work_sync() waited for the work to finish execution,
2509 * %false if it was already idle.
2510 */
2511bool flush_work_sync(struct work_struct *work)
2512{
2513 struct wq_barrier barr;
2514 bool pending, waited;
2515
2516 /* we'll wait for executions separately, queue barr only if pending */
2517 pending = start_flush_work(work, &barr, false);
2518
2519 /* wait for executions to finish */
2520 waited = wait_on_work(work);
2521
2522 /* wait for the pending one */
2523 if (pending) {
2524 wait_for_completion(&barr.done);
2525 destroy_work_on_stack(&barr.work);
2526 }
2527
2528 return pending || waited;
2529}
2530EXPORT_SYMBOL_GPL(flush_work_sync);
2531
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002532/*
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002533 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002534 * so this work can't be re-armed in any way.
2535 */
2536static int try_to_grab_pending(struct work_struct *work)
2537{
Tejun Heo8b03ae32010-06-29 10:07:12 +02002538 struct global_cwq *gcwq;
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002539 int ret = -1;
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002540
Tejun Heo22df02b2010-06-29 10:07:10 +02002541 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002542 return 0;
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002543
2544 /*
2545 * The queueing is in progress, or it is already queued. Try to
2546 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
2547 */
Tejun Heo7a22ad72010-06-29 10:07:13 +02002548 gcwq = get_work_gcwq(work);
2549 if (!gcwq)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002550 return ret;
2551
Tejun Heo8b03ae32010-06-29 10:07:12 +02002552 spin_lock_irq(&gcwq->lock);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002553 if (!list_empty(&work->entry)) {
2554 /*
Tejun Heo7a22ad72010-06-29 10:07:13 +02002555 * This work is queued, but perhaps we locked the wrong gcwq.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002556 * In that case we must see the new value after rmb(), see
2557 * insert_work()->wmb().
2558 */
2559 smp_rmb();
Tejun Heo7a22ad72010-06-29 10:07:13 +02002560 if (gcwq == get_work_gcwq(work)) {
Thomas Gleixnerdc186ad2009-11-16 01:09:48 +09002561 debug_work_deactivate(work);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002562 list_del_init(&work->entry);
Tejun Heo7a22ad72010-06-29 10:07:13 +02002563 cwq_dec_nr_in_flight(get_work_cwq(work),
Tejun Heo8a2e8e5d2010-08-25 10:33:56 +02002564 get_work_color(work),
2565 *work_data_bits(work) & WORK_STRUCT_DELAYED);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002566 ret = 1;
2567 }
2568 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02002569 spin_unlock_irq(&gcwq->lock);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002570
2571 return ret;
2572}
2573
Tejun Heo401a8d02010-09-16 10:36:00 +02002574static bool __cancel_work_timer(struct work_struct *work,
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002575 struct timer_list* timer)
2576{
2577 int ret;
2578
2579 do {
2580 ret = (timer && likely(del_timer(timer)));
2581 if (!ret)
2582 ret = try_to_grab_pending(work);
2583 wait_on_work(work);
2584 } while (unlikely(ret < 0));
2585
Tejun Heo7a22ad72010-06-29 10:07:13 +02002586 clear_work_data(work);
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002587 return ret;
2588}
2589
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002590/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002591 * cancel_work_sync - cancel a work and wait for it to finish
2592 * @work: the work to cancel
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002593 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002594 * Cancel @work and wait for its execution to finish. This function
2595 * can be used even if the work re-queues itself or migrates to
2596 * another workqueue. On return from this function, @work is
2597 * guaranteed to be not pending or executing on any CPU.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002598 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002599 * cancel_work_sync(&delayed_work->work) must not be used for
2600 * delayed_work's. Use cancel_delayed_work_sync() instead.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002601 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002602 * The caller must ensure that the workqueue on which @work was last
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002603 * queued can't be destroyed before this function returns.
Tejun Heo401a8d02010-09-16 10:36:00 +02002604 *
2605 * RETURNS:
2606 * %true if @work was pending, %false otherwise.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002607 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002608bool cancel_work_sync(struct work_struct *work)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002609{
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002610 return __cancel_work_timer(work, NULL);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002611}
Oleg Nesterov28e53bd2007-05-09 02:34:22 -07002612EXPORT_SYMBOL_GPL(cancel_work_sync);
Oleg Nesterovb89deed2007-05-09 02:33:52 -07002613
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002614/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002615 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2616 * @dwork: the delayed work to flush
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002617 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002618 * Delayed timer is cancelled and the pending work is queued for
2619 * immediate execution. Like flush_work(), this function only
2620 * considers the last queueing instance of @dwork.
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002621 *
Tejun Heo401a8d02010-09-16 10:36:00 +02002622 * RETURNS:
2623 * %true if flush_work() waited for the work to finish execution,
2624 * %false if it was already idle.
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002625 */
Tejun Heo401a8d02010-09-16 10:36:00 +02002626bool flush_delayed_work(struct delayed_work *dwork)
2627{
2628 if (del_timer_sync(&dwork->timer))
2629 __queue_work(raw_smp_processor_id(),
2630 get_work_cwq(&dwork->work)->wq, &dwork->work);
2631 return flush_work(&dwork->work);
2632}
2633EXPORT_SYMBOL(flush_delayed_work);
2634
2635/**
Tejun Heo09383492010-09-16 10:48:29 +02002636 * flush_delayed_work_sync - wait for a dwork to finish
2637 * @dwork: the delayed work to flush
2638 *
2639 * Delayed timer is cancelled and the pending work is queued for
2640 * execution immediately. Other than timer handling, its behavior
2641 * is identical to flush_work_sync().
2642 *
2643 * RETURNS:
2644 * %true if flush_work_sync() waited for the work to finish execution,
2645 * %false if it was already idle.
2646 */
2647bool flush_delayed_work_sync(struct delayed_work *dwork)
2648{
2649 if (del_timer_sync(&dwork->timer))
2650 __queue_work(raw_smp_processor_id(),
2651 get_work_cwq(&dwork->work)->wq, &dwork->work);
2652 return flush_work_sync(&dwork->work);
2653}
2654EXPORT_SYMBOL(flush_delayed_work_sync);
2655
2656/**
Tejun Heo401a8d02010-09-16 10:36:00 +02002657 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2658 * @dwork: the delayed work cancel
2659 *
2660 * This is cancel_work_sync() for delayed works.
2661 *
2662 * RETURNS:
2663 * %true if @dwork was pending, %false otherwise.
2664 */
2665bool cancel_delayed_work_sync(struct delayed_work *dwork)
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002666{
Oleg Nesterov1f1f6422007-07-15 23:41:44 -07002667 return __cancel_work_timer(&dwork->work, &dwork->timer);
Oleg Nesterov6e84d642007-05-09 02:34:46 -07002668}
Oleg Nesterovf5a421a2007-07-15 23:41:44 -07002669EXPORT_SYMBOL(cancel_delayed_work_sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002671/**
2672 * schedule_work - put work task in global workqueue
2673 * @work: job to be done
2674 *
Bart Van Assche5b0f437d2009-07-30 19:00:53 +02002675 * Returns zero if @work was already on the kernel-global workqueue and
2676 * non-zero otherwise.
2677 *
2678 * This puts a job in the kernel-global workqueue if it was not already
2679 * queued and leaves it in the same position on the kernel-global
2680 * workqueue otherwise.
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002681 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002682int schedule_work(struct work_struct *work)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683{
Tejun Heod320c032010-06-29 10:07:14 +02002684 return queue_work(system_wq, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685}
Dave Jonesae90dd52006-06-30 01:40:45 -04002686EXPORT_SYMBOL(schedule_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
Zhang Ruic1a220e2008-07-23 21:28:39 -07002688/*
2689 * schedule_work_on - put work task on a specific cpu
2690 * @cpu: cpu to put the work task on
2691 * @work: job to be done
2692 *
2693 * This puts a job on a specific cpu
2694 */
2695int schedule_work_on(int cpu, struct work_struct *work)
2696{
Tejun Heod320c032010-06-29 10:07:14 +02002697 return queue_work_on(cpu, system_wq, work);
Zhang Ruic1a220e2008-07-23 21:28:39 -07002698}
2699EXPORT_SYMBOL(schedule_work_on);
2700
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002701/**
2702 * schedule_delayed_work - put work task in global workqueue after delay
David Howells52bad642006-11-22 14:54:01 +00002703 * @dwork: job to be done
2704 * @delay: number of jiffies to wait or 0 for immediate execution
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002705 *
2706 * After waiting for a given time this puts a job in the kernel-global
2707 * workqueue.
2708 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002709int schedule_delayed_work(struct delayed_work *dwork,
Ingo Molnar82f67cd2007-02-16 01:28:13 -08002710 unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711{
Tejun Heod320c032010-06-29 10:07:14 +02002712 return queue_delayed_work(system_wq, dwork, delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713}
Dave Jonesae90dd52006-06-30 01:40:45 -04002714EXPORT_SYMBOL(schedule_delayed_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002716/**
2717 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2718 * @cpu: cpu to use
David Howells52bad642006-11-22 14:54:01 +00002719 * @dwork: job to be done
Rolf Eike Beer0fcb78c2006-07-30 03:03:42 -07002720 * @delay: number of jiffies to wait
2721 *
2722 * After waiting for a given time this puts a job in the kernel-global
2723 * workqueue on the specified CPU.
2724 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725int schedule_delayed_work_on(int cpu,
David Howells52bad642006-11-22 14:54:01 +00002726 struct delayed_work *dwork, unsigned long delay)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727{
Tejun Heod320c032010-06-29 10:07:14 +02002728 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729}
Dave Jonesae90dd52006-06-30 01:40:45 -04002730EXPORT_SYMBOL(schedule_delayed_work_on);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731
Andrew Mortonb6136772006-06-25 05:47:49 -07002732/**
Tejun Heo31ddd872010-10-19 11:14:49 +02002733 * schedule_on_each_cpu - execute a function synchronously on each online CPU
Andrew Mortonb6136772006-06-25 05:47:49 -07002734 * @func: the function to call
Andrew Mortonb6136772006-06-25 05:47:49 -07002735 *
Tejun Heo31ddd872010-10-19 11:14:49 +02002736 * schedule_on_each_cpu() executes @func on each online CPU using the
2737 * system workqueue and blocks until all CPUs have completed.
Andrew Mortonb6136772006-06-25 05:47:49 -07002738 * schedule_on_each_cpu() is very slow.
Tejun Heo31ddd872010-10-19 11:14:49 +02002739 *
2740 * RETURNS:
2741 * 0 on success, -errno on failure.
Andrew Mortonb6136772006-06-25 05:47:49 -07002742 */
David Howells65f27f32006-11-22 14:55:48 +00002743int schedule_on_each_cpu(work_func_t func)
Christoph Lameter15316ba2006-01-08 01:00:43 -08002744{
2745 int cpu;
Namhyung Kim38f51562010-08-08 14:24:09 +02002746 struct work_struct __percpu *works;
Christoph Lameter15316ba2006-01-08 01:00:43 -08002747
Andrew Mortonb6136772006-06-25 05:47:49 -07002748 works = alloc_percpu(struct work_struct);
2749 if (!works)
Christoph Lameter15316ba2006-01-08 01:00:43 -08002750 return -ENOMEM;
Andrew Mortonb6136772006-06-25 05:47:49 -07002751
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002752 get_online_cpus();
Tejun Heo93981802009-11-17 14:06:20 -08002753
Christoph Lameter15316ba2006-01-08 01:00:43 -08002754 for_each_online_cpu(cpu) {
Ingo Molnar9bfb1832006-12-18 20:05:09 +01002755 struct work_struct *work = per_cpu_ptr(works, cpu);
2756
2757 INIT_WORK(work, func);
Tejun Heob71ab8c2010-06-29 10:07:14 +02002758 schedule_work_on(cpu, work);
Andi Kleen65a64462009-10-14 06:22:47 +02002759 }
Tejun Heo93981802009-11-17 14:06:20 -08002760
2761 for_each_online_cpu(cpu)
2762 flush_work(per_cpu_ptr(works, cpu));
2763
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002764 put_online_cpus();
Andrew Mortonb6136772006-06-25 05:47:49 -07002765 free_percpu(works);
Christoph Lameter15316ba2006-01-08 01:00:43 -08002766 return 0;
2767}
2768
Alan Sterneef6a7d2010-02-12 17:39:21 +09002769/**
2770 * flush_scheduled_work - ensure that any scheduled work has run to completion.
2771 *
2772 * Forces execution of the kernel-global workqueue and blocks until its
2773 * completion.
2774 *
2775 * Think twice before calling this function! It's very easy to get into
2776 * trouble if you don't take great care. Either of the following situations
2777 * will lead to deadlock:
2778 *
2779 * One of the work items currently on the workqueue needs to acquire
2780 * a lock held by your code or its caller.
2781 *
2782 * Your code is running in the context of a work routine.
2783 *
2784 * They will be detected by lockdep when they occur, but the first might not
2785 * occur very often. It depends on what work items are on the workqueue and
2786 * what locks they need, which you have no control over.
2787 *
2788 * In most situations flushing the entire workqueue is overkill; you merely
2789 * need to know that a particular work item isn't queued and isn't running.
2790 * In such cases you should use cancel_delayed_work_sync() or
2791 * cancel_work_sync() instead.
2792 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793void flush_scheduled_work(void)
2794{
Tejun Heod320c032010-06-29 10:07:14 +02002795 flush_workqueue(system_wq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796}
Dave Jonesae90dd52006-06-30 01:40:45 -04002797EXPORT_SYMBOL(flush_scheduled_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798
2799/**
James Bottomley1fa44ec2006-02-23 12:43:43 -06002800 * execute_in_process_context - reliably execute the routine with user context
2801 * @fn: the function to execute
James Bottomley1fa44ec2006-02-23 12:43:43 -06002802 * @ew: guaranteed storage for the execute work structure (must
2803 * be available when the work executes)
2804 *
2805 * Executes the function immediately if process context is available,
2806 * otherwise schedules the function for delayed execution.
2807 *
2808 * Returns: 0 - function was executed
2809 * 1 - function was scheduled for execution
2810 */
David Howells65f27f32006-11-22 14:55:48 +00002811int execute_in_process_context(work_func_t fn, struct execute_work *ew)
James Bottomley1fa44ec2006-02-23 12:43:43 -06002812{
2813 if (!in_interrupt()) {
David Howells65f27f32006-11-22 14:55:48 +00002814 fn(&ew->work);
James Bottomley1fa44ec2006-02-23 12:43:43 -06002815 return 0;
2816 }
2817
David Howells65f27f32006-11-22 14:55:48 +00002818 INIT_WORK(&ew->work, fn);
James Bottomley1fa44ec2006-02-23 12:43:43 -06002819 schedule_work(&ew->work);
2820
2821 return 1;
2822}
2823EXPORT_SYMBOL_GPL(execute_in_process_context);
2824
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825int keventd_up(void)
2826{
Tejun Heod320c032010-06-29 10:07:14 +02002827 return system_wq != NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828}
2829
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002830static int alloc_cwqs(struct workqueue_struct *wq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831{
Oleg Nesterov3af244332007-05-09 02:34:09 -07002832 /*
Tejun Heo0f900042010-06-29 10:07:11 +02002833 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
2834 * Make sure that the alignment isn't lower than that of
2835 * unsigned long long.
Oleg Nesterov3af244332007-05-09 02:34:09 -07002836 */
Tejun Heo0f900042010-06-29 10:07:11 +02002837 const size_t size = sizeof(struct cpu_workqueue_struct);
2838 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
2839 __alignof__(unsigned long long));
Tejun Heo931ac772010-07-20 11:07:48 +02002840#ifdef CONFIG_SMP
2841 bool percpu = !(wq->flags & WQ_UNBOUND);
2842#else
2843 bool percpu = false;
2844#endif
Oleg Nesterov3af244332007-05-09 02:34:09 -07002845
Tejun Heo931ac772010-07-20 11:07:48 +02002846 if (percpu)
Tejun Heof3421792010-07-02 10:03:51 +02002847 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
Tejun Heo931ac772010-07-20 11:07:48 +02002848 else {
Tejun Heof3421792010-07-02 10:03:51 +02002849 void *ptr;
Frederic Weisbeckere1d8aa92009-01-12 23:15:46 +01002850
Tejun Heof3421792010-07-02 10:03:51 +02002851 /*
2852 * Allocate enough room to align cwq and put an extra
2853 * pointer at the end pointing back to the originally
2854 * allocated pointer which will be used for free.
2855 */
2856 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
2857 if (ptr) {
2858 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
2859 *(void **)(wq->cpu_wq.single + 1) = ptr;
2860 }
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002861 }
Tejun Heof3421792010-07-02 10:03:51 +02002862
David Howells52605622010-10-25 23:41:11 +01002863 /* just in case, make sure it's actually aligned
2864 * - this is affected by PERCPU() alignment in vmlinux.lds.S
2865 */
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002866 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
2867 return wq->cpu_wq.v ? 0 : -ENOMEM;
Oleg Nesterov3af244332007-05-09 02:34:09 -07002868}
2869
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002870static void free_cwqs(struct workqueue_struct *wq)
Oleg Nesterov06ba38a2007-05-09 02:34:15 -07002871{
Tejun Heo931ac772010-07-20 11:07:48 +02002872#ifdef CONFIG_SMP
2873 bool percpu = !(wq->flags & WQ_UNBOUND);
2874#else
2875 bool percpu = false;
2876#endif
Oleg Nesterov06ba38a2007-05-09 02:34:15 -07002877
Tejun Heo931ac772010-07-20 11:07:48 +02002878 if (percpu)
Tejun Heof3421792010-07-02 10:03:51 +02002879 free_percpu(wq->cpu_wq.pcpu);
2880 else if (wq->cpu_wq.single) {
2881 /* the pointer to free is stored right after the cwq */
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002882 kfree(*(void **)(wq->cpu_wq.single + 1));
Oleg Nesterov06ba38a2007-05-09 02:34:15 -07002883 }
2884}
2885
Tejun Heof3421792010-07-02 10:03:51 +02002886static int wq_clamp_max_active(int max_active, unsigned int flags,
2887 const char *name)
Tejun Heob71ab8c2010-06-29 10:07:14 +02002888{
Tejun Heof3421792010-07-02 10:03:51 +02002889 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
2890
2891 if (max_active < 1 || max_active > lim)
Tejun Heob71ab8c2010-06-29 10:07:14 +02002892 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
2893 "is out of range, clamping between %d and %d\n",
Tejun Heof3421792010-07-02 10:03:51 +02002894 max_active, name, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02002895
Tejun Heof3421792010-07-02 10:03:51 +02002896 return clamp_val(max_active, 1, lim);
Tejun Heob71ab8c2010-06-29 10:07:14 +02002897}
2898
Tejun Heod320c032010-06-29 10:07:14 +02002899struct workqueue_struct *__alloc_workqueue_key(const char *name,
2900 unsigned int flags,
2901 int max_active,
2902 struct lock_class_key *key,
2903 const char *lock_name)
Oleg Nesterov3af244332007-05-09 02:34:09 -07002904{
2905 struct workqueue_struct *wq;
Tejun Heoc34056a2010-06-29 10:07:11 +02002906 unsigned int cpu;
Oleg Nesterov3af244332007-05-09 02:34:09 -07002907
Tejun Heof3421792010-07-02 10:03:51 +02002908 /*
Tejun Heo6370a6a2010-10-11 15:12:27 +02002909 * Workqueues which may be used during memory reclaim should
2910 * have a rescuer to guarantee forward progress.
2911 */
2912 if (flags & WQ_MEM_RECLAIM)
2913 flags |= WQ_RESCUER;
2914
2915 /*
Tejun Heof3421792010-07-02 10:03:51 +02002916 * Unbound workqueues aren't concurrency managed and should be
2917 * dispatched to workers immediately.
2918 */
2919 if (flags & WQ_UNBOUND)
2920 flags |= WQ_HIGHPRI;
2921
Tejun Heod320c032010-06-29 10:07:14 +02002922 max_active = max_active ?: WQ_DFL_ACTIVE;
Tejun Heof3421792010-07-02 10:03:51 +02002923 max_active = wq_clamp_max_active(max_active, flags, name);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002924
2925 wq = kzalloc(sizeof(*wq), GFP_KERNEL);
2926 if (!wq)
Tejun Heo4690c4a2010-06-29 10:07:10 +02002927 goto err;
Oleg Nesterov3af244332007-05-09 02:34:09 -07002928
Tejun Heo97e37d72010-06-29 10:07:10 +02002929 wq->flags = flags;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02002930 wq->saved_max_active = max_active;
Tejun Heo73f53c42010-06-29 10:07:11 +02002931 mutex_init(&wq->flush_mutex);
2932 atomic_set(&wq->nr_cwqs_to_flush, 0);
2933 INIT_LIST_HEAD(&wq->flusher_queue);
2934 INIT_LIST_HEAD(&wq->flusher_overflow);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002935
2936 wq->name = name;
Johannes Bergeb13ba82008-01-16 09:51:58 +01002937 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
Oleg Nesterovcce1a162007-05-09 02:34:13 -07002938 INIT_LIST_HEAD(&wq->list);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002939
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002940 if (alloc_cwqs(wq) < 0)
2941 goto err;
2942
Tejun Heof3421792010-07-02 10:03:51 +02002943 for_each_cwq_cpu(cpu, wq) {
Tejun Heo15376632010-06-29 10:07:11 +02002944 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
Tejun Heo8b03ae32010-06-29 10:07:12 +02002945 struct global_cwq *gcwq = get_gcwq(cpu);
Tejun Heo15376632010-06-29 10:07:11 +02002946
Tejun Heo0f900042010-06-29 10:07:11 +02002947 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
Tejun Heo8b03ae32010-06-29 10:07:12 +02002948 cwq->gcwq = gcwq;
Tejun Heoc34056a2010-06-29 10:07:11 +02002949 cwq->wq = wq;
Tejun Heo73f53c42010-06-29 10:07:11 +02002950 cwq->flush_color = -1;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02002951 cwq->max_active = max_active;
Tejun Heo1e19ffc2010-06-29 10:07:12 +02002952 INIT_LIST_HEAD(&cwq->delayed_works);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002953 }
2954
Tejun Heoe22bee72010-06-29 10:07:14 +02002955 if (flags & WQ_RESCUER) {
2956 struct worker *rescuer;
2957
Tejun Heof2e005a2010-07-20 15:59:09 +02002958 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
Tejun Heoe22bee72010-06-29 10:07:14 +02002959 goto err;
2960
2961 wq->rescuer = rescuer = alloc_worker();
2962 if (!rescuer)
2963 goto err;
2964
2965 rescuer->task = kthread_create(rescuer_thread, wq, "%s", name);
2966 if (IS_ERR(rescuer->task))
2967 goto err;
2968
Tejun Heoe22bee72010-06-29 10:07:14 +02002969 rescuer->task->flags |= PF_THREAD_BOUND;
2970 wake_up_process(rescuer->task);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002971 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07002972
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02002973 /*
2974 * workqueue_lock protects global freeze state and workqueues
2975 * list. Grab it, set max_active accordingly and add the new
2976 * workqueue to workqueues list.
2977 */
Tejun Heo15376632010-06-29 10:07:11 +02002978 spin_lock(&workqueue_lock);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02002979
Tejun Heo58a69cb2011-02-16 09:25:31 +01002980 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
Tejun Heof3421792010-07-02 10:03:51 +02002981 for_each_cwq_cpu(cpu, wq)
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02002982 get_cwq(cpu, wq)->max_active = 0;
2983
Tejun Heo15376632010-06-29 10:07:11 +02002984 list_add(&wq->list, &workqueues);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02002985
Tejun Heo15376632010-06-29 10:07:11 +02002986 spin_unlock(&workqueue_lock);
2987
Oleg Nesterov3af244332007-05-09 02:34:09 -07002988 return wq;
Tejun Heo4690c4a2010-06-29 10:07:10 +02002989err:
2990 if (wq) {
Tejun Heobdbc5dd2010-07-02 10:03:51 +02002991 free_cwqs(wq);
Tejun Heof2e005a2010-07-20 15:59:09 +02002992 free_mayday_mask(wq->mayday_mask);
Tejun Heoe22bee72010-06-29 10:07:14 +02002993 kfree(wq->rescuer);
Tejun Heo4690c4a2010-06-29 10:07:10 +02002994 kfree(wq);
2995 }
2996 return NULL;
Oleg Nesterov3af244332007-05-09 02:34:09 -07002997}
Tejun Heod320c032010-06-29 10:07:14 +02002998EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
Oleg Nesterov3af244332007-05-09 02:34:09 -07002999
3000/**
3001 * destroy_workqueue - safely terminate a workqueue
3002 * @wq: target workqueue
3003 *
3004 * Safely destroy a workqueue. All work currently pending will be done first.
3005 */
3006void destroy_workqueue(struct workqueue_struct *wq)
3007{
Tejun Heoc8efcc22010-12-20 19:32:04 +01003008 unsigned int flush_cnt = 0;
Tejun Heoc8e55f32010-06-29 10:07:12 +02003009 unsigned int cpu;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003010
Tejun Heoc8efcc22010-12-20 19:32:04 +01003011 /*
3012 * Mark @wq dying and drain all pending works. Once WQ_DYING is
3013 * set, only chain queueing is allowed. IOW, only currently
3014 * pending or running work items on @wq can queue further work
3015 * items on it. @wq is flushed repeatedly until it becomes empty.
3016 * The number of flushing is detemined by the depth of chaining and
3017 * should be relatively short. Whine if it takes too long.
3018 */
Tejun Heoe41e7042010-08-24 14:22:47 +02003019 wq->flags |= WQ_DYING;
Tejun Heoc8efcc22010-12-20 19:32:04 +01003020reflush:
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003021 flush_workqueue(wq);
3022
Tejun Heoc8efcc22010-12-20 19:32:04 +01003023 for_each_cwq_cpu(cpu, wq) {
3024 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3025
3026 if (!cwq->nr_active && list_empty(&cwq->delayed_works))
3027 continue;
3028
3029 if (++flush_cnt == 10 ||
3030 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
3031 printk(KERN_WARNING "workqueue %s: flush on "
3032 "destruction isn't complete after %u tries\n",
3033 wq->name, flush_cnt);
3034 goto reflush;
3035 }
3036
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003037 /*
3038 * wq list is used to freeze wq, remove from list after
3039 * flushing is complete in case freeze races us.
3040 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003041 spin_lock(&workqueue_lock);
Oleg Nesterovb1f4ec12007-05-09 02:34:12 -07003042 list_del(&wq->list);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01003043 spin_unlock(&workqueue_lock);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003044
Tejun Heoe22bee72010-06-29 10:07:14 +02003045 /* sanity check */
Tejun Heof3421792010-07-02 10:03:51 +02003046 for_each_cwq_cpu(cpu, wq) {
Tejun Heo73f53c42010-06-29 10:07:11 +02003047 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3048 int i;
Oleg Nesterov3af244332007-05-09 02:34:09 -07003049
Tejun Heo73f53c42010-06-29 10:07:11 +02003050 for (i = 0; i < WORK_NR_COLORS; i++)
3051 BUG_ON(cwq->nr_in_flight[i]);
Tejun Heo1e19ffc2010-06-29 10:07:12 +02003052 BUG_ON(cwq->nr_active);
3053 BUG_ON(!list_empty(&cwq->delayed_works));
Tejun Heo73f53c42010-06-29 10:07:11 +02003054 }
Oleg Nesterov3af244332007-05-09 02:34:09 -07003055
Tejun Heoe22bee72010-06-29 10:07:14 +02003056 if (wq->flags & WQ_RESCUER) {
3057 kthread_stop(wq->rescuer->task);
Tejun Heof2e005a2010-07-20 15:59:09 +02003058 free_mayday_mask(wq->mayday_mask);
Xiaotian Feng8d9df9f2010-08-16 09:54:28 +02003059 kfree(wq->rescuer);
Tejun Heoe22bee72010-06-29 10:07:14 +02003060 }
3061
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003062 free_cwqs(wq);
Oleg Nesterov3af244332007-05-09 02:34:09 -07003063 kfree(wq);
3064}
3065EXPORT_SYMBOL_GPL(destroy_workqueue);
3066
Tejun Heodcd989c2010-06-29 10:07:14 +02003067/**
3068 * workqueue_set_max_active - adjust max_active of a workqueue
3069 * @wq: target workqueue
3070 * @max_active: new max_active value.
3071 *
3072 * Set max_active of @wq to @max_active.
3073 *
3074 * CONTEXT:
3075 * Don't call from IRQ context.
3076 */
3077void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3078{
3079 unsigned int cpu;
3080
Tejun Heof3421792010-07-02 10:03:51 +02003081 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
Tejun Heodcd989c2010-06-29 10:07:14 +02003082
3083 spin_lock(&workqueue_lock);
3084
3085 wq->saved_max_active = max_active;
3086
Tejun Heof3421792010-07-02 10:03:51 +02003087 for_each_cwq_cpu(cpu, wq) {
Tejun Heodcd989c2010-06-29 10:07:14 +02003088 struct global_cwq *gcwq = get_gcwq(cpu);
3089
3090 spin_lock_irq(&gcwq->lock);
3091
Tejun Heo58a69cb2011-02-16 09:25:31 +01003092 if (!(wq->flags & WQ_FREEZABLE) ||
Tejun Heodcd989c2010-06-29 10:07:14 +02003093 !(gcwq->flags & GCWQ_FREEZING))
3094 get_cwq(gcwq->cpu, wq)->max_active = max_active;
3095
3096 spin_unlock_irq(&gcwq->lock);
3097 }
3098
3099 spin_unlock(&workqueue_lock);
3100}
3101EXPORT_SYMBOL_GPL(workqueue_set_max_active);
3102
3103/**
3104 * workqueue_congested - test whether a workqueue is congested
3105 * @cpu: CPU in question
3106 * @wq: target workqueue
3107 *
3108 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3109 * no synchronization around this function and the test result is
3110 * unreliable and only useful as advisory hints or for debugging.
3111 *
3112 * RETURNS:
3113 * %true if congested, %false otherwise.
3114 */
3115bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
3116{
3117 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3118
3119 return !list_empty(&cwq->delayed_works);
3120}
3121EXPORT_SYMBOL_GPL(workqueue_congested);
3122
3123/**
3124 * work_cpu - return the last known associated cpu for @work
3125 * @work: the work of interest
3126 *
3127 * RETURNS:
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003128 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
Tejun Heodcd989c2010-06-29 10:07:14 +02003129 */
3130unsigned int work_cpu(struct work_struct *work)
3131{
3132 struct global_cwq *gcwq = get_work_gcwq(work);
3133
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003134 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
Tejun Heodcd989c2010-06-29 10:07:14 +02003135}
3136EXPORT_SYMBOL_GPL(work_cpu);
3137
3138/**
3139 * work_busy - test whether a work is currently pending or running
3140 * @work: the work to be tested
3141 *
3142 * Test whether @work is currently pending or running. There is no
3143 * synchronization around this function and the test result is
3144 * unreliable and only useful as advisory hints or for debugging.
3145 * Especially for reentrant wqs, the pending state might hide the
3146 * running state.
3147 *
3148 * RETURNS:
3149 * OR'd bitmask of WORK_BUSY_* bits.
3150 */
3151unsigned int work_busy(struct work_struct *work)
3152{
3153 struct global_cwq *gcwq = get_work_gcwq(work);
3154 unsigned long flags;
3155 unsigned int ret = 0;
3156
3157 if (!gcwq)
3158 return false;
3159
3160 spin_lock_irqsave(&gcwq->lock, flags);
3161
3162 if (work_pending(work))
3163 ret |= WORK_BUSY_PENDING;
3164 if (find_worker_executing_work(gcwq, work))
3165 ret |= WORK_BUSY_RUNNING;
3166
3167 spin_unlock_irqrestore(&gcwq->lock, flags);
3168
3169 return ret;
3170}
3171EXPORT_SYMBOL_GPL(work_busy);
3172
Tejun Heodb7bccf2010-06-29 10:07:12 +02003173/*
3174 * CPU hotplug.
3175 *
Tejun Heoe22bee72010-06-29 10:07:14 +02003176 * There are two challenges in supporting CPU hotplug. Firstly, there
3177 * are a lot of assumptions on strong associations among work, cwq and
3178 * gcwq which make migrating pending and scheduled works very
3179 * difficult to implement without impacting hot paths. Secondly,
3180 * gcwqs serve mix of short, long and very long running works making
3181 * blocked draining impractical.
3182 *
3183 * This is solved by allowing a gcwq to be detached from CPU, running
3184 * it with unbound (rogue) workers and allowing it to be reattached
3185 * later if the cpu comes back online. A separate thread is created
3186 * to govern a gcwq in such state and is called the trustee of the
3187 * gcwq.
Tejun Heodb7bccf2010-06-29 10:07:12 +02003188 *
3189 * Trustee states and their descriptions.
3190 *
3191 * START Command state used on startup. On CPU_DOWN_PREPARE, a
3192 * new trustee is started with this state.
3193 *
3194 * IN_CHARGE Once started, trustee will enter this state after
Tejun Heoe22bee72010-06-29 10:07:14 +02003195 * assuming the manager role and making all existing
3196 * workers rogue. DOWN_PREPARE waits for trustee to
3197 * enter this state. After reaching IN_CHARGE, trustee
3198 * tries to execute the pending worklist until it's empty
3199 * and the state is set to BUTCHER, or the state is set
3200 * to RELEASE.
Tejun Heodb7bccf2010-06-29 10:07:12 +02003201 *
3202 * BUTCHER Command state which is set by the cpu callback after
3203 * the cpu has went down. Once this state is set trustee
3204 * knows that there will be no new works on the worklist
3205 * and once the worklist is empty it can proceed to
3206 * killing idle workers.
3207 *
3208 * RELEASE Command state which is set by the cpu callback if the
3209 * cpu down has been canceled or it has come online
3210 * again. After recognizing this state, trustee stops
Tejun Heoe22bee72010-06-29 10:07:14 +02003211 * trying to drain or butcher and clears ROGUE, rebinds
3212 * all remaining workers back to the cpu and releases
3213 * manager role.
Tejun Heodb7bccf2010-06-29 10:07:12 +02003214 *
3215 * DONE Trustee will enter this state after BUTCHER or RELEASE
3216 * is complete.
3217 *
3218 * trustee CPU draining
3219 * took over down complete
3220 * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE
3221 * | | ^
3222 * | CPU is back online v return workers |
3223 * ----------------> RELEASE --------------
3224 */
3225
3226/**
3227 * trustee_wait_event_timeout - timed event wait for trustee
3228 * @cond: condition to wait for
3229 * @timeout: timeout in jiffies
3230 *
3231 * wait_event_timeout() for trustee to use. Handles locking and
3232 * checks for RELEASE request.
3233 *
3234 * CONTEXT:
3235 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3236 * multiple times. To be used by trustee.
3237 *
3238 * RETURNS:
3239 * Positive indicating left time if @cond is satisfied, 0 if timed
3240 * out, -1 if canceled.
3241 */
3242#define trustee_wait_event_timeout(cond, timeout) ({ \
3243 long __ret = (timeout); \
3244 while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \
3245 __ret) { \
3246 spin_unlock_irq(&gcwq->lock); \
3247 __wait_event_timeout(gcwq->trustee_wait, (cond) || \
3248 (gcwq->trustee_state == TRUSTEE_RELEASE), \
3249 __ret); \
3250 spin_lock_irq(&gcwq->lock); \
3251 } \
3252 gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \
3253})
3254
3255/**
3256 * trustee_wait_event - event wait for trustee
3257 * @cond: condition to wait for
3258 *
3259 * wait_event() for trustee to use. Automatically handles locking and
3260 * checks for CANCEL request.
3261 *
3262 * CONTEXT:
3263 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3264 * multiple times. To be used by trustee.
3265 *
3266 * RETURNS:
3267 * 0 if @cond is satisfied, -1 if canceled.
3268 */
3269#define trustee_wait_event(cond) ({ \
3270 long __ret1; \
3271 __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\
3272 __ret1 < 0 ? -1 : 0; \
3273})
3274
3275static int __cpuinit trustee_thread(void *__gcwq)
3276{
3277 struct global_cwq *gcwq = __gcwq;
3278 struct worker *worker;
Tejun Heoe22bee72010-06-29 10:07:14 +02003279 struct work_struct *work;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003280 struct hlist_node *pos;
Tejun Heoe22bee72010-06-29 10:07:14 +02003281 long rc;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003282 int i;
3283
3284 BUG_ON(gcwq->cpu != smp_processor_id());
3285
3286 spin_lock_irq(&gcwq->lock);
3287 /*
Tejun Heoe22bee72010-06-29 10:07:14 +02003288 * Claim the manager position and make all workers rogue.
3289 * Trustee must be bound to the target cpu and can't be
3290 * cancelled.
Tejun Heodb7bccf2010-06-29 10:07:12 +02003291 */
3292 BUG_ON(gcwq->cpu != smp_processor_id());
Tejun Heoe22bee72010-06-29 10:07:14 +02003293 rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS));
3294 BUG_ON(rc < 0);
3295
3296 gcwq->flags |= GCWQ_MANAGING_WORKERS;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003297
3298 list_for_each_entry(worker, &gcwq->idle_list, entry)
Tejun Heocb444762010-07-02 10:03:50 +02003299 worker->flags |= WORKER_ROGUE;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003300
3301 for_each_busy_worker(worker, i, pos, gcwq)
Tejun Heocb444762010-07-02 10:03:50 +02003302 worker->flags |= WORKER_ROGUE;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003303
3304 /*
Tejun Heoe22bee72010-06-29 10:07:14 +02003305 * Call schedule() so that we cross rq->lock and thus can
3306 * guarantee sched callbacks see the rogue flag. This is
3307 * necessary as scheduler callbacks may be invoked from other
3308 * cpus.
3309 */
3310 spin_unlock_irq(&gcwq->lock);
3311 schedule();
3312 spin_lock_irq(&gcwq->lock);
3313
3314 /*
Tejun Heocb444762010-07-02 10:03:50 +02003315 * Sched callbacks are disabled now. Zap nr_running. After
3316 * this, nr_running stays zero and need_more_worker() and
3317 * keep_working() are always true as long as the worklist is
3318 * not empty.
Tejun Heoe22bee72010-06-29 10:07:14 +02003319 */
Tejun Heocb444762010-07-02 10:03:50 +02003320 atomic_set(get_gcwq_nr_running(gcwq->cpu), 0);
Tejun Heoe22bee72010-06-29 10:07:14 +02003321
3322 spin_unlock_irq(&gcwq->lock);
3323 del_timer_sync(&gcwq->idle_timer);
3324 spin_lock_irq(&gcwq->lock);
3325
3326 /*
Tejun Heodb7bccf2010-06-29 10:07:12 +02003327 * We're now in charge. Notify and proceed to drain. We need
3328 * to keep the gcwq running during the whole CPU down
3329 * procedure as other cpu hotunplug callbacks may need to
3330 * flush currently running tasks.
3331 */
3332 gcwq->trustee_state = TRUSTEE_IN_CHARGE;
3333 wake_up_all(&gcwq->trustee_wait);
3334
3335 /*
3336 * The original cpu is in the process of dying and may go away
3337 * anytime now. When that happens, we and all workers would
Tejun Heoe22bee72010-06-29 10:07:14 +02003338 * be migrated to other cpus. Try draining any left work. We
3339 * want to get it over with ASAP - spam rescuers, wake up as
3340 * many idlers as necessary and create new ones till the
3341 * worklist is empty. Note that if the gcwq is frozen, there
Tejun Heo58a69cb2011-02-16 09:25:31 +01003342 * may be frozen works in freezable cwqs. Don't declare
Tejun Heoe22bee72010-06-29 10:07:14 +02003343 * completion while frozen.
Tejun Heodb7bccf2010-06-29 10:07:12 +02003344 */
3345 while (gcwq->nr_workers != gcwq->nr_idle ||
3346 gcwq->flags & GCWQ_FREEZING ||
3347 gcwq->trustee_state == TRUSTEE_IN_CHARGE) {
Tejun Heoe22bee72010-06-29 10:07:14 +02003348 int nr_works = 0;
3349
3350 list_for_each_entry(work, &gcwq->worklist, entry) {
3351 send_mayday(work);
3352 nr_works++;
3353 }
3354
3355 list_for_each_entry(worker, &gcwq->idle_list, entry) {
3356 if (!nr_works--)
3357 break;
3358 wake_up_process(worker->task);
3359 }
3360
3361 if (need_to_create_worker(gcwq)) {
3362 spin_unlock_irq(&gcwq->lock);
3363 worker = create_worker(gcwq, false);
3364 spin_lock_irq(&gcwq->lock);
3365 if (worker) {
Tejun Heocb444762010-07-02 10:03:50 +02003366 worker->flags |= WORKER_ROGUE;
Tejun Heoe22bee72010-06-29 10:07:14 +02003367 start_worker(worker);
3368 }
3369 }
3370
Tejun Heodb7bccf2010-06-29 10:07:12 +02003371 /* give a breather */
3372 if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0)
3373 break;
3374 }
3375
Tejun Heoe22bee72010-06-29 10:07:14 +02003376 /*
3377 * Either all works have been scheduled and cpu is down, or
3378 * cpu down has already been canceled. Wait for and butcher
3379 * all workers till we're canceled.
3380 */
3381 do {
3382 rc = trustee_wait_event(!list_empty(&gcwq->idle_list));
3383 while (!list_empty(&gcwq->idle_list))
3384 destroy_worker(list_first_entry(&gcwq->idle_list,
3385 struct worker, entry));
3386 } while (gcwq->nr_workers && rc >= 0);
3387
3388 /*
3389 * At this point, either draining has completed and no worker
3390 * is left, or cpu down has been canceled or the cpu is being
3391 * brought back up. There shouldn't be any idle one left.
3392 * Tell the remaining busy ones to rebind once it finishes the
3393 * currently scheduled works by scheduling the rebind_work.
3394 */
3395 WARN_ON(!list_empty(&gcwq->idle_list));
3396
3397 for_each_busy_worker(worker, i, pos, gcwq) {
3398 struct work_struct *rebind_work = &worker->rebind_work;
3399
3400 /*
3401 * Rebind_work may race with future cpu hotplug
3402 * operations. Use a separate flag to mark that
3403 * rebinding is scheduled.
3404 */
Tejun Heocb444762010-07-02 10:03:50 +02003405 worker->flags |= WORKER_REBIND;
3406 worker->flags &= ~WORKER_ROGUE;
Tejun Heoe22bee72010-06-29 10:07:14 +02003407
3408 /* queue rebind_work, wq doesn't matter, use the default one */
3409 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
3410 work_data_bits(rebind_work)))
3411 continue;
3412
3413 debug_work_activate(rebind_work);
Tejun Heod320c032010-06-29 10:07:14 +02003414 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
Tejun Heoe22bee72010-06-29 10:07:14 +02003415 worker->scheduled.next,
3416 work_color_to_flags(WORK_NO_COLOR));
3417 }
3418
3419 /* relinquish manager role */
3420 gcwq->flags &= ~GCWQ_MANAGING_WORKERS;
3421
Tejun Heodb7bccf2010-06-29 10:07:12 +02003422 /* notify completion */
3423 gcwq->trustee = NULL;
3424 gcwq->trustee_state = TRUSTEE_DONE;
3425 wake_up_all(&gcwq->trustee_wait);
3426 spin_unlock_irq(&gcwq->lock);
3427 return 0;
3428}
3429
3430/**
3431 * wait_trustee_state - wait for trustee to enter the specified state
3432 * @gcwq: gcwq the trustee of interest belongs to
3433 * @state: target state to wait for
3434 *
3435 * Wait for the trustee to reach @state. DONE is already matched.
3436 *
3437 * CONTEXT:
3438 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
3439 * multiple times. To be used by cpu_callback.
3440 */
3441static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state)
Namhyung Kim06bd6eb2010-08-22 23:19:42 +09003442__releases(&gcwq->lock)
3443__acquires(&gcwq->lock)
Tejun Heodb7bccf2010-06-29 10:07:12 +02003444{
3445 if (!(gcwq->trustee_state == state ||
3446 gcwq->trustee_state == TRUSTEE_DONE)) {
3447 spin_unlock_irq(&gcwq->lock);
3448 __wait_event(gcwq->trustee_wait,
3449 gcwq->trustee_state == state ||
3450 gcwq->trustee_state == TRUSTEE_DONE);
3451 spin_lock_irq(&gcwq->lock);
3452 }
3453}
3454
Oleg Nesterov3af244332007-05-09 02:34:09 -07003455static int __devinit workqueue_cpu_callback(struct notifier_block *nfb,
3456 unsigned long action,
3457 void *hcpu)
3458{
3459 unsigned int cpu = (unsigned long)hcpu;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003460 struct global_cwq *gcwq = get_gcwq(cpu);
3461 struct task_struct *new_trustee = NULL;
Tejun Heoe22bee72010-06-29 10:07:14 +02003462 struct worker *uninitialized_var(new_worker);
Tejun Heodb7bccf2010-06-29 10:07:12 +02003463 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003465 action &= ~CPU_TASKS_FROZEN;
3466
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 switch (action) {
Tejun Heodb7bccf2010-06-29 10:07:12 +02003468 case CPU_DOWN_PREPARE:
3469 new_trustee = kthread_create(trustee_thread, gcwq,
3470 "workqueue_trustee/%d\n", cpu);
3471 if (IS_ERR(new_trustee))
3472 return notifier_from_errno(PTR_ERR(new_trustee));
3473 kthread_bind(new_trustee, cpu);
Tejun Heoe22bee72010-06-29 10:07:14 +02003474 /* fall through */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 case CPU_UP_PREPARE:
Tejun Heoe22bee72010-06-29 10:07:14 +02003476 BUG_ON(gcwq->first_idle);
3477 new_worker = create_worker(gcwq, false);
3478 if (!new_worker) {
3479 if (new_trustee)
3480 kthread_stop(new_trustee);
3481 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 }
3484
Tejun Heodb7bccf2010-06-29 10:07:12 +02003485 /* some are called w/ irq disabled, don't disturb irq status */
3486 spin_lock_irqsave(&gcwq->lock, flags);
3487
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07003488 switch (action) {
Tejun Heodb7bccf2010-06-29 10:07:12 +02003489 case CPU_DOWN_PREPARE:
3490 /* initialize trustee and tell it to acquire the gcwq */
3491 BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE);
3492 gcwq->trustee = new_trustee;
3493 gcwq->trustee_state = TRUSTEE_START;
3494 wake_up_process(gcwq->trustee);
3495 wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE);
Tejun Heoe22bee72010-06-29 10:07:14 +02003496 /* fall through */
3497 case CPU_UP_PREPARE:
3498 BUG_ON(gcwq->first_idle);
3499 gcwq->first_idle = new_worker;
3500 break;
3501
3502 case CPU_DYING:
3503 /*
3504 * Before this, the trustee and all workers except for
3505 * the ones which are still executing works from
3506 * before the last CPU down must be on the cpu. After
3507 * this, they'll all be diasporas.
3508 */
3509 gcwq->flags |= GCWQ_DISASSOCIATED;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003510 break;
3511
Oleg Nesterov3da1c842008-07-25 01:47:50 -07003512 case CPU_POST_DEAD:
Tejun Heodb7bccf2010-06-29 10:07:12 +02003513 gcwq->trustee_state = TRUSTEE_BUTCHER;
Tejun Heoe22bee72010-06-29 10:07:14 +02003514 /* fall through */
3515 case CPU_UP_CANCELED:
3516 destroy_worker(gcwq->first_idle);
3517 gcwq->first_idle = NULL;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003518 break;
3519
3520 case CPU_DOWN_FAILED:
3521 case CPU_ONLINE:
Tejun Heoe22bee72010-06-29 10:07:14 +02003522 gcwq->flags &= ~GCWQ_DISASSOCIATED;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003523 if (gcwq->trustee_state != TRUSTEE_DONE) {
3524 gcwq->trustee_state = TRUSTEE_RELEASE;
3525 wake_up_process(gcwq->trustee);
3526 wait_trustee_state(gcwq, TRUSTEE_DONE);
3527 }
3528
Tejun Heoe22bee72010-06-29 10:07:14 +02003529 /*
3530 * Trustee is done and there might be no worker left.
3531 * Put the first_idle in and request a real manager to
3532 * take a look.
3533 */
3534 spin_unlock_irq(&gcwq->lock);
3535 kthread_bind(gcwq->first_idle->task, cpu);
3536 spin_lock_irq(&gcwq->lock);
3537 gcwq->flags |= GCWQ_MANAGE_WORKERS;
3538 start_worker(gcwq->first_idle);
3539 gcwq->first_idle = NULL;
Tejun Heodb7bccf2010-06-29 10:07:12 +02003540 break;
Oleg Nesterov00dfcaf2008-04-29 01:00:27 -07003541 }
3542
Tejun Heodb7bccf2010-06-29 10:07:12 +02003543 spin_unlock_irqrestore(&gcwq->lock, flags);
3544
Tejun Heo15376632010-06-29 10:07:11 +02003545 return notifier_from_errno(0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547
Rusty Russell2d3854a2008-11-05 13:39:10 +11003548#ifdef CONFIG_SMP
Rusty Russell8ccad402009-01-16 15:31:15 -08003549
Rusty Russell2d3854a2008-11-05 13:39:10 +11003550struct work_for_cpu {
Andrew Morton6b44003e2009-04-09 09:50:37 -06003551 struct completion completion;
Rusty Russell2d3854a2008-11-05 13:39:10 +11003552 long (*fn)(void *);
3553 void *arg;
3554 long ret;
3555};
3556
Andrew Morton6b44003e2009-04-09 09:50:37 -06003557static int do_work_for_cpu(void *_wfc)
Rusty Russell2d3854a2008-11-05 13:39:10 +11003558{
Andrew Morton6b44003e2009-04-09 09:50:37 -06003559 struct work_for_cpu *wfc = _wfc;
Rusty Russell2d3854a2008-11-05 13:39:10 +11003560 wfc->ret = wfc->fn(wfc->arg);
Andrew Morton6b44003e2009-04-09 09:50:37 -06003561 complete(&wfc->completion);
3562 return 0;
Rusty Russell2d3854a2008-11-05 13:39:10 +11003563}
3564
3565/**
3566 * work_on_cpu - run a function in user context on a particular cpu
3567 * @cpu: the cpu to run on
3568 * @fn: the function to run
3569 * @arg: the function arg
3570 *
Rusty Russell31ad9082009-01-16 15:31:15 -08003571 * This will return the value @fn returns.
3572 * It is up to the caller to ensure that the cpu doesn't go offline.
Andrew Morton6b44003e2009-04-09 09:50:37 -06003573 * The caller must not hold any locks which would prevent @fn from completing.
Rusty Russell2d3854a2008-11-05 13:39:10 +11003574 */
3575long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3576{
Andrew Morton6b44003e2009-04-09 09:50:37 -06003577 struct task_struct *sub_thread;
3578 struct work_for_cpu wfc = {
3579 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3580 .fn = fn,
3581 .arg = arg,
3582 };
Rusty Russell2d3854a2008-11-05 13:39:10 +11003583
Andrew Morton6b44003e2009-04-09 09:50:37 -06003584 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3585 if (IS_ERR(sub_thread))
3586 return PTR_ERR(sub_thread);
3587 kthread_bind(sub_thread, cpu);
3588 wake_up_process(sub_thread);
3589 wait_for_completion(&wfc.completion);
Rusty Russell2d3854a2008-11-05 13:39:10 +11003590 return wfc.ret;
3591}
3592EXPORT_SYMBOL_GPL(work_on_cpu);
3593#endif /* CONFIG_SMP */
3594
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003595#ifdef CONFIG_FREEZER
Rusty Russelle7577c52009-01-01 10:12:25 +10303596
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003597/**
3598 * freeze_workqueues_begin - begin freezing workqueues
3599 *
Tejun Heo58a69cb2011-02-16 09:25:31 +01003600 * Start freezing workqueues. After this function returns, all freezable
3601 * workqueues will queue new works to their frozen_works list instead of
3602 * gcwq->worklist.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003603 *
3604 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02003605 * Grabs and releases workqueue_lock and gcwq->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003606 */
3607void freeze_workqueues_begin(void)
3608{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003609 unsigned int cpu;
3610
3611 spin_lock(&workqueue_lock);
3612
3613 BUG_ON(workqueue_freezing);
3614 workqueue_freezing = true;
3615
Tejun Heof3421792010-07-02 10:03:51 +02003616 for_each_gcwq_cpu(cpu) {
Tejun Heo8b03ae32010-06-29 10:07:12 +02003617 struct global_cwq *gcwq = get_gcwq(cpu);
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003618 struct workqueue_struct *wq;
Tejun Heo8b03ae32010-06-29 10:07:12 +02003619
3620 spin_lock_irq(&gcwq->lock);
3621
Tejun Heodb7bccf2010-06-29 10:07:12 +02003622 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3623 gcwq->flags |= GCWQ_FREEZING;
3624
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003625 list_for_each_entry(wq, &workqueues, list) {
3626 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3627
Tejun Heo58a69cb2011-02-16 09:25:31 +01003628 if (cwq && wq->flags & WQ_FREEZABLE)
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003629 cwq->max_active = 0;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003630 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02003631
3632 spin_unlock_irq(&gcwq->lock);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003633 }
3634
3635 spin_unlock(&workqueue_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636}
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003637
3638/**
Tejun Heo58a69cb2011-02-16 09:25:31 +01003639 * freeze_workqueues_busy - are freezable workqueues still busy?
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003640 *
3641 * Check whether freezing is complete. This function must be called
3642 * between freeze_workqueues_begin() and thaw_workqueues().
3643 *
3644 * CONTEXT:
3645 * Grabs and releases workqueue_lock.
3646 *
3647 * RETURNS:
Tejun Heo58a69cb2011-02-16 09:25:31 +01003648 * %true if some freezable workqueues are still busy. %false if freezing
3649 * is complete.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003650 */
3651bool freeze_workqueues_busy(void)
3652{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003653 unsigned int cpu;
3654 bool busy = false;
3655
3656 spin_lock(&workqueue_lock);
3657
3658 BUG_ON(!workqueue_freezing);
3659
Tejun Heof3421792010-07-02 10:03:51 +02003660 for_each_gcwq_cpu(cpu) {
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003661 struct workqueue_struct *wq;
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003662 /*
3663 * nr_active is monotonically decreasing. It's safe
3664 * to peek without lock.
3665 */
3666 list_for_each_entry(wq, &workqueues, list) {
3667 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3668
Tejun Heo58a69cb2011-02-16 09:25:31 +01003669 if (!cwq || !(wq->flags & WQ_FREEZABLE))
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003670 continue;
3671
3672 BUG_ON(cwq->nr_active < 0);
3673 if (cwq->nr_active) {
3674 busy = true;
3675 goto out_unlock;
3676 }
3677 }
3678 }
3679out_unlock:
3680 spin_unlock(&workqueue_lock);
3681 return busy;
3682}
3683
3684/**
3685 * thaw_workqueues - thaw workqueues
3686 *
3687 * Thaw workqueues. Normal queueing is restored and all collected
Tejun Heo7e116292010-06-29 10:07:13 +02003688 * frozen works are transferred to their respective gcwq worklists.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003689 *
3690 * CONTEXT:
Tejun Heo8b03ae32010-06-29 10:07:12 +02003691 * Grabs and releases workqueue_lock and gcwq->lock's.
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003692 */
3693void thaw_workqueues(void)
3694{
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003695 unsigned int cpu;
3696
3697 spin_lock(&workqueue_lock);
3698
3699 if (!workqueue_freezing)
3700 goto out_unlock;
3701
Tejun Heof3421792010-07-02 10:03:51 +02003702 for_each_gcwq_cpu(cpu) {
Tejun Heo8b03ae32010-06-29 10:07:12 +02003703 struct global_cwq *gcwq = get_gcwq(cpu);
Tejun Heobdbc5dd2010-07-02 10:03:51 +02003704 struct workqueue_struct *wq;
Tejun Heo8b03ae32010-06-29 10:07:12 +02003705
3706 spin_lock_irq(&gcwq->lock);
3707
Tejun Heodb7bccf2010-06-29 10:07:12 +02003708 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3709 gcwq->flags &= ~GCWQ_FREEZING;
3710
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003711 list_for_each_entry(wq, &workqueues, list) {
3712 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3713
Tejun Heo58a69cb2011-02-16 09:25:31 +01003714 if (!cwq || !(wq->flags & WQ_FREEZABLE))
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003715 continue;
3716
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003717 /* restore max_active and repopulate worklist */
3718 cwq->max_active = wq->saved_max_active;
3719
3720 while (!list_empty(&cwq->delayed_works) &&
3721 cwq->nr_active < cwq->max_active)
3722 cwq_activate_first_delayed(cwq);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003723 }
Tejun Heo8b03ae32010-06-29 10:07:12 +02003724
Tejun Heoe22bee72010-06-29 10:07:14 +02003725 wake_up_worker(gcwq);
3726
Tejun Heo8b03ae32010-06-29 10:07:12 +02003727 spin_unlock_irq(&gcwq->lock);
Tejun Heoa0a1a5f2010-06-29 10:07:12 +02003728 }
3729
3730 workqueue_freezing = false;
3731out_unlock:
3732 spin_unlock(&workqueue_lock);
3733}
3734#endif /* CONFIG_FREEZER */
3735
Suresh Siddha6ee05782010-07-30 14:57:37 -07003736static int __init init_workqueues(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737{
Tejun Heoc34056a2010-06-29 10:07:11 +02003738 unsigned int cpu;
Tejun Heoc8e55f32010-06-29 10:07:12 +02003739 int i;
Tejun Heoc34056a2010-06-29 10:07:11 +02003740
Tejun Heof6500942010-08-09 11:50:34 +02003741 cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE);
Tejun Heo8b03ae32010-06-29 10:07:12 +02003742
3743 /* initialize gcwqs */
Tejun Heof3421792010-07-02 10:03:51 +02003744 for_each_gcwq_cpu(cpu) {
Tejun Heo8b03ae32010-06-29 10:07:12 +02003745 struct global_cwq *gcwq = get_gcwq(cpu);
3746
3747 spin_lock_init(&gcwq->lock);
Tejun Heo7e116292010-06-29 10:07:13 +02003748 INIT_LIST_HEAD(&gcwq->worklist);
Tejun Heo8b03ae32010-06-29 10:07:12 +02003749 gcwq->cpu = cpu;
Tejun Heo477a3c32010-08-31 10:54:35 +02003750 gcwq->flags |= GCWQ_DISASSOCIATED;
Tejun Heo8b03ae32010-06-29 10:07:12 +02003751
Tejun Heoc8e55f32010-06-29 10:07:12 +02003752 INIT_LIST_HEAD(&gcwq->idle_list);
3753 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3754 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3755
Tejun Heoe22bee72010-06-29 10:07:14 +02003756 init_timer_deferrable(&gcwq->idle_timer);
3757 gcwq->idle_timer.function = idle_worker_timeout;
3758 gcwq->idle_timer.data = (unsigned long)gcwq;
3759
3760 setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout,
3761 (unsigned long)gcwq);
3762
Tejun Heo8b03ae32010-06-29 10:07:12 +02003763 ida_init(&gcwq->worker_ida);
Tejun Heodb7bccf2010-06-29 10:07:12 +02003764
3765 gcwq->trustee_state = TRUSTEE_DONE;
3766 init_waitqueue_head(&gcwq->trustee_wait);
Tejun Heo8b03ae32010-06-29 10:07:12 +02003767 }
3768
Tejun Heoe22bee72010-06-29 10:07:14 +02003769 /* create the initial worker */
Tejun Heof3421792010-07-02 10:03:51 +02003770 for_each_online_gcwq_cpu(cpu) {
Tejun Heoe22bee72010-06-29 10:07:14 +02003771 struct global_cwq *gcwq = get_gcwq(cpu);
3772 struct worker *worker;
3773
Tejun Heo477a3c32010-08-31 10:54:35 +02003774 if (cpu != WORK_CPU_UNBOUND)
3775 gcwq->flags &= ~GCWQ_DISASSOCIATED;
Tejun Heoe22bee72010-06-29 10:07:14 +02003776 worker = create_worker(gcwq, true);
3777 BUG_ON(!worker);
3778 spin_lock_irq(&gcwq->lock);
3779 start_worker(worker);
3780 spin_unlock_irq(&gcwq->lock);
3781 }
3782
Tejun Heod320c032010-06-29 10:07:14 +02003783 system_wq = alloc_workqueue("events", 0, 0);
3784 system_long_wq = alloc_workqueue("events_long", 0, 0);
3785 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
Tejun Heof3421792010-07-02 10:03:51 +02003786 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3787 WQ_UNBOUND_MAX_ACTIVE);
Tejun Heo24d51ad2011-02-21 09:52:50 +01003788 system_freezable_wq = alloc_workqueue("events_freezable",
3789 WQ_FREEZABLE, 0);
Hitoshi Mitakee5cba242010-11-26 12:06:44 +01003790 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
Tejun Heo24d51ad2011-02-21 09:52:50 +01003791 !system_unbound_wq || !system_freezable_wq);
Suresh Siddha6ee05782010-07-30 14:57:37 -07003792 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793}
Suresh Siddha6ee05782010-07-30 14:57:37 -07003794early_initcall(init_workqueues);