]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - kernel/workqueue.c
workqueue: factor out __queue_delayed_work() from queue_delayed_work_on()
[mirror_ubuntu-artful-kernel.git] / kernel / workqueue.c
CommitLineData
1da177e4 1/*
c54fce6e 2 * kernel/workqueue.c - generic async execution with shared worker pool
1da177e4 3 *
c54fce6e 4 * Copyright (C) 2002 Ingo Molnar
1da177e4 5 *
c54fce6e
TH
6 * Derived from the taskqueue/keventd code by:
7 * David Woodhouse <dwmw2@infradead.org>
8 * Andrew Morton
9 * Kai Petzke <wpp@marie.physik.tu-berlin.de>
10 * Theodore Ts'o <tytso@mit.edu>
1da177e4 11 *
c54fce6e 12 * Made to use alloc_percpu by Christoph Lameter.
1da177e4 13 *
c54fce6e
TH
14 * Copyright (C) 2010 SUSE Linux Products GmbH
15 * Copyright (C) 2010 Tejun Heo <tj@kernel.org>
89ada679 16 *
c54fce6e
TH
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.
1da177e4
LT
24 */
25
9984de1a 26#include <linux/export.h>
1da177e4
LT
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>
1fa44eca 37#include <linux/hardirq.h>
46934023 38#include <linux/mempolicy.h>
341a5958 39#include <linux/freezer.h>
d5abe669
PZ
40#include <linux/kallsyms.h>
41#include <linux/debug_locks.h>
4e6045f1 42#include <linux/lockdep.h>
c34056a3 43#include <linux/idr.h>
e22bee78
TH
44
45#include "workqueue_sched.h"
1da177e4 46
c8e55f36 47enum {
bc2ae0f5
TH
48 /*
49 * global_cwq flags
50 *
51 * A bound gcwq is either associated or disassociated with its CPU.
52 * While associated (!DISASSOCIATED), all workers are bound to the
53 * CPU and none has %WORKER_UNBOUND set and concurrency management
54 * is in effect.
55 *
56 * While DISASSOCIATED, the cpu may be offline and all workers have
57 * %WORKER_UNBOUND set and concurrency management disabled, and may
58 * be executing on any CPU. The gcwq behaves as an unbound one.
59 *
60 * Note that DISASSOCIATED can be flipped only while holding
61 * managership of all pools on the gcwq to avoid changing binding
62 * state while create_worker() is in progress.
63 */
11ebea50
TH
64 GCWQ_DISASSOCIATED = 1 << 0, /* cpu can't serve workers */
65 GCWQ_FREEZING = 1 << 1, /* freeze in progress */
66
67 /* pool flags */
68 POOL_MANAGE_WORKERS = 1 << 0, /* need to manage workers */
db7bccf4 69
c8e55f36
TH
70 /* worker flags */
71 WORKER_STARTED = 1 << 0, /* started */
72 WORKER_DIE = 1 << 1, /* die die die */
73 WORKER_IDLE = 1 << 2, /* is idle */
e22bee78 74 WORKER_PREP = 1 << 3, /* preparing to run works */
e22bee78 75 WORKER_REBIND = 1 << 5, /* mom is home, come back */
fb0e7beb 76 WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */
f3421797 77 WORKER_UNBOUND = 1 << 7, /* worker is unbound */
e22bee78 78
403c821d
TH
79 WORKER_NOT_RUNNING = WORKER_PREP | WORKER_REBIND | WORKER_UNBOUND |
80 WORKER_CPU_INTENSIVE,
db7bccf4 81
3270476a 82 NR_WORKER_POOLS = 2, /* # worker pools per gcwq */
4ce62e9e 83
c8e55f36
TH
84 BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */
85 BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER,
86 BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1,
db7bccf4 87
e22bee78
TH
88 MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */
89 IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */
90
3233cdbd
TH
91 MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2,
92 /* call for help after 10ms
93 (min two ticks) */
e22bee78
TH
94 MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */
95 CREATE_COOLDOWN = HZ, /* time to breath after fail */
e22bee78
TH
96
97 /*
98 * Rescue workers are used only on emergencies and shared by
99 * all cpus. Give -20.
100 */
101 RESCUER_NICE_LEVEL = -20,
3270476a 102 HIGHPRI_NICE_LEVEL = -20,
c8e55f36 103};
1da177e4
LT
104
105/*
4690c4ab
TH
106 * Structure fields follow one of the following exclusion rules.
107 *
e41e704b
TH
108 * I: Modifiable by initialization/destruction paths and read-only for
109 * everyone else.
4690c4ab 110 *
e22bee78
TH
111 * P: Preemption protected. Disabling preemption is enough and should
112 * only be modified and accessed from the local cpu.
113 *
8b03ae3c 114 * L: gcwq->lock protected. Access with gcwq->lock held.
4690c4ab 115 *
e22bee78
TH
116 * X: During normal operation, modification requires gcwq->lock and
117 * should be done only from local cpu. Either disabling preemption
118 * on local cpu or grabbing gcwq->lock is enough for read access.
f3421797 119 * If GCWQ_DISASSOCIATED is set, it's identical to L.
e22bee78 120 *
73f53c4a
TH
121 * F: wq->flush_mutex protected.
122 *
4690c4ab 123 * W: workqueue_lock protected.
1da177e4 124 */
1da177e4 125
8b03ae3c 126struct global_cwq;
bd7bdd43 127struct worker_pool;
25511a47 128struct idle_rebind;
1da177e4 129
e22bee78
TH
130/*
131 * The poor guys doing the actual heavy lifting. All on-duty workers
132 * are either serving the manager role, on idle list or on busy hash.
133 */
c34056a3 134struct worker {
c8e55f36
TH
135 /* on idle list while idle, on busy hash table while busy */
136 union {
137 struct list_head entry; /* L: while idle */
138 struct hlist_node hentry; /* L: while busy */
139 };
1da177e4 140
c34056a3 141 struct work_struct *current_work; /* L: work being processed */
8cca0eea 142 struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */
affee4b2 143 struct list_head scheduled; /* L: scheduled works */
c34056a3 144 struct task_struct *task; /* I: worker task */
bd7bdd43 145 struct worker_pool *pool; /* I: the associated pool */
e22bee78
TH
146 /* 64 bytes boundary on 64bit, 32 on 32bit */
147 unsigned long last_active; /* L: last active timestamp */
148 unsigned int flags; /* X: flags */
c34056a3 149 int id; /* I: worker id */
25511a47
TH
150
151 /* for rebinding worker to CPU */
152 struct idle_rebind *idle_rebind; /* L: for idle worker */
153 struct work_struct rebind_work; /* L: for busy worker */
c34056a3
TH
154};
155
bd7bdd43
TH
156struct worker_pool {
157 struct global_cwq *gcwq; /* I: the owning gcwq */
11ebea50 158 unsigned int flags; /* X: flags */
bd7bdd43
TH
159
160 struct list_head worklist; /* L: list of pending works */
161 int nr_workers; /* L: total number of workers */
162 int nr_idle; /* L: currently idle ones */
163
164 struct list_head idle_list; /* X: list of idle workers */
165 struct timer_list idle_timer; /* L: worker idle timeout */
166 struct timer_list mayday_timer; /* L: SOS timer for workers */
167
60373152 168 struct mutex manager_mutex; /* mutex manager should hold */
bd7bdd43 169 struct ida worker_ida; /* L: for worker IDs */
bd7bdd43
TH
170};
171
8b03ae3c 172/*
e22bee78
TH
173 * Global per-cpu workqueue. There's one and only one for each cpu
174 * and all works are queued and processed here regardless of their
175 * target workqueues.
8b03ae3c
TH
176 */
177struct global_cwq {
178 spinlock_t lock; /* the gcwq lock */
179 unsigned int cpu; /* I: the associated cpu */
db7bccf4 180 unsigned int flags; /* L: GCWQ_* flags */
c8e55f36 181
bd7bdd43 182 /* workers are chained either in busy_hash or pool idle_list */
c8e55f36
TH
183 struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE];
184 /* L: hash of busy workers */
185
3270476a 186 struct worker_pool pools[2]; /* normal and highpri pools */
db7bccf4 187
25511a47 188 wait_queue_head_t rebind_hold; /* rebind hold wait */
8b03ae3c
TH
189} ____cacheline_aligned_in_smp;
190
1da177e4 191/*
502ca9d8 192 * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of
0f900049
TH
193 * work_struct->data are used for flags and thus cwqs need to be
194 * aligned at two's power of the number of flag bits.
1da177e4
LT
195 */
196struct cpu_workqueue_struct {
bd7bdd43 197 struct worker_pool *pool; /* I: the associated pool */
4690c4ab 198 struct workqueue_struct *wq; /* I: the owning workqueue */
73f53c4a
TH
199 int work_color; /* L: current color */
200 int flush_color; /* L: flushing color */
201 int nr_in_flight[WORK_NR_COLORS];
202 /* L: nr of in_flight works */
1e19ffc6 203 int nr_active; /* L: nr of active works */
a0a1a5fd 204 int max_active; /* L: max active works */
1e19ffc6 205 struct list_head delayed_works; /* L: delayed works */
0f900049 206};
1da177e4 207
73f53c4a
TH
208/*
209 * Structure used to wait for workqueue flush.
210 */
211struct wq_flusher {
212 struct list_head list; /* F: list of flushers */
213 int flush_color; /* F: flush color waiting for */
214 struct completion done; /* flush completion */
215};
216
f2e005aa
TH
217/*
218 * All cpumasks are assumed to be always set on UP and thus can't be
219 * used to determine whether there's something to be done.
220 */
221#ifdef CONFIG_SMP
222typedef cpumask_var_t mayday_mask_t;
223#define mayday_test_and_set_cpu(cpu, mask) \
224 cpumask_test_and_set_cpu((cpu), (mask))
225#define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask))
226#define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask))
9c37547a 227#define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp))
f2e005aa
TH
228#define free_mayday_mask(mask) free_cpumask_var((mask))
229#else
230typedef unsigned long mayday_mask_t;
231#define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask))
232#define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask))
233#define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask))
234#define alloc_mayday_mask(maskp, gfp) true
235#define free_mayday_mask(mask) do { } while (0)
236#endif
1da177e4
LT
237
238/*
239 * The externally visible workqueue abstraction is an array of
240 * per-CPU workqueues:
241 */
242struct workqueue_struct {
9c5a2ba7 243 unsigned int flags; /* W: WQ_* flags */
bdbc5dd7
TH
244 union {
245 struct cpu_workqueue_struct __percpu *pcpu;
246 struct cpu_workqueue_struct *single;
247 unsigned long v;
248 } cpu_wq; /* I: cwq's */
4690c4ab 249 struct list_head list; /* W: list of all workqueues */
73f53c4a
TH
250
251 struct mutex flush_mutex; /* protects wq flushing */
252 int work_color; /* F: current work color */
253 int flush_color; /* F: current flush color */
254 atomic_t nr_cwqs_to_flush; /* flush in progress */
255 struct wq_flusher *first_flusher; /* F: first flusher */
256 struct list_head flusher_queue; /* F: flush waiters */
257 struct list_head flusher_overflow; /* F: flush overflow list */
258
f2e005aa 259 mayday_mask_t mayday_mask; /* cpus requesting rescue */
e22bee78
TH
260 struct worker *rescuer; /* I: rescue worker */
261
9c5a2ba7 262 int nr_drainers; /* W: drain in progress */
dcd989cb 263 int saved_max_active; /* W: saved cwq max_active */
4e6045f1 264#ifdef CONFIG_LOCKDEP
4690c4ab 265 struct lockdep_map lockdep_map;
4e6045f1 266#endif
b196be89 267 char name[]; /* I: workqueue name */
1da177e4
LT
268};
269
d320c038
TH
270struct workqueue_struct *system_wq __read_mostly;
271struct workqueue_struct *system_long_wq __read_mostly;
272struct workqueue_struct *system_nrt_wq __read_mostly;
f3421797 273struct workqueue_struct *system_unbound_wq __read_mostly;
24d51add 274struct workqueue_struct *system_freezable_wq __read_mostly;
62d3c543 275struct workqueue_struct *system_nrt_freezable_wq __read_mostly;
d320c038
TH
276EXPORT_SYMBOL_GPL(system_wq);
277EXPORT_SYMBOL_GPL(system_long_wq);
278EXPORT_SYMBOL_GPL(system_nrt_wq);
f3421797 279EXPORT_SYMBOL_GPL(system_unbound_wq);
24d51add 280EXPORT_SYMBOL_GPL(system_freezable_wq);
62d3c543 281EXPORT_SYMBOL_GPL(system_nrt_freezable_wq);
d320c038 282
97bd2347
TH
283#define CREATE_TRACE_POINTS
284#include <trace/events/workqueue.h>
285
4ce62e9e 286#define for_each_worker_pool(pool, gcwq) \
3270476a
TH
287 for ((pool) = &(gcwq)->pools[0]; \
288 (pool) < &(gcwq)->pools[NR_WORKER_POOLS]; (pool)++)
4ce62e9e 289
db7bccf4
TH
290#define for_each_busy_worker(worker, i, pos, gcwq) \
291 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \
292 hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry)
293
f3421797
TH
294static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask,
295 unsigned int sw)
296{
297 if (cpu < nr_cpu_ids) {
298 if (sw & 1) {
299 cpu = cpumask_next(cpu, mask);
300 if (cpu < nr_cpu_ids)
301 return cpu;
302 }
303 if (sw & 2)
304 return WORK_CPU_UNBOUND;
305 }
306 return WORK_CPU_NONE;
307}
308
309static inline int __next_wq_cpu(int cpu, const struct cpumask *mask,
310 struct workqueue_struct *wq)
311{
312 return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2);
313}
314
09884951
TH
315/*
316 * CPU iterators
317 *
318 * An extra gcwq is defined for an invalid cpu number
319 * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any
320 * specific CPU. The following iterators are similar to
321 * for_each_*_cpu() iterators but also considers the unbound gcwq.
322 *
323 * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND
324 * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND
325 * for_each_cwq_cpu() : possible CPUs for bound workqueues,
326 * WORK_CPU_UNBOUND for unbound workqueues
327 */
f3421797
TH
328#define for_each_gcwq_cpu(cpu) \
329 for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \
330 (cpu) < WORK_CPU_NONE; \
331 (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3))
332
333#define for_each_online_gcwq_cpu(cpu) \
334 for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \
335 (cpu) < WORK_CPU_NONE; \
336 (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3))
337
338#define for_each_cwq_cpu(cpu, wq) \
339 for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \
340 (cpu) < WORK_CPU_NONE; \
341 (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq)))
342
dc186ad7
TG
343#ifdef CONFIG_DEBUG_OBJECTS_WORK
344
345static struct debug_obj_descr work_debug_descr;
346
99777288
SG
347static void *work_debug_hint(void *addr)
348{
349 return ((struct work_struct *) addr)->func;
350}
351
dc186ad7
TG
352/*
353 * fixup_init is called when:
354 * - an active object is initialized
355 */
356static int work_fixup_init(void *addr, enum debug_obj_state state)
357{
358 struct work_struct *work = addr;
359
360 switch (state) {
361 case ODEBUG_STATE_ACTIVE:
362 cancel_work_sync(work);
363 debug_object_init(work, &work_debug_descr);
364 return 1;
365 default:
366 return 0;
367 }
368}
369
370/*
371 * fixup_activate is called when:
372 * - an active object is activated
373 * - an unknown object is activated (might be a statically initialized object)
374 */
375static int work_fixup_activate(void *addr, enum debug_obj_state state)
376{
377 struct work_struct *work = addr;
378
379 switch (state) {
380
381 case ODEBUG_STATE_NOTAVAILABLE:
382 /*
383 * This is not really a fixup. The work struct was
384 * statically initialized. We just make sure that it
385 * is tracked in the object tracker.
386 */
22df02bb 387 if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) {
dc186ad7
TG
388 debug_object_init(work, &work_debug_descr);
389 debug_object_activate(work, &work_debug_descr);
390 return 0;
391 }
392 WARN_ON_ONCE(1);
393 return 0;
394
395 case ODEBUG_STATE_ACTIVE:
396 WARN_ON(1);
397
398 default:
399 return 0;
400 }
401}
402
403/*
404 * fixup_free is called when:
405 * - an active object is freed
406 */
407static int work_fixup_free(void *addr, enum debug_obj_state state)
408{
409 struct work_struct *work = addr;
410
411 switch (state) {
412 case ODEBUG_STATE_ACTIVE:
413 cancel_work_sync(work);
414 debug_object_free(work, &work_debug_descr);
415 return 1;
416 default:
417 return 0;
418 }
419}
420
421static struct debug_obj_descr work_debug_descr = {
422 .name = "work_struct",
99777288 423 .debug_hint = work_debug_hint,
dc186ad7
TG
424 .fixup_init = work_fixup_init,
425 .fixup_activate = work_fixup_activate,
426 .fixup_free = work_fixup_free,
427};
428
429static inline void debug_work_activate(struct work_struct *work)
430{
431 debug_object_activate(work, &work_debug_descr);
432}
433
434static inline void debug_work_deactivate(struct work_struct *work)
435{
436 debug_object_deactivate(work, &work_debug_descr);
437}
438
439void __init_work(struct work_struct *work, int onstack)
440{
441 if (onstack)
442 debug_object_init_on_stack(work, &work_debug_descr);
443 else
444 debug_object_init(work, &work_debug_descr);
445}
446EXPORT_SYMBOL_GPL(__init_work);
447
448void destroy_work_on_stack(struct work_struct *work)
449{
450 debug_object_free(work, &work_debug_descr);
451}
452EXPORT_SYMBOL_GPL(destroy_work_on_stack);
453
454#else
455static inline void debug_work_activate(struct work_struct *work) { }
456static inline void debug_work_deactivate(struct work_struct *work) { }
457#endif
458
95402b38
GS
459/* Serializes the accesses to the list of workqueues. */
460static DEFINE_SPINLOCK(workqueue_lock);
1da177e4 461static LIST_HEAD(workqueues);
a0a1a5fd 462static bool workqueue_freezing; /* W: have wqs started freezing? */
c34056a3 463
e22bee78
TH
464/*
465 * The almighty global cpu workqueues. nr_running is the only field
466 * which is expected to be used frequently by other cpus via
467 * try_to_wake_up(). Put it in a separate cacheline.
468 */
8b03ae3c 469static DEFINE_PER_CPU(struct global_cwq, global_cwq);
4ce62e9e 470static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, pool_nr_running[NR_WORKER_POOLS]);
8b03ae3c 471
f3421797
TH
472/*
473 * Global cpu workqueue and nr_running counter for unbound gcwq. The
474 * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its
475 * workers have WORKER_UNBOUND set.
476 */
477static struct global_cwq unbound_global_cwq;
4ce62e9e
TH
478static atomic_t unbound_pool_nr_running[NR_WORKER_POOLS] = {
479 [0 ... NR_WORKER_POOLS - 1] = ATOMIC_INIT(0), /* always 0 */
480};
f3421797 481
c34056a3 482static int worker_thread(void *__worker);
1da177e4 483
3270476a
TH
484static int worker_pool_pri(struct worker_pool *pool)
485{
486 return pool - pool->gcwq->pools;
487}
488
8b03ae3c
TH
489static struct global_cwq *get_gcwq(unsigned int cpu)
490{
f3421797
TH
491 if (cpu != WORK_CPU_UNBOUND)
492 return &per_cpu(global_cwq, cpu);
493 else
494 return &unbound_global_cwq;
8b03ae3c
TH
495}
496
63d95a91 497static atomic_t *get_pool_nr_running(struct worker_pool *pool)
e22bee78 498{
63d95a91 499 int cpu = pool->gcwq->cpu;
3270476a 500 int idx = worker_pool_pri(pool);
63d95a91 501
f3421797 502 if (cpu != WORK_CPU_UNBOUND)
4ce62e9e 503 return &per_cpu(pool_nr_running, cpu)[idx];
f3421797 504 else
4ce62e9e 505 return &unbound_pool_nr_running[idx];
e22bee78
TH
506}
507
1537663f
TH
508static struct cpu_workqueue_struct *get_cwq(unsigned int cpu,
509 struct workqueue_struct *wq)
b1f4ec17 510{
f3421797 511 if (!(wq->flags & WQ_UNBOUND)) {
e06ffa1e 512 if (likely(cpu < nr_cpu_ids))
f3421797 513 return per_cpu_ptr(wq->cpu_wq.pcpu, cpu);
f3421797
TH
514 } else if (likely(cpu == WORK_CPU_UNBOUND))
515 return wq->cpu_wq.single;
516 return NULL;
b1f4ec17
ON
517}
518
73f53c4a
TH
519static unsigned int work_color_to_flags(int color)
520{
521 return color << WORK_STRUCT_COLOR_SHIFT;
522}
523
524static int get_work_color(struct work_struct *work)
525{
526 return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) &
527 ((1 << WORK_STRUCT_COLOR_BITS) - 1);
528}
529
530static int work_next_color(int color)
531{
532 return (color + 1) % WORK_NR_COLORS;
533}
1da177e4 534
14441960 535/*
b5490077
TH
536 * While queued, %WORK_STRUCT_CWQ is set and non flag bits of a work's data
537 * contain the pointer to the queued cwq. Once execution starts, the flag
538 * is cleared and the high bits contain OFFQ flags and CPU number.
7a22ad75 539 *
8930caba
TH
540 * set_work_cwq(), set_work_cpu_and_clear_pending() and clear_work_data()
541 * can be used to set the cwq, cpu or clear work->data. These functions
542 * should only be called while the work is owned - ie. while the PENDING
543 * bit is set.
7a22ad75
TH
544 *
545 * get_work_[g]cwq() can be used to obtain the gcwq or cwq
546 * corresponding to a work. gcwq is available once the work has been
547 * queued anywhere after initialization. cwq is available only from
548 * queueing until execution starts.
14441960 549 */
7a22ad75
TH
550static inline void set_work_data(struct work_struct *work, unsigned long data,
551 unsigned long flags)
365970a1 552{
4594bf15 553 BUG_ON(!work_pending(work));
7a22ad75
TH
554 atomic_long_set(&work->data, data | flags | work_static(work));
555}
365970a1 556
7a22ad75
TH
557static void set_work_cwq(struct work_struct *work,
558 struct cpu_workqueue_struct *cwq,
559 unsigned long extra_flags)
560{
561 set_work_data(work, (unsigned long)cwq,
e120153d 562 WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags);
365970a1
DH
563}
564
8930caba
TH
565static void set_work_cpu_and_clear_pending(struct work_struct *work,
566 unsigned int cpu)
7a22ad75 567{
b5490077 568 set_work_data(work, (unsigned long)cpu << WORK_OFFQ_CPU_SHIFT, 0);
7a22ad75 569}
f756d5e2 570
7a22ad75 571static void clear_work_data(struct work_struct *work)
1da177e4 572{
7a22ad75 573 set_work_data(work, WORK_STRUCT_NO_CPU, 0);
1da177e4
LT
574}
575
7a22ad75 576static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work)
b1f4ec17 577{
e120153d 578 unsigned long data = atomic_long_read(&work->data);
7a22ad75 579
e120153d
TH
580 if (data & WORK_STRUCT_CWQ)
581 return (void *)(data & WORK_STRUCT_WQ_DATA_MASK);
582 else
583 return NULL;
4d707b9f
ON
584}
585
7a22ad75 586static struct global_cwq *get_work_gcwq(struct work_struct *work)
365970a1 587{
e120153d 588 unsigned long data = atomic_long_read(&work->data);
7a22ad75
TH
589 unsigned int cpu;
590
e120153d
TH
591 if (data & WORK_STRUCT_CWQ)
592 return ((struct cpu_workqueue_struct *)
bd7bdd43 593 (data & WORK_STRUCT_WQ_DATA_MASK))->pool->gcwq;
7a22ad75 594
b5490077 595 cpu = data >> WORK_OFFQ_CPU_SHIFT;
bdbc5dd7 596 if (cpu == WORK_CPU_NONE)
7a22ad75
TH
597 return NULL;
598
f3421797 599 BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND);
7a22ad75 600 return get_gcwq(cpu);
b1f4ec17
ON
601}
602
e22bee78 603/*
3270476a
TH
604 * Policy functions. These define the policies on how the global worker
605 * pools are managed. Unless noted otherwise, these functions assume that
606 * they're being called with gcwq->lock held.
e22bee78
TH
607 */
608
63d95a91 609static bool __need_more_worker(struct worker_pool *pool)
a848e3b6 610{
3270476a 611 return !atomic_read(get_pool_nr_running(pool));
a848e3b6
ON
612}
613
4594bf15 614/*
e22bee78
TH
615 * Need to wake up a worker? Called from anything but currently
616 * running workers.
974271c4
TH
617 *
618 * Note that, because unbound workers never contribute to nr_running, this
619 * function will always return %true for unbound gcwq as long as the
620 * worklist isn't empty.
4594bf15 621 */
63d95a91 622static bool need_more_worker(struct worker_pool *pool)
365970a1 623{
63d95a91 624 return !list_empty(&pool->worklist) && __need_more_worker(pool);
e22bee78 625}
4594bf15 626
e22bee78 627/* Can I start working? Called from busy but !running workers. */
63d95a91 628static bool may_start_working(struct worker_pool *pool)
e22bee78 629{
63d95a91 630 return pool->nr_idle;
e22bee78
TH
631}
632
633/* Do I need to keep working? Called from currently running workers. */
63d95a91 634static bool keep_working(struct worker_pool *pool)
e22bee78 635{
63d95a91 636 atomic_t *nr_running = get_pool_nr_running(pool);
e22bee78 637
3270476a 638 return !list_empty(&pool->worklist) && atomic_read(nr_running) <= 1;
e22bee78
TH
639}
640
641/* Do we need a new worker? Called from manager. */
63d95a91 642static bool need_to_create_worker(struct worker_pool *pool)
e22bee78 643{
63d95a91 644 return need_more_worker(pool) && !may_start_working(pool);
e22bee78 645}
365970a1 646
e22bee78 647/* Do I need to be the manager? */
63d95a91 648static bool need_to_manage_workers(struct worker_pool *pool)
e22bee78 649{
63d95a91 650 return need_to_create_worker(pool) ||
11ebea50 651 (pool->flags & POOL_MANAGE_WORKERS);
e22bee78
TH
652}
653
654/* Do we have too many workers and should some go away? */
63d95a91 655static bool too_many_workers(struct worker_pool *pool)
e22bee78 656{
60373152 657 bool managing = mutex_is_locked(&pool->manager_mutex);
63d95a91
TH
658 int nr_idle = pool->nr_idle + managing; /* manager is considered idle */
659 int nr_busy = pool->nr_workers - nr_idle;
e22bee78
TH
660
661 return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy;
365970a1
DH
662}
663
4d707b9f 664/*
e22bee78
TH
665 * Wake up functions.
666 */
667
7e11629d 668/* Return the first worker. Safe with preemption disabled */
63d95a91 669static struct worker *first_worker(struct worker_pool *pool)
7e11629d 670{
63d95a91 671 if (unlikely(list_empty(&pool->idle_list)))
7e11629d
TH
672 return NULL;
673
63d95a91 674 return list_first_entry(&pool->idle_list, struct worker, entry);
7e11629d
TH
675}
676
677/**
678 * wake_up_worker - wake up an idle worker
63d95a91 679 * @pool: worker pool to wake worker from
7e11629d 680 *
63d95a91 681 * Wake up the first idle worker of @pool.
7e11629d
TH
682 *
683 * CONTEXT:
684 * spin_lock_irq(gcwq->lock).
685 */
63d95a91 686static void wake_up_worker(struct worker_pool *pool)
7e11629d 687{
63d95a91 688 struct worker *worker = first_worker(pool);
7e11629d
TH
689
690 if (likely(worker))
691 wake_up_process(worker->task);
692}
693
d302f017 694/**
e22bee78
TH
695 * wq_worker_waking_up - a worker is waking up
696 * @task: task waking up
697 * @cpu: CPU @task is waking up to
698 *
699 * This function is called during try_to_wake_up() when a worker is
700 * being awoken.
701 *
702 * CONTEXT:
703 * spin_lock_irq(rq->lock)
704 */
705void wq_worker_waking_up(struct task_struct *task, unsigned int cpu)
706{
707 struct worker *worker = kthread_data(task);
708
2d64672e 709 if (!(worker->flags & WORKER_NOT_RUNNING))
63d95a91 710 atomic_inc(get_pool_nr_running(worker->pool));
e22bee78
TH
711}
712
713/**
714 * wq_worker_sleeping - a worker is going to sleep
715 * @task: task going to sleep
716 * @cpu: CPU in question, must be the current CPU number
717 *
718 * This function is called during schedule() when a busy worker is
719 * going to sleep. Worker on the same cpu can be woken up by
720 * returning pointer to its task.
721 *
722 * CONTEXT:
723 * spin_lock_irq(rq->lock)
724 *
725 * RETURNS:
726 * Worker task on @cpu to wake up, %NULL if none.
727 */
728struct task_struct *wq_worker_sleeping(struct task_struct *task,
729 unsigned int cpu)
730{
731 struct worker *worker = kthread_data(task), *to_wakeup = NULL;
bd7bdd43 732 struct worker_pool *pool = worker->pool;
63d95a91 733 atomic_t *nr_running = get_pool_nr_running(pool);
e22bee78 734
2d64672e 735 if (worker->flags & WORKER_NOT_RUNNING)
e22bee78
TH
736 return NULL;
737
738 /* this can only happen on the local cpu */
739 BUG_ON(cpu != raw_smp_processor_id());
740
741 /*
742 * The counterpart of the following dec_and_test, implied mb,
743 * worklist not empty test sequence is in insert_work().
744 * Please read comment there.
745 *
628c78e7
TH
746 * NOT_RUNNING is clear. This means that we're bound to and
747 * running on the local cpu w/ rq lock held and preemption
748 * disabled, which in turn means that none else could be
749 * manipulating idle_list, so dereferencing idle_list without gcwq
750 * lock is safe.
e22bee78 751 */
bd7bdd43 752 if (atomic_dec_and_test(nr_running) && !list_empty(&pool->worklist))
63d95a91 753 to_wakeup = first_worker(pool);
e22bee78
TH
754 return to_wakeup ? to_wakeup->task : NULL;
755}
756
757/**
758 * worker_set_flags - set worker flags and adjust nr_running accordingly
cb444766 759 * @worker: self
d302f017
TH
760 * @flags: flags to set
761 * @wakeup: wakeup an idle worker if necessary
762 *
e22bee78
TH
763 * Set @flags in @worker->flags and adjust nr_running accordingly. If
764 * nr_running becomes zero and @wakeup is %true, an idle worker is
765 * woken up.
d302f017 766 *
cb444766
TH
767 * CONTEXT:
768 * spin_lock_irq(gcwq->lock)
d302f017
TH
769 */
770static inline void worker_set_flags(struct worker *worker, unsigned int flags,
771 bool wakeup)
772{
bd7bdd43 773 struct worker_pool *pool = worker->pool;
e22bee78 774
cb444766
TH
775 WARN_ON_ONCE(worker->task != current);
776
e22bee78
TH
777 /*
778 * If transitioning into NOT_RUNNING, adjust nr_running and
779 * wake up an idle worker as necessary if requested by
780 * @wakeup.
781 */
782 if ((flags & WORKER_NOT_RUNNING) &&
783 !(worker->flags & WORKER_NOT_RUNNING)) {
63d95a91 784 atomic_t *nr_running = get_pool_nr_running(pool);
e22bee78
TH
785
786 if (wakeup) {
787 if (atomic_dec_and_test(nr_running) &&
bd7bdd43 788 !list_empty(&pool->worklist))
63d95a91 789 wake_up_worker(pool);
e22bee78
TH
790 } else
791 atomic_dec(nr_running);
792 }
793
d302f017
TH
794 worker->flags |= flags;
795}
796
797/**
e22bee78 798 * worker_clr_flags - clear worker flags and adjust nr_running accordingly
cb444766 799 * @worker: self
d302f017
TH
800 * @flags: flags to clear
801 *
e22bee78 802 * Clear @flags in @worker->flags and adjust nr_running accordingly.
d302f017 803 *
cb444766
TH
804 * CONTEXT:
805 * spin_lock_irq(gcwq->lock)
d302f017
TH
806 */
807static inline void worker_clr_flags(struct worker *worker, unsigned int flags)
808{
63d95a91 809 struct worker_pool *pool = worker->pool;
e22bee78
TH
810 unsigned int oflags = worker->flags;
811
cb444766
TH
812 WARN_ON_ONCE(worker->task != current);
813
d302f017 814 worker->flags &= ~flags;
e22bee78 815
42c025f3
TH
816 /*
817 * If transitioning out of NOT_RUNNING, increment nr_running. Note
818 * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask
819 * of multiple flags, not a single flag.
820 */
e22bee78
TH
821 if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING))
822 if (!(worker->flags & WORKER_NOT_RUNNING))
63d95a91 823 atomic_inc(get_pool_nr_running(pool));
d302f017
TH
824}
825
c8e55f36
TH
826/**
827 * busy_worker_head - return the busy hash head for a work
828 * @gcwq: gcwq of interest
829 * @work: work to be hashed
830 *
831 * Return hash head of @gcwq for @work.
832 *
833 * CONTEXT:
834 * spin_lock_irq(gcwq->lock).
835 *
836 * RETURNS:
837 * Pointer to the hash head.
838 */
839static struct hlist_head *busy_worker_head(struct global_cwq *gcwq,
840 struct work_struct *work)
841{
842 const int base_shift = ilog2(sizeof(struct work_struct));
843 unsigned long v = (unsigned long)work;
844
845 /* simple shift and fold hash, do we need something better? */
846 v >>= base_shift;
847 v += v >> BUSY_WORKER_HASH_ORDER;
848 v &= BUSY_WORKER_HASH_MASK;
849
850 return &gcwq->busy_hash[v];
851}
852
8cca0eea
TH
853/**
854 * __find_worker_executing_work - find worker which is executing a work
855 * @gcwq: gcwq of interest
856 * @bwh: hash head as returned by busy_worker_head()
857 * @work: work to find worker for
858 *
859 * Find a worker which is executing @work on @gcwq. @bwh should be
860 * the hash head obtained by calling busy_worker_head() with the same
861 * work.
862 *
863 * CONTEXT:
864 * spin_lock_irq(gcwq->lock).
865 *
866 * RETURNS:
867 * Pointer to worker which is executing @work if found, NULL
868 * otherwise.
869 */
870static struct worker *__find_worker_executing_work(struct global_cwq *gcwq,
871 struct hlist_head *bwh,
872 struct work_struct *work)
873{
874 struct worker *worker;
875 struct hlist_node *tmp;
876
877 hlist_for_each_entry(worker, tmp, bwh, hentry)
878 if (worker->current_work == work)
879 return worker;
880 return NULL;
881}
882
883/**
884 * find_worker_executing_work - find worker which is executing a work
885 * @gcwq: gcwq of interest
886 * @work: work to find worker for
887 *
888 * Find a worker which is executing @work on @gcwq. This function is
889 * identical to __find_worker_executing_work() except that this
890 * function calculates @bwh itself.
891 *
892 * CONTEXT:
893 * spin_lock_irq(gcwq->lock).
894 *
895 * RETURNS:
896 * Pointer to worker which is executing @work if found, NULL
897 * otherwise.
4d707b9f 898 */
8cca0eea
TH
899static struct worker *find_worker_executing_work(struct global_cwq *gcwq,
900 struct work_struct *work)
4d707b9f 901{
8cca0eea
TH
902 return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work),
903 work);
4d707b9f
ON
904}
905
bf4ede01
TH
906/**
907 * move_linked_works - move linked works to a list
908 * @work: start of series of works to be scheduled
909 * @head: target list to append @work to
910 * @nextp: out paramter for nested worklist walking
911 *
912 * Schedule linked works starting from @work to @head. Work series to
913 * be scheduled starts at @work and includes any consecutive work with
914 * WORK_STRUCT_LINKED set in its predecessor.
915 *
916 * If @nextp is not NULL, it's updated to point to the next work of
917 * the last scheduled work. This allows move_linked_works() to be
918 * nested inside outer list_for_each_entry_safe().
919 *
920 * CONTEXT:
921 * spin_lock_irq(gcwq->lock).
922 */
923static void move_linked_works(struct work_struct *work, struct list_head *head,
924 struct work_struct **nextp)
925{
926 struct work_struct *n;
927
928 /*
929 * Linked worklist will always end before the end of the list,
930 * use NULL for list head.
931 */
932 list_for_each_entry_safe_from(work, n, NULL, entry) {
933 list_move_tail(&work->entry, head);
934 if (!(*work_data_bits(work) & WORK_STRUCT_LINKED))
935 break;
936 }
937
938 /*
939 * If we're already inside safe list traversal and have moved
940 * multiple works to the scheduled queue, the next position
941 * needs to be updated.
942 */
943 if (nextp)
944 *nextp = n;
945}
946
947static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq)
948{
949 struct work_struct *work = list_first_entry(&cwq->delayed_works,
950 struct work_struct, entry);
951
952 trace_workqueue_activate_work(work);
953 move_linked_works(work, &cwq->pool->worklist, NULL);
954 __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work));
955 cwq->nr_active++;
956}
957
958/**
959 * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight
960 * @cwq: cwq of interest
961 * @color: color of work which left the queue
962 * @delayed: for a delayed work
963 *
964 * A work either has completed or is removed from pending queue,
965 * decrement nr_in_flight of its cwq and handle workqueue flushing.
966 *
967 * CONTEXT:
968 * spin_lock_irq(gcwq->lock).
969 */
970static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color,
971 bool delayed)
972{
973 /* ignore uncolored works */
974 if (color == WORK_NO_COLOR)
975 return;
976
977 cwq->nr_in_flight[color]--;
978
979 if (!delayed) {
980 cwq->nr_active--;
981 if (!list_empty(&cwq->delayed_works)) {
982 /* one down, submit a delayed one */
983 if (cwq->nr_active < cwq->max_active)
984 cwq_activate_first_delayed(cwq);
985 }
986 }
987
988 /* is flush in progress and are we at the flushing tip? */
989 if (likely(cwq->flush_color != color))
990 return;
991
992 /* are there still in-flight works? */
993 if (cwq->nr_in_flight[color])
994 return;
995
996 /* this cwq is done, clear flush_color */
997 cwq->flush_color = -1;
998
999 /*
1000 * If this was the last cwq, wake up the first flusher. It
1001 * will handle the rest.
1002 */
1003 if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush))
1004 complete(&cwq->wq->first_flusher->done);
1005}
1006
1007/*
1008 * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit,
1009 * so this work can't be re-armed in any way.
1010 */
1011static int try_to_grab_pending(struct work_struct *work)
1012{
1013 struct global_cwq *gcwq;
1014 int ret = -1;
1015
1016 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work)))
1017 return 0;
1018
1019 /*
1020 * The queueing is in progress, or it is already queued. Try to
1021 * steal it from ->worklist without clearing WORK_STRUCT_PENDING.
1022 */
1023 gcwq = get_work_gcwq(work);
1024 if (!gcwq)
1025 return ret;
1026
1027 spin_lock_irq(&gcwq->lock);
1028 if (!list_empty(&work->entry)) {
1029 /*
1030 * This work is queued, but perhaps we locked the wrong gcwq.
1031 * In that case we must see the new value after rmb(), see
1032 * insert_work()->wmb().
1033 */
1034 smp_rmb();
1035 if (gcwq == get_work_gcwq(work)) {
1036 debug_work_deactivate(work);
1037 list_del_init(&work->entry);
1038 cwq_dec_nr_in_flight(get_work_cwq(work),
1039 get_work_color(work),
1040 *work_data_bits(work) & WORK_STRUCT_DELAYED);
1041 ret = 1;
1042 }
1043 }
1044 spin_unlock_irq(&gcwq->lock);
1045
1046 return ret;
1047}
1048
4690c4ab 1049/**
7e11629d 1050 * insert_work - insert a work into gcwq
4690c4ab
TH
1051 * @cwq: cwq @work belongs to
1052 * @work: work to insert
1053 * @head: insertion point
1054 * @extra_flags: extra WORK_STRUCT_* flags to set
1055 *
7e11629d
TH
1056 * Insert @work which belongs to @cwq into @gcwq after @head.
1057 * @extra_flags is or'd to work_struct flags.
4690c4ab
TH
1058 *
1059 * CONTEXT:
8b03ae3c 1060 * spin_lock_irq(gcwq->lock).
4690c4ab 1061 */
b89deed3 1062static void insert_work(struct cpu_workqueue_struct *cwq,
4690c4ab
TH
1063 struct work_struct *work, struct list_head *head,
1064 unsigned int extra_flags)
b89deed3 1065{
63d95a91 1066 struct worker_pool *pool = cwq->pool;
e22bee78 1067
4690c4ab 1068 /* we own @work, set data and link */
7a22ad75 1069 set_work_cwq(work, cwq, extra_flags);
e1d8aa9f 1070
6e84d644
ON
1071 /*
1072 * Ensure that we get the right work->data if we see the
1073 * result of list_add() below, see try_to_grab_pending().
1074 */
1075 smp_wmb();
4690c4ab 1076
1a4d9b0a 1077 list_add_tail(&work->entry, head);
e22bee78
TH
1078
1079 /*
1080 * Ensure either worker_sched_deactivated() sees the above
1081 * list_add_tail() or we see zero nr_running to avoid workers
1082 * lying around lazily while there are works to be processed.
1083 */
1084 smp_mb();
1085
63d95a91
TH
1086 if (__need_more_worker(pool))
1087 wake_up_worker(pool);
b89deed3
ON
1088}
1089
c8efcc25
TH
1090/*
1091 * Test whether @work is being queued from another work executing on the
1092 * same workqueue. This is rather expensive and should only be used from
1093 * cold paths.
1094 */
1095static bool is_chained_work(struct workqueue_struct *wq)
1096{
1097 unsigned long flags;
1098 unsigned int cpu;
1099
1100 for_each_gcwq_cpu(cpu) {
1101 struct global_cwq *gcwq = get_gcwq(cpu);
1102 struct worker *worker;
1103 struct hlist_node *pos;
1104 int i;
1105
1106 spin_lock_irqsave(&gcwq->lock, flags);
1107 for_each_busy_worker(worker, i, pos, gcwq) {
1108 if (worker->task != current)
1109 continue;
1110 spin_unlock_irqrestore(&gcwq->lock, flags);
1111 /*
1112 * I'm @worker, no locking necessary. See if @work
1113 * is headed to the same workqueue.
1114 */
1115 return worker->current_cwq->wq == wq;
1116 }
1117 spin_unlock_irqrestore(&gcwq->lock, flags);
1118 }
1119 return false;
1120}
1121
4690c4ab 1122static void __queue_work(unsigned int cpu, struct workqueue_struct *wq,
1da177e4
LT
1123 struct work_struct *work)
1124{
502ca9d8
TH
1125 struct global_cwq *gcwq;
1126 struct cpu_workqueue_struct *cwq;
1e19ffc6 1127 struct list_head *worklist;
8a2e8e5d 1128 unsigned int work_flags;
8930caba
TH
1129
1130 /*
1131 * While a work item is PENDING && off queue, a task trying to
1132 * steal the PENDING will busy-loop waiting for it to either get
1133 * queued or lose PENDING. Grabbing PENDING and queueing should
1134 * happen with IRQ disabled.
1135 */
1136 WARN_ON_ONCE(!irqs_disabled());
1da177e4 1137
dc186ad7 1138 debug_work_activate(work);
1e19ffc6 1139
c8efcc25 1140 /* if dying, only works from the same workqueue are allowed */
9c5a2ba7 1141 if (unlikely(wq->flags & WQ_DRAINING) &&
c8efcc25 1142 WARN_ON_ONCE(!is_chained_work(wq)))
e41e704b
TH
1143 return;
1144
c7fc77f7
TH
1145 /* determine gcwq to use */
1146 if (!(wq->flags & WQ_UNBOUND)) {
18aa9eff
TH
1147 struct global_cwq *last_gcwq;
1148
57469821 1149 if (cpu == WORK_CPU_UNBOUND)
c7fc77f7
TH
1150 cpu = raw_smp_processor_id();
1151
18aa9eff
TH
1152 /*
1153 * It's multi cpu. If @wq is non-reentrant and @work
1154 * was previously on a different cpu, it might still
1155 * be running there, in which case the work needs to
1156 * be queued on that cpu to guarantee non-reentrance.
1157 */
502ca9d8 1158 gcwq = get_gcwq(cpu);
18aa9eff
TH
1159 if (wq->flags & WQ_NON_REENTRANT &&
1160 (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) {
1161 struct worker *worker;
1162
8930caba 1163 spin_lock(&last_gcwq->lock);
18aa9eff
TH
1164
1165 worker = find_worker_executing_work(last_gcwq, work);
1166
1167 if (worker && worker->current_cwq->wq == wq)
1168 gcwq = last_gcwq;
1169 else {
1170 /* meh... not running there, queue here */
8930caba
TH
1171 spin_unlock(&last_gcwq->lock);
1172 spin_lock(&gcwq->lock);
18aa9eff 1173 }
8930caba
TH
1174 } else {
1175 spin_lock(&gcwq->lock);
1176 }
f3421797
TH
1177 } else {
1178 gcwq = get_gcwq(WORK_CPU_UNBOUND);
8930caba 1179 spin_lock(&gcwq->lock);
502ca9d8
TH
1180 }
1181
1182 /* gcwq determined, get cwq and queue */
1183 cwq = get_cwq(gcwq->cpu, wq);
cdadf009 1184 trace_workqueue_queue_work(cpu, cwq, work);
502ca9d8 1185
f5b2552b 1186 if (WARN_ON(!list_empty(&work->entry))) {
8930caba 1187 spin_unlock(&gcwq->lock);
f5b2552b
DC
1188 return;
1189 }
1e19ffc6 1190
73f53c4a 1191 cwq->nr_in_flight[cwq->work_color]++;
8a2e8e5d 1192 work_flags = work_color_to_flags(cwq->work_color);
1e19ffc6
TH
1193
1194 if (likely(cwq->nr_active < cwq->max_active)) {
cdadf009 1195 trace_workqueue_activate_work(work);
1e19ffc6 1196 cwq->nr_active++;
3270476a 1197 worklist = &cwq->pool->worklist;
8a2e8e5d
TH
1198 } else {
1199 work_flags |= WORK_STRUCT_DELAYED;
1e19ffc6 1200 worklist = &cwq->delayed_works;
8a2e8e5d 1201 }
1e19ffc6 1202
8a2e8e5d 1203 insert_work(cwq, work, worklist, work_flags);
1e19ffc6 1204
8930caba 1205 spin_unlock(&gcwq->lock);
1da177e4
LT
1206}
1207
c1a220e7
ZR
1208/**
1209 * queue_work_on - queue work on specific cpu
1210 * @cpu: CPU number to execute work on
1211 * @wq: workqueue to use
1212 * @work: work to queue
1213 *
d4283e93 1214 * Returns %false if @work was already on a queue, %true otherwise.
c1a220e7
ZR
1215 *
1216 * We queue the work to a specific CPU, the caller must ensure it
1217 * can't go away.
1218 */
d4283e93
TH
1219bool queue_work_on(int cpu, struct workqueue_struct *wq,
1220 struct work_struct *work)
c1a220e7 1221{
d4283e93 1222 bool ret = false;
8930caba
TH
1223 unsigned long flags;
1224
1225 local_irq_save(flags);
c1a220e7 1226
22df02bb 1227 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
4690c4ab 1228 __queue_work(cpu, wq, work);
d4283e93 1229 ret = true;
c1a220e7 1230 }
8930caba
TH
1231
1232 local_irq_restore(flags);
c1a220e7
ZR
1233 return ret;
1234}
1235EXPORT_SYMBOL_GPL(queue_work_on);
1236
0fcb78c2 1237/**
0a13c00e 1238 * queue_work - queue work on a workqueue
0fcb78c2 1239 * @wq: workqueue to use
0a13c00e 1240 * @work: work to queue
0fcb78c2 1241 *
d4283e93 1242 * Returns %false if @work was already on a queue, %true otherwise.
0a13c00e
TH
1243 *
1244 * We queue the work to the CPU on which it was submitted, but if the CPU dies
1245 * it can be processed by another CPU.
0fcb78c2 1246 */
d4283e93 1247bool queue_work(struct workqueue_struct *wq, struct work_struct *work)
1da177e4 1248{
57469821 1249 return queue_work_on(WORK_CPU_UNBOUND, wq, work);
0a13c00e
TH
1250}
1251EXPORT_SYMBOL_GPL(queue_work);
1252
d8e794df 1253void delayed_work_timer_fn(unsigned long __data)
0a13c00e
TH
1254{
1255 struct delayed_work *dwork = (struct delayed_work *)__data;
1256 struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work);
1257
8930caba 1258 local_irq_disable();
57469821 1259 __queue_work(WORK_CPU_UNBOUND, cwq->wq, &dwork->work);
8930caba 1260 local_irq_enable();
1da177e4 1261}
d8e794df 1262EXPORT_SYMBOL_GPL(delayed_work_timer_fn);
1da177e4 1263
7beb2edf
TH
1264static void __queue_delayed_work(int cpu, struct workqueue_struct *wq,
1265 struct delayed_work *dwork, unsigned long delay)
1266{
1267 struct timer_list *timer = &dwork->timer;
1268 struct work_struct *work = &dwork->work;
1269 unsigned int lcpu;
1270
1271 WARN_ON_ONCE(timer->function != delayed_work_timer_fn ||
1272 timer->data != (unsigned long)dwork);
1273 BUG_ON(timer_pending(timer));
1274 BUG_ON(!list_empty(&work->entry));
1275
1276 timer_stats_timer_set_start_info(&dwork->timer);
1277
1278 /*
1279 * This stores cwq for the moment, for the timer_fn. Note that the
1280 * work's gcwq is preserved to allow reentrance detection for
1281 * delayed works.
1282 */
1283 if (!(wq->flags & WQ_UNBOUND)) {
1284 struct global_cwq *gcwq = get_work_gcwq(work);
1285
1286 if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND)
1287 lcpu = gcwq->cpu;
1288 else
1289 lcpu = raw_smp_processor_id();
1290 } else {
1291 lcpu = WORK_CPU_UNBOUND;
1292 }
1293
1294 set_work_cwq(work, get_cwq(lcpu, wq), 0);
1295
1296 timer->expires = jiffies + delay;
1297
1298 if (unlikely(cpu != WORK_CPU_UNBOUND))
1299 add_timer_on(timer, cpu);
1300 else
1301 add_timer(timer);
1302}
1303
0fcb78c2
REB
1304/**
1305 * queue_delayed_work_on - queue work on specific CPU after delay
1306 * @cpu: CPU number to execute work on
1307 * @wq: workqueue to use
af9997e4 1308 * @dwork: work to queue
0fcb78c2
REB
1309 * @delay: number of jiffies to wait before queueing
1310 *
715f1300
TH
1311 * Returns %false if @work was already on a queue, %true otherwise. If
1312 * @delay is zero and @dwork is idle, it will be scheduled for immediate
1313 * execution.
0fcb78c2 1314 */
d4283e93
TH
1315bool queue_delayed_work_on(int cpu, struct workqueue_struct *wq,
1316 struct delayed_work *dwork, unsigned long delay)
7a6bc1cd 1317{
52bad64d 1318 struct work_struct *work = &dwork->work;
d4283e93 1319 bool ret = false;
8930caba
TH
1320 unsigned long flags;
1321
715f1300
TH
1322 if (!delay)
1323 return queue_work_on(cpu, wq, &dwork->work);
1324
8930caba
TH
1325 /* read the comment in __queue_work() */
1326 local_irq_save(flags);
7a6bc1cd 1327
22df02bb 1328 if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) {
7beb2edf 1329 __queue_delayed_work(cpu, wq, dwork, delay);
d4283e93 1330 ret = true;
7a6bc1cd 1331 }
8930caba
TH
1332
1333 local_irq_restore(flags);
7a6bc1cd
VP
1334 return ret;
1335}
ae90dd5d 1336EXPORT_SYMBOL_GPL(queue_delayed_work_on);
1da177e4 1337
0a13c00e
TH
1338/**
1339 * queue_delayed_work - queue work on a workqueue after delay
1340 * @wq: workqueue to use
1341 * @dwork: delayable work to queue
1342 * @delay: number of jiffies to wait before queueing
1343 *
715f1300 1344 * Equivalent to queue_delayed_work_on() but tries to use the local CPU.
0a13c00e 1345 */
d4283e93 1346bool queue_delayed_work(struct workqueue_struct *wq,
0a13c00e
TH
1347 struct delayed_work *dwork, unsigned long delay)
1348{
57469821 1349 return queue_delayed_work_on(WORK_CPU_UNBOUND, wq, dwork, delay);
0a13c00e
TH
1350}
1351EXPORT_SYMBOL_GPL(queue_delayed_work);
1352
c8e55f36
TH
1353/**
1354 * worker_enter_idle - enter idle state
1355 * @worker: worker which is entering idle state
1356 *
1357 * @worker is entering idle state. Update stats and idle timer if
1358 * necessary.
1359 *
1360 * LOCKING:
1361 * spin_lock_irq(gcwq->lock).
1362 */
1363static void worker_enter_idle(struct worker *worker)
1da177e4 1364{
bd7bdd43
TH
1365 struct worker_pool *pool = worker->pool;
1366 struct global_cwq *gcwq = pool->gcwq;
c8e55f36
TH
1367
1368 BUG_ON(worker->flags & WORKER_IDLE);
1369 BUG_ON(!list_empty(&worker->entry) &&
1370 (worker->hentry.next || worker->hentry.pprev));
1371
cb444766
TH
1372 /* can't use worker_set_flags(), also called from start_worker() */
1373 worker->flags |= WORKER_IDLE;
bd7bdd43 1374 pool->nr_idle++;
e22bee78 1375 worker->last_active = jiffies;
c8e55f36
TH
1376
1377 /* idle_list is LIFO */
bd7bdd43 1378 list_add(&worker->entry, &pool->idle_list);
db7bccf4 1379
628c78e7
TH
1380 if (too_many_workers(pool) && !timer_pending(&pool->idle_timer))
1381 mod_timer(&pool->idle_timer, jiffies + IDLE_WORKER_TIMEOUT);
cb444766 1382
544ecf31 1383 /*
628c78e7
TH
1384 * Sanity check nr_running. Because gcwq_unbind_fn() releases
1385 * gcwq->lock between setting %WORKER_UNBOUND and zapping
1386 * nr_running, the warning may trigger spuriously. Check iff
1387 * unbind is not in progress.
544ecf31 1388 */
628c78e7 1389 WARN_ON_ONCE(!(gcwq->flags & GCWQ_DISASSOCIATED) &&
bd7bdd43 1390 pool->nr_workers == pool->nr_idle &&
63d95a91 1391 atomic_read(get_pool_nr_running(pool)));
c8e55f36
TH
1392}
1393
1394/**
1395 * worker_leave_idle - leave idle state
1396 * @worker: worker which is leaving idle state
1397 *
1398 * @worker is leaving idle state. Update stats.
1399 *
1400 * LOCKING:
1401 * spin_lock_irq(gcwq->lock).
1402 */
1403static void worker_leave_idle(struct worker *worker)
1404{
bd7bdd43 1405 struct worker_pool *pool = worker->pool;
c8e55f36
TH
1406
1407 BUG_ON(!(worker->flags & WORKER_IDLE));
d302f017 1408 worker_clr_flags(worker, WORKER_IDLE);
bd7bdd43 1409 pool->nr_idle--;
c8e55f36
TH
1410 list_del_init(&worker->entry);
1411}
1412
e22bee78
TH
1413/**
1414 * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq
1415 * @worker: self
1416 *
1417 * Works which are scheduled while the cpu is online must at least be
1418 * scheduled to a worker which is bound to the cpu so that if they are
1419 * flushed from cpu callbacks while cpu is going down, they are
1420 * guaranteed to execute on the cpu.
1421 *
1422 * This function is to be used by rogue workers and rescuers to bind
1423 * themselves to the target cpu and may race with cpu going down or
1424 * coming online. kthread_bind() can't be used because it may put the
1425 * worker to already dead cpu and set_cpus_allowed_ptr() can't be used
1426 * verbatim as it's best effort and blocking and gcwq may be
1427 * [dis]associated in the meantime.
1428 *
f2d5a0ee
TH
1429 * This function tries set_cpus_allowed() and locks gcwq and verifies the
1430 * binding against %GCWQ_DISASSOCIATED which is set during
1431 * %CPU_DOWN_PREPARE and cleared during %CPU_ONLINE, so if the worker
1432 * enters idle state or fetches works without dropping lock, it can
1433 * guarantee the scheduling requirement described in the first paragraph.
e22bee78
TH
1434 *
1435 * CONTEXT:
1436 * Might sleep. Called without any lock but returns with gcwq->lock
1437 * held.
1438 *
1439 * RETURNS:
1440 * %true if the associated gcwq is online (@worker is successfully
1441 * bound), %false if offline.
1442 */
1443static bool worker_maybe_bind_and_lock(struct worker *worker)
972fa1c5 1444__acquires(&gcwq->lock)
e22bee78 1445{
bd7bdd43 1446 struct global_cwq *gcwq = worker->pool->gcwq;
e22bee78
TH
1447 struct task_struct *task = worker->task;
1448
1449 while (true) {
4e6045f1 1450 /*
e22bee78
TH
1451 * The following call may fail, succeed or succeed
1452 * without actually migrating the task to the cpu if
1453 * it races with cpu hotunplug operation. Verify
1454 * against GCWQ_DISASSOCIATED.
4e6045f1 1455 */
f3421797
TH
1456 if (!(gcwq->flags & GCWQ_DISASSOCIATED))
1457 set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu));
e22bee78
TH
1458
1459 spin_lock_irq(&gcwq->lock);
1460 if (gcwq->flags & GCWQ_DISASSOCIATED)
1461 return false;
1462 if (task_cpu(task) == gcwq->cpu &&
1463 cpumask_equal(&current->cpus_allowed,
1464 get_cpu_mask(gcwq->cpu)))
1465 return true;
1466 spin_unlock_irq(&gcwq->lock);
1467
5035b20f
TH
1468 /*
1469 * We've raced with CPU hot[un]plug. Give it a breather
1470 * and retry migration. cond_resched() is required here;
1471 * otherwise, we might deadlock against cpu_stop trying to
1472 * bring down the CPU on non-preemptive kernel.
1473 */
e22bee78 1474 cpu_relax();
5035b20f 1475 cond_resched();
e22bee78
TH
1476 }
1477}
1478
25511a47
TH
1479struct idle_rebind {
1480 int cnt; /* # workers to be rebound */
1481 struct completion done; /* all workers rebound */
1482};
1483
1484/*
1485 * Rebind an idle @worker to its CPU. During CPU onlining, this has to
1486 * happen synchronously for idle workers. worker_thread() will test
1487 * %WORKER_REBIND before leaving idle and call this function.
1488 */
1489static void idle_worker_rebind(struct worker *worker)
1490{
1491 struct global_cwq *gcwq = worker->pool->gcwq;
1492
1493 /* CPU must be online at this point */
1494 WARN_ON(!worker_maybe_bind_and_lock(worker));
1495 if (!--worker->idle_rebind->cnt)
1496 complete(&worker->idle_rebind->done);
1497 spin_unlock_irq(&worker->pool->gcwq->lock);
1498
1499 /* we did our part, wait for rebind_workers() to finish up */
1500 wait_event(gcwq->rebind_hold, !(worker->flags & WORKER_REBIND));
1501}
1502
e22bee78 1503/*
25511a47 1504 * Function for @worker->rebind.work used to rebind unbound busy workers to
403c821d
TH
1505 * the associated cpu which is coming back online. This is scheduled by
1506 * cpu up but can race with other cpu hotplug operations and may be
1507 * executed twice without intervening cpu down.
e22bee78 1508 */
25511a47 1509static void busy_worker_rebind_fn(struct work_struct *work)
e22bee78
TH
1510{
1511 struct worker *worker = container_of(work, struct worker, rebind_work);
bd7bdd43 1512 struct global_cwq *gcwq = worker->pool->gcwq;
e22bee78
TH
1513
1514 if (worker_maybe_bind_and_lock(worker))
1515 worker_clr_flags(worker, WORKER_REBIND);
1516
1517 spin_unlock_irq(&gcwq->lock);
1518}
1519
25511a47
TH
1520/**
1521 * rebind_workers - rebind all workers of a gcwq to the associated CPU
1522 * @gcwq: gcwq of interest
1523 *
1524 * @gcwq->cpu is coming online. Rebind all workers to the CPU. Rebinding
1525 * is different for idle and busy ones.
1526 *
1527 * The idle ones should be rebound synchronously and idle rebinding should
1528 * be complete before any worker starts executing work items with
1529 * concurrency management enabled; otherwise, scheduler may oops trying to
1530 * wake up non-local idle worker from wq_worker_sleeping().
1531 *
1532 * This is achieved by repeatedly requesting rebinding until all idle
1533 * workers are known to have been rebound under @gcwq->lock and holding all
1534 * idle workers from becoming busy until idle rebinding is complete.
1535 *
1536 * Once idle workers are rebound, busy workers can be rebound as they
1537 * finish executing their current work items. Queueing the rebind work at
1538 * the head of their scheduled lists is enough. Note that nr_running will
1539 * be properbly bumped as busy workers rebind.
1540 *
1541 * On return, all workers are guaranteed to either be bound or have rebind
1542 * work item scheduled.
1543 */
1544static void rebind_workers(struct global_cwq *gcwq)
1545 __releases(&gcwq->lock) __acquires(&gcwq->lock)
1546{
1547 struct idle_rebind idle_rebind;
1548 struct worker_pool *pool;
1549 struct worker *worker;
1550 struct hlist_node *pos;
1551 int i;
1552
1553 lockdep_assert_held(&gcwq->lock);
1554
1555 for_each_worker_pool(pool, gcwq)
1556 lockdep_assert_held(&pool->manager_mutex);
1557
1558 /*
1559 * Rebind idle workers. Interlocked both ways. We wait for
1560 * workers to rebind via @idle_rebind.done. Workers will wait for
1561 * us to finish up by watching %WORKER_REBIND.
1562 */
1563 init_completion(&idle_rebind.done);
1564retry:
1565 idle_rebind.cnt = 1;
1566 INIT_COMPLETION(idle_rebind.done);
1567
1568 /* set REBIND and kick idle ones, we'll wait for these later */
1569 for_each_worker_pool(pool, gcwq) {
1570 list_for_each_entry(worker, &pool->idle_list, entry) {
1571 if (worker->flags & WORKER_REBIND)
1572 continue;
1573
1574 /* morph UNBOUND to REBIND */
1575 worker->flags &= ~WORKER_UNBOUND;
1576 worker->flags |= WORKER_REBIND;
1577
1578 idle_rebind.cnt++;
1579 worker->idle_rebind = &idle_rebind;
1580
1581 /* worker_thread() will call idle_worker_rebind() */
1582 wake_up_process(worker->task);
1583 }
1584 }
1585
1586 if (--idle_rebind.cnt) {
1587 spin_unlock_irq(&gcwq->lock);
1588 wait_for_completion(&idle_rebind.done);
1589 spin_lock_irq(&gcwq->lock);
1590 /* busy ones might have become idle while waiting, retry */
1591 goto retry;
1592 }
1593
1594 /*
1595 * All idle workers are rebound and waiting for %WORKER_REBIND to
1596 * be cleared inside idle_worker_rebind(). Clear and release.
1597 * Clearing %WORKER_REBIND from this foreign context is safe
1598 * because these workers are still guaranteed to be idle.
1599 */
1600 for_each_worker_pool(pool, gcwq)
1601 list_for_each_entry(worker, &pool->idle_list, entry)
1602 worker->flags &= ~WORKER_REBIND;
1603
1604 wake_up_all(&gcwq->rebind_hold);
1605
1606 /* rebind busy workers */
1607 for_each_busy_worker(worker, i, pos, gcwq) {
1608 struct work_struct *rebind_work = &worker->rebind_work;
1609
1610 /* morph UNBOUND to REBIND */
1611 worker->flags &= ~WORKER_UNBOUND;
1612 worker->flags |= WORKER_REBIND;
1613
1614 if (test_and_set_bit(WORK_STRUCT_PENDING_BIT,
1615 work_data_bits(rebind_work)))
1616 continue;
1617
1618 /* wq doesn't matter, use the default one */
1619 debug_work_activate(rebind_work);
1620 insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work,
1621 worker->scheduled.next,
1622 work_color_to_flags(WORK_NO_COLOR));
1623 }
1624}
1625
c34056a3
TH
1626static struct worker *alloc_worker(void)
1627{
1628 struct worker *worker;
1629
1630 worker = kzalloc(sizeof(*worker), GFP_KERNEL);
c8e55f36
TH
1631 if (worker) {
1632 INIT_LIST_HEAD(&worker->entry);
affee4b2 1633 INIT_LIST_HEAD(&worker->scheduled);
25511a47 1634 INIT_WORK(&worker->rebind_work, busy_worker_rebind_fn);
e22bee78
TH
1635 /* on creation a worker is in !idle && prep state */
1636 worker->flags = WORKER_PREP;
c8e55f36 1637 }
c34056a3
TH
1638 return worker;
1639}
1640
1641/**
1642 * create_worker - create a new workqueue worker
63d95a91 1643 * @pool: pool the new worker will belong to
c34056a3 1644 *
63d95a91 1645 * Create a new worker which is bound to @pool. The returned worker
c34056a3
TH
1646 * can be started by calling start_worker() or destroyed using
1647 * destroy_worker().
1648 *
1649 * CONTEXT:
1650 * Might sleep. Does GFP_KERNEL allocations.
1651 *
1652 * RETURNS:
1653 * Pointer to the newly created worker.
1654 */
bc2ae0f5 1655static struct worker *create_worker(struct worker_pool *pool)
c34056a3 1656{
63d95a91 1657 struct global_cwq *gcwq = pool->gcwq;
3270476a 1658 const char *pri = worker_pool_pri(pool) ? "H" : "";
c34056a3 1659 struct worker *worker = NULL;
f3421797 1660 int id = -1;
c34056a3 1661
8b03ae3c 1662 spin_lock_irq(&gcwq->lock);
bd7bdd43 1663 while (ida_get_new(&pool->worker_ida, &id)) {
8b03ae3c 1664 spin_unlock_irq(&gcwq->lock);
bd7bdd43 1665 if (!ida_pre_get(&pool->worker_ida, GFP_KERNEL))
c34056a3 1666 goto fail;
8b03ae3c 1667 spin_lock_irq(&gcwq->lock);
c34056a3 1668 }
8b03ae3c 1669 spin_unlock_irq(&gcwq->lock);
c34056a3
TH
1670
1671 worker = alloc_worker();
1672 if (!worker)
1673 goto fail;
1674
bd7bdd43 1675 worker->pool = pool;
c34056a3
TH
1676 worker->id = id;
1677
bc2ae0f5 1678 if (gcwq->cpu != WORK_CPU_UNBOUND)
94dcf29a 1679 worker->task = kthread_create_on_node(worker_thread,
3270476a
TH
1680 worker, cpu_to_node(gcwq->cpu),
1681 "kworker/%u:%d%s", gcwq->cpu, id, pri);
f3421797
TH
1682 else
1683 worker->task = kthread_create(worker_thread, worker,
3270476a 1684 "kworker/u:%d%s", id, pri);
c34056a3
TH
1685 if (IS_ERR(worker->task))
1686 goto fail;
1687
3270476a
TH
1688 if (worker_pool_pri(pool))
1689 set_user_nice(worker->task, HIGHPRI_NICE_LEVEL);
1690
db7bccf4 1691 /*
bc2ae0f5
TH
1692 * Determine CPU binding of the new worker depending on
1693 * %GCWQ_DISASSOCIATED. The caller is responsible for ensuring the
1694 * flag remains stable across this function. See the comments
1695 * above the flag definition for details.
1696 *
1697 * As an unbound worker may later become a regular one if CPU comes
1698 * online, make sure every worker has %PF_THREAD_BOUND set.
db7bccf4 1699 */
bc2ae0f5 1700 if (!(gcwq->flags & GCWQ_DISASSOCIATED)) {
8b03ae3c 1701 kthread_bind(worker->task, gcwq->cpu);
bc2ae0f5 1702 } else {
db7bccf4 1703 worker->task->flags |= PF_THREAD_BOUND;
bc2ae0f5 1704 worker->flags |= WORKER_UNBOUND;
f3421797 1705 }
c34056a3
TH
1706
1707 return worker;
1708fail:
1709 if (id >= 0) {
8b03ae3c 1710 spin_lock_irq(&gcwq->lock);
bd7bdd43 1711 ida_remove(&pool->worker_ida, id);
8b03ae3c 1712 spin_unlock_irq(&gcwq->lock);
c34056a3
TH
1713 }
1714 kfree(worker);
1715 return NULL;
1716}
1717
1718/**
1719 * start_worker - start a newly created worker
1720 * @worker: worker to start
1721 *
c8e55f36 1722 * Make the gcwq aware of @worker and start it.
c34056a3
TH
1723 *
1724 * CONTEXT:
8b03ae3c 1725 * spin_lock_irq(gcwq->lock).
c34056a3
TH
1726 */
1727static void start_worker(struct worker *worker)
1728{
cb444766 1729 worker->flags |= WORKER_STARTED;
bd7bdd43 1730 worker->pool->nr_workers++;
c8e55f36 1731 worker_enter_idle(worker);
c34056a3
TH
1732 wake_up_process(worker->task);
1733}
1734
1735/**
1736 * destroy_worker - destroy a workqueue worker
1737 * @worker: worker to be destroyed
1738 *
c8e55f36
TH
1739 * Destroy @worker and adjust @gcwq stats accordingly.
1740 *
1741 * CONTEXT:
1742 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
c34056a3
TH
1743 */
1744static void destroy_worker(struct worker *worker)
1745{
bd7bdd43
TH
1746 struct worker_pool *pool = worker->pool;
1747 struct global_cwq *gcwq = pool->gcwq;
c34056a3
TH
1748 int id = worker->id;
1749
1750 /* sanity check frenzy */
1751 BUG_ON(worker->current_work);
affee4b2 1752 BUG_ON(!list_empty(&worker->scheduled));
c34056a3 1753
c8e55f36 1754 if (worker->flags & WORKER_STARTED)
bd7bdd43 1755 pool->nr_workers--;
c8e55f36 1756 if (worker->flags & WORKER_IDLE)
bd7bdd43 1757 pool->nr_idle--;
c8e55f36
TH
1758
1759 list_del_init(&worker->entry);
cb444766 1760 worker->flags |= WORKER_DIE;
c8e55f36
TH
1761
1762 spin_unlock_irq(&gcwq->lock);
1763
c34056a3
TH
1764 kthread_stop(worker->task);
1765 kfree(worker);
1766
8b03ae3c 1767 spin_lock_irq(&gcwq->lock);
bd7bdd43 1768 ida_remove(&pool->worker_ida, id);
c34056a3
TH
1769}
1770
63d95a91 1771static void idle_worker_timeout(unsigned long __pool)
e22bee78 1772{
63d95a91
TH
1773 struct worker_pool *pool = (void *)__pool;
1774 struct global_cwq *gcwq = pool->gcwq;
e22bee78
TH
1775
1776 spin_lock_irq(&gcwq->lock);
1777
63d95a91 1778 if (too_many_workers(pool)) {
e22bee78
TH
1779 struct worker *worker;
1780 unsigned long expires;
1781
1782 /* idle_list is kept in LIFO order, check the last one */
63d95a91 1783 worker = list_entry(pool->idle_list.prev, struct worker, entry);
e22bee78
TH
1784 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
1785
1786 if (time_before(jiffies, expires))
63d95a91 1787 mod_timer(&pool->idle_timer, expires);
e22bee78
TH
1788 else {
1789 /* it's been idle for too long, wake up manager */
11ebea50 1790 pool->flags |= POOL_MANAGE_WORKERS;
63d95a91 1791 wake_up_worker(pool);
d5abe669 1792 }
e22bee78
TH
1793 }
1794
1795 spin_unlock_irq(&gcwq->lock);
1796}
d5abe669 1797
e22bee78
TH
1798static bool send_mayday(struct work_struct *work)
1799{
1800 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
1801 struct workqueue_struct *wq = cwq->wq;
f3421797 1802 unsigned int cpu;
e22bee78
TH
1803
1804 if (!(wq->flags & WQ_RESCUER))
1805 return false;
1806
1807 /* mayday mayday mayday */
bd7bdd43 1808 cpu = cwq->pool->gcwq->cpu;
f3421797
TH
1809 /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */
1810 if (cpu == WORK_CPU_UNBOUND)
1811 cpu = 0;
f2e005aa 1812 if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask))
e22bee78
TH
1813 wake_up_process(wq->rescuer->task);
1814 return true;
1815}
1816
63d95a91 1817static void gcwq_mayday_timeout(unsigned long __pool)
e22bee78 1818{
63d95a91
TH
1819 struct worker_pool *pool = (void *)__pool;
1820 struct global_cwq *gcwq = pool->gcwq;
e22bee78
TH
1821 struct work_struct *work;
1822
1823 spin_lock_irq(&gcwq->lock);
1824
63d95a91 1825 if (need_to_create_worker(pool)) {
e22bee78
TH
1826 /*
1827 * We've been trying to create a new worker but
1828 * haven't been successful. We might be hitting an
1829 * allocation deadlock. Send distress signals to
1830 * rescuers.
1831 */
63d95a91 1832 list_for_each_entry(work, &pool->worklist, entry)
e22bee78 1833 send_mayday(work);
1da177e4 1834 }
e22bee78
TH
1835
1836 spin_unlock_irq(&gcwq->lock);
1837
63d95a91 1838 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INTERVAL);
1da177e4
LT
1839}
1840
e22bee78
TH
1841/**
1842 * maybe_create_worker - create a new worker if necessary
63d95a91 1843 * @pool: pool to create a new worker for
e22bee78 1844 *
63d95a91 1845 * Create a new worker for @pool if necessary. @pool is guaranteed to
e22bee78
TH
1846 * have at least one idle worker on return from this function. If
1847 * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is
63d95a91 1848 * sent to all rescuers with works scheduled on @pool to resolve
e22bee78
TH
1849 * possible allocation deadlock.
1850 *
1851 * On return, need_to_create_worker() is guaranteed to be false and
1852 * may_start_working() true.
1853 *
1854 * LOCKING:
1855 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1856 * multiple times. Does GFP_KERNEL allocations. Called only from
1857 * manager.
1858 *
1859 * RETURNS:
1860 * false if no action was taken and gcwq->lock stayed locked, true
1861 * otherwise.
1862 */
63d95a91 1863static bool maybe_create_worker(struct worker_pool *pool)
06bd6ebf
NK
1864__releases(&gcwq->lock)
1865__acquires(&gcwq->lock)
1da177e4 1866{
63d95a91
TH
1867 struct global_cwq *gcwq = pool->gcwq;
1868
1869 if (!need_to_create_worker(pool))
e22bee78
TH
1870 return false;
1871restart:
9f9c2364
TH
1872 spin_unlock_irq(&gcwq->lock);
1873
e22bee78 1874 /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */
63d95a91 1875 mod_timer(&pool->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT);
e22bee78
TH
1876
1877 while (true) {
1878 struct worker *worker;
1879
bc2ae0f5 1880 worker = create_worker(pool);
e22bee78 1881 if (worker) {
63d95a91 1882 del_timer_sync(&pool->mayday_timer);
e22bee78
TH
1883 spin_lock_irq(&gcwq->lock);
1884 start_worker(worker);
63d95a91 1885 BUG_ON(need_to_create_worker(pool));
e22bee78
TH
1886 return true;
1887 }
1888
63d95a91 1889 if (!need_to_create_worker(pool))
e22bee78 1890 break;
1da177e4 1891
e22bee78
TH
1892 __set_current_state(TASK_INTERRUPTIBLE);
1893 schedule_timeout(CREATE_COOLDOWN);
9f9c2364 1894
63d95a91 1895 if (!need_to_create_worker(pool))
e22bee78
TH
1896 break;
1897 }
1898
63d95a91 1899 del_timer_sync(&pool->mayday_timer);
e22bee78 1900 spin_lock_irq(&gcwq->lock);
63d95a91 1901 if (need_to_create_worker(pool))
e22bee78
TH
1902 goto restart;
1903 return true;
1904}
1905
1906/**
1907 * maybe_destroy_worker - destroy workers which have been idle for a while
63d95a91 1908 * @pool: pool to destroy workers for
e22bee78 1909 *
63d95a91 1910 * Destroy @pool workers which have been idle for longer than
e22bee78
TH
1911 * IDLE_WORKER_TIMEOUT.
1912 *
1913 * LOCKING:
1914 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1915 * multiple times. Called only from manager.
1916 *
1917 * RETURNS:
1918 * false if no action was taken and gcwq->lock stayed locked, true
1919 * otherwise.
1920 */
63d95a91 1921static bool maybe_destroy_workers(struct worker_pool *pool)
e22bee78
TH
1922{
1923 bool ret = false;
1da177e4 1924
63d95a91 1925 while (too_many_workers(pool)) {
e22bee78
TH
1926 struct worker *worker;
1927 unsigned long expires;
3af24433 1928
63d95a91 1929 worker = list_entry(pool->idle_list.prev, struct worker, entry);
e22bee78 1930 expires = worker->last_active + IDLE_WORKER_TIMEOUT;
85f4186a 1931
e22bee78 1932 if (time_before(jiffies, expires)) {
63d95a91 1933 mod_timer(&pool->idle_timer, expires);
3af24433 1934 break;
e22bee78 1935 }
1da177e4 1936
e22bee78
TH
1937 destroy_worker(worker);
1938 ret = true;
1da177e4 1939 }
3af24433 1940
e22bee78
TH
1941 return ret;
1942}
1943
1944/**
1945 * manage_workers - manage worker pool
1946 * @worker: self
1947 *
1948 * Assume the manager role and manage gcwq worker pool @worker belongs
1949 * to. At any given time, there can be only zero or one manager per
1950 * gcwq. The exclusion is handled automatically by this function.
1951 *
1952 * The caller can safely start processing works on false return. On
1953 * true return, it's guaranteed that need_to_create_worker() is false
1954 * and may_start_working() is true.
1955 *
1956 * CONTEXT:
1957 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
1958 * multiple times. Does GFP_KERNEL allocations.
1959 *
1960 * RETURNS:
1961 * false if no action was taken and gcwq->lock stayed locked, true if
1962 * some action was taken.
1963 */
1964static bool manage_workers(struct worker *worker)
1965{
63d95a91 1966 struct worker_pool *pool = worker->pool;
e22bee78
TH
1967 bool ret = false;
1968
60373152 1969 if (!mutex_trylock(&pool->manager_mutex))
e22bee78
TH
1970 return ret;
1971
11ebea50 1972 pool->flags &= ~POOL_MANAGE_WORKERS;
e22bee78
TH
1973
1974 /*
1975 * Destroy and then create so that may_start_working() is true
1976 * on return.
1977 */
63d95a91
TH
1978 ret |= maybe_destroy_workers(pool);
1979 ret |= maybe_create_worker(pool);
e22bee78 1980
60373152 1981 mutex_unlock(&pool->manager_mutex);
e22bee78
TH
1982 return ret;
1983}
1984
a62428c0
TH
1985/**
1986 * process_one_work - process single work
c34056a3 1987 * @worker: self
a62428c0
TH
1988 * @work: work to process
1989 *
1990 * Process @work. This function contains all the logics necessary to
1991 * process a single work including synchronization against and
1992 * interaction with other workers on the same cpu, queueing and
1993 * flushing. As long as context requirement is met, any worker can
1994 * call this function to process a work.
1995 *
1996 * CONTEXT:
8b03ae3c 1997 * spin_lock_irq(gcwq->lock) which is released and regrabbed.
a62428c0 1998 */
c34056a3 1999static void process_one_work(struct worker *worker, struct work_struct *work)
06bd6ebf
NK
2000__releases(&gcwq->lock)
2001__acquires(&gcwq->lock)
a62428c0 2002{
7e11629d 2003 struct cpu_workqueue_struct *cwq = get_work_cwq(work);
bd7bdd43
TH
2004 struct worker_pool *pool = worker->pool;
2005 struct global_cwq *gcwq = pool->gcwq;
c8e55f36 2006 struct hlist_head *bwh = busy_worker_head(gcwq, work);
fb0e7beb 2007 bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE;
a62428c0 2008 work_func_t f = work->func;
73f53c4a 2009 int work_color;
7e11629d 2010 struct worker *collision;
a62428c0
TH
2011#ifdef CONFIG_LOCKDEP
2012 /*
2013 * It is permissible to free the struct work_struct from
2014 * inside the function that is called from it, this we need to
2015 * take into account for lockdep too. To avoid bogus "held
2016 * lock freed" warnings as well as problems when looking into
2017 * work->lockdep_map, make a copy and use that here.
2018 */
4d82a1de
PZ
2019 struct lockdep_map lockdep_map;
2020
2021 lockdep_copy_map(&lockdep_map, &work->lockdep_map);
a62428c0 2022#endif
6fec10a1
TH
2023 /*
2024 * Ensure we're on the correct CPU. DISASSOCIATED test is
2025 * necessary to avoid spurious warnings from rescuers servicing the
2026 * unbound or a disassociated gcwq.
2027 */
25511a47 2028 WARN_ON_ONCE(!(worker->flags & (WORKER_UNBOUND | WORKER_REBIND)) &&
6fec10a1 2029 !(gcwq->flags & GCWQ_DISASSOCIATED) &&
25511a47
TH
2030 raw_smp_processor_id() != gcwq->cpu);
2031
7e11629d
TH
2032 /*
2033 * A single work shouldn't be executed concurrently by
2034 * multiple workers on a single cpu. Check whether anyone is
2035 * already processing the work. If so, defer the work to the
2036 * currently executing one.
2037 */
2038 collision = __find_worker_executing_work(gcwq, bwh, work);
2039 if (unlikely(collision)) {
2040 move_linked_works(work, &collision->scheduled, NULL);
2041 return;
2042 }
2043
8930caba 2044 /* claim and dequeue */
a62428c0 2045 debug_work_deactivate(work);
c8e55f36 2046 hlist_add_head(&worker->hentry, bwh);
c34056a3 2047 worker->current_work = work;
8cca0eea 2048 worker->current_cwq = cwq;
73f53c4a 2049 work_color = get_work_color(work);
7a22ad75 2050
a62428c0
TH
2051 list_del_init(&work->entry);
2052
fb0e7beb
TH
2053 /*
2054 * CPU intensive works don't participate in concurrency
2055 * management. They're the scheduler's responsibility.
2056 */
2057 if (unlikely(cpu_intensive))
2058 worker_set_flags(worker, WORKER_CPU_INTENSIVE, true);
2059
974271c4
TH
2060 /*
2061 * Unbound gcwq isn't concurrency managed and work items should be
2062 * executed ASAP. Wake up another worker if necessary.
2063 */
63d95a91
TH
2064 if ((worker->flags & WORKER_UNBOUND) && need_more_worker(pool))
2065 wake_up_worker(pool);
974271c4 2066
8930caba
TH
2067 /*
2068 * Record the last CPU and clear PENDING. The following wmb is
2069 * paired with the implied mb in test_and_set_bit(PENDING) and
2070 * ensures all updates to @work made here are visible to and
2071 * precede any updates by the next PENDING owner. Also, clear
2072 * PENDING inside @gcwq->lock so that PENDING and queued state
2073 * changes happen together while IRQ is disabled.
2074 */
2075 smp_wmb();
2076 set_work_cpu_and_clear_pending(work, gcwq->cpu);
a62428c0 2077
8930caba 2078 spin_unlock_irq(&gcwq->lock);
959d1af8 2079
e159489b 2080 lock_map_acquire_read(&cwq->wq->lockdep_map);
a62428c0 2081 lock_map_acquire(&lockdep_map);
e36c886a 2082 trace_workqueue_execute_start(work);
a62428c0 2083 f(work);
e36c886a
AV
2084 /*
2085 * While we must be careful to not use "work" after this, the trace
2086 * point will only record its address.
2087 */
2088 trace_workqueue_execute_end(work);
a62428c0
TH
2089 lock_map_release(&lockdep_map);
2090 lock_map_release(&cwq->wq->lockdep_map);
2091
2092 if (unlikely(in_atomic() || lockdep_depth(current) > 0)) {
2093 printk(KERN_ERR "BUG: workqueue leaked lock or atomic: "
2094 "%s/0x%08x/%d\n",
2095 current->comm, preempt_count(), task_pid_nr(current));
2096 printk(KERN_ERR " last function: ");
2097 print_symbol("%s\n", (unsigned long)f);
2098 debug_show_held_locks(current);
2099 dump_stack();
2100 }
2101
8b03ae3c 2102 spin_lock_irq(&gcwq->lock);
a62428c0 2103
fb0e7beb
TH
2104 /* clear cpu intensive status */
2105 if (unlikely(cpu_intensive))
2106 worker_clr_flags(worker, WORKER_CPU_INTENSIVE);
2107
a62428c0 2108 /* we're done with it, release */
c8e55f36 2109 hlist_del_init(&worker->hentry);
c34056a3 2110 worker->current_work = NULL;
8cca0eea 2111 worker->current_cwq = NULL;
8a2e8e5d 2112 cwq_dec_nr_in_flight(cwq, work_color, false);
a62428c0
TH
2113}
2114
affee4b2
TH
2115/**
2116 * process_scheduled_works - process scheduled works
2117 * @worker: self
2118 *
2119 * Process all scheduled works. Please note that the scheduled list
2120 * may change while processing a work, so this function repeatedly
2121 * fetches a work from the top and executes it.
2122 *
2123 * CONTEXT:
8b03ae3c 2124 * spin_lock_irq(gcwq->lock) which may be released and regrabbed
affee4b2
TH
2125 * multiple times.
2126 */
2127static void process_scheduled_works(struct worker *worker)
1da177e4 2128{
affee4b2
TH
2129 while (!list_empty(&worker->scheduled)) {
2130 struct work_struct *work = list_first_entry(&worker->scheduled,
1da177e4 2131 struct work_struct, entry);
c34056a3 2132 process_one_work(worker, work);
1da177e4 2133 }
1da177e4
LT
2134}
2135
4690c4ab
TH
2136/**
2137 * worker_thread - the worker thread function
c34056a3 2138 * @__worker: self
4690c4ab 2139 *
e22bee78
TH
2140 * The gcwq worker thread function. There's a single dynamic pool of
2141 * these per each cpu. These workers process all works regardless of
2142 * their specific target workqueue. The only exception is works which
2143 * belong to workqueues with a rescuer which will be explained in
2144 * rescuer_thread().
4690c4ab 2145 */
c34056a3 2146static int worker_thread(void *__worker)
1da177e4 2147{
c34056a3 2148 struct worker *worker = __worker;
bd7bdd43
TH
2149 struct worker_pool *pool = worker->pool;
2150 struct global_cwq *gcwq = pool->gcwq;
1da177e4 2151
e22bee78
TH
2152 /* tell the scheduler that this is a workqueue worker */
2153 worker->task->flags |= PF_WQ_WORKER;
c8e55f36 2154woke_up:
c8e55f36 2155 spin_lock_irq(&gcwq->lock);
1da177e4 2156
25511a47
TH
2157 /*
2158 * DIE can be set only while idle and REBIND set while busy has
2159 * @worker->rebind_work scheduled. Checking here is enough.
2160 */
2161 if (unlikely(worker->flags & (WORKER_REBIND | WORKER_DIE))) {
c8e55f36 2162 spin_unlock_irq(&gcwq->lock);
25511a47
TH
2163
2164 if (worker->flags & WORKER_DIE) {
2165 worker->task->flags &= ~PF_WQ_WORKER;
2166 return 0;
2167 }
2168
2169 idle_worker_rebind(worker);
2170 goto woke_up;
c8e55f36 2171 }
affee4b2 2172
c8e55f36 2173 worker_leave_idle(worker);
db7bccf4 2174recheck:
e22bee78 2175 /* no more worker necessary? */
63d95a91 2176 if (!need_more_worker(pool))
e22bee78
TH
2177 goto sleep;
2178
2179 /* do we need to manage? */
63d95a91 2180 if (unlikely(!may_start_working(pool)) && manage_workers(worker))
e22bee78
TH
2181 goto recheck;
2182
c8e55f36
TH
2183 /*
2184 * ->scheduled list can only be filled while a worker is
2185 * preparing to process a work or actually processing it.
2186 * Make sure nobody diddled with it while I was sleeping.
2187 */
2188 BUG_ON(!list_empty(&worker->scheduled));
2189
e22bee78
TH
2190 /*
2191 * When control reaches this point, we're guaranteed to have
2192 * at least one idle worker or that someone else has already
2193 * assumed the manager role.
2194 */
2195 worker_clr_flags(worker, WORKER_PREP);
2196
2197 do {
c8e55f36 2198 struct work_struct *work =
bd7bdd43 2199 list_first_entry(&pool->worklist,
c8e55f36
TH
2200 struct work_struct, entry);
2201
2202 if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) {
2203 /* optimization path, not strictly necessary */
2204 process_one_work(worker, work);
2205 if (unlikely(!list_empty(&worker->scheduled)))
affee4b2 2206 process_scheduled_works(worker);
c8e55f36
TH
2207 } else {
2208 move_linked_works(work, &worker->scheduled, NULL);
2209 process_scheduled_works(worker);
affee4b2 2210 }
63d95a91 2211 } while (keep_working(pool));
e22bee78
TH
2212
2213 worker_set_flags(worker, WORKER_PREP, false);
d313dd85 2214sleep:
63d95a91 2215 if (unlikely(need_to_manage_workers(pool)) && manage_workers(worker))
e22bee78 2216 goto recheck;
d313dd85 2217
c8e55f36 2218 /*
e22bee78
TH
2219 * gcwq->lock is held and there's no work to process and no
2220 * need to manage, sleep. Workers are woken up only while
2221 * holding gcwq->lock or from local cpu, so setting the
2222 * current state before releasing gcwq->lock is enough to
2223 * prevent losing any event.
c8e55f36
TH
2224 */
2225 worker_enter_idle(worker);
2226 __set_current_state(TASK_INTERRUPTIBLE);
2227 spin_unlock_irq(&gcwq->lock);
2228 schedule();
2229 goto woke_up;
1da177e4
LT
2230}
2231
e22bee78
TH
2232/**
2233 * rescuer_thread - the rescuer thread function
2234 * @__wq: the associated workqueue
2235 *
2236 * Workqueue rescuer thread function. There's one rescuer for each
2237 * workqueue which has WQ_RESCUER set.
2238 *
2239 * Regular work processing on a gcwq may block trying to create a new
2240 * worker which uses GFP_KERNEL allocation which has slight chance of
2241 * developing into deadlock if some works currently on the same queue
2242 * need to be processed to satisfy the GFP_KERNEL allocation. This is
2243 * the problem rescuer solves.
2244 *
2245 * When such condition is possible, the gcwq summons rescuers of all
2246 * workqueues which have works queued on the gcwq and let them process
2247 * those works so that forward progress can be guaranteed.
2248 *
2249 * This should happen rarely.
2250 */
2251static int rescuer_thread(void *__wq)
2252{
2253 struct workqueue_struct *wq = __wq;
2254 struct worker *rescuer = wq->rescuer;
2255 struct list_head *scheduled = &rescuer->scheduled;
f3421797 2256 bool is_unbound = wq->flags & WQ_UNBOUND;
e22bee78
TH
2257 unsigned int cpu;
2258
2259 set_user_nice(current, RESCUER_NICE_LEVEL);
2260repeat:
2261 set_current_state(TASK_INTERRUPTIBLE);
2262
2263 if (kthread_should_stop())
2264 return 0;
2265
f3421797
TH
2266 /*
2267 * See whether any cpu is asking for help. Unbounded
2268 * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND.
2269 */
f2e005aa 2270 for_each_mayday_cpu(cpu, wq->mayday_mask) {
f3421797
TH
2271 unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu;
2272 struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq);
bd7bdd43
TH
2273 struct worker_pool *pool = cwq->pool;
2274 struct global_cwq *gcwq = pool->gcwq;
e22bee78
TH
2275 struct work_struct *work, *n;
2276
2277 __set_current_state(TASK_RUNNING);
f2e005aa 2278 mayday_clear_cpu(cpu, wq->mayday_mask);
e22bee78
TH
2279
2280 /* migrate to the target cpu if possible */
bd7bdd43 2281 rescuer->pool = pool;
e22bee78
TH
2282 worker_maybe_bind_and_lock(rescuer);
2283
2284 /*
2285 * Slurp in all works issued via this workqueue and
2286 * process'em.
2287 */
2288 BUG_ON(!list_empty(&rescuer->scheduled));
bd7bdd43 2289 list_for_each_entry_safe(work, n, &pool->worklist, entry)
e22bee78
TH
2290 if (get_work_cwq(work) == cwq)
2291 move_linked_works(work, scheduled, &n);
2292
2293 process_scheduled_works(rescuer);
7576958a
TH
2294
2295 /*
2296 * Leave this gcwq. If keep_working() is %true, notify a
2297 * regular worker; otherwise, we end up with 0 concurrency
2298 * and stalling the execution.
2299 */
63d95a91
TH
2300 if (keep_working(pool))
2301 wake_up_worker(pool);
7576958a 2302
e22bee78
TH
2303 spin_unlock_irq(&gcwq->lock);
2304 }
2305
2306 schedule();
2307 goto repeat;
1da177e4
LT
2308}
2309
fc2e4d70
ON
2310struct wq_barrier {
2311 struct work_struct work;
2312 struct completion done;
2313};
2314
2315static void wq_barrier_func(struct work_struct *work)
2316{
2317 struct wq_barrier *barr = container_of(work, struct wq_barrier, work);
2318 complete(&barr->done);
2319}
2320
4690c4ab
TH
2321/**
2322 * insert_wq_barrier - insert a barrier work
2323 * @cwq: cwq to insert barrier into
2324 * @barr: wq_barrier to insert
affee4b2
TH
2325 * @target: target work to attach @barr to
2326 * @worker: worker currently executing @target, NULL if @target is not executing
4690c4ab 2327 *
affee4b2
TH
2328 * @barr is linked to @target such that @barr is completed only after
2329 * @target finishes execution. Please note that the ordering
2330 * guarantee is observed only with respect to @target and on the local
2331 * cpu.
2332 *
2333 * Currently, a queued barrier can't be canceled. This is because
2334 * try_to_grab_pending() can't determine whether the work to be
2335 * grabbed is at the head of the queue and thus can't clear LINKED
2336 * flag of the previous work while there must be a valid next work
2337 * after a work with LINKED flag set.
2338 *
2339 * Note that when @worker is non-NULL, @target may be modified
2340 * underneath us, so we can't reliably determine cwq from @target.
4690c4ab
TH
2341 *
2342 * CONTEXT:
8b03ae3c 2343 * spin_lock_irq(gcwq->lock).
4690c4ab 2344 */
83c22520 2345static void insert_wq_barrier(struct cpu_workqueue_struct *cwq,
affee4b2
TH
2346 struct wq_barrier *barr,
2347 struct work_struct *target, struct worker *worker)
fc2e4d70 2348{
affee4b2
TH
2349 struct list_head *head;
2350 unsigned int linked = 0;
2351
dc186ad7 2352 /*
8b03ae3c 2353 * debugobject calls are safe here even with gcwq->lock locked
dc186ad7
TG
2354 * as we know for sure that this will not trigger any of the
2355 * checks and call back into the fixup functions where we
2356 * might deadlock.
2357 */
ca1cab37 2358 INIT_WORK_ONSTACK(&barr->work, wq_barrier_func);
22df02bb 2359 __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work));
fc2e4d70 2360 init_completion(&barr->done);
83c22520 2361
affee4b2
TH
2362 /*
2363 * If @target is currently being executed, schedule the
2364 * barrier to the worker; otherwise, put it after @target.
2365 */
2366 if (worker)
2367 head = worker->scheduled.next;
2368 else {
2369 unsigned long *bits = work_data_bits(target);
2370
2371 head = target->entry.next;
2372 /* there can already be other linked works, inherit and set */
2373 linked = *bits & WORK_STRUCT_LINKED;
2374 __set_bit(WORK_STRUCT_LINKED_BIT, bits);
2375 }
2376
dc186ad7 2377 debug_work_activate(&barr->work);
affee4b2
TH
2378 insert_work(cwq, &barr->work, head,
2379 work_color_to_flags(WORK_NO_COLOR) | linked);
fc2e4d70
ON
2380}
2381
73f53c4a
TH
2382/**
2383 * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing
2384 * @wq: workqueue being flushed
2385 * @flush_color: new flush color, < 0 for no-op
2386 * @work_color: new work color, < 0 for no-op
2387 *
2388 * Prepare cwqs for workqueue flushing.
2389 *
2390 * If @flush_color is non-negative, flush_color on all cwqs should be
2391 * -1. If no cwq has in-flight commands at the specified color, all
2392 * cwq->flush_color's stay at -1 and %false is returned. If any cwq
2393 * has in flight commands, its cwq->flush_color is set to
2394 * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq
2395 * wakeup logic is armed and %true is returned.
2396 *
2397 * The caller should have initialized @wq->first_flusher prior to
2398 * calling this function with non-negative @flush_color. If
2399 * @flush_color is negative, no flush color update is done and %false
2400 * is returned.
2401 *
2402 * If @work_color is non-negative, all cwqs should have the same
2403 * work_color which is previous to @work_color and all will be
2404 * advanced to @work_color.
2405 *
2406 * CONTEXT:
2407 * mutex_lock(wq->flush_mutex).
2408 *
2409 * RETURNS:
2410 * %true if @flush_color >= 0 and there's something to flush. %false
2411 * otherwise.
2412 */
2413static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq,
2414 int flush_color, int work_color)
1da177e4 2415{
73f53c4a
TH
2416 bool wait = false;
2417 unsigned int cpu;
1da177e4 2418
73f53c4a
TH
2419 if (flush_color >= 0) {
2420 BUG_ON(atomic_read(&wq->nr_cwqs_to_flush));
2421 atomic_set(&wq->nr_cwqs_to_flush, 1);
1da177e4 2422 }
2355b70f 2423
f3421797 2424 for_each_cwq_cpu(cpu, wq) {
73f53c4a 2425 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
bd7bdd43 2426 struct global_cwq *gcwq = cwq->pool->gcwq;
fc2e4d70 2427
8b03ae3c 2428 spin_lock_irq(&gcwq->lock);
83c22520 2429
73f53c4a
TH
2430 if (flush_color >= 0) {
2431 BUG_ON(cwq->flush_color != -1);
fc2e4d70 2432
73f53c4a
TH
2433 if (cwq->nr_in_flight[flush_color]) {
2434 cwq->flush_color = flush_color;
2435 atomic_inc(&wq->nr_cwqs_to_flush);
2436 wait = true;
2437 }
2438 }
1da177e4 2439
73f53c4a
TH
2440 if (work_color >= 0) {
2441 BUG_ON(work_color != work_next_color(cwq->work_color));
2442 cwq->work_color = work_color;
2443 }
1da177e4 2444
8b03ae3c 2445 spin_unlock_irq(&gcwq->lock);
1da177e4 2446 }
2355b70f 2447
73f53c4a
TH
2448 if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush))
2449 complete(&wq->first_flusher->done);
14441960 2450
73f53c4a 2451 return wait;
1da177e4
LT
2452}
2453
0fcb78c2 2454/**
1da177e4 2455 * flush_workqueue - ensure that any scheduled work has run to completion.
0fcb78c2 2456 * @wq: workqueue to flush
1da177e4
LT
2457 *
2458 * Forces execution of the workqueue and blocks until its completion.
2459 * This is typically used in driver shutdown handlers.
2460 *
fc2e4d70
ON
2461 * We sleep until all works which were queued on entry have been handled,
2462 * but we are not livelocked by new incoming ones.
1da177e4 2463 */
7ad5b3a5 2464void flush_workqueue(struct workqueue_struct *wq)
1da177e4 2465{
73f53c4a
TH
2466 struct wq_flusher this_flusher = {
2467 .list = LIST_HEAD_INIT(this_flusher.list),
2468 .flush_color = -1,
2469 .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done),
2470 };
2471 int next_color;
1da177e4 2472
3295f0ef
IM
2473 lock_map_acquire(&wq->lockdep_map);
2474 lock_map_release(&wq->lockdep_map);
73f53c4a
TH
2475
2476 mutex_lock(&wq->flush_mutex);
2477
2478 /*
2479 * Start-to-wait phase
2480 */
2481 next_color = work_next_color(wq->work_color);
2482
2483 if (next_color != wq->flush_color) {
2484 /*
2485 * Color space is not full. The current work_color
2486 * becomes our flush_color and work_color is advanced
2487 * by one.
2488 */
2489 BUG_ON(!list_empty(&wq->flusher_overflow));
2490 this_flusher.flush_color = wq->work_color;
2491 wq->work_color = next_color;
2492
2493 if (!wq->first_flusher) {
2494 /* no flush in progress, become the first flusher */
2495 BUG_ON(wq->flush_color != this_flusher.flush_color);
2496
2497 wq->first_flusher = &this_flusher;
2498
2499 if (!flush_workqueue_prep_cwqs(wq, wq->flush_color,
2500 wq->work_color)) {
2501 /* nothing to flush, done */
2502 wq->flush_color = next_color;
2503 wq->first_flusher = NULL;
2504 goto out_unlock;
2505 }
2506 } else {
2507 /* wait in queue */
2508 BUG_ON(wq->flush_color == this_flusher.flush_color);
2509 list_add_tail(&this_flusher.list, &wq->flusher_queue);
2510 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2511 }
2512 } else {
2513 /*
2514 * Oops, color space is full, wait on overflow queue.
2515 * The next flush completion will assign us
2516 * flush_color and transfer to flusher_queue.
2517 */
2518 list_add_tail(&this_flusher.list, &wq->flusher_overflow);
2519 }
2520
2521 mutex_unlock(&wq->flush_mutex);
2522
2523 wait_for_completion(&this_flusher.done);
2524
2525 /*
2526 * Wake-up-and-cascade phase
2527 *
2528 * First flushers are responsible for cascading flushes and
2529 * handling overflow. Non-first flushers can simply return.
2530 */
2531 if (wq->first_flusher != &this_flusher)
2532 return;
2533
2534 mutex_lock(&wq->flush_mutex);
2535
4ce48b37
TH
2536 /* we might have raced, check again with mutex held */
2537 if (wq->first_flusher != &this_flusher)
2538 goto out_unlock;
2539
73f53c4a
TH
2540 wq->first_flusher = NULL;
2541
2542 BUG_ON(!list_empty(&this_flusher.list));
2543 BUG_ON(wq->flush_color != this_flusher.flush_color);
2544
2545 while (true) {
2546 struct wq_flusher *next, *tmp;
2547
2548 /* complete all the flushers sharing the current flush color */
2549 list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) {
2550 if (next->flush_color != wq->flush_color)
2551 break;
2552 list_del_init(&next->list);
2553 complete(&next->done);
2554 }
2555
2556 BUG_ON(!list_empty(&wq->flusher_overflow) &&
2557 wq->flush_color != work_next_color(wq->work_color));
2558
2559 /* this flush_color is finished, advance by one */
2560 wq->flush_color = work_next_color(wq->flush_color);
2561
2562 /* one color has been freed, handle overflow queue */
2563 if (!list_empty(&wq->flusher_overflow)) {
2564 /*
2565 * Assign the same color to all overflowed
2566 * flushers, advance work_color and append to
2567 * flusher_queue. This is the start-to-wait
2568 * phase for these overflowed flushers.
2569 */
2570 list_for_each_entry(tmp, &wq->flusher_overflow, list)
2571 tmp->flush_color = wq->work_color;
2572
2573 wq->work_color = work_next_color(wq->work_color);
2574
2575 list_splice_tail_init(&wq->flusher_overflow,
2576 &wq->flusher_queue);
2577 flush_workqueue_prep_cwqs(wq, -1, wq->work_color);
2578 }
2579
2580 if (list_empty(&wq->flusher_queue)) {
2581 BUG_ON(wq->flush_color != wq->work_color);
2582 break;
2583 }
2584
2585 /*
2586 * Need to flush more colors. Make the next flusher
2587 * the new first flusher and arm cwqs.
2588 */
2589 BUG_ON(wq->flush_color == wq->work_color);
2590 BUG_ON(wq->flush_color != next->flush_color);
2591
2592 list_del_init(&next->list);
2593 wq->first_flusher = next;
2594
2595 if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1))
2596 break;
2597
2598 /*
2599 * Meh... this color is already done, clear first
2600 * flusher and repeat cascading.
2601 */
2602 wq->first_flusher = NULL;
2603 }
2604
2605out_unlock:
2606 mutex_unlock(&wq->flush_mutex);
1da177e4 2607}
ae90dd5d 2608EXPORT_SYMBOL_GPL(flush_workqueue);
1da177e4 2609
9c5a2ba7
TH
2610/**
2611 * drain_workqueue - drain a workqueue
2612 * @wq: workqueue to drain
2613 *
2614 * Wait until the workqueue becomes empty. While draining is in progress,
2615 * only chain queueing is allowed. IOW, only currently pending or running
2616 * work items on @wq can queue further work items on it. @wq is flushed
2617 * repeatedly until it becomes empty. The number of flushing is detemined
2618 * by the depth of chaining and should be relatively short. Whine if it
2619 * takes too long.
2620 */
2621void drain_workqueue(struct workqueue_struct *wq)
2622{
2623 unsigned int flush_cnt = 0;
2624 unsigned int cpu;
2625
2626 /*
2627 * __queue_work() needs to test whether there are drainers, is much
2628 * hotter than drain_workqueue() and already looks at @wq->flags.
2629 * Use WQ_DRAINING so that queue doesn't have to check nr_drainers.
2630 */
2631 spin_lock(&workqueue_lock);
2632 if (!wq->nr_drainers++)
2633 wq->flags |= WQ_DRAINING;
2634 spin_unlock(&workqueue_lock);
2635reflush:
2636 flush_workqueue(wq);
2637
2638 for_each_cwq_cpu(cpu, wq) {
2639 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
fa2563e4 2640 bool drained;
9c5a2ba7 2641
bd7bdd43 2642 spin_lock_irq(&cwq->pool->gcwq->lock);
fa2563e4 2643 drained = !cwq->nr_active && list_empty(&cwq->delayed_works);
bd7bdd43 2644 spin_unlock_irq(&cwq->pool->gcwq->lock);
fa2563e4
TT
2645
2646 if (drained)
9c5a2ba7
TH
2647 continue;
2648
2649 if (++flush_cnt == 10 ||
2650 (flush_cnt % 100 == 0 && flush_cnt <= 1000))
2651 pr_warning("workqueue %s: flush on destruction isn't complete after %u tries\n",
2652 wq->name, flush_cnt);
2653 goto reflush;
2654 }
2655
2656 spin_lock(&workqueue_lock);
2657 if (!--wq->nr_drainers)
2658 wq->flags &= ~WQ_DRAINING;
2659 spin_unlock(&workqueue_lock);
2660}
2661EXPORT_SYMBOL_GPL(drain_workqueue);
2662
baf59022
TH
2663static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr,
2664 bool wait_executing)
db700897 2665{
affee4b2 2666 struct worker *worker = NULL;
8b03ae3c 2667 struct global_cwq *gcwq;
db700897 2668 struct cpu_workqueue_struct *cwq;
db700897
ON
2669
2670 might_sleep();
7a22ad75
TH
2671 gcwq = get_work_gcwq(work);
2672 if (!gcwq)
baf59022 2673 return false;
db700897 2674
8b03ae3c 2675 spin_lock_irq(&gcwq->lock);
db700897
ON
2676 if (!list_empty(&work->entry)) {
2677 /*
2678 * See the comment near try_to_grab_pending()->smp_rmb().
7a22ad75
TH
2679 * If it was re-queued to a different gcwq under us, we
2680 * are not going to wait.
db700897
ON
2681 */
2682 smp_rmb();
7a22ad75 2683 cwq = get_work_cwq(work);
bd7bdd43 2684 if (unlikely(!cwq || gcwq != cwq->pool->gcwq))
4690c4ab 2685 goto already_gone;
baf59022 2686 } else if (wait_executing) {
7a22ad75 2687 worker = find_worker_executing_work(gcwq, work);
affee4b2 2688 if (!worker)
4690c4ab 2689 goto already_gone;
7a22ad75 2690 cwq = worker->current_cwq;
baf59022
TH
2691 } else
2692 goto already_gone;
db700897 2693
baf59022 2694 insert_wq_barrier(cwq, barr, work, worker);
8b03ae3c 2695 spin_unlock_irq(&gcwq->lock);
7a22ad75 2696
e159489b
TH
2697 /*
2698 * If @max_active is 1 or rescuer is in use, flushing another work
2699 * item on the same workqueue may lead to deadlock. Make sure the
2700 * flusher is not running on the same workqueue by verifying write
2701 * access.
2702 */
2703 if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER)
2704 lock_map_acquire(&cwq->wq->lockdep_map);
2705 else
2706 lock_map_acquire_read(&cwq->wq->lockdep_map);
7a22ad75 2707 lock_map_release(&cwq->wq->lockdep_map);
e159489b 2708
401a8d04 2709 return true;
4690c4ab 2710already_gone:
8b03ae3c 2711 spin_unlock_irq(&gcwq->lock);
401a8d04 2712 return false;
db700897 2713}
baf59022
TH
2714
2715/**
2716 * flush_work - wait for a work to finish executing the last queueing instance
2717 * @work: the work to flush
2718 *
2719 * Wait until @work has finished execution. This function considers
2720 * only the last queueing instance of @work. If @work has been
2721 * enqueued across different CPUs on a non-reentrant workqueue or on
2722 * multiple workqueues, @work might still be executing on return on
2723 * some of the CPUs from earlier queueing.
2724 *
2725 * If @work was queued only on a non-reentrant, ordered or unbound
2726 * workqueue, @work is guaranteed to be idle on return if it hasn't
2727 * been requeued since flush started.
2728 *
2729 * RETURNS:
2730 * %true if flush_work() waited for the work to finish execution,
2731 * %false if it was already idle.
2732 */
2733bool flush_work(struct work_struct *work)
2734{
2735 struct wq_barrier barr;
2736
0976dfc1
SB
2737 lock_map_acquire(&work->lockdep_map);
2738 lock_map_release(&work->lockdep_map);
2739
baf59022
TH
2740 if (start_flush_work(work, &barr, true)) {
2741 wait_for_completion(&barr.done);
2742 destroy_work_on_stack(&barr.work);
2743 return true;
2744 } else
2745 return false;
2746}
db700897
ON
2747EXPORT_SYMBOL_GPL(flush_work);
2748
401a8d04
TH
2749static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work)
2750{
2751 struct wq_barrier barr;
2752 struct worker *worker;
2753
2754 spin_lock_irq(&gcwq->lock);
2755
2756 worker = find_worker_executing_work(gcwq, work);
2757 if (unlikely(worker))
2758 insert_wq_barrier(worker->current_cwq, &barr, work, worker);
2759
2760 spin_unlock_irq(&gcwq->lock);
2761
2762 if (unlikely(worker)) {
2763 wait_for_completion(&barr.done);
2764 destroy_work_on_stack(&barr.work);
2765 return true;
2766 } else
2767 return false;
2768}
2769
2770static bool wait_on_work(struct work_struct *work)
2771{
2772 bool ret = false;
2773 int cpu;
2774
2775 might_sleep();
2776
2777 lock_map_acquire(&work->lockdep_map);
2778 lock_map_release(&work->lockdep_map);
2779
2780 for_each_gcwq_cpu(cpu)
2781 ret |= wait_on_cpu_work(get_gcwq(cpu), work);
2782 return ret;
2783}
2784
09383498
TH
2785/**
2786 * flush_work_sync - wait until a work has finished execution
2787 * @work: the work to flush
2788 *
2789 * Wait until @work has finished execution. On return, it's
2790 * guaranteed that all queueing instances of @work which happened
2791 * before this function is called are finished. In other words, if
2792 * @work hasn't been requeued since this function was called, @work is
2793 * guaranteed to be idle on return.
2794 *
2795 * RETURNS:
2796 * %true if flush_work_sync() waited for the work to finish execution,
2797 * %false if it was already idle.
2798 */
2799bool flush_work_sync(struct work_struct *work)
2800{
2801 struct wq_barrier barr;
2802 bool pending, waited;
2803
2804 /* we'll wait for executions separately, queue barr only if pending */
2805 pending = start_flush_work(work, &barr, false);
2806
2807 /* wait for executions to finish */
2808 waited = wait_on_work(work);
2809
2810 /* wait for the pending one */
2811 if (pending) {
2812 wait_for_completion(&barr.done);
2813 destroy_work_on_stack(&barr.work);
2814 }
2815
2816 return pending || waited;
2817}
2818EXPORT_SYMBOL_GPL(flush_work_sync);
2819
401a8d04 2820static bool __cancel_work_timer(struct work_struct *work,
1f1f642e
ON
2821 struct timer_list* timer)
2822{
2823 int ret;
2824
2825 do {
2826 ret = (timer && likely(del_timer(timer)));
2827 if (!ret)
2828 ret = try_to_grab_pending(work);
2829 wait_on_work(work);
2830 } while (unlikely(ret < 0));
2831
7a22ad75 2832 clear_work_data(work);
1f1f642e
ON
2833 return ret;
2834}
2835
6e84d644 2836/**
401a8d04
TH
2837 * cancel_work_sync - cancel a work and wait for it to finish
2838 * @work: the work to cancel
6e84d644 2839 *
401a8d04
TH
2840 * Cancel @work and wait for its execution to finish. This function
2841 * can be used even if the work re-queues itself or migrates to
2842 * another workqueue. On return from this function, @work is
2843 * guaranteed to be not pending or executing on any CPU.
1f1f642e 2844 *
401a8d04
TH
2845 * cancel_work_sync(&delayed_work->work) must not be used for
2846 * delayed_work's. Use cancel_delayed_work_sync() instead.
6e84d644 2847 *
401a8d04 2848 * The caller must ensure that the workqueue on which @work was last
6e84d644 2849 * queued can't be destroyed before this function returns.
401a8d04
TH
2850 *
2851 * RETURNS:
2852 * %true if @work was pending, %false otherwise.
6e84d644 2853 */
401a8d04 2854bool cancel_work_sync(struct work_struct *work)
6e84d644 2855{
1f1f642e 2856 return __cancel_work_timer(work, NULL);
b89deed3 2857}
28e53bdd 2858EXPORT_SYMBOL_GPL(cancel_work_sync);
b89deed3 2859
6e84d644 2860/**
401a8d04
TH
2861 * flush_delayed_work - wait for a dwork to finish executing the last queueing
2862 * @dwork: the delayed work to flush
6e84d644 2863 *
401a8d04
TH
2864 * Delayed timer is cancelled and the pending work is queued for
2865 * immediate execution. Like flush_work(), this function only
2866 * considers the last queueing instance of @dwork.
1f1f642e 2867 *
401a8d04
TH
2868 * RETURNS:
2869 * %true if flush_work() waited for the work to finish execution,
2870 * %false if it was already idle.
6e84d644 2871 */
401a8d04
TH
2872bool flush_delayed_work(struct delayed_work *dwork)
2873{
8930caba 2874 local_irq_disable();
401a8d04 2875 if (del_timer_sync(&dwork->timer))
57469821 2876 __queue_work(WORK_CPU_UNBOUND,
401a8d04 2877 get_work_cwq(&dwork->work)->wq, &dwork->work);
8930caba 2878 local_irq_enable();
401a8d04
TH
2879 return flush_work(&dwork->work);
2880}
2881EXPORT_SYMBOL(flush_delayed_work);
2882
09383498
TH
2883/**
2884 * flush_delayed_work_sync - wait for a dwork to finish
2885 * @dwork: the delayed work to flush
2886 *
2887 * Delayed timer is cancelled and the pending work is queued for
2888 * execution immediately. Other than timer handling, its behavior
2889 * is identical to flush_work_sync().
2890 *
2891 * RETURNS:
2892 * %true if flush_work_sync() waited for the work to finish execution,
2893 * %false if it was already idle.
2894 */
2895bool flush_delayed_work_sync(struct delayed_work *dwork)
2896{
8930caba 2897 local_irq_disable();
09383498 2898 if (del_timer_sync(&dwork->timer))
57469821 2899 __queue_work(WORK_CPU_UNBOUND,
09383498 2900 get_work_cwq(&dwork->work)->wq, &dwork->work);
8930caba 2901 local_irq_enable();
09383498
TH
2902 return flush_work_sync(&dwork->work);
2903}
2904EXPORT_SYMBOL(flush_delayed_work_sync);
2905
401a8d04
TH
2906/**
2907 * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish
2908 * @dwork: the delayed work cancel
2909 *
2910 * This is cancel_work_sync() for delayed works.
2911 *
2912 * RETURNS:
2913 * %true if @dwork was pending, %false otherwise.
2914 */
2915bool cancel_delayed_work_sync(struct delayed_work *dwork)
6e84d644 2916{
1f1f642e 2917 return __cancel_work_timer(&dwork->work, &dwork->timer);
6e84d644 2918}
f5a421a4 2919EXPORT_SYMBOL(cancel_delayed_work_sync);
1da177e4 2920
d4283e93 2921/**
0a13c00e
TH
2922 * schedule_work_on - put work task on a specific cpu
2923 * @cpu: cpu to put the work task on
2924 * @work: job to be done
2925 *
2926 * This puts a job on a specific cpu
2927 */
d4283e93 2928bool schedule_work_on(int cpu, struct work_struct *work)
0a13c00e
TH
2929{
2930 return queue_work_on(cpu, system_wq, work);
2931}
2932EXPORT_SYMBOL(schedule_work_on);
2933
0fcb78c2
REB
2934/**
2935 * schedule_work - put work task in global workqueue
2936 * @work: job to be done
2937 *
d4283e93
TH
2938 * Returns %false if @work was already on the kernel-global workqueue and
2939 * %true otherwise.
5b0f437d
BVA
2940 *
2941 * This puts a job in the kernel-global workqueue if it was not already
2942 * queued and leaves it in the same position on the kernel-global
2943 * workqueue otherwise.
0fcb78c2 2944 */
d4283e93 2945bool schedule_work(struct work_struct *work)
1da177e4 2946{
d320c038 2947 return queue_work(system_wq, work);
1da177e4 2948}
ae90dd5d 2949EXPORT_SYMBOL(schedule_work);
1da177e4 2950
0a13c00e
TH
2951/**
2952 * schedule_delayed_work_on - queue work in global workqueue on CPU after delay
2953 * @cpu: cpu to use
2954 * @dwork: job to be done
2955 * @delay: number of jiffies to wait
c1a220e7 2956 *
0a13c00e
TH
2957 * After waiting for a given time this puts a job in the kernel-global
2958 * workqueue on the specified CPU.
c1a220e7 2959 */
d4283e93
TH
2960bool schedule_delayed_work_on(int cpu, struct delayed_work *dwork,
2961 unsigned long delay)
c1a220e7 2962{
0a13c00e 2963 return queue_delayed_work_on(cpu, system_wq, dwork, delay);
c1a220e7 2964}
0a13c00e 2965EXPORT_SYMBOL(schedule_delayed_work_on);
c1a220e7 2966
0fcb78c2
REB
2967/**
2968 * schedule_delayed_work - put work task in global workqueue after delay
52bad64d
DH
2969 * @dwork: job to be done
2970 * @delay: number of jiffies to wait or 0 for immediate execution
0fcb78c2
REB
2971 *
2972 * After waiting for a given time this puts a job in the kernel-global
2973 * workqueue.
2974 */
d4283e93 2975bool schedule_delayed_work(struct delayed_work *dwork, unsigned long delay)
1da177e4 2976{
d320c038 2977 return queue_delayed_work(system_wq, dwork, delay);
1da177e4 2978}
ae90dd5d 2979EXPORT_SYMBOL(schedule_delayed_work);
1da177e4 2980
b6136773 2981/**
31ddd871 2982 * schedule_on_each_cpu - execute a function synchronously on each online CPU
b6136773 2983 * @func: the function to call
b6136773 2984 *
31ddd871
TH
2985 * schedule_on_each_cpu() executes @func on each online CPU using the
2986 * system workqueue and blocks until all CPUs have completed.
b6136773 2987 * schedule_on_each_cpu() is very slow.
31ddd871
TH
2988 *
2989 * RETURNS:
2990 * 0 on success, -errno on failure.
b6136773 2991 */
65f27f38 2992int schedule_on_each_cpu(work_func_t func)
15316ba8
CL
2993{
2994 int cpu;
38f51568 2995 struct work_struct __percpu *works;
15316ba8 2996
b6136773
AM
2997 works = alloc_percpu(struct work_struct);
2998 if (!works)
15316ba8 2999 return -ENOMEM;
b6136773 3000
93981800
TH
3001 get_online_cpus();
3002
15316ba8 3003 for_each_online_cpu(cpu) {
9bfb1839
IM
3004 struct work_struct *work = per_cpu_ptr(works, cpu);
3005
3006 INIT_WORK(work, func);
b71ab8c2 3007 schedule_work_on(cpu, work);
65a64464 3008 }
93981800
TH
3009
3010 for_each_online_cpu(cpu)
3011 flush_work(per_cpu_ptr(works, cpu));
3012
95402b38 3013 put_online_cpus();
b6136773 3014 free_percpu(works);
15316ba8
CL
3015 return 0;
3016}
3017
eef6a7d5
AS
3018/**
3019 * flush_scheduled_work - ensure that any scheduled work has run to completion.
3020 *
3021 * Forces execution of the kernel-global workqueue and blocks until its
3022 * completion.
3023 *
3024 * Think twice before calling this function! It's very easy to get into
3025 * trouble if you don't take great care. Either of the following situations
3026 * will lead to deadlock:
3027 *
3028 * One of the work items currently on the workqueue needs to acquire
3029 * a lock held by your code or its caller.
3030 *
3031 * Your code is running in the context of a work routine.
3032 *
3033 * They will be detected by lockdep when they occur, but the first might not
3034 * occur very often. It depends on what work items are on the workqueue and
3035 * what locks they need, which you have no control over.
3036 *
3037 * In most situations flushing the entire workqueue is overkill; you merely
3038 * need to know that a particular work item isn't queued and isn't running.
3039 * In such cases you should use cancel_delayed_work_sync() or
3040 * cancel_work_sync() instead.
3041 */
1da177e4
LT
3042void flush_scheduled_work(void)
3043{
d320c038 3044 flush_workqueue(system_wq);
1da177e4 3045}
ae90dd5d 3046EXPORT_SYMBOL(flush_scheduled_work);
1da177e4 3047
1fa44eca
JB
3048/**
3049 * execute_in_process_context - reliably execute the routine with user context
3050 * @fn: the function to execute
1fa44eca
JB
3051 * @ew: guaranteed storage for the execute work structure (must
3052 * be available when the work executes)
3053 *
3054 * Executes the function immediately if process context is available,
3055 * otherwise schedules the function for delayed execution.
3056 *
3057 * Returns: 0 - function was executed
3058 * 1 - function was scheduled for execution
3059 */
65f27f38 3060int execute_in_process_context(work_func_t fn, struct execute_work *ew)
1fa44eca
JB
3061{
3062 if (!in_interrupt()) {
65f27f38 3063 fn(&ew->work);
1fa44eca
JB
3064 return 0;
3065 }
3066
65f27f38 3067 INIT_WORK(&ew->work, fn);
1fa44eca
JB
3068 schedule_work(&ew->work);
3069
3070 return 1;
3071}
3072EXPORT_SYMBOL_GPL(execute_in_process_context);
3073
1da177e4
LT
3074int keventd_up(void)
3075{
d320c038 3076 return system_wq != NULL;
1da177e4
LT
3077}
3078
bdbc5dd7 3079static int alloc_cwqs(struct workqueue_struct *wq)
0f900049 3080{
65a64464 3081 /*
0f900049
TH
3082 * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS.
3083 * Make sure that the alignment isn't lower than that of
3084 * unsigned long long.
65a64464 3085 */
0f900049
TH
3086 const size_t size = sizeof(struct cpu_workqueue_struct);
3087 const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS,
3088 __alignof__(unsigned long long));
65a64464 3089
e06ffa1e 3090 if (!(wq->flags & WQ_UNBOUND))
f3421797 3091 wq->cpu_wq.pcpu = __alloc_percpu(size, align);
931ac77e 3092 else {
f3421797
TH
3093 void *ptr;
3094
3095 /*
3096 * Allocate enough room to align cwq and put an extra
3097 * pointer at the end pointing back to the originally
3098 * allocated pointer which will be used for free.
3099 */
3100 ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL);
3101 if (ptr) {
3102 wq->cpu_wq.single = PTR_ALIGN(ptr, align);
3103 *(void **)(wq->cpu_wq.single + 1) = ptr;
3104 }
bdbc5dd7 3105 }
f3421797 3106
0415b00d 3107 /* just in case, make sure it's actually aligned */
bdbc5dd7
TH
3108 BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align));
3109 return wq->cpu_wq.v ? 0 : -ENOMEM;
0f900049
TH
3110}
3111
bdbc5dd7 3112static void free_cwqs(struct workqueue_struct *wq)
0f900049 3113{
e06ffa1e 3114 if (!(wq->flags & WQ_UNBOUND))
f3421797
TH
3115 free_percpu(wq->cpu_wq.pcpu);
3116 else if (wq->cpu_wq.single) {
3117 /* the pointer to free is stored right after the cwq */
bdbc5dd7 3118 kfree(*(void **)(wq->cpu_wq.single + 1));
f3421797 3119 }
0f900049
TH
3120}
3121
f3421797
TH
3122static int wq_clamp_max_active(int max_active, unsigned int flags,
3123 const char *name)
b71ab8c2 3124{
f3421797
TH
3125 int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE;
3126
3127 if (max_active < 1 || max_active > lim)
b71ab8c2
TH
3128 printk(KERN_WARNING "workqueue: max_active %d requested for %s "
3129 "is out of range, clamping between %d and %d\n",
f3421797 3130 max_active, name, 1, lim);
b71ab8c2 3131
f3421797 3132 return clamp_val(max_active, 1, lim);
b71ab8c2
TH
3133}
3134
b196be89 3135struct workqueue_struct *__alloc_workqueue_key(const char *fmt,
d320c038
TH
3136 unsigned int flags,
3137 int max_active,
3138 struct lock_class_key *key,
b196be89 3139 const char *lock_name, ...)
1da177e4 3140{
b196be89 3141 va_list args, args1;
1da177e4 3142 struct workqueue_struct *wq;
c34056a3 3143 unsigned int cpu;
b196be89
TH
3144 size_t namelen;
3145
3146 /* determine namelen, allocate wq and format name */
3147 va_start(args, lock_name);
3148 va_copy(args1, args);
3149 namelen = vsnprintf(NULL, 0, fmt, args) + 1;
3150
3151 wq = kzalloc(sizeof(*wq) + namelen, GFP_KERNEL);
3152 if (!wq)
3153 goto err;
3154
3155 vsnprintf(wq->name, namelen, fmt, args1);
3156 va_end(args);
3157 va_end(args1);
1da177e4 3158
6370a6ad
TH
3159 /*
3160 * Workqueues which may be used during memory reclaim should
3161 * have a rescuer to guarantee forward progress.
3162 */
3163 if (flags & WQ_MEM_RECLAIM)
3164 flags |= WQ_RESCUER;
3165
d320c038 3166 max_active = max_active ?: WQ_DFL_ACTIVE;
b196be89 3167 max_active = wq_clamp_max_active(max_active, flags, wq->name);
3af24433 3168
b196be89 3169 /* init wq */
97e37d7b 3170 wq->flags = flags;
a0a1a5fd 3171 wq->saved_max_active = max_active;
73f53c4a
TH
3172 mutex_init(&wq->flush_mutex);
3173 atomic_set(&wq->nr_cwqs_to_flush, 0);
3174 INIT_LIST_HEAD(&wq->flusher_queue);
3175 INIT_LIST_HEAD(&wq->flusher_overflow);
502ca9d8 3176
eb13ba87 3177 lockdep_init_map(&wq->lockdep_map, lock_name, key, 0);
cce1a165 3178 INIT_LIST_HEAD(&wq->list);
3af24433 3179
bdbc5dd7
TH
3180 if (alloc_cwqs(wq) < 0)
3181 goto err;
3182
f3421797 3183 for_each_cwq_cpu(cpu, wq) {
1537663f 3184 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
8b03ae3c 3185 struct global_cwq *gcwq = get_gcwq(cpu);
3270476a 3186 int pool_idx = (bool)(flags & WQ_HIGHPRI);
1537663f 3187
0f900049 3188 BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK);
3270476a 3189 cwq->pool = &gcwq->pools[pool_idx];
c34056a3 3190 cwq->wq = wq;
73f53c4a 3191 cwq->flush_color = -1;
1e19ffc6 3192 cwq->max_active = max_active;
1e19ffc6 3193 INIT_LIST_HEAD(&cwq->delayed_works);
e22bee78 3194 }
1537663f 3195
e22bee78
TH
3196 if (flags & WQ_RESCUER) {
3197 struct worker *rescuer;
3198
f2e005aa 3199 if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL))
e22bee78
TH
3200 goto err;
3201
3202 wq->rescuer = rescuer = alloc_worker();
3203 if (!rescuer)
3204 goto err;
3205
b196be89
TH
3206 rescuer->task = kthread_create(rescuer_thread, wq, "%s",
3207 wq->name);
e22bee78
TH
3208 if (IS_ERR(rescuer->task))
3209 goto err;
3210
e22bee78
TH
3211 rescuer->task->flags |= PF_THREAD_BOUND;
3212 wake_up_process(rescuer->task);
3af24433
ON
3213 }
3214
a0a1a5fd
TH
3215 /*
3216 * workqueue_lock protects global freeze state and workqueues
3217 * list. Grab it, set max_active accordingly and add the new
3218 * workqueue to workqueues list.
3219 */
1537663f 3220 spin_lock(&workqueue_lock);
a0a1a5fd 3221
58a69cb4 3222 if (workqueue_freezing && wq->flags & WQ_FREEZABLE)
f3421797 3223 for_each_cwq_cpu(cpu, wq)
a0a1a5fd
TH
3224 get_cwq(cpu, wq)->max_active = 0;
3225
1537663f 3226 list_add(&wq->list, &workqueues);
a0a1a5fd 3227
1537663f
TH
3228 spin_unlock(&workqueue_lock);
3229
3af24433 3230 return wq;
4690c4ab
TH
3231err:
3232 if (wq) {
bdbc5dd7 3233 free_cwqs(wq);
f2e005aa 3234 free_mayday_mask(wq->mayday_mask);
e22bee78 3235 kfree(wq->rescuer);
4690c4ab
TH
3236 kfree(wq);
3237 }
3238 return NULL;
3af24433 3239}
d320c038 3240EXPORT_SYMBOL_GPL(__alloc_workqueue_key);
1da177e4 3241
3af24433
ON
3242/**
3243 * destroy_workqueue - safely terminate a workqueue
3244 * @wq: target workqueue
3245 *
3246 * Safely destroy a workqueue. All work currently pending will be done first.
3247 */
3248void destroy_workqueue(struct workqueue_struct *wq)
3249{
c8e55f36 3250 unsigned int cpu;
3af24433 3251
9c5a2ba7
TH
3252 /* drain it before proceeding with destruction */
3253 drain_workqueue(wq);
c8efcc25 3254
a0a1a5fd
TH
3255 /*
3256 * wq list is used to freeze wq, remove from list after
3257 * flushing is complete in case freeze races us.
3258 */
95402b38 3259 spin_lock(&workqueue_lock);
b1f4ec17 3260 list_del(&wq->list);
95402b38 3261 spin_unlock(&workqueue_lock);
3af24433 3262
e22bee78 3263 /* sanity check */
f3421797 3264 for_each_cwq_cpu(cpu, wq) {
73f53c4a
TH
3265 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3266 int i;
3267
73f53c4a
TH
3268 for (i = 0; i < WORK_NR_COLORS; i++)
3269 BUG_ON(cwq->nr_in_flight[i]);
1e19ffc6
TH
3270 BUG_ON(cwq->nr_active);
3271 BUG_ON(!list_empty(&cwq->delayed_works));
73f53c4a 3272 }
9b41ea72 3273
e22bee78
TH
3274 if (wq->flags & WQ_RESCUER) {
3275 kthread_stop(wq->rescuer->task);
f2e005aa 3276 free_mayday_mask(wq->mayday_mask);
8d9df9f0 3277 kfree(wq->rescuer);
e22bee78
TH
3278 }
3279
bdbc5dd7 3280 free_cwqs(wq);
3af24433
ON
3281 kfree(wq);
3282}
3283EXPORT_SYMBOL_GPL(destroy_workqueue);
3284
dcd989cb
TH
3285/**
3286 * workqueue_set_max_active - adjust max_active of a workqueue
3287 * @wq: target workqueue
3288 * @max_active: new max_active value.
3289 *
3290 * Set max_active of @wq to @max_active.
3291 *
3292 * CONTEXT:
3293 * Don't call from IRQ context.
3294 */
3295void workqueue_set_max_active(struct workqueue_struct *wq, int max_active)
3296{
3297 unsigned int cpu;
3298
f3421797 3299 max_active = wq_clamp_max_active(max_active, wq->flags, wq->name);
dcd989cb
TH
3300
3301 spin_lock(&workqueue_lock);
3302
3303 wq->saved_max_active = max_active;
3304
f3421797 3305 for_each_cwq_cpu(cpu, wq) {
dcd989cb
TH
3306 struct global_cwq *gcwq = get_gcwq(cpu);
3307
3308 spin_lock_irq(&gcwq->lock);
3309
58a69cb4 3310 if (!(wq->flags & WQ_FREEZABLE) ||
dcd989cb
TH
3311 !(gcwq->flags & GCWQ_FREEZING))
3312 get_cwq(gcwq->cpu, wq)->max_active = max_active;
9bfb1839 3313
dcd989cb 3314 spin_unlock_irq(&gcwq->lock);
65a64464 3315 }
93981800 3316
dcd989cb 3317 spin_unlock(&workqueue_lock);
15316ba8 3318}
dcd989cb 3319EXPORT_SYMBOL_GPL(workqueue_set_max_active);
15316ba8 3320
eef6a7d5 3321/**
dcd989cb
TH
3322 * workqueue_congested - test whether a workqueue is congested
3323 * @cpu: CPU in question
3324 * @wq: target workqueue
eef6a7d5 3325 *
dcd989cb
TH
3326 * Test whether @wq's cpu workqueue for @cpu is congested. There is
3327 * no synchronization around this function and the test result is
3328 * unreliable and only useful as advisory hints or for debugging.
eef6a7d5 3329 *
dcd989cb
TH
3330 * RETURNS:
3331 * %true if congested, %false otherwise.
eef6a7d5 3332 */
dcd989cb 3333bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq)
1da177e4 3334{
dcd989cb
TH
3335 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3336
3337 return !list_empty(&cwq->delayed_works);
1da177e4 3338}
dcd989cb 3339EXPORT_SYMBOL_GPL(workqueue_congested);
1da177e4 3340
1fa44eca 3341/**
dcd989cb
TH
3342 * work_cpu - return the last known associated cpu for @work
3343 * @work: the work of interest
1fa44eca 3344 *
dcd989cb 3345 * RETURNS:
bdbc5dd7 3346 * CPU number if @work was ever queued. WORK_CPU_NONE otherwise.
1fa44eca 3347 */
dcd989cb 3348unsigned int work_cpu(struct work_struct *work)
1fa44eca 3349{
dcd989cb 3350 struct global_cwq *gcwq = get_work_gcwq(work);
1fa44eca 3351
bdbc5dd7 3352 return gcwq ? gcwq->cpu : WORK_CPU_NONE;
1fa44eca 3353}
dcd989cb 3354EXPORT_SYMBOL_GPL(work_cpu);
1fa44eca 3355
dcd989cb
TH
3356/**
3357 * work_busy - test whether a work is currently pending or running
3358 * @work: the work to be tested
3359 *
3360 * Test whether @work is currently pending or running. There is no
3361 * synchronization around this function and the test result is
3362 * unreliable and only useful as advisory hints or for debugging.
3363 * Especially for reentrant wqs, the pending state might hide the
3364 * running state.
3365 *
3366 * RETURNS:
3367 * OR'd bitmask of WORK_BUSY_* bits.
3368 */
3369unsigned int work_busy(struct work_struct *work)
1da177e4 3370{
dcd989cb
TH
3371 struct global_cwq *gcwq = get_work_gcwq(work);
3372 unsigned long flags;
3373 unsigned int ret = 0;
1da177e4 3374
dcd989cb
TH
3375 if (!gcwq)
3376 return false;
1da177e4 3377
dcd989cb 3378 spin_lock_irqsave(&gcwq->lock, flags);
1da177e4 3379
dcd989cb
TH
3380 if (work_pending(work))
3381 ret |= WORK_BUSY_PENDING;
3382 if (find_worker_executing_work(gcwq, work))
3383 ret |= WORK_BUSY_RUNNING;
1da177e4 3384
dcd989cb 3385 spin_unlock_irqrestore(&gcwq->lock, flags);
1da177e4 3386
dcd989cb 3387 return ret;
1da177e4 3388}
dcd989cb 3389EXPORT_SYMBOL_GPL(work_busy);
1da177e4 3390
db7bccf4
TH
3391/*
3392 * CPU hotplug.
3393 *
e22bee78
TH
3394 * There are two challenges in supporting CPU hotplug. Firstly, there
3395 * are a lot of assumptions on strong associations among work, cwq and
3396 * gcwq which make migrating pending and scheduled works very
3397 * difficult to implement without impacting hot paths. Secondly,
3398 * gcwqs serve mix of short, long and very long running works making
3399 * blocked draining impractical.
3400 *
628c78e7
TH
3401 * This is solved by allowing a gcwq to be disassociated from the CPU
3402 * running as an unbound one and allowing it to be reattached later if the
3403 * cpu comes back online.
db7bccf4 3404 */
1da177e4 3405
60373152 3406/* claim manager positions of all pools */
8db25e78 3407static void gcwq_claim_management_and_lock(struct global_cwq *gcwq)
60373152
TH
3408{
3409 struct worker_pool *pool;
3410
3411 for_each_worker_pool(pool, gcwq)
3412 mutex_lock_nested(&pool->manager_mutex, pool - gcwq->pools);
8db25e78 3413 spin_lock_irq(&gcwq->lock);
60373152
TH
3414}
3415
3416/* release manager positions */
8db25e78 3417static void gcwq_release_management_and_unlock(struct global_cwq *gcwq)
60373152
TH
3418{
3419 struct worker_pool *pool;
3420
8db25e78 3421 spin_unlock_irq(&gcwq->lock);
60373152
TH
3422 for_each_worker_pool(pool, gcwq)
3423 mutex_unlock(&pool->manager_mutex);
3424}
3425
628c78e7 3426static void gcwq_unbind_fn(struct work_struct *work)
3af24433 3427{
628c78e7 3428 struct global_cwq *gcwq = get_gcwq(smp_processor_id());
4ce62e9e 3429 struct worker_pool *pool;
db7bccf4
TH
3430 struct worker *worker;
3431 struct hlist_node *pos;
3432 int i;
3af24433 3433
db7bccf4
TH
3434 BUG_ON(gcwq->cpu != smp_processor_id());
3435
8db25e78 3436 gcwq_claim_management_and_lock(gcwq);
3af24433 3437
f2d5a0ee
TH
3438 /*
3439 * We've claimed all manager positions. Make all workers unbound
3440 * and set DISASSOCIATED. Before this, all workers except for the
3441 * ones which are still executing works from before the last CPU
3442 * down must be on the cpu. After this, they may become diasporas.
3443 */
60373152 3444 for_each_worker_pool(pool, gcwq)
4ce62e9e 3445 list_for_each_entry(worker, &pool->idle_list, entry)
403c821d 3446 worker->flags |= WORKER_UNBOUND;
3af24433 3447
db7bccf4 3448 for_each_busy_worker(worker, i, pos, gcwq)
403c821d 3449 worker->flags |= WORKER_UNBOUND;
06ba38a9 3450
f2d5a0ee
TH
3451 gcwq->flags |= GCWQ_DISASSOCIATED;
3452
8db25e78 3453 gcwq_release_management_and_unlock(gcwq);
628c78e7 3454
e22bee78 3455 /*
403c821d 3456 * Call schedule() so that we cross rq->lock and thus can guarantee
628c78e7
TH
3457 * sched callbacks see the %WORKER_UNBOUND flag. This is necessary
3458 * as scheduler callbacks may be invoked from other cpus.
e22bee78 3459 */
e22bee78 3460 schedule();
06ba38a9 3461
e22bee78 3462 /*
628c78e7
TH
3463 * Sched callbacks are disabled now. Zap nr_running. After this,
3464 * nr_running stays zero and need_more_worker() and keep_working()
3465 * are always true as long as the worklist is not empty. @gcwq now
3466 * behaves as unbound (in terms of concurrency management) gcwq
3467 * which is served by workers tied to the CPU.
3468 *
3469 * On return from this function, the current worker would trigger
3470 * unbound chain execution of pending work items if other workers
3471 * didn't already.
e22bee78 3472 */
4ce62e9e
TH
3473 for_each_worker_pool(pool, gcwq)
3474 atomic_set(get_pool_nr_running(pool), 0);
3af24433 3475}
3af24433 3476
8db25e78
TH
3477/*
3478 * Workqueues should be brought up before normal priority CPU notifiers.
3479 * This will be registered high priority CPU notifier.
3480 */
3481static int __devinit workqueue_cpu_up_callback(struct notifier_block *nfb,
3482 unsigned long action,
3483 void *hcpu)
3af24433
ON
3484{
3485 unsigned int cpu = (unsigned long)hcpu;
db7bccf4 3486 struct global_cwq *gcwq = get_gcwq(cpu);
4ce62e9e 3487 struct worker_pool *pool;
3ce63377 3488
8db25e78 3489 switch (action & ~CPU_TASKS_FROZEN) {
3af24433 3490 case CPU_UP_PREPARE:
4ce62e9e 3491 for_each_worker_pool(pool, gcwq) {
3ce63377
TH
3492 struct worker *worker;
3493
3494 if (pool->nr_workers)
3495 continue;
3496
3497 worker = create_worker(pool);
3498 if (!worker)
3499 return NOTIFY_BAD;
3500
3501 spin_lock_irq(&gcwq->lock);
3502 start_worker(worker);
3503 spin_unlock_irq(&gcwq->lock);
3af24433 3504 }
8db25e78 3505 break;
3af24433 3506
db7bccf4
TH
3507 case CPU_DOWN_FAILED:
3508 case CPU_ONLINE:
8db25e78 3509 gcwq_claim_management_and_lock(gcwq);
bc2ae0f5 3510 gcwq->flags &= ~GCWQ_DISASSOCIATED;
25511a47 3511 rebind_workers(gcwq);
8db25e78 3512 gcwq_release_management_and_unlock(gcwq);
db7bccf4 3513 break;
00dfcaf7 3514 }
65758202
TH
3515 return NOTIFY_OK;
3516}
3517
3518/*
3519 * Workqueues should be brought down after normal priority CPU notifiers.
3520 * This will be registered as low priority CPU notifier.
3521 */
3522static int __devinit workqueue_cpu_down_callback(struct notifier_block *nfb,
3523 unsigned long action,
3524 void *hcpu)
3525{
8db25e78
TH
3526 unsigned int cpu = (unsigned long)hcpu;
3527 struct work_struct unbind_work;
3528
65758202
TH
3529 switch (action & ~CPU_TASKS_FROZEN) {
3530 case CPU_DOWN_PREPARE:
8db25e78
TH
3531 /* unbinding should happen on the local CPU */
3532 INIT_WORK_ONSTACK(&unbind_work, gcwq_unbind_fn);
3533 schedule_work_on(cpu, &unbind_work);
3534 flush_work(&unbind_work);
3535 break;
65758202
TH
3536 }
3537 return NOTIFY_OK;
3538}
3539
2d3854a3 3540#ifdef CONFIG_SMP
8ccad40d 3541
2d3854a3 3542struct work_for_cpu {
6b44003e 3543 struct completion completion;
2d3854a3
RR
3544 long (*fn)(void *);
3545 void *arg;
3546 long ret;
3547};
3548
6b44003e 3549static int do_work_for_cpu(void *_wfc)
2d3854a3 3550{
6b44003e 3551 struct work_for_cpu *wfc = _wfc;
2d3854a3 3552 wfc->ret = wfc->fn(wfc->arg);
6b44003e
AM
3553 complete(&wfc->completion);
3554 return 0;
2d3854a3
RR
3555}
3556
3557/**
3558 * work_on_cpu - run a function in user context on a particular cpu
3559 * @cpu: the cpu to run on
3560 * @fn: the function to run
3561 * @arg: the function arg
3562 *
31ad9081
RR
3563 * This will return the value @fn returns.
3564 * It is up to the caller to ensure that the cpu doesn't go offline.
6b44003e 3565 * The caller must not hold any locks which would prevent @fn from completing.
2d3854a3
RR
3566 */
3567long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg)
3568{
6b44003e
AM
3569 struct task_struct *sub_thread;
3570 struct work_for_cpu wfc = {
3571 .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion),
3572 .fn = fn,
3573 .arg = arg,
3574 };
3575
3576 sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu");
3577 if (IS_ERR(sub_thread))
3578 return PTR_ERR(sub_thread);
3579 kthread_bind(sub_thread, cpu);
3580 wake_up_process(sub_thread);
3581 wait_for_completion(&wfc.completion);
2d3854a3
RR
3582 return wfc.ret;
3583}
3584EXPORT_SYMBOL_GPL(work_on_cpu);
3585#endif /* CONFIG_SMP */
3586
a0a1a5fd
TH
3587#ifdef CONFIG_FREEZER
3588
3589/**
3590 * freeze_workqueues_begin - begin freezing workqueues
3591 *
58a69cb4
TH
3592 * Start freezing workqueues. After this function returns, all freezable
3593 * workqueues will queue new works to their frozen_works list instead of
3594 * gcwq->worklist.
a0a1a5fd
TH
3595 *
3596 * CONTEXT:
8b03ae3c 3597 * Grabs and releases workqueue_lock and gcwq->lock's.
a0a1a5fd
TH
3598 */
3599void freeze_workqueues_begin(void)
3600{
a0a1a5fd
TH
3601 unsigned int cpu;
3602
3603 spin_lock(&workqueue_lock);
3604
3605 BUG_ON(workqueue_freezing);
3606 workqueue_freezing = true;
3607
f3421797 3608 for_each_gcwq_cpu(cpu) {
8b03ae3c 3609 struct global_cwq *gcwq = get_gcwq(cpu);
bdbc5dd7 3610 struct workqueue_struct *wq;
8b03ae3c
TH
3611
3612 spin_lock_irq(&gcwq->lock);
3613
db7bccf4
TH
3614 BUG_ON(gcwq->flags & GCWQ_FREEZING);
3615 gcwq->flags |= GCWQ_FREEZING;
3616
a0a1a5fd
TH
3617 list_for_each_entry(wq, &workqueues, list) {
3618 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3619
58a69cb4 3620 if (cwq && wq->flags & WQ_FREEZABLE)
a0a1a5fd 3621 cwq->max_active = 0;
a0a1a5fd 3622 }
8b03ae3c
TH
3623
3624 spin_unlock_irq(&gcwq->lock);
a0a1a5fd
TH
3625 }
3626
3627 spin_unlock(&workqueue_lock);
3628}
3629
3630/**
58a69cb4 3631 * freeze_workqueues_busy - are freezable workqueues still busy?
a0a1a5fd
TH
3632 *
3633 * Check whether freezing is complete. This function must be called
3634 * between freeze_workqueues_begin() and thaw_workqueues().
3635 *
3636 * CONTEXT:
3637 * Grabs and releases workqueue_lock.
3638 *
3639 * RETURNS:
58a69cb4
TH
3640 * %true if some freezable workqueues are still busy. %false if freezing
3641 * is complete.
a0a1a5fd
TH
3642 */
3643bool freeze_workqueues_busy(void)
3644{
a0a1a5fd
TH
3645 unsigned int cpu;
3646 bool busy = false;
3647
3648 spin_lock(&workqueue_lock);
3649
3650 BUG_ON(!workqueue_freezing);
3651
f3421797 3652 for_each_gcwq_cpu(cpu) {
bdbc5dd7 3653 struct workqueue_struct *wq;
a0a1a5fd
TH
3654 /*
3655 * nr_active is monotonically decreasing. It's safe
3656 * to peek without lock.
3657 */
3658 list_for_each_entry(wq, &workqueues, list) {
3659 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3660
58a69cb4 3661 if (!cwq || !(wq->flags & WQ_FREEZABLE))
a0a1a5fd
TH
3662 continue;
3663
3664 BUG_ON(cwq->nr_active < 0);
3665 if (cwq->nr_active) {
3666 busy = true;
3667 goto out_unlock;
3668 }
3669 }
3670 }
3671out_unlock:
3672 spin_unlock(&workqueue_lock);
3673 return busy;
3674}
3675
3676/**
3677 * thaw_workqueues - thaw workqueues
3678 *
3679 * Thaw workqueues. Normal queueing is restored and all collected
7e11629d 3680 * frozen works are transferred to their respective gcwq worklists.
a0a1a5fd
TH
3681 *
3682 * CONTEXT:
8b03ae3c 3683 * Grabs and releases workqueue_lock and gcwq->lock's.
a0a1a5fd
TH
3684 */
3685void thaw_workqueues(void)
3686{
a0a1a5fd
TH
3687 unsigned int cpu;
3688
3689 spin_lock(&workqueue_lock);
3690
3691 if (!workqueue_freezing)
3692 goto out_unlock;
3693
f3421797 3694 for_each_gcwq_cpu(cpu) {
8b03ae3c 3695 struct global_cwq *gcwq = get_gcwq(cpu);
4ce62e9e 3696 struct worker_pool *pool;
bdbc5dd7 3697 struct workqueue_struct *wq;
8b03ae3c
TH
3698
3699 spin_lock_irq(&gcwq->lock);
3700
db7bccf4
TH
3701 BUG_ON(!(gcwq->flags & GCWQ_FREEZING));
3702 gcwq->flags &= ~GCWQ_FREEZING;
3703
a0a1a5fd
TH
3704 list_for_each_entry(wq, &workqueues, list) {
3705 struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq);
3706
58a69cb4 3707 if (!cwq || !(wq->flags & WQ_FREEZABLE))
a0a1a5fd
TH
3708 continue;
3709
a0a1a5fd
TH
3710 /* restore max_active and repopulate worklist */
3711 cwq->max_active = wq->saved_max_active;
3712
3713 while (!list_empty(&cwq->delayed_works) &&
3714 cwq->nr_active < cwq->max_active)
3715 cwq_activate_first_delayed(cwq);
a0a1a5fd 3716 }
8b03ae3c 3717
4ce62e9e
TH
3718 for_each_worker_pool(pool, gcwq)
3719 wake_up_worker(pool);
e22bee78 3720
8b03ae3c 3721 spin_unlock_irq(&gcwq->lock);
a0a1a5fd
TH
3722 }
3723
3724 workqueue_freezing = false;
3725out_unlock:
3726 spin_unlock(&workqueue_lock);
3727}
3728#endif /* CONFIG_FREEZER */
3729
6ee0578b 3730static int __init init_workqueues(void)
1da177e4 3731{
c34056a3 3732 unsigned int cpu;
c8e55f36 3733 int i;
c34056a3 3734
b5490077
TH
3735 /* make sure we have enough bits for OFFQ CPU number */
3736 BUILD_BUG_ON((1LU << (BITS_PER_LONG - WORK_OFFQ_CPU_SHIFT)) <
3737 WORK_CPU_LAST);
3738
65758202
TH
3739 cpu_notifier(workqueue_cpu_up_callback, CPU_PRI_WORKQUEUE_UP);
3740 cpu_notifier(workqueue_cpu_down_callback, CPU_PRI_WORKQUEUE_DOWN);
8b03ae3c
TH
3741
3742 /* initialize gcwqs */
f3421797 3743 for_each_gcwq_cpu(cpu) {
8b03ae3c 3744 struct global_cwq *gcwq = get_gcwq(cpu);
4ce62e9e 3745 struct worker_pool *pool;
8b03ae3c
TH
3746
3747 spin_lock_init(&gcwq->lock);
3748 gcwq->cpu = cpu;
477a3c33 3749 gcwq->flags |= GCWQ_DISASSOCIATED;
8b03ae3c 3750
c8e55f36
TH
3751 for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++)
3752 INIT_HLIST_HEAD(&gcwq->busy_hash[i]);
3753
4ce62e9e
TH
3754 for_each_worker_pool(pool, gcwq) {
3755 pool->gcwq = gcwq;
3756 INIT_LIST_HEAD(&pool->worklist);
3757 INIT_LIST_HEAD(&pool->idle_list);
e7577c50 3758
4ce62e9e
TH
3759 init_timer_deferrable(&pool->idle_timer);
3760 pool->idle_timer.function = idle_worker_timeout;
3761 pool->idle_timer.data = (unsigned long)pool;
e22bee78 3762
4ce62e9e
TH
3763 setup_timer(&pool->mayday_timer, gcwq_mayday_timeout,
3764 (unsigned long)pool);
3765
60373152 3766 mutex_init(&pool->manager_mutex);
4ce62e9e
TH
3767 ida_init(&pool->worker_ida);
3768 }
db7bccf4 3769
25511a47 3770 init_waitqueue_head(&gcwq->rebind_hold);
8b03ae3c
TH
3771 }
3772
e22bee78 3773 /* create the initial worker */
f3421797 3774 for_each_online_gcwq_cpu(cpu) {
e22bee78 3775 struct global_cwq *gcwq = get_gcwq(cpu);
4ce62e9e 3776 struct worker_pool *pool;
e22bee78 3777
477a3c33
TH
3778 if (cpu != WORK_CPU_UNBOUND)
3779 gcwq->flags &= ~GCWQ_DISASSOCIATED;
4ce62e9e
TH
3780
3781 for_each_worker_pool(pool, gcwq) {
3782 struct worker *worker;
3783
bc2ae0f5 3784 worker = create_worker(pool);
4ce62e9e
TH
3785 BUG_ON(!worker);
3786 spin_lock_irq(&gcwq->lock);
3787 start_worker(worker);
3788 spin_unlock_irq(&gcwq->lock);
3789 }
e22bee78
TH
3790 }
3791
d320c038
TH
3792 system_wq = alloc_workqueue("events", 0, 0);
3793 system_long_wq = alloc_workqueue("events_long", 0, 0);
3794 system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0);
f3421797
TH
3795 system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND,
3796 WQ_UNBOUND_MAX_ACTIVE);
24d51add
TH
3797 system_freezable_wq = alloc_workqueue("events_freezable",
3798 WQ_FREEZABLE, 0);
62d3c543
AS
3799 system_nrt_freezable_wq = alloc_workqueue("events_nrt_freezable",
3800 WQ_NON_REENTRANT | WQ_FREEZABLE, 0);
e5cba24e 3801 BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq ||
62d3c543
AS
3802 !system_unbound_wq || !system_freezable_wq ||
3803 !system_nrt_freezable_wq);
6ee0578b 3804 return 0;
1da177e4 3805}
6ee0578b 3806early_initcall(init_workqueues);