]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/kernel/workqueue.c | |
3 | * | |
4 | * Generic mechanism for defining kernel helper threads for running | |
5 | * arbitrary tasks in process context. | |
6 | * | |
7 | * Started by Ingo Molnar, Copyright (C) 2002 | |
8 | * | |
9 | * Derived from the taskqueue/keventd code by: | |
10 | * | |
11 | * David Woodhouse <dwmw2@infradead.org> | |
e1f8e874 | 12 | * Andrew Morton |
1da177e4 LT |
13 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> |
14 | * Theodore Ts'o <tytso@mit.edu> | |
89ada679 | 15 | * |
cde53535 | 16 | * Made to use alloc_percpu by Christoph Lameter. |
1da177e4 LT |
17 | */ |
18 | ||
19 | #include <linux/module.h> | |
20 | #include <linux/kernel.h> | |
21 | #include <linux/sched.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/signal.h> | |
24 | #include <linux/completion.h> | |
25 | #include <linux/workqueue.h> | |
26 | #include <linux/slab.h> | |
27 | #include <linux/cpu.h> | |
28 | #include <linux/notifier.h> | |
29 | #include <linux/kthread.h> | |
1fa44eca | 30 | #include <linux/hardirq.h> |
46934023 | 31 | #include <linux/mempolicy.h> |
341a5958 | 32 | #include <linux/freezer.h> |
d5abe669 PZ |
33 | #include <linux/kallsyms.h> |
34 | #include <linux/debug_locks.h> | |
4e6045f1 | 35 | #include <linux/lockdep.h> |
c34056a3 | 36 | #include <linux/idr.h> |
e22bee78 TH |
37 | |
38 | #include "workqueue_sched.h" | |
1da177e4 | 39 | |
c8e55f36 | 40 | enum { |
db7bccf4 | 41 | /* global_cwq flags */ |
e22bee78 TH |
42 | GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ |
43 | GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */ | |
44 | GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ | |
db7bccf4 | 45 | GCWQ_FREEZING = 1 << 3, /* freeze in progress */ |
649027d7 | 46 | GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */ |
db7bccf4 | 47 | |
c8e55f36 TH |
48 | /* worker flags */ |
49 | WORKER_STARTED = 1 << 0, /* started */ | |
50 | WORKER_DIE = 1 << 1, /* die die die */ | |
51 | WORKER_IDLE = 1 << 2, /* is idle */ | |
e22bee78 | 52 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
db7bccf4 | 53 | WORKER_ROGUE = 1 << 4, /* not bound to any cpu */ |
e22bee78 | 54 | WORKER_REBIND = 1 << 5, /* mom is home, come back */ |
fb0e7beb | 55 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
e22bee78 | 56 | |
fb0e7beb TH |
57 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND | |
58 | WORKER_CPU_INTENSIVE, | |
db7bccf4 TH |
59 | |
60 | /* gcwq->trustee_state */ | |
61 | TRUSTEE_START = 0, /* start */ | |
62 | TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */ | |
63 | TRUSTEE_BUTCHER = 2, /* butcher workers */ | |
64 | TRUSTEE_RELEASE = 3, /* release workers */ | |
65 | TRUSTEE_DONE = 4, /* trustee is done */ | |
c8e55f36 TH |
66 | |
67 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ | |
68 | BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, | |
69 | BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, | |
db7bccf4 | 70 | |
e22bee78 TH |
71 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
72 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ | |
73 | ||
74 | MAYDAY_INITIAL_TIMEOUT = HZ / 100, /* call for help after 10ms */ | |
75 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ | |
76 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ | |
db7bccf4 | 77 | TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ |
e22bee78 TH |
78 | |
79 | /* | |
80 | * Rescue workers are used only on emergencies and shared by | |
81 | * all cpus. Give -20. | |
82 | */ | |
83 | RESCUER_NICE_LEVEL = -20, | |
c8e55f36 TH |
84 | }; |
85 | ||
4690c4ab TH |
86 | /* |
87 | * Structure fields follow one of the following exclusion rules. | |
88 | * | |
89 | * I: Set during initialization and read-only afterwards. | |
90 | * | |
e22bee78 TH |
91 | * P: Preemption protected. Disabling preemption is enough and should |
92 | * only be modified and accessed from the local cpu. | |
93 | * | |
8b03ae3c | 94 | * L: gcwq->lock protected. Access with gcwq->lock held. |
4690c4ab | 95 | * |
e22bee78 TH |
96 | * X: During normal operation, modification requires gcwq->lock and |
97 | * should be done only from local cpu. Either disabling preemption | |
98 | * on local cpu or grabbing gcwq->lock is enough for read access. | |
99 | * While trustee is in charge, it's identical to L. | |
100 | * | |
73f53c4a TH |
101 | * F: wq->flush_mutex protected. |
102 | * | |
4690c4ab TH |
103 | * W: workqueue_lock protected. |
104 | */ | |
105 | ||
8b03ae3c | 106 | struct global_cwq; |
c34056a3 | 107 | |
e22bee78 TH |
108 | /* |
109 | * The poor guys doing the actual heavy lifting. All on-duty workers | |
110 | * are either serving the manager role, on idle list or on busy hash. | |
111 | */ | |
c34056a3 | 112 | struct worker { |
c8e55f36 TH |
113 | /* on idle list while idle, on busy hash table while busy */ |
114 | union { | |
115 | struct list_head entry; /* L: while idle */ | |
116 | struct hlist_node hentry; /* L: while busy */ | |
117 | }; | |
118 | ||
c34056a3 | 119 | struct work_struct *current_work; /* L: work being processed */ |
8cca0eea | 120 | struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ |
affee4b2 | 121 | struct list_head scheduled; /* L: scheduled works */ |
c34056a3 | 122 | struct task_struct *task; /* I: worker task */ |
8b03ae3c | 123 | struct global_cwq *gcwq; /* I: the associated gcwq */ |
e22bee78 TH |
124 | /* 64 bytes boundary on 64bit, 32 on 32bit */ |
125 | unsigned long last_active; /* L: last active timestamp */ | |
126 | unsigned int flags; /* X: flags */ | |
c34056a3 | 127 | int id; /* I: worker id */ |
e22bee78 | 128 | struct work_struct rebind_work; /* L: rebind worker to cpu */ |
c34056a3 TH |
129 | }; |
130 | ||
8b03ae3c | 131 | /* |
e22bee78 TH |
132 | * Global per-cpu workqueue. There's one and only one for each cpu |
133 | * and all works are queued and processed here regardless of their | |
134 | * target workqueues. | |
8b03ae3c TH |
135 | */ |
136 | struct global_cwq { | |
137 | spinlock_t lock; /* the gcwq lock */ | |
7e11629d | 138 | struct list_head worklist; /* L: list of pending works */ |
8b03ae3c | 139 | unsigned int cpu; /* I: the associated cpu */ |
db7bccf4 | 140 | unsigned int flags; /* L: GCWQ_* flags */ |
c8e55f36 TH |
141 | |
142 | int nr_workers; /* L: total number of workers */ | |
143 | int nr_idle; /* L: currently idle ones */ | |
144 | ||
145 | /* workers are chained either in the idle_list or busy_hash */ | |
e22bee78 | 146 | struct list_head idle_list; /* X: list of idle workers */ |
c8e55f36 TH |
147 | struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; |
148 | /* L: hash of busy workers */ | |
149 | ||
e22bee78 TH |
150 | struct timer_list idle_timer; /* L: worker idle timeout */ |
151 | struct timer_list mayday_timer; /* L: SOS timer for dworkers */ | |
152 | ||
8b03ae3c | 153 | struct ida worker_ida; /* L: for worker IDs */ |
db7bccf4 TH |
154 | |
155 | struct task_struct *trustee; /* L: for gcwq shutdown */ | |
156 | unsigned int trustee_state; /* L: trustee state */ | |
157 | wait_queue_head_t trustee_wait; /* trustee wait */ | |
e22bee78 | 158 | struct worker *first_idle; /* L: first idle worker */ |
8b03ae3c TH |
159 | } ____cacheline_aligned_in_smp; |
160 | ||
1da177e4 | 161 | /* |
502ca9d8 | 162 | * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of |
0f900049 TH |
163 | * work_struct->data are used for flags and thus cwqs need to be |
164 | * aligned at two's power of the number of flag bits. | |
1da177e4 LT |
165 | */ |
166 | struct cpu_workqueue_struct { | |
8b03ae3c | 167 | struct global_cwq *gcwq; /* I: the associated gcwq */ |
4690c4ab | 168 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
73f53c4a TH |
169 | int work_color; /* L: current color */ |
170 | int flush_color; /* L: flushing color */ | |
171 | int nr_in_flight[WORK_NR_COLORS]; | |
172 | /* L: nr of in_flight works */ | |
1e19ffc6 | 173 | int nr_active; /* L: nr of active works */ |
a0a1a5fd | 174 | int max_active; /* L: max active works */ |
1e19ffc6 | 175 | struct list_head delayed_works; /* L: delayed works */ |
0f900049 | 176 | }; |
1da177e4 | 177 | |
73f53c4a TH |
178 | /* |
179 | * Structure used to wait for workqueue flush. | |
180 | */ | |
181 | struct wq_flusher { | |
182 | struct list_head list; /* F: list of flushers */ | |
183 | int flush_color; /* F: flush color waiting for */ | |
184 | struct completion done; /* flush completion */ | |
185 | }; | |
186 | ||
1da177e4 LT |
187 | /* |
188 | * The externally visible workqueue abstraction is an array of | |
189 | * per-CPU workqueues: | |
190 | */ | |
191 | struct workqueue_struct { | |
97e37d7b | 192 | unsigned int flags; /* I: WQ_* flags */ |
4690c4ab TH |
193 | struct cpu_workqueue_struct *cpu_wq; /* I: cwq's */ |
194 | struct list_head list; /* W: list of all workqueues */ | |
73f53c4a TH |
195 | |
196 | struct mutex flush_mutex; /* protects wq flushing */ | |
197 | int work_color; /* F: current work color */ | |
198 | int flush_color; /* F: current flush color */ | |
199 | atomic_t nr_cwqs_to_flush; /* flush in progress */ | |
200 | struct wq_flusher *first_flusher; /* F: first flusher */ | |
201 | struct list_head flusher_queue; /* F: flush waiters */ | |
202 | struct list_head flusher_overflow; /* F: flush overflow list */ | |
203 | ||
502ca9d8 TH |
204 | unsigned long single_cpu; /* cpu for single cpu wq */ |
205 | ||
e22bee78 TH |
206 | cpumask_var_t mayday_mask; /* cpus requesting rescue */ |
207 | struct worker *rescuer; /* I: rescue worker */ | |
208 | ||
dcd989cb | 209 | int saved_max_active; /* W: saved cwq max_active */ |
4690c4ab | 210 | const char *name; /* I: workqueue name */ |
4e6045f1 | 211 | #ifdef CONFIG_LOCKDEP |
4690c4ab | 212 | struct lockdep_map lockdep_map; |
4e6045f1 | 213 | #endif |
1da177e4 LT |
214 | }; |
215 | ||
d320c038 TH |
216 | struct workqueue_struct *system_wq __read_mostly; |
217 | struct workqueue_struct *system_long_wq __read_mostly; | |
218 | struct workqueue_struct *system_nrt_wq __read_mostly; | |
219 | EXPORT_SYMBOL_GPL(system_wq); | |
220 | EXPORT_SYMBOL_GPL(system_long_wq); | |
221 | EXPORT_SYMBOL_GPL(system_nrt_wq); | |
222 | ||
db7bccf4 TH |
223 | #define for_each_busy_worker(worker, i, pos, gcwq) \ |
224 | for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ | |
225 | hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) | |
226 | ||
dc186ad7 TG |
227 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
228 | ||
229 | static struct debug_obj_descr work_debug_descr; | |
230 | ||
231 | /* | |
232 | * fixup_init is called when: | |
233 | * - an active object is initialized | |
234 | */ | |
235 | static int work_fixup_init(void *addr, enum debug_obj_state state) | |
236 | { | |
237 | struct work_struct *work = addr; | |
238 | ||
239 | switch (state) { | |
240 | case ODEBUG_STATE_ACTIVE: | |
241 | cancel_work_sync(work); | |
242 | debug_object_init(work, &work_debug_descr); | |
243 | return 1; | |
244 | default: | |
245 | return 0; | |
246 | } | |
247 | } | |
248 | ||
249 | /* | |
250 | * fixup_activate is called when: | |
251 | * - an active object is activated | |
252 | * - an unknown object is activated (might be a statically initialized object) | |
253 | */ | |
254 | static int work_fixup_activate(void *addr, enum debug_obj_state state) | |
255 | { | |
256 | struct work_struct *work = addr; | |
257 | ||
258 | switch (state) { | |
259 | ||
260 | case ODEBUG_STATE_NOTAVAILABLE: | |
261 | /* | |
262 | * This is not really a fixup. The work struct was | |
263 | * statically initialized. We just make sure that it | |
264 | * is tracked in the object tracker. | |
265 | */ | |
22df02bb | 266 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
dc186ad7 TG |
267 | debug_object_init(work, &work_debug_descr); |
268 | debug_object_activate(work, &work_debug_descr); | |
269 | return 0; | |
270 | } | |
271 | WARN_ON_ONCE(1); | |
272 | return 0; | |
273 | ||
274 | case ODEBUG_STATE_ACTIVE: | |
275 | WARN_ON(1); | |
276 | ||
277 | default: | |
278 | return 0; | |
279 | } | |
280 | } | |
281 | ||
282 | /* | |
283 | * fixup_free is called when: | |
284 | * - an active object is freed | |
285 | */ | |
286 | static int work_fixup_free(void *addr, enum debug_obj_state state) | |
287 | { | |
288 | struct work_struct *work = addr; | |
289 | ||
290 | switch (state) { | |
291 | case ODEBUG_STATE_ACTIVE: | |
292 | cancel_work_sync(work); | |
293 | debug_object_free(work, &work_debug_descr); | |
294 | return 1; | |
295 | default: | |
296 | return 0; | |
297 | } | |
298 | } | |
299 | ||
300 | static struct debug_obj_descr work_debug_descr = { | |
301 | .name = "work_struct", | |
302 | .fixup_init = work_fixup_init, | |
303 | .fixup_activate = work_fixup_activate, | |
304 | .fixup_free = work_fixup_free, | |
305 | }; | |
306 | ||
307 | static inline void debug_work_activate(struct work_struct *work) | |
308 | { | |
309 | debug_object_activate(work, &work_debug_descr); | |
310 | } | |
311 | ||
312 | static inline void debug_work_deactivate(struct work_struct *work) | |
313 | { | |
314 | debug_object_deactivate(work, &work_debug_descr); | |
315 | } | |
316 | ||
317 | void __init_work(struct work_struct *work, int onstack) | |
318 | { | |
319 | if (onstack) | |
320 | debug_object_init_on_stack(work, &work_debug_descr); | |
321 | else | |
322 | debug_object_init(work, &work_debug_descr); | |
323 | } | |
324 | EXPORT_SYMBOL_GPL(__init_work); | |
325 | ||
326 | void destroy_work_on_stack(struct work_struct *work) | |
327 | { | |
328 | debug_object_free(work, &work_debug_descr); | |
329 | } | |
330 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); | |
331 | ||
332 | #else | |
333 | static inline void debug_work_activate(struct work_struct *work) { } | |
334 | static inline void debug_work_deactivate(struct work_struct *work) { } | |
335 | #endif | |
336 | ||
95402b38 GS |
337 | /* Serializes the accesses to the list of workqueues. */ |
338 | static DEFINE_SPINLOCK(workqueue_lock); | |
1da177e4 | 339 | static LIST_HEAD(workqueues); |
a0a1a5fd | 340 | static bool workqueue_freezing; /* W: have wqs started freezing? */ |
c34056a3 | 341 | |
e22bee78 TH |
342 | /* |
343 | * The almighty global cpu workqueues. nr_running is the only field | |
344 | * which is expected to be used frequently by other cpus via | |
345 | * try_to_wake_up(). Put it in a separate cacheline. | |
346 | */ | |
8b03ae3c | 347 | static DEFINE_PER_CPU(struct global_cwq, global_cwq); |
e22bee78 | 348 | static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running); |
8b03ae3c | 349 | |
c34056a3 | 350 | static int worker_thread(void *__worker); |
1da177e4 | 351 | |
8b03ae3c TH |
352 | static struct global_cwq *get_gcwq(unsigned int cpu) |
353 | { | |
354 | return &per_cpu(global_cwq, cpu); | |
355 | } | |
356 | ||
e22bee78 TH |
357 | static atomic_t *get_gcwq_nr_running(unsigned int cpu) |
358 | { | |
359 | return &per_cpu(gcwq_nr_running, cpu); | |
360 | } | |
361 | ||
1537663f TH |
362 | static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, |
363 | struct workqueue_struct *wq) | |
b1f4ec17 | 364 | { |
1537663f | 365 | return per_cpu_ptr(wq->cpu_wq, cpu); |
b1f4ec17 ON |
366 | } |
367 | ||
73f53c4a TH |
368 | static unsigned int work_color_to_flags(int color) |
369 | { | |
370 | return color << WORK_STRUCT_COLOR_SHIFT; | |
371 | } | |
372 | ||
373 | static int get_work_color(struct work_struct *work) | |
374 | { | |
375 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & | |
376 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); | |
377 | } | |
378 | ||
379 | static int work_next_color(int color) | |
380 | { | |
381 | return (color + 1) % WORK_NR_COLORS; | |
382 | } | |
383 | ||
4594bf15 | 384 | /* |
7a22ad75 TH |
385 | * Work data points to the cwq while a work is on queue. Once |
386 | * execution starts, it points to the cpu the work was last on. This | |
387 | * can be distinguished by comparing the data value against | |
388 | * PAGE_OFFSET. | |
389 | * | |
390 | * set_work_{cwq|cpu}() and clear_work_data() can be used to set the | |
391 | * cwq, cpu or clear work->data. These functions should only be | |
392 | * called while the work is owned - ie. while the PENDING bit is set. | |
393 | * | |
394 | * get_work_[g]cwq() can be used to obtain the gcwq or cwq | |
395 | * corresponding to a work. gcwq is available once the work has been | |
396 | * queued anywhere after initialization. cwq is available only from | |
397 | * queueing until execution starts. | |
4594bf15 | 398 | */ |
7a22ad75 TH |
399 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
400 | unsigned long flags) | |
365970a1 | 401 | { |
4594bf15 | 402 | BUG_ON(!work_pending(work)); |
7a22ad75 TH |
403 | atomic_long_set(&work->data, data | flags | work_static(work)); |
404 | } | |
365970a1 | 405 | |
7a22ad75 TH |
406 | static void set_work_cwq(struct work_struct *work, |
407 | struct cpu_workqueue_struct *cwq, | |
408 | unsigned long extra_flags) | |
409 | { | |
410 | set_work_data(work, (unsigned long)cwq, | |
411 | WORK_STRUCT_PENDING | extra_flags); | |
365970a1 DH |
412 | } |
413 | ||
7a22ad75 TH |
414 | static void set_work_cpu(struct work_struct *work, unsigned int cpu) |
415 | { | |
416 | set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); | |
417 | } | |
418 | ||
419 | static void clear_work_data(struct work_struct *work) | |
420 | { | |
421 | set_work_data(work, WORK_STRUCT_NO_CPU, 0); | |
422 | } | |
423 | ||
424 | static inline unsigned long get_work_data(struct work_struct *work) | |
425 | { | |
426 | return atomic_long_read(&work->data) & WORK_STRUCT_WQ_DATA_MASK; | |
427 | } | |
428 | ||
429 | static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) | |
4d707b9f | 430 | { |
7a22ad75 TH |
431 | unsigned long data = get_work_data(work); |
432 | ||
433 | return data >= PAGE_OFFSET ? (void *)data : NULL; | |
4d707b9f ON |
434 | } |
435 | ||
7a22ad75 | 436 | static struct global_cwq *get_work_gcwq(struct work_struct *work) |
365970a1 | 437 | { |
7a22ad75 TH |
438 | unsigned long data = get_work_data(work); |
439 | unsigned int cpu; | |
440 | ||
441 | if (data >= PAGE_OFFSET) | |
442 | return ((struct cpu_workqueue_struct *)data)->gcwq; | |
443 | ||
444 | cpu = data >> WORK_STRUCT_FLAG_BITS; | |
445 | if (cpu == NR_CPUS) | |
446 | return NULL; | |
447 | ||
a1e453d2 | 448 | BUG_ON(cpu >= nr_cpu_ids); |
7a22ad75 | 449 | return get_gcwq(cpu); |
365970a1 DH |
450 | } |
451 | ||
e22bee78 TH |
452 | /* |
453 | * Policy functions. These define the policies on how the global | |
454 | * worker pool is managed. Unless noted otherwise, these functions | |
455 | * assume that they're being called with gcwq->lock held. | |
456 | */ | |
457 | ||
649027d7 TH |
458 | static bool __need_more_worker(struct global_cwq *gcwq) |
459 | { | |
460 | return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) || | |
461 | gcwq->flags & GCWQ_HIGHPRI_PENDING; | |
462 | } | |
463 | ||
e22bee78 TH |
464 | /* |
465 | * Need to wake up a worker? Called from anything but currently | |
466 | * running workers. | |
467 | */ | |
468 | static bool need_more_worker(struct global_cwq *gcwq) | |
469 | { | |
649027d7 | 470 | return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq); |
e22bee78 TH |
471 | } |
472 | ||
473 | /* Can I start working? Called from busy but !running workers. */ | |
474 | static bool may_start_working(struct global_cwq *gcwq) | |
475 | { | |
476 | return gcwq->nr_idle; | |
477 | } | |
478 | ||
479 | /* Do I need to keep working? Called from currently running workers. */ | |
480 | static bool keep_working(struct global_cwq *gcwq) | |
481 | { | |
482 | atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); | |
483 | ||
484 | return !list_empty(&gcwq->worklist) && atomic_read(nr_running) <= 1; | |
485 | } | |
486 | ||
487 | /* Do we need a new worker? Called from manager. */ | |
488 | static bool need_to_create_worker(struct global_cwq *gcwq) | |
489 | { | |
490 | return need_more_worker(gcwq) && !may_start_working(gcwq); | |
491 | } | |
492 | ||
493 | /* Do I need to be the manager? */ | |
494 | static bool need_to_manage_workers(struct global_cwq *gcwq) | |
495 | { | |
496 | return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS; | |
497 | } | |
498 | ||
499 | /* Do we have too many workers and should some go away? */ | |
500 | static bool too_many_workers(struct global_cwq *gcwq) | |
501 | { | |
502 | bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS; | |
503 | int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */ | |
504 | int nr_busy = gcwq->nr_workers - nr_idle; | |
505 | ||
506 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; | |
507 | } | |
508 | ||
509 | /* | |
510 | * Wake up functions. | |
511 | */ | |
512 | ||
7e11629d TH |
513 | /* Return the first worker. Safe with preemption disabled */ |
514 | static struct worker *first_worker(struct global_cwq *gcwq) | |
515 | { | |
516 | if (unlikely(list_empty(&gcwq->idle_list))) | |
517 | return NULL; | |
518 | ||
519 | return list_first_entry(&gcwq->idle_list, struct worker, entry); | |
520 | } | |
521 | ||
522 | /** | |
523 | * wake_up_worker - wake up an idle worker | |
524 | * @gcwq: gcwq to wake worker for | |
525 | * | |
526 | * Wake up the first idle worker of @gcwq. | |
527 | * | |
528 | * CONTEXT: | |
529 | * spin_lock_irq(gcwq->lock). | |
530 | */ | |
531 | static void wake_up_worker(struct global_cwq *gcwq) | |
532 | { | |
533 | struct worker *worker = first_worker(gcwq); | |
534 | ||
535 | if (likely(worker)) | |
536 | wake_up_process(worker->task); | |
537 | } | |
538 | ||
d302f017 | 539 | /** |
e22bee78 TH |
540 | * wq_worker_waking_up - a worker is waking up |
541 | * @task: task waking up | |
542 | * @cpu: CPU @task is waking up to | |
543 | * | |
544 | * This function is called during try_to_wake_up() when a worker is | |
545 | * being awoken. | |
546 | * | |
547 | * CONTEXT: | |
548 | * spin_lock_irq(rq->lock) | |
549 | */ | |
550 | void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) | |
551 | { | |
552 | struct worker *worker = kthread_data(task); | |
553 | ||
554 | if (likely(!(worker->flags & WORKER_NOT_RUNNING))) | |
555 | atomic_inc(get_gcwq_nr_running(cpu)); | |
556 | } | |
557 | ||
558 | /** | |
559 | * wq_worker_sleeping - a worker is going to sleep | |
560 | * @task: task going to sleep | |
561 | * @cpu: CPU in question, must be the current CPU number | |
562 | * | |
563 | * This function is called during schedule() when a busy worker is | |
564 | * going to sleep. Worker on the same cpu can be woken up by | |
565 | * returning pointer to its task. | |
566 | * | |
567 | * CONTEXT: | |
568 | * spin_lock_irq(rq->lock) | |
569 | * | |
570 | * RETURNS: | |
571 | * Worker task on @cpu to wake up, %NULL if none. | |
572 | */ | |
573 | struct task_struct *wq_worker_sleeping(struct task_struct *task, | |
574 | unsigned int cpu) | |
575 | { | |
576 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; | |
577 | struct global_cwq *gcwq = get_gcwq(cpu); | |
578 | atomic_t *nr_running = get_gcwq_nr_running(cpu); | |
579 | ||
580 | if (unlikely(worker->flags & WORKER_NOT_RUNNING)) | |
581 | return NULL; | |
582 | ||
583 | /* this can only happen on the local cpu */ | |
584 | BUG_ON(cpu != raw_smp_processor_id()); | |
585 | ||
586 | /* | |
587 | * The counterpart of the following dec_and_test, implied mb, | |
588 | * worklist not empty test sequence is in insert_work(). | |
589 | * Please read comment there. | |
590 | * | |
591 | * NOT_RUNNING is clear. This means that trustee is not in | |
592 | * charge and we're running on the local cpu w/ rq lock held | |
593 | * and preemption disabled, which in turn means that none else | |
594 | * could be manipulating idle_list, so dereferencing idle_list | |
595 | * without gcwq lock is safe. | |
596 | */ | |
597 | if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist)) | |
598 | to_wakeup = first_worker(gcwq); | |
599 | return to_wakeup ? to_wakeup->task : NULL; | |
600 | } | |
601 | ||
602 | /** | |
603 | * worker_set_flags - set worker flags and adjust nr_running accordingly | |
cb444766 | 604 | * @worker: self |
d302f017 TH |
605 | * @flags: flags to set |
606 | * @wakeup: wakeup an idle worker if necessary | |
607 | * | |
e22bee78 TH |
608 | * Set @flags in @worker->flags and adjust nr_running accordingly. If |
609 | * nr_running becomes zero and @wakeup is %true, an idle worker is | |
610 | * woken up. | |
d302f017 | 611 | * |
cb444766 TH |
612 | * CONTEXT: |
613 | * spin_lock_irq(gcwq->lock) | |
d302f017 TH |
614 | */ |
615 | static inline void worker_set_flags(struct worker *worker, unsigned int flags, | |
616 | bool wakeup) | |
617 | { | |
e22bee78 TH |
618 | struct global_cwq *gcwq = worker->gcwq; |
619 | ||
cb444766 TH |
620 | WARN_ON_ONCE(worker->task != current); |
621 | ||
e22bee78 TH |
622 | /* |
623 | * If transitioning into NOT_RUNNING, adjust nr_running and | |
624 | * wake up an idle worker as necessary if requested by | |
625 | * @wakeup. | |
626 | */ | |
627 | if ((flags & WORKER_NOT_RUNNING) && | |
628 | !(worker->flags & WORKER_NOT_RUNNING)) { | |
629 | atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); | |
630 | ||
631 | if (wakeup) { | |
632 | if (atomic_dec_and_test(nr_running) && | |
633 | !list_empty(&gcwq->worklist)) | |
634 | wake_up_worker(gcwq); | |
635 | } else | |
636 | atomic_dec(nr_running); | |
637 | } | |
638 | ||
d302f017 TH |
639 | worker->flags |= flags; |
640 | } | |
641 | ||
642 | /** | |
e22bee78 | 643 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
cb444766 | 644 | * @worker: self |
d302f017 TH |
645 | * @flags: flags to clear |
646 | * | |
e22bee78 | 647 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
d302f017 | 648 | * |
cb444766 TH |
649 | * CONTEXT: |
650 | * spin_lock_irq(gcwq->lock) | |
d302f017 TH |
651 | */ |
652 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) | |
653 | { | |
e22bee78 TH |
654 | struct global_cwq *gcwq = worker->gcwq; |
655 | unsigned int oflags = worker->flags; | |
656 | ||
cb444766 TH |
657 | WARN_ON_ONCE(worker->task != current); |
658 | ||
d302f017 | 659 | worker->flags &= ~flags; |
e22bee78 TH |
660 | |
661 | /* if transitioning out of NOT_RUNNING, increment nr_running */ | |
662 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) | |
663 | if (!(worker->flags & WORKER_NOT_RUNNING)) | |
664 | atomic_inc(get_gcwq_nr_running(gcwq->cpu)); | |
d302f017 TH |
665 | } |
666 | ||
c8e55f36 TH |
667 | /** |
668 | * busy_worker_head - return the busy hash head for a work | |
669 | * @gcwq: gcwq of interest | |
670 | * @work: work to be hashed | |
671 | * | |
672 | * Return hash head of @gcwq for @work. | |
673 | * | |
674 | * CONTEXT: | |
675 | * spin_lock_irq(gcwq->lock). | |
676 | * | |
677 | * RETURNS: | |
678 | * Pointer to the hash head. | |
679 | */ | |
680 | static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, | |
681 | struct work_struct *work) | |
682 | { | |
683 | const int base_shift = ilog2(sizeof(struct work_struct)); | |
684 | unsigned long v = (unsigned long)work; | |
685 | ||
686 | /* simple shift and fold hash, do we need something better? */ | |
687 | v >>= base_shift; | |
688 | v += v >> BUSY_WORKER_HASH_ORDER; | |
689 | v &= BUSY_WORKER_HASH_MASK; | |
690 | ||
691 | return &gcwq->busy_hash[v]; | |
692 | } | |
693 | ||
8cca0eea TH |
694 | /** |
695 | * __find_worker_executing_work - find worker which is executing a work | |
696 | * @gcwq: gcwq of interest | |
697 | * @bwh: hash head as returned by busy_worker_head() | |
698 | * @work: work to find worker for | |
699 | * | |
700 | * Find a worker which is executing @work on @gcwq. @bwh should be | |
701 | * the hash head obtained by calling busy_worker_head() with the same | |
702 | * work. | |
703 | * | |
704 | * CONTEXT: | |
705 | * spin_lock_irq(gcwq->lock). | |
706 | * | |
707 | * RETURNS: | |
708 | * Pointer to worker which is executing @work if found, NULL | |
709 | * otherwise. | |
710 | */ | |
711 | static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, | |
712 | struct hlist_head *bwh, | |
713 | struct work_struct *work) | |
714 | { | |
715 | struct worker *worker; | |
716 | struct hlist_node *tmp; | |
717 | ||
718 | hlist_for_each_entry(worker, tmp, bwh, hentry) | |
719 | if (worker->current_work == work) | |
720 | return worker; | |
721 | return NULL; | |
722 | } | |
723 | ||
724 | /** | |
725 | * find_worker_executing_work - find worker which is executing a work | |
726 | * @gcwq: gcwq of interest | |
727 | * @work: work to find worker for | |
728 | * | |
729 | * Find a worker which is executing @work on @gcwq. This function is | |
730 | * identical to __find_worker_executing_work() except that this | |
731 | * function calculates @bwh itself. | |
732 | * | |
733 | * CONTEXT: | |
734 | * spin_lock_irq(gcwq->lock). | |
735 | * | |
736 | * RETURNS: | |
737 | * Pointer to worker which is executing @work if found, NULL | |
738 | * otherwise. | |
739 | */ | |
740 | static struct worker *find_worker_executing_work(struct global_cwq *gcwq, | |
741 | struct work_struct *work) | |
742 | { | |
743 | return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), | |
744 | work); | |
745 | } | |
746 | ||
649027d7 TH |
747 | /** |
748 | * gcwq_determine_ins_pos - find insertion position | |
749 | * @gcwq: gcwq of interest | |
750 | * @cwq: cwq a work is being queued for | |
751 | * | |
752 | * A work for @cwq is about to be queued on @gcwq, determine insertion | |
753 | * position for the work. If @cwq is for HIGHPRI wq, the work is | |
754 | * queued at the head of the queue but in FIFO order with respect to | |
755 | * other HIGHPRI works; otherwise, at the end of the queue. This | |
756 | * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that | |
757 | * there are HIGHPRI works pending. | |
758 | * | |
759 | * CONTEXT: | |
760 | * spin_lock_irq(gcwq->lock). | |
761 | * | |
762 | * RETURNS: | |
763 | * Pointer to inserstion position. | |
764 | */ | |
765 | static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq, | |
766 | struct cpu_workqueue_struct *cwq) | |
767 | { | |
768 | struct work_struct *twork; | |
769 | ||
770 | if (likely(!(cwq->wq->flags & WQ_HIGHPRI))) | |
771 | return &gcwq->worklist; | |
772 | ||
773 | list_for_each_entry(twork, &gcwq->worklist, entry) { | |
774 | struct cpu_workqueue_struct *tcwq = get_work_cwq(twork); | |
775 | ||
776 | if (!(tcwq->wq->flags & WQ_HIGHPRI)) | |
777 | break; | |
778 | } | |
779 | ||
780 | gcwq->flags |= GCWQ_HIGHPRI_PENDING; | |
781 | return &twork->entry; | |
782 | } | |
783 | ||
4690c4ab | 784 | /** |
7e11629d | 785 | * insert_work - insert a work into gcwq |
4690c4ab TH |
786 | * @cwq: cwq @work belongs to |
787 | * @work: work to insert | |
788 | * @head: insertion point | |
789 | * @extra_flags: extra WORK_STRUCT_* flags to set | |
790 | * | |
7e11629d TH |
791 | * Insert @work which belongs to @cwq into @gcwq after @head. |
792 | * @extra_flags is or'd to work_struct flags. | |
4690c4ab TH |
793 | * |
794 | * CONTEXT: | |
8b03ae3c | 795 | * spin_lock_irq(gcwq->lock). |
4690c4ab | 796 | */ |
b89deed3 | 797 | static void insert_work(struct cpu_workqueue_struct *cwq, |
4690c4ab TH |
798 | struct work_struct *work, struct list_head *head, |
799 | unsigned int extra_flags) | |
b89deed3 | 800 | { |
e22bee78 TH |
801 | struct global_cwq *gcwq = cwq->gcwq; |
802 | ||
4690c4ab | 803 | /* we own @work, set data and link */ |
7a22ad75 | 804 | set_work_cwq(work, cwq, extra_flags); |
4690c4ab | 805 | |
6e84d644 ON |
806 | /* |
807 | * Ensure that we get the right work->data if we see the | |
808 | * result of list_add() below, see try_to_grab_pending(). | |
809 | */ | |
810 | smp_wmb(); | |
4690c4ab | 811 | |
1a4d9b0a | 812 | list_add_tail(&work->entry, head); |
e22bee78 TH |
813 | |
814 | /* | |
815 | * Ensure either worker_sched_deactivated() sees the above | |
816 | * list_add_tail() or we see zero nr_running to avoid workers | |
817 | * lying around lazily while there are works to be processed. | |
818 | */ | |
819 | smp_mb(); | |
820 | ||
649027d7 | 821 | if (__need_more_worker(gcwq)) |
e22bee78 | 822 | wake_up_worker(gcwq); |
b89deed3 ON |
823 | } |
824 | ||
502ca9d8 TH |
825 | /** |
826 | * cwq_unbind_single_cpu - unbind cwq from single cpu workqueue processing | |
827 | * @cwq: cwq to unbind | |
828 | * | |
829 | * Try to unbind @cwq from single cpu workqueue processing. If | |
830 | * @cwq->wq is frozen, unbind is delayed till the workqueue is thawed. | |
831 | * | |
832 | * CONTEXT: | |
833 | * spin_lock_irq(gcwq->lock). | |
834 | */ | |
835 | static void cwq_unbind_single_cpu(struct cpu_workqueue_struct *cwq) | |
836 | { | |
837 | struct workqueue_struct *wq = cwq->wq; | |
838 | struct global_cwq *gcwq = cwq->gcwq; | |
839 | ||
840 | BUG_ON(wq->single_cpu != gcwq->cpu); | |
841 | /* | |
842 | * Unbind from workqueue if @cwq is not frozen. If frozen, | |
843 | * thaw_workqueues() will either restart processing on this | |
844 | * cpu or unbind if empty. This keeps works queued while | |
845 | * frozen fully ordered and flushable. | |
846 | */ | |
847 | if (likely(!(gcwq->flags & GCWQ_FREEZING))) { | |
848 | smp_wmb(); /* paired with cmpxchg() in __queue_work() */ | |
849 | wq->single_cpu = NR_CPUS; | |
850 | } | |
851 | } | |
852 | ||
4690c4ab | 853 | static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, |
1da177e4 LT |
854 | struct work_struct *work) |
855 | { | |
502ca9d8 TH |
856 | struct global_cwq *gcwq; |
857 | struct cpu_workqueue_struct *cwq; | |
1e19ffc6 | 858 | struct list_head *worklist; |
1da177e4 | 859 | unsigned long flags; |
502ca9d8 | 860 | bool arbitrate; |
1da177e4 | 861 | |
dc186ad7 | 862 | debug_work_activate(work); |
1e19ffc6 | 863 | |
18aa9eff TH |
864 | /* |
865 | * Determine gcwq to use. SINGLE_CPU is inherently | |
866 | * NON_REENTRANT, so test it first. | |
867 | */ | |
502ca9d8 | 868 | if (!(wq->flags & WQ_SINGLE_CPU)) { |
18aa9eff TH |
869 | struct global_cwq *last_gcwq; |
870 | ||
871 | /* | |
872 | * It's multi cpu. If @wq is non-reentrant and @work | |
873 | * was previously on a different cpu, it might still | |
874 | * be running there, in which case the work needs to | |
875 | * be queued on that cpu to guarantee non-reentrance. | |
876 | */ | |
502ca9d8 | 877 | gcwq = get_gcwq(cpu); |
18aa9eff TH |
878 | if (wq->flags & WQ_NON_REENTRANT && |
879 | (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { | |
880 | struct worker *worker; | |
881 | ||
882 | spin_lock_irqsave(&last_gcwq->lock, flags); | |
883 | ||
884 | worker = find_worker_executing_work(last_gcwq, work); | |
885 | ||
886 | if (worker && worker->current_cwq->wq == wq) | |
887 | gcwq = last_gcwq; | |
888 | else { | |
889 | /* meh... not running there, queue here */ | |
890 | spin_unlock_irqrestore(&last_gcwq->lock, flags); | |
891 | spin_lock_irqsave(&gcwq->lock, flags); | |
892 | } | |
893 | } else | |
894 | spin_lock_irqsave(&gcwq->lock, flags); | |
502ca9d8 TH |
895 | } else { |
896 | unsigned int req_cpu = cpu; | |
897 | ||
898 | /* | |
899 | * It's a bit more complex for single cpu workqueues. | |
900 | * We first need to determine which cpu is going to be | |
901 | * used. If no cpu is currently serving this | |
902 | * workqueue, arbitrate using atomic accesses to | |
903 | * wq->single_cpu; otherwise, use the current one. | |
904 | */ | |
905 | retry: | |
906 | cpu = wq->single_cpu; | |
907 | arbitrate = cpu == NR_CPUS; | |
908 | if (arbitrate) | |
909 | cpu = req_cpu; | |
910 | ||
911 | gcwq = get_gcwq(cpu); | |
912 | spin_lock_irqsave(&gcwq->lock, flags); | |
913 | ||
914 | /* | |
915 | * The following cmpxchg() is a full barrier paired | |
916 | * with smp_wmb() in cwq_unbind_single_cpu() and | |
917 | * guarantees that all changes to wq->st_* fields are | |
918 | * visible on the new cpu after this point. | |
919 | */ | |
920 | if (arbitrate) | |
921 | cmpxchg(&wq->single_cpu, NR_CPUS, cpu); | |
922 | ||
923 | if (unlikely(wq->single_cpu != cpu)) { | |
924 | spin_unlock_irqrestore(&gcwq->lock, flags); | |
925 | goto retry; | |
926 | } | |
927 | } | |
928 | ||
929 | /* gcwq determined, get cwq and queue */ | |
930 | cwq = get_cwq(gcwq->cpu, wq); | |
931 | ||
4690c4ab | 932 | BUG_ON(!list_empty(&work->entry)); |
1e19ffc6 | 933 | |
73f53c4a | 934 | cwq->nr_in_flight[cwq->work_color]++; |
1e19ffc6 TH |
935 | |
936 | if (likely(cwq->nr_active < cwq->max_active)) { | |
937 | cwq->nr_active++; | |
649027d7 | 938 | worklist = gcwq_determine_ins_pos(gcwq, cwq); |
1e19ffc6 TH |
939 | } else |
940 | worklist = &cwq->delayed_works; | |
941 | ||
942 | insert_work(cwq, work, worklist, work_color_to_flags(cwq->work_color)); | |
943 | ||
8b03ae3c | 944 | spin_unlock_irqrestore(&gcwq->lock, flags); |
1da177e4 LT |
945 | } |
946 | ||
0fcb78c2 REB |
947 | /** |
948 | * queue_work - queue work on a workqueue | |
949 | * @wq: workqueue to use | |
950 | * @work: work to queue | |
951 | * | |
057647fc | 952 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
1da177e4 | 953 | * |
00dfcaf7 ON |
954 | * We queue the work to the CPU on which it was submitted, but if the CPU dies |
955 | * it can be processed by another CPU. | |
1da177e4 | 956 | */ |
7ad5b3a5 | 957 | int queue_work(struct workqueue_struct *wq, struct work_struct *work) |
1da177e4 | 958 | { |
ef1ca236 ON |
959 | int ret; |
960 | ||
961 | ret = queue_work_on(get_cpu(), wq, work); | |
962 | put_cpu(); | |
963 | ||
1da177e4 LT |
964 | return ret; |
965 | } | |
ae90dd5d | 966 | EXPORT_SYMBOL_GPL(queue_work); |
1da177e4 | 967 | |
c1a220e7 ZR |
968 | /** |
969 | * queue_work_on - queue work on specific cpu | |
970 | * @cpu: CPU number to execute work on | |
971 | * @wq: workqueue to use | |
972 | * @work: work to queue | |
973 | * | |
974 | * Returns 0 if @work was already on a queue, non-zero otherwise. | |
975 | * | |
976 | * We queue the work to a specific CPU, the caller must ensure it | |
977 | * can't go away. | |
978 | */ | |
979 | int | |
980 | queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) | |
981 | { | |
982 | int ret = 0; | |
983 | ||
22df02bb | 984 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
4690c4ab | 985 | __queue_work(cpu, wq, work); |
c1a220e7 ZR |
986 | ret = 1; |
987 | } | |
988 | return ret; | |
989 | } | |
990 | EXPORT_SYMBOL_GPL(queue_work_on); | |
991 | ||
6d141c3f | 992 | static void delayed_work_timer_fn(unsigned long __data) |
1da177e4 | 993 | { |
52bad64d | 994 | struct delayed_work *dwork = (struct delayed_work *)__data; |
7a22ad75 | 995 | struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); |
1da177e4 | 996 | |
4690c4ab | 997 | __queue_work(smp_processor_id(), cwq->wq, &dwork->work); |
1da177e4 LT |
998 | } |
999 | ||
0fcb78c2 REB |
1000 | /** |
1001 | * queue_delayed_work - queue work on a workqueue after delay | |
1002 | * @wq: workqueue to use | |
af9997e4 | 1003 | * @dwork: delayable work to queue |
0fcb78c2 REB |
1004 | * @delay: number of jiffies to wait before queueing |
1005 | * | |
057647fc | 1006 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
0fcb78c2 | 1007 | */ |
7ad5b3a5 | 1008 | int queue_delayed_work(struct workqueue_struct *wq, |
52bad64d | 1009 | struct delayed_work *dwork, unsigned long delay) |
1da177e4 | 1010 | { |
52bad64d | 1011 | if (delay == 0) |
63bc0362 | 1012 | return queue_work(wq, &dwork->work); |
1da177e4 | 1013 | |
63bc0362 | 1014 | return queue_delayed_work_on(-1, wq, dwork, delay); |
1da177e4 | 1015 | } |
ae90dd5d | 1016 | EXPORT_SYMBOL_GPL(queue_delayed_work); |
1da177e4 | 1017 | |
0fcb78c2 REB |
1018 | /** |
1019 | * queue_delayed_work_on - queue work on specific CPU after delay | |
1020 | * @cpu: CPU number to execute work on | |
1021 | * @wq: workqueue to use | |
af9997e4 | 1022 | * @dwork: work to queue |
0fcb78c2 REB |
1023 | * @delay: number of jiffies to wait before queueing |
1024 | * | |
057647fc | 1025 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
0fcb78c2 | 1026 | */ |
7a6bc1cd | 1027 | int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
52bad64d | 1028 | struct delayed_work *dwork, unsigned long delay) |
7a6bc1cd VP |
1029 | { |
1030 | int ret = 0; | |
52bad64d DH |
1031 | struct timer_list *timer = &dwork->timer; |
1032 | struct work_struct *work = &dwork->work; | |
7a6bc1cd | 1033 | |
22df02bb | 1034 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
7a22ad75 TH |
1035 | struct global_cwq *gcwq = get_work_gcwq(work); |
1036 | unsigned int lcpu = gcwq ? gcwq->cpu : raw_smp_processor_id(); | |
1037 | ||
7a6bc1cd VP |
1038 | BUG_ON(timer_pending(timer)); |
1039 | BUG_ON(!list_empty(&work->entry)); | |
1040 | ||
8a3e77cc | 1041 | timer_stats_timer_set_start_info(&dwork->timer); |
7a22ad75 TH |
1042 | /* |
1043 | * This stores cwq for the moment, for the timer_fn. | |
1044 | * Note that the work's gcwq is preserved to allow | |
1045 | * reentrance detection for delayed works. | |
1046 | */ | |
1047 | set_work_cwq(work, get_cwq(lcpu, wq), 0); | |
7a6bc1cd | 1048 | timer->expires = jiffies + delay; |
52bad64d | 1049 | timer->data = (unsigned long)dwork; |
7a6bc1cd | 1050 | timer->function = delayed_work_timer_fn; |
63bc0362 ON |
1051 | |
1052 | if (unlikely(cpu >= 0)) | |
1053 | add_timer_on(timer, cpu); | |
1054 | else | |
1055 | add_timer(timer); | |
7a6bc1cd VP |
1056 | ret = 1; |
1057 | } | |
1058 | return ret; | |
1059 | } | |
ae90dd5d | 1060 | EXPORT_SYMBOL_GPL(queue_delayed_work_on); |
1da177e4 | 1061 | |
c8e55f36 TH |
1062 | /** |
1063 | * worker_enter_idle - enter idle state | |
1064 | * @worker: worker which is entering idle state | |
1065 | * | |
1066 | * @worker is entering idle state. Update stats and idle timer if | |
1067 | * necessary. | |
1068 | * | |
1069 | * LOCKING: | |
1070 | * spin_lock_irq(gcwq->lock). | |
1071 | */ | |
1072 | static void worker_enter_idle(struct worker *worker) | |
1073 | { | |
1074 | struct global_cwq *gcwq = worker->gcwq; | |
1075 | ||
1076 | BUG_ON(worker->flags & WORKER_IDLE); | |
1077 | BUG_ON(!list_empty(&worker->entry) && | |
1078 | (worker->hentry.next || worker->hentry.pprev)); | |
1079 | ||
cb444766 TH |
1080 | /* can't use worker_set_flags(), also called from start_worker() */ |
1081 | worker->flags |= WORKER_IDLE; | |
c8e55f36 | 1082 | gcwq->nr_idle++; |
e22bee78 | 1083 | worker->last_active = jiffies; |
c8e55f36 TH |
1084 | |
1085 | /* idle_list is LIFO */ | |
1086 | list_add(&worker->entry, &gcwq->idle_list); | |
db7bccf4 | 1087 | |
e22bee78 TH |
1088 | if (likely(!(worker->flags & WORKER_ROGUE))) { |
1089 | if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer)) | |
1090 | mod_timer(&gcwq->idle_timer, | |
1091 | jiffies + IDLE_WORKER_TIMEOUT); | |
1092 | } else | |
db7bccf4 | 1093 | wake_up_all(&gcwq->trustee_wait); |
cb444766 TH |
1094 | |
1095 | /* sanity check nr_running */ | |
1096 | WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle && | |
1097 | atomic_read(get_gcwq_nr_running(gcwq->cpu))); | |
c8e55f36 TH |
1098 | } |
1099 | ||
1100 | /** | |
1101 | * worker_leave_idle - leave idle state | |
1102 | * @worker: worker which is leaving idle state | |
1103 | * | |
1104 | * @worker is leaving idle state. Update stats. | |
1105 | * | |
1106 | * LOCKING: | |
1107 | * spin_lock_irq(gcwq->lock). | |
1108 | */ | |
1109 | static void worker_leave_idle(struct worker *worker) | |
1110 | { | |
1111 | struct global_cwq *gcwq = worker->gcwq; | |
1112 | ||
1113 | BUG_ON(!(worker->flags & WORKER_IDLE)); | |
d302f017 | 1114 | worker_clr_flags(worker, WORKER_IDLE); |
c8e55f36 TH |
1115 | gcwq->nr_idle--; |
1116 | list_del_init(&worker->entry); | |
1117 | } | |
1118 | ||
e22bee78 TH |
1119 | /** |
1120 | * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq | |
1121 | * @worker: self | |
1122 | * | |
1123 | * Works which are scheduled while the cpu is online must at least be | |
1124 | * scheduled to a worker which is bound to the cpu so that if they are | |
1125 | * flushed from cpu callbacks while cpu is going down, they are | |
1126 | * guaranteed to execute on the cpu. | |
1127 | * | |
1128 | * This function is to be used by rogue workers and rescuers to bind | |
1129 | * themselves to the target cpu and may race with cpu going down or | |
1130 | * coming online. kthread_bind() can't be used because it may put the | |
1131 | * worker to already dead cpu and set_cpus_allowed_ptr() can't be used | |
1132 | * verbatim as it's best effort and blocking and gcwq may be | |
1133 | * [dis]associated in the meantime. | |
1134 | * | |
1135 | * This function tries set_cpus_allowed() and locks gcwq and verifies | |
1136 | * the binding against GCWQ_DISASSOCIATED which is set during | |
1137 | * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters | |
1138 | * idle state or fetches works without dropping lock, it can guarantee | |
1139 | * the scheduling requirement described in the first paragraph. | |
1140 | * | |
1141 | * CONTEXT: | |
1142 | * Might sleep. Called without any lock but returns with gcwq->lock | |
1143 | * held. | |
1144 | * | |
1145 | * RETURNS: | |
1146 | * %true if the associated gcwq is online (@worker is successfully | |
1147 | * bound), %false if offline. | |
1148 | */ | |
1149 | static bool worker_maybe_bind_and_lock(struct worker *worker) | |
1150 | { | |
1151 | struct global_cwq *gcwq = worker->gcwq; | |
1152 | struct task_struct *task = worker->task; | |
1153 | ||
1154 | while (true) { | |
1155 | /* | |
1156 | * The following call may fail, succeed or succeed | |
1157 | * without actually migrating the task to the cpu if | |
1158 | * it races with cpu hotunplug operation. Verify | |
1159 | * against GCWQ_DISASSOCIATED. | |
1160 | */ | |
1161 | set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); | |
1162 | ||
1163 | spin_lock_irq(&gcwq->lock); | |
1164 | if (gcwq->flags & GCWQ_DISASSOCIATED) | |
1165 | return false; | |
1166 | if (task_cpu(task) == gcwq->cpu && | |
1167 | cpumask_equal(¤t->cpus_allowed, | |
1168 | get_cpu_mask(gcwq->cpu))) | |
1169 | return true; | |
1170 | spin_unlock_irq(&gcwq->lock); | |
1171 | ||
1172 | /* CPU has come up inbetween, retry migration */ | |
1173 | cpu_relax(); | |
1174 | } | |
1175 | } | |
1176 | ||
1177 | /* | |
1178 | * Function for worker->rebind_work used to rebind rogue busy workers | |
1179 | * to the associated cpu which is coming back online. This is | |
1180 | * scheduled by cpu up but can race with other cpu hotplug operations | |
1181 | * and may be executed twice without intervening cpu down. | |
1182 | */ | |
1183 | static void worker_rebind_fn(struct work_struct *work) | |
1184 | { | |
1185 | struct worker *worker = container_of(work, struct worker, rebind_work); | |
1186 | struct global_cwq *gcwq = worker->gcwq; | |
1187 | ||
1188 | if (worker_maybe_bind_and_lock(worker)) | |
1189 | worker_clr_flags(worker, WORKER_REBIND); | |
1190 | ||
1191 | spin_unlock_irq(&gcwq->lock); | |
1192 | } | |
1193 | ||
c34056a3 TH |
1194 | static struct worker *alloc_worker(void) |
1195 | { | |
1196 | struct worker *worker; | |
1197 | ||
1198 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); | |
c8e55f36 TH |
1199 | if (worker) { |
1200 | INIT_LIST_HEAD(&worker->entry); | |
affee4b2 | 1201 | INIT_LIST_HEAD(&worker->scheduled); |
e22bee78 TH |
1202 | INIT_WORK(&worker->rebind_work, worker_rebind_fn); |
1203 | /* on creation a worker is in !idle && prep state */ | |
1204 | worker->flags = WORKER_PREP; | |
c8e55f36 | 1205 | } |
c34056a3 TH |
1206 | return worker; |
1207 | } | |
1208 | ||
1209 | /** | |
1210 | * create_worker - create a new workqueue worker | |
7e11629d | 1211 | * @gcwq: gcwq the new worker will belong to |
c34056a3 TH |
1212 | * @bind: whether to set affinity to @cpu or not |
1213 | * | |
7e11629d | 1214 | * Create a new worker which is bound to @gcwq. The returned worker |
c34056a3 TH |
1215 | * can be started by calling start_worker() or destroyed using |
1216 | * destroy_worker(). | |
1217 | * | |
1218 | * CONTEXT: | |
1219 | * Might sleep. Does GFP_KERNEL allocations. | |
1220 | * | |
1221 | * RETURNS: | |
1222 | * Pointer to the newly created worker. | |
1223 | */ | |
7e11629d | 1224 | static struct worker *create_worker(struct global_cwq *gcwq, bool bind) |
c34056a3 TH |
1225 | { |
1226 | int id = -1; | |
1227 | struct worker *worker = NULL; | |
1228 | ||
8b03ae3c TH |
1229 | spin_lock_irq(&gcwq->lock); |
1230 | while (ida_get_new(&gcwq->worker_ida, &id)) { | |
1231 | spin_unlock_irq(&gcwq->lock); | |
1232 | if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL)) | |
c34056a3 | 1233 | goto fail; |
8b03ae3c | 1234 | spin_lock_irq(&gcwq->lock); |
c34056a3 | 1235 | } |
8b03ae3c | 1236 | spin_unlock_irq(&gcwq->lock); |
c34056a3 TH |
1237 | |
1238 | worker = alloc_worker(); | |
1239 | if (!worker) | |
1240 | goto fail; | |
1241 | ||
8b03ae3c | 1242 | worker->gcwq = gcwq; |
c34056a3 TH |
1243 | worker->id = id; |
1244 | ||
1245 | worker->task = kthread_create(worker_thread, worker, "kworker/%u:%d", | |
8b03ae3c | 1246 | gcwq->cpu, id); |
c34056a3 TH |
1247 | if (IS_ERR(worker->task)) |
1248 | goto fail; | |
1249 | ||
db7bccf4 TH |
1250 | /* |
1251 | * A rogue worker will become a regular one if CPU comes | |
1252 | * online later on. Make sure every worker has | |
1253 | * PF_THREAD_BOUND set. | |
1254 | */ | |
c34056a3 | 1255 | if (bind) |
8b03ae3c | 1256 | kthread_bind(worker->task, gcwq->cpu); |
db7bccf4 TH |
1257 | else |
1258 | worker->task->flags |= PF_THREAD_BOUND; | |
c34056a3 TH |
1259 | |
1260 | return worker; | |
1261 | fail: | |
1262 | if (id >= 0) { | |
8b03ae3c TH |
1263 | spin_lock_irq(&gcwq->lock); |
1264 | ida_remove(&gcwq->worker_ida, id); | |
1265 | spin_unlock_irq(&gcwq->lock); | |
c34056a3 TH |
1266 | } |
1267 | kfree(worker); | |
1268 | return NULL; | |
1269 | } | |
1270 | ||
1271 | /** | |
1272 | * start_worker - start a newly created worker | |
1273 | * @worker: worker to start | |
1274 | * | |
c8e55f36 | 1275 | * Make the gcwq aware of @worker and start it. |
c34056a3 TH |
1276 | * |
1277 | * CONTEXT: | |
8b03ae3c | 1278 | * spin_lock_irq(gcwq->lock). |
c34056a3 TH |
1279 | */ |
1280 | static void start_worker(struct worker *worker) | |
1281 | { | |
cb444766 | 1282 | worker->flags |= WORKER_STARTED; |
c8e55f36 TH |
1283 | worker->gcwq->nr_workers++; |
1284 | worker_enter_idle(worker); | |
c34056a3 TH |
1285 | wake_up_process(worker->task); |
1286 | } | |
1287 | ||
1288 | /** | |
1289 | * destroy_worker - destroy a workqueue worker | |
1290 | * @worker: worker to be destroyed | |
1291 | * | |
c8e55f36 TH |
1292 | * Destroy @worker and adjust @gcwq stats accordingly. |
1293 | * | |
1294 | * CONTEXT: | |
1295 | * spin_lock_irq(gcwq->lock) which is released and regrabbed. | |
c34056a3 TH |
1296 | */ |
1297 | static void destroy_worker(struct worker *worker) | |
1298 | { | |
8b03ae3c | 1299 | struct global_cwq *gcwq = worker->gcwq; |
c34056a3 TH |
1300 | int id = worker->id; |
1301 | ||
1302 | /* sanity check frenzy */ | |
1303 | BUG_ON(worker->current_work); | |
affee4b2 | 1304 | BUG_ON(!list_empty(&worker->scheduled)); |
c34056a3 | 1305 | |
c8e55f36 TH |
1306 | if (worker->flags & WORKER_STARTED) |
1307 | gcwq->nr_workers--; | |
1308 | if (worker->flags & WORKER_IDLE) | |
1309 | gcwq->nr_idle--; | |
1310 | ||
1311 | list_del_init(&worker->entry); | |
cb444766 | 1312 | worker->flags |= WORKER_DIE; |
c8e55f36 TH |
1313 | |
1314 | spin_unlock_irq(&gcwq->lock); | |
1315 | ||
c34056a3 TH |
1316 | kthread_stop(worker->task); |
1317 | kfree(worker); | |
1318 | ||
8b03ae3c TH |
1319 | spin_lock_irq(&gcwq->lock); |
1320 | ida_remove(&gcwq->worker_ida, id); | |
c34056a3 TH |
1321 | } |
1322 | ||
e22bee78 TH |
1323 | static void idle_worker_timeout(unsigned long __gcwq) |
1324 | { | |
1325 | struct global_cwq *gcwq = (void *)__gcwq; | |
1326 | ||
1327 | spin_lock_irq(&gcwq->lock); | |
1328 | ||
1329 | if (too_many_workers(gcwq)) { | |
1330 | struct worker *worker; | |
1331 | unsigned long expires; | |
1332 | ||
1333 | /* idle_list is kept in LIFO order, check the last one */ | |
1334 | worker = list_entry(gcwq->idle_list.prev, struct worker, entry); | |
1335 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
1336 | ||
1337 | if (time_before(jiffies, expires)) | |
1338 | mod_timer(&gcwq->idle_timer, expires); | |
1339 | else { | |
1340 | /* it's been idle for too long, wake up manager */ | |
1341 | gcwq->flags |= GCWQ_MANAGE_WORKERS; | |
1342 | wake_up_worker(gcwq); | |
1343 | } | |
1344 | } | |
1345 | ||
1346 | spin_unlock_irq(&gcwq->lock); | |
1347 | } | |
1348 | ||
1349 | static bool send_mayday(struct work_struct *work) | |
1350 | { | |
1351 | struct cpu_workqueue_struct *cwq = get_work_cwq(work); | |
1352 | struct workqueue_struct *wq = cwq->wq; | |
1353 | ||
1354 | if (!(wq->flags & WQ_RESCUER)) | |
1355 | return false; | |
1356 | ||
1357 | /* mayday mayday mayday */ | |
1358 | if (!cpumask_test_and_set_cpu(cwq->gcwq->cpu, wq->mayday_mask)) | |
1359 | wake_up_process(wq->rescuer->task); | |
1360 | return true; | |
1361 | } | |
1362 | ||
1363 | static void gcwq_mayday_timeout(unsigned long __gcwq) | |
1364 | { | |
1365 | struct global_cwq *gcwq = (void *)__gcwq; | |
1366 | struct work_struct *work; | |
1367 | ||
1368 | spin_lock_irq(&gcwq->lock); | |
1369 | ||
1370 | if (need_to_create_worker(gcwq)) { | |
1371 | /* | |
1372 | * We've been trying to create a new worker but | |
1373 | * haven't been successful. We might be hitting an | |
1374 | * allocation deadlock. Send distress signals to | |
1375 | * rescuers. | |
1376 | */ | |
1377 | list_for_each_entry(work, &gcwq->worklist, entry) | |
1378 | send_mayday(work); | |
1379 | } | |
1380 | ||
1381 | spin_unlock_irq(&gcwq->lock); | |
1382 | ||
1383 | mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL); | |
1384 | } | |
1385 | ||
1386 | /** | |
1387 | * maybe_create_worker - create a new worker if necessary | |
1388 | * @gcwq: gcwq to create a new worker for | |
1389 | * | |
1390 | * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to | |
1391 | * have at least one idle worker on return from this function. If | |
1392 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is | |
1393 | * sent to all rescuers with works scheduled on @gcwq to resolve | |
1394 | * possible allocation deadlock. | |
1395 | * | |
1396 | * On return, need_to_create_worker() is guaranteed to be false and | |
1397 | * may_start_working() true. | |
1398 | * | |
1399 | * LOCKING: | |
1400 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
1401 | * multiple times. Does GFP_KERNEL allocations. Called only from | |
1402 | * manager. | |
1403 | * | |
1404 | * RETURNS: | |
1405 | * false if no action was taken and gcwq->lock stayed locked, true | |
1406 | * otherwise. | |
1407 | */ | |
1408 | static bool maybe_create_worker(struct global_cwq *gcwq) | |
1409 | { | |
1410 | if (!need_to_create_worker(gcwq)) | |
1411 | return false; | |
1412 | restart: | |
1413 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ | |
1414 | mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); | |
1415 | ||
1416 | while (true) { | |
1417 | struct worker *worker; | |
1418 | ||
1419 | spin_unlock_irq(&gcwq->lock); | |
1420 | ||
1421 | worker = create_worker(gcwq, true); | |
1422 | if (worker) { | |
1423 | del_timer_sync(&gcwq->mayday_timer); | |
1424 | spin_lock_irq(&gcwq->lock); | |
1425 | start_worker(worker); | |
1426 | BUG_ON(need_to_create_worker(gcwq)); | |
1427 | return true; | |
1428 | } | |
1429 | ||
1430 | if (!need_to_create_worker(gcwq)) | |
1431 | break; | |
1432 | ||
1433 | spin_unlock_irq(&gcwq->lock); | |
1434 | __set_current_state(TASK_INTERRUPTIBLE); | |
1435 | schedule_timeout(CREATE_COOLDOWN); | |
1436 | spin_lock_irq(&gcwq->lock); | |
1437 | if (!need_to_create_worker(gcwq)) | |
1438 | break; | |
1439 | } | |
1440 | ||
1441 | spin_unlock_irq(&gcwq->lock); | |
1442 | del_timer_sync(&gcwq->mayday_timer); | |
1443 | spin_lock_irq(&gcwq->lock); | |
1444 | if (need_to_create_worker(gcwq)) | |
1445 | goto restart; | |
1446 | return true; | |
1447 | } | |
1448 | ||
1449 | /** | |
1450 | * maybe_destroy_worker - destroy workers which have been idle for a while | |
1451 | * @gcwq: gcwq to destroy workers for | |
1452 | * | |
1453 | * Destroy @gcwq workers which have been idle for longer than | |
1454 | * IDLE_WORKER_TIMEOUT. | |
1455 | * | |
1456 | * LOCKING: | |
1457 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
1458 | * multiple times. Called only from manager. | |
1459 | * | |
1460 | * RETURNS: | |
1461 | * false if no action was taken and gcwq->lock stayed locked, true | |
1462 | * otherwise. | |
1463 | */ | |
1464 | static bool maybe_destroy_workers(struct global_cwq *gcwq) | |
1465 | { | |
1466 | bool ret = false; | |
1467 | ||
1468 | while (too_many_workers(gcwq)) { | |
1469 | struct worker *worker; | |
1470 | unsigned long expires; | |
1471 | ||
1472 | worker = list_entry(gcwq->idle_list.prev, struct worker, entry); | |
1473 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; | |
1474 | ||
1475 | if (time_before(jiffies, expires)) { | |
1476 | mod_timer(&gcwq->idle_timer, expires); | |
1477 | break; | |
1478 | } | |
1479 | ||
1480 | destroy_worker(worker); | |
1481 | ret = true; | |
1482 | } | |
1483 | ||
1484 | return ret; | |
1485 | } | |
1486 | ||
1487 | /** | |
1488 | * manage_workers - manage worker pool | |
1489 | * @worker: self | |
1490 | * | |
1491 | * Assume the manager role and manage gcwq worker pool @worker belongs | |
1492 | * to. At any given time, there can be only zero or one manager per | |
1493 | * gcwq. The exclusion is handled automatically by this function. | |
1494 | * | |
1495 | * The caller can safely start processing works on false return. On | |
1496 | * true return, it's guaranteed that need_to_create_worker() is false | |
1497 | * and may_start_working() is true. | |
1498 | * | |
1499 | * CONTEXT: | |
1500 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
1501 | * multiple times. Does GFP_KERNEL allocations. | |
1502 | * | |
1503 | * RETURNS: | |
1504 | * false if no action was taken and gcwq->lock stayed locked, true if | |
1505 | * some action was taken. | |
1506 | */ | |
1507 | static bool manage_workers(struct worker *worker) | |
1508 | { | |
1509 | struct global_cwq *gcwq = worker->gcwq; | |
1510 | bool ret = false; | |
1511 | ||
1512 | if (gcwq->flags & GCWQ_MANAGING_WORKERS) | |
1513 | return ret; | |
1514 | ||
1515 | gcwq->flags &= ~GCWQ_MANAGE_WORKERS; | |
1516 | gcwq->flags |= GCWQ_MANAGING_WORKERS; | |
1517 | ||
1518 | /* | |
1519 | * Destroy and then create so that may_start_working() is true | |
1520 | * on return. | |
1521 | */ | |
1522 | ret |= maybe_destroy_workers(gcwq); | |
1523 | ret |= maybe_create_worker(gcwq); | |
1524 | ||
1525 | gcwq->flags &= ~GCWQ_MANAGING_WORKERS; | |
1526 | ||
1527 | /* | |
1528 | * The trustee might be waiting to take over the manager | |
1529 | * position, tell it we're done. | |
1530 | */ | |
1531 | if (unlikely(gcwq->trustee)) | |
1532 | wake_up_all(&gcwq->trustee_wait); | |
1533 | ||
1534 | return ret; | |
1535 | } | |
1536 | ||
affee4b2 TH |
1537 | /** |
1538 | * move_linked_works - move linked works to a list | |
1539 | * @work: start of series of works to be scheduled | |
1540 | * @head: target list to append @work to | |
1541 | * @nextp: out paramter for nested worklist walking | |
1542 | * | |
1543 | * Schedule linked works starting from @work to @head. Work series to | |
1544 | * be scheduled starts at @work and includes any consecutive work with | |
1545 | * WORK_STRUCT_LINKED set in its predecessor. | |
1546 | * | |
1547 | * If @nextp is not NULL, it's updated to point to the next work of | |
1548 | * the last scheduled work. This allows move_linked_works() to be | |
1549 | * nested inside outer list_for_each_entry_safe(). | |
1550 | * | |
1551 | * CONTEXT: | |
8b03ae3c | 1552 | * spin_lock_irq(gcwq->lock). |
affee4b2 TH |
1553 | */ |
1554 | static void move_linked_works(struct work_struct *work, struct list_head *head, | |
1555 | struct work_struct **nextp) | |
1556 | { | |
1557 | struct work_struct *n; | |
1558 | ||
1559 | /* | |
1560 | * Linked worklist will always end before the end of the list, | |
1561 | * use NULL for list head. | |
1562 | */ | |
1563 | list_for_each_entry_safe_from(work, n, NULL, entry) { | |
1564 | list_move_tail(&work->entry, head); | |
1565 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) | |
1566 | break; | |
1567 | } | |
1568 | ||
1569 | /* | |
1570 | * If we're already inside safe list traversal and have moved | |
1571 | * multiple works to the scheduled queue, the next position | |
1572 | * needs to be updated. | |
1573 | */ | |
1574 | if (nextp) | |
1575 | *nextp = n; | |
1576 | } | |
1577 | ||
1e19ffc6 TH |
1578 | static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) |
1579 | { | |
1580 | struct work_struct *work = list_first_entry(&cwq->delayed_works, | |
1581 | struct work_struct, entry); | |
649027d7 | 1582 | struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq); |
1e19ffc6 | 1583 | |
649027d7 | 1584 | move_linked_works(work, pos, NULL); |
1e19ffc6 TH |
1585 | cwq->nr_active++; |
1586 | } | |
1587 | ||
73f53c4a TH |
1588 | /** |
1589 | * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight | |
1590 | * @cwq: cwq of interest | |
1591 | * @color: color of work which left the queue | |
1592 | * | |
1593 | * A work either has completed or is removed from pending queue, | |
1594 | * decrement nr_in_flight of its cwq and handle workqueue flushing. | |
1595 | * | |
1596 | * CONTEXT: | |
8b03ae3c | 1597 | * spin_lock_irq(gcwq->lock). |
73f53c4a TH |
1598 | */ |
1599 | static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color) | |
1600 | { | |
1601 | /* ignore uncolored works */ | |
1602 | if (color == WORK_NO_COLOR) | |
1603 | return; | |
1604 | ||
1605 | cwq->nr_in_flight[color]--; | |
1e19ffc6 TH |
1606 | cwq->nr_active--; |
1607 | ||
502ca9d8 TH |
1608 | if (!list_empty(&cwq->delayed_works)) { |
1609 | /* one down, submit a delayed one */ | |
1610 | if (cwq->nr_active < cwq->max_active) | |
1611 | cwq_activate_first_delayed(cwq); | |
1612 | } else if (!cwq->nr_active && cwq->wq->flags & WQ_SINGLE_CPU) { | |
1613 | /* this was the last work, unbind from single cpu */ | |
1614 | cwq_unbind_single_cpu(cwq); | |
1615 | } | |
73f53c4a TH |
1616 | |
1617 | /* is flush in progress and are we at the flushing tip? */ | |
1618 | if (likely(cwq->flush_color != color)) | |
1619 | return; | |
1620 | ||
1621 | /* are there still in-flight works? */ | |
1622 | if (cwq->nr_in_flight[color]) | |
1623 | return; | |
1624 | ||
1625 | /* this cwq is done, clear flush_color */ | |
1626 | cwq->flush_color = -1; | |
1627 | ||
1628 | /* | |
1629 | * If this was the last cwq, wake up the first flusher. It | |
1630 | * will handle the rest. | |
1631 | */ | |
1632 | if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) | |
1633 | complete(&cwq->wq->first_flusher->done); | |
1634 | } | |
1635 | ||
a62428c0 TH |
1636 | /** |
1637 | * process_one_work - process single work | |
c34056a3 | 1638 | * @worker: self |
a62428c0 TH |
1639 | * @work: work to process |
1640 | * | |
1641 | * Process @work. This function contains all the logics necessary to | |
1642 | * process a single work including synchronization against and | |
1643 | * interaction with other workers on the same cpu, queueing and | |
1644 | * flushing. As long as context requirement is met, any worker can | |
1645 | * call this function to process a work. | |
1646 | * | |
1647 | * CONTEXT: | |
8b03ae3c | 1648 | * spin_lock_irq(gcwq->lock) which is released and regrabbed. |
a62428c0 | 1649 | */ |
c34056a3 | 1650 | static void process_one_work(struct worker *worker, struct work_struct *work) |
a62428c0 | 1651 | { |
7e11629d | 1652 | struct cpu_workqueue_struct *cwq = get_work_cwq(work); |
8b03ae3c | 1653 | struct global_cwq *gcwq = cwq->gcwq; |
c8e55f36 | 1654 | struct hlist_head *bwh = busy_worker_head(gcwq, work); |
fb0e7beb | 1655 | bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; |
a62428c0 | 1656 | work_func_t f = work->func; |
73f53c4a | 1657 | int work_color; |
7e11629d | 1658 | struct worker *collision; |
a62428c0 TH |
1659 | #ifdef CONFIG_LOCKDEP |
1660 | /* | |
1661 | * It is permissible to free the struct work_struct from | |
1662 | * inside the function that is called from it, this we need to | |
1663 | * take into account for lockdep too. To avoid bogus "held | |
1664 | * lock freed" warnings as well as problems when looking into | |
1665 | * work->lockdep_map, make a copy and use that here. | |
1666 | */ | |
1667 | struct lockdep_map lockdep_map = work->lockdep_map; | |
1668 | #endif | |
7e11629d TH |
1669 | /* |
1670 | * A single work shouldn't be executed concurrently by | |
1671 | * multiple workers on a single cpu. Check whether anyone is | |
1672 | * already processing the work. If so, defer the work to the | |
1673 | * currently executing one. | |
1674 | */ | |
1675 | collision = __find_worker_executing_work(gcwq, bwh, work); | |
1676 | if (unlikely(collision)) { | |
1677 | move_linked_works(work, &collision->scheduled, NULL); | |
1678 | return; | |
1679 | } | |
1680 | ||
a62428c0 | 1681 | /* claim and process */ |
a62428c0 | 1682 | debug_work_deactivate(work); |
c8e55f36 | 1683 | hlist_add_head(&worker->hentry, bwh); |
c34056a3 | 1684 | worker->current_work = work; |
8cca0eea | 1685 | worker->current_cwq = cwq; |
73f53c4a | 1686 | work_color = get_work_color(work); |
7a22ad75 | 1687 | |
7a22ad75 TH |
1688 | /* record the current cpu number in the work data and dequeue */ |
1689 | set_work_cpu(work, gcwq->cpu); | |
a62428c0 TH |
1690 | list_del_init(&work->entry); |
1691 | ||
649027d7 TH |
1692 | /* |
1693 | * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI, | |
1694 | * wake up another worker; otherwise, clear HIGHPRI_PENDING. | |
1695 | */ | |
1696 | if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) { | |
1697 | struct work_struct *nwork = list_first_entry(&gcwq->worklist, | |
1698 | struct work_struct, entry); | |
1699 | ||
1700 | if (!list_empty(&gcwq->worklist) && | |
1701 | get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI) | |
1702 | wake_up_worker(gcwq); | |
1703 | else | |
1704 | gcwq->flags &= ~GCWQ_HIGHPRI_PENDING; | |
1705 | } | |
1706 | ||
fb0e7beb TH |
1707 | /* |
1708 | * CPU intensive works don't participate in concurrency | |
1709 | * management. They're the scheduler's responsibility. | |
1710 | */ | |
1711 | if (unlikely(cpu_intensive)) | |
1712 | worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); | |
1713 | ||
8b03ae3c | 1714 | spin_unlock_irq(&gcwq->lock); |
a62428c0 | 1715 | |
a62428c0 TH |
1716 | work_clear_pending(work); |
1717 | lock_map_acquire(&cwq->wq->lockdep_map); | |
1718 | lock_map_acquire(&lockdep_map); | |
1719 | f(work); | |
1720 | lock_map_release(&lockdep_map); | |
1721 | lock_map_release(&cwq->wq->lockdep_map); | |
1722 | ||
1723 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { | |
1724 | printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " | |
1725 | "%s/0x%08x/%d\n", | |
1726 | current->comm, preempt_count(), task_pid_nr(current)); | |
1727 | printk(KERN_ERR " last function: "); | |
1728 | print_symbol("%s\n", (unsigned long)f); | |
1729 | debug_show_held_locks(current); | |
1730 | dump_stack(); | |
1731 | } | |
1732 | ||
8b03ae3c | 1733 | spin_lock_irq(&gcwq->lock); |
a62428c0 | 1734 | |
fb0e7beb TH |
1735 | /* clear cpu intensive status */ |
1736 | if (unlikely(cpu_intensive)) | |
1737 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); | |
1738 | ||
a62428c0 | 1739 | /* we're done with it, release */ |
c8e55f36 | 1740 | hlist_del_init(&worker->hentry); |
c34056a3 | 1741 | worker->current_work = NULL; |
8cca0eea | 1742 | worker->current_cwq = NULL; |
73f53c4a | 1743 | cwq_dec_nr_in_flight(cwq, work_color); |
a62428c0 TH |
1744 | } |
1745 | ||
affee4b2 TH |
1746 | /** |
1747 | * process_scheduled_works - process scheduled works | |
1748 | * @worker: self | |
1749 | * | |
1750 | * Process all scheduled works. Please note that the scheduled list | |
1751 | * may change while processing a work, so this function repeatedly | |
1752 | * fetches a work from the top and executes it. | |
1753 | * | |
1754 | * CONTEXT: | |
8b03ae3c | 1755 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
affee4b2 TH |
1756 | * multiple times. |
1757 | */ | |
1758 | static void process_scheduled_works(struct worker *worker) | |
1da177e4 | 1759 | { |
affee4b2 TH |
1760 | while (!list_empty(&worker->scheduled)) { |
1761 | struct work_struct *work = list_first_entry(&worker->scheduled, | |
1da177e4 | 1762 | struct work_struct, entry); |
c34056a3 | 1763 | process_one_work(worker, work); |
1da177e4 | 1764 | } |
1da177e4 LT |
1765 | } |
1766 | ||
4690c4ab TH |
1767 | /** |
1768 | * worker_thread - the worker thread function | |
c34056a3 | 1769 | * @__worker: self |
4690c4ab | 1770 | * |
e22bee78 TH |
1771 | * The gcwq worker thread function. There's a single dynamic pool of |
1772 | * these per each cpu. These workers process all works regardless of | |
1773 | * their specific target workqueue. The only exception is works which | |
1774 | * belong to workqueues with a rescuer which will be explained in | |
1775 | * rescuer_thread(). | |
4690c4ab | 1776 | */ |
c34056a3 | 1777 | static int worker_thread(void *__worker) |
1da177e4 | 1778 | { |
c34056a3 | 1779 | struct worker *worker = __worker; |
8b03ae3c | 1780 | struct global_cwq *gcwq = worker->gcwq; |
1da177e4 | 1781 | |
e22bee78 TH |
1782 | /* tell the scheduler that this is a workqueue worker */ |
1783 | worker->task->flags |= PF_WQ_WORKER; | |
c8e55f36 | 1784 | woke_up: |
c8e55f36 | 1785 | spin_lock_irq(&gcwq->lock); |
1da177e4 | 1786 | |
c8e55f36 TH |
1787 | /* DIE can be set only while we're idle, checking here is enough */ |
1788 | if (worker->flags & WORKER_DIE) { | |
1789 | spin_unlock_irq(&gcwq->lock); | |
e22bee78 | 1790 | worker->task->flags &= ~PF_WQ_WORKER; |
c8e55f36 TH |
1791 | return 0; |
1792 | } | |
affee4b2 | 1793 | |
c8e55f36 | 1794 | worker_leave_idle(worker); |
db7bccf4 | 1795 | recheck: |
e22bee78 TH |
1796 | /* no more worker necessary? */ |
1797 | if (!need_more_worker(gcwq)) | |
1798 | goto sleep; | |
1799 | ||
1800 | /* do we need to manage? */ | |
1801 | if (unlikely(!may_start_working(gcwq)) && manage_workers(worker)) | |
1802 | goto recheck; | |
1803 | ||
c8e55f36 TH |
1804 | /* |
1805 | * ->scheduled list can only be filled while a worker is | |
1806 | * preparing to process a work or actually processing it. | |
1807 | * Make sure nobody diddled with it while I was sleeping. | |
1808 | */ | |
1809 | BUG_ON(!list_empty(&worker->scheduled)); | |
1810 | ||
e22bee78 TH |
1811 | /* |
1812 | * When control reaches this point, we're guaranteed to have | |
1813 | * at least one idle worker or that someone else has already | |
1814 | * assumed the manager role. | |
1815 | */ | |
1816 | worker_clr_flags(worker, WORKER_PREP); | |
1817 | ||
1818 | do { | |
c8e55f36 | 1819 | struct work_struct *work = |
7e11629d | 1820 | list_first_entry(&gcwq->worklist, |
c8e55f36 TH |
1821 | struct work_struct, entry); |
1822 | ||
1823 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { | |
1824 | /* optimization path, not strictly necessary */ | |
1825 | process_one_work(worker, work); | |
1826 | if (unlikely(!list_empty(&worker->scheduled))) | |
affee4b2 | 1827 | process_scheduled_works(worker); |
c8e55f36 TH |
1828 | } else { |
1829 | move_linked_works(work, &worker->scheduled, NULL); | |
1830 | process_scheduled_works(worker); | |
affee4b2 | 1831 | } |
e22bee78 TH |
1832 | } while (keep_working(gcwq)); |
1833 | ||
1834 | worker_set_flags(worker, WORKER_PREP, false); | |
3af24433 | 1835 | |
e22bee78 TH |
1836 | if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker)) |
1837 | goto recheck; | |
1838 | sleep: | |
c8e55f36 | 1839 | /* |
e22bee78 TH |
1840 | * gcwq->lock is held and there's no work to process and no |
1841 | * need to manage, sleep. Workers are woken up only while | |
1842 | * holding gcwq->lock or from local cpu, so setting the | |
1843 | * current state before releasing gcwq->lock is enough to | |
1844 | * prevent losing any event. | |
c8e55f36 TH |
1845 | */ |
1846 | worker_enter_idle(worker); | |
1847 | __set_current_state(TASK_INTERRUPTIBLE); | |
1848 | spin_unlock_irq(&gcwq->lock); | |
1849 | schedule(); | |
1850 | goto woke_up; | |
1da177e4 LT |
1851 | } |
1852 | ||
e22bee78 TH |
1853 | /** |
1854 | * rescuer_thread - the rescuer thread function | |
1855 | * @__wq: the associated workqueue | |
1856 | * | |
1857 | * Workqueue rescuer thread function. There's one rescuer for each | |
1858 | * workqueue which has WQ_RESCUER set. | |
1859 | * | |
1860 | * Regular work processing on a gcwq may block trying to create a new | |
1861 | * worker which uses GFP_KERNEL allocation which has slight chance of | |
1862 | * developing into deadlock if some works currently on the same queue | |
1863 | * need to be processed to satisfy the GFP_KERNEL allocation. This is | |
1864 | * the problem rescuer solves. | |
1865 | * | |
1866 | * When such condition is possible, the gcwq summons rescuers of all | |
1867 | * workqueues which have works queued on the gcwq and let them process | |
1868 | * those works so that forward progress can be guaranteed. | |
1869 | * | |
1870 | * This should happen rarely. | |
1871 | */ | |
1872 | static int rescuer_thread(void *__wq) | |
1873 | { | |
1874 | struct workqueue_struct *wq = __wq; | |
1875 | struct worker *rescuer = wq->rescuer; | |
1876 | struct list_head *scheduled = &rescuer->scheduled; | |
1877 | unsigned int cpu; | |
1878 | ||
1879 | set_user_nice(current, RESCUER_NICE_LEVEL); | |
1880 | repeat: | |
1881 | set_current_state(TASK_INTERRUPTIBLE); | |
1882 | ||
1883 | if (kthread_should_stop()) | |
1884 | return 0; | |
1885 | ||
1886 | for_each_cpu(cpu, wq->mayday_mask) { | |
1887 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
1888 | struct global_cwq *gcwq = cwq->gcwq; | |
1889 | struct work_struct *work, *n; | |
1890 | ||
1891 | __set_current_state(TASK_RUNNING); | |
1892 | cpumask_clear_cpu(cpu, wq->mayday_mask); | |
1893 | ||
1894 | /* migrate to the target cpu if possible */ | |
1895 | rescuer->gcwq = gcwq; | |
1896 | worker_maybe_bind_and_lock(rescuer); | |
1897 | ||
1898 | /* | |
1899 | * Slurp in all works issued via this workqueue and | |
1900 | * process'em. | |
1901 | */ | |
1902 | BUG_ON(!list_empty(&rescuer->scheduled)); | |
1903 | list_for_each_entry_safe(work, n, &gcwq->worklist, entry) | |
1904 | if (get_work_cwq(work) == cwq) | |
1905 | move_linked_works(work, scheduled, &n); | |
1906 | ||
1907 | process_scheduled_works(rescuer); | |
1908 | spin_unlock_irq(&gcwq->lock); | |
1909 | } | |
1910 | ||
1911 | schedule(); | |
1912 | goto repeat; | |
1913 | } | |
1914 | ||
fc2e4d70 ON |
1915 | struct wq_barrier { |
1916 | struct work_struct work; | |
1917 | struct completion done; | |
1918 | }; | |
1919 | ||
1920 | static void wq_barrier_func(struct work_struct *work) | |
1921 | { | |
1922 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); | |
1923 | complete(&barr->done); | |
1924 | } | |
1925 | ||
4690c4ab TH |
1926 | /** |
1927 | * insert_wq_barrier - insert a barrier work | |
1928 | * @cwq: cwq to insert barrier into | |
1929 | * @barr: wq_barrier to insert | |
affee4b2 TH |
1930 | * @target: target work to attach @barr to |
1931 | * @worker: worker currently executing @target, NULL if @target is not executing | |
4690c4ab | 1932 | * |
affee4b2 TH |
1933 | * @barr is linked to @target such that @barr is completed only after |
1934 | * @target finishes execution. Please note that the ordering | |
1935 | * guarantee is observed only with respect to @target and on the local | |
1936 | * cpu. | |
1937 | * | |
1938 | * Currently, a queued barrier can't be canceled. This is because | |
1939 | * try_to_grab_pending() can't determine whether the work to be | |
1940 | * grabbed is at the head of the queue and thus can't clear LINKED | |
1941 | * flag of the previous work while there must be a valid next work | |
1942 | * after a work with LINKED flag set. | |
1943 | * | |
1944 | * Note that when @worker is non-NULL, @target may be modified | |
1945 | * underneath us, so we can't reliably determine cwq from @target. | |
4690c4ab TH |
1946 | * |
1947 | * CONTEXT: | |
8b03ae3c | 1948 | * spin_lock_irq(gcwq->lock). |
4690c4ab | 1949 | */ |
83c22520 | 1950 | static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, |
affee4b2 TH |
1951 | struct wq_barrier *barr, |
1952 | struct work_struct *target, struct worker *worker) | |
fc2e4d70 | 1953 | { |
affee4b2 TH |
1954 | struct list_head *head; |
1955 | unsigned int linked = 0; | |
1956 | ||
dc186ad7 | 1957 | /* |
8b03ae3c | 1958 | * debugobject calls are safe here even with gcwq->lock locked |
dc186ad7 TG |
1959 | * as we know for sure that this will not trigger any of the |
1960 | * checks and call back into the fixup functions where we | |
1961 | * might deadlock. | |
1962 | */ | |
1963 | INIT_WORK_ON_STACK(&barr->work, wq_barrier_func); | |
22df02bb | 1964 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
fc2e4d70 | 1965 | init_completion(&barr->done); |
83c22520 | 1966 | |
affee4b2 TH |
1967 | /* |
1968 | * If @target is currently being executed, schedule the | |
1969 | * barrier to the worker; otherwise, put it after @target. | |
1970 | */ | |
1971 | if (worker) | |
1972 | head = worker->scheduled.next; | |
1973 | else { | |
1974 | unsigned long *bits = work_data_bits(target); | |
1975 | ||
1976 | head = target->entry.next; | |
1977 | /* there can already be other linked works, inherit and set */ | |
1978 | linked = *bits & WORK_STRUCT_LINKED; | |
1979 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); | |
1980 | } | |
1981 | ||
dc186ad7 | 1982 | debug_work_activate(&barr->work); |
affee4b2 TH |
1983 | insert_work(cwq, &barr->work, head, |
1984 | work_color_to_flags(WORK_NO_COLOR) | linked); | |
fc2e4d70 ON |
1985 | } |
1986 | ||
73f53c4a TH |
1987 | /** |
1988 | * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing | |
1989 | * @wq: workqueue being flushed | |
1990 | * @flush_color: new flush color, < 0 for no-op | |
1991 | * @work_color: new work color, < 0 for no-op | |
1992 | * | |
1993 | * Prepare cwqs for workqueue flushing. | |
1994 | * | |
1995 | * If @flush_color is non-negative, flush_color on all cwqs should be | |
1996 | * -1. If no cwq has in-flight commands at the specified color, all | |
1997 | * cwq->flush_color's stay at -1 and %false is returned. If any cwq | |
1998 | * has in flight commands, its cwq->flush_color is set to | |
1999 | * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq | |
2000 | * wakeup logic is armed and %true is returned. | |
2001 | * | |
2002 | * The caller should have initialized @wq->first_flusher prior to | |
2003 | * calling this function with non-negative @flush_color. If | |
2004 | * @flush_color is negative, no flush color update is done and %false | |
2005 | * is returned. | |
2006 | * | |
2007 | * If @work_color is non-negative, all cwqs should have the same | |
2008 | * work_color which is previous to @work_color and all will be | |
2009 | * advanced to @work_color. | |
2010 | * | |
2011 | * CONTEXT: | |
2012 | * mutex_lock(wq->flush_mutex). | |
2013 | * | |
2014 | * RETURNS: | |
2015 | * %true if @flush_color >= 0 and there's something to flush. %false | |
2016 | * otherwise. | |
2017 | */ | |
2018 | static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, | |
2019 | int flush_color, int work_color) | |
1da177e4 | 2020 | { |
73f53c4a TH |
2021 | bool wait = false; |
2022 | unsigned int cpu; | |
1da177e4 | 2023 | |
73f53c4a TH |
2024 | if (flush_color >= 0) { |
2025 | BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); | |
2026 | atomic_set(&wq->nr_cwqs_to_flush, 1); | |
1da177e4 | 2027 | } |
2355b70f | 2028 | |
73f53c4a TH |
2029 | for_each_possible_cpu(cpu) { |
2030 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
8b03ae3c | 2031 | struct global_cwq *gcwq = cwq->gcwq; |
73f53c4a | 2032 | |
8b03ae3c | 2033 | spin_lock_irq(&gcwq->lock); |
73f53c4a TH |
2034 | |
2035 | if (flush_color >= 0) { | |
2036 | BUG_ON(cwq->flush_color != -1); | |
2037 | ||
2038 | if (cwq->nr_in_flight[flush_color]) { | |
2039 | cwq->flush_color = flush_color; | |
2040 | atomic_inc(&wq->nr_cwqs_to_flush); | |
2041 | wait = true; | |
2042 | } | |
2043 | } | |
2044 | ||
2045 | if (work_color >= 0) { | |
2046 | BUG_ON(work_color != work_next_color(cwq->work_color)); | |
2047 | cwq->work_color = work_color; | |
2048 | } | |
2049 | ||
8b03ae3c | 2050 | spin_unlock_irq(&gcwq->lock); |
dc186ad7 | 2051 | } |
14441960 | 2052 | |
73f53c4a TH |
2053 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) |
2054 | complete(&wq->first_flusher->done); | |
2055 | ||
2056 | return wait; | |
1da177e4 LT |
2057 | } |
2058 | ||
0fcb78c2 | 2059 | /** |
1da177e4 | 2060 | * flush_workqueue - ensure that any scheduled work has run to completion. |
0fcb78c2 | 2061 | * @wq: workqueue to flush |
1da177e4 LT |
2062 | * |
2063 | * Forces execution of the workqueue and blocks until its completion. | |
2064 | * This is typically used in driver shutdown handlers. | |
2065 | * | |
fc2e4d70 ON |
2066 | * We sleep until all works which were queued on entry have been handled, |
2067 | * but we are not livelocked by new incoming ones. | |
1da177e4 | 2068 | */ |
7ad5b3a5 | 2069 | void flush_workqueue(struct workqueue_struct *wq) |
1da177e4 | 2070 | { |
73f53c4a TH |
2071 | struct wq_flusher this_flusher = { |
2072 | .list = LIST_HEAD_INIT(this_flusher.list), | |
2073 | .flush_color = -1, | |
2074 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), | |
2075 | }; | |
2076 | int next_color; | |
1da177e4 | 2077 | |
3295f0ef IM |
2078 | lock_map_acquire(&wq->lockdep_map); |
2079 | lock_map_release(&wq->lockdep_map); | |
73f53c4a TH |
2080 | |
2081 | mutex_lock(&wq->flush_mutex); | |
2082 | ||
2083 | /* | |
2084 | * Start-to-wait phase | |
2085 | */ | |
2086 | next_color = work_next_color(wq->work_color); | |
2087 | ||
2088 | if (next_color != wq->flush_color) { | |
2089 | /* | |
2090 | * Color space is not full. The current work_color | |
2091 | * becomes our flush_color and work_color is advanced | |
2092 | * by one. | |
2093 | */ | |
2094 | BUG_ON(!list_empty(&wq->flusher_overflow)); | |
2095 | this_flusher.flush_color = wq->work_color; | |
2096 | wq->work_color = next_color; | |
2097 | ||
2098 | if (!wq->first_flusher) { | |
2099 | /* no flush in progress, become the first flusher */ | |
2100 | BUG_ON(wq->flush_color != this_flusher.flush_color); | |
2101 | ||
2102 | wq->first_flusher = &this_flusher; | |
2103 | ||
2104 | if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, | |
2105 | wq->work_color)) { | |
2106 | /* nothing to flush, done */ | |
2107 | wq->flush_color = next_color; | |
2108 | wq->first_flusher = NULL; | |
2109 | goto out_unlock; | |
2110 | } | |
2111 | } else { | |
2112 | /* wait in queue */ | |
2113 | BUG_ON(wq->flush_color == this_flusher.flush_color); | |
2114 | list_add_tail(&this_flusher.list, &wq->flusher_queue); | |
2115 | flush_workqueue_prep_cwqs(wq, -1, wq->work_color); | |
2116 | } | |
2117 | } else { | |
2118 | /* | |
2119 | * Oops, color space is full, wait on overflow queue. | |
2120 | * The next flush completion will assign us | |
2121 | * flush_color and transfer to flusher_queue. | |
2122 | */ | |
2123 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); | |
2124 | } | |
2125 | ||
2126 | mutex_unlock(&wq->flush_mutex); | |
2127 | ||
2128 | wait_for_completion(&this_flusher.done); | |
2129 | ||
2130 | /* | |
2131 | * Wake-up-and-cascade phase | |
2132 | * | |
2133 | * First flushers are responsible for cascading flushes and | |
2134 | * handling overflow. Non-first flushers can simply return. | |
2135 | */ | |
2136 | if (wq->first_flusher != &this_flusher) | |
2137 | return; | |
2138 | ||
2139 | mutex_lock(&wq->flush_mutex); | |
2140 | ||
4ce48b37 TH |
2141 | /* we might have raced, check again with mutex held */ |
2142 | if (wq->first_flusher != &this_flusher) | |
2143 | goto out_unlock; | |
2144 | ||
73f53c4a TH |
2145 | wq->first_flusher = NULL; |
2146 | ||
2147 | BUG_ON(!list_empty(&this_flusher.list)); | |
2148 | BUG_ON(wq->flush_color != this_flusher.flush_color); | |
2149 | ||
2150 | while (true) { | |
2151 | struct wq_flusher *next, *tmp; | |
2152 | ||
2153 | /* complete all the flushers sharing the current flush color */ | |
2154 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { | |
2155 | if (next->flush_color != wq->flush_color) | |
2156 | break; | |
2157 | list_del_init(&next->list); | |
2158 | complete(&next->done); | |
2159 | } | |
2160 | ||
2161 | BUG_ON(!list_empty(&wq->flusher_overflow) && | |
2162 | wq->flush_color != work_next_color(wq->work_color)); | |
2163 | ||
2164 | /* this flush_color is finished, advance by one */ | |
2165 | wq->flush_color = work_next_color(wq->flush_color); | |
2166 | ||
2167 | /* one color has been freed, handle overflow queue */ | |
2168 | if (!list_empty(&wq->flusher_overflow)) { | |
2169 | /* | |
2170 | * Assign the same color to all overflowed | |
2171 | * flushers, advance work_color and append to | |
2172 | * flusher_queue. This is the start-to-wait | |
2173 | * phase for these overflowed flushers. | |
2174 | */ | |
2175 | list_for_each_entry(tmp, &wq->flusher_overflow, list) | |
2176 | tmp->flush_color = wq->work_color; | |
2177 | ||
2178 | wq->work_color = work_next_color(wq->work_color); | |
2179 | ||
2180 | list_splice_tail_init(&wq->flusher_overflow, | |
2181 | &wq->flusher_queue); | |
2182 | flush_workqueue_prep_cwqs(wq, -1, wq->work_color); | |
2183 | } | |
2184 | ||
2185 | if (list_empty(&wq->flusher_queue)) { | |
2186 | BUG_ON(wq->flush_color != wq->work_color); | |
2187 | break; | |
2188 | } | |
2189 | ||
2190 | /* | |
2191 | * Need to flush more colors. Make the next flusher | |
2192 | * the new first flusher and arm cwqs. | |
2193 | */ | |
2194 | BUG_ON(wq->flush_color == wq->work_color); | |
2195 | BUG_ON(wq->flush_color != next->flush_color); | |
2196 | ||
2197 | list_del_init(&next->list); | |
2198 | wq->first_flusher = next; | |
2199 | ||
2200 | if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) | |
2201 | break; | |
2202 | ||
2203 | /* | |
2204 | * Meh... this color is already done, clear first | |
2205 | * flusher and repeat cascading. | |
2206 | */ | |
2207 | wq->first_flusher = NULL; | |
2208 | } | |
2209 | ||
2210 | out_unlock: | |
2211 | mutex_unlock(&wq->flush_mutex); | |
1da177e4 | 2212 | } |
ae90dd5d | 2213 | EXPORT_SYMBOL_GPL(flush_workqueue); |
1da177e4 | 2214 | |
db700897 ON |
2215 | /** |
2216 | * flush_work - block until a work_struct's callback has terminated | |
2217 | * @work: the work which is to be flushed | |
2218 | * | |
a67da70d ON |
2219 | * Returns false if @work has already terminated. |
2220 | * | |
db700897 ON |
2221 | * It is expected that, prior to calling flush_work(), the caller has |
2222 | * arranged for the work to not be requeued, otherwise it doesn't make | |
2223 | * sense to use this function. | |
2224 | */ | |
2225 | int flush_work(struct work_struct *work) | |
2226 | { | |
affee4b2 | 2227 | struct worker *worker = NULL; |
8b03ae3c | 2228 | struct global_cwq *gcwq; |
7a22ad75 | 2229 | struct cpu_workqueue_struct *cwq; |
db700897 ON |
2230 | struct wq_barrier barr; |
2231 | ||
2232 | might_sleep(); | |
7a22ad75 TH |
2233 | gcwq = get_work_gcwq(work); |
2234 | if (!gcwq) | |
db700897 | 2235 | return 0; |
a67da70d | 2236 | |
8b03ae3c | 2237 | spin_lock_irq(&gcwq->lock); |
db700897 ON |
2238 | if (!list_empty(&work->entry)) { |
2239 | /* | |
2240 | * See the comment near try_to_grab_pending()->smp_rmb(). | |
7a22ad75 TH |
2241 | * If it was re-queued to a different gcwq under us, we |
2242 | * are not going to wait. | |
db700897 ON |
2243 | */ |
2244 | smp_rmb(); | |
7a22ad75 TH |
2245 | cwq = get_work_cwq(work); |
2246 | if (unlikely(!cwq || gcwq != cwq->gcwq)) | |
4690c4ab | 2247 | goto already_gone; |
db700897 | 2248 | } else { |
7a22ad75 | 2249 | worker = find_worker_executing_work(gcwq, work); |
affee4b2 | 2250 | if (!worker) |
4690c4ab | 2251 | goto already_gone; |
7a22ad75 | 2252 | cwq = worker->current_cwq; |
db700897 | 2253 | } |
db700897 | 2254 | |
affee4b2 | 2255 | insert_wq_barrier(cwq, &barr, work, worker); |
8b03ae3c | 2256 | spin_unlock_irq(&gcwq->lock); |
7a22ad75 TH |
2257 | |
2258 | lock_map_acquire(&cwq->wq->lockdep_map); | |
2259 | lock_map_release(&cwq->wq->lockdep_map); | |
2260 | ||
db700897 | 2261 | wait_for_completion(&barr.done); |
dc186ad7 | 2262 | destroy_work_on_stack(&barr.work); |
db700897 | 2263 | return 1; |
4690c4ab | 2264 | already_gone: |
8b03ae3c | 2265 | spin_unlock_irq(&gcwq->lock); |
4690c4ab | 2266 | return 0; |
db700897 ON |
2267 | } |
2268 | EXPORT_SYMBOL_GPL(flush_work); | |
2269 | ||
6e84d644 | 2270 | /* |
1f1f642e | 2271 | * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, |
6e84d644 ON |
2272 | * so this work can't be re-armed in any way. |
2273 | */ | |
2274 | static int try_to_grab_pending(struct work_struct *work) | |
2275 | { | |
8b03ae3c | 2276 | struct global_cwq *gcwq; |
1f1f642e | 2277 | int ret = -1; |
6e84d644 | 2278 | |
22df02bb | 2279 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
1f1f642e | 2280 | return 0; |
6e84d644 ON |
2281 | |
2282 | /* | |
2283 | * The queueing is in progress, or it is already queued. Try to | |
2284 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. | |
2285 | */ | |
7a22ad75 TH |
2286 | gcwq = get_work_gcwq(work); |
2287 | if (!gcwq) | |
6e84d644 ON |
2288 | return ret; |
2289 | ||
8b03ae3c | 2290 | spin_lock_irq(&gcwq->lock); |
6e84d644 ON |
2291 | if (!list_empty(&work->entry)) { |
2292 | /* | |
7a22ad75 | 2293 | * This work is queued, but perhaps we locked the wrong gcwq. |
6e84d644 ON |
2294 | * In that case we must see the new value after rmb(), see |
2295 | * insert_work()->wmb(). | |
2296 | */ | |
2297 | smp_rmb(); | |
7a22ad75 | 2298 | if (gcwq == get_work_gcwq(work)) { |
dc186ad7 | 2299 | debug_work_deactivate(work); |
6e84d644 | 2300 | list_del_init(&work->entry); |
7a22ad75 TH |
2301 | cwq_dec_nr_in_flight(get_work_cwq(work), |
2302 | get_work_color(work)); | |
6e84d644 ON |
2303 | ret = 1; |
2304 | } | |
2305 | } | |
8b03ae3c | 2306 | spin_unlock_irq(&gcwq->lock); |
6e84d644 ON |
2307 | |
2308 | return ret; | |
2309 | } | |
2310 | ||
7a22ad75 | 2311 | static void wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) |
b89deed3 ON |
2312 | { |
2313 | struct wq_barrier barr; | |
affee4b2 | 2314 | struct worker *worker; |
b89deed3 | 2315 | |
8b03ae3c | 2316 | spin_lock_irq(&gcwq->lock); |
affee4b2 | 2317 | |
7a22ad75 TH |
2318 | worker = find_worker_executing_work(gcwq, work); |
2319 | if (unlikely(worker)) | |
2320 | insert_wq_barrier(worker->current_cwq, &barr, work, worker); | |
affee4b2 | 2321 | |
8b03ae3c | 2322 | spin_unlock_irq(&gcwq->lock); |
b89deed3 | 2323 | |
affee4b2 | 2324 | if (unlikely(worker)) { |
b89deed3 | 2325 | wait_for_completion(&barr.done); |
dc186ad7 TG |
2326 | destroy_work_on_stack(&barr.work); |
2327 | } | |
b89deed3 ON |
2328 | } |
2329 | ||
6e84d644 | 2330 | static void wait_on_work(struct work_struct *work) |
b89deed3 | 2331 | { |
b1f4ec17 | 2332 | int cpu; |
b89deed3 | 2333 | |
f293ea92 ON |
2334 | might_sleep(); |
2335 | ||
3295f0ef IM |
2336 | lock_map_acquire(&work->lockdep_map); |
2337 | lock_map_release(&work->lockdep_map); | |
4e6045f1 | 2338 | |
1537663f | 2339 | for_each_possible_cpu(cpu) |
7a22ad75 | 2340 | wait_on_cpu_work(get_gcwq(cpu), work); |
6e84d644 ON |
2341 | } |
2342 | ||
1f1f642e ON |
2343 | static int __cancel_work_timer(struct work_struct *work, |
2344 | struct timer_list* timer) | |
2345 | { | |
2346 | int ret; | |
2347 | ||
2348 | do { | |
2349 | ret = (timer && likely(del_timer(timer))); | |
2350 | if (!ret) | |
2351 | ret = try_to_grab_pending(work); | |
2352 | wait_on_work(work); | |
2353 | } while (unlikely(ret < 0)); | |
2354 | ||
7a22ad75 | 2355 | clear_work_data(work); |
1f1f642e ON |
2356 | return ret; |
2357 | } | |
2358 | ||
6e84d644 ON |
2359 | /** |
2360 | * cancel_work_sync - block until a work_struct's callback has terminated | |
2361 | * @work: the work which is to be flushed | |
2362 | * | |
1f1f642e ON |
2363 | * Returns true if @work was pending. |
2364 | * | |
6e84d644 ON |
2365 | * cancel_work_sync() will cancel the work if it is queued. If the work's |
2366 | * callback appears to be running, cancel_work_sync() will block until it | |
2367 | * has completed. | |
2368 | * | |
2369 | * It is possible to use this function if the work re-queues itself. It can | |
2370 | * cancel the work even if it migrates to another workqueue, however in that | |
2371 | * case it only guarantees that work->func() has completed on the last queued | |
2372 | * workqueue. | |
2373 | * | |
2374 | * cancel_work_sync(&delayed_work->work) should be used only if ->timer is not | |
2375 | * pending, otherwise it goes into a busy-wait loop until the timer expires. | |
2376 | * | |
2377 | * The caller must ensure that workqueue_struct on which this work was last | |
2378 | * queued can't be destroyed before this function returns. | |
2379 | */ | |
1f1f642e | 2380 | int cancel_work_sync(struct work_struct *work) |
6e84d644 | 2381 | { |
1f1f642e | 2382 | return __cancel_work_timer(work, NULL); |
b89deed3 | 2383 | } |
28e53bdd | 2384 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
b89deed3 | 2385 | |
6e84d644 | 2386 | /** |
f5a421a4 | 2387 | * cancel_delayed_work_sync - reliably kill off a delayed work. |
6e84d644 ON |
2388 | * @dwork: the delayed work struct |
2389 | * | |
1f1f642e ON |
2390 | * Returns true if @dwork was pending. |
2391 | * | |
6e84d644 ON |
2392 | * It is possible to use this function if @dwork rearms itself via queue_work() |
2393 | * or queue_delayed_work(). See also the comment for cancel_work_sync(). | |
2394 | */ | |
1f1f642e | 2395 | int cancel_delayed_work_sync(struct delayed_work *dwork) |
6e84d644 | 2396 | { |
1f1f642e | 2397 | return __cancel_work_timer(&dwork->work, &dwork->timer); |
6e84d644 | 2398 | } |
f5a421a4 | 2399 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
1da177e4 | 2400 | |
0fcb78c2 REB |
2401 | /** |
2402 | * schedule_work - put work task in global workqueue | |
2403 | * @work: job to be done | |
2404 | * | |
5b0f437d BVA |
2405 | * Returns zero if @work was already on the kernel-global workqueue and |
2406 | * non-zero otherwise. | |
2407 | * | |
2408 | * This puts a job in the kernel-global workqueue if it was not already | |
2409 | * queued and leaves it in the same position on the kernel-global | |
2410 | * workqueue otherwise. | |
0fcb78c2 | 2411 | */ |
7ad5b3a5 | 2412 | int schedule_work(struct work_struct *work) |
1da177e4 | 2413 | { |
d320c038 | 2414 | return queue_work(system_wq, work); |
1da177e4 | 2415 | } |
ae90dd5d | 2416 | EXPORT_SYMBOL(schedule_work); |
1da177e4 | 2417 | |
c1a220e7 ZR |
2418 | /* |
2419 | * schedule_work_on - put work task on a specific cpu | |
2420 | * @cpu: cpu to put the work task on | |
2421 | * @work: job to be done | |
2422 | * | |
2423 | * This puts a job on a specific cpu | |
2424 | */ | |
2425 | int schedule_work_on(int cpu, struct work_struct *work) | |
2426 | { | |
d320c038 | 2427 | return queue_work_on(cpu, system_wq, work); |
c1a220e7 ZR |
2428 | } |
2429 | EXPORT_SYMBOL(schedule_work_on); | |
2430 | ||
0fcb78c2 REB |
2431 | /** |
2432 | * schedule_delayed_work - put work task in global workqueue after delay | |
52bad64d DH |
2433 | * @dwork: job to be done |
2434 | * @delay: number of jiffies to wait or 0 for immediate execution | |
0fcb78c2 REB |
2435 | * |
2436 | * After waiting for a given time this puts a job in the kernel-global | |
2437 | * workqueue. | |
2438 | */ | |
7ad5b3a5 | 2439 | int schedule_delayed_work(struct delayed_work *dwork, |
82f67cd9 | 2440 | unsigned long delay) |
1da177e4 | 2441 | { |
d320c038 | 2442 | return queue_delayed_work(system_wq, dwork, delay); |
1da177e4 | 2443 | } |
ae90dd5d | 2444 | EXPORT_SYMBOL(schedule_delayed_work); |
1da177e4 | 2445 | |
8c53e463 LT |
2446 | /** |
2447 | * flush_delayed_work - block until a dwork_struct's callback has terminated | |
2448 | * @dwork: the delayed work which is to be flushed | |
2449 | * | |
2450 | * Any timeout is cancelled, and any pending work is run immediately. | |
2451 | */ | |
2452 | void flush_delayed_work(struct delayed_work *dwork) | |
2453 | { | |
2454 | if (del_timer_sync(&dwork->timer)) { | |
7a22ad75 | 2455 | __queue_work(get_cpu(), get_work_cwq(&dwork->work)->wq, |
4690c4ab | 2456 | &dwork->work); |
8c53e463 LT |
2457 | put_cpu(); |
2458 | } | |
2459 | flush_work(&dwork->work); | |
2460 | } | |
2461 | EXPORT_SYMBOL(flush_delayed_work); | |
2462 | ||
0fcb78c2 REB |
2463 | /** |
2464 | * schedule_delayed_work_on - queue work in global workqueue on CPU after delay | |
2465 | * @cpu: cpu to use | |
52bad64d | 2466 | * @dwork: job to be done |
0fcb78c2 REB |
2467 | * @delay: number of jiffies to wait |
2468 | * | |
2469 | * After waiting for a given time this puts a job in the kernel-global | |
2470 | * workqueue on the specified CPU. | |
2471 | */ | |
1da177e4 | 2472 | int schedule_delayed_work_on(int cpu, |
52bad64d | 2473 | struct delayed_work *dwork, unsigned long delay) |
1da177e4 | 2474 | { |
d320c038 | 2475 | return queue_delayed_work_on(cpu, system_wq, dwork, delay); |
1da177e4 | 2476 | } |
ae90dd5d | 2477 | EXPORT_SYMBOL(schedule_delayed_work_on); |
1da177e4 | 2478 | |
b6136773 AM |
2479 | /** |
2480 | * schedule_on_each_cpu - call a function on each online CPU from keventd | |
2481 | * @func: the function to call | |
b6136773 AM |
2482 | * |
2483 | * Returns zero on success. | |
2484 | * Returns -ve errno on failure. | |
2485 | * | |
b6136773 AM |
2486 | * schedule_on_each_cpu() is very slow. |
2487 | */ | |
65f27f38 | 2488 | int schedule_on_each_cpu(work_func_t func) |
15316ba8 CL |
2489 | { |
2490 | int cpu; | |
b6136773 | 2491 | struct work_struct *works; |
15316ba8 | 2492 | |
b6136773 AM |
2493 | works = alloc_percpu(struct work_struct); |
2494 | if (!works) | |
15316ba8 | 2495 | return -ENOMEM; |
b6136773 | 2496 | |
93981800 TH |
2497 | get_online_cpus(); |
2498 | ||
15316ba8 | 2499 | for_each_online_cpu(cpu) { |
9bfb1839 IM |
2500 | struct work_struct *work = per_cpu_ptr(works, cpu); |
2501 | ||
2502 | INIT_WORK(work, func); | |
b71ab8c2 | 2503 | schedule_work_on(cpu, work); |
65a64464 | 2504 | } |
93981800 TH |
2505 | |
2506 | for_each_online_cpu(cpu) | |
2507 | flush_work(per_cpu_ptr(works, cpu)); | |
2508 | ||
95402b38 | 2509 | put_online_cpus(); |
b6136773 | 2510 | free_percpu(works); |
15316ba8 CL |
2511 | return 0; |
2512 | } | |
2513 | ||
eef6a7d5 AS |
2514 | /** |
2515 | * flush_scheduled_work - ensure that any scheduled work has run to completion. | |
2516 | * | |
2517 | * Forces execution of the kernel-global workqueue and blocks until its | |
2518 | * completion. | |
2519 | * | |
2520 | * Think twice before calling this function! It's very easy to get into | |
2521 | * trouble if you don't take great care. Either of the following situations | |
2522 | * will lead to deadlock: | |
2523 | * | |
2524 | * One of the work items currently on the workqueue needs to acquire | |
2525 | * a lock held by your code or its caller. | |
2526 | * | |
2527 | * Your code is running in the context of a work routine. | |
2528 | * | |
2529 | * They will be detected by lockdep when they occur, but the first might not | |
2530 | * occur very often. It depends on what work items are on the workqueue and | |
2531 | * what locks they need, which you have no control over. | |
2532 | * | |
2533 | * In most situations flushing the entire workqueue is overkill; you merely | |
2534 | * need to know that a particular work item isn't queued and isn't running. | |
2535 | * In such cases you should use cancel_delayed_work_sync() or | |
2536 | * cancel_work_sync() instead. | |
2537 | */ | |
1da177e4 LT |
2538 | void flush_scheduled_work(void) |
2539 | { | |
d320c038 | 2540 | flush_workqueue(system_wq); |
1da177e4 | 2541 | } |
ae90dd5d | 2542 | EXPORT_SYMBOL(flush_scheduled_work); |
1da177e4 | 2543 | |
1fa44eca JB |
2544 | /** |
2545 | * execute_in_process_context - reliably execute the routine with user context | |
2546 | * @fn: the function to execute | |
1fa44eca JB |
2547 | * @ew: guaranteed storage for the execute work structure (must |
2548 | * be available when the work executes) | |
2549 | * | |
2550 | * Executes the function immediately if process context is available, | |
2551 | * otherwise schedules the function for delayed execution. | |
2552 | * | |
2553 | * Returns: 0 - function was executed | |
2554 | * 1 - function was scheduled for execution | |
2555 | */ | |
65f27f38 | 2556 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
1fa44eca JB |
2557 | { |
2558 | if (!in_interrupt()) { | |
65f27f38 | 2559 | fn(&ew->work); |
1fa44eca JB |
2560 | return 0; |
2561 | } | |
2562 | ||
65f27f38 | 2563 | INIT_WORK(&ew->work, fn); |
1fa44eca JB |
2564 | schedule_work(&ew->work); |
2565 | ||
2566 | return 1; | |
2567 | } | |
2568 | EXPORT_SYMBOL_GPL(execute_in_process_context); | |
2569 | ||
1da177e4 LT |
2570 | int keventd_up(void) |
2571 | { | |
d320c038 | 2572 | return system_wq != NULL; |
1da177e4 LT |
2573 | } |
2574 | ||
0f900049 TH |
2575 | static struct cpu_workqueue_struct *alloc_cwqs(void) |
2576 | { | |
2577 | /* | |
2578 | * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. | |
2579 | * Make sure that the alignment isn't lower than that of | |
2580 | * unsigned long long. | |
2581 | */ | |
2582 | const size_t size = sizeof(struct cpu_workqueue_struct); | |
2583 | const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, | |
2584 | __alignof__(unsigned long long)); | |
2585 | struct cpu_workqueue_struct *cwqs; | |
2586 | #ifndef CONFIG_SMP | |
2587 | void *ptr; | |
2588 | ||
2589 | /* | |
2590 | * On UP, percpu allocator doesn't honor alignment parameter | |
2591 | * and simply uses arch-dependent default. Allocate enough | |
2592 | * room to align cwq and put an extra pointer at the end | |
2593 | * pointing back to the originally allocated pointer which | |
2594 | * will be used for free. | |
2595 | * | |
2596 | * FIXME: This really belongs to UP percpu code. Update UP | |
2597 | * percpu code to honor alignment and remove this ugliness. | |
2598 | */ | |
2599 | ptr = __alloc_percpu(size + align + sizeof(void *), 1); | |
2600 | cwqs = PTR_ALIGN(ptr, align); | |
2601 | *(void **)per_cpu_ptr(cwqs + 1, 0) = ptr; | |
2602 | #else | |
2603 | /* On SMP, percpu allocator can do it itself */ | |
2604 | cwqs = __alloc_percpu(size, align); | |
2605 | #endif | |
2606 | /* just in case, make sure it's actually aligned */ | |
2607 | BUG_ON(!IS_ALIGNED((unsigned long)cwqs, align)); | |
2608 | return cwqs; | |
2609 | } | |
2610 | ||
2611 | static void free_cwqs(struct cpu_workqueue_struct *cwqs) | |
2612 | { | |
2613 | #ifndef CONFIG_SMP | |
2614 | /* on UP, the pointer to free is stored right after the cwq */ | |
2615 | if (cwqs) | |
2616 | free_percpu(*(void **)per_cpu_ptr(cwqs + 1, 0)); | |
2617 | #else | |
2618 | free_percpu(cwqs); | |
2619 | #endif | |
2620 | } | |
2621 | ||
b71ab8c2 TH |
2622 | static int wq_clamp_max_active(int max_active, const char *name) |
2623 | { | |
2624 | if (max_active < 1 || max_active > WQ_MAX_ACTIVE) | |
2625 | printk(KERN_WARNING "workqueue: max_active %d requested for %s " | |
2626 | "is out of range, clamping between %d and %d\n", | |
2627 | max_active, name, 1, WQ_MAX_ACTIVE); | |
2628 | ||
2629 | return clamp_val(max_active, 1, WQ_MAX_ACTIVE); | |
2630 | } | |
2631 | ||
d320c038 TH |
2632 | struct workqueue_struct *__alloc_workqueue_key(const char *name, |
2633 | unsigned int flags, | |
2634 | int max_active, | |
2635 | struct lock_class_key *key, | |
2636 | const char *lock_name) | |
1da177e4 | 2637 | { |
1da177e4 | 2638 | struct workqueue_struct *wq; |
c34056a3 | 2639 | unsigned int cpu; |
1da177e4 | 2640 | |
d320c038 | 2641 | max_active = max_active ?: WQ_DFL_ACTIVE; |
b71ab8c2 | 2642 | max_active = wq_clamp_max_active(max_active, name); |
1e19ffc6 | 2643 | |
3af24433 ON |
2644 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
2645 | if (!wq) | |
4690c4ab | 2646 | goto err; |
3af24433 | 2647 | |
0f900049 | 2648 | wq->cpu_wq = alloc_cwqs(); |
4690c4ab TH |
2649 | if (!wq->cpu_wq) |
2650 | goto err; | |
3af24433 | 2651 | |
97e37d7b | 2652 | wq->flags = flags; |
a0a1a5fd | 2653 | wq->saved_max_active = max_active; |
73f53c4a TH |
2654 | mutex_init(&wq->flush_mutex); |
2655 | atomic_set(&wq->nr_cwqs_to_flush, 0); | |
2656 | INIT_LIST_HEAD(&wq->flusher_queue); | |
2657 | INIT_LIST_HEAD(&wq->flusher_overflow); | |
502ca9d8 TH |
2658 | wq->single_cpu = NR_CPUS; |
2659 | ||
3af24433 | 2660 | wq->name = name; |
eb13ba87 | 2661 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
cce1a165 | 2662 | INIT_LIST_HEAD(&wq->list); |
3af24433 | 2663 | |
1537663f TH |
2664 | for_each_possible_cpu(cpu) { |
2665 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
8b03ae3c | 2666 | struct global_cwq *gcwq = get_gcwq(cpu); |
1537663f | 2667 | |
0f900049 | 2668 | BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); |
8b03ae3c | 2669 | cwq->gcwq = gcwq; |
c34056a3 | 2670 | cwq->wq = wq; |
73f53c4a | 2671 | cwq->flush_color = -1; |
1e19ffc6 | 2672 | cwq->max_active = max_active; |
1e19ffc6 | 2673 | INIT_LIST_HEAD(&cwq->delayed_works); |
e22bee78 | 2674 | } |
1537663f | 2675 | |
e22bee78 TH |
2676 | if (flags & WQ_RESCUER) { |
2677 | struct worker *rescuer; | |
2678 | ||
2679 | if (!alloc_cpumask_var(&wq->mayday_mask, GFP_KERNEL)) | |
2680 | goto err; | |
2681 | ||
2682 | wq->rescuer = rescuer = alloc_worker(); | |
2683 | if (!rescuer) | |
2684 | goto err; | |
2685 | ||
2686 | rescuer->task = kthread_create(rescuer_thread, wq, "%s", name); | |
2687 | if (IS_ERR(rescuer->task)) | |
2688 | goto err; | |
2689 | ||
2690 | wq->rescuer = rescuer; | |
2691 | rescuer->task->flags |= PF_THREAD_BOUND; | |
2692 | wake_up_process(rescuer->task); | |
3af24433 ON |
2693 | } |
2694 | ||
a0a1a5fd TH |
2695 | /* |
2696 | * workqueue_lock protects global freeze state and workqueues | |
2697 | * list. Grab it, set max_active accordingly and add the new | |
2698 | * workqueue to workqueues list. | |
2699 | */ | |
1537663f | 2700 | spin_lock(&workqueue_lock); |
a0a1a5fd TH |
2701 | |
2702 | if (workqueue_freezing && wq->flags & WQ_FREEZEABLE) | |
2703 | for_each_possible_cpu(cpu) | |
2704 | get_cwq(cpu, wq)->max_active = 0; | |
2705 | ||
1537663f | 2706 | list_add(&wq->list, &workqueues); |
a0a1a5fd | 2707 | |
1537663f TH |
2708 | spin_unlock(&workqueue_lock); |
2709 | ||
3af24433 | 2710 | return wq; |
4690c4ab TH |
2711 | err: |
2712 | if (wq) { | |
0f900049 | 2713 | free_cwqs(wq->cpu_wq); |
e22bee78 TH |
2714 | free_cpumask_var(wq->mayday_mask); |
2715 | kfree(wq->rescuer); | |
4690c4ab TH |
2716 | kfree(wq); |
2717 | } | |
2718 | return NULL; | |
3af24433 | 2719 | } |
d320c038 | 2720 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
1da177e4 | 2721 | |
3af24433 ON |
2722 | /** |
2723 | * destroy_workqueue - safely terminate a workqueue | |
2724 | * @wq: target workqueue | |
2725 | * | |
2726 | * Safely destroy a workqueue. All work currently pending will be done first. | |
2727 | */ | |
2728 | void destroy_workqueue(struct workqueue_struct *wq) | |
2729 | { | |
c8e55f36 | 2730 | unsigned int cpu; |
3af24433 | 2731 | |
a0a1a5fd TH |
2732 | flush_workqueue(wq); |
2733 | ||
2734 | /* | |
2735 | * wq list is used to freeze wq, remove from list after | |
2736 | * flushing is complete in case freeze races us. | |
2737 | */ | |
95402b38 | 2738 | spin_lock(&workqueue_lock); |
b1f4ec17 | 2739 | list_del(&wq->list); |
95402b38 | 2740 | spin_unlock(&workqueue_lock); |
3af24433 | 2741 | |
e22bee78 | 2742 | /* sanity check */ |
73f53c4a TH |
2743 | for_each_possible_cpu(cpu) { |
2744 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
2745 | int i; | |
2746 | ||
73f53c4a TH |
2747 | for (i = 0; i < WORK_NR_COLORS; i++) |
2748 | BUG_ON(cwq->nr_in_flight[i]); | |
1e19ffc6 TH |
2749 | BUG_ON(cwq->nr_active); |
2750 | BUG_ON(!list_empty(&cwq->delayed_works)); | |
73f53c4a | 2751 | } |
9b41ea72 | 2752 | |
e22bee78 TH |
2753 | if (wq->flags & WQ_RESCUER) { |
2754 | kthread_stop(wq->rescuer->task); | |
2755 | free_cpumask_var(wq->mayday_mask); | |
2756 | } | |
2757 | ||
0f900049 | 2758 | free_cwqs(wq->cpu_wq); |
3af24433 ON |
2759 | kfree(wq); |
2760 | } | |
2761 | EXPORT_SYMBOL_GPL(destroy_workqueue); | |
2762 | ||
dcd989cb TH |
2763 | /** |
2764 | * workqueue_set_max_active - adjust max_active of a workqueue | |
2765 | * @wq: target workqueue | |
2766 | * @max_active: new max_active value. | |
2767 | * | |
2768 | * Set max_active of @wq to @max_active. | |
2769 | * | |
2770 | * CONTEXT: | |
2771 | * Don't call from IRQ context. | |
2772 | */ | |
2773 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) | |
2774 | { | |
2775 | unsigned int cpu; | |
2776 | ||
2777 | max_active = wq_clamp_max_active(max_active, wq->name); | |
2778 | ||
2779 | spin_lock(&workqueue_lock); | |
2780 | ||
2781 | wq->saved_max_active = max_active; | |
2782 | ||
2783 | for_each_possible_cpu(cpu) { | |
2784 | struct global_cwq *gcwq = get_gcwq(cpu); | |
2785 | ||
2786 | spin_lock_irq(&gcwq->lock); | |
2787 | ||
2788 | if (!(wq->flags & WQ_FREEZEABLE) || | |
2789 | !(gcwq->flags & GCWQ_FREEZING)) | |
2790 | get_cwq(gcwq->cpu, wq)->max_active = max_active; | |
2791 | ||
2792 | spin_unlock_irq(&gcwq->lock); | |
2793 | } | |
2794 | ||
2795 | spin_unlock(&workqueue_lock); | |
2796 | } | |
2797 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); | |
2798 | ||
2799 | /** | |
2800 | * workqueue_congested - test whether a workqueue is congested | |
2801 | * @cpu: CPU in question | |
2802 | * @wq: target workqueue | |
2803 | * | |
2804 | * Test whether @wq's cpu workqueue for @cpu is congested. There is | |
2805 | * no synchronization around this function and the test result is | |
2806 | * unreliable and only useful as advisory hints or for debugging. | |
2807 | * | |
2808 | * RETURNS: | |
2809 | * %true if congested, %false otherwise. | |
2810 | */ | |
2811 | bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) | |
2812 | { | |
2813 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
2814 | ||
2815 | return !list_empty(&cwq->delayed_works); | |
2816 | } | |
2817 | EXPORT_SYMBOL_GPL(workqueue_congested); | |
2818 | ||
2819 | /** | |
2820 | * work_cpu - return the last known associated cpu for @work | |
2821 | * @work: the work of interest | |
2822 | * | |
2823 | * RETURNS: | |
2824 | * CPU number if @work was ever queued. NR_CPUS otherwise. | |
2825 | */ | |
2826 | unsigned int work_cpu(struct work_struct *work) | |
2827 | { | |
2828 | struct global_cwq *gcwq = get_work_gcwq(work); | |
2829 | ||
2830 | return gcwq ? gcwq->cpu : NR_CPUS; | |
2831 | } | |
2832 | EXPORT_SYMBOL_GPL(work_cpu); | |
2833 | ||
2834 | /** | |
2835 | * work_busy - test whether a work is currently pending or running | |
2836 | * @work: the work to be tested | |
2837 | * | |
2838 | * Test whether @work is currently pending or running. There is no | |
2839 | * synchronization around this function and the test result is | |
2840 | * unreliable and only useful as advisory hints or for debugging. | |
2841 | * Especially for reentrant wqs, the pending state might hide the | |
2842 | * running state. | |
2843 | * | |
2844 | * RETURNS: | |
2845 | * OR'd bitmask of WORK_BUSY_* bits. | |
2846 | */ | |
2847 | unsigned int work_busy(struct work_struct *work) | |
2848 | { | |
2849 | struct global_cwq *gcwq = get_work_gcwq(work); | |
2850 | unsigned long flags; | |
2851 | unsigned int ret = 0; | |
2852 | ||
2853 | if (!gcwq) | |
2854 | return false; | |
2855 | ||
2856 | spin_lock_irqsave(&gcwq->lock, flags); | |
2857 | ||
2858 | if (work_pending(work)) | |
2859 | ret |= WORK_BUSY_PENDING; | |
2860 | if (find_worker_executing_work(gcwq, work)) | |
2861 | ret |= WORK_BUSY_RUNNING; | |
2862 | ||
2863 | spin_unlock_irqrestore(&gcwq->lock, flags); | |
2864 | ||
2865 | return ret; | |
2866 | } | |
2867 | EXPORT_SYMBOL_GPL(work_busy); | |
2868 | ||
db7bccf4 TH |
2869 | /* |
2870 | * CPU hotplug. | |
2871 | * | |
e22bee78 TH |
2872 | * There are two challenges in supporting CPU hotplug. Firstly, there |
2873 | * are a lot of assumptions on strong associations among work, cwq and | |
2874 | * gcwq which make migrating pending and scheduled works very | |
2875 | * difficult to implement without impacting hot paths. Secondly, | |
2876 | * gcwqs serve mix of short, long and very long running works making | |
2877 | * blocked draining impractical. | |
2878 | * | |
2879 | * This is solved by allowing a gcwq to be detached from CPU, running | |
2880 | * it with unbound (rogue) workers and allowing it to be reattached | |
2881 | * later if the cpu comes back online. A separate thread is created | |
2882 | * to govern a gcwq in such state and is called the trustee of the | |
2883 | * gcwq. | |
db7bccf4 TH |
2884 | * |
2885 | * Trustee states and their descriptions. | |
2886 | * | |
2887 | * START Command state used on startup. On CPU_DOWN_PREPARE, a | |
2888 | * new trustee is started with this state. | |
2889 | * | |
2890 | * IN_CHARGE Once started, trustee will enter this state after | |
e22bee78 TH |
2891 | * assuming the manager role and making all existing |
2892 | * workers rogue. DOWN_PREPARE waits for trustee to | |
2893 | * enter this state. After reaching IN_CHARGE, trustee | |
2894 | * tries to execute the pending worklist until it's empty | |
2895 | * and the state is set to BUTCHER, or the state is set | |
2896 | * to RELEASE. | |
db7bccf4 TH |
2897 | * |
2898 | * BUTCHER Command state which is set by the cpu callback after | |
2899 | * the cpu has went down. Once this state is set trustee | |
2900 | * knows that there will be no new works on the worklist | |
2901 | * and once the worklist is empty it can proceed to | |
2902 | * killing idle workers. | |
2903 | * | |
2904 | * RELEASE Command state which is set by the cpu callback if the | |
2905 | * cpu down has been canceled or it has come online | |
2906 | * again. After recognizing this state, trustee stops | |
e22bee78 TH |
2907 | * trying to drain or butcher and clears ROGUE, rebinds |
2908 | * all remaining workers back to the cpu and releases | |
2909 | * manager role. | |
db7bccf4 TH |
2910 | * |
2911 | * DONE Trustee will enter this state after BUTCHER or RELEASE | |
2912 | * is complete. | |
2913 | * | |
2914 | * trustee CPU draining | |
2915 | * took over down complete | |
2916 | * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE | |
2917 | * | | ^ | |
2918 | * | CPU is back online v return workers | | |
2919 | * ----------------> RELEASE -------------- | |
2920 | */ | |
2921 | ||
2922 | /** | |
2923 | * trustee_wait_event_timeout - timed event wait for trustee | |
2924 | * @cond: condition to wait for | |
2925 | * @timeout: timeout in jiffies | |
2926 | * | |
2927 | * wait_event_timeout() for trustee to use. Handles locking and | |
2928 | * checks for RELEASE request. | |
2929 | * | |
2930 | * CONTEXT: | |
2931 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
2932 | * multiple times. To be used by trustee. | |
2933 | * | |
2934 | * RETURNS: | |
2935 | * Positive indicating left time if @cond is satisfied, 0 if timed | |
2936 | * out, -1 if canceled. | |
2937 | */ | |
2938 | #define trustee_wait_event_timeout(cond, timeout) ({ \ | |
2939 | long __ret = (timeout); \ | |
2940 | while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ | |
2941 | __ret) { \ | |
2942 | spin_unlock_irq(&gcwq->lock); \ | |
2943 | __wait_event_timeout(gcwq->trustee_wait, (cond) || \ | |
2944 | (gcwq->trustee_state == TRUSTEE_RELEASE), \ | |
2945 | __ret); \ | |
2946 | spin_lock_irq(&gcwq->lock); \ | |
2947 | } \ | |
2948 | gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ | |
2949 | }) | |
2950 | ||
2951 | /** | |
2952 | * trustee_wait_event - event wait for trustee | |
2953 | * @cond: condition to wait for | |
2954 | * | |
2955 | * wait_event() for trustee to use. Automatically handles locking and | |
2956 | * checks for CANCEL request. | |
2957 | * | |
2958 | * CONTEXT: | |
2959 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
2960 | * multiple times. To be used by trustee. | |
2961 | * | |
2962 | * RETURNS: | |
2963 | * 0 if @cond is satisfied, -1 if canceled. | |
2964 | */ | |
2965 | #define trustee_wait_event(cond) ({ \ | |
2966 | long __ret1; \ | |
2967 | __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ | |
2968 | __ret1 < 0 ? -1 : 0; \ | |
2969 | }) | |
2970 | ||
2971 | static int __cpuinit trustee_thread(void *__gcwq) | |
2972 | { | |
2973 | struct global_cwq *gcwq = __gcwq; | |
2974 | struct worker *worker; | |
e22bee78 | 2975 | struct work_struct *work; |
db7bccf4 | 2976 | struct hlist_node *pos; |
e22bee78 | 2977 | long rc; |
db7bccf4 TH |
2978 | int i; |
2979 | ||
2980 | BUG_ON(gcwq->cpu != smp_processor_id()); | |
2981 | ||
2982 | spin_lock_irq(&gcwq->lock); | |
2983 | /* | |
e22bee78 TH |
2984 | * Claim the manager position and make all workers rogue. |
2985 | * Trustee must be bound to the target cpu and can't be | |
2986 | * cancelled. | |
db7bccf4 TH |
2987 | */ |
2988 | BUG_ON(gcwq->cpu != smp_processor_id()); | |
e22bee78 TH |
2989 | rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS)); |
2990 | BUG_ON(rc < 0); | |
2991 | ||
2992 | gcwq->flags |= GCWQ_MANAGING_WORKERS; | |
db7bccf4 TH |
2993 | |
2994 | list_for_each_entry(worker, &gcwq->idle_list, entry) | |
cb444766 | 2995 | worker->flags |= WORKER_ROGUE; |
db7bccf4 TH |
2996 | |
2997 | for_each_busy_worker(worker, i, pos, gcwq) | |
cb444766 | 2998 | worker->flags |= WORKER_ROGUE; |
db7bccf4 | 2999 | |
e22bee78 TH |
3000 | /* |
3001 | * Call schedule() so that we cross rq->lock and thus can | |
3002 | * guarantee sched callbacks see the rogue flag. This is | |
3003 | * necessary as scheduler callbacks may be invoked from other | |
3004 | * cpus. | |
3005 | */ | |
3006 | spin_unlock_irq(&gcwq->lock); | |
3007 | schedule(); | |
3008 | spin_lock_irq(&gcwq->lock); | |
3009 | ||
3010 | /* | |
cb444766 TH |
3011 | * Sched callbacks are disabled now. Zap nr_running. After |
3012 | * this, nr_running stays zero and need_more_worker() and | |
3013 | * keep_working() are always true as long as the worklist is | |
3014 | * not empty. | |
e22bee78 | 3015 | */ |
cb444766 | 3016 | atomic_set(get_gcwq_nr_running(gcwq->cpu), 0); |
e22bee78 TH |
3017 | |
3018 | spin_unlock_irq(&gcwq->lock); | |
3019 | del_timer_sync(&gcwq->idle_timer); | |
3020 | spin_lock_irq(&gcwq->lock); | |
3021 | ||
db7bccf4 TH |
3022 | /* |
3023 | * We're now in charge. Notify and proceed to drain. We need | |
3024 | * to keep the gcwq running during the whole CPU down | |
3025 | * procedure as other cpu hotunplug callbacks may need to | |
3026 | * flush currently running tasks. | |
3027 | */ | |
3028 | gcwq->trustee_state = TRUSTEE_IN_CHARGE; | |
3029 | wake_up_all(&gcwq->trustee_wait); | |
3030 | ||
3031 | /* | |
3032 | * The original cpu is in the process of dying and may go away | |
3033 | * anytime now. When that happens, we and all workers would | |
e22bee78 TH |
3034 | * be migrated to other cpus. Try draining any left work. We |
3035 | * want to get it over with ASAP - spam rescuers, wake up as | |
3036 | * many idlers as necessary and create new ones till the | |
3037 | * worklist is empty. Note that if the gcwq is frozen, there | |
3038 | * may be frozen works in freezeable cwqs. Don't declare | |
3039 | * completion while frozen. | |
db7bccf4 TH |
3040 | */ |
3041 | while (gcwq->nr_workers != gcwq->nr_idle || | |
3042 | gcwq->flags & GCWQ_FREEZING || | |
3043 | gcwq->trustee_state == TRUSTEE_IN_CHARGE) { | |
e22bee78 TH |
3044 | int nr_works = 0; |
3045 | ||
3046 | list_for_each_entry(work, &gcwq->worklist, entry) { | |
3047 | send_mayday(work); | |
3048 | nr_works++; | |
3049 | } | |
3050 | ||
3051 | list_for_each_entry(worker, &gcwq->idle_list, entry) { | |
3052 | if (!nr_works--) | |
3053 | break; | |
3054 | wake_up_process(worker->task); | |
3055 | } | |
3056 | ||
3057 | if (need_to_create_worker(gcwq)) { | |
3058 | spin_unlock_irq(&gcwq->lock); | |
3059 | worker = create_worker(gcwq, false); | |
3060 | spin_lock_irq(&gcwq->lock); | |
3061 | if (worker) { | |
cb444766 | 3062 | worker->flags |= WORKER_ROGUE; |
e22bee78 TH |
3063 | start_worker(worker); |
3064 | } | |
3065 | } | |
3066 | ||
db7bccf4 TH |
3067 | /* give a breather */ |
3068 | if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) | |
3069 | break; | |
3070 | } | |
3071 | ||
e22bee78 TH |
3072 | /* |
3073 | * Either all works have been scheduled and cpu is down, or | |
3074 | * cpu down has already been canceled. Wait for and butcher | |
3075 | * all workers till we're canceled. | |
3076 | */ | |
3077 | do { | |
3078 | rc = trustee_wait_event(!list_empty(&gcwq->idle_list)); | |
3079 | while (!list_empty(&gcwq->idle_list)) | |
3080 | destroy_worker(list_first_entry(&gcwq->idle_list, | |
3081 | struct worker, entry)); | |
3082 | } while (gcwq->nr_workers && rc >= 0); | |
3083 | ||
3084 | /* | |
3085 | * At this point, either draining has completed and no worker | |
3086 | * is left, or cpu down has been canceled or the cpu is being | |
3087 | * brought back up. There shouldn't be any idle one left. | |
3088 | * Tell the remaining busy ones to rebind once it finishes the | |
3089 | * currently scheduled works by scheduling the rebind_work. | |
3090 | */ | |
3091 | WARN_ON(!list_empty(&gcwq->idle_list)); | |
3092 | ||
3093 | for_each_busy_worker(worker, i, pos, gcwq) { | |
3094 | struct work_struct *rebind_work = &worker->rebind_work; | |
3095 | ||
3096 | /* | |
3097 | * Rebind_work may race with future cpu hotplug | |
3098 | * operations. Use a separate flag to mark that | |
3099 | * rebinding is scheduled. | |
3100 | */ | |
cb444766 TH |
3101 | worker->flags |= WORKER_REBIND; |
3102 | worker->flags &= ~WORKER_ROGUE; | |
e22bee78 TH |
3103 | |
3104 | /* queue rebind_work, wq doesn't matter, use the default one */ | |
3105 | if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, | |
3106 | work_data_bits(rebind_work))) | |
3107 | continue; | |
3108 | ||
3109 | debug_work_activate(rebind_work); | |
d320c038 | 3110 | insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, |
e22bee78 TH |
3111 | worker->scheduled.next, |
3112 | work_color_to_flags(WORK_NO_COLOR)); | |
3113 | } | |
3114 | ||
3115 | /* relinquish manager role */ | |
3116 | gcwq->flags &= ~GCWQ_MANAGING_WORKERS; | |
3117 | ||
db7bccf4 TH |
3118 | /* notify completion */ |
3119 | gcwq->trustee = NULL; | |
3120 | gcwq->trustee_state = TRUSTEE_DONE; | |
3121 | wake_up_all(&gcwq->trustee_wait); | |
3122 | spin_unlock_irq(&gcwq->lock); | |
3123 | return 0; | |
3124 | } | |
3125 | ||
3126 | /** | |
3127 | * wait_trustee_state - wait for trustee to enter the specified state | |
3128 | * @gcwq: gcwq the trustee of interest belongs to | |
3129 | * @state: target state to wait for | |
3130 | * | |
3131 | * Wait for the trustee to reach @state. DONE is already matched. | |
3132 | * | |
3133 | * CONTEXT: | |
3134 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed | |
3135 | * multiple times. To be used by cpu_callback. | |
3136 | */ | |
3137 | static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) | |
3138 | { | |
3139 | if (!(gcwq->trustee_state == state || | |
3140 | gcwq->trustee_state == TRUSTEE_DONE)) { | |
3141 | spin_unlock_irq(&gcwq->lock); | |
3142 | __wait_event(gcwq->trustee_wait, | |
3143 | gcwq->trustee_state == state || | |
3144 | gcwq->trustee_state == TRUSTEE_DONE); | |
3145 | spin_lock_irq(&gcwq->lock); | |
3146 | } | |
3147 | } | |
3148 | ||
3af24433 ON |
3149 | static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, |
3150 | unsigned long action, | |
3151 | void *hcpu) | |
3152 | { | |
3153 | unsigned int cpu = (unsigned long)hcpu; | |
db7bccf4 TH |
3154 | struct global_cwq *gcwq = get_gcwq(cpu); |
3155 | struct task_struct *new_trustee = NULL; | |
e22bee78 | 3156 | struct worker *uninitialized_var(new_worker); |
db7bccf4 | 3157 | unsigned long flags; |
3af24433 | 3158 | |
8bb78442 RW |
3159 | action &= ~CPU_TASKS_FROZEN; |
3160 | ||
db7bccf4 TH |
3161 | switch (action) { |
3162 | case CPU_DOWN_PREPARE: | |
3163 | new_trustee = kthread_create(trustee_thread, gcwq, | |
3164 | "workqueue_trustee/%d\n", cpu); | |
3165 | if (IS_ERR(new_trustee)) | |
3166 | return notifier_from_errno(PTR_ERR(new_trustee)); | |
3167 | kthread_bind(new_trustee, cpu); | |
e22bee78 TH |
3168 | /* fall through */ |
3169 | case CPU_UP_PREPARE: | |
3170 | BUG_ON(gcwq->first_idle); | |
3171 | new_worker = create_worker(gcwq, false); | |
3172 | if (!new_worker) { | |
3173 | if (new_trustee) | |
3174 | kthread_stop(new_trustee); | |
3175 | return NOTIFY_BAD; | |
3176 | } | |
db7bccf4 | 3177 | } |
3af24433 | 3178 | |
db7bccf4 TH |
3179 | /* some are called w/ irq disabled, don't disturb irq status */ |
3180 | spin_lock_irqsave(&gcwq->lock, flags); | |
3af24433 | 3181 | |
db7bccf4 TH |
3182 | switch (action) { |
3183 | case CPU_DOWN_PREPARE: | |
3184 | /* initialize trustee and tell it to acquire the gcwq */ | |
3185 | BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); | |
3186 | gcwq->trustee = new_trustee; | |
3187 | gcwq->trustee_state = TRUSTEE_START; | |
3188 | wake_up_process(gcwq->trustee); | |
3189 | wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); | |
e22bee78 TH |
3190 | /* fall through */ |
3191 | case CPU_UP_PREPARE: | |
3192 | BUG_ON(gcwq->first_idle); | |
3193 | gcwq->first_idle = new_worker; | |
3194 | break; | |
3195 | ||
3196 | case CPU_DYING: | |
3197 | /* | |
3198 | * Before this, the trustee and all workers except for | |
3199 | * the ones which are still executing works from | |
3200 | * before the last CPU down must be on the cpu. After | |
3201 | * this, they'll all be diasporas. | |
3202 | */ | |
3203 | gcwq->flags |= GCWQ_DISASSOCIATED; | |
db7bccf4 TH |
3204 | break; |
3205 | ||
3206 | case CPU_POST_DEAD: | |
3207 | gcwq->trustee_state = TRUSTEE_BUTCHER; | |
e22bee78 TH |
3208 | /* fall through */ |
3209 | case CPU_UP_CANCELED: | |
3210 | destroy_worker(gcwq->first_idle); | |
3211 | gcwq->first_idle = NULL; | |
db7bccf4 TH |
3212 | break; |
3213 | ||
3214 | case CPU_DOWN_FAILED: | |
3215 | case CPU_ONLINE: | |
e22bee78 | 3216 | gcwq->flags &= ~GCWQ_DISASSOCIATED; |
db7bccf4 TH |
3217 | if (gcwq->trustee_state != TRUSTEE_DONE) { |
3218 | gcwq->trustee_state = TRUSTEE_RELEASE; | |
3219 | wake_up_process(gcwq->trustee); | |
3220 | wait_trustee_state(gcwq, TRUSTEE_DONE); | |
3af24433 | 3221 | } |
db7bccf4 | 3222 | |
e22bee78 TH |
3223 | /* |
3224 | * Trustee is done and there might be no worker left. | |
3225 | * Put the first_idle in and request a real manager to | |
3226 | * take a look. | |
3227 | */ | |
3228 | spin_unlock_irq(&gcwq->lock); | |
3229 | kthread_bind(gcwq->first_idle->task, cpu); | |
3230 | spin_lock_irq(&gcwq->lock); | |
3231 | gcwq->flags |= GCWQ_MANAGE_WORKERS; | |
3232 | start_worker(gcwq->first_idle); | |
3233 | gcwq->first_idle = NULL; | |
db7bccf4 | 3234 | break; |
1da177e4 LT |
3235 | } |
3236 | ||
db7bccf4 TH |
3237 | spin_unlock_irqrestore(&gcwq->lock, flags); |
3238 | ||
1537663f | 3239 | return notifier_from_errno(0); |
1da177e4 | 3240 | } |
1da177e4 | 3241 | |
2d3854a3 | 3242 | #ifdef CONFIG_SMP |
8ccad40d | 3243 | |
2d3854a3 | 3244 | struct work_for_cpu { |
6b44003e | 3245 | struct completion completion; |
2d3854a3 RR |
3246 | long (*fn)(void *); |
3247 | void *arg; | |
3248 | long ret; | |
3249 | }; | |
3250 | ||
6b44003e | 3251 | static int do_work_for_cpu(void *_wfc) |
2d3854a3 | 3252 | { |
6b44003e | 3253 | struct work_for_cpu *wfc = _wfc; |
2d3854a3 | 3254 | wfc->ret = wfc->fn(wfc->arg); |
6b44003e AM |
3255 | complete(&wfc->completion); |
3256 | return 0; | |
2d3854a3 RR |
3257 | } |
3258 | ||
3259 | /** | |
3260 | * work_on_cpu - run a function in user context on a particular cpu | |
3261 | * @cpu: the cpu to run on | |
3262 | * @fn: the function to run | |
3263 | * @arg: the function arg | |
3264 | * | |
31ad9081 RR |
3265 | * This will return the value @fn returns. |
3266 | * It is up to the caller to ensure that the cpu doesn't go offline. | |
6b44003e | 3267 | * The caller must not hold any locks which would prevent @fn from completing. |
2d3854a3 RR |
3268 | */ |
3269 | long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) | |
3270 | { | |
6b44003e AM |
3271 | struct task_struct *sub_thread; |
3272 | struct work_for_cpu wfc = { | |
3273 | .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), | |
3274 | .fn = fn, | |
3275 | .arg = arg, | |
3276 | }; | |
3277 | ||
3278 | sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); | |
3279 | if (IS_ERR(sub_thread)) | |
3280 | return PTR_ERR(sub_thread); | |
3281 | kthread_bind(sub_thread, cpu); | |
3282 | wake_up_process(sub_thread); | |
3283 | wait_for_completion(&wfc.completion); | |
2d3854a3 RR |
3284 | return wfc.ret; |
3285 | } | |
3286 | EXPORT_SYMBOL_GPL(work_on_cpu); | |
3287 | #endif /* CONFIG_SMP */ | |
3288 | ||
a0a1a5fd TH |
3289 | #ifdef CONFIG_FREEZER |
3290 | ||
3291 | /** | |
3292 | * freeze_workqueues_begin - begin freezing workqueues | |
3293 | * | |
3294 | * Start freezing workqueues. After this function returns, all | |
3295 | * freezeable workqueues will queue new works to their frozen_works | |
7e11629d | 3296 | * list instead of gcwq->worklist. |
a0a1a5fd TH |
3297 | * |
3298 | * CONTEXT: | |
8b03ae3c | 3299 | * Grabs and releases workqueue_lock and gcwq->lock's. |
a0a1a5fd TH |
3300 | */ |
3301 | void freeze_workqueues_begin(void) | |
3302 | { | |
3303 | struct workqueue_struct *wq; | |
3304 | unsigned int cpu; | |
3305 | ||
3306 | spin_lock(&workqueue_lock); | |
3307 | ||
3308 | BUG_ON(workqueue_freezing); | |
3309 | workqueue_freezing = true; | |
3310 | ||
3311 | for_each_possible_cpu(cpu) { | |
8b03ae3c TH |
3312 | struct global_cwq *gcwq = get_gcwq(cpu); |
3313 | ||
3314 | spin_lock_irq(&gcwq->lock); | |
3315 | ||
db7bccf4 TH |
3316 | BUG_ON(gcwq->flags & GCWQ_FREEZING); |
3317 | gcwq->flags |= GCWQ_FREEZING; | |
3318 | ||
a0a1a5fd TH |
3319 | list_for_each_entry(wq, &workqueues, list) { |
3320 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
3321 | ||
a0a1a5fd TH |
3322 | if (wq->flags & WQ_FREEZEABLE) |
3323 | cwq->max_active = 0; | |
a0a1a5fd | 3324 | } |
8b03ae3c TH |
3325 | |
3326 | spin_unlock_irq(&gcwq->lock); | |
a0a1a5fd TH |
3327 | } |
3328 | ||
3329 | spin_unlock(&workqueue_lock); | |
3330 | } | |
3331 | ||
3332 | /** | |
3333 | * freeze_workqueues_busy - are freezeable workqueues still busy? | |
3334 | * | |
3335 | * Check whether freezing is complete. This function must be called | |
3336 | * between freeze_workqueues_begin() and thaw_workqueues(). | |
3337 | * | |
3338 | * CONTEXT: | |
3339 | * Grabs and releases workqueue_lock. | |
3340 | * | |
3341 | * RETURNS: | |
3342 | * %true if some freezeable workqueues are still busy. %false if | |
3343 | * freezing is complete. | |
3344 | */ | |
3345 | bool freeze_workqueues_busy(void) | |
3346 | { | |
3347 | struct workqueue_struct *wq; | |
3348 | unsigned int cpu; | |
3349 | bool busy = false; | |
3350 | ||
3351 | spin_lock(&workqueue_lock); | |
3352 | ||
3353 | BUG_ON(!workqueue_freezing); | |
3354 | ||
3355 | for_each_possible_cpu(cpu) { | |
3356 | /* | |
3357 | * nr_active is monotonically decreasing. It's safe | |
3358 | * to peek without lock. | |
3359 | */ | |
3360 | list_for_each_entry(wq, &workqueues, list) { | |
3361 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
3362 | ||
3363 | if (!(wq->flags & WQ_FREEZEABLE)) | |
3364 | continue; | |
3365 | ||
3366 | BUG_ON(cwq->nr_active < 0); | |
3367 | if (cwq->nr_active) { | |
3368 | busy = true; | |
3369 | goto out_unlock; | |
3370 | } | |
3371 | } | |
3372 | } | |
3373 | out_unlock: | |
3374 | spin_unlock(&workqueue_lock); | |
3375 | return busy; | |
3376 | } | |
3377 | ||
3378 | /** | |
3379 | * thaw_workqueues - thaw workqueues | |
3380 | * | |
3381 | * Thaw workqueues. Normal queueing is restored and all collected | |
7e11629d | 3382 | * frozen works are transferred to their respective gcwq worklists. |
a0a1a5fd TH |
3383 | * |
3384 | * CONTEXT: | |
8b03ae3c | 3385 | * Grabs and releases workqueue_lock and gcwq->lock's. |
a0a1a5fd TH |
3386 | */ |
3387 | void thaw_workqueues(void) | |
3388 | { | |
3389 | struct workqueue_struct *wq; | |
3390 | unsigned int cpu; | |
3391 | ||
3392 | spin_lock(&workqueue_lock); | |
3393 | ||
3394 | if (!workqueue_freezing) | |
3395 | goto out_unlock; | |
3396 | ||
3397 | for_each_possible_cpu(cpu) { | |
8b03ae3c TH |
3398 | struct global_cwq *gcwq = get_gcwq(cpu); |
3399 | ||
3400 | spin_lock_irq(&gcwq->lock); | |
3401 | ||
db7bccf4 TH |
3402 | BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); |
3403 | gcwq->flags &= ~GCWQ_FREEZING; | |
3404 | ||
a0a1a5fd TH |
3405 | list_for_each_entry(wq, &workqueues, list) { |
3406 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); | |
3407 | ||
3408 | if (!(wq->flags & WQ_FREEZEABLE)) | |
3409 | continue; | |
3410 | ||
a0a1a5fd TH |
3411 | /* restore max_active and repopulate worklist */ |
3412 | cwq->max_active = wq->saved_max_active; | |
3413 | ||
3414 | while (!list_empty(&cwq->delayed_works) && | |
3415 | cwq->nr_active < cwq->max_active) | |
3416 | cwq_activate_first_delayed(cwq); | |
3417 | ||
502ca9d8 TH |
3418 | /* perform delayed unbind from single cpu if empty */ |
3419 | if (wq->single_cpu == gcwq->cpu && | |
3420 | !cwq->nr_active && list_empty(&cwq->delayed_works)) | |
3421 | cwq_unbind_single_cpu(cwq); | |
a0a1a5fd | 3422 | } |
8b03ae3c | 3423 | |
e22bee78 TH |
3424 | wake_up_worker(gcwq); |
3425 | ||
8b03ae3c | 3426 | spin_unlock_irq(&gcwq->lock); |
a0a1a5fd TH |
3427 | } |
3428 | ||
3429 | workqueue_freezing = false; | |
3430 | out_unlock: | |
3431 | spin_unlock(&workqueue_lock); | |
3432 | } | |
3433 | #endif /* CONFIG_FREEZER */ | |
3434 | ||
c12920d1 | 3435 | void __init init_workqueues(void) |
1da177e4 | 3436 | { |
c34056a3 | 3437 | unsigned int cpu; |
c8e55f36 | 3438 | int i; |
c34056a3 | 3439 | |
7a22ad75 TH |
3440 | /* |
3441 | * The pointer part of work->data is either pointing to the | |
3442 | * cwq or contains the cpu number the work ran last on. Make | |
3443 | * sure cpu number won't overflow into kernel pointer area so | |
3444 | * that they can be distinguished. | |
3445 | */ | |
3446 | BUILD_BUG_ON(NR_CPUS << WORK_STRUCT_FLAG_BITS >= PAGE_OFFSET); | |
3447 | ||
db7bccf4 | 3448 | hotcpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE); |
8b03ae3c TH |
3449 | |
3450 | /* initialize gcwqs */ | |
3451 | for_each_possible_cpu(cpu) { | |
3452 | struct global_cwq *gcwq = get_gcwq(cpu); | |
3453 | ||
3454 | spin_lock_init(&gcwq->lock); | |
7e11629d | 3455 | INIT_LIST_HEAD(&gcwq->worklist); |
8b03ae3c TH |
3456 | gcwq->cpu = cpu; |
3457 | ||
c8e55f36 TH |
3458 | INIT_LIST_HEAD(&gcwq->idle_list); |
3459 | for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) | |
3460 | INIT_HLIST_HEAD(&gcwq->busy_hash[i]); | |
3461 | ||
e22bee78 TH |
3462 | init_timer_deferrable(&gcwq->idle_timer); |
3463 | gcwq->idle_timer.function = idle_worker_timeout; | |
3464 | gcwq->idle_timer.data = (unsigned long)gcwq; | |
3465 | ||
3466 | setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout, | |
3467 | (unsigned long)gcwq); | |
3468 | ||
8b03ae3c | 3469 | ida_init(&gcwq->worker_ida); |
db7bccf4 TH |
3470 | |
3471 | gcwq->trustee_state = TRUSTEE_DONE; | |
3472 | init_waitqueue_head(&gcwq->trustee_wait); | |
8b03ae3c TH |
3473 | } |
3474 | ||
e22bee78 TH |
3475 | /* create the initial worker */ |
3476 | for_each_online_cpu(cpu) { | |
3477 | struct global_cwq *gcwq = get_gcwq(cpu); | |
3478 | struct worker *worker; | |
3479 | ||
3480 | worker = create_worker(gcwq, true); | |
3481 | BUG_ON(!worker); | |
3482 | spin_lock_irq(&gcwq->lock); | |
3483 | start_worker(worker); | |
3484 | spin_unlock_irq(&gcwq->lock); | |
3485 | } | |
3486 | ||
d320c038 TH |
3487 | system_wq = alloc_workqueue("events", 0, 0); |
3488 | system_long_wq = alloc_workqueue("events_long", 0, 0); | |
3489 | system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); | |
3490 | BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq); | |
1da177e4 | 3491 | } |