Root/
1 | /* |
2 | * kernel/workqueue.c - generic async execution with shared worker pool |
3 | * |
4 | * Copyright (C) 2002 Ingo Molnar |
5 | * |
6 | * Derived from the taskqueue/keventd code by: |
7 | * David Woodhouse <dwmw2@infradead.org> |
8 | * Andrew Morton |
9 | * Kai Petzke <wpp@marie.physik.tu-berlin.de> |
10 | * Theodore Ts'o <tytso@mit.edu> |
11 | * |
12 | * Made to use alloc_percpu by Christoph Lameter. |
13 | * |
14 | * Copyright (C) 2010 SUSE Linux Products GmbH |
15 | * Copyright (C) 2010 Tejun Heo <tj@kernel.org> |
16 | * |
17 | * This is the generic async execution mechanism. Work items as are |
18 | * executed in process context. The worker pool is shared and |
19 | * automatically managed. There is one worker pool for each CPU and |
20 | * one extra for works which are better served by workers which are |
21 | * not bound to any specific CPU. |
22 | * |
23 | * Please read Documentation/workqueue.txt for details. |
24 | */ |
25 | |
26 | #include <linux/module.h> |
27 | #include <linux/kernel.h> |
28 | #include <linux/sched.h> |
29 | #include <linux/init.h> |
30 | #include <linux/signal.h> |
31 | #include <linux/completion.h> |
32 | #include <linux/workqueue.h> |
33 | #include <linux/slab.h> |
34 | #include <linux/cpu.h> |
35 | #include <linux/notifier.h> |
36 | #include <linux/kthread.h> |
37 | #include <linux/hardirq.h> |
38 | #include <linux/mempolicy.h> |
39 | #include <linux/freezer.h> |
40 | #include <linux/kallsyms.h> |
41 | #include <linux/debug_locks.h> |
42 | #include <linux/lockdep.h> |
43 | #include <linux/idr.h> |
44 | |
45 | #include "workqueue_sched.h" |
46 | |
47 | enum { |
48 | /* global_cwq flags */ |
49 | GCWQ_MANAGE_WORKERS = 1 << 0, /* need to manage workers */ |
50 | GCWQ_MANAGING_WORKERS = 1 << 1, /* managing workers */ |
51 | GCWQ_DISASSOCIATED = 1 << 2, /* cpu can't serve workers */ |
52 | GCWQ_FREEZING = 1 << 3, /* freeze in progress */ |
53 | GCWQ_HIGHPRI_PENDING = 1 << 4, /* highpri works on queue */ |
54 | |
55 | /* worker flags */ |
56 | WORKER_STARTED = 1 << 0, /* started */ |
57 | WORKER_DIE = 1 << 1, /* die die die */ |
58 | WORKER_IDLE = 1 << 2, /* is idle */ |
59 | WORKER_PREP = 1 << 3, /* preparing to run works */ |
60 | WORKER_ROGUE = 1 << 4, /* not bound to any cpu */ |
61 | WORKER_REBIND = 1 << 5, /* mom is home, come back */ |
62 | WORKER_CPU_INTENSIVE = 1 << 6, /* cpu intensive */ |
63 | WORKER_UNBOUND = 1 << 7, /* worker is unbound */ |
64 | |
65 | WORKER_NOT_RUNNING = WORKER_PREP | WORKER_ROGUE | WORKER_REBIND | |
66 | WORKER_CPU_INTENSIVE | WORKER_UNBOUND, |
67 | |
68 | /* gcwq->trustee_state */ |
69 | TRUSTEE_START = 0, /* start */ |
70 | TRUSTEE_IN_CHARGE = 1, /* trustee in charge of gcwq */ |
71 | TRUSTEE_BUTCHER = 2, /* butcher workers */ |
72 | TRUSTEE_RELEASE = 3, /* release workers */ |
73 | TRUSTEE_DONE = 4, /* trustee is done */ |
74 | |
75 | BUSY_WORKER_HASH_ORDER = 6, /* 64 pointers */ |
76 | BUSY_WORKER_HASH_SIZE = 1 << BUSY_WORKER_HASH_ORDER, |
77 | BUSY_WORKER_HASH_MASK = BUSY_WORKER_HASH_SIZE - 1, |
78 | |
79 | MAX_IDLE_WORKERS_RATIO = 4, /* 1/4 of busy can be idle */ |
80 | IDLE_WORKER_TIMEOUT = 300 * HZ, /* keep idle ones for 5 mins */ |
81 | |
82 | MAYDAY_INITIAL_TIMEOUT = HZ / 100 >= 2 ? HZ / 100 : 2, |
83 | /* call for help after 10ms |
84 | (min two ticks) */ |
85 | MAYDAY_INTERVAL = HZ / 10, /* and then every 100ms */ |
86 | CREATE_COOLDOWN = HZ, /* time to breath after fail */ |
87 | TRUSTEE_COOLDOWN = HZ / 10, /* for trustee draining */ |
88 | |
89 | /* |
90 | * Rescue workers are used only on emergencies and shared by |
91 | * all cpus. Give -20. |
92 | */ |
93 | RESCUER_NICE_LEVEL = -20, |
94 | }; |
95 | |
96 | /* |
97 | * Structure fields follow one of the following exclusion rules. |
98 | * |
99 | * I: Modifiable by initialization/destruction paths and read-only for |
100 | * everyone else. |
101 | * |
102 | * P: Preemption protected. Disabling preemption is enough and should |
103 | * only be modified and accessed from the local cpu. |
104 | * |
105 | * L: gcwq->lock protected. Access with gcwq->lock held. |
106 | * |
107 | * X: During normal operation, modification requires gcwq->lock and |
108 | * should be done only from local cpu. Either disabling preemption |
109 | * on local cpu or grabbing gcwq->lock is enough for read access. |
110 | * If GCWQ_DISASSOCIATED is set, it's identical to L. |
111 | * |
112 | * F: wq->flush_mutex protected. |
113 | * |
114 | * W: workqueue_lock protected. |
115 | */ |
116 | |
117 | struct global_cwq; |
118 | |
119 | /* |
120 | * The poor guys doing the actual heavy lifting. All on-duty workers |
121 | * are either serving the manager role, on idle list or on busy hash. |
122 | */ |
123 | struct worker { |
124 | /* on idle list while idle, on busy hash table while busy */ |
125 | union { |
126 | struct list_head entry; /* L: while idle */ |
127 | struct hlist_node hentry; /* L: while busy */ |
128 | }; |
129 | |
130 | struct work_struct *current_work; /* L: work being processed */ |
131 | struct cpu_workqueue_struct *current_cwq; /* L: current_work's cwq */ |
132 | struct list_head scheduled; /* L: scheduled works */ |
133 | struct task_struct *task; /* I: worker task */ |
134 | struct global_cwq *gcwq; /* I: the associated gcwq */ |
135 | /* 64 bytes boundary on 64bit, 32 on 32bit */ |
136 | unsigned long last_active; /* L: last active timestamp */ |
137 | unsigned int flags; /* X: flags */ |
138 | int id; /* I: worker id */ |
139 | struct work_struct rebind_work; /* L: rebind worker to cpu */ |
140 | }; |
141 | |
142 | /* |
143 | * Global per-cpu workqueue. There's one and only one for each cpu |
144 | * and all works are queued and processed here regardless of their |
145 | * target workqueues. |
146 | */ |
147 | struct global_cwq { |
148 | spinlock_t lock; /* the gcwq lock */ |
149 | struct list_head worklist; /* L: list of pending works */ |
150 | unsigned int cpu; /* I: the associated cpu */ |
151 | unsigned int flags; /* L: GCWQ_* flags */ |
152 | |
153 | int nr_workers; /* L: total number of workers */ |
154 | int nr_idle; /* L: currently idle ones */ |
155 | |
156 | /* workers are chained either in the idle_list or busy_hash */ |
157 | struct list_head idle_list; /* X: list of idle workers */ |
158 | struct hlist_head busy_hash[BUSY_WORKER_HASH_SIZE]; |
159 | /* L: hash of busy workers */ |
160 | |
161 | struct timer_list idle_timer; /* L: worker idle timeout */ |
162 | struct timer_list mayday_timer; /* L: SOS timer for dworkers */ |
163 | |
164 | struct ida worker_ida; /* L: for worker IDs */ |
165 | |
166 | struct task_struct *trustee; /* L: for gcwq shutdown */ |
167 | unsigned int trustee_state; /* L: trustee state */ |
168 | wait_queue_head_t trustee_wait; /* trustee wait */ |
169 | struct worker *first_idle; /* L: first idle worker */ |
170 | } ____cacheline_aligned_in_smp; |
171 | |
172 | /* |
173 | * The per-CPU workqueue. The lower WORK_STRUCT_FLAG_BITS of |
174 | * work_struct->data are used for flags and thus cwqs need to be |
175 | * aligned at two's power of the number of flag bits. |
176 | */ |
177 | struct cpu_workqueue_struct { |
178 | struct global_cwq *gcwq; /* I: the associated gcwq */ |
179 | struct workqueue_struct *wq; /* I: the owning workqueue */ |
180 | int work_color; /* L: current color */ |
181 | int flush_color; /* L: flushing color */ |
182 | int nr_in_flight[WORK_NR_COLORS]; |
183 | /* L: nr of in_flight works */ |
184 | int nr_active; /* L: nr of active works */ |
185 | int max_active; /* L: max active works */ |
186 | struct list_head delayed_works; /* L: delayed works */ |
187 | }; |
188 | |
189 | /* |
190 | * Structure used to wait for workqueue flush. |
191 | */ |
192 | struct wq_flusher { |
193 | struct list_head list; /* F: list of flushers */ |
194 | int flush_color; /* F: flush color waiting for */ |
195 | struct completion done; /* flush completion */ |
196 | }; |
197 | |
198 | /* |
199 | * All cpumasks are assumed to be always set on UP and thus can't be |
200 | * used to determine whether there's something to be done. |
201 | */ |
202 | #ifdef CONFIG_SMP |
203 | typedef cpumask_var_t mayday_mask_t; |
204 | #define mayday_test_and_set_cpu(cpu, mask) \ |
205 | cpumask_test_and_set_cpu((cpu), (mask)) |
206 | #define mayday_clear_cpu(cpu, mask) cpumask_clear_cpu((cpu), (mask)) |
207 | #define for_each_mayday_cpu(cpu, mask) for_each_cpu((cpu), (mask)) |
208 | #define alloc_mayday_mask(maskp, gfp) zalloc_cpumask_var((maskp), (gfp)) |
209 | #define free_mayday_mask(mask) free_cpumask_var((mask)) |
210 | #else |
211 | typedef unsigned long mayday_mask_t; |
212 | #define mayday_test_and_set_cpu(cpu, mask) test_and_set_bit(0, &(mask)) |
213 | #define mayday_clear_cpu(cpu, mask) clear_bit(0, &(mask)) |
214 | #define for_each_mayday_cpu(cpu, mask) if ((cpu) = 0, (mask)) |
215 | #define alloc_mayday_mask(maskp, gfp) true |
216 | #define free_mayday_mask(mask) do { } while (0) |
217 | #endif |
218 | |
219 | /* |
220 | * The externally visible workqueue abstraction is an array of |
221 | * per-CPU workqueues: |
222 | */ |
223 | struct workqueue_struct { |
224 | unsigned int flags; /* I: WQ_* flags */ |
225 | union { |
226 | struct cpu_workqueue_struct __percpu *pcpu; |
227 | struct cpu_workqueue_struct *single; |
228 | unsigned long v; |
229 | } cpu_wq; /* I: cwq's */ |
230 | struct list_head list; /* W: list of all workqueues */ |
231 | |
232 | struct mutex flush_mutex; /* protects wq flushing */ |
233 | int work_color; /* F: current work color */ |
234 | int flush_color; /* F: current flush color */ |
235 | atomic_t nr_cwqs_to_flush; /* flush in progress */ |
236 | struct wq_flusher *first_flusher; /* F: first flusher */ |
237 | struct list_head flusher_queue; /* F: flush waiters */ |
238 | struct list_head flusher_overflow; /* F: flush overflow list */ |
239 | |
240 | mayday_mask_t mayday_mask; /* cpus requesting rescue */ |
241 | struct worker *rescuer; /* I: rescue worker */ |
242 | |
243 | int saved_max_active; /* W: saved cwq max_active */ |
244 | const char *name; /* I: workqueue name */ |
245 | #ifdef CONFIG_LOCKDEP |
246 | struct lockdep_map lockdep_map; |
247 | #endif |
248 | }; |
249 | |
250 | struct workqueue_struct *system_wq __read_mostly; |
251 | struct workqueue_struct *system_long_wq __read_mostly; |
252 | struct workqueue_struct *system_nrt_wq __read_mostly; |
253 | struct workqueue_struct *system_unbound_wq __read_mostly; |
254 | EXPORT_SYMBOL_GPL(system_wq); |
255 | EXPORT_SYMBOL_GPL(system_long_wq); |
256 | EXPORT_SYMBOL_GPL(system_nrt_wq); |
257 | EXPORT_SYMBOL_GPL(system_unbound_wq); |
258 | |
259 | #define CREATE_TRACE_POINTS |
260 | #include <trace/events/workqueue.h> |
261 | |
262 | #define for_each_busy_worker(worker, i, pos, gcwq) \ |
263 | for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) \ |
264 | hlist_for_each_entry(worker, pos, &gcwq->busy_hash[i], hentry) |
265 | |
266 | static inline int __next_gcwq_cpu(int cpu, const struct cpumask *mask, |
267 | unsigned int sw) |
268 | { |
269 | if (cpu < nr_cpu_ids) { |
270 | if (sw & 1) { |
271 | cpu = cpumask_next(cpu, mask); |
272 | if (cpu < nr_cpu_ids) |
273 | return cpu; |
274 | } |
275 | if (sw & 2) |
276 | return WORK_CPU_UNBOUND; |
277 | } |
278 | return WORK_CPU_NONE; |
279 | } |
280 | |
281 | static inline int __next_wq_cpu(int cpu, const struct cpumask *mask, |
282 | struct workqueue_struct *wq) |
283 | { |
284 | return __next_gcwq_cpu(cpu, mask, !(wq->flags & WQ_UNBOUND) ? 1 : 2); |
285 | } |
286 | |
287 | /* |
288 | * CPU iterators |
289 | * |
290 | * An extra gcwq is defined for an invalid cpu number |
291 | * (WORK_CPU_UNBOUND) to host workqueues which are not bound to any |
292 | * specific CPU. The following iterators are similar to |
293 | * for_each_*_cpu() iterators but also considers the unbound gcwq. |
294 | * |
295 | * for_each_gcwq_cpu() : possible CPUs + WORK_CPU_UNBOUND |
296 | * for_each_online_gcwq_cpu() : online CPUs + WORK_CPU_UNBOUND |
297 | * for_each_cwq_cpu() : possible CPUs for bound workqueues, |
298 | * WORK_CPU_UNBOUND for unbound workqueues |
299 | */ |
300 | #define for_each_gcwq_cpu(cpu) \ |
301 | for ((cpu) = __next_gcwq_cpu(-1, cpu_possible_mask, 3); \ |
302 | (cpu) < WORK_CPU_NONE; \ |
303 | (cpu) = __next_gcwq_cpu((cpu), cpu_possible_mask, 3)) |
304 | |
305 | #define for_each_online_gcwq_cpu(cpu) \ |
306 | for ((cpu) = __next_gcwq_cpu(-1, cpu_online_mask, 3); \ |
307 | (cpu) < WORK_CPU_NONE; \ |
308 | (cpu) = __next_gcwq_cpu((cpu), cpu_online_mask, 3)) |
309 | |
310 | #define for_each_cwq_cpu(cpu, wq) \ |
311 | for ((cpu) = __next_wq_cpu(-1, cpu_possible_mask, (wq)); \ |
312 | (cpu) < WORK_CPU_NONE; \ |
313 | (cpu) = __next_wq_cpu((cpu), cpu_possible_mask, (wq))) |
314 | |
315 | #ifdef CONFIG_DEBUG_OBJECTS_WORK |
316 | |
317 | static struct debug_obj_descr work_debug_descr; |
318 | |
319 | /* |
320 | * fixup_init is called when: |
321 | * - an active object is initialized |
322 | */ |
323 | static int work_fixup_init(void *addr, enum debug_obj_state state) |
324 | { |
325 | struct work_struct *work = addr; |
326 | |
327 | switch (state) { |
328 | case ODEBUG_STATE_ACTIVE: |
329 | cancel_work_sync(work); |
330 | debug_object_init(work, &work_debug_descr); |
331 | return 1; |
332 | default: |
333 | return 0; |
334 | } |
335 | } |
336 | |
337 | /* |
338 | * fixup_activate is called when: |
339 | * - an active object is activated |
340 | * - an unknown object is activated (might be a statically initialized object) |
341 | */ |
342 | static int work_fixup_activate(void *addr, enum debug_obj_state state) |
343 | { |
344 | struct work_struct *work = addr; |
345 | |
346 | switch (state) { |
347 | |
348 | case ODEBUG_STATE_NOTAVAILABLE: |
349 | /* |
350 | * This is not really a fixup. The work struct was |
351 | * statically initialized. We just make sure that it |
352 | * is tracked in the object tracker. |
353 | */ |
354 | if (test_bit(WORK_STRUCT_STATIC_BIT, work_data_bits(work))) { |
355 | debug_object_init(work, &work_debug_descr); |
356 | debug_object_activate(work, &work_debug_descr); |
357 | return 0; |
358 | } |
359 | WARN_ON_ONCE(1); |
360 | return 0; |
361 | |
362 | case ODEBUG_STATE_ACTIVE: |
363 | WARN_ON(1); |
364 | |
365 | default: |
366 | return 0; |
367 | } |
368 | } |
369 | |
370 | /* |
371 | * fixup_free is called when: |
372 | * - an active object is freed |
373 | */ |
374 | static int work_fixup_free(void *addr, enum debug_obj_state state) |
375 | { |
376 | struct work_struct *work = addr; |
377 | |
378 | switch (state) { |
379 | case ODEBUG_STATE_ACTIVE: |
380 | cancel_work_sync(work); |
381 | debug_object_free(work, &work_debug_descr); |
382 | return 1; |
383 | default: |
384 | return 0; |
385 | } |
386 | } |
387 | |
388 | static struct debug_obj_descr work_debug_descr = { |
389 | .name = "work_struct", |
390 | .fixup_init = work_fixup_init, |
391 | .fixup_activate = work_fixup_activate, |
392 | .fixup_free = work_fixup_free, |
393 | }; |
394 | |
395 | static inline void debug_work_activate(struct work_struct *work) |
396 | { |
397 | debug_object_activate(work, &work_debug_descr); |
398 | } |
399 | |
400 | static inline void debug_work_deactivate(struct work_struct *work) |
401 | { |
402 | debug_object_deactivate(work, &work_debug_descr); |
403 | } |
404 | |
405 | void __init_work(struct work_struct *work, int onstack) |
406 | { |
407 | if (onstack) |
408 | debug_object_init_on_stack(work, &work_debug_descr); |
409 | else |
410 | debug_object_init(work, &work_debug_descr); |
411 | } |
412 | EXPORT_SYMBOL_GPL(__init_work); |
413 | |
414 | void destroy_work_on_stack(struct work_struct *work) |
415 | { |
416 | debug_object_free(work, &work_debug_descr); |
417 | } |
418 | EXPORT_SYMBOL_GPL(destroy_work_on_stack); |
419 | |
420 | #else |
421 | static inline void debug_work_activate(struct work_struct *work) { } |
422 | static inline void debug_work_deactivate(struct work_struct *work) { } |
423 | #endif |
424 | |
425 | /* Serializes the accesses to the list of workqueues. */ |
426 | static DEFINE_SPINLOCK(workqueue_lock); |
427 | static LIST_HEAD(workqueues); |
428 | static bool workqueue_freezing; /* W: have wqs started freezing? */ |
429 | |
430 | /* |
431 | * The almighty global cpu workqueues. nr_running is the only field |
432 | * which is expected to be used frequently by other cpus via |
433 | * try_to_wake_up(). Put it in a separate cacheline. |
434 | */ |
435 | static DEFINE_PER_CPU(struct global_cwq, global_cwq); |
436 | static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, gcwq_nr_running); |
437 | |
438 | /* |
439 | * Global cpu workqueue and nr_running counter for unbound gcwq. The |
440 | * gcwq is always online, has GCWQ_DISASSOCIATED set, and all its |
441 | * workers have WORKER_UNBOUND set. |
442 | */ |
443 | static struct global_cwq unbound_global_cwq; |
444 | static atomic_t unbound_gcwq_nr_running = ATOMIC_INIT(0); /* always 0 */ |
445 | |
446 | static int worker_thread(void *__worker); |
447 | |
448 | static struct global_cwq *get_gcwq(unsigned int cpu) |
449 | { |
450 | if (cpu != WORK_CPU_UNBOUND) |
451 | return &per_cpu(global_cwq, cpu); |
452 | else |
453 | return &unbound_global_cwq; |
454 | } |
455 | |
456 | static atomic_t *get_gcwq_nr_running(unsigned int cpu) |
457 | { |
458 | if (cpu != WORK_CPU_UNBOUND) |
459 | return &per_cpu(gcwq_nr_running, cpu); |
460 | else |
461 | return &unbound_gcwq_nr_running; |
462 | } |
463 | |
464 | static struct cpu_workqueue_struct *get_cwq(unsigned int cpu, |
465 | struct workqueue_struct *wq) |
466 | { |
467 | if (!(wq->flags & WQ_UNBOUND)) { |
468 | if (likely(cpu < nr_cpu_ids)) { |
469 | #ifdef CONFIG_SMP |
470 | return per_cpu_ptr(wq->cpu_wq.pcpu, cpu); |
471 | #else |
472 | return wq->cpu_wq.single; |
473 | #endif |
474 | } |
475 | } else if (likely(cpu == WORK_CPU_UNBOUND)) |
476 | return wq->cpu_wq.single; |
477 | return NULL; |
478 | } |
479 | |
480 | static unsigned int work_color_to_flags(int color) |
481 | { |
482 | return color << WORK_STRUCT_COLOR_SHIFT; |
483 | } |
484 | |
485 | static int get_work_color(struct work_struct *work) |
486 | { |
487 | return (*work_data_bits(work) >> WORK_STRUCT_COLOR_SHIFT) & |
488 | ((1 << WORK_STRUCT_COLOR_BITS) - 1); |
489 | } |
490 | |
491 | static int work_next_color(int color) |
492 | { |
493 | return (color + 1) % WORK_NR_COLORS; |
494 | } |
495 | |
496 | /* |
497 | * A work's data points to the cwq with WORK_STRUCT_CWQ set while the |
498 | * work is on queue. Once execution starts, WORK_STRUCT_CWQ is |
499 | * cleared and the work data contains the cpu number it was last on. |
500 | * |
501 | * set_work_{cwq|cpu}() and clear_work_data() can be used to set the |
502 | * cwq, cpu or clear work->data. These functions should only be |
503 | * called while the work is owned - ie. while the PENDING bit is set. |
504 | * |
505 | * get_work_[g]cwq() can be used to obtain the gcwq or cwq |
506 | * corresponding to a work. gcwq is available once the work has been |
507 | * queued anywhere after initialization. cwq is available only from |
508 | * queueing until execution starts. |
509 | */ |
510 | static inline void set_work_data(struct work_struct *work, unsigned long data, |
511 | unsigned long flags) |
512 | { |
513 | BUG_ON(!work_pending(work)); |
514 | atomic_long_set(&work->data, data | flags | work_static(work)); |
515 | } |
516 | |
517 | static void set_work_cwq(struct work_struct *work, |
518 | struct cpu_workqueue_struct *cwq, |
519 | unsigned long extra_flags) |
520 | { |
521 | set_work_data(work, (unsigned long)cwq, |
522 | WORK_STRUCT_PENDING | WORK_STRUCT_CWQ | extra_flags); |
523 | } |
524 | |
525 | static void set_work_cpu(struct work_struct *work, unsigned int cpu) |
526 | { |
527 | set_work_data(work, cpu << WORK_STRUCT_FLAG_BITS, WORK_STRUCT_PENDING); |
528 | } |
529 | |
530 | static void clear_work_data(struct work_struct *work) |
531 | { |
532 | set_work_data(work, WORK_STRUCT_NO_CPU, 0); |
533 | } |
534 | |
535 | static struct cpu_workqueue_struct *get_work_cwq(struct work_struct *work) |
536 | { |
537 | unsigned long data = atomic_long_read(&work->data); |
538 | |
539 | if (data & WORK_STRUCT_CWQ) |
540 | return (void *)(data & WORK_STRUCT_WQ_DATA_MASK); |
541 | else |
542 | return NULL; |
543 | } |
544 | |
545 | static struct global_cwq *get_work_gcwq(struct work_struct *work) |
546 | { |
547 | unsigned long data = atomic_long_read(&work->data); |
548 | unsigned int cpu; |
549 | |
550 | if (data & WORK_STRUCT_CWQ) |
551 | return ((struct cpu_workqueue_struct *) |
552 | (data & WORK_STRUCT_WQ_DATA_MASK))->gcwq; |
553 | |
554 | cpu = data >> WORK_STRUCT_FLAG_BITS; |
555 | if (cpu == WORK_CPU_NONE) |
556 | return NULL; |
557 | |
558 | BUG_ON(cpu >= nr_cpu_ids && cpu != WORK_CPU_UNBOUND); |
559 | return get_gcwq(cpu); |
560 | } |
561 | |
562 | /* |
563 | * Policy functions. These define the policies on how the global |
564 | * worker pool is managed. Unless noted otherwise, these functions |
565 | * assume that they're being called with gcwq->lock held. |
566 | */ |
567 | |
568 | static bool __need_more_worker(struct global_cwq *gcwq) |
569 | { |
570 | return !atomic_read(get_gcwq_nr_running(gcwq->cpu)) || |
571 | gcwq->flags & GCWQ_HIGHPRI_PENDING; |
572 | } |
573 | |
574 | /* |
575 | * Need to wake up a worker? Called from anything but currently |
576 | * running workers. |
577 | */ |
578 | static bool need_more_worker(struct global_cwq *gcwq) |
579 | { |
580 | return !list_empty(&gcwq->worklist) && __need_more_worker(gcwq); |
581 | } |
582 | |
583 | /* Can I start working? Called from busy but !running workers. */ |
584 | static bool may_start_working(struct global_cwq *gcwq) |
585 | { |
586 | return gcwq->nr_idle; |
587 | } |
588 | |
589 | /* Do I need to keep working? Called from currently running workers. */ |
590 | static bool keep_working(struct global_cwq *gcwq) |
591 | { |
592 | atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); |
593 | |
594 | return !list_empty(&gcwq->worklist) && |
595 | (atomic_read(nr_running) <= 1 || |
596 | gcwq->flags & GCWQ_HIGHPRI_PENDING); |
597 | } |
598 | |
599 | /* Do we need a new worker? Called from manager. */ |
600 | static bool need_to_create_worker(struct global_cwq *gcwq) |
601 | { |
602 | return need_more_worker(gcwq) && !may_start_working(gcwq); |
603 | } |
604 | |
605 | /* Do I need to be the manager? */ |
606 | static bool need_to_manage_workers(struct global_cwq *gcwq) |
607 | { |
608 | return need_to_create_worker(gcwq) || gcwq->flags & GCWQ_MANAGE_WORKERS; |
609 | } |
610 | |
611 | /* Do we have too many workers and should some go away? */ |
612 | static bool too_many_workers(struct global_cwq *gcwq) |
613 | { |
614 | bool managing = gcwq->flags & GCWQ_MANAGING_WORKERS; |
615 | int nr_idle = gcwq->nr_idle + managing; /* manager is considered idle */ |
616 | int nr_busy = gcwq->nr_workers - nr_idle; |
617 | |
618 | return nr_idle > 2 && (nr_idle - 2) * MAX_IDLE_WORKERS_RATIO >= nr_busy; |
619 | } |
620 | |
621 | /* |
622 | * Wake up functions. |
623 | */ |
624 | |
625 | /* Return the first worker. Safe with preemption disabled */ |
626 | static struct worker *first_worker(struct global_cwq *gcwq) |
627 | { |
628 | if (unlikely(list_empty(&gcwq->idle_list))) |
629 | return NULL; |
630 | |
631 | return list_first_entry(&gcwq->idle_list, struct worker, entry); |
632 | } |
633 | |
634 | /** |
635 | * wake_up_worker - wake up an idle worker |
636 | * @gcwq: gcwq to wake worker for |
637 | * |
638 | * Wake up the first idle worker of @gcwq. |
639 | * |
640 | * CONTEXT: |
641 | * spin_lock_irq(gcwq->lock). |
642 | */ |
643 | static void wake_up_worker(struct global_cwq *gcwq) |
644 | { |
645 | struct worker *worker = first_worker(gcwq); |
646 | |
647 | if (likely(worker)) |
648 | wake_up_process(worker->task); |
649 | } |
650 | |
651 | /** |
652 | * wq_worker_waking_up - a worker is waking up |
653 | * @task: task waking up |
654 | * @cpu: CPU @task is waking up to |
655 | * |
656 | * This function is called during try_to_wake_up() when a worker is |
657 | * being awoken. |
658 | * |
659 | * CONTEXT: |
660 | * spin_lock_irq(rq->lock) |
661 | */ |
662 | void wq_worker_waking_up(struct task_struct *task, unsigned int cpu) |
663 | { |
664 | struct worker *worker = kthread_data(task); |
665 | |
666 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
667 | atomic_inc(get_gcwq_nr_running(cpu)); |
668 | } |
669 | |
670 | /** |
671 | * wq_worker_sleeping - a worker is going to sleep |
672 | * @task: task going to sleep |
673 | * @cpu: CPU in question, must be the current CPU number |
674 | * |
675 | * This function is called during schedule() when a busy worker is |
676 | * going to sleep. Worker on the same cpu can be woken up by |
677 | * returning pointer to its task. |
678 | * |
679 | * CONTEXT: |
680 | * spin_lock_irq(rq->lock) |
681 | * |
682 | * RETURNS: |
683 | * Worker task on @cpu to wake up, %NULL if none. |
684 | */ |
685 | struct task_struct *wq_worker_sleeping(struct task_struct *task, |
686 | unsigned int cpu) |
687 | { |
688 | struct worker *worker = kthread_data(task), *to_wakeup = NULL; |
689 | struct global_cwq *gcwq = get_gcwq(cpu); |
690 | atomic_t *nr_running = get_gcwq_nr_running(cpu); |
691 | |
692 | if (worker->flags & WORKER_NOT_RUNNING) |
693 | return NULL; |
694 | |
695 | /* this can only happen on the local cpu */ |
696 | BUG_ON(cpu != raw_smp_processor_id()); |
697 | |
698 | /* |
699 | * The counterpart of the following dec_and_test, implied mb, |
700 | * worklist not empty test sequence is in insert_work(). |
701 | * Please read comment there. |
702 | * |
703 | * NOT_RUNNING is clear. This means that trustee is not in |
704 | * charge and we're running on the local cpu w/ rq lock held |
705 | * and preemption disabled, which in turn means that none else |
706 | * could be manipulating idle_list, so dereferencing idle_list |
707 | * without gcwq lock is safe. |
708 | */ |
709 | if (atomic_dec_and_test(nr_running) && !list_empty(&gcwq->worklist)) |
710 | to_wakeup = first_worker(gcwq); |
711 | return to_wakeup ? to_wakeup->task : NULL; |
712 | } |
713 | |
714 | /** |
715 | * worker_set_flags - set worker flags and adjust nr_running accordingly |
716 | * @worker: self |
717 | * @flags: flags to set |
718 | * @wakeup: wakeup an idle worker if necessary |
719 | * |
720 | * Set @flags in @worker->flags and adjust nr_running accordingly. If |
721 | * nr_running becomes zero and @wakeup is %true, an idle worker is |
722 | * woken up. |
723 | * |
724 | * CONTEXT: |
725 | * spin_lock_irq(gcwq->lock) |
726 | */ |
727 | static inline void worker_set_flags(struct worker *worker, unsigned int flags, |
728 | bool wakeup) |
729 | { |
730 | struct global_cwq *gcwq = worker->gcwq; |
731 | |
732 | WARN_ON_ONCE(worker->task != current); |
733 | |
734 | /* |
735 | * If transitioning into NOT_RUNNING, adjust nr_running and |
736 | * wake up an idle worker as necessary if requested by |
737 | * @wakeup. |
738 | */ |
739 | if ((flags & WORKER_NOT_RUNNING) && |
740 | !(worker->flags & WORKER_NOT_RUNNING)) { |
741 | atomic_t *nr_running = get_gcwq_nr_running(gcwq->cpu); |
742 | |
743 | if (wakeup) { |
744 | if (atomic_dec_and_test(nr_running) && |
745 | !list_empty(&gcwq->worklist)) |
746 | wake_up_worker(gcwq); |
747 | } else |
748 | atomic_dec(nr_running); |
749 | } |
750 | |
751 | worker->flags |= flags; |
752 | } |
753 | |
754 | /** |
755 | * worker_clr_flags - clear worker flags and adjust nr_running accordingly |
756 | * @worker: self |
757 | * @flags: flags to clear |
758 | * |
759 | * Clear @flags in @worker->flags and adjust nr_running accordingly. |
760 | * |
761 | * CONTEXT: |
762 | * spin_lock_irq(gcwq->lock) |
763 | */ |
764 | static inline void worker_clr_flags(struct worker *worker, unsigned int flags) |
765 | { |
766 | struct global_cwq *gcwq = worker->gcwq; |
767 | unsigned int oflags = worker->flags; |
768 | |
769 | WARN_ON_ONCE(worker->task != current); |
770 | |
771 | worker->flags &= ~flags; |
772 | |
773 | /* |
774 | * If transitioning out of NOT_RUNNING, increment nr_running. Note |
775 | * that the nested NOT_RUNNING is not a noop. NOT_RUNNING is mask |
776 | * of multiple flags, not a single flag. |
777 | */ |
778 | if ((flags & WORKER_NOT_RUNNING) && (oflags & WORKER_NOT_RUNNING)) |
779 | if (!(worker->flags & WORKER_NOT_RUNNING)) |
780 | atomic_inc(get_gcwq_nr_running(gcwq->cpu)); |
781 | } |
782 | |
783 | /** |
784 | * busy_worker_head - return the busy hash head for a work |
785 | * @gcwq: gcwq of interest |
786 | * @work: work to be hashed |
787 | * |
788 | * Return hash head of @gcwq for @work. |
789 | * |
790 | * CONTEXT: |
791 | * spin_lock_irq(gcwq->lock). |
792 | * |
793 | * RETURNS: |
794 | * Pointer to the hash head. |
795 | */ |
796 | static struct hlist_head *busy_worker_head(struct global_cwq *gcwq, |
797 | struct work_struct *work) |
798 | { |
799 | const int base_shift = ilog2(sizeof(struct work_struct)); |
800 | unsigned long v = (unsigned long)work; |
801 | |
802 | /* simple shift and fold hash, do we need something better? */ |
803 | v >>= base_shift; |
804 | v += v >> BUSY_WORKER_HASH_ORDER; |
805 | v &= BUSY_WORKER_HASH_MASK; |
806 | |
807 | return &gcwq->busy_hash[v]; |
808 | } |
809 | |
810 | /** |
811 | * __find_worker_executing_work - find worker which is executing a work |
812 | * @gcwq: gcwq of interest |
813 | * @bwh: hash head as returned by busy_worker_head() |
814 | * @work: work to find worker for |
815 | * |
816 | * Find a worker which is executing @work on @gcwq. @bwh should be |
817 | * the hash head obtained by calling busy_worker_head() with the same |
818 | * work. |
819 | * |
820 | * CONTEXT: |
821 | * spin_lock_irq(gcwq->lock). |
822 | * |
823 | * RETURNS: |
824 | * Pointer to worker which is executing @work if found, NULL |
825 | * otherwise. |
826 | */ |
827 | static struct worker *__find_worker_executing_work(struct global_cwq *gcwq, |
828 | struct hlist_head *bwh, |
829 | struct work_struct *work) |
830 | { |
831 | struct worker *worker; |
832 | struct hlist_node *tmp; |
833 | |
834 | hlist_for_each_entry(worker, tmp, bwh, hentry) |
835 | if (worker->current_work == work) |
836 | return worker; |
837 | return NULL; |
838 | } |
839 | |
840 | /** |
841 | * find_worker_executing_work - find worker which is executing a work |
842 | * @gcwq: gcwq of interest |
843 | * @work: work to find worker for |
844 | * |
845 | * Find a worker which is executing @work on @gcwq. This function is |
846 | * identical to __find_worker_executing_work() except that this |
847 | * function calculates @bwh itself. |
848 | * |
849 | * CONTEXT: |
850 | * spin_lock_irq(gcwq->lock). |
851 | * |
852 | * RETURNS: |
853 | * Pointer to worker which is executing @work if found, NULL |
854 | * otherwise. |
855 | */ |
856 | static struct worker *find_worker_executing_work(struct global_cwq *gcwq, |
857 | struct work_struct *work) |
858 | { |
859 | return __find_worker_executing_work(gcwq, busy_worker_head(gcwq, work), |
860 | work); |
861 | } |
862 | |
863 | /** |
864 | * gcwq_determine_ins_pos - find insertion position |
865 | * @gcwq: gcwq of interest |
866 | * @cwq: cwq a work is being queued for |
867 | * |
868 | * A work for @cwq is about to be queued on @gcwq, determine insertion |
869 | * position for the work. If @cwq is for HIGHPRI wq, the work is |
870 | * queued at the head of the queue but in FIFO order with respect to |
871 | * other HIGHPRI works; otherwise, at the end of the queue. This |
872 | * function also sets GCWQ_HIGHPRI_PENDING flag to hint @gcwq that |
873 | * there are HIGHPRI works pending. |
874 | * |
875 | * CONTEXT: |
876 | * spin_lock_irq(gcwq->lock). |
877 | * |
878 | * RETURNS: |
879 | * Pointer to inserstion position. |
880 | */ |
881 | static inline struct list_head *gcwq_determine_ins_pos(struct global_cwq *gcwq, |
882 | struct cpu_workqueue_struct *cwq) |
883 | { |
884 | struct work_struct *twork; |
885 | |
886 | if (likely(!(cwq->wq->flags & WQ_HIGHPRI))) |
887 | return &gcwq->worklist; |
888 | |
889 | list_for_each_entry(twork, &gcwq->worklist, entry) { |
890 | struct cpu_workqueue_struct *tcwq = get_work_cwq(twork); |
891 | |
892 | if (!(tcwq->wq->flags & WQ_HIGHPRI)) |
893 | break; |
894 | } |
895 | |
896 | gcwq->flags |= GCWQ_HIGHPRI_PENDING; |
897 | return &twork->entry; |
898 | } |
899 | |
900 | /** |
901 | * insert_work - insert a work into gcwq |
902 | * @cwq: cwq @work belongs to |
903 | * @work: work to insert |
904 | * @head: insertion point |
905 | * @extra_flags: extra WORK_STRUCT_* flags to set |
906 | * |
907 | * Insert @work which belongs to @cwq into @gcwq after @head. |
908 | * @extra_flags is or'd to work_struct flags. |
909 | * |
910 | * CONTEXT: |
911 | * spin_lock_irq(gcwq->lock). |
912 | */ |
913 | static void insert_work(struct cpu_workqueue_struct *cwq, |
914 | struct work_struct *work, struct list_head *head, |
915 | unsigned int extra_flags) |
916 | { |
917 | struct global_cwq *gcwq = cwq->gcwq; |
918 | |
919 | /* we own @work, set data and link */ |
920 | set_work_cwq(work, cwq, extra_flags); |
921 | |
922 | /* |
923 | * Ensure that we get the right work->data if we see the |
924 | * result of list_add() below, see try_to_grab_pending(). |
925 | */ |
926 | smp_wmb(); |
927 | |
928 | list_add_tail(&work->entry, head); |
929 | |
930 | /* |
931 | * Ensure either worker_sched_deactivated() sees the above |
932 | * list_add_tail() or we see zero nr_running to avoid workers |
933 | * lying around lazily while there are works to be processed. |
934 | */ |
935 | smp_mb(); |
936 | |
937 | if (__need_more_worker(gcwq)) |
938 | wake_up_worker(gcwq); |
939 | } |
940 | |
941 | /* |
942 | * Test whether @work is being queued from another work executing on the |
943 | * same workqueue. This is rather expensive and should only be used from |
944 | * cold paths. |
945 | */ |
946 | static bool is_chained_work(struct workqueue_struct *wq) |
947 | { |
948 | unsigned long flags; |
949 | unsigned int cpu; |
950 | |
951 | for_each_gcwq_cpu(cpu) { |
952 | struct global_cwq *gcwq = get_gcwq(cpu); |
953 | struct worker *worker; |
954 | struct hlist_node *pos; |
955 | int i; |
956 | |
957 | spin_lock_irqsave(&gcwq->lock, flags); |
958 | for_each_busy_worker(worker, i, pos, gcwq) { |
959 | if (worker->task != current) |
960 | continue; |
961 | spin_unlock_irqrestore(&gcwq->lock, flags); |
962 | /* |
963 | * I'm @worker, no locking necessary. See if @work |
964 | * is headed to the same workqueue. |
965 | */ |
966 | return worker->current_cwq->wq == wq; |
967 | } |
968 | spin_unlock_irqrestore(&gcwq->lock, flags); |
969 | } |
970 | return false; |
971 | } |
972 | |
973 | static void __queue_work(unsigned int cpu, struct workqueue_struct *wq, |
974 | struct work_struct *work) |
975 | { |
976 | struct global_cwq *gcwq; |
977 | struct cpu_workqueue_struct *cwq; |
978 | struct list_head *worklist; |
979 | unsigned int work_flags; |
980 | unsigned long flags; |
981 | |
982 | debug_work_activate(work); |
983 | |
984 | /* if dying, only works from the same workqueue are allowed */ |
985 | if (unlikely(wq->flags & WQ_DYING) && |
986 | WARN_ON_ONCE(!is_chained_work(wq))) |
987 | return; |
988 | |
989 | /* determine gcwq to use */ |
990 | if (!(wq->flags & WQ_UNBOUND)) { |
991 | struct global_cwq *last_gcwq; |
992 | |
993 | if (unlikely(cpu == WORK_CPU_UNBOUND)) |
994 | cpu = raw_smp_processor_id(); |
995 | |
996 | /* |
997 | * It's multi cpu. If @wq is non-reentrant and @work |
998 | * was previously on a different cpu, it might still |
999 | * be running there, in which case the work needs to |
1000 | * be queued on that cpu to guarantee non-reentrance. |
1001 | */ |
1002 | gcwq = get_gcwq(cpu); |
1003 | if (wq->flags & WQ_NON_REENTRANT && |
1004 | (last_gcwq = get_work_gcwq(work)) && last_gcwq != gcwq) { |
1005 | struct worker *worker; |
1006 | |
1007 | spin_lock_irqsave(&last_gcwq->lock, flags); |
1008 | |
1009 | worker = find_worker_executing_work(last_gcwq, work); |
1010 | |
1011 | if (worker && worker->current_cwq->wq == wq) |
1012 | gcwq = last_gcwq; |
1013 | else { |
1014 | /* meh... not running there, queue here */ |
1015 | spin_unlock_irqrestore(&last_gcwq->lock, flags); |
1016 | spin_lock_irqsave(&gcwq->lock, flags); |
1017 | } |
1018 | } else |
1019 | spin_lock_irqsave(&gcwq->lock, flags); |
1020 | } else { |
1021 | gcwq = get_gcwq(WORK_CPU_UNBOUND); |
1022 | spin_lock_irqsave(&gcwq->lock, flags); |
1023 | } |
1024 | |
1025 | /* gcwq determined, get cwq and queue */ |
1026 | cwq = get_cwq(gcwq->cpu, wq); |
1027 | trace_workqueue_queue_work(cpu, cwq, work); |
1028 | |
1029 | BUG_ON(!list_empty(&work->entry)); |
1030 | |
1031 | cwq->nr_in_flight[cwq->work_color]++; |
1032 | work_flags = work_color_to_flags(cwq->work_color); |
1033 | |
1034 | if (likely(cwq->nr_active < cwq->max_active)) { |
1035 | trace_workqueue_activate_work(work); |
1036 | cwq->nr_active++; |
1037 | worklist = gcwq_determine_ins_pos(gcwq, cwq); |
1038 | } else { |
1039 | work_flags |= WORK_STRUCT_DELAYED; |
1040 | worklist = &cwq->delayed_works; |
1041 | } |
1042 | |
1043 | insert_work(cwq, work, worklist, work_flags); |
1044 | |
1045 | spin_unlock_irqrestore(&gcwq->lock, flags); |
1046 | } |
1047 | |
1048 | /** |
1049 | * queue_work - queue work on a workqueue |
1050 | * @wq: workqueue to use |
1051 | * @work: work to queue |
1052 | * |
1053 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
1054 | * |
1055 | * We queue the work to the CPU on which it was submitted, but if the CPU dies |
1056 | * it can be processed by another CPU. |
1057 | */ |
1058 | int queue_work(struct workqueue_struct *wq, struct work_struct *work) |
1059 | { |
1060 | int ret; |
1061 | |
1062 | ret = queue_work_on(get_cpu(), wq, work); |
1063 | put_cpu(); |
1064 | |
1065 | return ret; |
1066 | } |
1067 | EXPORT_SYMBOL_GPL(queue_work); |
1068 | |
1069 | /** |
1070 | * queue_work_on - queue work on specific cpu |
1071 | * @cpu: CPU number to execute work on |
1072 | * @wq: workqueue to use |
1073 | * @work: work to queue |
1074 | * |
1075 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
1076 | * |
1077 | * We queue the work to a specific CPU, the caller must ensure it |
1078 | * can't go away. |
1079 | */ |
1080 | int |
1081 | queue_work_on(int cpu, struct workqueue_struct *wq, struct work_struct *work) |
1082 | { |
1083 | int ret = 0; |
1084 | |
1085 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
1086 | __queue_work(cpu, wq, work); |
1087 | ret = 1; |
1088 | } |
1089 | return ret; |
1090 | } |
1091 | EXPORT_SYMBOL_GPL(queue_work_on); |
1092 | |
1093 | static void delayed_work_timer_fn(unsigned long __data) |
1094 | { |
1095 | struct delayed_work *dwork = (struct delayed_work *)__data; |
1096 | struct cpu_workqueue_struct *cwq = get_work_cwq(&dwork->work); |
1097 | |
1098 | __queue_work(smp_processor_id(), cwq->wq, &dwork->work); |
1099 | } |
1100 | |
1101 | /** |
1102 | * queue_delayed_work - queue work on a workqueue after delay |
1103 | * @wq: workqueue to use |
1104 | * @dwork: delayable work to queue |
1105 | * @delay: number of jiffies to wait before queueing |
1106 | * |
1107 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
1108 | */ |
1109 | int queue_delayed_work(struct workqueue_struct *wq, |
1110 | struct delayed_work *dwork, unsigned long delay) |
1111 | { |
1112 | if (delay == 0) |
1113 | return queue_work(wq, &dwork->work); |
1114 | |
1115 | return queue_delayed_work_on(-1, wq, dwork, delay); |
1116 | } |
1117 | EXPORT_SYMBOL_GPL(queue_delayed_work); |
1118 | |
1119 | /** |
1120 | * queue_delayed_work_on - queue work on specific CPU after delay |
1121 | * @cpu: CPU number to execute work on |
1122 | * @wq: workqueue to use |
1123 | * @dwork: work to queue |
1124 | * @delay: number of jiffies to wait before queueing |
1125 | * |
1126 | * Returns 0 if @work was already on a queue, non-zero otherwise. |
1127 | */ |
1128 | int queue_delayed_work_on(int cpu, struct workqueue_struct *wq, |
1129 | struct delayed_work *dwork, unsigned long delay) |
1130 | { |
1131 | int ret = 0; |
1132 | struct timer_list *timer = &dwork->timer; |
1133 | struct work_struct *work = &dwork->work; |
1134 | |
1135 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) { |
1136 | unsigned int lcpu; |
1137 | |
1138 | BUG_ON(timer_pending(timer)); |
1139 | BUG_ON(!list_empty(&work->entry)); |
1140 | |
1141 | timer_stats_timer_set_start_info(&dwork->timer); |
1142 | |
1143 | /* |
1144 | * This stores cwq for the moment, for the timer_fn. |
1145 | * Note that the work's gcwq is preserved to allow |
1146 | * reentrance detection for delayed works. |
1147 | */ |
1148 | if (!(wq->flags & WQ_UNBOUND)) { |
1149 | struct global_cwq *gcwq = get_work_gcwq(work); |
1150 | |
1151 | if (gcwq && gcwq->cpu != WORK_CPU_UNBOUND) |
1152 | lcpu = gcwq->cpu; |
1153 | else |
1154 | lcpu = raw_smp_processor_id(); |
1155 | } else |
1156 | lcpu = WORK_CPU_UNBOUND; |
1157 | |
1158 | set_work_cwq(work, get_cwq(lcpu, wq), 0); |
1159 | |
1160 | timer->expires = jiffies + delay; |
1161 | timer->data = (unsigned long)dwork; |
1162 | timer->function = delayed_work_timer_fn; |
1163 | |
1164 | if (unlikely(cpu >= 0)) |
1165 | add_timer_on(timer, cpu); |
1166 | else |
1167 | add_timer(timer); |
1168 | ret = 1; |
1169 | } |
1170 | return ret; |
1171 | } |
1172 | EXPORT_SYMBOL_GPL(queue_delayed_work_on); |
1173 | |
1174 | /** |
1175 | * worker_enter_idle - enter idle state |
1176 | * @worker: worker which is entering idle state |
1177 | * |
1178 | * @worker is entering idle state. Update stats and idle timer if |
1179 | * necessary. |
1180 | * |
1181 | * LOCKING: |
1182 | * spin_lock_irq(gcwq->lock). |
1183 | */ |
1184 | static void worker_enter_idle(struct worker *worker) |
1185 | { |
1186 | struct global_cwq *gcwq = worker->gcwq; |
1187 | |
1188 | BUG_ON(worker->flags & WORKER_IDLE); |
1189 | BUG_ON(!list_empty(&worker->entry) && |
1190 | (worker->hentry.next || worker->hentry.pprev)); |
1191 | |
1192 | /* can't use worker_set_flags(), also called from start_worker() */ |
1193 | worker->flags |= WORKER_IDLE; |
1194 | gcwq->nr_idle++; |
1195 | worker->last_active = jiffies; |
1196 | |
1197 | /* idle_list is LIFO */ |
1198 | list_add(&worker->entry, &gcwq->idle_list); |
1199 | |
1200 | if (likely(!(worker->flags & WORKER_ROGUE))) { |
1201 | if (too_many_workers(gcwq) && !timer_pending(&gcwq->idle_timer)) |
1202 | mod_timer(&gcwq->idle_timer, |
1203 | jiffies + IDLE_WORKER_TIMEOUT); |
1204 | } else |
1205 | wake_up_all(&gcwq->trustee_wait); |
1206 | |
1207 | /* sanity check nr_running */ |
1208 | WARN_ON_ONCE(gcwq->nr_workers == gcwq->nr_idle && |
1209 | atomic_read(get_gcwq_nr_running(gcwq->cpu))); |
1210 | } |
1211 | |
1212 | /** |
1213 | * worker_leave_idle - leave idle state |
1214 | * @worker: worker which is leaving idle state |
1215 | * |
1216 | * @worker is leaving idle state. Update stats. |
1217 | * |
1218 | * LOCKING: |
1219 | * spin_lock_irq(gcwq->lock). |
1220 | */ |
1221 | static void worker_leave_idle(struct worker *worker) |
1222 | { |
1223 | struct global_cwq *gcwq = worker->gcwq; |
1224 | |
1225 | BUG_ON(!(worker->flags & WORKER_IDLE)); |
1226 | worker_clr_flags(worker, WORKER_IDLE); |
1227 | gcwq->nr_idle--; |
1228 | list_del_init(&worker->entry); |
1229 | } |
1230 | |
1231 | /** |
1232 | * worker_maybe_bind_and_lock - bind worker to its cpu if possible and lock gcwq |
1233 | * @worker: self |
1234 | * |
1235 | * Works which are scheduled while the cpu is online must at least be |
1236 | * scheduled to a worker which is bound to the cpu so that if they are |
1237 | * flushed from cpu callbacks while cpu is going down, they are |
1238 | * guaranteed to execute on the cpu. |
1239 | * |
1240 | * This function is to be used by rogue workers and rescuers to bind |
1241 | * themselves to the target cpu and may race with cpu going down or |
1242 | * coming online. kthread_bind() can't be used because it may put the |
1243 | * worker to already dead cpu and set_cpus_allowed_ptr() can't be used |
1244 | * verbatim as it's best effort and blocking and gcwq may be |
1245 | * [dis]associated in the meantime. |
1246 | * |
1247 | * This function tries set_cpus_allowed() and locks gcwq and verifies |
1248 | * the binding against GCWQ_DISASSOCIATED which is set during |
1249 | * CPU_DYING and cleared during CPU_ONLINE, so if the worker enters |
1250 | * idle state or fetches works without dropping lock, it can guarantee |
1251 | * the scheduling requirement described in the first paragraph. |
1252 | * |
1253 | * CONTEXT: |
1254 | * Might sleep. Called without any lock but returns with gcwq->lock |
1255 | * held. |
1256 | * |
1257 | * RETURNS: |
1258 | * %true if the associated gcwq is online (@worker is successfully |
1259 | * bound), %false if offline. |
1260 | */ |
1261 | static bool worker_maybe_bind_and_lock(struct worker *worker) |
1262 | __acquires(&gcwq->lock) |
1263 | { |
1264 | struct global_cwq *gcwq = worker->gcwq; |
1265 | struct task_struct *task = worker->task; |
1266 | |
1267 | while (true) { |
1268 | /* |
1269 | * The following call may fail, succeed or succeed |
1270 | * without actually migrating the task to the cpu if |
1271 | * it races with cpu hotunplug operation. Verify |
1272 | * against GCWQ_DISASSOCIATED. |
1273 | */ |
1274 | if (!(gcwq->flags & GCWQ_DISASSOCIATED)) |
1275 | set_cpus_allowed_ptr(task, get_cpu_mask(gcwq->cpu)); |
1276 | |
1277 | spin_lock_irq(&gcwq->lock); |
1278 | if (gcwq->flags & GCWQ_DISASSOCIATED) |
1279 | return false; |
1280 | if (task_cpu(task) == gcwq->cpu && |
1281 | cpumask_equal(¤t->cpus_allowed, |
1282 | get_cpu_mask(gcwq->cpu))) |
1283 | return true; |
1284 | spin_unlock_irq(&gcwq->lock); |
1285 | |
1286 | /* CPU has come up inbetween, retry migration */ |
1287 | cpu_relax(); |
1288 | } |
1289 | } |
1290 | |
1291 | /* |
1292 | * Function for worker->rebind_work used to rebind rogue busy workers |
1293 | * to the associated cpu which is coming back online. This is |
1294 | * scheduled by cpu up but can race with other cpu hotplug operations |
1295 | * and may be executed twice without intervening cpu down. |
1296 | */ |
1297 | static void worker_rebind_fn(struct work_struct *work) |
1298 | { |
1299 | struct worker *worker = container_of(work, struct worker, rebind_work); |
1300 | struct global_cwq *gcwq = worker->gcwq; |
1301 | |
1302 | if (worker_maybe_bind_and_lock(worker)) |
1303 | worker_clr_flags(worker, WORKER_REBIND); |
1304 | |
1305 | spin_unlock_irq(&gcwq->lock); |
1306 | } |
1307 | |
1308 | static struct worker *alloc_worker(void) |
1309 | { |
1310 | struct worker *worker; |
1311 | |
1312 | worker = kzalloc(sizeof(*worker), GFP_KERNEL); |
1313 | if (worker) { |
1314 | INIT_LIST_HEAD(&worker->entry); |
1315 | INIT_LIST_HEAD(&worker->scheduled); |
1316 | INIT_WORK(&worker->rebind_work, worker_rebind_fn); |
1317 | /* on creation a worker is in !idle && prep state */ |
1318 | worker->flags = WORKER_PREP; |
1319 | } |
1320 | return worker; |
1321 | } |
1322 | |
1323 | /** |
1324 | * create_worker - create a new workqueue worker |
1325 | * @gcwq: gcwq the new worker will belong to |
1326 | * @bind: whether to set affinity to @cpu or not |
1327 | * |
1328 | * Create a new worker which is bound to @gcwq. The returned worker |
1329 | * can be started by calling start_worker() or destroyed using |
1330 | * destroy_worker(). |
1331 | * |
1332 | * CONTEXT: |
1333 | * Might sleep. Does GFP_KERNEL allocations. |
1334 | * |
1335 | * RETURNS: |
1336 | * Pointer to the newly created worker. |
1337 | */ |
1338 | static struct worker *create_worker(struct global_cwq *gcwq, bool bind) |
1339 | { |
1340 | bool on_unbound_cpu = gcwq->cpu == WORK_CPU_UNBOUND; |
1341 | struct worker *worker = NULL; |
1342 | int id = -1; |
1343 | |
1344 | spin_lock_irq(&gcwq->lock); |
1345 | while (ida_get_new(&gcwq->worker_ida, &id)) { |
1346 | spin_unlock_irq(&gcwq->lock); |
1347 | if (!ida_pre_get(&gcwq->worker_ida, GFP_KERNEL)) |
1348 | goto fail; |
1349 | spin_lock_irq(&gcwq->lock); |
1350 | } |
1351 | spin_unlock_irq(&gcwq->lock); |
1352 | |
1353 | worker = alloc_worker(); |
1354 | if (!worker) |
1355 | goto fail; |
1356 | |
1357 | worker->gcwq = gcwq; |
1358 | worker->id = id; |
1359 | |
1360 | if (!on_unbound_cpu) |
1361 | worker->task = kthread_create(worker_thread, worker, |
1362 | "kworker/%u:%d", gcwq->cpu, id); |
1363 | else |
1364 | worker->task = kthread_create(worker_thread, worker, |
1365 | "kworker/u:%d", id); |
1366 | if (IS_ERR(worker->task)) |
1367 | goto fail; |
1368 | |
1369 | /* |
1370 | * A rogue worker will become a regular one if CPU comes |
1371 | * online later on. Make sure every worker has |
1372 | * PF_THREAD_BOUND set. |
1373 | */ |
1374 | if (bind && !on_unbound_cpu) |
1375 | kthread_bind(worker->task, gcwq->cpu); |
1376 | else { |
1377 | worker->task->flags |= PF_THREAD_BOUND; |
1378 | if (on_unbound_cpu) |
1379 | worker->flags |= WORKER_UNBOUND; |
1380 | } |
1381 | |
1382 | return worker; |
1383 | fail: |
1384 | if (id >= 0) { |
1385 | spin_lock_irq(&gcwq->lock); |
1386 | ida_remove(&gcwq->worker_ida, id); |
1387 | spin_unlock_irq(&gcwq->lock); |
1388 | } |
1389 | kfree(worker); |
1390 | return NULL; |
1391 | } |
1392 | |
1393 | /** |
1394 | * start_worker - start a newly created worker |
1395 | * @worker: worker to start |
1396 | * |
1397 | * Make the gcwq aware of @worker and start it. |
1398 | * |
1399 | * CONTEXT: |
1400 | * spin_lock_irq(gcwq->lock). |
1401 | */ |
1402 | static void start_worker(struct worker *worker) |
1403 | { |
1404 | worker->flags |= WORKER_STARTED; |
1405 | worker->gcwq->nr_workers++; |
1406 | worker_enter_idle(worker); |
1407 | wake_up_process(worker->task); |
1408 | } |
1409 | |
1410 | /** |
1411 | * destroy_worker - destroy a workqueue worker |
1412 | * @worker: worker to be destroyed |
1413 | * |
1414 | * Destroy @worker and adjust @gcwq stats accordingly. |
1415 | * |
1416 | * CONTEXT: |
1417 | * spin_lock_irq(gcwq->lock) which is released and regrabbed. |
1418 | */ |
1419 | static void destroy_worker(struct worker *worker) |
1420 | { |
1421 | struct global_cwq *gcwq = worker->gcwq; |
1422 | int id = worker->id; |
1423 | |
1424 | /* sanity check frenzy */ |
1425 | BUG_ON(worker->current_work); |
1426 | BUG_ON(!list_empty(&worker->scheduled)); |
1427 | |
1428 | if (worker->flags & WORKER_STARTED) |
1429 | gcwq->nr_workers--; |
1430 | if (worker->flags & WORKER_IDLE) |
1431 | gcwq->nr_idle--; |
1432 | |
1433 | list_del_init(&worker->entry); |
1434 | worker->flags |= WORKER_DIE; |
1435 | |
1436 | spin_unlock_irq(&gcwq->lock); |
1437 | |
1438 | kthread_stop(worker->task); |
1439 | kfree(worker); |
1440 | |
1441 | spin_lock_irq(&gcwq->lock); |
1442 | ida_remove(&gcwq->worker_ida, id); |
1443 | } |
1444 | |
1445 | static void idle_worker_timeout(unsigned long __gcwq) |
1446 | { |
1447 | struct global_cwq *gcwq = (void *)__gcwq; |
1448 | |
1449 | spin_lock_irq(&gcwq->lock); |
1450 | |
1451 | if (too_many_workers(gcwq)) { |
1452 | struct worker *worker; |
1453 | unsigned long expires; |
1454 | |
1455 | /* idle_list is kept in LIFO order, check the last one */ |
1456 | worker = list_entry(gcwq->idle_list.prev, struct worker, entry); |
1457 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1458 | |
1459 | if (time_before(jiffies, expires)) |
1460 | mod_timer(&gcwq->idle_timer, expires); |
1461 | else { |
1462 | /* it's been idle for too long, wake up manager */ |
1463 | gcwq->flags |= GCWQ_MANAGE_WORKERS; |
1464 | wake_up_worker(gcwq); |
1465 | } |
1466 | } |
1467 | |
1468 | spin_unlock_irq(&gcwq->lock); |
1469 | } |
1470 | |
1471 | static bool send_mayday(struct work_struct *work) |
1472 | { |
1473 | struct cpu_workqueue_struct *cwq = get_work_cwq(work); |
1474 | struct workqueue_struct *wq = cwq->wq; |
1475 | unsigned int cpu; |
1476 | |
1477 | if (!(wq->flags & WQ_RESCUER)) |
1478 | return false; |
1479 | |
1480 | /* mayday mayday mayday */ |
1481 | cpu = cwq->gcwq->cpu; |
1482 | /* WORK_CPU_UNBOUND can't be set in cpumask, use cpu 0 instead */ |
1483 | if (cpu == WORK_CPU_UNBOUND) |
1484 | cpu = 0; |
1485 | if (!mayday_test_and_set_cpu(cpu, wq->mayday_mask)) |
1486 | wake_up_process(wq->rescuer->task); |
1487 | return true; |
1488 | } |
1489 | |
1490 | static void gcwq_mayday_timeout(unsigned long __gcwq) |
1491 | { |
1492 | struct global_cwq *gcwq = (void *)__gcwq; |
1493 | struct work_struct *work; |
1494 | |
1495 | spin_lock_irq(&gcwq->lock); |
1496 | |
1497 | if (need_to_create_worker(gcwq)) { |
1498 | /* |
1499 | * We've been trying to create a new worker but |
1500 | * haven't been successful. We might be hitting an |
1501 | * allocation deadlock. Send distress signals to |
1502 | * rescuers. |
1503 | */ |
1504 | list_for_each_entry(work, &gcwq->worklist, entry) |
1505 | send_mayday(work); |
1506 | } |
1507 | |
1508 | spin_unlock_irq(&gcwq->lock); |
1509 | |
1510 | mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INTERVAL); |
1511 | } |
1512 | |
1513 | /** |
1514 | * maybe_create_worker - create a new worker if necessary |
1515 | * @gcwq: gcwq to create a new worker for |
1516 | * |
1517 | * Create a new worker for @gcwq if necessary. @gcwq is guaranteed to |
1518 | * have at least one idle worker on return from this function. If |
1519 | * creating a new worker takes longer than MAYDAY_INTERVAL, mayday is |
1520 | * sent to all rescuers with works scheduled on @gcwq to resolve |
1521 | * possible allocation deadlock. |
1522 | * |
1523 | * On return, need_to_create_worker() is guaranteed to be false and |
1524 | * may_start_working() true. |
1525 | * |
1526 | * LOCKING: |
1527 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
1528 | * multiple times. Does GFP_KERNEL allocations. Called only from |
1529 | * manager. |
1530 | * |
1531 | * RETURNS: |
1532 | * false if no action was taken and gcwq->lock stayed locked, true |
1533 | * otherwise. |
1534 | */ |
1535 | static bool maybe_create_worker(struct global_cwq *gcwq) |
1536 | __releases(&gcwq->lock) |
1537 | __acquires(&gcwq->lock) |
1538 | { |
1539 | if (!need_to_create_worker(gcwq)) |
1540 | return false; |
1541 | restart: |
1542 | spin_unlock_irq(&gcwq->lock); |
1543 | |
1544 | /* if we don't make progress in MAYDAY_INITIAL_TIMEOUT, call for help */ |
1545 | mod_timer(&gcwq->mayday_timer, jiffies + MAYDAY_INITIAL_TIMEOUT); |
1546 | |
1547 | while (true) { |
1548 | struct worker *worker; |
1549 | |
1550 | worker = create_worker(gcwq, true); |
1551 | if (worker) { |
1552 | del_timer_sync(&gcwq->mayday_timer); |
1553 | spin_lock_irq(&gcwq->lock); |
1554 | start_worker(worker); |
1555 | BUG_ON(need_to_create_worker(gcwq)); |
1556 | return true; |
1557 | } |
1558 | |
1559 | if (!need_to_create_worker(gcwq)) |
1560 | break; |
1561 | |
1562 | __set_current_state(TASK_INTERRUPTIBLE); |
1563 | schedule_timeout(CREATE_COOLDOWN); |
1564 | |
1565 | if (!need_to_create_worker(gcwq)) |
1566 | break; |
1567 | } |
1568 | |
1569 | del_timer_sync(&gcwq->mayday_timer); |
1570 | spin_lock_irq(&gcwq->lock); |
1571 | if (need_to_create_worker(gcwq)) |
1572 | goto restart; |
1573 | return true; |
1574 | } |
1575 | |
1576 | /** |
1577 | * maybe_destroy_worker - destroy workers which have been idle for a while |
1578 | * @gcwq: gcwq to destroy workers for |
1579 | * |
1580 | * Destroy @gcwq workers which have been idle for longer than |
1581 | * IDLE_WORKER_TIMEOUT. |
1582 | * |
1583 | * LOCKING: |
1584 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
1585 | * multiple times. Called only from manager. |
1586 | * |
1587 | * RETURNS: |
1588 | * false if no action was taken and gcwq->lock stayed locked, true |
1589 | * otherwise. |
1590 | */ |
1591 | static bool maybe_destroy_workers(struct global_cwq *gcwq) |
1592 | { |
1593 | bool ret = false; |
1594 | |
1595 | while (too_many_workers(gcwq)) { |
1596 | struct worker *worker; |
1597 | unsigned long expires; |
1598 | |
1599 | worker = list_entry(gcwq->idle_list.prev, struct worker, entry); |
1600 | expires = worker->last_active + IDLE_WORKER_TIMEOUT; |
1601 | |
1602 | if (time_before(jiffies, expires)) { |
1603 | mod_timer(&gcwq->idle_timer, expires); |
1604 | break; |
1605 | } |
1606 | |
1607 | destroy_worker(worker); |
1608 | ret = true; |
1609 | } |
1610 | |
1611 | return ret; |
1612 | } |
1613 | |
1614 | /** |
1615 | * manage_workers - manage worker pool |
1616 | * @worker: self |
1617 | * |
1618 | * Assume the manager role and manage gcwq worker pool @worker belongs |
1619 | * to. At any given time, there can be only zero or one manager per |
1620 | * gcwq. The exclusion is handled automatically by this function. |
1621 | * |
1622 | * The caller can safely start processing works on false return. On |
1623 | * true return, it's guaranteed that need_to_create_worker() is false |
1624 | * and may_start_working() is true. |
1625 | * |
1626 | * CONTEXT: |
1627 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
1628 | * multiple times. Does GFP_KERNEL allocations. |
1629 | * |
1630 | * RETURNS: |
1631 | * false if no action was taken and gcwq->lock stayed locked, true if |
1632 | * some action was taken. |
1633 | */ |
1634 | static bool manage_workers(struct worker *worker) |
1635 | { |
1636 | struct global_cwq *gcwq = worker->gcwq; |
1637 | bool ret = false; |
1638 | |
1639 | if (gcwq->flags & GCWQ_MANAGING_WORKERS) |
1640 | return ret; |
1641 | |
1642 | gcwq->flags &= ~GCWQ_MANAGE_WORKERS; |
1643 | gcwq->flags |= GCWQ_MANAGING_WORKERS; |
1644 | |
1645 | /* |
1646 | * Destroy and then create so that may_start_working() is true |
1647 | * on return. |
1648 | */ |
1649 | ret |= maybe_destroy_workers(gcwq); |
1650 | ret |= maybe_create_worker(gcwq); |
1651 | |
1652 | gcwq->flags &= ~GCWQ_MANAGING_WORKERS; |
1653 | |
1654 | /* |
1655 | * The trustee might be waiting to take over the manager |
1656 | * position, tell it we're done. |
1657 | */ |
1658 | if (unlikely(gcwq->trustee)) |
1659 | wake_up_all(&gcwq->trustee_wait); |
1660 | |
1661 | return ret; |
1662 | } |
1663 | |
1664 | /** |
1665 | * move_linked_works - move linked works to a list |
1666 | * @work: start of series of works to be scheduled |
1667 | * @head: target list to append @work to |
1668 | * @nextp: out paramter for nested worklist walking |
1669 | * |
1670 | * Schedule linked works starting from @work to @head. Work series to |
1671 | * be scheduled starts at @work and includes any consecutive work with |
1672 | * WORK_STRUCT_LINKED set in its predecessor. |
1673 | * |
1674 | * If @nextp is not NULL, it's updated to point to the next work of |
1675 | * the last scheduled work. This allows move_linked_works() to be |
1676 | * nested inside outer list_for_each_entry_safe(). |
1677 | * |
1678 | * CONTEXT: |
1679 | * spin_lock_irq(gcwq->lock). |
1680 | */ |
1681 | static void move_linked_works(struct work_struct *work, struct list_head *head, |
1682 | struct work_struct **nextp) |
1683 | { |
1684 | struct work_struct *n; |
1685 | |
1686 | /* |
1687 | * Linked worklist will always end before the end of the list, |
1688 | * use NULL for list head. |
1689 | */ |
1690 | list_for_each_entry_safe_from(work, n, NULL, entry) { |
1691 | list_move_tail(&work->entry, head); |
1692 | if (!(*work_data_bits(work) & WORK_STRUCT_LINKED)) |
1693 | break; |
1694 | } |
1695 | |
1696 | /* |
1697 | * If we're already inside safe list traversal and have moved |
1698 | * multiple works to the scheduled queue, the next position |
1699 | * needs to be updated. |
1700 | */ |
1701 | if (nextp) |
1702 | *nextp = n; |
1703 | } |
1704 | |
1705 | static void cwq_activate_first_delayed(struct cpu_workqueue_struct *cwq) |
1706 | { |
1707 | struct work_struct *work = list_first_entry(&cwq->delayed_works, |
1708 | struct work_struct, entry); |
1709 | struct list_head *pos = gcwq_determine_ins_pos(cwq->gcwq, cwq); |
1710 | |
1711 | trace_workqueue_activate_work(work); |
1712 | move_linked_works(work, pos, NULL); |
1713 | __clear_bit(WORK_STRUCT_DELAYED_BIT, work_data_bits(work)); |
1714 | cwq->nr_active++; |
1715 | } |
1716 | |
1717 | /** |
1718 | * cwq_dec_nr_in_flight - decrement cwq's nr_in_flight |
1719 | * @cwq: cwq of interest |
1720 | * @color: color of work which left the queue |
1721 | * @delayed: for a delayed work |
1722 | * |
1723 | * A work either has completed or is removed from pending queue, |
1724 | * decrement nr_in_flight of its cwq and handle workqueue flushing. |
1725 | * |
1726 | * CONTEXT: |
1727 | * spin_lock_irq(gcwq->lock). |
1728 | */ |
1729 | static void cwq_dec_nr_in_flight(struct cpu_workqueue_struct *cwq, int color, |
1730 | bool delayed) |
1731 | { |
1732 | /* ignore uncolored works */ |
1733 | if (color == WORK_NO_COLOR) |
1734 | return; |
1735 | |
1736 | cwq->nr_in_flight[color]--; |
1737 | |
1738 | if (!delayed) { |
1739 | cwq->nr_active--; |
1740 | if (!list_empty(&cwq->delayed_works)) { |
1741 | /* one down, submit a delayed one */ |
1742 | if (cwq->nr_active < cwq->max_active) |
1743 | cwq_activate_first_delayed(cwq); |
1744 | } |
1745 | } |
1746 | |
1747 | /* is flush in progress and are we at the flushing tip? */ |
1748 | if (likely(cwq->flush_color != color)) |
1749 | return; |
1750 | |
1751 | /* are there still in-flight works? */ |
1752 | if (cwq->nr_in_flight[color]) |
1753 | return; |
1754 | |
1755 | /* this cwq is done, clear flush_color */ |
1756 | cwq->flush_color = -1; |
1757 | |
1758 | /* |
1759 | * If this was the last cwq, wake up the first flusher. It |
1760 | * will handle the rest. |
1761 | */ |
1762 | if (atomic_dec_and_test(&cwq->wq->nr_cwqs_to_flush)) |
1763 | complete(&cwq->wq->first_flusher->done); |
1764 | } |
1765 | |
1766 | /** |
1767 | * process_one_work - process single work |
1768 | * @worker: self |
1769 | * @work: work to process |
1770 | * |
1771 | * Process @work. This function contains all the logics necessary to |
1772 | * process a single work including synchronization against and |
1773 | * interaction with other workers on the same cpu, queueing and |
1774 | * flushing. As long as context requirement is met, any worker can |
1775 | * call this function to process a work. |
1776 | * |
1777 | * CONTEXT: |
1778 | * spin_lock_irq(gcwq->lock) which is released and regrabbed. |
1779 | */ |
1780 | static void process_one_work(struct worker *worker, struct work_struct *work) |
1781 | __releases(&gcwq->lock) |
1782 | __acquires(&gcwq->lock) |
1783 | { |
1784 | struct cpu_workqueue_struct *cwq = get_work_cwq(work); |
1785 | struct global_cwq *gcwq = cwq->gcwq; |
1786 | struct hlist_head *bwh = busy_worker_head(gcwq, work); |
1787 | bool cpu_intensive = cwq->wq->flags & WQ_CPU_INTENSIVE; |
1788 | work_func_t f = work->func; |
1789 | int work_color; |
1790 | struct worker *collision; |
1791 | #ifdef CONFIG_LOCKDEP |
1792 | /* |
1793 | * It is permissible to free the struct work_struct from |
1794 | * inside the function that is called from it, this we need to |
1795 | * take into account for lockdep too. To avoid bogus "held |
1796 | * lock freed" warnings as well as problems when looking into |
1797 | * work->lockdep_map, make a copy and use that here. |
1798 | */ |
1799 | struct lockdep_map lockdep_map = work->lockdep_map; |
1800 | #endif |
1801 | /* |
1802 | * A single work shouldn't be executed concurrently by |
1803 | * multiple workers on a single cpu. Check whether anyone is |
1804 | * already processing the work. If so, defer the work to the |
1805 | * currently executing one. |
1806 | */ |
1807 | collision = __find_worker_executing_work(gcwq, bwh, work); |
1808 | if (unlikely(collision)) { |
1809 | move_linked_works(work, &collision->scheduled, NULL); |
1810 | return; |
1811 | } |
1812 | |
1813 | /* claim and process */ |
1814 | debug_work_deactivate(work); |
1815 | hlist_add_head(&worker->hentry, bwh); |
1816 | worker->current_work = work; |
1817 | worker->current_cwq = cwq; |
1818 | work_color = get_work_color(work); |
1819 | |
1820 | /* record the current cpu number in the work data and dequeue */ |
1821 | set_work_cpu(work, gcwq->cpu); |
1822 | list_del_init(&work->entry); |
1823 | |
1824 | /* |
1825 | * If HIGHPRI_PENDING, check the next work, and, if HIGHPRI, |
1826 | * wake up another worker; otherwise, clear HIGHPRI_PENDING. |
1827 | */ |
1828 | if (unlikely(gcwq->flags & GCWQ_HIGHPRI_PENDING)) { |
1829 | struct work_struct *nwork = list_first_entry(&gcwq->worklist, |
1830 | struct work_struct, entry); |
1831 | |
1832 | if (!list_empty(&gcwq->worklist) && |
1833 | get_work_cwq(nwork)->wq->flags & WQ_HIGHPRI) |
1834 | wake_up_worker(gcwq); |
1835 | else |
1836 | gcwq->flags &= ~GCWQ_HIGHPRI_PENDING; |
1837 | } |
1838 | |
1839 | /* |
1840 | * CPU intensive works don't participate in concurrency |
1841 | * management. They're the scheduler's responsibility. |
1842 | */ |
1843 | if (unlikely(cpu_intensive)) |
1844 | worker_set_flags(worker, WORKER_CPU_INTENSIVE, true); |
1845 | |
1846 | spin_unlock_irq(&gcwq->lock); |
1847 | |
1848 | work_clear_pending(work); |
1849 | lock_map_acquire_read(&cwq->wq->lockdep_map); |
1850 | lock_map_acquire(&lockdep_map); |
1851 | trace_workqueue_execute_start(work); |
1852 | f(work); |
1853 | /* |
1854 | * While we must be careful to not use "work" after this, the trace |
1855 | * point will only record its address. |
1856 | */ |
1857 | trace_workqueue_execute_end(work); |
1858 | lock_map_release(&lockdep_map); |
1859 | lock_map_release(&cwq->wq->lockdep_map); |
1860 | |
1861 | if (unlikely(in_atomic() || lockdep_depth(current) > 0)) { |
1862 | printk(KERN_ERR "BUG: workqueue leaked lock or atomic: " |
1863 | "%s/0x%08x/%d\n", |
1864 | current->comm, preempt_count(), task_pid_nr(current)); |
1865 | printk(KERN_ERR " last function: "); |
1866 | print_symbol("%s\n", (unsigned long)f); |
1867 | debug_show_held_locks(current); |
1868 | dump_stack(); |
1869 | } |
1870 | |
1871 | spin_lock_irq(&gcwq->lock); |
1872 | |
1873 | /* clear cpu intensive status */ |
1874 | if (unlikely(cpu_intensive)) |
1875 | worker_clr_flags(worker, WORKER_CPU_INTENSIVE); |
1876 | |
1877 | /* we're done with it, release */ |
1878 | hlist_del_init(&worker->hentry); |
1879 | worker->current_work = NULL; |
1880 | worker->current_cwq = NULL; |
1881 | cwq_dec_nr_in_flight(cwq, work_color, false); |
1882 | } |
1883 | |
1884 | /** |
1885 | * process_scheduled_works - process scheduled works |
1886 | * @worker: self |
1887 | * |
1888 | * Process all scheduled works. Please note that the scheduled list |
1889 | * may change while processing a work, so this function repeatedly |
1890 | * fetches a work from the top and executes it. |
1891 | * |
1892 | * CONTEXT: |
1893 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
1894 | * multiple times. |
1895 | */ |
1896 | static void process_scheduled_works(struct worker *worker) |
1897 | { |
1898 | while (!list_empty(&worker->scheduled)) { |
1899 | struct work_struct *work = list_first_entry(&worker->scheduled, |
1900 | struct work_struct, entry); |
1901 | process_one_work(worker, work); |
1902 | } |
1903 | } |
1904 | |
1905 | /** |
1906 | * worker_thread - the worker thread function |
1907 | * @__worker: self |
1908 | * |
1909 | * The gcwq worker thread function. There's a single dynamic pool of |
1910 | * these per each cpu. These workers process all works regardless of |
1911 | * their specific target workqueue. The only exception is works which |
1912 | * belong to workqueues with a rescuer which will be explained in |
1913 | * rescuer_thread(). |
1914 | */ |
1915 | static int worker_thread(void *__worker) |
1916 | { |
1917 | struct worker *worker = __worker; |
1918 | struct global_cwq *gcwq = worker->gcwq; |
1919 | |
1920 | /* tell the scheduler that this is a workqueue worker */ |
1921 | worker->task->flags |= PF_WQ_WORKER; |
1922 | woke_up: |
1923 | spin_lock_irq(&gcwq->lock); |
1924 | |
1925 | /* DIE can be set only while we're idle, checking here is enough */ |
1926 | if (worker->flags & WORKER_DIE) { |
1927 | spin_unlock_irq(&gcwq->lock); |
1928 | worker->task->flags &= ~PF_WQ_WORKER; |
1929 | return 0; |
1930 | } |
1931 | |
1932 | worker_leave_idle(worker); |
1933 | recheck: |
1934 | /* no more worker necessary? */ |
1935 | if (!need_more_worker(gcwq)) |
1936 | goto sleep; |
1937 | |
1938 | /* do we need to manage? */ |
1939 | if (unlikely(!may_start_working(gcwq)) && manage_workers(worker)) |
1940 | goto recheck; |
1941 | |
1942 | /* |
1943 | * ->scheduled list can only be filled while a worker is |
1944 | * preparing to process a work or actually processing it. |
1945 | * Make sure nobody diddled with it while I was sleeping. |
1946 | */ |
1947 | BUG_ON(!list_empty(&worker->scheduled)); |
1948 | |
1949 | /* |
1950 | * When control reaches this point, we're guaranteed to have |
1951 | * at least one idle worker or that someone else has already |
1952 | * assumed the manager role. |
1953 | */ |
1954 | worker_clr_flags(worker, WORKER_PREP); |
1955 | |
1956 | do { |
1957 | struct work_struct *work = |
1958 | list_first_entry(&gcwq->worklist, |
1959 | struct work_struct, entry); |
1960 | |
1961 | if (likely(!(*work_data_bits(work) & WORK_STRUCT_LINKED))) { |
1962 | /* optimization path, not strictly necessary */ |
1963 | process_one_work(worker, work); |
1964 | if (unlikely(!list_empty(&worker->scheduled))) |
1965 | process_scheduled_works(worker); |
1966 | } else { |
1967 | move_linked_works(work, &worker->scheduled, NULL); |
1968 | process_scheduled_works(worker); |
1969 | } |
1970 | } while (keep_working(gcwq)); |
1971 | |
1972 | worker_set_flags(worker, WORKER_PREP, false); |
1973 | sleep: |
1974 | if (unlikely(need_to_manage_workers(gcwq)) && manage_workers(worker)) |
1975 | goto recheck; |
1976 | |
1977 | /* |
1978 | * gcwq->lock is held and there's no work to process and no |
1979 | * need to manage, sleep. Workers are woken up only while |
1980 | * holding gcwq->lock or from local cpu, so setting the |
1981 | * current state before releasing gcwq->lock is enough to |
1982 | * prevent losing any event. |
1983 | */ |
1984 | worker_enter_idle(worker); |
1985 | __set_current_state(TASK_INTERRUPTIBLE); |
1986 | spin_unlock_irq(&gcwq->lock); |
1987 | schedule(); |
1988 | goto woke_up; |
1989 | } |
1990 | |
1991 | /** |
1992 | * rescuer_thread - the rescuer thread function |
1993 | * @__wq: the associated workqueue |
1994 | * |
1995 | * Workqueue rescuer thread function. There's one rescuer for each |
1996 | * workqueue which has WQ_RESCUER set. |
1997 | * |
1998 | * Regular work processing on a gcwq may block trying to create a new |
1999 | * worker which uses GFP_KERNEL allocation which has slight chance of |
2000 | * developing into deadlock if some works currently on the same queue |
2001 | * need to be processed to satisfy the GFP_KERNEL allocation. This is |
2002 | * the problem rescuer solves. |
2003 | * |
2004 | * When such condition is possible, the gcwq summons rescuers of all |
2005 | * workqueues which have works queued on the gcwq and let them process |
2006 | * those works so that forward progress can be guaranteed. |
2007 | * |
2008 | * This should happen rarely. |
2009 | */ |
2010 | static int rescuer_thread(void *__wq) |
2011 | { |
2012 | struct workqueue_struct *wq = __wq; |
2013 | struct worker *rescuer = wq->rescuer; |
2014 | struct list_head *scheduled = &rescuer->scheduled; |
2015 | bool is_unbound = wq->flags & WQ_UNBOUND; |
2016 | unsigned int cpu; |
2017 | |
2018 | set_user_nice(current, RESCUER_NICE_LEVEL); |
2019 | repeat: |
2020 | set_current_state(TASK_INTERRUPTIBLE); |
2021 | |
2022 | if (kthread_should_stop()) |
2023 | return 0; |
2024 | |
2025 | /* |
2026 | * See whether any cpu is asking for help. Unbounded |
2027 | * workqueues use cpu 0 in mayday_mask for CPU_UNBOUND. |
2028 | */ |
2029 | for_each_mayday_cpu(cpu, wq->mayday_mask) { |
2030 | unsigned int tcpu = is_unbound ? WORK_CPU_UNBOUND : cpu; |
2031 | struct cpu_workqueue_struct *cwq = get_cwq(tcpu, wq); |
2032 | struct global_cwq *gcwq = cwq->gcwq; |
2033 | struct work_struct *work, *n; |
2034 | |
2035 | __set_current_state(TASK_RUNNING); |
2036 | mayday_clear_cpu(cpu, wq->mayday_mask); |
2037 | |
2038 | /* migrate to the target cpu if possible */ |
2039 | rescuer->gcwq = gcwq; |
2040 | worker_maybe_bind_and_lock(rescuer); |
2041 | |
2042 | /* |
2043 | * Slurp in all works issued via this workqueue and |
2044 | * process'em. |
2045 | */ |
2046 | BUG_ON(!list_empty(&rescuer->scheduled)); |
2047 | list_for_each_entry_safe(work, n, &gcwq->worklist, entry) |
2048 | if (get_work_cwq(work) == cwq) |
2049 | move_linked_works(work, scheduled, &n); |
2050 | |
2051 | process_scheduled_works(rescuer); |
2052 | |
2053 | /* |
2054 | * Leave this gcwq. If keep_working() is %true, notify a |
2055 | * regular worker; otherwise, we end up with 0 concurrency |
2056 | * and stalling the execution. |
2057 | */ |
2058 | if (keep_working(gcwq)) |
2059 | wake_up_worker(gcwq); |
2060 | |
2061 | spin_unlock_irq(&gcwq->lock); |
2062 | } |
2063 | |
2064 | schedule(); |
2065 | goto repeat; |
2066 | } |
2067 | |
2068 | struct wq_barrier { |
2069 | struct work_struct work; |
2070 | struct completion done; |
2071 | }; |
2072 | |
2073 | static void wq_barrier_func(struct work_struct *work) |
2074 | { |
2075 | struct wq_barrier *barr = container_of(work, struct wq_barrier, work); |
2076 | complete(&barr->done); |
2077 | } |
2078 | |
2079 | /** |
2080 | * insert_wq_barrier - insert a barrier work |
2081 | * @cwq: cwq to insert barrier into |
2082 | * @barr: wq_barrier to insert |
2083 | * @target: target work to attach @barr to |
2084 | * @worker: worker currently executing @target, NULL if @target is not executing |
2085 | * |
2086 | * @barr is linked to @target such that @barr is completed only after |
2087 | * @target finishes execution. Please note that the ordering |
2088 | * guarantee is observed only with respect to @target and on the local |
2089 | * cpu. |
2090 | * |
2091 | * Currently, a queued barrier can't be canceled. This is because |
2092 | * try_to_grab_pending() can't determine whether the work to be |
2093 | * grabbed is at the head of the queue and thus can't clear LINKED |
2094 | * flag of the previous work while there must be a valid next work |
2095 | * after a work with LINKED flag set. |
2096 | * |
2097 | * Note that when @worker is non-NULL, @target may be modified |
2098 | * underneath us, so we can't reliably determine cwq from @target. |
2099 | * |
2100 | * CONTEXT: |
2101 | * spin_lock_irq(gcwq->lock). |
2102 | */ |
2103 | static void insert_wq_barrier(struct cpu_workqueue_struct *cwq, |
2104 | struct wq_barrier *barr, |
2105 | struct work_struct *target, struct worker *worker) |
2106 | { |
2107 | struct list_head *head; |
2108 | unsigned int linked = 0; |
2109 | |
2110 | /* |
2111 | * debugobject calls are safe here even with gcwq->lock locked |
2112 | * as we know for sure that this will not trigger any of the |
2113 | * checks and call back into the fixup functions where we |
2114 | * might deadlock. |
2115 | */ |
2116 | INIT_WORK_ONSTACK(&barr->work, wq_barrier_func); |
2117 | __set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(&barr->work)); |
2118 | init_completion(&barr->done); |
2119 | |
2120 | /* |
2121 | * If @target is currently being executed, schedule the |
2122 | * barrier to the worker; otherwise, put it after @target. |
2123 | */ |
2124 | if (worker) |
2125 | head = worker->scheduled.next; |
2126 | else { |
2127 | unsigned long *bits = work_data_bits(target); |
2128 | |
2129 | head = target->entry.next; |
2130 | /* there can already be other linked works, inherit and set */ |
2131 | linked = *bits & WORK_STRUCT_LINKED; |
2132 | __set_bit(WORK_STRUCT_LINKED_BIT, bits); |
2133 | } |
2134 | |
2135 | debug_work_activate(&barr->work); |
2136 | insert_work(cwq, &barr->work, head, |
2137 | work_color_to_flags(WORK_NO_COLOR) | linked); |
2138 | } |
2139 | |
2140 | /** |
2141 | * flush_workqueue_prep_cwqs - prepare cwqs for workqueue flushing |
2142 | * @wq: workqueue being flushed |
2143 | * @flush_color: new flush color, < 0 for no-op |
2144 | * @work_color: new work color, < 0 for no-op |
2145 | * |
2146 | * Prepare cwqs for workqueue flushing. |
2147 | * |
2148 | * If @flush_color is non-negative, flush_color on all cwqs should be |
2149 | * -1. If no cwq has in-flight commands at the specified color, all |
2150 | * cwq->flush_color's stay at -1 and %false is returned. If any cwq |
2151 | * has in flight commands, its cwq->flush_color is set to |
2152 | * @flush_color, @wq->nr_cwqs_to_flush is updated accordingly, cwq |
2153 | * wakeup logic is armed and %true is returned. |
2154 | * |
2155 | * The caller should have initialized @wq->first_flusher prior to |
2156 | * calling this function with non-negative @flush_color. If |
2157 | * @flush_color is negative, no flush color update is done and %false |
2158 | * is returned. |
2159 | * |
2160 | * If @work_color is non-negative, all cwqs should have the same |
2161 | * work_color which is previous to @work_color and all will be |
2162 | * advanced to @work_color. |
2163 | * |
2164 | * CONTEXT: |
2165 | * mutex_lock(wq->flush_mutex). |
2166 | * |
2167 | * RETURNS: |
2168 | * %true if @flush_color >= 0 and there's something to flush. %false |
2169 | * otherwise. |
2170 | */ |
2171 | static bool flush_workqueue_prep_cwqs(struct workqueue_struct *wq, |
2172 | int flush_color, int work_color) |
2173 | { |
2174 | bool wait = false; |
2175 | unsigned int cpu; |
2176 | |
2177 | if (flush_color >= 0) { |
2178 | BUG_ON(atomic_read(&wq->nr_cwqs_to_flush)); |
2179 | atomic_set(&wq->nr_cwqs_to_flush, 1); |
2180 | } |
2181 | |
2182 | for_each_cwq_cpu(cpu, wq) { |
2183 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
2184 | struct global_cwq *gcwq = cwq->gcwq; |
2185 | |
2186 | spin_lock_irq(&gcwq->lock); |
2187 | |
2188 | if (flush_color >= 0) { |
2189 | BUG_ON(cwq->flush_color != -1); |
2190 | |
2191 | if (cwq->nr_in_flight[flush_color]) { |
2192 | cwq->flush_color = flush_color; |
2193 | atomic_inc(&wq->nr_cwqs_to_flush); |
2194 | wait = true; |
2195 | } |
2196 | } |
2197 | |
2198 | if (work_color >= 0) { |
2199 | BUG_ON(work_color != work_next_color(cwq->work_color)); |
2200 | cwq->work_color = work_color; |
2201 | } |
2202 | |
2203 | spin_unlock_irq(&gcwq->lock); |
2204 | } |
2205 | |
2206 | if (flush_color >= 0 && atomic_dec_and_test(&wq->nr_cwqs_to_flush)) |
2207 | complete(&wq->first_flusher->done); |
2208 | |
2209 | return wait; |
2210 | } |
2211 | |
2212 | /** |
2213 | * flush_workqueue - ensure that any scheduled work has run to completion. |
2214 | * @wq: workqueue to flush |
2215 | * |
2216 | * Forces execution of the workqueue and blocks until its completion. |
2217 | * This is typically used in driver shutdown handlers. |
2218 | * |
2219 | * We sleep until all works which were queued on entry have been handled, |
2220 | * but we are not livelocked by new incoming ones. |
2221 | */ |
2222 | void flush_workqueue(struct workqueue_struct *wq) |
2223 | { |
2224 | struct wq_flusher this_flusher = { |
2225 | .list = LIST_HEAD_INIT(this_flusher.list), |
2226 | .flush_color = -1, |
2227 | .done = COMPLETION_INITIALIZER_ONSTACK(this_flusher.done), |
2228 | }; |
2229 | int next_color; |
2230 | |
2231 | lock_map_acquire(&wq->lockdep_map); |
2232 | lock_map_release(&wq->lockdep_map); |
2233 | |
2234 | mutex_lock(&wq->flush_mutex); |
2235 | |
2236 | /* |
2237 | * Start-to-wait phase |
2238 | */ |
2239 | next_color = work_next_color(wq->work_color); |
2240 | |
2241 | if (next_color != wq->flush_color) { |
2242 | /* |
2243 | * Color space is not full. The current work_color |
2244 | * becomes our flush_color and work_color is advanced |
2245 | * by one. |
2246 | */ |
2247 | BUG_ON(!list_empty(&wq->flusher_overflow)); |
2248 | this_flusher.flush_color = wq->work_color; |
2249 | wq->work_color = next_color; |
2250 | |
2251 | if (!wq->first_flusher) { |
2252 | /* no flush in progress, become the first flusher */ |
2253 | BUG_ON(wq->flush_color != this_flusher.flush_color); |
2254 | |
2255 | wq->first_flusher = &this_flusher; |
2256 | |
2257 | if (!flush_workqueue_prep_cwqs(wq, wq->flush_color, |
2258 | wq->work_color)) { |
2259 | /* nothing to flush, done */ |
2260 | wq->flush_color = next_color; |
2261 | wq->first_flusher = NULL; |
2262 | goto out_unlock; |
2263 | } |
2264 | } else { |
2265 | /* wait in queue */ |
2266 | BUG_ON(wq->flush_color == this_flusher.flush_color); |
2267 | list_add_tail(&this_flusher.list, &wq->flusher_queue); |
2268 | flush_workqueue_prep_cwqs(wq, -1, wq->work_color); |
2269 | } |
2270 | } else { |
2271 | /* |
2272 | * Oops, color space is full, wait on overflow queue. |
2273 | * The next flush completion will assign us |
2274 | * flush_color and transfer to flusher_queue. |
2275 | */ |
2276 | list_add_tail(&this_flusher.list, &wq->flusher_overflow); |
2277 | } |
2278 | |
2279 | mutex_unlock(&wq->flush_mutex); |
2280 | |
2281 | wait_for_completion(&this_flusher.done); |
2282 | |
2283 | /* |
2284 | * Wake-up-and-cascade phase |
2285 | * |
2286 | * First flushers are responsible for cascading flushes and |
2287 | * handling overflow. Non-first flushers can simply return. |
2288 | */ |
2289 | if (wq->first_flusher != &this_flusher) |
2290 | return; |
2291 | |
2292 | mutex_lock(&wq->flush_mutex); |
2293 | |
2294 | /* we might have raced, check again with mutex held */ |
2295 | if (wq->first_flusher != &this_flusher) |
2296 | goto out_unlock; |
2297 | |
2298 | wq->first_flusher = NULL; |
2299 | |
2300 | BUG_ON(!list_empty(&this_flusher.list)); |
2301 | BUG_ON(wq->flush_color != this_flusher.flush_color); |
2302 | |
2303 | while (true) { |
2304 | struct wq_flusher *next, *tmp; |
2305 | |
2306 | /* complete all the flushers sharing the current flush color */ |
2307 | list_for_each_entry_safe(next, tmp, &wq->flusher_queue, list) { |
2308 | if (next->flush_color != wq->flush_color) |
2309 | break; |
2310 | list_del_init(&next->list); |
2311 | complete(&next->done); |
2312 | } |
2313 | |
2314 | BUG_ON(!list_empty(&wq->flusher_overflow) && |
2315 | wq->flush_color != work_next_color(wq->work_color)); |
2316 | |
2317 | /* this flush_color is finished, advance by one */ |
2318 | wq->flush_color = work_next_color(wq->flush_color); |
2319 | |
2320 | /* one color has been freed, handle overflow queue */ |
2321 | if (!list_empty(&wq->flusher_overflow)) { |
2322 | /* |
2323 | * Assign the same color to all overflowed |
2324 | * flushers, advance work_color and append to |
2325 | * flusher_queue. This is the start-to-wait |
2326 | * phase for these overflowed flushers. |
2327 | */ |
2328 | list_for_each_entry(tmp, &wq->flusher_overflow, list) |
2329 | tmp->flush_color = wq->work_color; |
2330 | |
2331 | wq->work_color = work_next_color(wq->work_color); |
2332 | |
2333 | list_splice_tail_init(&wq->flusher_overflow, |
2334 | &wq->flusher_queue); |
2335 | flush_workqueue_prep_cwqs(wq, -1, wq->work_color); |
2336 | } |
2337 | |
2338 | if (list_empty(&wq->flusher_queue)) { |
2339 | BUG_ON(wq->flush_color != wq->work_color); |
2340 | break; |
2341 | } |
2342 | |
2343 | /* |
2344 | * Need to flush more colors. Make the next flusher |
2345 | * the new first flusher and arm cwqs. |
2346 | */ |
2347 | BUG_ON(wq->flush_color == wq->work_color); |
2348 | BUG_ON(wq->flush_color != next->flush_color); |
2349 | |
2350 | list_del_init(&next->list); |
2351 | wq->first_flusher = next; |
2352 | |
2353 | if (flush_workqueue_prep_cwqs(wq, wq->flush_color, -1)) |
2354 | break; |
2355 | |
2356 | /* |
2357 | * Meh... this color is already done, clear first |
2358 | * flusher and repeat cascading. |
2359 | */ |
2360 | wq->first_flusher = NULL; |
2361 | } |
2362 | |
2363 | out_unlock: |
2364 | mutex_unlock(&wq->flush_mutex); |
2365 | } |
2366 | EXPORT_SYMBOL_GPL(flush_workqueue); |
2367 | |
2368 | static bool start_flush_work(struct work_struct *work, struct wq_barrier *barr, |
2369 | bool wait_executing) |
2370 | { |
2371 | struct worker *worker = NULL; |
2372 | struct global_cwq *gcwq; |
2373 | struct cpu_workqueue_struct *cwq; |
2374 | |
2375 | might_sleep(); |
2376 | gcwq = get_work_gcwq(work); |
2377 | if (!gcwq) |
2378 | return false; |
2379 | |
2380 | spin_lock_irq(&gcwq->lock); |
2381 | if (!list_empty(&work->entry)) { |
2382 | /* |
2383 | * See the comment near try_to_grab_pending()->smp_rmb(). |
2384 | * If it was re-queued to a different gcwq under us, we |
2385 | * are not going to wait. |
2386 | */ |
2387 | smp_rmb(); |
2388 | cwq = get_work_cwq(work); |
2389 | if (unlikely(!cwq || gcwq != cwq->gcwq)) |
2390 | goto already_gone; |
2391 | } else if (wait_executing) { |
2392 | worker = find_worker_executing_work(gcwq, work); |
2393 | if (!worker) |
2394 | goto already_gone; |
2395 | cwq = worker->current_cwq; |
2396 | } else |
2397 | goto already_gone; |
2398 | |
2399 | insert_wq_barrier(cwq, barr, work, worker); |
2400 | spin_unlock_irq(&gcwq->lock); |
2401 | |
2402 | /* |
2403 | * If @max_active is 1 or rescuer is in use, flushing another work |
2404 | * item on the same workqueue may lead to deadlock. Make sure the |
2405 | * flusher is not running on the same workqueue by verifying write |
2406 | * access. |
2407 | */ |
2408 | if (cwq->wq->saved_max_active == 1 || cwq->wq->flags & WQ_RESCUER) |
2409 | lock_map_acquire(&cwq->wq->lockdep_map); |
2410 | else |
2411 | lock_map_acquire_read(&cwq->wq->lockdep_map); |
2412 | lock_map_release(&cwq->wq->lockdep_map); |
2413 | |
2414 | return true; |
2415 | already_gone: |
2416 | spin_unlock_irq(&gcwq->lock); |
2417 | return false; |
2418 | } |
2419 | |
2420 | /** |
2421 | * flush_work - wait for a work to finish executing the last queueing instance |
2422 | * @work: the work to flush |
2423 | * |
2424 | * Wait until @work has finished execution. This function considers |
2425 | * only the last queueing instance of @work. If @work has been |
2426 | * enqueued across different CPUs on a non-reentrant workqueue or on |
2427 | * multiple workqueues, @work might still be executing on return on |
2428 | * some of the CPUs from earlier queueing. |
2429 | * |
2430 | * If @work was queued only on a non-reentrant, ordered or unbound |
2431 | * workqueue, @work is guaranteed to be idle on return if it hasn't |
2432 | * been requeued since flush started. |
2433 | * |
2434 | * RETURNS: |
2435 | * %true if flush_work() waited for the work to finish execution, |
2436 | * %false if it was already idle. |
2437 | */ |
2438 | bool flush_work(struct work_struct *work) |
2439 | { |
2440 | struct wq_barrier barr; |
2441 | |
2442 | if (start_flush_work(work, &barr, true)) { |
2443 | wait_for_completion(&barr.done); |
2444 | destroy_work_on_stack(&barr.work); |
2445 | return true; |
2446 | } else |
2447 | return false; |
2448 | } |
2449 | EXPORT_SYMBOL_GPL(flush_work); |
2450 | |
2451 | static bool wait_on_cpu_work(struct global_cwq *gcwq, struct work_struct *work) |
2452 | { |
2453 | struct wq_barrier barr; |
2454 | struct worker *worker; |
2455 | |
2456 | spin_lock_irq(&gcwq->lock); |
2457 | |
2458 | worker = find_worker_executing_work(gcwq, work); |
2459 | if (unlikely(worker)) |
2460 | insert_wq_barrier(worker->current_cwq, &barr, work, worker); |
2461 | |
2462 | spin_unlock_irq(&gcwq->lock); |
2463 | |
2464 | if (unlikely(worker)) { |
2465 | wait_for_completion(&barr.done); |
2466 | destroy_work_on_stack(&barr.work); |
2467 | return true; |
2468 | } else |
2469 | return false; |
2470 | } |
2471 | |
2472 | static bool wait_on_work(struct work_struct *work) |
2473 | { |
2474 | bool ret = false; |
2475 | int cpu; |
2476 | |
2477 | might_sleep(); |
2478 | |
2479 | lock_map_acquire(&work->lockdep_map); |
2480 | lock_map_release(&work->lockdep_map); |
2481 | |
2482 | for_each_gcwq_cpu(cpu) |
2483 | ret |= wait_on_cpu_work(get_gcwq(cpu), work); |
2484 | return ret; |
2485 | } |
2486 | |
2487 | /** |
2488 | * flush_work_sync - wait until a work has finished execution |
2489 | * @work: the work to flush |
2490 | * |
2491 | * Wait until @work has finished execution. On return, it's |
2492 | * guaranteed that all queueing instances of @work which happened |
2493 | * before this function is called are finished. In other words, if |
2494 | * @work hasn't been requeued since this function was called, @work is |
2495 | * guaranteed to be idle on return. |
2496 | * |
2497 | * RETURNS: |
2498 | * %true if flush_work_sync() waited for the work to finish execution, |
2499 | * %false if it was already idle. |
2500 | */ |
2501 | bool flush_work_sync(struct work_struct *work) |
2502 | { |
2503 | struct wq_barrier barr; |
2504 | bool pending, waited; |
2505 | |
2506 | /* we'll wait for executions separately, queue barr only if pending */ |
2507 | pending = start_flush_work(work, &barr, false); |
2508 | |
2509 | /* wait for executions to finish */ |
2510 | waited = wait_on_work(work); |
2511 | |
2512 | /* wait for the pending one */ |
2513 | if (pending) { |
2514 | wait_for_completion(&barr.done); |
2515 | destroy_work_on_stack(&barr.work); |
2516 | } |
2517 | |
2518 | return pending || waited; |
2519 | } |
2520 | EXPORT_SYMBOL_GPL(flush_work_sync); |
2521 | |
2522 | /* |
2523 | * Upon a successful return (>= 0), the caller "owns" WORK_STRUCT_PENDING bit, |
2524 | * so this work can't be re-armed in any way. |
2525 | */ |
2526 | static int try_to_grab_pending(struct work_struct *work) |
2527 | { |
2528 | struct global_cwq *gcwq; |
2529 | int ret = -1; |
2530 | |
2531 | if (!test_and_set_bit(WORK_STRUCT_PENDING_BIT, work_data_bits(work))) |
2532 | return 0; |
2533 | |
2534 | /* |
2535 | * The queueing is in progress, or it is already queued. Try to |
2536 | * steal it from ->worklist without clearing WORK_STRUCT_PENDING. |
2537 | */ |
2538 | gcwq = get_work_gcwq(work); |
2539 | if (!gcwq) |
2540 | return ret; |
2541 | |
2542 | spin_lock_irq(&gcwq->lock); |
2543 | if (!list_empty(&work->entry)) { |
2544 | /* |
2545 | * This work is queued, but perhaps we locked the wrong gcwq. |
2546 | * In that case we must see the new value after rmb(), see |
2547 | * insert_work()->wmb(). |
2548 | */ |
2549 | smp_rmb(); |
2550 | if (gcwq == get_work_gcwq(work)) { |
2551 | debug_work_deactivate(work); |
2552 | list_del_init(&work->entry); |
2553 | cwq_dec_nr_in_flight(get_work_cwq(work), |
2554 | get_work_color(work), |
2555 | *work_data_bits(work) & WORK_STRUCT_DELAYED); |
2556 | ret = 1; |
2557 | } |
2558 | } |
2559 | spin_unlock_irq(&gcwq->lock); |
2560 | |
2561 | return ret; |
2562 | } |
2563 | |
2564 | static bool __cancel_work_timer(struct work_struct *work, |
2565 | struct timer_list* timer) |
2566 | { |
2567 | int ret; |
2568 | |
2569 | do { |
2570 | ret = (timer && likely(del_timer(timer))); |
2571 | if (!ret) |
2572 | ret = try_to_grab_pending(work); |
2573 | wait_on_work(work); |
2574 | } while (unlikely(ret < 0)); |
2575 | |
2576 | clear_work_data(work); |
2577 | return ret; |
2578 | } |
2579 | |
2580 | /** |
2581 | * cancel_work_sync - cancel a work and wait for it to finish |
2582 | * @work: the work to cancel |
2583 | * |
2584 | * Cancel @work and wait for its execution to finish. This function |
2585 | * can be used even if the work re-queues itself or migrates to |
2586 | * another workqueue. On return from this function, @work is |
2587 | * guaranteed to be not pending or executing on any CPU. |
2588 | * |
2589 | * cancel_work_sync(&delayed_work->work) must not be used for |
2590 | * delayed_work's. Use cancel_delayed_work_sync() instead. |
2591 | * |
2592 | * The caller must ensure that the workqueue on which @work was last |
2593 | * queued can't be destroyed before this function returns. |
2594 | * |
2595 | * RETURNS: |
2596 | * %true if @work was pending, %false otherwise. |
2597 | */ |
2598 | bool cancel_work_sync(struct work_struct *work) |
2599 | { |
2600 | return __cancel_work_timer(work, NULL); |
2601 | } |
2602 | EXPORT_SYMBOL_GPL(cancel_work_sync); |
2603 | |
2604 | /** |
2605 | * flush_delayed_work - wait for a dwork to finish executing the last queueing |
2606 | * @dwork: the delayed work to flush |
2607 | * |
2608 | * Delayed timer is cancelled and the pending work is queued for |
2609 | * immediate execution. Like flush_work(), this function only |
2610 | * considers the last queueing instance of @dwork. |
2611 | * |
2612 | * RETURNS: |
2613 | * %true if flush_work() waited for the work to finish execution, |
2614 | * %false if it was already idle. |
2615 | */ |
2616 | bool flush_delayed_work(struct delayed_work *dwork) |
2617 | { |
2618 | if (del_timer_sync(&dwork->timer)) |
2619 | __queue_work(raw_smp_processor_id(), |
2620 | get_work_cwq(&dwork->work)->wq, &dwork->work); |
2621 | return flush_work(&dwork->work); |
2622 | } |
2623 | EXPORT_SYMBOL(flush_delayed_work); |
2624 | |
2625 | /** |
2626 | * flush_delayed_work_sync - wait for a dwork to finish |
2627 | * @dwork: the delayed work to flush |
2628 | * |
2629 | * Delayed timer is cancelled and the pending work is queued for |
2630 | * execution immediately. Other than timer handling, its behavior |
2631 | * is identical to flush_work_sync(). |
2632 | * |
2633 | * RETURNS: |
2634 | * %true if flush_work_sync() waited for the work to finish execution, |
2635 | * %false if it was already idle. |
2636 | */ |
2637 | bool flush_delayed_work_sync(struct delayed_work *dwork) |
2638 | { |
2639 | if (del_timer_sync(&dwork->timer)) |
2640 | __queue_work(raw_smp_processor_id(), |
2641 | get_work_cwq(&dwork->work)->wq, &dwork->work); |
2642 | return flush_work_sync(&dwork->work); |
2643 | } |
2644 | EXPORT_SYMBOL(flush_delayed_work_sync); |
2645 | |
2646 | /** |
2647 | * cancel_delayed_work_sync - cancel a delayed work and wait for it to finish |
2648 | * @dwork: the delayed work cancel |
2649 | * |
2650 | * This is cancel_work_sync() for delayed works. |
2651 | * |
2652 | * RETURNS: |
2653 | * %true if @dwork was pending, %false otherwise. |
2654 | */ |
2655 | bool cancel_delayed_work_sync(struct delayed_work *dwork) |
2656 | { |
2657 | return __cancel_work_timer(&dwork->work, &dwork->timer); |
2658 | } |
2659 | EXPORT_SYMBOL(cancel_delayed_work_sync); |
2660 | |
2661 | /** |
2662 | * schedule_work - put work task in global workqueue |
2663 | * @work: job to be done |
2664 | * |
2665 | * Returns zero if @work was already on the kernel-global workqueue and |
2666 | * non-zero otherwise. |
2667 | * |
2668 | * This puts a job in the kernel-global workqueue if it was not already |
2669 | * queued and leaves it in the same position on the kernel-global |
2670 | * workqueue otherwise. |
2671 | */ |
2672 | int schedule_work(struct work_struct *work) |
2673 | { |
2674 | return queue_work(system_wq, work); |
2675 | } |
2676 | EXPORT_SYMBOL(schedule_work); |
2677 | |
2678 | /* |
2679 | * schedule_work_on - put work task on a specific cpu |
2680 | * @cpu: cpu to put the work task on |
2681 | * @work: job to be done |
2682 | * |
2683 | * This puts a job on a specific cpu |
2684 | */ |
2685 | int schedule_work_on(int cpu, struct work_struct *work) |
2686 | { |
2687 | return queue_work_on(cpu, system_wq, work); |
2688 | } |
2689 | EXPORT_SYMBOL(schedule_work_on); |
2690 | |
2691 | /** |
2692 | * schedule_delayed_work - put work task in global workqueue after delay |
2693 | * @dwork: job to be done |
2694 | * @delay: number of jiffies to wait or 0 for immediate execution |
2695 | * |
2696 | * After waiting for a given time this puts a job in the kernel-global |
2697 | * workqueue. |
2698 | */ |
2699 | int schedule_delayed_work(struct delayed_work *dwork, |
2700 | unsigned long delay) |
2701 | { |
2702 | return queue_delayed_work(system_wq, dwork, delay); |
2703 | } |
2704 | EXPORT_SYMBOL(schedule_delayed_work); |
2705 | |
2706 | /** |
2707 | * schedule_delayed_work_on - queue work in global workqueue on CPU after delay |
2708 | * @cpu: cpu to use |
2709 | * @dwork: job to be done |
2710 | * @delay: number of jiffies to wait |
2711 | * |
2712 | * After waiting for a given time this puts a job in the kernel-global |
2713 | * workqueue on the specified CPU. |
2714 | */ |
2715 | int schedule_delayed_work_on(int cpu, |
2716 | struct delayed_work *dwork, unsigned long delay) |
2717 | { |
2718 | return queue_delayed_work_on(cpu, system_wq, dwork, delay); |
2719 | } |
2720 | EXPORT_SYMBOL(schedule_delayed_work_on); |
2721 | |
2722 | /** |
2723 | * schedule_on_each_cpu - execute a function synchronously on each online CPU |
2724 | * @func: the function to call |
2725 | * |
2726 | * schedule_on_each_cpu() executes @func on each online CPU using the |
2727 | * system workqueue and blocks until all CPUs have completed. |
2728 | * schedule_on_each_cpu() is very slow. |
2729 | * |
2730 | * RETURNS: |
2731 | * 0 on success, -errno on failure. |
2732 | */ |
2733 | int schedule_on_each_cpu(work_func_t func) |
2734 | { |
2735 | int cpu; |
2736 | struct work_struct __percpu *works; |
2737 | |
2738 | works = alloc_percpu(struct work_struct); |
2739 | if (!works) |
2740 | return -ENOMEM; |
2741 | |
2742 | get_online_cpus(); |
2743 | |
2744 | for_each_online_cpu(cpu) { |
2745 | struct work_struct *work = per_cpu_ptr(works, cpu); |
2746 | |
2747 | INIT_WORK(work, func); |
2748 | schedule_work_on(cpu, work); |
2749 | } |
2750 | |
2751 | for_each_online_cpu(cpu) |
2752 | flush_work(per_cpu_ptr(works, cpu)); |
2753 | |
2754 | put_online_cpus(); |
2755 | free_percpu(works); |
2756 | return 0; |
2757 | } |
2758 | |
2759 | /** |
2760 | * flush_scheduled_work - ensure that any scheduled work has run to completion. |
2761 | * |
2762 | * Forces execution of the kernel-global workqueue and blocks until its |
2763 | * completion. |
2764 | * |
2765 | * Think twice before calling this function! It's very easy to get into |
2766 | * trouble if you don't take great care. Either of the following situations |
2767 | * will lead to deadlock: |
2768 | * |
2769 | * One of the work items currently on the workqueue needs to acquire |
2770 | * a lock held by your code or its caller. |
2771 | * |
2772 | * Your code is running in the context of a work routine. |
2773 | * |
2774 | * They will be detected by lockdep when they occur, but the first might not |
2775 | * occur very often. It depends on what work items are on the workqueue and |
2776 | * what locks they need, which you have no control over. |
2777 | * |
2778 | * In most situations flushing the entire workqueue is overkill; you merely |
2779 | * need to know that a particular work item isn't queued and isn't running. |
2780 | * In such cases you should use cancel_delayed_work_sync() or |
2781 | * cancel_work_sync() instead. |
2782 | */ |
2783 | void flush_scheduled_work(void) |
2784 | { |
2785 | flush_workqueue(system_wq); |
2786 | } |
2787 | EXPORT_SYMBOL(flush_scheduled_work); |
2788 | |
2789 | /** |
2790 | * execute_in_process_context - reliably execute the routine with user context |
2791 | * @fn: the function to execute |
2792 | * @ew: guaranteed storage for the execute work structure (must |
2793 | * be available when the work executes) |
2794 | * |
2795 | * Executes the function immediately if process context is available, |
2796 | * otherwise schedules the function for delayed execution. |
2797 | * |
2798 | * Returns: 0 - function was executed |
2799 | * 1 - function was scheduled for execution |
2800 | */ |
2801 | int execute_in_process_context(work_func_t fn, struct execute_work *ew) |
2802 | { |
2803 | if (!in_interrupt()) { |
2804 | fn(&ew->work); |
2805 | return 0; |
2806 | } |
2807 | |
2808 | INIT_WORK(&ew->work, fn); |
2809 | schedule_work(&ew->work); |
2810 | |
2811 | return 1; |
2812 | } |
2813 | EXPORT_SYMBOL_GPL(execute_in_process_context); |
2814 | |
2815 | int keventd_up(void) |
2816 | { |
2817 | return system_wq != NULL; |
2818 | } |
2819 | |
2820 | static int alloc_cwqs(struct workqueue_struct *wq) |
2821 | { |
2822 | /* |
2823 | * cwqs are forced aligned according to WORK_STRUCT_FLAG_BITS. |
2824 | * Make sure that the alignment isn't lower than that of |
2825 | * unsigned long long. |
2826 | */ |
2827 | const size_t size = sizeof(struct cpu_workqueue_struct); |
2828 | const size_t align = max_t(size_t, 1 << WORK_STRUCT_FLAG_BITS, |
2829 | __alignof__(unsigned long long)); |
2830 | #ifdef CONFIG_SMP |
2831 | bool percpu = !(wq->flags & WQ_UNBOUND); |
2832 | #else |
2833 | bool percpu = false; |
2834 | #endif |
2835 | |
2836 | if (percpu) |
2837 | wq->cpu_wq.pcpu = __alloc_percpu(size, align); |
2838 | else { |
2839 | void *ptr; |
2840 | |
2841 | /* |
2842 | * Allocate enough room to align cwq and put an extra |
2843 | * pointer at the end pointing back to the originally |
2844 | * allocated pointer which will be used for free. |
2845 | */ |
2846 | ptr = kzalloc(size + align + sizeof(void *), GFP_KERNEL); |
2847 | if (ptr) { |
2848 | wq->cpu_wq.single = PTR_ALIGN(ptr, align); |
2849 | *(void **)(wq->cpu_wq.single + 1) = ptr; |
2850 | } |
2851 | } |
2852 | |
2853 | /* just in case, make sure it's actually aligned |
2854 | * - this is affected by PERCPU() alignment in vmlinux.lds.S |
2855 | */ |
2856 | BUG_ON(!IS_ALIGNED(wq->cpu_wq.v, align)); |
2857 | return wq->cpu_wq.v ? 0 : -ENOMEM; |
2858 | } |
2859 | |
2860 | static void free_cwqs(struct workqueue_struct *wq) |
2861 | { |
2862 | #ifdef CONFIG_SMP |
2863 | bool percpu = !(wq->flags & WQ_UNBOUND); |
2864 | #else |
2865 | bool percpu = false; |
2866 | #endif |
2867 | |
2868 | if (percpu) |
2869 | free_percpu(wq->cpu_wq.pcpu); |
2870 | else if (wq->cpu_wq.single) { |
2871 | /* the pointer to free is stored right after the cwq */ |
2872 | kfree(*(void **)(wq->cpu_wq.single + 1)); |
2873 | } |
2874 | } |
2875 | |
2876 | static int wq_clamp_max_active(int max_active, unsigned int flags, |
2877 | const char *name) |
2878 | { |
2879 | int lim = flags & WQ_UNBOUND ? WQ_UNBOUND_MAX_ACTIVE : WQ_MAX_ACTIVE; |
2880 | |
2881 | if (max_active < 1 || max_active > lim) |
2882 | printk(KERN_WARNING "workqueue: max_active %d requested for %s " |
2883 | "is out of range, clamping between %d and %d\n", |
2884 | max_active, name, 1, lim); |
2885 | |
2886 | return clamp_val(max_active, 1, lim); |
2887 | } |
2888 | |
2889 | struct workqueue_struct *__alloc_workqueue_key(const char *name, |
2890 | unsigned int flags, |
2891 | int max_active, |
2892 | struct lock_class_key *key, |
2893 | const char *lock_name) |
2894 | { |
2895 | struct workqueue_struct *wq; |
2896 | unsigned int cpu; |
2897 | |
2898 | /* |
2899 | * Workqueues which may be used during memory reclaim should |
2900 | * have a rescuer to guarantee forward progress. |
2901 | */ |
2902 | if (flags & WQ_MEM_RECLAIM) |
2903 | flags |= WQ_RESCUER; |
2904 | |
2905 | /* |
2906 | * Unbound workqueues aren't concurrency managed and should be |
2907 | * dispatched to workers immediately. |
2908 | */ |
2909 | if (flags & WQ_UNBOUND) |
2910 | flags |= WQ_HIGHPRI; |
2911 | |
2912 | max_active = max_active ?: WQ_DFL_ACTIVE; |
2913 | max_active = wq_clamp_max_active(max_active, flags, name); |
2914 | |
2915 | wq = kzalloc(sizeof(*wq), GFP_KERNEL); |
2916 | if (!wq) |
2917 | goto err; |
2918 | |
2919 | wq->flags = flags; |
2920 | wq->saved_max_active = max_active; |
2921 | mutex_init(&wq->flush_mutex); |
2922 | atomic_set(&wq->nr_cwqs_to_flush, 0); |
2923 | INIT_LIST_HEAD(&wq->flusher_queue); |
2924 | INIT_LIST_HEAD(&wq->flusher_overflow); |
2925 | |
2926 | wq->name = name; |
2927 | lockdep_init_map(&wq->lockdep_map, lock_name, key, 0); |
2928 | INIT_LIST_HEAD(&wq->list); |
2929 | |
2930 | if (alloc_cwqs(wq) < 0) |
2931 | goto err; |
2932 | |
2933 | for_each_cwq_cpu(cpu, wq) { |
2934 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
2935 | struct global_cwq *gcwq = get_gcwq(cpu); |
2936 | |
2937 | BUG_ON((unsigned long)cwq & WORK_STRUCT_FLAG_MASK); |
2938 | cwq->gcwq = gcwq; |
2939 | cwq->wq = wq; |
2940 | cwq->flush_color = -1; |
2941 | cwq->max_active = max_active; |
2942 | INIT_LIST_HEAD(&cwq->delayed_works); |
2943 | } |
2944 | |
2945 | if (flags & WQ_RESCUER) { |
2946 | struct worker *rescuer; |
2947 | |
2948 | if (!alloc_mayday_mask(&wq->mayday_mask, GFP_KERNEL)) |
2949 | goto err; |
2950 | |
2951 | wq->rescuer = rescuer = alloc_worker(); |
2952 | if (!rescuer) |
2953 | goto err; |
2954 | |
2955 | rescuer->task = kthread_create(rescuer_thread, wq, "%s", name); |
2956 | if (IS_ERR(rescuer->task)) |
2957 | goto err; |
2958 | |
2959 | rescuer->task->flags |= PF_THREAD_BOUND; |
2960 | wake_up_process(rescuer->task); |
2961 | } |
2962 | |
2963 | /* |
2964 | * workqueue_lock protects global freeze state and workqueues |
2965 | * list. Grab it, set max_active accordingly and add the new |
2966 | * workqueue to workqueues list. |
2967 | */ |
2968 | spin_lock(&workqueue_lock); |
2969 | |
2970 | if (workqueue_freezing && wq->flags & WQ_FREEZABLE) |
2971 | for_each_cwq_cpu(cpu, wq) |
2972 | get_cwq(cpu, wq)->max_active = 0; |
2973 | |
2974 | list_add(&wq->list, &workqueues); |
2975 | |
2976 | spin_unlock(&workqueue_lock); |
2977 | |
2978 | return wq; |
2979 | err: |
2980 | if (wq) { |
2981 | free_cwqs(wq); |
2982 | free_mayday_mask(wq->mayday_mask); |
2983 | kfree(wq->rescuer); |
2984 | kfree(wq); |
2985 | } |
2986 | return NULL; |
2987 | } |
2988 | EXPORT_SYMBOL_GPL(__alloc_workqueue_key); |
2989 | |
2990 | /** |
2991 | * destroy_workqueue - safely terminate a workqueue |
2992 | * @wq: target workqueue |
2993 | * |
2994 | * Safely destroy a workqueue. All work currently pending will be done first. |
2995 | */ |
2996 | void destroy_workqueue(struct workqueue_struct *wq) |
2997 | { |
2998 | unsigned int flush_cnt = 0; |
2999 | unsigned int cpu; |
3000 | |
3001 | /* |
3002 | * Mark @wq dying and drain all pending works. Once WQ_DYING is |
3003 | * set, only chain queueing is allowed. IOW, only currently |
3004 | * pending or running work items on @wq can queue further work |
3005 | * items on it. @wq is flushed repeatedly until it becomes empty. |
3006 | * The number of flushing is detemined by the depth of chaining and |
3007 | * should be relatively short. Whine if it takes too long. |
3008 | */ |
3009 | wq->flags |= WQ_DYING; |
3010 | reflush: |
3011 | flush_workqueue(wq); |
3012 | |
3013 | for_each_cwq_cpu(cpu, wq) { |
3014 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3015 | |
3016 | if (!cwq->nr_active && list_empty(&cwq->delayed_works)) |
3017 | continue; |
3018 | |
3019 | if (++flush_cnt == 10 || |
3020 | (flush_cnt % 100 == 0 && flush_cnt <= 1000)) |
3021 | printk(KERN_WARNING "workqueue %s: flush on " |
3022 | "destruction isn't complete after %u tries\n", |
3023 | wq->name, flush_cnt); |
3024 | goto reflush; |
3025 | } |
3026 | |
3027 | /* |
3028 | * wq list is used to freeze wq, remove from list after |
3029 | * flushing is complete in case freeze races us. |
3030 | */ |
3031 | spin_lock(&workqueue_lock); |
3032 | list_del(&wq->list); |
3033 | spin_unlock(&workqueue_lock); |
3034 | |
3035 | /* sanity check */ |
3036 | for_each_cwq_cpu(cpu, wq) { |
3037 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3038 | int i; |
3039 | |
3040 | for (i = 0; i < WORK_NR_COLORS; i++) |
3041 | BUG_ON(cwq->nr_in_flight[i]); |
3042 | BUG_ON(cwq->nr_active); |
3043 | BUG_ON(!list_empty(&cwq->delayed_works)); |
3044 | } |
3045 | |
3046 | if (wq->flags & WQ_RESCUER) { |
3047 | kthread_stop(wq->rescuer->task); |
3048 | free_mayday_mask(wq->mayday_mask); |
3049 | kfree(wq->rescuer); |
3050 | } |
3051 | |
3052 | free_cwqs(wq); |
3053 | kfree(wq); |
3054 | } |
3055 | EXPORT_SYMBOL_GPL(destroy_workqueue); |
3056 | |
3057 | /** |
3058 | * workqueue_set_max_active - adjust max_active of a workqueue |
3059 | * @wq: target workqueue |
3060 | * @max_active: new max_active value. |
3061 | * |
3062 | * Set max_active of @wq to @max_active. |
3063 | * |
3064 | * CONTEXT: |
3065 | * Don't call from IRQ context. |
3066 | */ |
3067 | void workqueue_set_max_active(struct workqueue_struct *wq, int max_active) |
3068 | { |
3069 | unsigned int cpu; |
3070 | |
3071 | max_active = wq_clamp_max_active(max_active, wq->flags, wq->name); |
3072 | |
3073 | spin_lock(&workqueue_lock); |
3074 | |
3075 | wq->saved_max_active = max_active; |
3076 | |
3077 | for_each_cwq_cpu(cpu, wq) { |
3078 | struct global_cwq *gcwq = get_gcwq(cpu); |
3079 | |
3080 | spin_lock_irq(&gcwq->lock); |
3081 | |
3082 | if (!(wq->flags & WQ_FREEZABLE) || |
3083 | !(gcwq->flags & GCWQ_FREEZING)) |
3084 | get_cwq(gcwq->cpu, wq)->max_active = max_active; |
3085 | |
3086 | spin_unlock_irq(&gcwq->lock); |
3087 | } |
3088 | |
3089 | spin_unlock(&workqueue_lock); |
3090 | } |
3091 | EXPORT_SYMBOL_GPL(workqueue_set_max_active); |
3092 | |
3093 | /** |
3094 | * workqueue_congested - test whether a workqueue is congested |
3095 | * @cpu: CPU in question |
3096 | * @wq: target workqueue |
3097 | * |
3098 | * Test whether @wq's cpu workqueue for @cpu is congested. There is |
3099 | * no synchronization around this function and the test result is |
3100 | * unreliable and only useful as advisory hints or for debugging. |
3101 | * |
3102 | * RETURNS: |
3103 | * %true if congested, %false otherwise. |
3104 | */ |
3105 | bool workqueue_congested(unsigned int cpu, struct workqueue_struct *wq) |
3106 | { |
3107 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3108 | |
3109 | return !list_empty(&cwq->delayed_works); |
3110 | } |
3111 | EXPORT_SYMBOL_GPL(workqueue_congested); |
3112 | |
3113 | /** |
3114 | * work_cpu - return the last known associated cpu for @work |
3115 | * @work: the work of interest |
3116 | * |
3117 | * RETURNS: |
3118 | * CPU number if @work was ever queued. WORK_CPU_NONE otherwise. |
3119 | */ |
3120 | unsigned int work_cpu(struct work_struct *work) |
3121 | { |
3122 | struct global_cwq *gcwq = get_work_gcwq(work); |
3123 | |
3124 | return gcwq ? gcwq->cpu : WORK_CPU_NONE; |
3125 | } |
3126 | EXPORT_SYMBOL_GPL(work_cpu); |
3127 | |
3128 | /** |
3129 | * work_busy - test whether a work is currently pending or running |
3130 | * @work: the work to be tested |
3131 | * |
3132 | * Test whether @work is currently pending or running. There is no |
3133 | * synchronization around this function and the test result is |
3134 | * unreliable and only useful as advisory hints or for debugging. |
3135 | * Especially for reentrant wqs, the pending state might hide the |
3136 | * running state. |
3137 | * |
3138 | * RETURNS: |
3139 | * OR'd bitmask of WORK_BUSY_* bits. |
3140 | */ |
3141 | unsigned int work_busy(struct work_struct *work) |
3142 | { |
3143 | struct global_cwq *gcwq = get_work_gcwq(work); |
3144 | unsigned long flags; |
3145 | unsigned int ret = 0; |
3146 | |
3147 | if (!gcwq) |
3148 | return false; |
3149 | |
3150 | spin_lock_irqsave(&gcwq->lock, flags); |
3151 | |
3152 | if (work_pending(work)) |
3153 | ret |= WORK_BUSY_PENDING; |
3154 | if (find_worker_executing_work(gcwq, work)) |
3155 | ret |= WORK_BUSY_RUNNING; |
3156 | |
3157 | spin_unlock_irqrestore(&gcwq->lock, flags); |
3158 | |
3159 | return ret; |
3160 | } |
3161 | EXPORT_SYMBOL_GPL(work_busy); |
3162 | |
3163 | /* |
3164 | * CPU hotplug. |
3165 | * |
3166 | * There are two challenges in supporting CPU hotplug. Firstly, there |
3167 | * are a lot of assumptions on strong associations among work, cwq and |
3168 | * gcwq which make migrating pending and scheduled works very |
3169 | * difficult to implement without impacting hot paths. Secondly, |
3170 | * gcwqs serve mix of short, long and very long running works making |
3171 | * blocked draining impractical. |
3172 | * |
3173 | * This is solved by allowing a gcwq to be detached from CPU, running |
3174 | * it with unbound (rogue) workers and allowing it to be reattached |
3175 | * later if the cpu comes back online. A separate thread is created |
3176 | * to govern a gcwq in such state and is called the trustee of the |
3177 | * gcwq. |
3178 | * |
3179 | * Trustee states and their descriptions. |
3180 | * |
3181 | * START Command state used on startup. On CPU_DOWN_PREPARE, a |
3182 | * new trustee is started with this state. |
3183 | * |
3184 | * IN_CHARGE Once started, trustee will enter this state after |
3185 | * assuming the manager role and making all existing |
3186 | * workers rogue. DOWN_PREPARE waits for trustee to |
3187 | * enter this state. After reaching IN_CHARGE, trustee |
3188 | * tries to execute the pending worklist until it's empty |
3189 | * and the state is set to BUTCHER, or the state is set |
3190 | * to RELEASE. |
3191 | * |
3192 | * BUTCHER Command state which is set by the cpu callback after |
3193 | * the cpu has went down. Once this state is set trustee |
3194 | * knows that there will be no new works on the worklist |
3195 | * and once the worklist is empty it can proceed to |
3196 | * killing idle workers. |
3197 | * |
3198 | * RELEASE Command state which is set by the cpu callback if the |
3199 | * cpu down has been canceled or it has come online |
3200 | * again. After recognizing this state, trustee stops |
3201 | * trying to drain or butcher and clears ROGUE, rebinds |
3202 | * all remaining workers back to the cpu and releases |
3203 | * manager role. |
3204 | * |
3205 | * DONE Trustee will enter this state after BUTCHER or RELEASE |
3206 | * is complete. |
3207 | * |
3208 | * trustee CPU draining |
3209 | * took over down complete |
3210 | * START -----------> IN_CHARGE -----------> BUTCHER -----------> DONE |
3211 | * | | ^ |
3212 | * | CPU is back online v return workers | |
3213 | * ----------------> RELEASE -------------- |
3214 | */ |
3215 | |
3216 | /** |
3217 | * trustee_wait_event_timeout - timed event wait for trustee |
3218 | * @cond: condition to wait for |
3219 | * @timeout: timeout in jiffies |
3220 | * |
3221 | * wait_event_timeout() for trustee to use. Handles locking and |
3222 | * checks for RELEASE request. |
3223 | * |
3224 | * CONTEXT: |
3225 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
3226 | * multiple times. To be used by trustee. |
3227 | * |
3228 | * RETURNS: |
3229 | * Positive indicating left time if @cond is satisfied, 0 if timed |
3230 | * out, -1 if canceled. |
3231 | */ |
3232 | #define trustee_wait_event_timeout(cond, timeout) ({ \ |
3233 | long __ret = (timeout); \ |
3234 | while (!((cond) || (gcwq->trustee_state == TRUSTEE_RELEASE)) && \ |
3235 | __ret) { \ |
3236 | spin_unlock_irq(&gcwq->lock); \ |
3237 | __wait_event_timeout(gcwq->trustee_wait, (cond) || \ |
3238 | (gcwq->trustee_state == TRUSTEE_RELEASE), \ |
3239 | __ret); \ |
3240 | spin_lock_irq(&gcwq->lock); \ |
3241 | } \ |
3242 | gcwq->trustee_state == TRUSTEE_RELEASE ? -1 : (__ret); \ |
3243 | }) |
3244 | |
3245 | /** |
3246 | * trustee_wait_event - event wait for trustee |
3247 | * @cond: condition to wait for |
3248 | * |
3249 | * wait_event() for trustee to use. Automatically handles locking and |
3250 | * checks for CANCEL request. |
3251 | * |
3252 | * CONTEXT: |
3253 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
3254 | * multiple times. To be used by trustee. |
3255 | * |
3256 | * RETURNS: |
3257 | * 0 if @cond is satisfied, -1 if canceled. |
3258 | */ |
3259 | #define trustee_wait_event(cond) ({ \ |
3260 | long __ret1; \ |
3261 | __ret1 = trustee_wait_event_timeout(cond, MAX_SCHEDULE_TIMEOUT);\ |
3262 | __ret1 < 0 ? -1 : 0; \ |
3263 | }) |
3264 | |
3265 | static int __cpuinit trustee_thread(void *__gcwq) |
3266 | { |
3267 | struct global_cwq *gcwq = __gcwq; |
3268 | struct worker *worker; |
3269 | struct work_struct *work; |
3270 | struct hlist_node *pos; |
3271 | long rc; |
3272 | int i; |
3273 | |
3274 | BUG_ON(gcwq->cpu != smp_processor_id()); |
3275 | |
3276 | spin_lock_irq(&gcwq->lock); |
3277 | /* |
3278 | * Claim the manager position and make all workers rogue. |
3279 | * Trustee must be bound to the target cpu and can't be |
3280 | * cancelled. |
3281 | */ |
3282 | BUG_ON(gcwq->cpu != smp_processor_id()); |
3283 | rc = trustee_wait_event(!(gcwq->flags & GCWQ_MANAGING_WORKERS)); |
3284 | BUG_ON(rc < 0); |
3285 | |
3286 | gcwq->flags |= GCWQ_MANAGING_WORKERS; |
3287 | |
3288 | list_for_each_entry(worker, &gcwq->idle_list, entry) |
3289 | worker->flags |= WORKER_ROGUE; |
3290 | |
3291 | for_each_busy_worker(worker, i, pos, gcwq) |
3292 | worker->flags |= WORKER_ROGUE; |
3293 | |
3294 | /* |
3295 | * Call schedule() so that we cross rq->lock and thus can |
3296 | * guarantee sched callbacks see the rogue flag. This is |
3297 | * necessary as scheduler callbacks may be invoked from other |
3298 | * cpus. |
3299 | */ |
3300 | spin_unlock_irq(&gcwq->lock); |
3301 | schedule(); |
3302 | spin_lock_irq(&gcwq->lock); |
3303 | |
3304 | /* |
3305 | * Sched callbacks are disabled now. Zap nr_running. After |
3306 | * this, nr_running stays zero and need_more_worker() and |
3307 | * keep_working() are always true as long as the worklist is |
3308 | * not empty. |
3309 | */ |
3310 | atomic_set(get_gcwq_nr_running(gcwq->cpu), 0); |
3311 | |
3312 | spin_unlock_irq(&gcwq->lock); |
3313 | del_timer_sync(&gcwq->idle_timer); |
3314 | spin_lock_irq(&gcwq->lock); |
3315 | |
3316 | /* |
3317 | * We're now in charge. Notify and proceed to drain. We need |
3318 | * to keep the gcwq running during the whole CPU down |
3319 | * procedure as other cpu hotunplug callbacks may need to |
3320 | * flush currently running tasks. |
3321 | */ |
3322 | gcwq->trustee_state = TRUSTEE_IN_CHARGE; |
3323 | wake_up_all(&gcwq->trustee_wait); |
3324 | |
3325 | /* |
3326 | * The original cpu is in the process of dying and may go away |
3327 | * anytime now. When that happens, we and all workers would |
3328 | * be migrated to other cpus. Try draining any left work. We |
3329 | * want to get it over with ASAP - spam rescuers, wake up as |
3330 | * many idlers as necessary and create new ones till the |
3331 | * worklist is empty. Note that if the gcwq is frozen, there |
3332 | * may be frozen works in freezable cwqs. Don't declare |
3333 | * completion while frozen. |
3334 | */ |
3335 | while (gcwq->nr_workers != gcwq->nr_idle || |
3336 | gcwq->flags & GCWQ_FREEZING || |
3337 | gcwq->trustee_state == TRUSTEE_IN_CHARGE) { |
3338 | int nr_works = 0; |
3339 | |
3340 | list_for_each_entry(work, &gcwq->worklist, entry) { |
3341 | send_mayday(work); |
3342 | nr_works++; |
3343 | } |
3344 | |
3345 | list_for_each_entry(worker, &gcwq->idle_list, entry) { |
3346 | if (!nr_works--) |
3347 | break; |
3348 | wake_up_process(worker->task); |
3349 | } |
3350 | |
3351 | if (need_to_create_worker(gcwq)) { |
3352 | spin_unlock_irq(&gcwq->lock); |
3353 | worker = create_worker(gcwq, false); |
3354 | spin_lock_irq(&gcwq->lock); |
3355 | if (worker) { |
3356 | worker->flags |= WORKER_ROGUE; |
3357 | start_worker(worker); |
3358 | } |
3359 | } |
3360 | |
3361 | /* give a breather */ |
3362 | if (trustee_wait_event_timeout(false, TRUSTEE_COOLDOWN) < 0) |
3363 | break; |
3364 | } |
3365 | |
3366 | /* |
3367 | * Either all works have been scheduled and cpu is down, or |
3368 | * cpu down has already been canceled. Wait for and butcher |
3369 | * all workers till we're canceled. |
3370 | */ |
3371 | do { |
3372 | rc = trustee_wait_event(!list_empty(&gcwq->idle_list)); |
3373 | while (!list_empty(&gcwq->idle_list)) |
3374 | destroy_worker(list_first_entry(&gcwq->idle_list, |
3375 | struct worker, entry)); |
3376 | } while (gcwq->nr_workers && rc >= 0); |
3377 | |
3378 | /* |
3379 | * At this point, either draining has completed and no worker |
3380 | * is left, or cpu down has been canceled or the cpu is being |
3381 | * brought back up. There shouldn't be any idle one left. |
3382 | * Tell the remaining busy ones to rebind once it finishes the |
3383 | * currently scheduled works by scheduling the rebind_work. |
3384 | */ |
3385 | WARN_ON(!list_empty(&gcwq->idle_list)); |
3386 | |
3387 | for_each_busy_worker(worker, i, pos, gcwq) { |
3388 | struct work_struct *rebind_work = &worker->rebind_work; |
3389 | |
3390 | /* |
3391 | * Rebind_work may race with future cpu hotplug |
3392 | * operations. Use a separate flag to mark that |
3393 | * rebinding is scheduled. |
3394 | */ |
3395 | worker->flags |= WORKER_REBIND; |
3396 | worker->flags &= ~WORKER_ROGUE; |
3397 | |
3398 | /* queue rebind_work, wq doesn't matter, use the default one */ |
3399 | if (test_and_set_bit(WORK_STRUCT_PENDING_BIT, |
3400 | work_data_bits(rebind_work))) |
3401 | continue; |
3402 | |
3403 | debug_work_activate(rebind_work); |
3404 | insert_work(get_cwq(gcwq->cpu, system_wq), rebind_work, |
3405 | worker->scheduled.next, |
3406 | work_color_to_flags(WORK_NO_COLOR)); |
3407 | } |
3408 | |
3409 | /* relinquish manager role */ |
3410 | gcwq->flags &= ~GCWQ_MANAGING_WORKERS; |
3411 | |
3412 | /* notify completion */ |
3413 | gcwq->trustee = NULL; |
3414 | gcwq->trustee_state = TRUSTEE_DONE; |
3415 | wake_up_all(&gcwq->trustee_wait); |
3416 | spin_unlock_irq(&gcwq->lock); |
3417 | return 0; |
3418 | } |
3419 | |
3420 | /** |
3421 | * wait_trustee_state - wait for trustee to enter the specified state |
3422 | * @gcwq: gcwq the trustee of interest belongs to |
3423 | * @state: target state to wait for |
3424 | * |
3425 | * Wait for the trustee to reach @state. DONE is already matched. |
3426 | * |
3427 | * CONTEXT: |
3428 | * spin_lock_irq(gcwq->lock) which may be released and regrabbed |
3429 | * multiple times. To be used by cpu_callback. |
3430 | */ |
3431 | static void __cpuinit wait_trustee_state(struct global_cwq *gcwq, int state) |
3432 | __releases(&gcwq->lock) |
3433 | __acquires(&gcwq->lock) |
3434 | { |
3435 | if (!(gcwq->trustee_state == state || |
3436 | gcwq->trustee_state == TRUSTEE_DONE)) { |
3437 | spin_unlock_irq(&gcwq->lock); |
3438 | __wait_event(gcwq->trustee_wait, |
3439 | gcwq->trustee_state == state || |
3440 | gcwq->trustee_state == TRUSTEE_DONE); |
3441 | spin_lock_irq(&gcwq->lock); |
3442 | } |
3443 | } |
3444 | |
3445 | static int __devinit workqueue_cpu_callback(struct notifier_block *nfb, |
3446 | unsigned long action, |
3447 | void *hcpu) |
3448 | { |
3449 | unsigned int cpu = (unsigned long)hcpu; |
3450 | struct global_cwq *gcwq = get_gcwq(cpu); |
3451 | struct task_struct *new_trustee = NULL; |
3452 | struct worker *uninitialized_var(new_worker); |
3453 | unsigned long flags; |
3454 | |
3455 | action &= ~CPU_TASKS_FROZEN; |
3456 | |
3457 | switch (action) { |
3458 | case CPU_DOWN_PREPARE: |
3459 | new_trustee = kthread_create(trustee_thread, gcwq, |
3460 | "workqueue_trustee/%d\n", cpu); |
3461 | if (IS_ERR(new_trustee)) |
3462 | return notifier_from_errno(PTR_ERR(new_trustee)); |
3463 | kthread_bind(new_trustee, cpu); |
3464 | /* fall through */ |
3465 | case CPU_UP_PREPARE: |
3466 | BUG_ON(gcwq->first_idle); |
3467 | new_worker = create_worker(gcwq, false); |
3468 | if (!new_worker) { |
3469 | if (new_trustee) |
3470 | kthread_stop(new_trustee); |
3471 | return NOTIFY_BAD; |
3472 | } |
3473 | } |
3474 | |
3475 | /* some are called w/ irq disabled, don't disturb irq status */ |
3476 | spin_lock_irqsave(&gcwq->lock, flags); |
3477 | |
3478 | switch (action) { |
3479 | case CPU_DOWN_PREPARE: |
3480 | /* initialize trustee and tell it to acquire the gcwq */ |
3481 | BUG_ON(gcwq->trustee || gcwq->trustee_state != TRUSTEE_DONE); |
3482 | gcwq->trustee = new_trustee; |
3483 | gcwq->trustee_state = TRUSTEE_START; |
3484 | wake_up_process(gcwq->trustee); |
3485 | wait_trustee_state(gcwq, TRUSTEE_IN_CHARGE); |
3486 | /* fall through */ |
3487 | case CPU_UP_PREPARE: |
3488 | BUG_ON(gcwq->first_idle); |
3489 | gcwq->first_idle = new_worker; |
3490 | break; |
3491 | |
3492 | case CPU_DYING: |
3493 | /* |
3494 | * Before this, the trustee and all workers except for |
3495 | * the ones which are still executing works from |
3496 | * before the last CPU down must be on the cpu. After |
3497 | * this, they'll all be diasporas. |
3498 | */ |
3499 | gcwq->flags |= GCWQ_DISASSOCIATED; |
3500 | break; |
3501 | |
3502 | case CPU_POST_DEAD: |
3503 | gcwq->trustee_state = TRUSTEE_BUTCHER; |
3504 | /* fall through */ |
3505 | case CPU_UP_CANCELED: |
3506 | destroy_worker(gcwq->first_idle); |
3507 | gcwq->first_idle = NULL; |
3508 | break; |
3509 | |
3510 | case CPU_DOWN_FAILED: |
3511 | case CPU_ONLINE: |
3512 | gcwq->flags &= ~GCWQ_DISASSOCIATED; |
3513 | if (gcwq->trustee_state != TRUSTEE_DONE) { |
3514 | gcwq->trustee_state = TRUSTEE_RELEASE; |
3515 | wake_up_process(gcwq->trustee); |
3516 | wait_trustee_state(gcwq, TRUSTEE_DONE); |
3517 | } |
3518 | |
3519 | /* |
3520 | * Trustee is done and there might be no worker left. |
3521 | * Put the first_idle in and request a real manager to |
3522 | * take a look. |
3523 | */ |
3524 | spin_unlock_irq(&gcwq->lock); |
3525 | kthread_bind(gcwq->first_idle->task, cpu); |
3526 | spin_lock_irq(&gcwq->lock); |
3527 | gcwq->flags |= GCWQ_MANAGE_WORKERS; |
3528 | start_worker(gcwq->first_idle); |
3529 | gcwq->first_idle = NULL; |
3530 | break; |
3531 | } |
3532 | |
3533 | spin_unlock_irqrestore(&gcwq->lock, flags); |
3534 | |
3535 | return notifier_from_errno(0); |
3536 | } |
3537 | |
3538 | #ifdef CONFIG_SMP |
3539 | |
3540 | struct work_for_cpu { |
3541 | struct completion completion; |
3542 | long (*fn)(void *); |
3543 | void *arg; |
3544 | long ret; |
3545 | }; |
3546 | |
3547 | static int do_work_for_cpu(void *_wfc) |
3548 | { |
3549 | struct work_for_cpu *wfc = _wfc; |
3550 | wfc->ret = wfc->fn(wfc->arg); |
3551 | complete(&wfc->completion); |
3552 | return 0; |
3553 | } |
3554 | |
3555 | /** |
3556 | * work_on_cpu - run a function in user context on a particular cpu |
3557 | * @cpu: the cpu to run on |
3558 | * @fn: the function to run |
3559 | * @arg: the function arg |
3560 | * |
3561 | * This will return the value @fn returns. |
3562 | * It is up to the caller to ensure that the cpu doesn't go offline. |
3563 | * The caller must not hold any locks which would prevent @fn from completing. |
3564 | */ |
3565 | long work_on_cpu(unsigned int cpu, long (*fn)(void *), void *arg) |
3566 | { |
3567 | struct task_struct *sub_thread; |
3568 | struct work_for_cpu wfc = { |
3569 | .completion = COMPLETION_INITIALIZER_ONSTACK(wfc.completion), |
3570 | .fn = fn, |
3571 | .arg = arg, |
3572 | }; |
3573 | |
3574 | sub_thread = kthread_create(do_work_for_cpu, &wfc, "work_for_cpu"); |
3575 | if (IS_ERR(sub_thread)) |
3576 | return PTR_ERR(sub_thread); |
3577 | kthread_bind(sub_thread, cpu); |
3578 | wake_up_process(sub_thread); |
3579 | wait_for_completion(&wfc.completion); |
3580 | return wfc.ret; |
3581 | } |
3582 | EXPORT_SYMBOL_GPL(work_on_cpu); |
3583 | #endif /* CONFIG_SMP */ |
3584 | |
3585 | #ifdef CONFIG_FREEZER |
3586 | |
3587 | /** |
3588 | * freeze_workqueues_begin - begin freezing workqueues |
3589 | * |
3590 | * Start freezing workqueues. After this function returns, all freezable |
3591 | * workqueues will queue new works to their frozen_works list instead of |
3592 | * gcwq->worklist. |
3593 | * |
3594 | * CONTEXT: |
3595 | * Grabs and releases workqueue_lock and gcwq->lock's. |
3596 | */ |
3597 | void freeze_workqueues_begin(void) |
3598 | { |
3599 | unsigned int cpu; |
3600 | |
3601 | spin_lock(&workqueue_lock); |
3602 | |
3603 | BUG_ON(workqueue_freezing); |
3604 | workqueue_freezing = true; |
3605 | |
3606 | for_each_gcwq_cpu(cpu) { |
3607 | struct global_cwq *gcwq = get_gcwq(cpu); |
3608 | struct workqueue_struct *wq; |
3609 | |
3610 | spin_lock_irq(&gcwq->lock); |
3611 | |
3612 | BUG_ON(gcwq->flags & GCWQ_FREEZING); |
3613 | gcwq->flags |= GCWQ_FREEZING; |
3614 | |
3615 | list_for_each_entry(wq, &workqueues, list) { |
3616 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3617 | |
3618 | if (cwq && wq->flags & WQ_FREEZABLE) |
3619 | cwq->max_active = 0; |
3620 | } |
3621 | |
3622 | spin_unlock_irq(&gcwq->lock); |
3623 | } |
3624 | |
3625 | spin_unlock(&workqueue_lock); |
3626 | } |
3627 | |
3628 | /** |
3629 | * freeze_workqueues_busy - are freezable workqueues still busy? |
3630 | * |
3631 | * Check whether freezing is complete. This function must be called |
3632 | * between freeze_workqueues_begin() and thaw_workqueues(). |
3633 | * |
3634 | * CONTEXT: |
3635 | * Grabs and releases workqueue_lock. |
3636 | * |
3637 | * RETURNS: |
3638 | * %true if some freezable workqueues are still busy. %false if freezing |
3639 | * is complete. |
3640 | */ |
3641 | bool freeze_workqueues_busy(void) |
3642 | { |
3643 | unsigned int cpu; |
3644 | bool busy = false; |
3645 | |
3646 | spin_lock(&workqueue_lock); |
3647 | |
3648 | BUG_ON(!workqueue_freezing); |
3649 | |
3650 | for_each_gcwq_cpu(cpu) { |
3651 | struct workqueue_struct *wq; |
3652 | /* |
3653 | * nr_active is monotonically decreasing. It's safe |
3654 | * to peek without lock. |
3655 | */ |
3656 | list_for_each_entry(wq, &workqueues, list) { |
3657 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3658 | |
3659 | if (!cwq || !(wq->flags & WQ_FREEZABLE)) |
3660 | continue; |
3661 | |
3662 | BUG_ON(cwq->nr_active < 0); |
3663 | if (cwq->nr_active) { |
3664 | busy = true; |
3665 | goto out_unlock; |
3666 | } |
3667 | } |
3668 | } |
3669 | out_unlock: |
3670 | spin_unlock(&workqueue_lock); |
3671 | return busy; |
3672 | } |
3673 | |
3674 | /** |
3675 | * thaw_workqueues - thaw workqueues |
3676 | * |
3677 | * Thaw workqueues. Normal queueing is restored and all collected |
3678 | * frozen works are transferred to their respective gcwq worklists. |
3679 | * |
3680 | * CONTEXT: |
3681 | * Grabs and releases workqueue_lock and gcwq->lock's. |
3682 | */ |
3683 | void thaw_workqueues(void) |
3684 | { |
3685 | unsigned int cpu; |
3686 | |
3687 | spin_lock(&workqueue_lock); |
3688 | |
3689 | if (!workqueue_freezing) |
3690 | goto out_unlock; |
3691 | |
3692 | for_each_gcwq_cpu(cpu) { |
3693 | struct global_cwq *gcwq = get_gcwq(cpu); |
3694 | struct workqueue_struct *wq; |
3695 | |
3696 | spin_lock_irq(&gcwq->lock); |
3697 | |
3698 | BUG_ON(!(gcwq->flags & GCWQ_FREEZING)); |
3699 | gcwq->flags &= ~GCWQ_FREEZING; |
3700 | |
3701 | list_for_each_entry(wq, &workqueues, list) { |
3702 | struct cpu_workqueue_struct *cwq = get_cwq(cpu, wq); |
3703 | |
3704 | if (!cwq || !(wq->flags & WQ_FREEZABLE)) |
3705 | continue; |
3706 | |
3707 | /* restore max_active and repopulate worklist */ |
3708 | cwq->max_active = wq->saved_max_active; |
3709 | |
3710 | while (!list_empty(&cwq->delayed_works) && |
3711 | cwq->nr_active < cwq->max_active) |
3712 | cwq_activate_first_delayed(cwq); |
3713 | } |
3714 | |
3715 | wake_up_worker(gcwq); |
3716 | |
3717 | spin_unlock_irq(&gcwq->lock); |
3718 | } |
3719 | |
3720 | workqueue_freezing = false; |
3721 | out_unlock: |
3722 | spin_unlock(&workqueue_lock); |
3723 | } |
3724 | #endif /* CONFIG_FREEZER */ |
3725 | |
3726 | static int __init init_workqueues(void) |
3727 | { |
3728 | unsigned int cpu; |
3729 | int i; |
3730 | |
3731 | cpu_notifier(workqueue_cpu_callback, CPU_PRI_WORKQUEUE); |
3732 | |
3733 | /* initialize gcwqs */ |
3734 | for_each_gcwq_cpu(cpu) { |
3735 | struct global_cwq *gcwq = get_gcwq(cpu); |
3736 | |
3737 | spin_lock_init(&gcwq->lock); |
3738 | INIT_LIST_HEAD(&gcwq->worklist); |
3739 | gcwq->cpu = cpu; |
3740 | gcwq->flags |= GCWQ_DISASSOCIATED; |
3741 | |
3742 | INIT_LIST_HEAD(&gcwq->idle_list); |
3743 | for (i = 0; i < BUSY_WORKER_HASH_SIZE; i++) |
3744 | INIT_HLIST_HEAD(&gcwq->busy_hash[i]); |
3745 | |
3746 | init_timer_deferrable(&gcwq->idle_timer); |
3747 | gcwq->idle_timer.function = idle_worker_timeout; |
3748 | gcwq->idle_timer.data = (unsigned long)gcwq; |
3749 | |
3750 | setup_timer(&gcwq->mayday_timer, gcwq_mayday_timeout, |
3751 | (unsigned long)gcwq); |
3752 | |
3753 | ida_init(&gcwq->worker_ida); |
3754 | |
3755 | gcwq->trustee_state = TRUSTEE_DONE; |
3756 | init_waitqueue_head(&gcwq->trustee_wait); |
3757 | } |
3758 | |
3759 | /* create the initial worker */ |
3760 | for_each_online_gcwq_cpu(cpu) { |
3761 | struct global_cwq *gcwq = get_gcwq(cpu); |
3762 | struct worker *worker; |
3763 | |
3764 | if (cpu != WORK_CPU_UNBOUND) |
3765 | gcwq->flags &= ~GCWQ_DISASSOCIATED; |
3766 | worker = create_worker(gcwq, true); |
3767 | BUG_ON(!worker); |
3768 | spin_lock_irq(&gcwq->lock); |
3769 | start_worker(worker); |
3770 | spin_unlock_irq(&gcwq->lock); |
3771 | } |
3772 | |
3773 | system_wq = alloc_workqueue("events", 0, 0); |
3774 | system_long_wq = alloc_workqueue("events_long", 0, 0); |
3775 | system_nrt_wq = alloc_workqueue("events_nrt", WQ_NON_REENTRANT, 0); |
3776 | system_unbound_wq = alloc_workqueue("events_unbound", WQ_UNBOUND, |
3777 | WQ_UNBOUND_MAX_ACTIVE); |
3778 | BUG_ON(!system_wq || !system_long_wq || !system_nrt_wq || |
3779 | !system_unbound_wq); |
3780 | return 0; |
3781 | } |
3782 | early_initcall(init_workqueues); |
3783 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9