Commit | Line | Data |
---|---|---|
07fe7cb7 DH |
1 | /* Worker thread pool for slow items, such as filesystem lookups or mkdirs |
2 | * | |
3 | * Copyright (C) 2008 Red Hat, Inc. All Rights Reserved. | |
4 | * Written by David Howells (dhowells@redhat.com) | |
5 | * | |
6 | * This program is free software; you can redistribute it and/or | |
7 | * modify it under the terms of the GNU General Public Licence | |
8 | * as published by the Free Software Foundation; either version | |
9 | * 2 of the Licence, or (at your option) any later version. | |
8f0aa2f2 DH |
10 | * |
11 | * See Documentation/slow-work.txt | |
07fe7cb7 DH |
12 | */ |
13 | ||
14 | #include <linux/module.h> | |
15 | #include <linux/slow-work.h> | |
16 | #include <linux/kthread.h> | |
17 | #include <linux/freezer.h> | |
18 | #include <linux/wait.h> | |
07fe7cb7 | 19 | |
109d9272 DH |
20 | #define SLOW_WORK_CULL_TIMEOUT (5 * HZ) /* cull threads 5s after running out of |
21 | * things to do */ | |
22 | #define SLOW_WORK_OOM_TIMEOUT (5 * HZ) /* can't start new threads for 5s after | |
23 | * OOM */ | |
24 | ||
3d7a641e DH |
25 | #define SLOW_WORK_THREAD_LIMIT 255 /* abs maximum number of slow-work threads */ |
26 | ||
109d9272 DH |
27 | static void slow_work_cull_timeout(unsigned long); |
28 | static void slow_work_oom_timeout(unsigned long); | |
29 | ||
12e22c5e | 30 | #ifdef CONFIG_SYSCTL |
8d65af78 | 31 | static int slow_work_min_threads_sysctl(struct ctl_table *, int, |
12e22c5e DH |
32 | void __user *, size_t *, loff_t *); |
33 | ||
8d65af78 | 34 | static int slow_work_max_threads_sysctl(struct ctl_table *, int , |
12e22c5e DH |
35 | void __user *, size_t *, loff_t *); |
36 | #endif | |
37 | ||
07fe7cb7 DH |
38 | /* |
39 | * The pool of threads has at least min threads in it as long as someone is | |
40 | * using the facility, and may have as many as max. | |
41 | * | |
42 | * A portion of the pool may be processing very slow operations. | |
43 | */ | |
44 | static unsigned slow_work_min_threads = 2; | |
45 | static unsigned slow_work_max_threads = 4; | |
46 | static unsigned vslow_work_proportion = 50; /* % of threads that may process | |
47 | * very slow work */ | |
12e22c5e DH |
48 | |
49 | #ifdef CONFIG_SYSCTL | |
50 | static const int slow_work_min_min_threads = 2; | |
3d7a641e | 51 | static int slow_work_max_max_threads = SLOW_WORK_THREAD_LIMIT; |
12e22c5e DH |
52 | static const int slow_work_min_vslow = 1; |
53 | static const int slow_work_max_vslow = 99; | |
54 | ||
55 | ctl_table slow_work_sysctls[] = { | |
56 | { | |
57 | .ctl_name = CTL_UNNUMBERED, | |
58 | .procname = "min-threads", | |
59 | .data = &slow_work_min_threads, | |
60 | .maxlen = sizeof(unsigned), | |
61 | .mode = 0644, | |
62 | .proc_handler = slow_work_min_threads_sysctl, | |
63 | .extra1 = (void *) &slow_work_min_min_threads, | |
64 | .extra2 = &slow_work_max_threads, | |
65 | }, | |
66 | { | |
67 | .ctl_name = CTL_UNNUMBERED, | |
68 | .procname = "max-threads", | |
69 | .data = &slow_work_max_threads, | |
70 | .maxlen = sizeof(unsigned), | |
71 | .mode = 0644, | |
72 | .proc_handler = slow_work_max_threads_sysctl, | |
73 | .extra1 = &slow_work_min_threads, | |
74 | .extra2 = (void *) &slow_work_max_max_threads, | |
75 | }, | |
76 | { | |
77 | .ctl_name = CTL_UNNUMBERED, | |
78 | .procname = "vslow-percentage", | |
79 | .data = &vslow_work_proportion, | |
80 | .maxlen = sizeof(unsigned), | |
81 | .mode = 0644, | |
82 | .proc_handler = &proc_dointvec_minmax, | |
83 | .extra1 = (void *) &slow_work_min_vslow, | |
84 | .extra2 = (void *) &slow_work_max_vslow, | |
85 | }, | |
86 | { .ctl_name = 0 } | |
87 | }; | |
88 | #endif | |
89 | ||
90 | /* | |
91 | * The active state of the thread pool | |
92 | */ | |
07fe7cb7 DH |
93 | static atomic_t slow_work_thread_count; |
94 | static atomic_t vslow_work_executing_count; | |
95 | ||
109d9272 DH |
96 | static bool slow_work_may_not_start_new_thread; |
97 | static bool slow_work_cull; /* cull a thread due to lack of activity */ | |
98 | static DEFINE_TIMER(slow_work_cull_timer, slow_work_cull_timeout, 0, 0); | |
99 | static DEFINE_TIMER(slow_work_oom_timer, slow_work_oom_timeout, 0, 0); | |
100 | static struct slow_work slow_work_new_thread; /* new thread starter */ | |
101 | ||
3d7a641e DH |
102 | /* |
103 | * slow work ID allocation (use slow_work_queue_lock) | |
104 | */ | |
105 | static DECLARE_BITMAP(slow_work_ids, SLOW_WORK_THREAD_LIMIT); | |
106 | ||
107 | /* | |
108 | * Unregistration tracking to prevent put_ref() from disappearing during module | |
109 | * unload | |
110 | */ | |
111 | #ifdef CONFIG_MODULES | |
112 | static struct module *slow_work_thread_processing[SLOW_WORK_THREAD_LIMIT]; | |
113 | static struct module *slow_work_unreg_module; | |
114 | static struct slow_work *slow_work_unreg_work_item; | |
115 | static DECLARE_WAIT_QUEUE_HEAD(slow_work_unreg_wq); | |
116 | static DEFINE_MUTEX(slow_work_unreg_sync_lock); | |
117 | #endif | |
118 | ||
07fe7cb7 DH |
119 | /* |
120 | * The queues of work items and the lock governing access to them. These are | |
121 | * shared between all the CPUs. It doesn't make sense to have per-CPU queues | |
122 | * as the number of threads bears no relation to the number of CPUs. | |
123 | * | |
124 | * There are two queues of work items: one for slow work items, and one for | |
125 | * very slow work items. | |
126 | */ | |
127 | static LIST_HEAD(slow_work_queue); | |
128 | static LIST_HEAD(vslow_work_queue); | |
129 | static DEFINE_SPINLOCK(slow_work_queue_lock); | |
130 | ||
131 | /* | |
132 | * The thread controls. A variable used to signal to the threads that they | |
133 | * should exit when the queue is empty, a waitqueue used by the threads to wait | |
134 | * for signals, and a completion set by the last thread to exit. | |
135 | */ | |
136 | static bool slow_work_threads_should_exit; | |
137 | static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq); | |
138 | static DECLARE_COMPLETION(slow_work_last_thread_exited); | |
139 | ||
140 | /* | |
141 | * The number of users of the thread pool and its lock. Whilst this is zero we | |
142 | * have no threads hanging around, and when this reaches zero, we wait for all | |
143 | * active or queued work items to complete and kill all the threads we do have. | |
144 | */ | |
145 | static int slow_work_user_count; | |
146 | static DEFINE_MUTEX(slow_work_user_lock); | |
147 | ||
4d8bb2cb JA |
148 | static inline int slow_work_get_ref(struct slow_work *work) |
149 | { | |
150 | if (work->ops->get_ref) | |
151 | return work->ops->get_ref(work); | |
152 | ||
153 | return 0; | |
154 | } | |
155 | ||
156 | static inline void slow_work_put_ref(struct slow_work *work) | |
157 | { | |
158 | if (work->ops->put_ref) | |
159 | work->ops->put_ref(work); | |
160 | } | |
161 | ||
07fe7cb7 DH |
162 | /* |
163 | * Calculate the maximum number of active threads in the pool that are | |
164 | * permitted to process very slow work items. | |
165 | * | |
166 | * The answer is rounded up to at least 1, but may not equal or exceed the | |
167 | * maximum number of the threads in the pool. This means we always have at | |
168 | * least one thread that can process slow work items, and we always have at | |
169 | * least one thread that won't get tied up doing so. | |
170 | */ | |
171 | static unsigned slow_work_calc_vsmax(void) | |
172 | { | |
173 | unsigned vsmax; | |
174 | ||
175 | vsmax = atomic_read(&slow_work_thread_count) * vslow_work_proportion; | |
176 | vsmax /= 100; | |
177 | vsmax = max(vsmax, 1U); | |
178 | return min(vsmax, slow_work_max_threads - 1); | |
179 | } | |
180 | ||
181 | /* | |
182 | * Attempt to execute stuff queued on a slow thread. Return true if we managed | |
183 | * it, false if there was nothing to do. | |
184 | */ | |
3d7a641e | 185 | static bool slow_work_execute(int id) |
07fe7cb7 | 186 | { |
3d7a641e DH |
187 | #ifdef CONFIG_MODULES |
188 | struct module *module; | |
189 | #endif | |
07fe7cb7 DH |
190 | struct slow_work *work = NULL; |
191 | unsigned vsmax; | |
192 | bool very_slow; | |
193 | ||
194 | vsmax = slow_work_calc_vsmax(); | |
195 | ||
109d9272 DH |
196 | /* see if we can schedule a new thread to be started if we're not |
197 | * keeping up with the work */ | |
198 | if (!waitqueue_active(&slow_work_thread_wq) && | |
199 | (!list_empty(&slow_work_queue) || !list_empty(&vslow_work_queue)) && | |
200 | atomic_read(&slow_work_thread_count) < slow_work_max_threads && | |
201 | !slow_work_may_not_start_new_thread) | |
202 | slow_work_enqueue(&slow_work_new_thread); | |
203 | ||
07fe7cb7 DH |
204 | /* find something to execute */ |
205 | spin_lock_irq(&slow_work_queue_lock); | |
206 | if (!list_empty(&vslow_work_queue) && | |
207 | atomic_read(&vslow_work_executing_count) < vsmax) { | |
208 | work = list_entry(vslow_work_queue.next, | |
209 | struct slow_work, link); | |
210 | if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags)) | |
211 | BUG(); | |
212 | list_del_init(&work->link); | |
213 | atomic_inc(&vslow_work_executing_count); | |
214 | very_slow = true; | |
215 | } else if (!list_empty(&slow_work_queue)) { | |
216 | work = list_entry(slow_work_queue.next, | |
217 | struct slow_work, link); | |
218 | if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags)) | |
219 | BUG(); | |
220 | list_del_init(&work->link); | |
221 | very_slow = false; | |
222 | } else { | |
223 | very_slow = false; /* avoid the compiler warning */ | |
224 | } | |
3d7a641e DH |
225 | |
226 | #ifdef CONFIG_MODULES | |
227 | if (work) | |
228 | slow_work_thread_processing[id] = work->owner; | |
229 | #endif | |
230 | ||
07fe7cb7 DH |
231 | spin_unlock_irq(&slow_work_queue_lock); |
232 | ||
233 | if (!work) | |
234 | return false; | |
235 | ||
236 | if (!test_and_clear_bit(SLOW_WORK_PENDING, &work->flags)) | |
237 | BUG(); | |
238 | ||
01609502 JA |
239 | /* don't execute if the work is in the process of being cancelled */ |
240 | if (!test_bit(SLOW_WORK_CANCELLING, &work->flags)) | |
241 | work->ops->execute(work); | |
07fe7cb7 DH |
242 | |
243 | if (very_slow) | |
244 | atomic_dec(&vslow_work_executing_count); | |
245 | clear_bit_unlock(SLOW_WORK_EXECUTING, &work->flags); | |
246 | ||
01609502 JA |
247 | /* wake up anyone waiting for this work to be complete */ |
248 | wake_up_bit(&work->flags, SLOW_WORK_EXECUTING); | |
249 | ||
07fe7cb7 DH |
250 | /* if someone tried to enqueue the item whilst we were executing it, |
251 | * then it'll be left unenqueued to avoid multiple threads trying to | |
252 | * execute it simultaneously | |
253 | * | |
254 | * there is, however, a race between us testing the pending flag and | |
255 | * getting the spinlock, and between the enqueuer setting the pending | |
256 | * flag and getting the spinlock, so we use a deferral bit to tell us | |
257 | * if the enqueuer got there first | |
258 | */ | |
259 | if (test_bit(SLOW_WORK_PENDING, &work->flags)) { | |
260 | spin_lock_irq(&slow_work_queue_lock); | |
261 | ||
262 | if (!test_bit(SLOW_WORK_EXECUTING, &work->flags) && | |
263 | test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags)) | |
264 | goto auto_requeue; | |
265 | ||
266 | spin_unlock_irq(&slow_work_queue_lock); | |
267 | } | |
268 | ||
3d7a641e | 269 | /* sort out the race between module unloading and put_ref() */ |
4d8bb2cb | 270 | slow_work_put_ref(work); |
3d7a641e DH |
271 | |
272 | #ifdef CONFIG_MODULES | |
273 | module = slow_work_thread_processing[id]; | |
274 | slow_work_thread_processing[id] = NULL; | |
275 | smp_mb(); | |
276 | if (slow_work_unreg_work_item == work || | |
277 | slow_work_unreg_module == module) | |
278 | wake_up_all(&slow_work_unreg_wq); | |
279 | #endif | |
280 | ||
07fe7cb7 DH |
281 | return true; |
282 | ||
283 | auto_requeue: | |
284 | /* we must complete the enqueue operation | |
285 | * - we transfer our ref on the item back to the appropriate queue | |
286 | * - don't wake another thread up as we're awake already | |
287 | */ | |
288 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) | |
289 | list_add_tail(&work->link, &vslow_work_queue); | |
290 | else | |
291 | list_add_tail(&work->link, &slow_work_queue); | |
292 | spin_unlock_irq(&slow_work_queue_lock); | |
3d7a641e | 293 | slow_work_thread_processing[id] = NULL; |
07fe7cb7 DH |
294 | return true; |
295 | } | |
296 | ||
297 | /** | |
298 | * slow_work_enqueue - Schedule a slow work item for processing | |
299 | * @work: The work item to queue | |
300 | * | |
301 | * Schedule a slow work item for processing. If the item is already undergoing | |
302 | * execution, this guarantees not to re-enter the execution routine until the | |
303 | * first execution finishes. | |
304 | * | |
305 | * The item is pinned by this function as it retains a reference to it, managed | |
306 | * through the item operations. The item is unpinned once it has been | |
307 | * executed. | |
308 | * | |
309 | * An item may hog the thread that is running it for a relatively large amount | |
310 | * of time, sufficient, for example, to perform several lookup, mkdir, create | |
311 | * and setxattr operations. It may sleep on I/O and may sleep to obtain locks. | |
312 | * | |
313 | * Conversely, if a number of items are awaiting processing, it may take some | |
314 | * time before any given item is given attention. The number of threads in the | |
315 | * pool may be increased to deal with demand, but only up to a limit. | |
316 | * | |
317 | * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in | |
318 | * the very slow queue, from which only a portion of the threads will be | |
319 | * allowed to pick items to execute. This ensures that very slow items won't | |
320 | * overly block ones that are just ordinarily slow. | |
321 | * | |
01609502 JA |
322 | * Returns 0 if successful, -EAGAIN if not (or -ECANCELED if cancelled work is |
323 | * attempted queued) | |
07fe7cb7 DH |
324 | */ |
325 | int slow_work_enqueue(struct slow_work *work) | |
326 | { | |
327 | unsigned long flags; | |
01609502 JA |
328 | int ret; |
329 | ||
330 | if (test_bit(SLOW_WORK_CANCELLING, &work->flags)) | |
331 | return -ECANCELED; | |
07fe7cb7 DH |
332 | |
333 | BUG_ON(slow_work_user_count <= 0); | |
334 | BUG_ON(!work); | |
335 | BUG_ON(!work->ops); | |
07fe7cb7 DH |
336 | |
337 | /* when honouring an enqueue request, we only promise that we will run | |
338 | * the work function in the future; we do not promise to run it once | |
339 | * per enqueue request | |
340 | * | |
341 | * we use the PENDING bit to merge together repeat requests without | |
342 | * having to disable IRQs and take the spinlock, whilst still | |
343 | * maintaining our promise | |
344 | */ | |
345 | if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) { | |
346 | spin_lock_irqsave(&slow_work_queue_lock, flags); | |
347 | ||
01609502 JA |
348 | if (unlikely(test_bit(SLOW_WORK_CANCELLING, &work->flags))) |
349 | goto cancelled; | |
350 | ||
07fe7cb7 DH |
351 | /* we promise that we will not attempt to execute the work |
352 | * function in more than one thread simultaneously | |
353 | * | |
354 | * this, however, leaves us with a problem if we're asked to | |
355 | * enqueue the work whilst someone is executing the work | |
356 | * function as simply queueing the work immediately means that | |
357 | * another thread may try executing it whilst it is already | |
358 | * under execution | |
359 | * | |
360 | * to deal with this, we set the ENQ_DEFERRED bit instead of | |
361 | * enqueueing, and the thread currently executing the work | |
362 | * function will enqueue the work item when the work function | |
363 | * returns and it has cleared the EXECUTING bit | |
364 | */ | |
365 | if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) { | |
366 | set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags); | |
367 | } else { | |
01609502 JA |
368 | ret = slow_work_get_ref(work); |
369 | if (ret < 0) | |
370 | goto failed; | |
07fe7cb7 DH |
371 | if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags)) |
372 | list_add_tail(&work->link, &vslow_work_queue); | |
373 | else | |
374 | list_add_tail(&work->link, &slow_work_queue); | |
375 | wake_up(&slow_work_thread_wq); | |
376 | } | |
377 | ||
378 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); | |
379 | } | |
380 | return 0; | |
381 | ||
01609502 JA |
382 | cancelled: |
383 | ret = -ECANCELED; | |
384 | failed: | |
07fe7cb7 | 385 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); |
01609502 | 386 | return ret; |
07fe7cb7 DH |
387 | } |
388 | EXPORT_SYMBOL(slow_work_enqueue); | |
389 | ||
01609502 JA |
390 | static int slow_work_wait(void *word) |
391 | { | |
392 | schedule(); | |
393 | return 0; | |
394 | } | |
395 | ||
396 | /** | |
397 | * slow_work_cancel - Cancel a slow work item | |
398 | * @work: The work item to cancel | |
399 | * | |
400 | * This function will cancel a previously enqueued work item. If we cannot | |
401 | * cancel the work item, it is guarenteed to have run when this function | |
402 | * returns. | |
403 | */ | |
404 | void slow_work_cancel(struct slow_work *work) | |
405 | { | |
406 | bool wait = true, put = false; | |
407 | ||
408 | set_bit(SLOW_WORK_CANCELLING, &work->flags); | |
409 | ||
410 | spin_lock_irq(&slow_work_queue_lock); | |
411 | ||
412 | if (test_bit(SLOW_WORK_PENDING, &work->flags) && | |
413 | !list_empty(&work->link)) { | |
414 | /* the link in the pending queue holds a reference on the item | |
415 | * that we will need to release */ | |
416 | list_del_init(&work->link); | |
417 | wait = false; | |
418 | put = true; | |
419 | clear_bit(SLOW_WORK_PENDING, &work->flags); | |
420 | ||
421 | } else if (test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags)) { | |
422 | /* the executor is holding our only reference on the item, so | |
423 | * we merely need to wait for it to finish executing */ | |
424 | clear_bit(SLOW_WORK_PENDING, &work->flags); | |
425 | } | |
426 | ||
427 | spin_unlock_irq(&slow_work_queue_lock); | |
428 | ||
429 | /* the EXECUTING flag is set by the executor whilst the spinlock is set | |
430 | * and before the item is dequeued - so assuming the above doesn't | |
431 | * actually dequeue it, simply waiting for the EXECUTING flag to be | |
432 | * released here should be sufficient */ | |
433 | if (wait) | |
434 | wait_on_bit(&work->flags, SLOW_WORK_EXECUTING, slow_work_wait, | |
435 | TASK_UNINTERRUPTIBLE); | |
436 | ||
437 | clear_bit(SLOW_WORK_CANCELLING, &work->flags); | |
438 | if (put) | |
439 | slow_work_put_ref(work); | |
440 | } | |
441 | EXPORT_SYMBOL(slow_work_cancel); | |
442 | ||
009789f0 CP |
443 | /* |
444 | * Schedule a cull of the thread pool at some time in the near future | |
445 | */ | |
446 | static void slow_work_schedule_cull(void) | |
447 | { | |
448 | mod_timer(&slow_work_cull_timer, | |
449 | round_jiffies(jiffies + SLOW_WORK_CULL_TIMEOUT)); | |
450 | } | |
451 | ||
109d9272 DH |
452 | /* |
453 | * Worker thread culling algorithm | |
454 | */ | |
455 | static bool slow_work_cull_thread(void) | |
456 | { | |
457 | unsigned long flags; | |
458 | bool do_cull = false; | |
459 | ||
460 | spin_lock_irqsave(&slow_work_queue_lock, flags); | |
461 | ||
462 | if (slow_work_cull) { | |
463 | slow_work_cull = false; | |
464 | ||
465 | if (list_empty(&slow_work_queue) && | |
466 | list_empty(&vslow_work_queue) && | |
467 | atomic_read(&slow_work_thread_count) > | |
468 | slow_work_min_threads) { | |
009789f0 | 469 | slow_work_schedule_cull(); |
109d9272 DH |
470 | do_cull = true; |
471 | } | |
472 | } | |
473 | ||
474 | spin_unlock_irqrestore(&slow_work_queue_lock, flags); | |
475 | return do_cull; | |
476 | } | |
477 | ||
07fe7cb7 DH |
478 | /* |
479 | * Determine if there is slow work available for dispatch | |
480 | */ | |
481 | static inline bool slow_work_available(int vsmax) | |
482 | { | |
483 | return !list_empty(&slow_work_queue) || | |
484 | (!list_empty(&vslow_work_queue) && | |
485 | atomic_read(&vslow_work_executing_count) < vsmax); | |
486 | } | |
487 | ||
488 | /* | |
489 | * Worker thread dispatcher | |
490 | */ | |
491 | static int slow_work_thread(void *_data) | |
492 | { | |
3d7a641e | 493 | int vsmax, id; |
07fe7cb7 DH |
494 | |
495 | DEFINE_WAIT(wait); | |
496 | ||
497 | set_freezable(); | |
498 | set_user_nice(current, -5); | |
499 | ||
3d7a641e DH |
500 | /* allocate ourselves an ID */ |
501 | spin_lock_irq(&slow_work_queue_lock); | |
502 | id = find_first_zero_bit(slow_work_ids, SLOW_WORK_THREAD_LIMIT); | |
503 | BUG_ON(id < 0 || id >= SLOW_WORK_THREAD_LIMIT); | |
504 | __set_bit(id, slow_work_ids); | |
505 | spin_unlock_irq(&slow_work_queue_lock); | |
506 | ||
507 | sprintf(current->comm, "kslowd%03u", id); | |
508 | ||
07fe7cb7 DH |
509 | for (;;) { |
510 | vsmax = vslow_work_proportion; | |
511 | vsmax *= atomic_read(&slow_work_thread_count); | |
512 | vsmax /= 100; | |
513 | ||
b415c49a ON |
514 | prepare_to_wait_exclusive(&slow_work_thread_wq, &wait, |
515 | TASK_INTERRUPTIBLE); | |
07fe7cb7 DH |
516 | if (!freezing(current) && |
517 | !slow_work_threads_should_exit && | |
109d9272 DH |
518 | !slow_work_available(vsmax) && |
519 | !slow_work_cull) | |
07fe7cb7 DH |
520 | schedule(); |
521 | finish_wait(&slow_work_thread_wq, &wait); | |
522 | ||
523 | try_to_freeze(); | |
524 | ||
525 | vsmax = vslow_work_proportion; | |
526 | vsmax *= atomic_read(&slow_work_thread_count); | |
527 | vsmax /= 100; | |
528 | ||
3d7a641e | 529 | if (slow_work_available(vsmax) && slow_work_execute(id)) { |
07fe7cb7 | 530 | cond_resched(); |
109d9272 DH |
531 | if (list_empty(&slow_work_queue) && |
532 | list_empty(&vslow_work_queue) && | |
533 | atomic_read(&slow_work_thread_count) > | |
534 | slow_work_min_threads) | |
009789f0 | 535 | slow_work_schedule_cull(); |
07fe7cb7 DH |
536 | continue; |
537 | } | |
538 | ||
539 | if (slow_work_threads_should_exit) | |
540 | break; | |
109d9272 DH |
541 | |
542 | if (slow_work_cull && slow_work_cull_thread()) | |
543 | break; | |
07fe7cb7 DH |
544 | } |
545 | ||
3d7a641e DH |
546 | spin_lock_irq(&slow_work_queue_lock); |
547 | __clear_bit(id, slow_work_ids); | |
548 | spin_unlock_irq(&slow_work_queue_lock); | |
549 | ||
07fe7cb7 DH |
550 | if (atomic_dec_and_test(&slow_work_thread_count)) |
551 | complete_and_exit(&slow_work_last_thread_exited, 0); | |
552 | return 0; | |
553 | } | |
554 | ||
109d9272 DH |
555 | /* |
556 | * Handle thread cull timer expiration | |
557 | */ | |
558 | static void slow_work_cull_timeout(unsigned long data) | |
559 | { | |
560 | slow_work_cull = true; | |
561 | wake_up(&slow_work_thread_wq); | |
562 | } | |
563 | ||
109d9272 DH |
564 | /* |
565 | * Start a new slow work thread | |
566 | */ | |
567 | static void slow_work_new_thread_execute(struct slow_work *work) | |
568 | { | |
569 | struct task_struct *p; | |
570 | ||
571 | if (slow_work_threads_should_exit) | |
572 | return; | |
573 | ||
574 | if (atomic_read(&slow_work_thread_count) >= slow_work_max_threads) | |
575 | return; | |
576 | ||
577 | if (!mutex_trylock(&slow_work_user_lock)) | |
578 | return; | |
579 | ||
580 | slow_work_may_not_start_new_thread = true; | |
581 | atomic_inc(&slow_work_thread_count); | |
582 | p = kthread_run(slow_work_thread, NULL, "kslowd"); | |
583 | if (IS_ERR(p)) { | |
584 | printk(KERN_DEBUG "Slow work thread pool: OOM\n"); | |
585 | if (atomic_dec_and_test(&slow_work_thread_count)) | |
586 | BUG(); /* we're running on a slow work thread... */ | |
587 | mod_timer(&slow_work_oom_timer, | |
009789f0 | 588 | round_jiffies(jiffies + SLOW_WORK_OOM_TIMEOUT)); |
109d9272 DH |
589 | } else { |
590 | /* ratelimit the starting of new threads */ | |
591 | mod_timer(&slow_work_oom_timer, jiffies + 1); | |
592 | } | |
593 | ||
594 | mutex_unlock(&slow_work_user_lock); | |
595 | } | |
596 | ||
597 | static const struct slow_work_ops slow_work_new_thread_ops = { | |
3d7a641e | 598 | .owner = THIS_MODULE, |
109d9272 DH |
599 | .execute = slow_work_new_thread_execute, |
600 | }; | |
601 | ||
602 | /* | |
603 | * post-OOM new thread start suppression expiration | |
604 | */ | |
605 | static void slow_work_oom_timeout(unsigned long data) | |
606 | { | |
607 | slow_work_may_not_start_new_thread = false; | |
608 | } | |
609 | ||
12e22c5e DH |
610 | #ifdef CONFIG_SYSCTL |
611 | /* | |
612 | * Handle adjustment of the minimum number of threads | |
613 | */ | |
614 | static int slow_work_min_threads_sysctl(struct ctl_table *table, int write, | |
8d65af78 | 615 | void __user *buffer, |
12e22c5e DH |
616 | size_t *lenp, loff_t *ppos) |
617 | { | |
8d65af78 | 618 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
12e22c5e DH |
619 | int n; |
620 | ||
621 | if (ret == 0) { | |
622 | mutex_lock(&slow_work_user_lock); | |
623 | if (slow_work_user_count > 0) { | |
624 | /* see if we need to start or stop threads */ | |
625 | n = atomic_read(&slow_work_thread_count) - | |
626 | slow_work_min_threads; | |
627 | ||
628 | if (n < 0 && !slow_work_may_not_start_new_thread) | |
629 | slow_work_enqueue(&slow_work_new_thread); | |
630 | else if (n > 0) | |
009789f0 | 631 | slow_work_schedule_cull(); |
12e22c5e DH |
632 | } |
633 | mutex_unlock(&slow_work_user_lock); | |
634 | } | |
635 | ||
636 | return ret; | |
637 | } | |
638 | ||
639 | /* | |
640 | * Handle adjustment of the maximum number of threads | |
641 | */ | |
642 | static int slow_work_max_threads_sysctl(struct ctl_table *table, int write, | |
8d65af78 | 643 | void __user *buffer, |
12e22c5e DH |
644 | size_t *lenp, loff_t *ppos) |
645 | { | |
8d65af78 | 646 | int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); |
12e22c5e DH |
647 | int n; |
648 | ||
649 | if (ret == 0) { | |
650 | mutex_lock(&slow_work_user_lock); | |
651 | if (slow_work_user_count > 0) { | |
652 | /* see if we need to stop threads */ | |
653 | n = slow_work_max_threads - | |
654 | atomic_read(&slow_work_thread_count); | |
655 | ||
656 | if (n < 0) | |
009789f0 | 657 | slow_work_schedule_cull(); |
12e22c5e DH |
658 | } |
659 | mutex_unlock(&slow_work_user_lock); | |
660 | } | |
661 | ||
662 | return ret; | |
663 | } | |
664 | #endif /* CONFIG_SYSCTL */ | |
665 | ||
07fe7cb7 DH |
666 | /** |
667 | * slow_work_register_user - Register a user of the facility | |
3d7a641e | 668 | * @module: The module about to make use of the facility |
07fe7cb7 DH |
669 | * |
670 | * Register a user of the facility, starting up the initial threads if there | |
671 | * aren't any other users at this point. This will return 0 if successful, or | |
672 | * an error if not. | |
673 | */ | |
3d7a641e | 674 | int slow_work_register_user(struct module *module) |
07fe7cb7 DH |
675 | { |
676 | struct task_struct *p; | |
677 | int loop; | |
678 | ||
679 | mutex_lock(&slow_work_user_lock); | |
680 | ||
681 | if (slow_work_user_count == 0) { | |
682 | printk(KERN_NOTICE "Slow work thread pool: Starting up\n"); | |
683 | init_completion(&slow_work_last_thread_exited); | |
684 | ||
685 | slow_work_threads_should_exit = false; | |
109d9272 DH |
686 | slow_work_init(&slow_work_new_thread, |
687 | &slow_work_new_thread_ops); | |
688 | slow_work_may_not_start_new_thread = false; | |
689 | slow_work_cull = false; | |
07fe7cb7 DH |
690 | |
691 | /* start the minimum number of threads */ | |
692 | for (loop = 0; loop < slow_work_min_threads; loop++) { | |
693 | atomic_inc(&slow_work_thread_count); | |
694 | p = kthread_run(slow_work_thread, NULL, "kslowd"); | |
695 | if (IS_ERR(p)) | |
696 | goto error; | |
697 | } | |
698 | printk(KERN_NOTICE "Slow work thread pool: Ready\n"); | |
699 | } | |
700 | ||
701 | slow_work_user_count++; | |
702 | mutex_unlock(&slow_work_user_lock); | |
703 | return 0; | |
704 | ||
705 | error: | |
706 | if (atomic_dec_and_test(&slow_work_thread_count)) | |
707 | complete(&slow_work_last_thread_exited); | |
708 | if (loop > 0) { | |
709 | printk(KERN_ERR "Slow work thread pool:" | |
710 | " Aborting startup on ENOMEM\n"); | |
711 | slow_work_threads_should_exit = true; | |
712 | wake_up_all(&slow_work_thread_wq); | |
713 | wait_for_completion(&slow_work_last_thread_exited); | |
714 | printk(KERN_ERR "Slow work thread pool: Aborted\n"); | |
715 | } | |
716 | mutex_unlock(&slow_work_user_lock); | |
717 | return PTR_ERR(p); | |
718 | } | |
719 | EXPORT_SYMBOL(slow_work_register_user); | |
720 | ||
3d7a641e DH |
721 | /* |
722 | * wait for all outstanding items from the calling module to complete | |
723 | * - note that more items may be queued whilst we're waiting | |
724 | */ | |
725 | static void slow_work_wait_for_items(struct module *module) | |
726 | { | |
727 | DECLARE_WAITQUEUE(myself, current); | |
728 | struct slow_work *work; | |
729 | int loop; | |
730 | ||
731 | mutex_lock(&slow_work_unreg_sync_lock); | |
732 | add_wait_queue(&slow_work_unreg_wq, &myself); | |
733 | ||
734 | for (;;) { | |
735 | spin_lock_irq(&slow_work_queue_lock); | |
736 | ||
737 | /* first of all, we wait for the last queued item in each list | |
738 | * to be processed */ | |
739 | list_for_each_entry_reverse(work, &vslow_work_queue, link) { | |
740 | if (work->owner == module) { | |
741 | set_current_state(TASK_UNINTERRUPTIBLE); | |
742 | slow_work_unreg_work_item = work; | |
743 | goto do_wait; | |
744 | } | |
745 | } | |
746 | list_for_each_entry_reverse(work, &slow_work_queue, link) { | |
747 | if (work->owner == module) { | |
748 | set_current_state(TASK_UNINTERRUPTIBLE); | |
749 | slow_work_unreg_work_item = work; | |
750 | goto do_wait; | |
751 | } | |
752 | } | |
753 | ||
754 | /* then we wait for the items being processed to finish */ | |
755 | slow_work_unreg_module = module; | |
756 | smp_mb(); | |
757 | for (loop = 0; loop < SLOW_WORK_THREAD_LIMIT; loop++) { | |
758 | if (slow_work_thread_processing[loop] == module) | |
759 | goto do_wait; | |
760 | } | |
761 | spin_unlock_irq(&slow_work_queue_lock); | |
762 | break; /* okay, we're done */ | |
763 | ||
764 | do_wait: | |
765 | spin_unlock_irq(&slow_work_queue_lock); | |
766 | schedule(); | |
767 | slow_work_unreg_work_item = NULL; | |
768 | slow_work_unreg_module = NULL; | |
769 | } | |
770 | ||
771 | remove_wait_queue(&slow_work_unreg_wq, &myself); | |
772 | mutex_unlock(&slow_work_unreg_sync_lock); | |
773 | } | |
774 | ||
07fe7cb7 DH |
775 | /** |
776 | * slow_work_unregister_user - Unregister a user of the facility | |
3d7a641e | 777 | * @module: The module whose items should be cleared |
07fe7cb7 DH |
778 | * |
779 | * Unregister a user of the facility, killing all the threads if this was the | |
780 | * last one. | |
3d7a641e DH |
781 | * |
782 | * This waits for all the work items belonging to the nominated module to go | |
783 | * away before proceeding. | |
07fe7cb7 | 784 | */ |
3d7a641e | 785 | void slow_work_unregister_user(struct module *module) |
07fe7cb7 | 786 | { |
3d7a641e DH |
787 | /* first of all, wait for all outstanding items from the calling module |
788 | * to complete */ | |
789 | if (module) | |
790 | slow_work_wait_for_items(module); | |
791 | ||
792 | /* then we can actually go about shutting down the facility if need | |
793 | * be */ | |
07fe7cb7 DH |
794 | mutex_lock(&slow_work_user_lock); |
795 | ||
796 | BUG_ON(slow_work_user_count <= 0); | |
797 | ||
798 | slow_work_user_count--; | |
799 | if (slow_work_user_count == 0) { | |
800 | printk(KERN_NOTICE "Slow work thread pool: Shutting down\n"); | |
801 | slow_work_threads_should_exit = true; | |
418df63c JC |
802 | del_timer_sync(&slow_work_cull_timer); |
803 | del_timer_sync(&slow_work_oom_timer); | |
07fe7cb7 DH |
804 | wake_up_all(&slow_work_thread_wq); |
805 | wait_for_completion(&slow_work_last_thread_exited); | |
806 | printk(KERN_NOTICE "Slow work thread pool:" | |
807 | " Shut down complete\n"); | |
808 | } | |
809 | ||
810 | mutex_unlock(&slow_work_user_lock); | |
811 | } | |
812 | EXPORT_SYMBOL(slow_work_unregister_user); | |
813 | ||
814 | /* | |
815 | * Initialise the slow work facility | |
816 | */ | |
817 | static int __init init_slow_work(void) | |
818 | { | |
819 | unsigned nr_cpus = num_possible_cpus(); | |
820 | ||
12e22c5e | 821 | if (slow_work_max_threads < nr_cpus) |
07fe7cb7 | 822 | slow_work_max_threads = nr_cpus; |
12e22c5e DH |
823 | #ifdef CONFIG_SYSCTL |
824 | if (slow_work_max_max_threads < nr_cpus * 2) | |
825 | slow_work_max_max_threads = nr_cpus * 2; | |
826 | #endif | |
07fe7cb7 DH |
827 | return 0; |
828 | } | |
829 | ||
830 | subsys_initcall(init_slow_work); |