blob: fccf421eb5c1e9fc89c79102e9da4643fea35358 [file] [log] [blame]
David Howells07fe7cb2009-04-03 16:42:35 +01001/* 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.
David Howells8f0aa2f2009-04-03 16:42:35 +010010 *
11 * See Documentation/slow-work.txt
David Howells07fe7cb2009-04-03 16:42:35 +010012 */
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>
David Howells07fe7cb2009-04-03 16:42:35 +010019
David Howells109d9272009-04-03 16:42:35 +010020#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
David Howells3d7a6412009-11-19 18:10:23 +000025#define SLOW_WORK_THREAD_LIMIT 255 /* abs maximum number of slow-work threads */
26
David Howells109d9272009-04-03 16:42:35 +010027static void slow_work_cull_timeout(unsigned long);
28static void slow_work_oom_timeout(unsigned long);
29
David Howells12e22c52009-04-03 16:42:35 +010030#ifdef CONFIG_SYSCTL
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070031static int slow_work_min_threads_sysctl(struct ctl_table *, int,
David Howells12e22c52009-04-03 16:42:35 +010032 void __user *, size_t *, loff_t *);
33
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070034static int slow_work_max_threads_sysctl(struct ctl_table *, int ,
David Howells12e22c52009-04-03 16:42:35 +010035 void __user *, size_t *, loff_t *);
36#endif
37
David Howells07fe7cb2009-04-03 16:42:35 +010038/*
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 */
44static unsigned slow_work_min_threads = 2;
45static unsigned slow_work_max_threads = 4;
46static unsigned vslow_work_proportion = 50; /* % of threads that may process
47 * very slow work */
David Howells12e22c52009-04-03 16:42:35 +010048
49#ifdef CONFIG_SYSCTL
50static const int slow_work_min_min_threads = 2;
David Howells3d7a6412009-11-19 18:10:23 +000051static int slow_work_max_max_threads = SLOW_WORK_THREAD_LIMIT;
David Howells12e22c52009-04-03 16:42:35 +010052static const int slow_work_min_vslow = 1;
53static const int slow_work_max_vslow = 99;
54
55ctl_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 */
David Howells07fe7cb2009-04-03 16:42:35 +010093static atomic_t slow_work_thread_count;
94static atomic_t vslow_work_executing_count;
95
David Howells109d9272009-04-03 16:42:35 +010096static bool slow_work_may_not_start_new_thread;
97static bool slow_work_cull; /* cull a thread due to lack of activity */
98static DEFINE_TIMER(slow_work_cull_timer, slow_work_cull_timeout, 0, 0);
99static DEFINE_TIMER(slow_work_oom_timer, slow_work_oom_timeout, 0, 0);
100static struct slow_work slow_work_new_thread; /* new thread starter */
101
David Howells07fe7cb2009-04-03 16:42:35 +0100102/*
David Howells3d7a6412009-11-19 18:10:23 +0000103 * slow work ID allocation (use slow_work_queue_lock)
104 */
105static 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
112static struct module *slow_work_thread_processing[SLOW_WORK_THREAD_LIMIT];
113static struct module *slow_work_unreg_module;
114static struct slow_work *slow_work_unreg_work_item;
115static DECLARE_WAIT_QUEUE_HEAD(slow_work_unreg_wq);
116static DEFINE_MUTEX(slow_work_unreg_sync_lock);
117#endif
118
119/*
David Howells07fe7cb2009-04-03 16:42:35 +0100120 * 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 */
127static LIST_HEAD(slow_work_queue);
128static LIST_HEAD(vslow_work_queue);
129static 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 */
136static bool slow_work_threads_should_exit;
137static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq);
138static 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 */
145static int slow_work_user_count;
146static DEFINE_MUTEX(slow_work_user_lock);
147
Jens Axboe4d8bb2c2009-11-19 18:10:39 +0000148static 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
156static 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
David Howells07fe7cb2009-04-03 16:42:35 +0100162/*
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 */
171static 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 */
David Howells3d7a6412009-11-19 18:10:23 +0000185static bool slow_work_execute(int id)
David Howells07fe7cb2009-04-03 16:42:35 +0100186{
David Howells3d7a6412009-11-19 18:10:23 +0000187#ifdef CONFIG_MODULES
188 struct module *module;
189#endif
David Howells07fe7cb2009-04-03 16:42:35 +0100190 struct slow_work *work = NULL;
191 unsigned vsmax;
192 bool very_slow;
193
194 vsmax = slow_work_calc_vsmax();
195
David Howells109d9272009-04-03 16:42:35 +0100196 /* 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
David Howells07fe7cb2009-04-03 16:42:35 +0100204 /* 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 }
David Howells3d7a6412009-11-19 18:10:23 +0000225
226#ifdef CONFIG_MODULES
227 if (work)
228 slow_work_thread_processing[id] = work->owner;
229#endif
230
David Howells07fe7cb2009-04-03 16:42:35 +0100231 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
239 work->ops->execute(work);
240
241 if (very_slow)
242 atomic_dec(&vslow_work_executing_count);
243 clear_bit_unlock(SLOW_WORK_EXECUTING, &work->flags);
244
245 /* if someone tried to enqueue the item whilst we were executing it,
246 * then it'll be left unenqueued to avoid multiple threads trying to
247 * execute it simultaneously
248 *
249 * there is, however, a race between us testing the pending flag and
250 * getting the spinlock, and between the enqueuer setting the pending
251 * flag and getting the spinlock, so we use a deferral bit to tell us
252 * if the enqueuer got there first
253 */
254 if (test_bit(SLOW_WORK_PENDING, &work->flags)) {
255 spin_lock_irq(&slow_work_queue_lock);
256
257 if (!test_bit(SLOW_WORK_EXECUTING, &work->flags) &&
258 test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags))
259 goto auto_requeue;
260
261 spin_unlock_irq(&slow_work_queue_lock);
262 }
263
David Howells3d7a6412009-11-19 18:10:23 +0000264 /* sort out the race between module unloading and put_ref() */
Jens Axboe4d8bb2c2009-11-19 18:10:39 +0000265 slow_work_put_ref(work);
David Howells3d7a6412009-11-19 18:10:23 +0000266
267#ifdef CONFIG_MODULES
268 module = slow_work_thread_processing[id];
269 slow_work_thread_processing[id] = NULL;
270 smp_mb();
271 if (slow_work_unreg_work_item == work ||
272 slow_work_unreg_module == module)
273 wake_up_all(&slow_work_unreg_wq);
274#endif
275
David Howells07fe7cb2009-04-03 16:42:35 +0100276 return true;
277
278auto_requeue:
279 /* we must complete the enqueue operation
280 * - we transfer our ref on the item back to the appropriate queue
281 * - don't wake another thread up as we're awake already
282 */
283 if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
284 list_add_tail(&work->link, &vslow_work_queue);
285 else
286 list_add_tail(&work->link, &slow_work_queue);
287 spin_unlock_irq(&slow_work_queue_lock);
David Howells3d7a6412009-11-19 18:10:23 +0000288 slow_work_thread_processing[id] = NULL;
David Howells07fe7cb2009-04-03 16:42:35 +0100289 return true;
290}
291
292/**
293 * slow_work_enqueue - Schedule a slow work item for processing
294 * @work: The work item to queue
295 *
296 * Schedule a slow work item for processing. If the item is already undergoing
297 * execution, this guarantees not to re-enter the execution routine until the
298 * first execution finishes.
299 *
300 * The item is pinned by this function as it retains a reference to it, managed
301 * through the item operations. The item is unpinned once it has been
302 * executed.
303 *
304 * An item may hog the thread that is running it for a relatively large amount
305 * of time, sufficient, for example, to perform several lookup, mkdir, create
306 * and setxattr operations. It may sleep on I/O and may sleep to obtain locks.
307 *
308 * Conversely, if a number of items are awaiting processing, it may take some
309 * time before any given item is given attention. The number of threads in the
310 * pool may be increased to deal with demand, but only up to a limit.
311 *
312 * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in
313 * the very slow queue, from which only a portion of the threads will be
314 * allowed to pick items to execute. This ensures that very slow items won't
315 * overly block ones that are just ordinarily slow.
316 *
317 * Returns 0 if successful, -EAGAIN if not.
318 */
319int slow_work_enqueue(struct slow_work *work)
320{
321 unsigned long flags;
322
323 BUG_ON(slow_work_user_count <= 0);
324 BUG_ON(!work);
325 BUG_ON(!work->ops);
David Howells07fe7cb2009-04-03 16:42:35 +0100326
327 /* when honouring an enqueue request, we only promise that we will run
328 * the work function in the future; we do not promise to run it once
329 * per enqueue request
330 *
331 * we use the PENDING bit to merge together repeat requests without
332 * having to disable IRQs and take the spinlock, whilst still
333 * maintaining our promise
334 */
335 if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) {
336 spin_lock_irqsave(&slow_work_queue_lock, flags);
337
338 /* we promise that we will not attempt to execute the work
339 * function in more than one thread simultaneously
340 *
341 * this, however, leaves us with a problem if we're asked to
342 * enqueue the work whilst someone is executing the work
343 * function as simply queueing the work immediately means that
344 * another thread may try executing it whilst it is already
345 * under execution
346 *
347 * to deal with this, we set the ENQ_DEFERRED bit instead of
348 * enqueueing, and the thread currently executing the work
349 * function will enqueue the work item when the work function
350 * returns and it has cleared the EXECUTING bit
351 */
352 if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) {
353 set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags);
354 } else {
Jens Axboe4d8bb2c2009-11-19 18:10:39 +0000355 if (slow_work_get_ref(work) < 0)
David Howells07fe7cb2009-04-03 16:42:35 +0100356 goto cant_get_ref;
357 if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
358 list_add_tail(&work->link, &vslow_work_queue);
359 else
360 list_add_tail(&work->link, &slow_work_queue);
361 wake_up(&slow_work_thread_wq);
362 }
363
364 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
365 }
366 return 0;
367
368cant_get_ref:
369 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
370 return -EAGAIN;
371}
372EXPORT_SYMBOL(slow_work_enqueue);
373
374/*
Chris Peterson009789f2009-06-16 15:33:43 -0700375 * Schedule a cull of the thread pool at some time in the near future
376 */
377static void slow_work_schedule_cull(void)
378{
379 mod_timer(&slow_work_cull_timer,
380 round_jiffies(jiffies + SLOW_WORK_CULL_TIMEOUT));
381}
382
383/*
David Howells109d9272009-04-03 16:42:35 +0100384 * Worker thread culling algorithm
385 */
386static bool slow_work_cull_thread(void)
387{
388 unsigned long flags;
389 bool do_cull = false;
390
391 spin_lock_irqsave(&slow_work_queue_lock, flags);
392
393 if (slow_work_cull) {
394 slow_work_cull = false;
395
396 if (list_empty(&slow_work_queue) &&
397 list_empty(&vslow_work_queue) &&
398 atomic_read(&slow_work_thread_count) >
399 slow_work_min_threads) {
Chris Peterson009789f2009-06-16 15:33:43 -0700400 slow_work_schedule_cull();
David Howells109d9272009-04-03 16:42:35 +0100401 do_cull = true;
402 }
403 }
404
405 spin_unlock_irqrestore(&slow_work_queue_lock, flags);
406 return do_cull;
407}
408
409/*
David Howells07fe7cb2009-04-03 16:42:35 +0100410 * Determine if there is slow work available for dispatch
411 */
412static inline bool slow_work_available(int vsmax)
413{
414 return !list_empty(&slow_work_queue) ||
415 (!list_empty(&vslow_work_queue) &&
416 atomic_read(&vslow_work_executing_count) < vsmax);
417}
418
419/*
420 * Worker thread dispatcher
421 */
422static int slow_work_thread(void *_data)
423{
David Howells3d7a6412009-11-19 18:10:23 +0000424 int vsmax, id;
David Howells07fe7cb2009-04-03 16:42:35 +0100425
426 DEFINE_WAIT(wait);
427
428 set_freezable();
429 set_user_nice(current, -5);
430
David Howells3d7a6412009-11-19 18:10:23 +0000431 /* allocate ourselves an ID */
432 spin_lock_irq(&slow_work_queue_lock);
433 id = find_first_zero_bit(slow_work_ids, SLOW_WORK_THREAD_LIMIT);
434 BUG_ON(id < 0 || id >= SLOW_WORK_THREAD_LIMIT);
435 __set_bit(id, slow_work_ids);
436 spin_unlock_irq(&slow_work_queue_lock);
437
438 sprintf(current->comm, "kslowd%03u", id);
439
David Howells07fe7cb2009-04-03 16:42:35 +0100440 for (;;) {
441 vsmax = vslow_work_proportion;
442 vsmax *= atomic_read(&slow_work_thread_count);
443 vsmax /= 100;
444
Oleg Nesterovb415c492009-06-11 13:12:55 +0100445 prepare_to_wait_exclusive(&slow_work_thread_wq, &wait,
446 TASK_INTERRUPTIBLE);
David Howells07fe7cb2009-04-03 16:42:35 +0100447 if (!freezing(current) &&
448 !slow_work_threads_should_exit &&
David Howells109d9272009-04-03 16:42:35 +0100449 !slow_work_available(vsmax) &&
450 !slow_work_cull)
David Howells07fe7cb2009-04-03 16:42:35 +0100451 schedule();
452 finish_wait(&slow_work_thread_wq, &wait);
453
454 try_to_freeze();
455
456 vsmax = vslow_work_proportion;
457 vsmax *= atomic_read(&slow_work_thread_count);
458 vsmax /= 100;
459
David Howells3d7a6412009-11-19 18:10:23 +0000460 if (slow_work_available(vsmax) && slow_work_execute(id)) {
David Howells07fe7cb2009-04-03 16:42:35 +0100461 cond_resched();
David Howells109d9272009-04-03 16:42:35 +0100462 if (list_empty(&slow_work_queue) &&
463 list_empty(&vslow_work_queue) &&
464 atomic_read(&slow_work_thread_count) >
465 slow_work_min_threads)
Chris Peterson009789f2009-06-16 15:33:43 -0700466 slow_work_schedule_cull();
David Howells07fe7cb2009-04-03 16:42:35 +0100467 continue;
468 }
469
470 if (slow_work_threads_should_exit)
471 break;
David Howells109d9272009-04-03 16:42:35 +0100472
473 if (slow_work_cull && slow_work_cull_thread())
474 break;
David Howells07fe7cb2009-04-03 16:42:35 +0100475 }
476
David Howells3d7a6412009-11-19 18:10:23 +0000477 spin_lock_irq(&slow_work_queue_lock);
478 __clear_bit(id, slow_work_ids);
479 spin_unlock_irq(&slow_work_queue_lock);
480
David Howells07fe7cb2009-04-03 16:42:35 +0100481 if (atomic_dec_and_test(&slow_work_thread_count))
482 complete_and_exit(&slow_work_last_thread_exited, 0);
483 return 0;
484}
485
David Howells109d9272009-04-03 16:42:35 +0100486/*
487 * Handle thread cull timer expiration
488 */
489static void slow_work_cull_timeout(unsigned long data)
490{
491 slow_work_cull = true;
492 wake_up(&slow_work_thread_wq);
493}
494
495/*
David Howells109d9272009-04-03 16:42:35 +0100496 * Start a new slow work thread
497 */
498static void slow_work_new_thread_execute(struct slow_work *work)
499{
500 struct task_struct *p;
501
502 if (slow_work_threads_should_exit)
503 return;
504
505 if (atomic_read(&slow_work_thread_count) >= slow_work_max_threads)
506 return;
507
508 if (!mutex_trylock(&slow_work_user_lock))
509 return;
510
511 slow_work_may_not_start_new_thread = true;
512 atomic_inc(&slow_work_thread_count);
513 p = kthread_run(slow_work_thread, NULL, "kslowd");
514 if (IS_ERR(p)) {
515 printk(KERN_DEBUG "Slow work thread pool: OOM\n");
516 if (atomic_dec_and_test(&slow_work_thread_count))
517 BUG(); /* we're running on a slow work thread... */
518 mod_timer(&slow_work_oom_timer,
Chris Peterson009789f2009-06-16 15:33:43 -0700519 round_jiffies(jiffies + SLOW_WORK_OOM_TIMEOUT));
David Howells109d9272009-04-03 16:42:35 +0100520 } else {
521 /* ratelimit the starting of new threads */
522 mod_timer(&slow_work_oom_timer, jiffies + 1);
523 }
524
525 mutex_unlock(&slow_work_user_lock);
526}
527
528static const struct slow_work_ops slow_work_new_thread_ops = {
David Howells3d7a6412009-11-19 18:10:23 +0000529 .owner = THIS_MODULE,
David Howells109d9272009-04-03 16:42:35 +0100530 .execute = slow_work_new_thread_execute,
531};
532
533/*
534 * post-OOM new thread start suppression expiration
535 */
536static void slow_work_oom_timeout(unsigned long data)
537{
538 slow_work_may_not_start_new_thread = false;
539}
540
David Howells12e22c52009-04-03 16:42:35 +0100541#ifdef CONFIG_SYSCTL
542/*
543 * Handle adjustment of the minimum number of threads
544 */
545static int slow_work_min_threads_sysctl(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700546 void __user *buffer,
David Howells12e22c52009-04-03 16:42:35 +0100547 size_t *lenp, loff_t *ppos)
548{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700549 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Howells12e22c52009-04-03 16:42:35 +0100550 int n;
551
552 if (ret == 0) {
553 mutex_lock(&slow_work_user_lock);
554 if (slow_work_user_count > 0) {
555 /* see if we need to start or stop threads */
556 n = atomic_read(&slow_work_thread_count) -
557 slow_work_min_threads;
558
559 if (n < 0 && !slow_work_may_not_start_new_thread)
560 slow_work_enqueue(&slow_work_new_thread);
561 else if (n > 0)
Chris Peterson009789f2009-06-16 15:33:43 -0700562 slow_work_schedule_cull();
David Howells12e22c52009-04-03 16:42:35 +0100563 }
564 mutex_unlock(&slow_work_user_lock);
565 }
566
567 return ret;
568}
569
570/*
571 * Handle adjustment of the maximum number of threads
572 */
573static int slow_work_max_threads_sysctl(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700574 void __user *buffer,
David Howells12e22c52009-04-03 16:42:35 +0100575 size_t *lenp, loff_t *ppos)
576{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -0700577 int ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
David Howells12e22c52009-04-03 16:42:35 +0100578 int n;
579
580 if (ret == 0) {
581 mutex_lock(&slow_work_user_lock);
582 if (slow_work_user_count > 0) {
583 /* see if we need to stop threads */
584 n = slow_work_max_threads -
585 atomic_read(&slow_work_thread_count);
586
587 if (n < 0)
Chris Peterson009789f2009-06-16 15:33:43 -0700588 slow_work_schedule_cull();
David Howells12e22c52009-04-03 16:42:35 +0100589 }
590 mutex_unlock(&slow_work_user_lock);
591 }
592
593 return ret;
594}
595#endif /* CONFIG_SYSCTL */
596
David Howells07fe7cb2009-04-03 16:42:35 +0100597/**
598 * slow_work_register_user - Register a user of the facility
David Howells3d7a6412009-11-19 18:10:23 +0000599 * @module: The module about to make use of the facility
David Howells07fe7cb2009-04-03 16:42:35 +0100600 *
601 * Register a user of the facility, starting up the initial threads if there
602 * aren't any other users at this point. This will return 0 if successful, or
603 * an error if not.
604 */
David Howells3d7a6412009-11-19 18:10:23 +0000605int slow_work_register_user(struct module *module)
David Howells07fe7cb2009-04-03 16:42:35 +0100606{
607 struct task_struct *p;
608 int loop;
609
610 mutex_lock(&slow_work_user_lock);
611
612 if (slow_work_user_count == 0) {
613 printk(KERN_NOTICE "Slow work thread pool: Starting up\n");
614 init_completion(&slow_work_last_thread_exited);
615
616 slow_work_threads_should_exit = false;
David Howells109d9272009-04-03 16:42:35 +0100617 slow_work_init(&slow_work_new_thread,
618 &slow_work_new_thread_ops);
619 slow_work_may_not_start_new_thread = false;
620 slow_work_cull = false;
David Howells07fe7cb2009-04-03 16:42:35 +0100621
622 /* start the minimum number of threads */
623 for (loop = 0; loop < slow_work_min_threads; loop++) {
624 atomic_inc(&slow_work_thread_count);
625 p = kthread_run(slow_work_thread, NULL, "kslowd");
626 if (IS_ERR(p))
627 goto error;
628 }
629 printk(KERN_NOTICE "Slow work thread pool: Ready\n");
630 }
631
632 slow_work_user_count++;
633 mutex_unlock(&slow_work_user_lock);
634 return 0;
635
636error:
637 if (atomic_dec_and_test(&slow_work_thread_count))
638 complete(&slow_work_last_thread_exited);
639 if (loop > 0) {
640 printk(KERN_ERR "Slow work thread pool:"
641 " Aborting startup on ENOMEM\n");
642 slow_work_threads_should_exit = true;
643 wake_up_all(&slow_work_thread_wq);
644 wait_for_completion(&slow_work_last_thread_exited);
645 printk(KERN_ERR "Slow work thread pool: Aborted\n");
646 }
647 mutex_unlock(&slow_work_user_lock);
648 return PTR_ERR(p);
649}
650EXPORT_SYMBOL(slow_work_register_user);
651
David Howells3d7a6412009-11-19 18:10:23 +0000652/*
653 * wait for all outstanding items from the calling module to complete
654 * - note that more items may be queued whilst we're waiting
655 */
656static void slow_work_wait_for_items(struct module *module)
657{
658 DECLARE_WAITQUEUE(myself, current);
659 struct slow_work *work;
660 int loop;
661
662 mutex_lock(&slow_work_unreg_sync_lock);
663 add_wait_queue(&slow_work_unreg_wq, &myself);
664
665 for (;;) {
666 spin_lock_irq(&slow_work_queue_lock);
667
668 /* first of all, we wait for the last queued item in each list
669 * to be processed */
670 list_for_each_entry_reverse(work, &vslow_work_queue, link) {
671 if (work->owner == module) {
672 set_current_state(TASK_UNINTERRUPTIBLE);
673 slow_work_unreg_work_item = work;
674 goto do_wait;
675 }
676 }
677 list_for_each_entry_reverse(work, &slow_work_queue, link) {
678 if (work->owner == module) {
679 set_current_state(TASK_UNINTERRUPTIBLE);
680 slow_work_unreg_work_item = work;
681 goto do_wait;
682 }
683 }
684
685 /* then we wait for the items being processed to finish */
686 slow_work_unreg_module = module;
687 smp_mb();
688 for (loop = 0; loop < SLOW_WORK_THREAD_LIMIT; loop++) {
689 if (slow_work_thread_processing[loop] == module)
690 goto do_wait;
691 }
692 spin_unlock_irq(&slow_work_queue_lock);
693 break; /* okay, we're done */
694
695 do_wait:
696 spin_unlock_irq(&slow_work_queue_lock);
697 schedule();
698 slow_work_unreg_work_item = NULL;
699 slow_work_unreg_module = NULL;
700 }
701
702 remove_wait_queue(&slow_work_unreg_wq, &myself);
703 mutex_unlock(&slow_work_unreg_sync_lock);
704}
705
David Howells07fe7cb2009-04-03 16:42:35 +0100706/**
707 * slow_work_unregister_user - Unregister a user of the facility
David Howells3d7a6412009-11-19 18:10:23 +0000708 * @module: The module whose items should be cleared
David Howells07fe7cb2009-04-03 16:42:35 +0100709 *
710 * Unregister a user of the facility, killing all the threads if this was the
711 * last one.
David Howells3d7a6412009-11-19 18:10:23 +0000712 *
713 * This waits for all the work items belonging to the nominated module to go
714 * away before proceeding.
David Howells07fe7cb2009-04-03 16:42:35 +0100715 */
David Howells3d7a6412009-11-19 18:10:23 +0000716void slow_work_unregister_user(struct module *module)
David Howells07fe7cb2009-04-03 16:42:35 +0100717{
David Howells3d7a6412009-11-19 18:10:23 +0000718 /* first of all, wait for all outstanding items from the calling module
719 * to complete */
720 if (module)
721 slow_work_wait_for_items(module);
722
723 /* then we can actually go about shutting down the facility if need
724 * be */
David Howells07fe7cb2009-04-03 16:42:35 +0100725 mutex_lock(&slow_work_user_lock);
726
727 BUG_ON(slow_work_user_count <= 0);
728
729 slow_work_user_count--;
730 if (slow_work_user_count == 0) {
731 printk(KERN_NOTICE "Slow work thread pool: Shutting down\n");
732 slow_work_threads_should_exit = true;
Jonathan Corbet418df632009-04-22 12:01:49 +0100733 del_timer_sync(&slow_work_cull_timer);
734 del_timer_sync(&slow_work_oom_timer);
David Howells07fe7cb2009-04-03 16:42:35 +0100735 wake_up_all(&slow_work_thread_wq);
736 wait_for_completion(&slow_work_last_thread_exited);
737 printk(KERN_NOTICE "Slow work thread pool:"
738 " Shut down complete\n");
739 }
740
David Howells07fe7cb2009-04-03 16:42:35 +0100741 mutex_unlock(&slow_work_user_lock);
742}
743EXPORT_SYMBOL(slow_work_unregister_user);
744
745/*
746 * Initialise the slow work facility
747 */
748static int __init init_slow_work(void)
749{
750 unsigned nr_cpus = num_possible_cpus();
751
David Howells12e22c52009-04-03 16:42:35 +0100752 if (slow_work_max_threads < nr_cpus)
David Howells07fe7cb2009-04-03 16:42:35 +0100753 slow_work_max_threads = nr_cpus;
David Howells12e22c52009-04-03 16:42:35 +0100754#ifdef CONFIG_SYSCTL
755 if (slow_work_max_max_threads < nr_cpus * 2)
756 slow_work_max_max_threads = nr_cpus * 2;
757#endif
David Howells07fe7cb2009-04-03 16:42:35 +0100758 return 0;
759}
760
761subsys_initcall(init_slow_work);