summaryrefslogtreecommitdiff
path: root/kernel/slow-work.c
blob: 454abb21c8bd9e96747131ab644412508594f732 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
/* Worker thread pool for slow items, such as filesystem lookups or mkdirs
 *
 * Copyright (C) 2008 Red Hat, Inc. All Rights Reserved.
 * Written by David Howells (dhowells@redhat.com)
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public Licence
 * as published by the Free Software Foundation; either version
 * 2 of the Licence, or (at your option) any later version.
 */

#include <linux/module.h>
#include <linux/slow-work.h>
#include <linux/kthread.h>
#include <linux/freezer.h>
#include <linux/wait.h>
#include <asm/system.h>

#define SLOW_WORK_CULL_TIMEOUT (5 * HZ)	/* cull threads 5s after running out of
					 * things to do */
#define SLOW_WORK_OOM_TIMEOUT (5 * HZ)	/* can't start new threads for 5s after
					 * OOM */

static void slow_work_cull_timeout(unsigned long);
static void slow_work_oom_timeout(unsigned long);

/*
 * The pool of threads has at least min threads in it as long as someone is
 * using the facility, and may have as many as max.
 *
 * A portion of the pool may be processing very slow operations.
 */
static unsigned slow_work_min_threads = 2;
static unsigned slow_work_max_threads = 4;
static unsigned vslow_work_proportion = 50; /* % of threads that may process
					     * very slow work */
static atomic_t slow_work_thread_count;
static atomic_t vslow_work_executing_count;

static bool slow_work_may_not_start_new_thread;
static bool slow_work_cull; /* cull a thread due to lack of activity */
static DEFINE_TIMER(slow_work_cull_timer, slow_work_cull_timeout, 0, 0);
static DEFINE_TIMER(slow_work_oom_timer, slow_work_oom_timeout, 0, 0);
static struct slow_work slow_work_new_thread; /* new thread starter */

/*
 * The queues of work items and the lock governing access to them.  These are
 * shared between all the CPUs.  It doesn't make sense to have per-CPU queues
 * as the number of threads bears no relation to the number of CPUs.
 *
 * There are two queues of work items: one for slow work items, and one for
 * very slow work items.
 */
static LIST_HEAD(slow_work_queue);
static LIST_HEAD(vslow_work_queue);
static DEFINE_SPINLOCK(slow_work_queue_lock);

/*
 * The thread controls.  A variable used to signal to the threads that they
 * should exit when the queue is empty, a waitqueue used by the threads to wait
 * for signals, and a completion set by the last thread to exit.
 */
static bool slow_work_threads_should_exit;
static DECLARE_WAIT_QUEUE_HEAD(slow_work_thread_wq);
static DECLARE_COMPLETION(slow_work_last_thread_exited);

/*
 * The number of users of the thread pool and its lock.  Whilst this is zero we
 * have no threads hanging around, and when this reaches zero, we wait for all
 * active or queued work items to complete and kill all the threads we do have.
 */
static int slow_work_user_count;
static DEFINE_MUTEX(slow_work_user_lock);

/*
 * Calculate the maximum number of active threads in the pool that are
 * permitted to process very slow work items.
 *
 * The answer is rounded up to at least 1, but may not equal or exceed the
 * maximum number of the threads in the pool.  This means we always have at
 * least one thread that can process slow work items, and we always have at
 * least one thread that won't get tied up doing so.
 */
static unsigned slow_work_calc_vsmax(void)
{
	unsigned vsmax;

	vsmax = atomic_read(&slow_work_thread_count) * vslow_work_proportion;
	vsmax /= 100;
	vsmax = max(vsmax, 1U);
	return min(vsmax, slow_work_max_threads - 1);
}

/*
 * Attempt to execute stuff queued on a slow thread.  Return true if we managed
 * it, false if there was nothing to do.
 */
static bool slow_work_execute(void)
{
	struct slow_work *work = NULL;
	unsigned vsmax;
	bool very_slow;

	vsmax = slow_work_calc_vsmax();

	/* see if we can schedule a new thread to be started if we're not
	 * keeping up with the work */
	if (!waitqueue_active(&slow_work_thread_wq) &&
	    (!list_empty(&slow_work_queue) || !list_empty(&vslow_work_queue)) &&
	    atomic_read(&slow_work_thread_count) < slow_work_max_threads &&
	    !slow_work_may_not_start_new_thread)
		slow_work_enqueue(&slow_work_new_thread);

	/* find something to execute */
	spin_lock_irq(&slow_work_queue_lock);
	if (!list_empty(&vslow_work_queue) &&
	    atomic_read(&vslow_work_executing_count) < vsmax) {
		work = list_entry(vslow_work_queue.next,
				  struct slow_work, link);
		if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags))
			BUG();
		list_del_init(&work->link);
		atomic_inc(&vslow_work_executing_count);
		very_slow = true;
	} else if (!list_empty(&slow_work_queue)) {
		work = list_entry(slow_work_queue.next,
				  struct slow_work, link);
		if (test_and_set_bit_lock(SLOW_WORK_EXECUTING, &work->flags))
			BUG();
		list_del_init(&work->link);
		very_slow = false;
	} else {
		very_slow = false; /* avoid the compiler warning */
	}
	spin_unlock_irq(&slow_work_queue_lock);

	if (!work)
		return false;

	if (!test_and_clear_bit(SLOW_WORK_PENDING, &work->flags))
		BUG();

	work->ops->execute(work);

	if (very_slow)
		atomic_dec(&vslow_work_executing_count);
	clear_bit_unlock(SLOW_WORK_EXECUTING, &work->flags);

	/* if someone tried to enqueue the item whilst we were executing it,
	 * then it'll be left unenqueued to avoid multiple threads trying to
	 * execute it simultaneously
	 *
	 * there is, however, a race between us testing the pending flag and
	 * getting the spinlock, and between the enqueuer setting the pending
	 * flag and getting the spinlock, so we use a deferral bit to tell us
	 * if the enqueuer got there first
	 */
	if (test_bit(SLOW_WORK_PENDING, &work->flags)) {
		spin_lock_irq(&slow_work_queue_lock);

		if (!test_bit(SLOW_WORK_EXECUTING, &work->flags) &&
		    test_and_clear_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags))
			goto auto_requeue;

		spin_unlock_irq(&slow_work_queue_lock);
	}

	work->ops->put_ref(work);
	return true;

auto_requeue:
	/* we must complete the enqueue operation
	 * - we transfer our ref on the item back to the appropriate queue
	 * - don't wake another thread up as we're awake already
	 */
	if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
		list_add_tail(&work->link, &vslow_work_queue);
	else
		list_add_tail(&work->link, &slow_work_queue);
	spin_unlock_irq(&slow_work_queue_lock);
	return true;
}

/**
 * slow_work_enqueue - Schedule a slow work item for processing
 * @work: The work item to queue
 *
 * Schedule a slow work item for processing.  If the item is already undergoing
 * execution, this guarantees not to re-enter the execution routine until the
 * first execution finishes.
 *
 * The item is pinned by this function as it retains a reference to it, managed
 * through the item operations.  The item is unpinned once it has been
 * executed.
 *
 * An item may hog the thread that is running it for a relatively large amount
 * of time, sufficient, for example, to perform several lookup, mkdir, create
 * and setxattr operations.  It may sleep on I/O and may sleep to obtain locks.
 *
 * Conversely, if a number of items are awaiting processing, it may take some
 * time before any given item is given attention.  The number of threads in the
 * pool may be increased to deal with demand, but only up to a limit.
 *
 * If SLOW_WORK_VERY_SLOW is set on the work item, then it will be placed in
 * the very slow queue, from which only a portion of the threads will be
 * allowed to pick items to execute.  This ensures that very slow items won't
 * overly block ones that are just ordinarily slow.
 *
 * Returns 0 if successful, -EAGAIN if not.
 */
int slow_work_enqueue(struct slow_work *work)
{
	unsigned long flags;

	BUG_ON(slow_work_user_count <= 0);
	BUG_ON(!work);
	BUG_ON(!work->ops);
	BUG_ON(!work->ops->get_ref);

	/* when honouring an enqueue request, we only promise that we will run
	 * the work function in the future; we do not promise to run it once
	 * per enqueue request
	 *
	 * we use the PENDING bit to merge together repeat requests without
	 * having to disable IRQs and take the spinlock, whilst still
	 * maintaining our promise
	 */
	if (!test_and_set_bit_lock(SLOW_WORK_PENDING, &work->flags)) {
		spin_lock_irqsave(&slow_work_queue_lock, flags);

		/* we promise that we will not attempt to execute the work
		 * function in more than one thread simultaneously
		 *
		 * this, however, leaves us with a problem if we're asked to
		 * enqueue the work whilst someone is executing the work
		 * function as simply queueing the work immediately means that
		 * another thread may try executing it whilst it is already
		 * under execution
		 *
		 * to deal with this, we set the ENQ_DEFERRED bit instead of
		 * enqueueing, and the thread currently executing the work
		 * function will enqueue the work item when the work function
		 * returns and it has cleared the EXECUTING bit
		 */
		if (test_bit(SLOW_WORK_EXECUTING, &work->flags)) {
			set_bit(SLOW_WORK_ENQ_DEFERRED, &work->flags);
		} else {
			if (work->ops->get_ref(work) < 0)
				goto cant_get_ref;
			if (test_bit(SLOW_WORK_VERY_SLOW, &work->flags))
				list_add_tail(&work->link, &vslow_work_queue);
			else
				list_add_tail(&work->link, &slow_work_queue);
			wake_up(&slow_work_thread_wq);
		}

		spin_unlock_irqrestore(&slow_work_queue_lock, flags);
	}
	return 0;

cant_get_ref:
	spin_unlock_irqrestore(&slow_work_queue_lock, flags);
	return -EAGAIN;
}
EXPORT_SYMBOL(slow_work_enqueue);

/*
 * Worker thread culling algorithm
 */
static bool slow_work_cull_thread(void)
{
	unsigned long flags;
	bool do_cull = false;

	spin_lock_irqsave(&slow_work_queue_lock, flags);

	if (slow_work_cull) {
		slow_work_cull = false;

		if (list_empty(&slow_work_queue) &&
		    list_empty(&vslow_work_queue) &&
		    atomic_read(&slow_work_thread_count) >
		    slow_work_min_threads) {
			mod_timer(&slow_work_cull_timer,
				  jiffies + SLOW_WORK_CULL_TIMEOUT);
			do_cull = true;
		}
	}

	spin_unlock_irqrestore(&slow_work_queue_lock, flags);
	return do_cull;
}

/*
 * Determine if there is slow work available for dispatch
 */
static inline bool slow_work_available(int vsmax)
{
	return !list_empty(&slow_work_queue) ||
		(!list_empty(&vslow_work_queue) &&
		 atomic_read(&vslow_work_executing_count) < vsmax);
}

/*
 * Worker thread dispatcher
 */
static int slow_work_thread(void *_data)
{
	int vsmax;

	DEFINE_WAIT(wait);

	set_freezable();
	set_user_nice(current, -5);

	for (;;) {
		vsmax = vslow_work_proportion;
		vsmax *= atomic_read(&slow_work_thread_count);
		vsmax /= 100;

		prepare_to_wait(&slow_work_thread_wq, &wait,
				TASK_INTERRUPTIBLE);
		if (!freezing(current) &&
		    !slow_work_threads_should_exit &&
		    !slow_work_available(vsmax) &&
		    !slow_work_cull)
			schedule();
		finish_wait(&slow_work_thread_wq, &wait);

		try_to_freeze();

		vsmax = vslow_work_proportion;
		vsmax *= atomic_read(&slow_work_thread_count);
		vsmax /= 100;

		if (slow_work_available(vsmax) && slow_work_execute()) {
			cond_resched();
			if (list_empty(&slow_work_queue) &&
			    list_empty(&vslow_work_queue) &&
			    atomic_read(&slow_work_thread_count) >
			    slow_work_min_threads)
				mod_timer(&slow_work_cull_timer,
					  jiffies + SLOW_WORK_CULL_TIMEOUT);
			continue;
		}

		if (slow_work_threads_should_exit)
			break;

		if (slow_work_cull && slow_work_cull_thread())
			break;
	}

	if (atomic_dec_and_test(&slow_work_thread_count))
		complete_and_exit(&slow_work_last_thread_exited, 0);
	return 0;
}

/*
 * Handle thread cull timer expiration
 */
static void slow_work_cull_timeout(unsigned long data)
{
	slow_work_cull = true;
	wake_up(&slow_work_thread_wq);
}

/*
 * Get a reference on slow work thread starter
 */
static int slow_work_new_thread_get_ref(struct slow_work *work)
{
	return 0;
}

/*
 * Drop a reference on slow work thread starter
 */
static void slow_work_new_thread_put_ref(struct slow_work *work)
{
}

/*
 * Start a new slow work thread
 */
static void slow_work_new_thread_execute(struct slow_work *work)
{
	struct task_struct *p;

	if (slow_work_threads_should_exit)
		return;

	if (atomic_read(&slow_work_thread_count) >= slow_work_max_threads)
		return;

	if (!mutex_trylock(&slow_work_user_lock))
		return;

	slow_work_may_not_start_new_thread = true;
	atomic_inc(&slow_work_thread_count);
	p = kthread_run(slow_work_thread, NULL, "kslowd");
	if (IS_ERR(p)) {
		printk(KERN_DEBUG "Slow work thread pool: OOM\n");
		if (atomic_dec_and_test(&slow_work_thread_count))
			BUG(); /* we're running on a slow work thread... */
		mod_timer(&slow_work_oom_timer,
			  jiffies + SLOW_WORK_OOM_TIMEOUT);
	} else {
		/* ratelimit the starting of new threads */
		mod_timer(&slow_work_oom_timer, jiffies + 1);
	}

	mutex_unlock(&slow_work_user_lock);
}

static const struct slow_work_ops slow_work_new_thread_ops = {
	.get_ref	= slow_work_new_thread_get_ref,
	.put_ref	= slow_work_new_thread_put_ref,
	.execute	= slow_work_new_thread_execute,
};

/*
 * post-OOM new thread start suppression expiration
 */
static void slow_work_oom_timeout(unsigned long data)
{
	slow_work_may_not_start_new_thread = false;
}

/**
 * slow_work_register_user - Register a user of the facility
 *
 * Register a user of the facility, starting up the initial threads if there
 * aren't any other users at this point.  This will return 0 if successful, or
 * an error if not.
 */
int slow_work_register_user(void)
{
	struct task_struct *p;
	int loop;

	mutex_lock(&slow_work_user_lock);

	if (slow_work_user_count == 0) {
		printk(KERN_NOTICE "Slow work thread pool: Starting up\n");
		init_completion(&slow_work_last_thread_exited);

		slow_work_threads_should_exit = false;
		slow_work_init(&slow_work_new_thread,
			       &slow_work_new_thread_ops);
		slow_work_may_not_start_new_thread = false;
		slow_work_cull = false;

		/* start the minimum number of threads */
		for (loop = 0; loop < slow_work_min_threads; loop++) {
			atomic_inc(&slow_work_thread_count);
			p = kthread_run(slow_work_thread, NULL, "kslowd");
			if (IS_ERR(p))
				goto error;
		}
		printk(KERN_NOTICE "Slow work thread pool: Ready\n");
	}

	slow_work_user_count++;
	mutex_unlock(&slow_work_user_lock);
	return 0;

error:
	if (atomic_dec_and_test(&slow_work_thread_count))
		complete(&slow_work_last_thread_exited);
	if (loop > 0) {
		printk(KERN_ERR "Slow work thread pool:"
		       " Aborting startup on ENOMEM\n");
		slow_work_threads_should_exit = true;
		wake_up_all(&slow_work_thread_wq);
		wait_for_completion(&slow_work_last_thread_exited);
		printk(KERN_ERR "Slow work thread pool: Aborted\n");
	}
	mutex_unlock(&slow_work_user_lock);
	return PTR_ERR(p);
}
EXPORT_SYMBOL(slow_work_register_user);

/**
 * slow_work_unregister_user - Unregister a user of the facility
 *
 * Unregister a user of the facility, killing all the threads if this was the
 * last one.
 */
void slow_work_unregister_user(void)
{
	mutex_lock(&slow_work_user_lock);

	BUG_ON(slow_work_user_count <= 0);

	slow_work_user_count--;
	if (slow_work_user_count == 0) {
		printk(KERN_NOTICE "Slow work thread pool: Shutting down\n");
		slow_work_threads_should_exit = true;
		wake_up_all(&slow_work_thread_wq);
		wait_for_completion(&slow_work_last_thread_exited);
		printk(KERN_NOTICE "Slow work thread pool:"
		       " Shut down complete\n");
	}

	del_timer_sync(&slow_work_cull_timer);

	mutex_unlock(&slow_work_user_lock);
}
EXPORT_SYMBOL(slow_work_unregister_user);

/*
 * Initialise the slow work facility
 */
static int __init init_slow_work(void)
{
	unsigned nr_cpus = num_possible_cpus();

	if (nr_cpus > slow_work_max_threads)
		slow_work_max_threads = nr_cpus;
	return 0;
}

subsys_initcall(init_slow_work);