summaryrefslogtreecommitdiff
path: root/fs/xfs/xfs_mru_cache.c
blob: 4aa31664472123ff4bfca273fb5f6f9415f67c2d (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
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
/*
 * Copyright (c) 2006-2007 Silicon Graphics, Inc.
 * All Rights Reserved.
 *
 * This program is free software; you can redistribute it and/or
 * modify it under the terms of the GNU General Public License as
 * published by the Free Software Foundation.
 *
 * This program is distributed in the hope that it would be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program; if not, write the Free Software Foundation,
 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
 */
#include "xfs.h"
#include "xfs_mru_cache.h"

/*
 * The MRU Cache data structure consists of a data store, an array of lists and
 * a lock to protect its internal state.  At initialisation time, the client
 * supplies an element lifetime in milliseconds and a group count, as well as a
 * function pointer to call when deleting elements.  A data structure for
 * queueing up work in the form of timed callbacks is also included.
 *
 * The group count controls how many lists are created, and thereby how finely
 * the elements are grouped in time.  When reaping occurs, all the elements in
 * all the lists whose time has expired are deleted.
 *
 * To give an example of how this works in practice, consider a client that
 * initialises an MRU Cache with a lifetime of ten seconds and a group count of
 * five.  Five internal lists will be created, each representing a two second
 * period in time.  When the first element is added, time zero for the data
 * structure is initialised to the current time.
 *
 * All the elements added in the first two seconds are appended to the first
 * list.  Elements added in the third second go into the second list, and so on.
 * If an element is accessed at any point, it is removed from its list and
 * inserted at the head of the current most-recently-used list.
 *
 * The reaper function will have nothing to do until at least twelve seconds
 * have elapsed since the first element was added.  The reason for this is that
 * if it were called at t=11s, there could be elements in the first list that
 * have only been inactive for nine seconds, so it still does nothing.  If it is
 * called anywhere between t=12 and t=14 seconds, it will delete all the
 * elements that remain in the first list.  It's therefore possible for elements
 * to remain in the data store even after they've been inactive for up to
 * (t + t/g) seconds, where t is the inactive element lifetime and g is the
 * number of groups.
 *
 * The above example assumes that the reaper function gets called at least once
 * every (t/g) seconds.  If it is called less frequently, unused elements will
 * accumulate in the reap list until the reaper function is eventually called.
 * The current implementation uses work queue callbacks to carefully time the
 * reaper function calls, so this should happen rarely, if at all.
 *
 * From a design perspective, the primary reason for the choice of a list array
 * representing discrete time intervals is that it's only practical to reap
 * expired elements in groups of some appreciable size.  This automatically
 * introduces a granularity to element lifetimes, so there's no point storing an
 * individual timeout with each element that specifies a more precise reap time.
 * The bonus is a saving of sizeof(long) bytes of memory per element stored.
 *
 * The elements could have been stored in just one list, but an array of
 * counters or pointers would need to be maintained to allow them to be divided
 * up into discrete time groups.  More critically, the process of touching or
 * removing an element would involve walking large portions of the entire list,
 * which would have a detrimental effect on performance.  The additional memory
 * requirement for the array of list heads is minimal.
 *
 * When an element is touched or deleted, it needs to be removed from its
 * current list.  Doubly linked lists are used to make the list maintenance
 * portion of these operations O(1).  Since reaper timing can be imprecise,
 * inserts and lookups can occur when there are no free lists available.  When
 * this happens, all the elements on the LRU list need to be migrated to the end
 * of the reap list.  To keep the list maintenance portion of these operations
 * O(1) also, list tails need to be accessible without walking the entire list.
 * This is the reason why doubly linked list heads are used.
 */

/*
 * An MRU Cache is a dynamic data structure that stores its elements in a way
 * that allows efficient lookups, but also groups them into discrete time
 * intervals based on insertion time.  This allows elements to be efficiently
 * and automatically reaped after a fixed period of inactivity.
 *
 * When a client data pointer is stored in the MRU Cache it needs to be added to
 * both the data store and to one of the lists.  It must also be possible to
 * access each of these entries via the other, i.e. to:
 *
 *    a) Walk a list, removing the corresponding data store entry for each item.
 *    b) Look up a data store entry, then access its list entry directly.
 *
 * To achieve both of these goals, each entry must contain both a list entry and
 * a key, in addition to the user's data pointer.  Note that it's not a good
 * idea to have the client embed one of these structures at the top of their own
 * data structure, because inserting the same item more than once would most
 * likely result in a loop in one of the lists.  That's a sure-fire recipe for
 * an infinite loop in the code.
 */
typedef struct xfs_mru_cache_elem
{
	struct list_head list_node;
	unsigned long	key;
	void		*value;
} xfs_mru_cache_elem_t;

static kmem_zone_t		*xfs_mru_elem_zone;
static struct workqueue_struct	*xfs_mru_reap_wq;

/*
 * When inserting, destroying or reaping, it's first necessary to update the
 * lists relative to a particular time.  In the case of destroying, that time
 * will be well in the future to ensure that all items are moved to the reap
 * list.  In all other cases though, the time will be the current time.
 *
 * This function enters a loop, moving the contents of the LRU list to the reap
 * list again and again until either a) the lists are all empty, or b) time zero
 * has been advanced sufficiently to be within the immediate element lifetime.
 *
 * Case a) above is detected by counting how many groups are migrated and
 * stopping when they've all been moved.  Case b) is detected by monitoring the
 * time_zero field, which is updated as each group is migrated.
 *
 * The return value is the earliest time that more migration could be needed, or
 * zero if there's no need to schedule more work because the lists are empty.
 */
STATIC unsigned long
_xfs_mru_cache_migrate(
	xfs_mru_cache_t	*mru,
	unsigned long	now)
{
	unsigned int	grp;
	unsigned int	migrated = 0;
	struct list_head *lru_list;

	/* Nothing to do if the data store is empty. */
	if (!mru->time_zero)
		return 0;

	/* While time zero is older than the time spanned by all the lists. */
	while (mru->time_zero <= now - mru->grp_count * mru->grp_time) {

		/*
		 * If the LRU list isn't empty, migrate its elements to the tail
		 * of the reap list.
		 */
		lru_list = mru->lists + mru->lru_grp;
		if (!list_empty(lru_list))
			list_splice_init(lru_list, mru->reap_list.prev);

		/*
		 * Advance the LRU group number, freeing the old LRU list to
		 * become the new MRU list; advance time zero accordingly.
		 */
		mru->lru_grp = (mru->lru_grp + 1) % mru->grp_count;
		mru->time_zero += mru->grp_time;

		/*
		 * If reaping is so far behind that all the elements on all the
		 * lists have been migrated to the reap list, it's now empty.
		 */
		if (++migrated == mru->grp_count) {
			mru->lru_grp = 0;
			mru->time_zero = 0;
			return 0;
		}
	}

	/* Find the first non-empty list from the LRU end. */
	for (grp = 0; grp < mru->grp_count; grp++) {

		/* Check the grp'th list from the LRU end. */
		lru_list = mru->lists + ((mru->lru_grp + grp) % mru->grp_count);
		if (!list_empty(lru_list))
			return mru->time_zero +
			       (mru->grp_count + grp) * mru->grp_time;
	}

	/* All the lists must be empty. */
	mru->lru_grp = 0;
	mru->time_zero = 0;
	return 0;
}

/*
 * When inserting or doing a lookup, an element needs to be inserted into the
 * MRU list.  The lists must be migrated first to ensure that they're
 * up-to-date, otherwise the new element could be given a shorter lifetime in
 * the cache than it should.
 */
STATIC void
_xfs_mru_cache_list_insert(
	xfs_mru_cache_t		*mru,
	xfs_mru_cache_elem_t	*elem)
{
	unsigned int	grp = 0;
	unsigned long	now = jiffies;

	/*
	 * If the data store is empty, initialise time zero, leave grp set to
	 * zero and start the work queue timer if necessary.  Otherwise, set grp
	 * to the number of group times that have elapsed since time zero.
	 */
	if (!_xfs_mru_cache_migrate(mru, now)) {
		mru->time_zero = now;
		if (!mru->queued) {
			mru->queued = 1;
			queue_delayed_work(xfs_mru_reap_wq, &mru->work,
			                   mru->grp_count * mru->grp_time);
		}
	} else {
		grp = (now - mru->time_zero) / mru->grp_time;
		grp = (mru->lru_grp + grp) % mru->grp_count;
	}

	/* Insert the element at the tail of the corresponding list. */
	list_add_tail(&elem->list_node, mru->lists + grp);
}

/*
 * When destroying or reaping, all the elements that were migrated to the reap
 * list need to be deleted.  For each element this involves removing it from the
 * data store, removing it from the reap list, calling the client's free
 * function and deleting the element from the element zone.
 *
 * We get called holding the mru->lock, which we drop and then reacquire.
 * Sparse need special help with this to tell it we know what we are doing.
 */
STATIC void
_xfs_mru_cache_clear_reap_list(
	xfs_mru_cache_t		*mru) __releases(mru->lock) __acquires(mru->lock)

{
	xfs_mru_cache_elem_t	*elem, *next;
	struct list_head	tmp;

	INIT_LIST_HEAD(&tmp);
	list_for_each_entry_safe(elem, next, &mru->reap_list, list_node) {

		/* Remove the element from the data store. */
		radix_tree_delete(&mru->store, elem->key);

		/*
		 * remove to temp list so it can be freed without
		 * needing to hold the lock
		 */
		list_move(&elem->list_node, &tmp);
	}
	spin_unlock(&mru->lock);

	list_for_each_entry_safe(elem, next, &tmp, list_node) {

		/* Remove the element from the reap list. */
		list_del_init(&elem->list_node);

		/* Call the client's free function with the key and value pointer. */
		mru->free_func(elem->key, elem->value);

		/* Free the element structure. */
		kmem_zone_free(xfs_mru_elem_zone, elem);
	}

	spin_lock(&mru->lock);
}

/*
 * We fire the reap timer every group expiry interval so
 * we always have a reaper ready to run. This makes shutdown
 * and flushing of the reaper easy to do. Hence we need to
 * keep when the next reap must occur so we can determine
 * at each interval whether there is anything we need to do.
 */
STATIC void
_xfs_mru_cache_reap(
	struct work_struct	*work)
{
	xfs_mru_cache_t		*mru = container_of(work, xfs_mru_cache_t, work.work);
	unsigned long		now, next;

	ASSERT(mru && mru->lists);
	if (!mru || !mru->lists)
		return;

	spin_lock(&mru->lock);
	next = _xfs_mru_cache_migrate(mru, jiffies);
	_xfs_mru_cache_clear_reap_list(mru);

	mru->queued = next;
	if ((mru->queued > 0)) {
		now = jiffies;
		if (next <= now)
			next = 0;
		else
			next -= now;
		queue_delayed_work(xfs_mru_reap_wq, &mru->work, next);
	}

	spin_unlock(&mru->lock);
}

int
xfs_mru_cache_init(void)
{
	xfs_mru_elem_zone = kmem_zone_init(sizeof(xfs_mru_cache_elem_t),
	                                 "xfs_mru_cache_elem");
	if (!xfs_mru_elem_zone)
		goto out;

	xfs_mru_reap_wq = alloc_workqueue("xfs_mru_cache", WQ_MEM_RECLAIM, 1);
	if (!xfs_mru_reap_wq)
		goto out_destroy_mru_elem_zone;

	return 0;

 out_destroy_mru_elem_zone:
	kmem_zone_destroy(xfs_mru_elem_zone);
 out:
	return -ENOMEM;
}

void
xfs_mru_cache_uninit(void)
{
	destroy_workqueue(xfs_mru_reap_wq);
	kmem_zone_destroy(xfs_mru_elem_zone);
}

/*
 * To initialise a struct xfs_mru_cache pointer, call xfs_mru_cache_create()
 * with the address of the pointer, a lifetime value in milliseconds, a group
 * count and a free function to use when deleting elements.  This function
 * returns 0 if the initialisation was successful.
 */
int
xfs_mru_cache_create(
	xfs_mru_cache_t		**mrup,
	unsigned int		lifetime_ms,
	unsigned int		grp_count,
	xfs_mru_cache_free_func_t free_func)
{
	xfs_mru_cache_t	*mru = NULL;
	int		err = 0, grp;
	unsigned int	grp_time;

	if (mrup)
		*mrup = NULL;

	if (!mrup || !grp_count || !lifetime_ms || !free_func)
		return EINVAL;

	if (!(grp_time = msecs_to_jiffies(lifetime_ms) / grp_count))
		return EINVAL;

	if (!(mru = kmem_zalloc(sizeof(*mru), KM_SLEEP)))
		return ENOMEM;

	/* An extra list is needed to avoid reaping up to a grp_time early. */
	mru->grp_count = grp_count + 1;
	mru->lists = kmem_zalloc(mru->grp_count * sizeof(*mru->lists), KM_SLEEP);

	if (!mru->lists) {
		err = ENOMEM;
		goto exit;
	}

	for (grp = 0; grp < mru->grp_count; grp++)
		INIT_LIST_HEAD(mru->lists + grp);

	/*
	 * We use GFP_KERNEL radix tree preload and do inserts under a
	 * spinlock so GFP_ATOMIC is appropriate for the radix tree itself.
	 */
	INIT_RADIX_TREE(&mru->store, GFP_ATOMIC);
	INIT_LIST_HEAD(&mru->reap_list);
	spin_lock_init(&mru->lock);
	INIT_DELAYED_WORK(&mru->work, _xfs_mru_cache_reap);

	mru->grp_time  = grp_time;
	mru->free_func = free_func;

	*mrup = mru;

exit:
	if (err && mru && mru->lists)
		kmem_free(mru->lists);
	if (err && mru)
		kmem_free(mru);

	return err;
}

/*
 * Call xfs_mru_cache_flush() to flush out all cached entries, calling their
 * free functions as they're deleted.  When this function returns, the caller is
 * guaranteed that all the free functions for all the elements have finished
 * executing and the reaper is not running.
 */
static void
xfs_mru_cache_flush(
	xfs_mru_cache_t		*mru)
{
	if (!mru || !mru->lists)
		return;

	spin_lock(&mru->lock);
	if (mru->queued) {
		spin_unlock(&mru->lock);
		cancel_delayed_work_sync(&mru->work);
		spin_lock(&mru->lock);
	}

	_xfs_mru_cache_migrate(mru, jiffies + mru->grp_count * mru->grp_time);
	_xfs_mru_cache_clear_reap_list(mru);

	spin_unlock(&mru->lock);
}

void
xfs_mru_cache_destroy(
	xfs_mru_cache_t		*mru)
{
	if (!mru || !mru->lists)
		return;

	xfs_mru_cache_flush(mru);

	kmem_free(mru->lists);
	kmem_free(mru);
}

/*
 * To insert an element, call xfs_mru_cache_insert() with the data store, the
 * element's key and the client data pointer.  This function returns 0 on
 * success or ENOMEM if memory for the data element couldn't be allocated.
 */
int
xfs_mru_cache_insert(
	xfs_mru_cache_t	*mru,
	unsigned long	key,
	void		*value)
{
	xfs_mru_cache_elem_t *elem;
	int error;

	ASSERT(mru && mru->lists);
	if (!mru || !mru->lists)
		return EINVAL;

	elem = kmem_zone_zalloc(xfs_mru_elem_zone, KM_SLEEP);
	if (!elem)
		return ENOMEM;

	if (radix_tree_preload(GFP_KERNEL)) {
		error = ENOMEM;
		goto out_free_item;
	}

	INIT_LIST_HEAD(&elem->list_node);
	elem->key = key;
	elem->value = value;

	spin_lock(&mru->lock);

	error = -radix_tree_insert(&mru->store, key, elem);
	radix_tree_preload_end();
	if (error) {
		spin_unlock(&mru->lock);
		goto out_free_item;
	}
	_xfs_mru_cache_list_insert(mru, elem);

	spin_unlock(&mru->lock);

	return 0;
out_free_item:
	kmem_zone_free(xfs_mru_elem_zone, elem);
	return error;
}

/*
 * To remove an element without calling the free function, call
 * xfs_mru_cache_remove() with the data store and the element's key.  On success
 * the client data pointer for the removed element is returned, otherwise this
 * function will return a NULL pointer.
 */
void *
xfs_mru_cache_remove(
	xfs_mru_cache_t	*mru,
	unsigned long	key)
{
	xfs_mru_cache_elem_t *elem;
	void		*value = NULL;

	ASSERT(mru && mru->lists);
	if (!mru || !mru->lists)
		return NULL;

	spin_lock(&mru->lock);
	elem = radix_tree_delete(&mru->store, key);
	if (elem) {
		value = elem->value;
		list_del(&elem->list_node);
	}

	spin_unlock(&mru->lock);

	if (elem)
		kmem_zone_free(xfs_mru_elem_zone, elem);

	return value;
}

/*
 * To remove and element and call the free function, call xfs_mru_cache_delete()
 * with the data store and the element's key.
 */
void
xfs_mru_cache_delete(
	xfs_mru_cache_t	*mru,
	unsigned long	key)
{
	void		*value = xfs_mru_cache_remove(mru, key);

	if (value)
		mru->free_func(key, value);
}

/*
 * To look up an element using its key, call xfs_mru_cache_lookup() with the
 * data store and the element's key.  If found, the element will be moved to the
 * head of the MRU list to indicate that it's been touched.
 *
 * The internal data structures are protected by a spinlock that is STILL HELD
 * when this function returns.  Call xfs_mru_cache_done() to release it.  Note
 * that it is not safe to call any function that might sleep in the interim.
 *
 * The implementation could have used reference counting to avoid this
 * restriction, but since most clients simply want to get, set or test a member
 * of the returned data structure, the extra per-element memory isn't warranted.
 *
 * If the element isn't found, this function returns NULL and the spinlock is
 * released.  xfs_mru_cache_done() should NOT be called when this occurs.
 *
 * Because sparse isn't smart enough to know about conditional lock return
 * status, we need to help it get it right by annotating the path that does
 * not release the lock.
 */
void *
xfs_mru_cache_lookup(
	xfs_mru_cache_t	*mru,
	unsigned long	key)
{
	xfs_mru_cache_elem_t *elem;

	ASSERT(mru && mru->lists);
	if (!mru || !mru->lists)
		return NULL;

	spin_lock(&mru->lock);
	elem = radix_tree_lookup(&mru->store, key);
	if (elem) {
		list_del(&elem->list_node);
		_xfs_mru_cache_list_insert(mru, elem);
		__release(mru_lock); /* help sparse not be stupid */
	} else
		spin_unlock(&mru->lock);

	return elem ? elem->value : NULL;
}

/*
 * To release the internal data structure spinlock after having performed an
 * xfs_mru_cache_lookup() or an xfs_mru_cache_peek(), call xfs_mru_cache_done()
 * with the data store pointer.
 */
void
xfs_mru_cache_done(
	xfs_mru_cache_t	*mru) __releases(mru->lock)
{
	spin_unlock(&mru->lock);
}