xref: /linux/lib/lru_cache.c (revision 2363088eba2ecccfb643725e4864af73c4226a04)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3    lru_cache.c
4 
5    This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
6 
7    Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
8    Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
9    Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
10 
11 
12  */
13 
14 #include <linux/module.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/string.h> /* for memset */
18 #include <linux/seq_file.h> /* for seq_printf */
19 #include <linux/lru_cache.h>
20 
21 MODULE_AUTHOR("Philipp Reisner <phil@linbit.com>, "
22 	      "Lars Ellenberg <lars@linbit.com>");
23 MODULE_DESCRIPTION("lru_cache - Track sets of hot objects");
24 MODULE_LICENSE("GPL");
25 
26 /* this is developers aid only.
27  * it catches concurrent access (lack of locking on the users part) */
28 #define PARANOIA_ENTRY() do {		\
29 	BUG_ON(!lc);			\
30 	BUG_ON(!lc->nr_elements);	\
31 	BUG_ON(test_and_set_bit(__LC_PARANOIA, &lc->flags)); \
32 } while (0)
33 
34 #define RETURN(x...)     do { \
35 	clear_bit_unlock(__LC_PARANOIA, &lc->flags); \
36 	return x ; } while (0)
37 
38 /* BUG() if e is not one of the elements tracked by lc */
39 #define PARANOIA_LC_ELEMENT(lc, e) do {	\
40 	struct lru_cache *lc_ = (lc);	\
41 	struct lc_element *e_ = (e);	\
42 	unsigned i = e_->lc_index;	\
43 	BUG_ON(i >= lc_->nr_elements);	\
44 	BUG_ON(lc_->lc_element[i] != e_); } while (0)
45 
46 
47 /* We need to atomically
48  *  - try to grab the lock (set LC_LOCKED)
49  *  - only if there is no pending transaction
50  *    (neither LC_DIRTY nor LC_STARVING is set)
51  * Because of PARANOIA_ENTRY() above abusing lc->flags as well,
52  * it is not sufficient to just say
53  *	return 0 == cmpxchg(&lc->flags, 0, LC_LOCKED);
54  */
55 int lc_try_lock(struct lru_cache *lc)
56 {
57 	unsigned long val;
58 	do {
59 		val = cmpxchg(&lc->flags, 0, LC_LOCKED);
60 	} while (unlikely (val == LC_PARANOIA));
61 	/* Spin until no-one is inside a PARANOIA_ENTRY()/RETURN() section. */
62 	return 0 == val;
63 }
64 
65 /**
66  * lc_create - prepares to track objects in an active set
67  * @name: descriptive name only used in lc_seq_printf_stats and lc_seq_dump_details
68  * @cache: cache root pointer
69  * @max_pending_changes: maximum changes to accumulate until a transaction is required
70  * @e_count: number of elements allowed to be active simultaneously
71  * @e_size: size of the tracked objects
72  * @e_off: offset to the &struct lc_element member in a tracked object
73  *
74  * Returns a pointer to a newly initialized struct lru_cache on success,
75  * or NULL on (allocation) failure.
76  */
77 struct lru_cache *lc_create(const char *name, struct kmem_cache *cache,
78 		unsigned max_pending_changes,
79 		unsigned e_count, size_t e_size, size_t e_off)
80 {
81 	struct hlist_head *slot = NULL;
82 	struct lc_element **element = NULL;
83 	struct lru_cache *lc;
84 	struct lc_element *e;
85 	unsigned cache_obj_size = kmem_cache_size(cache);
86 	unsigned i;
87 
88 	WARN_ON(cache_obj_size < e_size);
89 	if (cache_obj_size < e_size)
90 		return NULL;
91 
92 	/* e_count too big; would probably fail the allocation below anyways.
93 	 * for typical use cases, e_count should be few thousand at most. */
94 	if (e_count > LC_MAX_ACTIVE)
95 		return NULL;
96 
97 	slot = kcalloc(e_count, sizeof(struct hlist_head), GFP_KERNEL);
98 	if (!slot)
99 		goto out_fail;
100 	element = kcalloc(e_count, sizeof(struct lc_element *), GFP_KERNEL);
101 	if (!element)
102 		goto out_fail;
103 
104 	lc = kzalloc(sizeof(*lc), GFP_KERNEL);
105 	if (!lc)
106 		goto out_fail;
107 
108 	INIT_LIST_HEAD(&lc->in_use);
109 	INIT_LIST_HEAD(&lc->lru);
110 	INIT_LIST_HEAD(&lc->free);
111 	INIT_LIST_HEAD(&lc->to_be_changed);
112 
113 	lc->name = name;
114 	lc->element_size = e_size;
115 	lc->element_off = e_off;
116 	lc->nr_elements = e_count;
117 	lc->max_pending_changes = max_pending_changes;
118 	lc->lc_cache = cache;
119 	lc->lc_element = element;
120 	lc->lc_slot = slot;
121 
122 	/* preallocate all objects */
123 	for (i = 0; i < e_count; i++) {
124 		void *p = kmem_cache_alloc(cache, GFP_KERNEL);
125 		if (!p)
126 			break;
127 		memset(p, 0, lc->element_size);
128 		e = p + e_off;
129 		e->lc_index = i;
130 		e->lc_number = LC_FREE;
131 		e->lc_new_number = LC_FREE;
132 		list_add(&e->list, &lc->free);
133 		element[i] = e;
134 	}
135 	if (i == e_count)
136 		return lc;
137 
138 	/* else: could not allocate all elements, give up */
139 	while (i) {
140 		void *p = element[--i];
141 		kmem_cache_free(cache, p - e_off);
142 	}
143 	kfree(lc);
144 out_fail:
145 	kfree(element);
146 	kfree(slot);
147 	return NULL;
148 }
149 
150 static void lc_free_by_index(struct lru_cache *lc, unsigned i)
151 {
152 	void *p = lc->lc_element[i];
153 	WARN_ON(!p);
154 	if (p) {
155 		p -= lc->element_off;
156 		kmem_cache_free(lc->lc_cache, p);
157 	}
158 }
159 
160 /**
161  * lc_destroy - frees memory allocated by lc_create()
162  * @lc: the lru cache to destroy
163  */
164 void lc_destroy(struct lru_cache *lc)
165 {
166 	unsigned i;
167 	if (!lc)
168 		return;
169 	for (i = 0; i < lc->nr_elements; i++)
170 		lc_free_by_index(lc, i);
171 	kfree(lc->lc_element);
172 	kfree(lc->lc_slot);
173 	kfree(lc);
174 }
175 
176 /**
177  * lc_reset - does a full reset for @lc and the hash table slots.
178  * @lc: the lru cache to operate on
179  *
180  * It is roughly the equivalent of re-allocating a fresh lru_cache object,
181  * basically a short cut to lc_destroy(lc); lc = lc_create(...);
182  */
183 void lc_reset(struct lru_cache *lc)
184 {
185 	unsigned i;
186 
187 	INIT_LIST_HEAD(&lc->in_use);
188 	INIT_LIST_HEAD(&lc->lru);
189 	INIT_LIST_HEAD(&lc->free);
190 	INIT_LIST_HEAD(&lc->to_be_changed);
191 	lc->used = 0;
192 	lc->hits = 0;
193 	lc->misses = 0;
194 	lc->starving = 0;
195 	lc->locked = 0;
196 	lc->changed = 0;
197 	lc->pending_changes = 0;
198 	lc->flags = 0;
199 	memset(lc->lc_slot, 0, sizeof(struct hlist_head) * lc->nr_elements);
200 
201 	for (i = 0; i < lc->nr_elements; i++) {
202 		struct lc_element *e = lc->lc_element[i];
203 		void *p = e;
204 		p -= lc->element_off;
205 		memset(p, 0, lc->element_size);
206 		/* re-init it */
207 		e->lc_index = i;
208 		e->lc_number = LC_FREE;
209 		e->lc_new_number = LC_FREE;
210 		list_add(&e->list, &lc->free);
211 	}
212 }
213 
214 /**
215  * lc_seq_printf_stats - print stats about @lc into @seq
216  * @seq: the seq_file to print into
217  * @lc: the lru cache to print statistics of
218  */
219 void lc_seq_printf_stats(struct seq_file *seq, struct lru_cache *lc)
220 {
221 	/* NOTE:
222 	 * total calls to lc_get are
223 	 * (starving + hits + misses)
224 	 * misses include "locked" count (update from an other thread in
225 	 * progress) and "changed", when this in fact lead to an successful
226 	 * update of the cache.
227 	 */
228 	seq_printf(seq, "\t%s: used:%u/%u hits:%lu misses:%lu starving:%lu locked:%lu changed:%lu\n",
229 		   lc->name, lc->used, lc->nr_elements,
230 		   lc->hits, lc->misses, lc->starving, lc->locked, lc->changed);
231 }
232 
233 static struct hlist_head *lc_hash_slot(struct lru_cache *lc, unsigned int enr)
234 {
235 	return  lc->lc_slot + (enr % lc->nr_elements);
236 }
237 
238 
239 static struct lc_element *__lc_find(struct lru_cache *lc, unsigned int enr,
240 		bool include_changing)
241 {
242 	struct lc_element *e;
243 
244 	BUG_ON(!lc);
245 	BUG_ON(!lc->nr_elements);
246 	hlist_for_each_entry(e, lc_hash_slot(lc, enr), colision) {
247 		/* "about to be changed" elements, pending transaction commit,
248 		 * are hashed by their "new number". "Normal" elements have
249 		 * lc_number == lc_new_number. */
250 		if (e->lc_new_number != enr)
251 			continue;
252 		if (e->lc_new_number == e->lc_number || include_changing)
253 			return e;
254 		break;
255 	}
256 	return NULL;
257 }
258 
259 /**
260  * lc_find - find element by label, if present in the hash table
261  * @lc: The lru_cache object
262  * @enr: element number
263  *
264  * Returns the pointer to an element, if the element with the requested
265  * "label" or element number is present in the hash table,
266  * or NULL if not found. Does not change the refcnt.
267  * Ignores elements that are "about to be used", i.e. not yet in the active
268  * set, but still pending transaction commit.
269  */
270 struct lc_element *lc_find(struct lru_cache *lc, unsigned int enr)
271 {
272 	return __lc_find(lc, enr, 0);
273 }
274 
275 /**
276  * lc_is_used - find element by label
277  * @lc: The lru_cache object
278  * @enr: element number
279  *
280  * Returns true, if the element with the requested "label" or element number is
281  * present in the hash table, and is used (refcnt > 0).
282  * Also finds elements that are not _currently_ used but only "about to be
283  * used", i.e. on the "to_be_changed" list, pending transaction commit.
284  */
285 bool lc_is_used(struct lru_cache *lc, unsigned int enr)
286 {
287 	struct lc_element *e = __lc_find(lc, enr, 1);
288 	return e && e->refcnt;
289 }
290 
291 /**
292  * lc_del - removes an element from the cache
293  * @lc: The lru_cache object
294  * @e: The element to remove
295  *
296  * @e must be unused (refcnt == 0). Moves @e from "lru" to "free" list,
297  * sets @e->enr to %LC_FREE.
298  */
299 void lc_del(struct lru_cache *lc, struct lc_element *e)
300 {
301 	PARANOIA_ENTRY();
302 	PARANOIA_LC_ELEMENT(lc, e);
303 	BUG_ON(e->refcnt);
304 
305 	e->lc_number = e->lc_new_number = LC_FREE;
306 	hlist_del_init(&e->colision);
307 	list_move(&e->list, &lc->free);
308 	RETURN();
309 }
310 
311 static struct lc_element *lc_prepare_for_change(struct lru_cache *lc, unsigned new_number)
312 {
313 	struct list_head *n;
314 	struct lc_element *e;
315 
316 	if (!list_empty(&lc->free))
317 		n = lc->free.next;
318 	else if (!list_empty(&lc->lru))
319 		n = lc->lru.prev;
320 	else
321 		return NULL;
322 
323 	e = list_entry(n, struct lc_element, list);
324 	PARANOIA_LC_ELEMENT(lc, e);
325 
326 	e->lc_new_number = new_number;
327 	if (!hlist_unhashed(&e->colision))
328 		__hlist_del(&e->colision);
329 	hlist_add_head(&e->colision, lc_hash_slot(lc, new_number));
330 	list_move(&e->list, &lc->to_be_changed);
331 
332 	return e;
333 }
334 
335 static int lc_unused_element_available(struct lru_cache *lc)
336 {
337 	if (!list_empty(&lc->free))
338 		return 1; /* something on the free list */
339 	if (!list_empty(&lc->lru))
340 		return 1;  /* something to evict */
341 
342 	return 0;
343 }
344 
345 /* used as internal flags to __lc_get */
346 enum {
347 	LC_GET_MAY_CHANGE = 1,
348 	LC_GET_MAY_USE_UNCOMMITTED = 2,
349 };
350 
351 static struct lc_element *__lc_get(struct lru_cache *lc, unsigned int enr, unsigned int flags)
352 {
353 	struct lc_element *e;
354 
355 	PARANOIA_ENTRY();
356 	if (test_bit(__LC_STARVING, &lc->flags)) {
357 		++lc->starving;
358 		RETURN(NULL);
359 	}
360 
361 	e = __lc_find(lc, enr, 1);
362 	/* if lc_new_number != lc_number,
363 	 * this enr is currently being pulled in already,
364 	 * and will be available once the pending transaction
365 	 * has been committed. */
366 	if (e) {
367 		if (e->lc_new_number != e->lc_number) {
368 			/* It has been found above, but on the "to_be_changed"
369 			 * list, not yet committed.  Don't pull it in twice,
370 			 * wait for the transaction, then try again...
371 			 */
372 			if (!(flags & LC_GET_MAY_USE_UNCOMMITTED))
373 				RETURN(NULL);
374 			/* ... unless the caller is aware of the implications,
375 			 * probably preparing a cumulative transaction. */
376 			++e->refcnt;
377 			++lc->hits;
378 			RETURN(e);
379 		}
380 		/* else: lc_new_number == lc_number; a real hit. */
381 		++lc->hits;
382 		if (e->refcnt++ == 0)
383 			lc->used++;
384 		list_move(&e->list, &lc->in_use); /* Not evictable... */
385 		RETURN(e);
386 	}
387 	/* e == NULL */
388 
389 	++lc->misses;
390 	if (!(flags & LC_GET_MAY_CHANGE))
391 		RETURN(NULL);
392 
393 	/* To avoid races with lc_try_lock(), first, mark us dirty
394 	 * (using test_and_set_bit, as it implies memory barriers), ... */
395 	test_and_set_bit(__LC_DIRTY, &lc->flags);
396 
397 	/* ... only then check if it is locked anyways. If lc_unlock clears
398 	 * the dirty bit again, that's not a problem, we will come here again.
399 	 */
400 	if (test_bit(__LC_LOCKED, &lc->flags)) {
401 		++lc->locked;
402 		RETURN(NULL);
403 	}
404 
405 	/* In case there is nothing available and we can not kick out
406 	 * the LRU element, we have to wait ...
407 	 */
408 	if (!lc_unused_element_available(lc)) {
409 		set_bit(__LC_STARVING, &lc->flags);
410 		RETURN(NULL);
411 	}
412 
413 	/* It was not present in the active set.  We are going to recycle an
414 	 * unused (or even "free") element, but we won't accumulate more than
415 	 * max_pending_changes changes.  */
416 	if (lc->pending_changes >= lc->max_pending_changes)
417 		RETURN(NULL);
418 
419 	e = lc_prepare_for_change(lc, enr);
420 	BUG_ON(!e);
421 
422 	clear_bit(__LC_STARVING, &lc->flags);
423 	BUG_ON(++e->refcnt != 1);
424 	lc->used++;
425 	lc->pending_changes++;
426 
427 	RETURN(e);
428 }
429 
430 /**
431  * lc_get - get element by label, maybe change the active set
432  * @lc: the lru cache to operate on
433  * @enr: the label to look up
434  *
435  * Finds an element in the cache, increases its usage count,
436  * "touches" and returns it.
437  *
438  * In case the requested number is not present, it needs to be added to the
439  * cache. Therefore it is possible that an other element becomes evicted from
440  * the cache. In either case, the user is notified so he is able to e.g. keep
441  * a persistent log of the cache changes, and therefore the objects in use.
442  *
443  * Return values:
444  *  NULL
445  *     The cache was marked %LC_STARVING,
446  *     or the requested label was not in the active set
447  *     and a changing transaction is still pending (@lc was marked %LC_DIRTY).
448  *     Or no unused or free element could be recycled (@lc will be marked as
449  *     %LC_STARVING, blocking further lc_get() operations).
450  *
451  *  pointer to the element with the REQUESTED element number.
452  *     In this case, it can be used right away
453  *
454  *  pointer to an UNUSED element with some different element number,
455  *          where that different number may also be %LC_FREE.
456  *
457  *          In this case, the cache is marked %LC_DIRTY,
458  *          so lc_try_lock() will no longer succeed.
459  *          The returned element pointer is moved to the "to_be_changed" list,
460  *          and registered with the new element number on the hash collision chains,
461  *          so it is possible to pick it up from lc_is_used().
462  *          Up to "max_pending_changes" (see lc_create()) can be accumulated.
463  *          The user now should do whatever housekeeping is necessary,
464  *          typically serialize on lc_try_lock_for_transaction(), then call
465  *          lc_committed(lc) and lc_unlock(), to finish the change.
466  *
467  * NOTE: The user needs to check the lc_number on EACH use, so he recognizes
468  *       any cache set change.
469  */
470 struct lc_element *lc_get(struct lru_cache *lc, unsigned int enr)
471 {
472 	return __lc_get(lc, enr, LC_GET_MAY_CHANGE);
473 }
474 
475 /**
476  * lc_get_cumulative - like lc_get; also finds to-be-changed elements
477  * @lc: the lru cache to operate on
478  * @enr: the label to look up
479  *
480  * Unlike lc_get this also returns the element for @enr, if it is belonging to
481  * a pending transaction, so the return values are like for lc_get(),
482  * plus:
483  *
484  * pointer to an element already on the "to_be_changed" list.
485  * 	In this case, the cache was already marked %LC_DIRTY.
486  *
487  * Caller needs to make sure that the pending transaction is completed,
488  * before proceeding to actually use this element.
489  */
490 struct lc_element *lc_get_cumulative(struct lru_cache *lc, unsigned int enr)
491 {
492 	return __lc_get(lc, enr, LC_GET_MAY_CHANGE|LC_GET_MAY_USE_UNCOMMITTED);
493 }
494 
495 /**
496  * lc_try_get - get element by label, if present; do not change the active set
497  * @lc: the lru cache to operate on
498  * @enr: the label to look up
499  *
500  * Finds an element in the cache, increases its usage count,
501  * "touches" and returns it.
502  *
503  * Return values:
504  *  NULL
505  *     The cache was marked %LC_STARVING,
506  *     or the requested label was not in the active set
507  *
508  *  pointer to the element with the REQUESTED element number.
509  *     In this case, it can be used right away
510  */
511 struct lc_element *lc_try_get(struct lru_cache *lc, unsigned int enr)
512 {
513 	return __lc_get(lc, enr, 0);
514 }
515 
516 /**
517  * lc_committed - tell @lc that pending changes have been recorded
518  * @lc: the lru cache to operate on
519  *
520  * User is expected to serialize on explicit lc_try_lock_for_transaction()
521  * before the transaction is started, and later needs to lc_unlock() explicitly
522  * as well.
523  */
524 void lc_committed(struct lru_cache *lc)
525 {
526 	struct lc_element *e, *tmp;
527 
528 	PARANOIA_ENTRY();
529 	list_for_each_entry_safe(e, tmp, &lc->to_be_changed, list) {
530 		/* count number of changes, not number of transactions */
531 		++lc->changed;
532 		e->lc_number = e->lc_new_number;
533 		list_move(&e->list, &lc->in_use);
534 	}
535 	lc->pending_changes = 0;
536 	RETURN();
537 }
538 
539 
540 /**
541  * lc_put - give up refcnt of @e
542  * @lc: the lru cache to operate on
543  * @e: the element to put
544  *
545  * If refcnt reaches zero, the element is moved to the lru list,
546  * and a %LC_STARVING (if set) is cleared.
547  * Returns the new (post-decrement) refcnt.
548  */
549 unsigned int lc_put(struct lru_cache *lc, struct lc_element *e)
550 {
551 	PARANOIA_ENTRY();
552 	PARANOIA_LC_ELEMENT(lc, e);
553 	BUG_ON(e->refcnt == 0);
554 	BUG_ON(e->lc_number != e->lc_new_number);
555 	if (--e->refcnt == 0) {
556 		/* move it to the front of LRU. */
557 		list_move(&e->list, &lc->lru);
558 		lc->used--;
559 		clear_bit_unlock(__LC_STARVING, &lc->flags);
560 	}
561 	RETURN(e->refcnt);
562 }
563 
564 /**
565  * lc_element_by_index
566  * @lc: the lru cache to operate on
567  * @i: the index of the element to return
568  */
569 struct lc_element *lc_element_by_index(struct lru_cache *lc, unsigned i)
570 {
571 	BUG_ON(i >= lc->nr_elements);
572 	BUG_ON(lc->lc_element[i] == NULL);
573 	BUG_ON(lc->lc_element[i]->lc_index != i);
574 	return lc->lc_element[i];
575 }
576 
577 /**
578  * lc_seq_dump_details - Dump a complete LRU cache to seq in textual form.
579  * @lc: the lru cache to operate on
580  * @seq: the &struct seq_file pointer to seq_printf into
581  * @utext: user supplied additional "heading" or other info
582  * @detail: function pointer the user may provide to dump further details
583  * of the object the lc_element is embedded in. May be NULL.
584  * Note: a leading space ' ' and trailing newline '\n' is implied.
585  */
586 void lc_seq_dump_details(struct seq_file *seq, struct lru_cache *lc, char *utext,
587 	     void (*detail) (struct seq_file *, struct lc_element *))
588 {
589 	unsigned int nr_elements = lc->nr_elements;
590 	struct lc_element *e;
591 	int i;
592 
593 	seq_printf(seq, "\tnn: lc_number (new nr) refcnt %s\n ", utext);
594 	for (i = 0; i < nr_elements; i++) {
595 		e = lc_element_by_index(lc, i);
596 		if (e->lc_number != e->lc_new_number)
597 			seq_printf(seq, "\t%5d: %6d %8d %6d ",
598 				i, e->lc_number, e->lc_new_number, e->refcnt);
599 		else
600 			seq_printf(seq, "\t%5d: %6d %-8s %6d ",
601 				i, e->lc_number, "-\"-", e->refcnt);
602 		if (detail)
603 			detail(seq, e);
604 		seq_putc(seq, '\n');
605 	}
606 }
607 
608 EXPORT_SYMBOL(lc_create);
609 EXPORT_SYMBOL(lc_reset);
610 EXPORT_SYMBOL(lc_destroy);
611 EXPORT_SYMBOL(lc_del);
612 EXPORT_SYMBOL(lc_try_get);
613 EXPORT_SYMBOL(lc_find);
614 EXPORT_SYMBOL(lc_get);
615 EXPORT_SYMBOL(lc_put);
616 EXPORT_SYMBOL(lc_committed);
617 EXPORT_SYMBOL(lc_element_by_index);
618 EXPORT_SYMBOL(lc_seq_printf_stats);
619 EXPORT_SYMBOL(lc_seq_dump_details);
620 EXPORT_SYMBOL(lc_try_lock);
621 EXPORT_SYMBOL(lc_is_used);
622 EXPORT_SYMBOL(lc_get_cumulative);
623