Loading...
1/*
2 lru_cache.c
3
4 This file is part of DRBD by Philipp Reisner and Lars Ellenberg.
5
6 Copyright (C) 2003-2008, LINBIT Information Technologies GmbH.
7 Copyright (C) 2003-2008, Philipp Reisner <philipp.reisner@linbit.com>.
8 Copyright (C) 2003-2008, Lars Ellenberg <lars.ellenberg@linbit.com>.
9
10 drbd is free software; you can redistribute it and/or modify
11 it under the terms of the GNU General Public License as published by
12 the Free Software Foundation; either version 2, or (at your option)
13 any later version.
14
15 drbd is distributed in the hope that it will be useful,
16 but WITHOUT ANY WARRANTY; without even the implied warranty of
17 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
18 GNU General Public License for more details.
19
20 You should have received a copy of the GNU General Public License
21 along with drbd; see the file COPYING. If not, write to
22 the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA.
23
24 */
25
26#include <linux/module.h>
27#include <linux/bitops.h>
28#include <linux/slab.h>
29#include <linux/string.h> /* for memset */
30#include <linux/seq_file.h> /* for seq_printf */
31#include <linux/lru_cache.h>
32
33MODULE_AUTHOR("Philipp Reisner <phil@linbit.com>, "
34 "Lars Ellenberg <lars@linbit.com>");
35MODULE_DESCRIPTION("lru_cache - Track sets of hot objects");
36MODULE_LICENSE("GPL");
37
38/* this is developers aid only.
39 * it catches concurrent access (lack of locking on the users part) */
40#define PARANOIA_ENTRY() do { \
41 BUG_ON(!lc); \
42 BUG_ON(!lc->nr_elements); \
43 BUG_ON(test_and_set_bit(__LC_PARANOIA, &lc->flags)); \
44} while (0)
45
46#define RETURN(x...) do { \
47 clear_bit_unlock(__LC_PARANOIA, &lc->flags); \
48 return x ; } while (0)
49
50/* BUG() if e is not one of the elements tracked by lc */
51#define PARANOIA_LC_ELEMENT(lc, e) do { \
52 struct lru_cache *lc_ = (lc); \
53 struct lc_element *e_ = (e); \
54 unsigned i = e_->lc_index; \
55 BUG_ON(i >= lc_->nr_elements); \
56 BUG_ON(lc_->lc_element[i] != e_); } while (0)
57
58
59/* We need to atomically
60 * - try to grab the lock (set LC_LOCKED)
61 * - only if there is no pending transaction
62 * (neither LC_DIRTY nor LC_STARVING is set)
63 * Because of PARANOIA_ENTRY() above abusing lc->flags as well,
64 * it is not sufficient to just say
65 * return 0 == cmpxchg(&lc->flags, 0, LC_LOCKED);
66 */
67int lc_try_lock(struct lru_cache *lc)
68{
69 unsigned long val;
70 do {
71 val = cmpxchg(&lc->flags, 0, LC_LOCKED);
72 } while (unlikely (val == LC_PARANOIA));
73 /* Spin until no-one is inside a PARANOIA_ENTRY()/RETURN() section. */
74 return 0 == val;
75#if 0
76 /* Alternative approach, spin in case someone enters or leaves a
77 * PARANOIA_ENTRY()/RETURN() section. */
78 unsigned long old, new, val;
79 do {
80 old = lc->flags & LC_PARANOIA;
81 new = old | LC_LOCKED;
82 val = cmpxchg(&lc->flags, old, new);
83 } while (unlikely (val == (old ^ LC_PARANOIA)));
84 return old == val;
85#endif
86}
87
88/**
89 * lc_create - prepares to track objects in an active set
90 * @name: descriptive name only used in lc_seq_printf_stats and lc_seq_dump_details
91 * @max_pending_changes: maximum changes to accumulate until a transaction is required
92 * @e_count: number of elements allowed to be active simultaneously
93 * @e_size: size of the tracked objects
94 * @e_off: offset to the &struct lc_element member in a tracked object
95 *
96 * Returns a pointer to a newly initialized struct lru_cache on success,
97 * or NULL on (allocation) failure.
98 */
99struct lru_cache *lc_create(const char *name, struct kmem_cache *cache,
100 unsigned max_pending_changes,
101 unsigned e_count, size_t e_size, size_t e_off)
102{
103 struct hlist_head *slot = NULL;
104 struct lc_element **element = NULL;
105 struct lru_cache *lc;
106 struct lc_element *e;
107 unsigned cache_obj_size = kmem_cache_size(cache);
108 unsigned i;
109
110 WARN_ON(cache_obj_size < e_size);
111 if (cache_obj_size < e_size)
112 return NULL;
113
114 /* e_count too big; would probably fail the allocation below anyways.
115 * for typical use cases, e_count should be few thousand at most. */
116 if (e_count > LC_MAX_ACTIVE)
117 return NULL;
118
119 slot = kcalloc(e_count, sizeof(struct hlist_head), GFP_KERNEL);
120 if (!slot)
121 goto out_fail;
122 element = kzalloc(e_count * sizeof(struct lc_element *), GFP_KERNEL);
123 if (!element)
124 goto out_fail;
125
126 lc = kzalloc(sizeof(*lc), GFP_KERNEL);
127 if (!lc)
128 goto out_fail;
129
130 INIT_LIST_HEAD(&lc->in_use);
131 INIT_LIST_HEAD(&lc->lru);
132 INIT_LIST_HEAD(&lc->free);
133 INIT_LIST_HEAD(&lc->to_be_changed);
134
135 lc->name = name;
136 lc->element_size = e_size;
137 lc->element_off = e_off;
138 lc->nr_elements = e_count;
139 lc->max_pending_changes = max_pending_changes;
140 lc->lc_cache = cache;
141 lc->lc_element = element;
142 lc->lc_slot = slot;
143
144 /* preallocate all objects */
145 for (i = 0; i < e_count; i++) {
146 void *p = kmem_cache_alloc(cache, GFP_KERNEL);
147 if (!p)
148 break;
149 memset(p, 0, lc->element_size);
150 e = p + e_off;
151 e->lc_index = i;
152 e->lc_number = LC_FREE;
153 e->lc_new_number = LC_FREE;
154 list_add(&e->list, &lc->free);
155 element[i] = e;
156 }
157 if (i == e_count)
158 return lc;
159
160 /* else: could not allocate all elements, give up */
161 for (i--; i; i--) {
162 void *p = element[i];
163 kmem_cache_free(cache, p - e_off);
164 }
165 kfree(lc);
166out_fail:
167 kfree(element);
168 kfree(slot);
169 return NULL;
170}
171
172static void lc_free_by_index(struct lru_cache *lc, unsigned i)
173{
174 void *p = lc->lc_element[i];
175 WARN_ON(!p);
176 if (p) {
177 p -= lc->element_off;
178 kmem_cache_free(lc->lc_cache, p);
179 }
180}
181
182/**
183 * lc_destroy - frees memory allocated by lc_create()
184 * @lc: the lru cache to destroy
185 */
186void lc_destroy(struct lru_cache *lc)
187{
188 unsigned i;
189 if (!lc)
190 return;
191 for (i = 0; i < lc->nr_elements; i++)
192 lc_free_by_index(lc, i);
193 kfree(lc->lc_element);
194 kfree(lc->lc_slot);
195 kfree(lc);
196}
197
198/**
199 * lc_reset - does a full reset for @lc and the hash table slots.
200 * @lc: the lru cache to operate on
201 *
202 * It is roughly the equivalent of re-allocating a fresh lru_cache object,
203 * basically a short cut to lc_destroy(lc); lc = lc_create(...);
204 */
205void lc_reset(struct lru_cache *lc)
206{
207 unsigned i;
208
209 INIT_LIST_HEAD(&lc->in_use);
210 INIT_LIST_HEAD(&lc->lru);
211 INIT_LIST_HEAD(&lc->free);
212 INIT_LIST_HEAD(&lc->to_be_changed);
213 lc->used = 0;
214 lc->hits = 0;
215 lc->misses = 0;
216 lc->starving = 0;
217 lc->locked = 0;
218 lc->changed = 0;
219 lc->pending_changes = 0;
220 lc->flags = 0;
221 memset(lc->lc_slot, 0, sizeof(struct hlist_head) * lc->nr_elements);
222
223 for (i = 0; i < lc->nr_elements; i++) {
224 struct lc_element *e = lc->lc_element[i];
225 void *p = e;
226 p -= lc->element_off;
227 memset(p, 0, lc->element_size);
228 /* re-init it */
229 e->lc_index = i;
230 e->lc_number = LC_FREE;
231 e->lc_new_number = LC_FREE;
232 list_add(&e->list, &lc->free);
233 }
234}
235
236/**
237 * lc_seq_printf_stats - print stats about @lc into @seq
238 * @seq: the seq_file to print into
239 * @lc: the lru cache to print statistics of
240 */
241void lc_seq_printf_stats(struct seq_file *seq, struct lru_cache *lc)
242{
243 /* NOTE:
244 * total calls to lc_get are
245 * (starving + hits + misses)
246 * misses include "locked" count (update from an other thread in
247 * progress) and "changed", when this in fact lead to an successful
248 * update of the cache.
249 */
250 seq_printf(seq, "\t%s: used:%u/%u hits:%lu misses:%lu starving:%lu locked:%lu changed:%lu\n",
251 lc->name, lc->used, lc->nr_elements,
252 lc->hits, lc->misses, lc->starving, lc->locked, lc->changed);
253}
254
255static struct hlist_head *lc_hash_slot(struct lru_cache *lc, unsigned int enr)
256{
257 return lc->lc_slot + (enr % lc->nr_elements);
258}
259
260
261static struct lc_element *__lc_find(struct lru_cache *lc, unsigned int enr,
262 bool include_changing)
263{
264 struct lc_element *e;
265
266 BUG_ON(!lc);
267 BUG_ON(!lc->nr_elements);
268 hlist_for_each_entry(e, lc_hash_slot(lc, enr), colision) {
269 /* "about to be changed" elements, pending transaction commit,
270 * are hashed by their "new number". "Normal" elements have
271 * lc_number == lc_new_number. */
272 if (e->lc_new_number != enr)
273 continue;
274 if (e->lc_new_number == e->lc_number || include_changing)
275 return e;
276 break;
277 }
278 return NULL;
279}
280
281/**
282 * lc_find - find element by label, if present in the hash table
283 * @lc: The lru_cache object
284 * @enr: element number
285 *
286 * Returns the pointer to an element, if the element with the requested
287 * "label" or element number is present in the hash table,
288 * or NULL if not found. Does not change the refcnt.
289 * Ignores elements that are "about to be used", i.e. not yet in the active
290 * set, but still pending transaction commit.
291 */
292struct lc_element *lc_find(struct lru_cache *lc, unsigned int enr)
293{
294 return __lc_find(lc, enr, 0);
295}
296
297/**
298 * lc_is_used - find element by label
299 * @lc: The lru_cache object
300 * @enr: element number
301 *
302 * Returns true, if the element with the requested "label" or element number is
303 * present in the hash table, and is used (refcnt > 0).
304 * Also finds elements that are not _currently_ used but only "about to be
305 * used", i.e. on the "to_be_changed" list, pending transaction commit.
306 */
307bool lc_is_used(struct lru_cache *lc, unsigned int enr)
308{
309 struct lc_element *e = __lc_find(lc, enr, 1);
310 return e && e->refcnt;
311}
312
313/**
314 * lc_del - removes an element from the cache
315 * @lc: The lru_cache object
316 * @e: The element to remove
317 *
318 * @e must be unused (refcnt == 0). Moves @e from "lru" to "free" list,
319 * sets @e->enr to %LC_FREE.
320 */
321void lc_del(struct lru_cache *lc, struct lc_element *e)
322{
323 PARANOIA_ENTRY();
324 PARANOIA_LC_ELEMENT(lc, e);
325 BUG_ON(e->refcnt);
326
327 e->lc_number = e->lc_new_number = LC_FREE;
328 hlist_del_init(&e->colision);
329 list_move(&e->list, &lc->free);
330 RETURN();
331}
332
333static struct lc_element *lc_prepare_for_change(struct lru_cache *lc, unsigned new_number)
334{
335 struct list_head *n;
336 struct lc_element *e;
337
338 if (!list_empty(&lc->free))
339 n = lc->free.next;
340 else if (!list_empty(&lc->lru))
341 n = lc->lru.prev;
342 else
343 return NULL;
344
345 e = list_entry(n, struct lc_element, list);
346 PARANOIA_LC_ELEMENT(lc, e);
347
348 e->lc_new_number = new_number;
349 if (!hlist_unhashed(&e->colision))
350 __hlist_del(&e->colision);
351 hlist_add_head(&e->colision, lc_hash_slot(lc, new_number));
352 list_move(&e->list, &lc->to_be_changed);
353
354 return e;
355}
356
357static int lc_unused_element_available(struct lru_cache *lc)
358{
359 if (!list_empty(&lc->free))
360 return 1; /* something on the free list */
361 if (!list_empty(&lc->lru))
362 return 1; /* something to evict */
363
364 return 0;
365}
366
367/* used as internal flags to __lc_get */
368enum {
369 LC_GET_MAY_CHANGE = 1,
370 LC_GET_MAY_USE_UNCOMMITTED = 2,
371};
372
373static struct lc_element *__lc_get(struct lru_cache *lc, unsigned int enr, unsigned int flags)
374{
375 struct lc_element *e;
376
377 PARANOIA_ENTRY();
378 if (lc->flags & LC_STARVING) {
379 ++lc->starving;
380 RETURN(NULL);
381 }
382
383 e = __lc_find(lc, enr, 1);
384 /* if lc_new_number != lc_number,
385 * this enr is currently being pulled in already,
386 * and will be available once the pending transaction
387 * has been committed. */
388 if (e) {
389 if (e->lc_new_number != e->lc_number) {
390 /* It has been found above, but on the "to_be_changed"
391 * list, not yet committed. Don't pull it in twice,
392 * wait for the transaction, then try again...
393 */
394 if (!(flags & LC_GET_MAY_USE_UNCOMMITTED))
395 RETURN(NULL);
396 /* ... unless the caller is aware of the implications,
397 * probably preparing a cumulative transaction. */
398 ++e->refcnt;
399 ++lc->hits;
400 RETURN(e);
401 }
402 /* else: lc_new_number == lc_number; a real hit. */
403 ++lc->hits;
404 if (e->refcnt++ == 0)
405 lc->used++;
406 list_move(&e->list, &lc->in_use); /* Not evictable... */
407 RETURN(e);
408 }
409 /* e == NULL */
410
411 ++lc->misses;
412 if (!(flags & LC_GET_MAY_CHANGE))
413 RETURN(NULL);
414
415 /* To avoid races with lc_try_lock(), first, mark us dirty
416 * (using test_and_set_bit, as it implies memory barriers), ... */
417 test_and_set_bit(__LC_DIRTY, &lc->flags);
418
419 /* ... only then check if it is locked anyways. If lc_unlock clears
420 * the dirty bit again, that's not a problem, we will come here again.
421 */
422 if (test_bit(__LC_LOCKED, &lc->flags)) {
423 ++lc->locked;
424 RETURN(NULL);
425 }
426
427 /* In case there is nothing available and we can not kick out
428 * the LRU element, we have to wait ...
429 */
430 if (!lc_unused_element_available(lc)) {
431 __set_bit(__LC_STARVING, &lc->flags);
432 RETURN(NULL);
433 }
434
435 /* It was not present in the active set. We are going to recycle an
436 * unused (or even "free") element, but we won't accumulate more than
437 * max_pending_changes changes. */
438 if (lc->pending_changes >= lc->max_pending_changes)
439 RETURN(NULL);
440
441 e = lc_prepare_for_change(lc, enr);
442 BUG_ON(!e);
443
444 clear_bit(__LC_STARVING, &lc->flags);
445 BUG_ON(++e->refcnt != 1);
446 lc->used++;
447 lc->pending_changes++;
448
449 RETURN(e);
450}
451
452/**
453 * lc_get - get element by label, maybe change the active set
454 * @lc: the lru cache to operate on
455 * @enr: the label to look up
456 *
457 * Finds an element in the cache, increases its usage count,
458 * "touches" and returns it.
459 *
460 * In case the requested number is not present, it needs to be added to the
461 * cache. Therefore it is possible that an other element becomes evicted from
462 * the cache. In either case, the user is notified so he is able to e.g. keep
463 * a persistent log of the cache changes, and therefore the objects in use.
464 *
465 * Return values:
466 * NULL
467 * The cache was marked %LC_STARVING,
468 * or the requested label was not in the active set
469 * and a changing transaction is still pending (@lc was marked %LC_DIRTY).
470 * Or no unused or free element could be recycled (@lc will be marked as
471 * %LC_STARVING, blocking further lc_get() operations).
472 *
473 * pointer to the element with the REQUESTED element number.
474 * In this case, it can be used right away
475 *
476 * pointer to an UNUSED element with some different element number,
477 * where that different number may also be %LC_FREE.
478 *
479 * In this case, the cache is marked %LC_DIRTY,
480 * so lc_try_lock() will no longer succeed.
481 * The returned element pointer is moved to the "to_be_changed" list,
482 * and registered with the new element number on the hash collision chains,
483 * so it is possible to pick it up from lc_is_used().
484 * Up to "max_pending_changes" (see lc_create()) can be accumulated.
485 * The user now should do whatever housekeeping is necessary,
486 * typically serialize on lc_try_lock_for_transaction(), then call
487 * lc_committed(lc) and lc_unlock(), to finish the change.
488 *
489 * NOTE: The user needs to check the lc_number on EACH use, so he recognizes
490 * any cache set change.
491 */
492struct lc_element *lc_get(struct lru_cache *lc, unsigned int enr)
493{
494 return __lc_get(lc, enr, LC_GET_MAY_CHANGE);
495}
496
497/**
498 * lc_get_cumulative - like lc_get; also finds to-be-changed elements
499 * @lc: the lru cache to operate on
500 * @enr: the label to look up
501 *
502 * Unlike lc_get this also returns the element for @enr, if it is belonging to
503 * a pending transaction, so the return values are like for lc_get(),
504 * plus:
505 *
506 * pointer to an element already on the "to_be_changed" list.
507 * In this case, the cache was already marked %LC_DIRTY.
508 *
509 * Caller needs to make sure that the pending transaction is completed,
510 * before proceeding to actually use this element.
511 */
512struct lc_element *lc_get_cumulative(struct lru_cache *lc, unsigned int enr)
513{
514 return __lc_get(lc, enr, LC_GET_MAY_CHANGE|LC_GET_MAY_USE_UNCOMMITTED);
515}
516
517/**
518 * lc_try_get - get element by label, if present; do not change the active set
519 * @lc: the lru cache to operate on
520 * @enr: the label to look up
521 *
522 * Finds an element in the cache, increases its usage count,
523 * "touches" and returns it.
524 *
525 * Return values:
526 * NULL
527 * The cache was marked %LC_STARVING,
528 * or the requested label was not in the active set
529 *
530 * pointer to the element with the REQUESTED element number.
531 * In this case, it can be used right away
532 */
533struct lc_element *lc_try_get(struct lru_cache *lc, unsigned int enr)
534{
535 return __lc_get(lc, enr, 0);
536}
537
538/**
539 * lc_committed - tell @lc that pending changes have been recorded
540 * @lc: the lru cache to operate on
541 *
542 * User is expected to serialize on explicit lc_try_lock_for_transaction()
543 * before the transaction is started, and later needs to lc_unlock() explicitly
544 * as well.
545 */
546void lc_committed(struct lru_cache *lc)
547{
548 struct lc_element *e, *tmp;
549
550 PARANOIA_ENTRY();
551 list_for_each_entry_safe(e, tmp, &lc->to_be_changed, list) {
552 /* count number of changes, not number of transactions */
553 ++lc->changed;
554 e->lc_number = e->lc_new_number;
555 list_move(&e->list, &lc->in_use);
556 }
557 lc->pending_changes = 0;
558 RETURN();
559}
560
561
562/**
563 * lc_put - give up refcnt of @e
564 * @lc: the lru cache to operate on
565 * @e: the element to put
566 *
567 * If refcnt reaches zero, the element is moved to the lru list,
568 * and a %LC_STARVING (if set) is cleared.
569 * Returns the new (post-decrement) refcnt.
570 */
571unsigned int lc_put(struct lru_cache *lc, struct lc_element *e)
572{
573 PARANOIA_ENTRY();
574 PARANOIA_LC_ELEMENT(lc, e);
575 BUG_ON(e->refcnt == 0);
576 BUG_ON(e->lc_number != e->lc_new_number);
577 if (--e->refcnt == 0) {
578 /* move it to the front of LRU. */
579 list_move(&e->list, &lc->lru);
580 lc->used--;
581 clear_bit_unlock(__LC_STARVING, &lc->flags);
582 }
583 RETURN(e->refcnt);
584}
585
586/**
587 * lc_element_by_index
588 * @lc: the lru cache to operate on
589 * @i: the index of the element to return
590 */
591struct lc_element *lc_element_by_index(struct lru_cache *lc, unsigned i)
592{
593 BUG_ON(i >= lc->nr_elements);
594 BUG_ON(lc->lc_element[i] == NULL);
595 BUG_ON(lc->lc_element[i]->lc_index != i);
596 return lc->lc_element[i];
597}
598
599/**
600 * lc_index_of
601 * @lc: the lru cache to operate on
602 * @e: the element to query for its index position in lc->element
603 */
604unsigned int lc_index_of(struct lru_cache *lc, struct lc_element *e)
605{
606 PARANOIA_LC_ELEMENT(lc, e);
607 return e->lc_index;
608}
609
610/**
611 * lc_set - associate index with label
612 * @lc: the lru cache to operate on
613 * @enr: the label to set
614 * @index: the element index to associate label with.
615 *
616 * Used to initialize the active set to some previously recorded state.
617 */
618void lc_set(struct lru_cache *lc, unsigned int enr, int index)
619{
620 struct lc_element *e;
621 struct list_head *lh;
622
623 if (index < 0 || index >= lc->nr_elements)
624 return;
625
626 e = lc_element_by_index(lc, index);
627 BUG_ON(e->lc_number != e->lc_new_number);
628 BUG_ON(e->refcnt != 0);
629
630 e->lc_number = e->lc_new_number = enr;
631 hlist_del_init(&e->colision);
632 if (enr == LC_FREE)
633 lh = &lc->free;
634 else {
635 hlist_add_head(&e->colision, lc_hash_slot(lc, enr));
636 lh = &lc->lru;
637 }
638 list_move(&e->list, lh);
639}
640
641/**
642 * lc_dump - Dump a complete LRU cache to seq in textual form.
643 * @lc: the lru cache to operate on
644 * @seq: the &struct seq_file pointer to seq_printf into
645 * @utext: user supplied additional "heading" or other info
646 * @detail: function pointer the user may provide to dump further details
647 * of the object the lc_element is embedded in. May be NULL.
648 * Note: a leading space ' ' and trailing newline '\n' is implied.
649 */
650void lc_seq_dump_details(struct seq_file *seq, struct lru_cache *lc, char *utext,
651 void (*detail) (struct seq_file *, struct lc_element *))
652{
653 unsigned int nr_elements = lc->nr_elements;
654 struct lc_element *e;
655 int i;
656
657 seq_printf(seq, "\tnn: lc_number (new nr) refcnt %s\n ", utext);
658 for (i = 0; i < nr_elements; i++) {
659 e = lc_element_by_index(lc, i);
660 if (e->lc_number != e->lc_new_number)
661 seq_printf(seq, "\t%5d: %6d %8d %6d ",
662 i, e->lc_number, e->lc_new_number, e->refcnt);
663 else
664 seq_printf(seq, "\t%5d: %6d %-8s %6d ",
665 i, e->lc_number, "-\"-", e->refcnt);
666 if (detail)
667 detail(seq, e);
668 seq_putc(seq, '\n');
669 }
670}
671
672EXPORT_SYMBOL(lc_create);
673EXPORT_SYMBOL(lc_reset);
674EXPORT_SYMBOL(lc_destroy);
675EXPORT_SYMBOL(lc_set);
676EXPORT_SYMBOL(lc_del);
677EXPORT_SYMBOL(lc_try_get);
678EXPORT_SYMBOL(lc_find);
679EXPORT_SYMBOL(lc_get);
680EXPORT_SYMBOL(lc_put);
681EXPORT_SYMBOL(lc_committed);
682EXPORT_SYMBOL(lc_element_by_index);
683EXPORT_SYMBOL(lc_index_of);
684EXPORT_SYMBOL(lc_seq_printf_stats);
685EXPORT_SYMBOL(lc_seq_dump_details);
686EXPORT_SYMBOL(lc_try_lock);
687EXPORT_SYMBOL(lc_is_used);
688EXPORT_SYMBOL(lc_get_cumulative);
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
21MODULE_AUTHOR("Philipp Reisner <phil@linbit.com>, "
22 "Lars Ellenberg <lars@linbit.com>");
23MODULE_DESCRIPTION("lru_cache - Track sets of hot objects");
24MODULE_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 */
55int 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#if 0
64 /* Alternative approach, spin in case someone enters or leaves a
65 * PARANOIA_ENTRY()/RETURN() section. */
66 unsigned long old, new, val;
67 do {
68 old = lc->flags & LC_PARANOIA;
69 new = old | LC_LOCKED;
70 val = cmpxchg(&lc->flags, old, new);
71 } while (unlikely (val == (old ^ LC_PARANOIA)));
72 return old == val;
73#endif
74}
75
76/**
77 * lc_create - prepares to track objects in an active set
78 * @name: descriptive name only used in lc_seq_printf_stats and lc_seq_dump_details
79 * @max_pending_changes: maximum changes to accumulate until a transaction is required
80 * @e_count: number of elements allowed to be active simultaneously
81 * @e_size: size of the tracked objects
82 * @e_off: offset to the &struct lc_element member in a tracked object
83 *
84 * Returns a pointer to a newly initialized struct lru_cache on success,
85 * or NULL on (allocation) failure.
86 */
87struct lru_cache *lc_create(const char *name, struct kmem_cache *cache,
88 unsigned max_pending_changes,
89 unsigned e_count, size_t e_size, size_t e_off)
90{
91 struct hlist_head *slot = NULL;
92 struct lc_element **element = NULL;
93 struct lru_cache *lc;
94 struct lc_element *e;
95 unsigned cache_obj_size = kmem_cache_size(cache);
96 unsigned i;
97
98 WARN_ON(cache_obj_size < e_size);
99 if (cache_obj_size < e_size)
100 return NULL;
101
102 /* e_count too big; would probably fail the allocation below anyways.
103 * for typical use cases, e_count should be few thousand at most. */
104 if (e_count > LC_MAX_ACTIVE)
105 return NULL;
106
107 slot = kcalloc(e_count, sizeof(struct hlist_head), GFP_KERNEL);
108 if (!slot)
109 goto out_fail;
110 element = kcalloc(e_count, sizeof(struct lc_element *), GFP_KERNEL);
111 if (!element)
112 goto out_fail;
113
114 lc = kzalloc(sizeof(*lc), GFP_KERNEL);
115 if (!lc)
116 goto out_fail;
117
118 INIT_LIST_HEAD(&lc->in_use);
119 INIT_LIST_HEAD(&lc->lru);
120 INIT_LIST_HEAD(&lc->free);
121 INIT_LIST_HEAD(&lc->to_be_changed);
122
123 lc->name = name;
124 lc->element_size = e_size;
125 lc->element_off = e_off;
126 lc->nr_elements = e_count;
127 lc->max_pending_changes = max_pending_changes;
128 lc->lc_cache = cache;
129 lc->lc_element = element;
130 lc->lc_slot = slot;
131
132 /* preallocate all objects */
133 for (i = 0; i < e_count; i++) {
134 void *p = kmem_cache_alloc(cache, GFP_KERNEL);
135 if (!p)
136 break;
137 memset(p, 0, lc->element_size);
138 e = p + e_off;
139 e->lc_index = i;
140 e->lc_number = LC_FREE;
141 e->lc_new_number = LC_FREE;
142 list_add(&e->list, &lc->free);
143 element[i] = e;
144 }
145 if (i == e_count)
146 return lc;
147
148 /* else: could not allocate all elements, give up */
149 for (i--; i; i--) {
150 void *p = element[i];
151 kmem_cache_free(cache, p - e_off);
152 }
153 kfree(lc);
154out_fail:
155 kfree(element);
156 kfree(slot);
157 return NULL;
158}
159
160static void lc_free_by_index(struct lru_cache *lc, unsigned i)
161{
162 void *p = lc->lc_element[i];
163 WARN_ON(!p);
164 if (p) {
165 p -= lc->element_off;
166 kmem_cache_free(lc->lc_cache, p);
167 }
168}
169
170/**
171 * lc_destroy - frees memory allocated by lc_create()
172 * @lc: the lru cache to destroy
173 */
174void lc_destroy(struct lru_cache *lc)
175{
176 unsigned i;
177 if (!lc)
178 return;
179 for (i = 0; i < lc->nr_elements; i++)
180 lc_free_by_index(lc, i);
181 kfree(lc->lc_element);
182 kfree(lc->lc_slot);
183 kfree(lc);
184}
185
186/**
187 * lc_reset - does a full reset for @lc and the hash table slots.
188 * @lc: the lru cache to operate on
189 *
190 * It is roughly the equivalent of re-allocating a fresh lru_cache object,
191 * basically a short cut to lc_destroy(lc); lc = lc_create(...);
192 */
193void lc_reset(struct lru_cache *lc)
194{
195 unsigned i;
196
197 INIT_LIST_HEAD(&lc->in_use);
198 INIT_LIST_HEAD(&lc->lru);
199 INIT_LIST_HEAD(&lc->free);
200 INIT_LIST_HEAD(&lc->to_be_changed);
201 lc->used = 0;
202 lc->hits = 0;
203 lc->misses = 0;
204 lc->starving = 0;
205 lc->locked = 0;
206 lc->changed = 0;
207 lc->pending_changes = 0;
208 lc->flags = 0;
209 memset(lc->lc_slot, 0, sizeof(struct hlist_head) * lc->nr_elements);
210
211 for (i = 0; i < lc->nr_elements; i++) {
212 struct lc_element *e = lc->lc_element[i];
213 void *p = e;
214 p -= lc->element_off;
215 memset(p, 0, lc->element_size);
216 /* re-init it */
217 e->lc_index = i;
218 e->lc_number = LC_FREE;
219 e->lc_new_number = LC_FREE;
220 list_add(&e->list, &lc->free);
221 }
222}
223
224/**
225 * lc_seq_printf_stats - print stats about @lc into @seq
226 * @seq: the seq_file to print into
227 * @lc: the lru cache to print statistics of
228 */
229void lc_seq_printf_stats(struct seq_file *seq, struct lru_cache *lc)
230{
231 /* NOTE:
232 * total calls to lc_get are
233 * (starving + hits + misses)
234 * misses include "locked" count (update from an other thread in
235 * progress) and "changed", when this in fact lead to an successful
236 * update of the cache.
237 */
238 seq_printf(seq, "\t%s: used:%u/%u hits:%lu misses:%lu starving:%lu locked:%lu changed:%lu\n",
239 lc->name, lc->used, lc->nr_elements,
240 lc->hits, lc->misses, lc->starving, lc->locked, lc->changed);
241}
242
243static struct hlist_head *lc_hash_slot(struct lru_cache *lc, unsigned int enr)
244{
245 return lc->lc_slot + (enr % lc->nr_elements);
246}
247
248
249static struct lc_element *__lc_find(struct lru_cache *lc, unsigned int enr,
250 bool include_changing)
251{
252 struct lc_element *e;
253
254 BUG_ON(!lc);
255 BUG_ON(!lc->nr_elements);
256 hlist_for_each_entry(e, lc_hash_slot(lc, enr), colision) {
257 /* "about to be changed" elements, pending transaction commit,
258 * are hashed by their "new number". "Normal" elements have
259 * lc_number == lc_new_number. */
260 if (e->lc_new_number != enr)
261 continue;
262 if (e->lc_new_number == e->lc_number || include_changing)
263 return e;
264 break;
265 }
266 return NULL;
267}
268
269/**
270 * lc_find - find element by label, if present in the hash table
271 * @lc: The lru_cache object
272 * @enr: element number
273 *
274 * Returns the pointer to an element, if the element with the requested
275 * "label" or element number is present in the hash table,
276 * or NULL if not found. Does not change the refcnt.
277 * Ignores elements that are "about to be used", i.e. not yet in the active
278 * set, but still pending transaction commit.
279 */
280struct lc_element *lc_find(struct lru_cache *lc, unsigned int enr)
281{
282 return __lc_find(lc, enr, 0);
283}
284
285/**
286 * lc_is_used - find element by label
287 * @lc: The lru_cache object
288 * @enr: element number
289 *
290 * Returns true, if the element with the requested "label" or element number is
291 * present in the hash table, and is used (refcnt > 0).
292 * Also finds elements that are not _currently_ used but only "about to be
293 * used", i.e. on the "to_be_changed" list, pending transaction commit.
294 */
295bool lc_is_used(struct lru_cache *lc, unsigned int enr)
296{
297 struct lc_element *e = __lc_find(lc, enr, 1);
298 return e && e->refcnt;
299}
300
301/**
302 * lc_del - removes an element from the cache
303 * @lc: The lru_cache object
304 * @e: The element to remove
305 *
306 * @e must be unused (refcnt == 0). Moves @e from "lru" to "free" list,
307 * sets @e->enr to %LC_FREE.
308 */
309void lc_del(struct lru_cache *lc, struct lc_element *e)
310{
311 PARANOIA_ENTRY();
312 PARANOIA_LC_ELEMENT(lc, e);
313 BUG_ON(e->refcnt);
314
315 e->lc_number = e->lc_new_number = LC_FREE;
316 hlist_del_init(&e->colision);
317 list_move(&e->list, &lc->free);
318 RETURN();
319}
320
321static struct lc_element *lc_prepare_for_change(struct lru_cache *lc, unsigned new_number)
322{
323 struct list_head *n;
324 struct lc_element *e;
325
326 if (!list_empty(&lc->free))
327 n = lc->free.next;
328 else if (!list_empty(&lc->lru))
329 n = lc->lru.prev;
330 else
331 return NULL;
332
333 e = list_entry(n, struct lc_element, list);
334 PARANOIA_LC_ELEMENT(lc, e);
335
336 e->lc_new_number = new_number;
337 if (!hlist_unhashed(&e->colision))
338 __hlist_del(&e->colision);
339 hlist_add_head(&e->colision, lc_hash_slot(lc, new_number));
340 list_move(&e->list, &lc->to_be_changed);
341
342 return e;
343}
344
345static int lc_unused_element_available(struct lru_cache *lc)
346{
347 if (!list_empty(&lc->free))
348 return 1; /* something on the free list */
349 if (!list_empty(&lc->lru))
350 return 1; /* something to evict */
351
352 return 0;
353}
354
355/* used as internal flags to __lc_get */
356enum {
357 LC_GET_MAY_CHANGE = 1,
358 LC_GET_MAY_USE_UNCOMMITTED = 2,
359};
360
361static struct lc_element *__lc_get(struct lru_cache *lc, unsigned int enr, unsigned int flags)
362{
363 struct lc_element *e;
364
365 PARANOIA_ENTRY();
366 if (lc->flags & LC_STARVING) {
367 ++lc->starving;
368 RETURN(NULL);
369 }
370
371 e = __lc_find(lc, enr, 1);
372 /* if lc_new_number != lc_number,
373 * this enr is currently being pulled in already,
374 * and will be available once the pending transaction
375 * has been committed. */
376 if (e) {
377 if (e->lc_new_number != e->lc_number) {
378 /* It has been found above, but on the "to_be_changed"
379 * list, not yet committed. Don't pull it in twice,
380 * wait for the transaction, then try again...
381 */
382 if (!(flags & LC_GET_MAY_USE_UNCOMMITTED))
383 RETURN(NULL);
384 /* ... unless the caller is aware of the implications,
385 * probably preparing a cumulative transaction. */
386 ++e->refcnt;
387 ++lc->hits;
388 RETURN(e);
389 }
390 /* else: lc_new_number == lc_number; a real hit. */
391 ++lc->hits;
392 if (e->refcnt++ == 0)
393 lc->used++;
394 list_move(&e->list, &lc->in_use); /* Not evictable... */
395 RETURN(e);
396 }
397 /* e == NULL */
398
399 ++lc->misses;
400 if (!(flags & LC_GET_MAY_CHANGE))
401 RETURN(NULL);
402
403 /* To avoid races with lc_try_lock(), first, mark us dirty
404 * (using test_and_set_bit, as it implies memory barriers), ... */
405 test_and_set_bit(__LC_DIRTY, &lc->flags);
406
407 /* ... only then check if it is locked anyways. If lc_unlock clears
408 * the dirty bit again, that's not a problem, we will come here again.
409 */
410 if (test_bit(__LC_LOCKED, &lc->flags)) {
411 ++lc->locked;
412 RETURN(NULL);
413 }
414
415 /* In case there is nothing available and we can not kick out
416 * the LRU element, we have to wait ...
417 */
418 if (!lc_unused_element_available(lc)) {
419 __set_bit(__LC_STARVING, &lc->flags);
420 RETURN(NULL);
421 }
422
423 /* It was not present in the active set. We are going to recycle an
424 * unused (or even "free") element, but we won't accumulate more than
425 * max_pending_changes changes. */
426 if (lc->pending_changes >= lc->max_pending_changes)
427 RETURN(NULL);
428
429 e = lc_prepare_for_change(lc, enr);
430 BUG_ON(!e);
431
432 clear_bit(__LC_STARVING, &lc->flags);
433 BUG_ON(++e->refcnt != 1);
434 lc->used++;
435 lc->pending_changes++;
436
437 RETURN(e);
438}
439
440/**
441 * lc_get - get element by label, maybe change the active set
442 * @lc: the lru cache to operate on
443 * @enr: the label to look up
444 *
445 * Finds an element in the cache, increases its usage count,
446 * "touches" and returns it.
447 *
448 * In case the requested number is not present, it needs to be added to the
449 * cache. Therefore it is possible that an other element becomes evicted from
450 * the cache. In either case, the user is notified so he is able to e.g. keep
451 * a persistent log of the cache changes, and therefore the objects in use.
452 *
453 * Return values:
454 * NULL
455 * The cache was marked %LC_STARVING,
456 * or the requested label was not in the active set
457 * and a changing transaction is still pending (@lc was marked %LC_DIRTY).
458 * Or no unused or free element could be recycled (@lc will be marked as
459 * %LC_STARVING, blocking further lc_get() operations).
460 *
461 * pointer to the element with the REQUESTED element number.
462 * In this case, it can be used right away
463 *
464 * pointer to an UNUSED element with some different element number,
465 * where that different number may also be %LC_FREE.
466 *
467 * In this case, the cache is marked %LC_DIRTY,
468 * so lc_try_lock() will no longer succeed.
469 * The returned element pointer is moved to the "to_be_changed" list,
470 * and registered with the new element number on the hash collision chains,
471 * so it is possible to pick it up from lc_is_used().
472 * Up to "max_pending_changes" (see lc_create()) can be accumulated.
473 * The user now should do whatever housekeeping is necessary,
474 * typically serialize on lc_try_lock_for_transaction(), then call
475 * lc_committed(lc) and lc_unlock(), to finish the change.
476 *
477 * NOTE: The user needs to check the lc_number on EACH use, so he recognizes
478 * any cache set change.
479 */
480struct lc_element *lc_get(struct lru_cache *lc, unsigned int enr)
481{
482 return __lc_get(lc, enr, LC_GET_MAY_CHANGE);
483}
484
485/**
486 * lc_get_cumulative - like lc_get; also finds to-be-changed elements
487 * @lc: the lru cache to operate on
488 * @enr: the label to look up
489 *
490 * Unlike lc_get this also returns the element for @enr, if it is belonging to
491 * a pending transaction, so the return values are like for lc_get(),
492 * plus:
493 *
494 * pointer to an element already on the "to_be_changed" list.
495 * In this case, the cache was already marked %LC_DIRTY.
496 *
497 * Caller needs to make sure that the pending transaction is completed,
498 * before proceeding to actually use this element.
499 */
500struct lc_element *lc_get_cumulative(struct lru_cache *lc, unsigned int enr)
501{
502 return __lc_get(lc, enr, LC_GET_MAY_CHANGE|LC_GET_MAY_USE_UNCOMMITTED);
503}
504
505/**
506 * lc_try_get - get element by label, if present; do not change the active set
507 * @lc: the lru cache to operate on
508 * @enr: the label to look up
509 *
510 * Finds an element in the cache, increases its usage count,
511 * "touches" and returns it.
512 *
513 * Return values:
514 * NULL
515 * The cache was marked %LC_STARVING,
516 * or the requested label was not in the active set
517 *
518 * pointer to the element with the REQUESTED element number.
519 * In this case, it can be used right away
520 */
521struct lc_element *lc_try_get(struct lru_cache *lc, unsigned int enr)
522{
523 return __lc_get(lc, enr, 0);
524}
525
526/**
527 * lc_committed - tell @lc that pending changes have been recorded
528 * @lc: the lru cache to operate on
529 *
530 * User is expected to serialize on explicit lc_try_lock_for_transaction()
531 * before the transaction is started, and later needs to lc_unlock() explicitly
532 * as well.
533 */
534void lc_committed(struct lru_cache *lc)
535{
536 struct lc_element *e, *tmp;
537
538 PARANOIA_ENTRY();
539 list_for_each_entry_safe(e, tmp, &lc->to_be_changed, list) {
540 /* count number of changes, not number of transactions */
541 ++lc->changed;
542 e->lc_number = e->lc_new_number;
543 list_move(&e->list, &lc->in_use);
544 }
545 lc->pending_changes = 0;
546 RETURN();
547}
548
549
550/**
551 * lc_put - give up refcnt of @e
552 * @lc: the lru cache to operate on
553 * @e: the element to put
554 *
555 * If refcnt reaches zero, the element is moved to the lru list,
556 * and a %LC_STARVING (if set) is cleared.
557 * Returns the new (post-decrement) refcnt.
558 */
559unsigned int lc_put(struct lru_cache *lc, struct lc_element *e)
560{
561 PARANOIA_ENTRY();
562 PARANOIA_LC_ELEMENT(lc, e);
563 BUG_ON(e->refcnt == 0);
564 BUG_ON(e->lc_number != e->lc_new_number);
565 if (--e->refcnt == 0) {
566 /* move it to the front of LRU. */
567 list_move(&e->list, &lc->lru);
568 lc->used--;
569 clear_bit_unlock(__LC_STARVING, &lc->flags);
570 }
571 RETURN(e->refcnt);
572}
573
574/**
575 * lc_element_by_index
576 * @lc: the lru cache to operate on
577 * @i: the index of the element to return
578 */
579struct lc_element *lc_element_by_index(struct lru_cache *lc, unsigned i)
580{
581 BUG_ON(i >= lc->nr_elements);
582 BUG_ON(lc->lc_element[i] == NULL);
583 BUG_ON(lc->lc_element[i]->lc_index != i);
584 return lc->lc_element[i];
585}
586
587/**
588 * lc_index_of
589 * @lc: the lru cache to operate on
590 * @e: the element to query for its index position in lc->element
591 */
592unsigned int lc_index_of(struct lru_cache *lc, struct lc_element *e)
593{
594 PARANOIA_LC_ELEMENT(lc, e);
595 return e->lc_index;
596}
597
598/**
599 * lc_set - associate index with label
600 * @lc: the lru cache to operate on
601 * @enr: the label to set
602 * @index: the element index to associate label with.
603 *
604 * Used to initialize the active set to some previously recorded state.
605 */
606void lc_set(struct lru_cache *lc, unsigned int enr, int index)
607{
608 struct lc_element *e;
609 struct list_head *lh;
610
611 if (index < 0 || index >= lc->nr_elements)
612 return;
613
614 e = lc_element_by_index(lc, index);
615 BUG_ON(e->lc_number != e->lc_new_number);
616 BUG_ON(e->refcnt != 0);
617
618 e->lc_number = e->lc_new_number = enr;
619 hlist_del_init(&e->colision);
620 if (enr == LC_FREE)
621 lh = &lc->free;
622 else {
623 hlist_add_head(&e->colision, lc_hash_slot(lc, enr));
624 lh = &lc->lru;
625 }
626 list_move(&e->list, lh);
627}
628
629/**
630 * lc_dump - Dump a complete LRU cache to seq in textual form.
631 * @lc: the lru cache to operate on
632 * @seq: the &struct seq_file pointer to seq_printf into
633 * @utext: user supplied additional "heading" or other info
634 * @detail: function pointer the user may provide to dump further details
635 * of the object the lc_element is embedded in. May be NULL.
636 * Note: a leading space ' ' and trailing newline '\n' is implied.
637 */
638void lc_seq_dump_details(struct seq_file *seq, struct lru_cache *lc, char *utext,
639 void (*detail) (struct seq_file *, struct lc_element *))
640{
641 unsigned int nr_elements = lc->nr_elements;
642 struct lc_element *e;
643 int i;
644
645 seq_printf(seq, "\tnn: lc_number (new nr) refcnt %s\n ", utext);
646 for (i = 0; i < nr_elements; i++) {
647 e = lc_element_by_index(lc, i);
648 if (e->lc_number != e->lc_new_number)
649 seq_printf(seq, "\t%5d: %6d %8d %6d ",
650 i, e->lc_number, e->lc_new_number, e->refcnt);
651 else
652 seq_printf(seq, "\t%5d: %6d %-8s %6d ",
653 i, e->lc_number, "-\"-", e->refcnt);
654 if (detail)
655 detail(seq, e);
656 seq_putc(seq, '\n');
657 }
658}
659
660EXPORT_SYMBOL(lc_create);
661EXPORT_SYMBOL(lc_reset);
662EXPORT_SYMBOL(lc_destroy);
663EXPORT_SYMBOL(lc_set);
664EXPORT_SYMBOL(lc_del);
665EXPORT_SYMBOL(lc_try_get);
666EXPORT_SYMBOL(lc_find);
667EXPORT_SYMBOL(lc_get);
668EXPORT_SYMBOL(lc_put);
669EXPORT_SYMBOL(lc_committed);
670EXPORT_SYMBOL(lc_element_by_index);
671EXPORT_SYMBOL(lc_index_of);
672EXPORT_SYMBOL(lc_seq_printf_stats);
673EXPORT_SYMBOL(lc_seq_dump_details);
674EXPORT_SYMBOL(lc_try_lock);
675EXPORT_SYMBOL(lc_is_used);
676EXPORT_SYMBOL(lc_get_cumulative);