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