blob: 964153255076bad8a1a24765ec9b7f884af4cd4b [file] [log] [blame]
Joe Thornberf2836352013-03-01 22:45:51 +00001/*
2 * Copyright (C) 2012 Red Hat. All rights reserved.
3 *
4 * This file is released under the GPL.
5 */
6
7#include "dm-cache-policy.h"
8#include "dm.h"
9
10#include <linux/hash.h>
11#include <linux/module.h>
12#include <linux/mutex.h>
13#include <linux/slab.h>
14#include <linux/vmalloc.h>
15
16#define DM_MSG_PREFIX "cache-policy-mq"
17#define MQ_VERSION "1.0.0"
18
19static struct kmem_cache *mq_entry_cache;
20
21/*----------------------------------------------------------------*/
22
23static unsigned next_power(unsigned n, unsigned min)
24{
25 return roundup_pow_of_two(max(n, min));
26}
27
28/*----------------------------------------------------------------*/
29
30static unsigned long *alloc_bitset(unsigned nr_entries)
31{
32 size_t s = sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG);
33 return vzalloc(s);
34}
35
36static void free_bitset(unsigned long *bits)
37{
38 vfree(bits);
39}
40
41/*----------------------------------------------------------------*/
42
43/*
44 * Large, sequential ios are probably better left on the origin device since
45 * spindles tend to have good bandwidth.
46 *
47 * The io_tracker tries to spot when the io is in one of these sequential
48 * modes.
49 *
50 * Two thresholds to switch between random and sequential io mode are defaulting
51 * as follows and can be adjusted via the constructor and message interfaces.
52 */
53#define RANDOM_THRESHOLD_DEFAULT 4
54#define SEQUENTIAL_THRESHOLD_DEFAULT 512
55
56enum io_pattern {
57 PATTERN_SEQUENTIAL,
58 PATTERN_RANDOM
59};
60
61struct io_tracker {
62 enum io_pattern pattern;
63
64 unsigned nr_seq_samples;
65 unsigned nr_rand_samples;
66 unsigned thresholds[2];
67
68 dm_oblock_t last_end_oblock;
69};
70
71static void iot_init(struct io_tracker *t,
72 int sequential_threshold, int random_threshold)
73{
74 t->pattern = PATTERN_RANDOM;
75 t->nr_seq_samples = 0;
76 t->nr_rand_samples = 0;
77 t->last_end_oblock = 0;
78 t->thresholds[PATTERN_RANDOM] = random_threshold;
79 t->thresholds[PATTERN_SEQUENTIAL] = sequential_threshold;
80}
81
82static enum io_pattern iot_pattern(struct io_tracker *t)
83{
84 return t->pattern;
85}
86
87static void iot_update_stats(struct io_tracker *t, struct bio *bio)
88{
89 if (bio->bi_sector == from_oblock(t->last_end_oblock) + 1)
90 t->nr_seq_samples++;
91 else {
92 /*
93 * Just one non-sequential IO is enough to reset the
94 * counters.
95 */
96 if (t->nr_seq_samples) {
97 t->nr_seq_samples = 0;
98 t->nr_rand_samples = 0;
99 }
100
101 t->nr_rand_samples++;
102 }
103
104 t->last_end_oblock = to_oblock(bio->bi_sector + bio_sectors(bio) - 1);
105}
106
107static void iot_check_for_pattern_switch(struct io_tracker *t)
108{
109 switch (t->pattern) {
110 case PATTERN_SEQUENTIAL:
111 if (t->nr_rand_samples >= t->thresholds[PATTERN_RANDOM]) {
112 t->pattern = PATTERN_RANDOM;
113 t->nr_seq_samples = t->nr_rand_samples = 0;
114 }
115 break;
116
117 case PATTERN_RANDOM:
118 if (t->nr_seq_samples >= t->thresholds[PATTERN_SEQUENTIAL]) {
119 t->pattern = PATTERN_SEQUENTIAL;
120 t->nr_seq_samples = t->nr_rand_samples = 0;
121 }
122 break;
123 }
124}
125
126static void iot_examine_bio(struct io_tracker *t, struct bio *bio)
127{
128 iot_update_stats(t, bio);
129 iot_check_for_pattern_switch(t);
130}
131
132/*----------------------------------------------------------------*/
133
134
135/*
136 * This queue is divided up into different levels. Allowing us to push
137 * entries to the back of any of the levels. Think of it as a partially
138 * sorted queue.
139 */
140#define NR_QUEUE_LEVELS 16u
141
142struct queue {
143 struct list_head qs[NR_QUEUE_LEVELS];
144};
145
146static void queue_init(struct queue *q)
147{
148 unsigned i;
149
150 for (i = 0; i < NR_QUEUE_LEVELS; i++)
151 INIT_LIST_HEAD(q->qs + i);
152}
153
154/*
155 * Insert an entry to the back of the given level.
156 */
157static void queue_push(struct queue *q, unsigned level, struct list_head *elt)
158{
159 list_add_tail(elt, q->qs + level);
160}
161
162static void queue_remove(struct list_head *elt)
163{
164 list_del(elt);
165}
166
167/*
168 * Shifts all regions down one level. This has no effect on the order of
169 * the queue.
170 */
171static void queue_shift_down(struct queue *q)
172{
173 unsigned level;
174
175 for (level = 1; level < NR_QUEUE_LEVELS; level++)
176 list_splice_init(q->qs + level, q->qs + level - 1);
177}
178
179/*
180 * Gives us the oldest entry of the lowest popoulated level. If the first
181 * level is emptied then we shift down one level.
182 */
183static struct list_head *queue_pop(struct queue *q)
184{
185 unsigned level;
186 struct list_head *r;
187
188 for (level = 0; level < NR_QUEUE_LEVELS; level++)
189 if (!list_empty(q->qs + level)) {
190 r = q->qs[level].next;
191 list_del(r);
192
193 /* have we just emptied the bottom level? */
194 if (level == 0 && list_empty(q->qs))
195 queue_shift_down(q);
196
197 return r;
198 }
199
200 return NULL;
201}
202
203static struct list_head *list_pop(struct list_head *lh)
204{
205 struct list_head *r = lh->next;
206
207 BUG_ON(!r);
208 list_del_init(r);
209
210 return r;
211}
212
213/*----------------------------------------------------------------*/
214
215/*
216 * Describes a cache entry. Used in both the cache and the pre_cache.
217 */
218struct entry {
219 struct hlist_node hlist;
220 struct list_head list;
221 dm_oblock_t oblock;
222 dm_cblock_t cblock; /* valid iff in_cache */
223
224 /*
225 * FIXME: pack these better
226 */
227 bool in_cache:1;
228 unsigned hit_count;
229 unsigned generation;
230 unsigned tick;
231};
232
233struct mq_policy {
234 struct dm_cache_policy policy;
235
236 /* protects everything */
237 struct mutex lock;
238 dm_cblock_t cache_size;
239 struct io_tracker tracker;
240
241 /*
242 * We maintain two queues of entries. The cache proper contains
243 * the currently active mappings. Whereas the pre_cache tracks
244 * blocks that are being hit frequently and potential candidates
245 * for promotion to the cache.
246 */
247 struct queue pre_cache;
248 struct queue cache;
249
250 /*
251 * Keeps track of time, incremented by the core. We use this to
252 * avoid attributing multiple hits within the same tick.
253 *
254 * Access to tick_protected should be done with the spin lock held.
255 * It's copied to tick at the start of the map function (within the
256 * mutex).
257 */
258 spinlock_t tick_lock;
259 unsigned tick_protected;
260 unsigned tick;
261
262 /*
263 * A count of the number of times the map function has been called
264 * and found an entry in the pre_cache or cache. Currently used to
265 * calculate the generation.
266 */
267 unsigned hit_count;
268
269 /*
270 * A generation is a longish period that is used to trigger some
271 * book keeping effects. eg, decrementing hit counts on entries.
272 * This is needed to allow the cache to evolve as io patterns
273 * change.
274 */
275 unsigned generation;
276 unsigned generation_period; /* in lookups (will probably change) */
277
278 /*
279 * Entries in the pre_cache whose hit count passes the promotion
280 * threshold move to the cache proper. Working out the correct
281 * value for the promotion_threshold is crucial to this policy.
282 */
283 unsigned promote_threshold;
284
285 /*
286 * We need cache_size entries for the cache, and choose to have
287 * cache_size entries for the pre_cache too. One motivation for
288 * using the same size is to make the hit counts directly
289 * comparable between pre_cache and cache.
290 */
291 unsigned nr_entries;
292 unsigned nr_entries_allocated;
293 struct list_head free;
294
295 /*
296 * Cache blocks may be unallocated. We store this info in a
297 * bitset.
298 */
299 unsigned long *allocation_bitset;
300 unsigned nr_cblocks_allocated;
301 unsigned find_free_nr_words;
302 unsigned find_free_last_word;
303
304 /*
305 * The hash table allows us to quickly find an entry by origin
306 * block. Both pre_cache and cache entries are in here.
307 */
308 unsigned nr_buckets;
309 dm_block_t hash_bits;
310 struct hlist_head *table;
311};
312
313/*----------------------------------------------------------------*/
314/* Free/alloc mq cache entry structures. */
315static void takeout_queue(struct list_head *lh, struct queue *q)
316{
317 unsigned level;
318
319 for (level = 0; level < NR_QUEUE_LEVELS; level++)
320 list_splice(q->qs + level, lh);
321}
322
323static void free_entries(struct mq_policy *mq)
324{
325 struct entry *e, *tmp;
326
327 takeout_queue(&mq->free, &mq->pre_cache);
328 takeout_queue(&mq->free, &mq->cache);
329
330 list_for_each_entry_safe(e, tmp, &mq->free, list)
331 kmem_cache_free(mq_entry_cache, e);
332}
333
334static int alloc_entries(struct mq_policy *mq, unsigned elts)
335{
336 unsigned u = mq->nr_entries;
337
338 INIT_LIST_HEAD(&mq->free);
339 mq->nr_entries_allocated = 0;
340
341 while (u--) {
342 struct entry *e = kmem_cache_zalloc(mq_entry_cache, GFP_KERNEL);
343
344 if (!e) {
345 free_entries(mq);
346 return -ENOMEM;
347 }
348
349
350 list_add(&e->list, &mq->free);
351 }
352
353 return 0;
354}
355
356/*----------------------------------------------------------------*/
357
358/*
359 * Simple hash table implementation. Should replace with the standard hash
360 * table that's making its way upstream.
361 */
362static void hash_insert(struct mq_policy *mq, struct entry *e)
363{
364 unsigned h = hash_64(from_oblock(e->oblock), mq->hash_bits);
365
366 hlist_add_head(&e->hlist, mq->table + h);
367}
368
369static struct entry *hash_lookup(struct mq_policy *mq, dm_oblock_t oblock)
370{
371 unsigned h = hash_64(from_oblock(oblock), mq->hash_bits);
372 struct hlist_head *bucket = mq->table + h;
373 struct entry *e;
374
375 hlist_for_each_entry(e, bucket, hlist)
376 if (e->oblock == oblock) {
377 hlist_del(&e->hlist);
378 hlist_add_head(&e->hlist, bucket);
379 return e;
380 }
381
382 return NULL;
383}
384
385static void hash_remove(struct entry *e)
386{
387 hlist_del(&e->hlist);
388}
389
390/*----------------------------------------------------------------*/
391
392/*
393 * Allocates a new entry structure. The memory is allocated in one lump,
394 * so we just handing it out here. Returns NULL if all entries have
395 * already been allocated. Cannot fail otherwise.
396 */
397static struct entry *alloc_entry(struct mq_policy *mq)
398{
399 struct entry *e;
400
401 if (mq->nr_entries_allocated >= mq->nr_entries) {
402 BUG_ON(!list_empty(&mq->free));
403 return NULL;
404 }
405
406 e = list_entry(list_pop(&mq->free), struct entry, list);
407 INIT_LIST_HEAD(&e->list);
408 INIT_HLIST_NODE(&e->hlist);
409
410 mq->nr_entries_allocated++;
411 return e;
412}
413
414/*----------------------------------------------------------------*/
415
416/*
417 * Mark cache blocks allocated or not in the bitset.
418 */
419static void alloc_cblock(struct mq_policy *mq, dm_cblock_t cblock)
420{
421 BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size));
422 BUG_ON(test_bit(from_cblock(cblock), mq->allocation_bitset));
423
424 set_bit(from_cblock(cblock), mq->allocation_bitset);
425 mq->nr_cblocks_allocated++;
426}
427
428static void free_cblock(struct mq_policy *mq, dm_cblock_t cblock)
429{
430 BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size));
431 BUG_ON(!test_bit(from_cblock(cblock), mq->allocation_bitset));
432
433 clear_bit(from_cblock(cblock), mq->allocation_bitset);
434 mq->nr_cblocks_allocated--;
435}
436
437static bool any_free_cblocks(struct mq_policy *mq)
438{
439 return mq->nr_cblocks_allocated < from_cblock(mq->cache_size);
440}
441
442/*
443 * Fills result out with a cache block that isn't in use, or return
444 * -ENOSPC. This does _not_ mark the cblock as allocated, the caller is
445 * reponsible for that.
446 */
447static int __find_free_cblock(struct mq_policy *mq, unsigned begin, unsigned end,
448 dm_cblock_t *result, unsigned *last_word)
449{
450 int r = -ENOSPC;
451 unsigned w;
452
453 for (w = begin; w < end; w++) {
454 /*
455 * ffz is undefined if no zero exists
456 */
457 if (mq->allocation_bitset[w] != ~0UL) {
458 *last_word = w;
459 *result = to_cblock((w * BITS_PER_LONG) + ffz(mq->allocation_bitset[w]));
460 if (from_cblock(*result) < from_cblock(mq->cache_size))
461 r = 0;
462
463 break;
464 }
465 }
466
467 return r;
468}
469
470static int find_free_cblock(struct mq_policy *mq, dm_cblock_t *result)
471{
472 int r;
473
474 if (!any_free_cblocks(mq))
475 return -ENOSPC;
476
477 r = __find_free_cblock(mq, mq->find_free_last_word, mq->find_free_nr_words, result, &mq->find_free_last_word);
478 if (r == -ENOSPC && mq->find_free_last_word)
479 r = __find_free_cblock(mq, 0, mq->find_free_last_word, result, &mq->find_free_last_word);
480
481 return r;
482}
483
484/*----------------------------------------------------------------*/
485
486/*
487 * Now we get to the meat of the policy. This section deals with deciding
488 * when to to add entries to the pre_cache and cache, and move between
489 * them.
490 */
491
492/*
493 * The queue level is based on the log2 of the hit count.
494 */
495static unsigned queue_level(struct entry *e)
496{
497 return min((unsigned) ilog2(e->hit_count), NR_QUEUE_LEVELS - 1u);
498}
499
500/*
501 * Inserts the entry into the pre_cache or the cache. Ensures the cache
502 * block is marked as allocated if necc. Inserts into the hash table. Sets the
503 * tick which records when the entry was last moved about.
504 */
505static void push(struct mq_policy *mq, struct entry *e)
506{
507 e->tick = mq->tick;
508 hash_insert(mq, e);
509
510 if (e->in_cache) {
511 alloc_cblock(mq, e->cblock);
512 queue_push(&mq->cache, queue_level(e), &e->list);
513 } else
514 queue_push(&mq->pre_cache, queue_level(e), &e->list);
515}
516
517/*
518 * Removes an entry from pre_cache or cache. Removes from the hash table.
519 * Frees off the cache block if necc.
520 */
521static void del(struct mq_policy *mq, struct entry *e)
522{
523 queue_remove(&e->list);
524 hash_remove(e);
525 if (e->in_cache)
526 free_cblock(mq, e->cblock);
527}
528
529/*
530 * Like del, except it removes the first entry in the queue (ie. the least
531 * recently used).
532 */
533static struct entry *pop(struct mq_policy *mq, struct queue *q)
534{
535 struct entry *e = container_of(queue_pop(q), struct entry, list);
536
537 if (e) {
538 hash_remove(e);
539
540 if (e->in_cache)
541 free_cblock(mq, e->cblock);
542 }
543
544 return e;
545}
546
547/*
548 * Has this entry already been updated?
549 */
550static bool updated_this_tick(struct mq_policy *mq, struct entry *e)
551{
552 return mq->tick == e->tick;
553}
554
555/*
556 * The promotion threshold is adjusted every generation. As are the counts
557 * of the entries.
558 *
559 * At the moment the threshold is taken by averaging the hit counts of some
560 * of the entries in the cache (the first 20 entries of the first level).
561 *
562 * We can be much cleverer than this though. For example, each promotion
563 * could bump up the threshold helping to prevent churn. Much more to do
564 * here.
565 */
566
567#define MAX_TO_AVERAGE 20
568
569static void check_generation(struct mq_policy *mq)
570{
571 unsigned total = 0, nr = 0, count = 0, level;
572 struct list_head *head;
573 struct entry *e;
574
575 if ((mq->hit_count >= mq->generation_period) &&
576 (mq->nr_cblocks_allocated == from_cblock(mq->cache_size))) {
577
578 mq->hit_count = 0;
579 mq->generation++;
580
581 for (level = 0; level < NR_QUEUE_LEVELS && count < MAX_TO_AVERAGE; level++) {
582 head = mq->cache.qs + level;
583 list_for_each_entry(e, head, list) {
584 nr++;
585 total += e->hit_count;
586
587 if (++count >= MAX_TO_AVERAGE)
588 break;
589 }
590 }
591
592 mq->promote_threshold = nr ? total / nr : 1;
593 if (mq->promote_threshold * nr < total)
594 mq->promote_threshold++;
595 }
596}
597
598/*
599 * Whenever we use an entry we bump up it's hit counter, and push it to the
600 * back to it's current level.
601 */
602static void requeue_and_update_tick(struct mq_policy *mq, struct entry *e)
603{
604 if (updated_this_tick(mq, e))
605 return;
606
607 e->hit_count++;
608 mq->hit_count++;
609 check_generation(mq);
610
611 /* generation adjustment, to stop the counts increasing forever. */
612 /* FIXME: divide? */
613 /* e->hit_count -= min(e->hit_count - 1, mq->generation - e->generation); */
614 e->generation = mq->generation;
615
616 del(mq, e);
617 push(mq, e);
618}
619
620/*
621 * Demote the least recently used entry from the cache to the pre_cache.
622 * Returns the new cache entry to use, and the old origin block it was
623 * mapped to.
624 *
625 * We drop the hit count on the demoted entry back to 1 to stop it bouncing
626 * straight back into the cache if it's subsequently hit. There are
627 * various options here, and more experimentation would be good:
628 *
629 * - just forget about the demoted entry completely (ie. don't insert it
630 into the pre_cache).
631 * - divide the hit count rather that setting to some hard coded value.
632 * - set the hit count to a hard coded value other than 1, eg, is it better
633 * if it goes in at level 2?
634 */
635static dm_cblock_t demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock)
636{
637 dm_cblock_t result;
638 struct entry *demoted = pop(mq, &mq->cache);
639
640 BUG_ON(!demoted);
641 result = demoted->cblock;
642 *oblock = demoted->oblock;
643 demoted->in_cache = false;
644 demoted->hit_count = 1;
645 push(mq, demoted);
646
647 return result;
648}
649
650/*
651 * We modify the basic promotion_threshold depending on the specific io.
652 *
653 * If the origin block has been discarded then there's no cost to copy it
654 * to the cache.
655 *
656 * We bias towards reads, since they can be demoted at no cost if they
657 * haven't been dirtied.
658 */
659#define DISCARDED_PROMOTE_THRESHOLD 1
660#define READ_PROMOTE_THRESHOLD 4
661#define WRITE_PROMOTE_THRESHOLD 8
662
663static unsigned adjusted_promote_threshold(struct mq_policy *mq,
664 bool discarded_oblock, int data_dir)
665{
666 if (discarded_oblock && any_free_cblocks(mq) && data_dir == WRITE)
667 /*
668 * We don't need to do any copying at all, so give this a
669 * very low threshold. In practice this only triggers
670 * during initial population after a format.
671 */
672 return DISCARDED_PROMOTE_THRESHOLD;
673
674 return data_dir == READ ?
675 (mq->promote_threshold + READ_PROMOTE_THRESHOLD) :
676 (mq->promote_threshold + WRITE_PROMOTE_THRESHOLD);
677}
678
679static bool should_promote(struct mq_policy *mq, struct entry *e,
680 bool discarded_oblock, int data_dir)
681{
682 return e->hit_count >=
683 adjusted_promote_threshold(mq, discarded_oblock, data_dir);
684}
685
686static int cache_entry_found(struct mq_policy *mq,
687 struct entry *e,
688 struct policy_result *result)
689{
690 requeue_and_update_tick(mq, e);
691
692 if (e->in_cache) {
693 result->op = POLICY_HIT;
694 result->cblock = e->cblock;
695 }
696
697 return 0;
698}
699
700/*
701 * Moves and entry from the pre_cache to the cache. The main work is
702 * finding which cache block to use.
703 */
704static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e,
705 struct policy_result *result)
706{
707 dm_cblock_t cblock;
708
709 if (find_free_cblock(mq, &cblock) == -ENOSPC) {
710 result->op = POLICY_REPLACE;
711 cblock = demote_cblock(mq, &result->old_oblock);
712 } else
713 result->op = POLICY_NEW;
714
715 result->cblock = e->cblock = cblock;
716
717 del(mq, e);
718 e->in_cache = true;
719 push(mq, e);
720
721 return 0;
722}
723
724static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e,
725 bool can_migrate, bool discarded_oblock,
726 int data_dir, struct policy_result *result)
727{
728 int r = 0;
729 bool updated = updated_this_tick(mq, e);
730
731 requeue_and_update_tick(mq, e);
732
733 if ((!discarded_oblock && updated) ||
734 !should_promote(mq, e, discarded_oblock, data_dir))
735 result->op = POLICY_MISS;
736 else if (!can_migrate)
737 r = -EWOULDBLOCK;
738 else
739 r = pre_cache_to_cache(mq, e, result);
740
741 return r;
742}
743
744static void insert_in_pre_cache(struct mq_policy *mq,
745 dm_oblock_t oblock)
746{
747 struct entry *e = alloc_entry(mq);
748
749 if (!e)
750 /*
751 * There's no spare entry structure, so we grab the least
752 * used one from the pre_cache.
753 */
754 e = pop(mq, &mq->pre_cache);
755
756 if (unlikely(!e)) {
757 DMWARN("couldn't pop from pre cache");
758 return;
759 }
760
761 e->in_cache = false;
762 e->oblock = oblock;
763 e->hit_count = 1;
764 e->generation = mq->generation;
765 push(mq, e);
766}
767
768static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock,
769 struct policy_result *result)
770{
771 struct entry *e;
772 dm_cblock_t cblock;
773
774 if (find_free_cblock(mq, &cblock) == -ENOSPC) {
775 result->op = POLICY_MISS;
776 insert_in_pre_cache(mq, oblock);
777 return;
778 }
779
780 e = alloc_entry(mq);
781 if (unlikely(!e)) {
782 result->op = POLICY_MISS;
783 return;
784 }
785
786 e->oblock = oblock;
787 e->cblock = cblock;
788 e->in_cache = true;
789 e->hit_count = 1;
790 e->generation = mq->generation;
791 push(mq, e);
792
793 result->op = POLICY_NEW;
794 result->cblock = e->cblock;
795}
796
797static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock,
798 bool can_migrate, bool discarded_oblock,
799 int data_dir, struct policy_result *result)
800{
801 if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) == 1) {
802 if (can_migrate)
803 insert_in_cache(mq, oblock, result);
804 else
805 return -EWOULDBLOCK;
806 } else {
807 insert_in_pre_cache(mq, oblock);
808 result->op = POLICY_MISS;
809 }
810
811 return 0;
812}
813
814/*
815 * Looks the oblock up in the hash table, then decides whether to put in
816 * pre_cache, or cache etc.
817 */
818static int map(struct mq_policy *mq, dm_oblock_t oblock,
819 bool can_migrate, bool discarded_oblock,
820 int data_dir, struct policy_result *result)
821{
822 int r = 0;
823 struct entry *e = hash_lookup(mq, oblock);
824
825 if (e && e->in_cache)
826 r = cache_entry_found(mq, e, result);
827 else if (iot_pattern(&mq->tracker) == PATTERN_SEQUENTIAL)
828 result->op = POLICY_MISS;
829 else if (e)
830 r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock,
831 data_dir, result);
832 else
833 r = no_entry_found(mq, oblock, can_migrate, discarded_oblock,
834 data_dir, result);
835
836 if (r == -EWOULDBLOCK)
837 result->op = POLICY_MISS;
838
839 return r;
840}
841
842/*----------------------------------------------------------------*/
843
844/*
845 * Public interface, via the policy struct. See dm-cache-policy.h for a
846 * description of these.
847 */
848
849static struct mq_policy *to_mq_policy(struct dm_cache_policy *p)
850{
851 return container_of(p, struct mq_policy, policy);
852}
853
854static void mq_destroy(struct dm_cache_policy *p)
855{
856 struct mq_policy *mq = to_mq_policy(p);
857
858 free_bitset(mq->allocation_bitset);
859 kfree(mq->table);
860 free_entries(mq);
861 kfree(mq);
862}
863
864static void copy_tick(struct mq_policy *mq)
865{
866 unsigned long flags;
867
868 spin_lock_irqsave(&mq->tick_lock, flags);
869 mq->tick = mq->tick_protected;
870 spin_unlock_irqrestore(&mq->tick_lock, flags);
871}
872
873static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock,
874 bool can_block, bool can_migrate, bool discarded_oblock,
875 struct bio *bio, struct policy_result *result)
876{
877 int r;
878 struct mq_policy *mq = to_mq_policy(p);
879
880 result->op = POLICY_MISS;
881
882 if (can_block)
883 mutex_lock(&mq->lock);
884 else if (!mutex_trylock(&mq->lock))
885 return -EWOULDBLOCK;
886
887 copy_tick(mq);
888
889 iot_examine_bio(&mq->tracker, bio);
890 r = map(mq, oblock, can_migrate, discarded_oblock,
891 bio_data_dir(bio), result);
892
893 mutex_unlock(&mq->lock);
894
895 return r;
896}
897
898static int mq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock)
899{
900 int r;
901 struct mq_policy *mq = to_mq_policy(p);
902 struct entry *e;
903
904 if (!mutex_trylock(&mq->lock))
905 return -EWOULDBLOCK;
906
907 e = hash_lookup(mq, oblock);
908 if (e && e->in_cache) {
909 *cblock = e->cblock;
910 r = 0;
911 } else
912 r = -ENOENT;
913
914 mutex_unlock(&mq->lock);
915
916 return r;
917}
918
919static int mq_load_mapping(struct dm_cache_policy *p,
920 dm_oblock_t oblock, dm_cblock_t cblock,
921 uint32_t hint, bool hint_valid)
922{
923 struct mq_policy *mq = to_mq_policy(p);
924 struct entry *e;
925
926 e = alloc_entry(mq);
927 if (!e)
928 return -ENOMEM;
929
930 e->cblock = cblock;
931 e->oblock = oblock;
932 e->in_cache = true;
933 e->hit_count = hint_valid ? hint : 1;
934 e->generation = mq->generation;
935 push(mq, e);
936
937 return 0;
938}
939
940static int mq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn,
941 void *context)
942{
943 struct mq_policy *mq = to_mq_policy(p);
944 int r = 0;
945 struct entry *e;
946 unsigned level;
947
948 mutex_lock(&mq->lock);
949
950 for (level = 0; level < NR_QUEUE_LEVELS; level++)
951 list_for_each_entry(e, &mq->cache.qs[level], list) {
952 r = fn(context, e->cblock, e->oblock, e->hit_count);
953 if (r)
954 goto out;
955 }
956
957out:
958 mutex_unlock(&mq->lock);
959
960 return r;
961}
962
963static void remove_mapping(struct mq_policy *mq, dm_oblock_t oblock)
964{
965 struct entry *e = hash_lookup(mq, oblock);
966
967 BUG_ON(!e || !e->in_cache);
968
969 del(mq, e);
970 e->in_cache = false;
971 push(mq, e);
972}
973
974static void mq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock)
975{
976 struct mq_policy *mq = to_mq_policy(p);
977
978 mutex_lock(&mq->lock);
979 remove_mapping(mq, oblock);
980 mutex_unlock(&mq->lock);
981}
982
983static void force_mapping(struct mq_policy *mq,
984 dm_oblock_t current_oblock, dm_oblock_t new_oblock)
985{
986 struct entry *e = hash_lookup(mq, current_oblock);
987
988 BUG_ON(!e || !e->in_cache);
989
990 del(mq, e);
991 e->oblock = new_oblock;
992 push(mq, e);
993}
994
995static void mq_force_mapping(struct dm_cache_policy *p,
996 dm_oblock_t current_oblock, dm_oblock_t new_oblock)
997{
998 struct mq_policy *mq = to_mq_policy(p);
999
1000 mutex_lock(&mq->lock);
1001 force_mapping(mq, current_oblock, new_oblock);
1002 mutex_unlock(&mq->lock);
1003}
1004
1005static dm_cblock_t mq_residency(struct dm_cache_policy *p)
1006{
1007 struct mq_policy *mq = to_mq_policy(p);
1008
1009 /* FIXME: lock mutex, not sure we can block here */
1010 return to_cblock(mq->nr_cblocks_allocated);
1011}
1012
1013static void mq_tick(struct dm_cache_policy *p)
1014{
1015 struct mq_policy *mq = to_mq_policy(p);
1016 unsigned long flags;
1017
1018 spin_lock_irqsave(&mq->tick_lock, flags);
1019 mq->tick_protected++;
1020 spin_unlock_irqrestore(&mq->tick_lock, flags);
1021}
1022
1023static int mq_set_config_value(struct dm_cache_policy *p,
1024 const char *key, const char *value)
1025{
1026 struct mq_policy *mq = to_mq_policy(p);
1027 enum io_pattern pattern;
1028 unsigned long tmp;
1029
1030 if (!strcasecmp(key, "random_threshold"))
1031 pattern = PATTERN_RANDOM;
1032 else if (!strcasecmp(key, "sequential_threshold"))
1033 pattern = PATTERN_SEQUENTIAL;
1034 else
1035 return -EINVAL;
1036
1037 if (kstrtoul(value, 10, &tmp))
1038 return -EINVAL;
1039
1040 mq->tracker.thresholds[pattern] = tmp;
1041
1042 return 0;
1043}
1044
1045static int mq_emit_config_values(struct dm_cache_policy *p, char *result, unsigned maxlen)
1046{
1047 ssize_t sz = 0;
1048 struct mq_policy *mq = to_mq_policy(p);
1049
1050 DMEMIT("4 random_threshold %u sequential_threshold %u",
1051 mq->tracker.thresholds[PATTERN_RANDOM],
1052 mq->tracker.thresholds[PATTERN_SEQUENTIAL]);
1053
1054 return 0;
1055}
1056
1057/* Init the policy plugin interface function pointers. */
1058static void init_policy_functions(struct mq_policy *mq)
1059{
1060 mq->policy.destroy = mq_destroy;
1061 mq->policy.map = mq_map;
1062 mq->policy.lookup = mq_lookup;
1063 mq->policy.load_mapping = mq_load_mapping;
1064 mq->policy.walk_mappings = mq_walk_mappings;
1065 mq->policy.remove_mapping = mq_remove_mapping;
1066 mq->policy.writeback_work = NULL;
1067 mq->policy.force_mapping = mq_force_mapping;
1068 mq->policy.residency = mq_residency;
1069 mq->policy.tick = mq_tick;
1070 mq->policy.emit_config_values = mq_emit_config_values;
1071 mq->policy.set_config_value = mq_set_config_value;
1072}
1073
1074static struct dm_cache_policy *mq_create(dm_cblock_t cache_size,
1075 sector_t origin_size,
1076 sector_t cache_block_size)
1077{
1078 int r;
1079 struct mq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL);
1080
1081 if (!mq)
1082 return NULL;
1083
1084 init_policy_functions(mq);
1085 iot_init(&mq->tracker, SEQUENTIAL_THRESHOLD_DEFAULT, RANDOM_THRESHOLD_DEFAULT);
1086
1087 mq->cache_size = cache_size;
1088 mq->tick_protected = 0;
1089 mq->tick = 0;
1090 mq->hit_count = 0;
1091 mq->generation = 0;
1092 mq->promote_threshold = 0;
1093 mutex_init(&mq->lock);
1094 spin_lock_init(&mq->tick_lock);
1095 mq->find_free_nr_words = dm_div_up(from_cblock(mq->cache_size), BITS_PER_LONG);
1096 mq->find_free_last_word = 0;
1097
1098 queue_init(&mq->pre_cache);
1099 queue_init(&mq->cache);
1100 mq->generation_period = max((unsigned) from_cblock(cache_size), 1024U);
1101
1102 mq->nr_entries = 2 * from_cblock(cache_size);
1103 r = alloc_entries(mq, mq->nr_entries);
1104 if (r)
1105 goto bad_cache_alloc;
1106
1107 mq->nr_entries_allocated = 0;
1108 mq->nr_cblocks_allocated = 0;
1109
1110 mq->nr_buckets = next_power(from_cblock(cache_size) / 2, 16);
1111 mq->hash_bits = ffs(mq->nr_buckets) - 1;
1112 mq->table = kzalloc(sizeof(*mq->table) * mq->nr_buckets, GFP_KERNEL);
1113 if (!mq->table)
1114 goto bad_alloc_table;
1115
1116 mq->allocation_bitset = alloc_bitset(from_cblock(cache_size));
1117 if (!mq->allocation_bitset)
1118 goto bad_alloc_bitset;
1119
1120 return &mq->policy;
1121
1122bad_alloc_bitset:
1123 kfree(mq->table);
1124bad_alloc_table:
1125 free_entries(mq);
1126bad_cache_alloc:
1127 kfree(mq);
1128
1129 return NULL;
1130}
1131
1132/*----------------------------------------------------------------*/
1133
1134static struct dm_cache_policy_type mq_policy_type = {
1135 .name = "mq",
1136 .hint_size = 4,
1137 .owner = THIS_MODULE,
1138 .create = mq_create
1139};
1140
1141static struct dm_cache_policy_type default_policy_type = {
1142 .name = "default",
1143 .hint_size = 4,
1144 .owner = THIS_MODULE,
1145 .create = mq_create
1146};
1147
1148static int __init mq_init(void)
1149{
1150 int r;
1151
1152 mq_entry_cache = kmem_cache_create("dm_mq_policy_cache_entry",
1153 sizeof(struct entry),
1154 __alignof__(struct entry),
1155 0, NULL);
1156 if (!mq_entry_cache)
1157 goto bad;
1158
1159 r = dm_cache_policy_register(&mq_policy_type);
1160 if (r) {
1161 DMERR("register failed %d", r);
1162 goto bad_register_mq;
1163 }
1164
1165 r = dm_cache_policy_register(&default_policy_type);
1166 if (!r) {
1167 DMINFO("version " MQ_VERSION " loaded");
1168 return 0;
1169 }
1170
1171 DMERR("register failed (as default) %d", r);
1172
1173 dm_cache_policy_unregister(&mq_policy_type);
1174bad_register_mq:
1175 kmem_cache_destroy(mq_entry_cache);
1176bad:
1177 return -ENOMEM;
1178}
1179
1180static void __exit mq_exit(void)
1181{
1182 dm_cache_policy_unregister(&mq_policy_type);
1183 dm_cache_policy_unregister(&default_policy_type);
1184
1185 kmem_cache_destroy(mq_entry_cache);
1186}
1187
1188module_init(mq_init);
1189module_exit(mq_exit);
1190
1191MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>");
1192MODULE_LICENSE("GPL");
1193MODULE_DESCRIPTION("mq cache policy");
1194
1195MODULE_ALIAS("dm-cache-default");