Commit | Line | Data |
---|---|---|
f2836352 JT |
1 | /* |
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" | |
f2836352 JT |
17 | |
18 | static struct kmem_cache *mq_entry_cache; | |
19 | ||
20 | /*----------------------------------------------------------------*/ | |
21 | ||
22 | static unsigned next_power(unsigned n, unsigned min) | |
23 | { | |
24 | return roundup_pow_of_two(max(n, min)); | |
25 | } | |
26 | ||
27 | /*----------------------------------------------------------------*/ | |
28 | ||
29 | static unsigned long *alloc_bitset(unsigned nr_entries) | |
30 | { | |
31 | size_t s = sizeof(unsigned long) * dm_div_up(nr_entries, BITS_PER_LONG); | |
32 | return vzalloc(s); | |
33 | } | |
34 | ||
35 | static void free_bitset(unsigned long *bits) | |
36 | { | |
37 | vfree(bits); | |
38 | } | |
39 | ||
40 | /*----------------------------------------------------------------*/ | |
41 | ||
42 | /* | |
43 | * Large, sequential ios are probably better left on the origin device since | |
44 | * spindles tend to have good bandwidth. | |
45 | * | |
46 | * The io_tracker tries to spot when the io is in one of these sequential | |
47 | * modes. | |
48 | * | |
49 | * Two thresholds to switch between random and sequential io mode are defaulting | |
50 | * as follows and can be adjusted via the constructor and message interfaces. | |
51 | */ | |
52 | #define RANDOM_THRESHOLD_DEFAULT 4 | |
53 | #define SEQUENTIAL_THRESHOLD_DEFAULT 512 | |
54 | ||
55 | enum io_pattern { | |
56 | PATTERN_SEQUENTIAL, | |
57 | PATTERN_RANDOM | |
58 | }; | |
59 | ||
60 | struct io_tracker { | |
61 | enum io_pattern pattern; | |
62 | ||
63 | unsigned nr_seq_samples; | |
64 | unsigned nr_rand_samples; | |
65 | unsigned thresholds[2]; | |
66 | ||
67 | dm_oblock_t last_end_oblock; | |
68 | }; | |
69 | ||
70 | static void iot_init(struct io_tracker *t, | |
71 | int sequential_threshold, int random_threshold) | |
72 | { | |
73 | t->pattern = PATTERN_RANDOM; | |
74 | t->nr_seq_samples = 0; | |
75 | t->nr_rand_samples = 0; | |
76 | t->last_end_oblock = 0; | |
77 | t->thresholds[PATTERN_RANDOM] = random_threshold; | |
78 | t->thresholds[PATTERN_SEQUENTIAL] = sequential_threshold; | |
79 | } | |
80 | ||
81 | static enum io_pattern iot_pattern(struct io_tracker *t) | |
82 | { | |
83 | return t->pattern; | |
84 | } | |
85 | ||
86 | static void iot_update_stats(struct io_tracker *t, struct bio *bio) | |
87 | { | |
88 | if (bio->bi_sector == from_oblock(t->last_end_oblock) + 1) | |
89 | t->nr_seq_samples++; | |
90 | else { | |
91 | /* | |
92 | * Just one non-sequential IO is enough to reset the | |
93 | * counters. | |
94 | */ | |
95 | if (t->nr_seq_samples) { | |
96 | t->nr_seq_samples = 0; | |
97 | t->nr_rand_samples = 0; | |
98 | } | |
99 | ||
100 | t->nr_rand_samples++; | |
101 | } | |
102 | ||
103 | t->last_end_oblock = to_oblock(bio->bi_sector + bio_sectors(bio) - 1); | |
104 | } | |
105 | ||
106 | static void iot_check_for_pattern_switch(struct io_tracker *t) | |
107 | { | |
108 | switch (t->pattern) { | |
109 | case PATTERN_SEQUENTIAL: | |
110 | if (t->nr_rand_samples >= t->thresholds[PATTERN_RANDOM]) { | |
111 | t->pattern = PATTERN_RANDOM; | |
112 | t->nr_seq_samples = t->nr_rand_samples = 0; | |
113 | } | |
114 | break; | |
115 | ||
116 | case PATTERN_RANDOM: | |
117 | if (t->nr_seq_samples >= t->thresholds[PATTERN_SEQUENTIAL]) { | |
118 | t->pattern = PATTERN_SEQUENTIAL; | |
119 | t->nr_seq_samples = t->nr_rand_samples = 0; | |
120 | } | |
121 | break; | |
122 | } | |
123 | } | |
124 | ||
125 | static void iot_examine_bio(struct io_tracker *t, struct bio *bio) | |
126 | { | |
127 | iot_update_stats(t, bio); | |
128 | iot_check_for_pattern_switch(t); | |
129 | } | |
130 | ||
131 | /*----------------------------------------------------------------*/ | |
132 | ||
133 | ||
134 | /* | |
135 | * This queue is divided up into different levels. Allowing us to push | |
136 | * entries to the back of any of the levels. Think of it as a partially | |
137 | * sorted queue. | |
138 | */ | |
139 | #define NR_QUEUE_LEVELS 16u | |
140 | ||
141 | struct queue { | |
142 | struct list_head qs[NR_QUEUE_LEVELS]; | |
143 | }; | |
144 | ||
145 | static void queue_init(struct queue *q) | |
146 | { | |
147 | unsigned i; | |
148 | ||
149 | for (i = 0; i < NR_QUEUE_LEVELS; i++) | |
150 | INIT_LIST_HEAD(q->qs + i); | |
151 | } | |
152 | ||
153 | /* | |
154 | * Insert an entry to the back of the given level. | |
155 | */ | |
156 | static void queue_push(struct queue *q, unsigned level, struct list_head *elt) | |
157 | { | |
158 | list_add_tail(elt, q->qs + level); | |
159 | } | |
160 | ||
161 | static void queue_remove(struct list_head *elt) | |
162 | { | |
163 | list_del(elt); | |
164 | } | |
165 | ||
166 | /* | |
167 | * Shifts all regions down one level. This has no effect on the order of | |
168 | * the queue. | |
169 | */ | |
170 | static void queue_shift_down(struct queue *q) | |
171 | { | |
172 | unsigned level; | |
173 | ||
174 | for (level = 1; level < NR_QUEUE_LEVELS; level++) | |
175 | list_splice_init(q->qs + level, q->qs + level - 1); | |
176 | } | |
177 | ||
178 | /* | |
179 | * Gives us the oldest entry of the lowest popoulated level. If the first | |
180 | * level is emptied then we shift down one level. | |
181 | */ | |
182 | static struct list_head *queue_pop(struct queue *q) | |
183 | { | |
184 | unsigned level; | |
185 | struct list_head *r; | |
186 | ||
187 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
188 | if (!list_empty(q->qs + level)) { | |
189 | r = q->qs[level].next; | |
190 | list_del(r); | |
191 | ||
192 | /* have we just emptied the bottom level? */ | |
193 | if (level == 0 && list_empty(q->qs)) | |
194 | queue_shift_down(q); | |
195 | ||
196 | return r; | |
197 | } | |
198 | ||
199 | return NULL; | |
200 | } | |
201 | ||
202 | static struct list_head *list_pop(struct list_head *lh) | |
203 | { | |
204 | struct list_head *r = lh->next; | |
205 | ||
206 | BUG_ON(!r); | |
207 | list_del_init(r); | |
208 | ||
209 | return r; | |
210 | } | |
211 | ||
212 | /*----------------------------------------------------------------*/ | |
213 | ||
214 | /* | |
215 | * Describes a cache entry. Used in both the cache and the pre_cache. | |
216 | */ | |
217 | struct entry { | |
218 | struct hlist_node hlist; | |
219 | struct list_head list; | |
220 | dm_oblock_t oblock; | |
221 | dm_cblock_t cblock; /* valid iff in_cache */ | |
222 | ||
223 | /* | |
224 | * FIXME: pack these better | |
225 | */ | |
226 | bool in_cache:1; | |
227 | unsigned hit_count; | |
228 | unsigned generation; | |
229 | unsigned tick; | |
230 | }; | |
231 | ||
232 | struct mq_policy { | |
233 | struct dm_cache_policy policy; | |
234 | ||
235 | /* protects everything */ | |
236 | struct mutex lock; | |
237 | dm_cblock_t cache_size; | |
238 | struct io_tracker tracker; | |
239 | ||
240 | /* | |
241 | * We maintain two queues of entries. The cache proper contains | |
242 | * the currently active mappings. Whereas the pre_cache tracks | |
243 | * blocks that are being hit frequently and potential candidates | |
244 | * for promotion to the cache. | |
245 | */ | |
246 | struct queue pre_cache; | |
247 | struct queue cache; | |
248 | ||
249 | /* | |
250 | * Keeps track of time, incremented by the core. We use this to | |
251 | * avoid attributing multiple hits within the same tick. | |
252 | * | |
253 | * Access to tick_protected should be done with the spin lock held. | |
254 | * It's copied to tick at the start of the map function (within the | |
255 | * mutex). | |
256 | */ | |
257 | spinlock_t tick_lock; | |
258 | unsigned tick_protected; | |
259 | unsigned tick; | |
260 | ||
261 | /* | |
262 | * A count of the number of times the map function has been called | |
263 | * and found an entry in the pre_cache or cache. Currently used to | |
264 | * calculate the generation. | |
265 | */ | |
266 | unsigned hit_count; | |
267 | ||
268 | /* | |
269 | * A generation is a longish period that is used to trigger some | |
270 | * book keeping effects. eg, decrementing hit counts on entries. | |
271 | * This is needed to allow the cache to evolve as io patterns | |
272 | * change. | |
273 | */ | |
274 | unsigned generation; | |
275 | unsigned generation_period; /* in lookups (will probably change) */ | |
276 | ||
277 | /* | |
278 | * Entries in the pre_cache whose hit count passes the promotion | |
279 | * threshold move to the cache proper. Working out the correct | |
280 | * value for the promotion_threshold is crucial to this policy. | |
281 | */ | |
282 | unsigned promote_threshold; | |
283 | ||
284 | /* | |
285 | * We need cache_size entries for the cache, and choose to have | |
286 | * cache_size entries for the pre_cache too. One motivation for | |
287 | * using the same size is to make the hit counts directly | |
288 | * comparable between pre_cache and cache. | |
289 | */ | |
290 | unsigned nr_entries; | |
291 | unsigned nr_entries_allocated; | |
292 | struct list_head free; | |
293 | ||
294 | /* | |
295 | * Cache blocks may be unallocated. We store this info in a | |
296 | * bitset. | |
297 | */ | |
298 | unsigned long *allocation_bitset; | |
299 | unsigned nr_cblocks_allocated; | |
300 | unsigned find_free_nr_words; | |
301 | unsigned find_free_last_word; | |
302 | ||
303 | /* | |
304 | * The hash table allows us to quickly find an entry by origin | |
305 | * block. Both pre_cache and cache entries are in here. | |
306 | */ | |
307 | unsigned nr_buckets; | |
308 | dm_block_t hash_bits; | |
309 | struct hlist_head *table; | |
310 | }; | |
311 | ||
312 | /*----------------------------------------------------------------*/ | |
313 | /* Free/alloc mq cache entry structures. */ | |
314 | static void takeout_queue(struct list_head *lh, struct queue *q) | |
315 | { | |
316 | unsigned level; | |
317 | ||
318 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
319 | list_splice(q->qs + level, lh); | |
320 | } | |
321 | ||
322 | static void free_entries(struct mq_policy *mq) | |
323 | { | |
324 | struct entry *e, *tmp; | |
325 | ||
326 | takeout_queue(&mq->free, &mq->pre_cache); | |
327 | takeout_queue(&mq->free, &mq->cache); | |
328 | ||
329 | list_for_each_entry_safe(e, tmp, &mq->free, list) | |
330 | kmem_cache_free(mq_entry_cache, e); | |
331 | } | |
332 | ||
333 | static int alloc_entries(struct mq_policy *mq, unsigned elts) | |
334 | { | |
335 | unsigned u = mq->nr_entries; | |
336 | ||
337 | INIT_LIST_HEAD(&mq->free); | |
338 | mq->nr_entries_allocated = 0; | |
339 | ||
340 | while (u--) { | |
341 | struct entry *e = kmem_cache_zalloc(mq_entry_cache, GFP_KERNEL); | |
342 | ||
343 | if (!e) { | |
344 | free_entries(mq); | |
345 | return -ENOMEM; | |
346 | } | |
347 | ||
348 | ||
349 | list_add(&e->list, &mq->free); | |
350 | } | |
351 | ||
352 | return 0; | |
353 | } | |
354 | ||
355 | /*----------------------------------------------------------------*/ | |
356 | ||
357 | /* | |
358 | * Simple hash table implementation. Should replace with the standard hash | |
359 | * table that's making its way upstream. | |
360 | */ | |
361 | static void hash_insert(struct mq_policy *mq, struct entry *e) | |
362 | { | |
363 | unsigned h = hash_64(from_oblock(e->oblock), mq->hash_bits); | |
364 | ||
365 | hlist_add_head(&e->hlist, mq->table + h); | |
366 | } | |
367 | ||
368 | static struct entry *hash_lookup(struct mq_policy *mq, dm_oblock_t oblock) | |
369 | { | |
370 | unsigned h = hash_64(from_oblock(oblock), mq->hash_bits); | |
371 | struct hlist_head *bucket = mq->table + h; | |
372 | struct entry *e; | |
373 | ||
374 | hlist_for_each_entry(e, bucket, hlist) | |
375 | if (e->oblock == oblock) { | |
376 | hlist_del(&e->hlist); | |
377 | hlist_add_head(&e->hlist, bucket); | |
378 | return e; | |
379 | } | |
380 | ||
381 | return NULL; | |
382 | } | |
383 | ||
384 | static void hash_remove(struct entry *e) | |
385 | { | |
386 | hlist_del(&e->hlist); | |
387 | } | |
388 | ||
389 | /*----------------------------------------------------------------*/ | |
390 | ||
391 | /* | |
392 | * Allocates a new entry structure. The memory is allocated in one lump, | |
393 | * so we just handing it out here. Returns NULL if all entries have | |
394 | * already been allocated. Cannot fail otherwise. | |
395 | */ | |
396 | static struct entry *alloc_entry(struct mq_policy *mq) | |
397 | { | |
398 | struct entry *e; | |
399 | ||
400 | if (mq->nr_entries_allocated >= mq->nr_entries) { | |
401 | BUG_ON(!list_empty(&mq->free)); | |
402 | return NULL; | |
403 | } | |
404 | ||
405 | e = list_entry(list_pop(&mq->free), struct entry, list); | |
406 | INIT_LIST_HEAD(&e->list); | |
407 | INIT_HLIST_NODE(&e->hlist); | |
408 | ||
409 | mq->nr_entries_allocated++; | |
410 | return e; | |
411 | } | |
412 | ||
413 | /*----------------------------------------------------------------*/ | |
414 | ||
415 | /* | |
416 | * Mark cache blocks allocated or not in the bitset. | |
417 | */ | |
418 | static void alloc_cblock(struct mq_policy *mq, dm_cblock_t cblock) | |
419 | { | |
420 | BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size)); | |
421 | BUG_ON(test_bit(from_cblock(cblock), mq->allocation_bitset)); | |
422 | ||
423 | set_bit(from_cblock(cblock), mq->allocation_bitset); | |
424 | mq->nr_cblocks_allocated++; | |
425 | } | |
426 | ||
427 | static void free_cblock(struct mq_policy *mq, dm_cblock_t cblock) | |
428 | { | |
429 | BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size)); | |
430 | BUG_ON(!test_bit(from_cblock(cblock), mq->allocation_bitset)); | |
431 | ||
432 | clear_bit(from_cblock(cblock), mq->allocation_bitset); | |
433 | mq->nr_cblocks_allocated--; | |
434 | } | |
435 | ||
436 | static bool any_free_cblocks(struct mq_policy *mq) | |
437 | { | |
438 | return mq->nr_cblocks_allocated < from_cblock(mq->cache_size); | |
439 | } | |
440 | ||
441 | /* | |
442 | * Fills result out with a cache block that isn't in use, or return | |
443 | * -ENOSPC. This does _not_ mark the cblock as allocated, the caller is | |
444 | * reponsible for that. | |
445 | */ | |
446 | static int __find_free_cblock(struct mq_policy *mq, unsigned begin, unsigned end, | |
447 | dm_cblock_t *result, unsigned *last_word) | |
448 | { | |
449 | int r = -ENOSPC; | |
450 | unsigned w; | |
451 | ||
452 | for (w = begin; w < end; w++) { | |
453 | /* | |
454 | * ffz is undefined if no zero exists | |
455 | */ | |
456 | if (mq->allocation_bitset[w] != ~0UL) { | |
457 | *last_word = w; | |
458 | *result = to_cblock((w * BITS_PER_LONG) + ffz(mq->allocation_bitset[w])); | |
459 | if (from_cblock(*result) < from_cblock(mq->cache_size)) | |
460 | r = 0; | |
461 | ||
462 | break; | |
463 | } | |
464 | } | |
465 | ||
466 | return r; | |
467 | } | |
468 | ||
469 | static int find_free_cblock(struct mq_policy *mq, dm_cblock_t *result) | |
470 | { | |
471 | int r; | |
472 | ||
473 | if (!any_free_cblocks(mq)) | |
474 | return -ENOSPC; | |
475 | ||
476 | r = __find_free_cblock(mq, mq->find_free_last_word, mq->find_free_nr_words, result, &mq->find_free_last_word); | |
477 | if (r == -ENOSPC && mq->find_free_last_word) | |
478 | r = __find_free_cblock(mq, 0, mq->find_free_last_word, result, &mq->find_free_last_word); | |
479 | ||
480 | return r; | |
481 | } | |
482 | ||
483 | /*----------------------------------------------------------------*/ | |
484 | ||
485 | /* | |
486 | * Now we get to the meat of the policy. This section deals with deciding | |
487 | * when to to add entries to the pre_cache and cache, and move between | |
488 | * them. | |
489 | */ | |
490 | ||
491 | /* | |
492 | * The queue level is based on the log2 of the hit count. | |
493 | */ | |
494 | static unsigned queue_level(struct entry *e) | |
495 | { | |
496 | return min((unsigned) ilog2(e->hit_count), NR_QUEUE_LEVELS - 1u); | |
497 | } | |
498 | ||
499 | /* | |
500 | * Inserts the entry into the pre_cache or the cache. Ensures the cache | |
501 | * block is marked as allocated if necc. Inserts into the hash table. Sets the | |
502 | * tick which records when the entry was last moved about. | |
503 | */ | |
504 | static void push(struct mq_policy *mq, struct entry *e) | |
505 | { | |
506 | e->tick = mq->tick; | |
507 | hash_insert(mq, e); | |
508 | ||
509 | if (e->in_cache) { | |
510 | alloc_cblock(mq, e->cblock); | |
511 | queue_push(&mq->cache, queue_level(e), &e->list); | |
512 | } else | |
513 | queue_push(&mq->pre_cache, queue_level(e), &e->list); | |
514 | } | |
515 | ||
516 | /* | |
517 | * Removes an entry from pre_cache or cache. Removes from the hash table. | |
518 | * Frees off the cache block if necc. | |
519 | */ | |
520 | static void del(struct mq_policy *mq, struct entry *e) | |
521 | { | |
522 | queue_remove(&e->list); | |
523 | hash_remove(e); | |
524 | if (e->in_cache) | |
525 | free_cblock(mq, e->cblock); | |
526 | } | |
527 | ||
528 | /* | |
529 | * Like del, except it removes the first entry in the queue (ie. the least | |
530 | * recently used). | |
531 | */ | |
532 | static struct entry *pop(struct mq_policy *mq, struct queue *q) | |
533 | { | |
534 | struct entry *e = container_of(queue_pop(q), struct entry, list); | |
535 | ||
536 | if (e) { | |
537 | hash_remove(e); | |
538 | ||
539 | if (e->in_cache) | |
540 | free_cblock(mq, e->cblock); | |
541 | } | |
542 | ||
543 | return e; | |
544 | } | |
545 | ||
546 | /* | |
547 | * Has this entry already been updated? | |
548 | */ | |
549 | static bool updated_this_tick(struct mq_policy *mq, struct entry *e) | |
550 | { | |
551 | return mq->tick == e->tick; | |
552 | } | |
553 | ||
554 | /* | |
555 | * The promotion threshold is adjusted every generation. As are the counts | |
556 | * of the entries. | |
557 | * | |
558 | * At the moment the threshold is taken by averaging the hit counts of some | |
559 | * of the entries in the cache (the first 20 entries of the first level). | |
560 | * | |
561 | * We can be much cleverer than this though. For example, each promotion | |
562 | * could bump up the threshold helping to prevent churn. Much more to do | |
563 | * here. | |
564 | */ | |
565 | ||
566 | #define MAX_TO_AVERAGE 20 | |
567 | ||
568 | static void check_generation(struct mq_policy *mq) | |
569 | { | |
570 | unsigned total = 0, nr = 0, count = 0, level; | |
571 | struct list_head *head; | |
572 | struct entry *e; | |
573 | ||
574 | if ((mq->hit_count >= mq->generation_period) && | |
575 | (mq->nr_cblocks_allocated == from_cblock(mq->cache_size))) { | |
576 | ||
577 | mq->hit_count = 0; | |
578 | mq->generation++; | |
579 | ||
580 | for (level = 0; level < NR_QUEUE_LEVELS && count < MAX_TO_AVERAGE; level++) { | |
581 | head = mq->cache.qs + level; | |
582 | list_for_each_entry(e, head, list) { | |
583 | nr++; | |
584 | total += e->hit_count; | |
585 | ||
586 | if (++count >= MAX_TO_AVERAGE) | |
587 | break; | |
588 | } | |
589 | } | |
590 | ||
591 | mq->promote_threshold = nr ? total / nr : 1; | |
592 | if (mq->promote_threshold * nr < total) | |
593 | mq->promote_threshold++; | |
594 | } | |
595 | } | |
596 | ||
597 | /* | |
598 | * Whenever we use an entry we bump up it's hit counter, and push it to the | |
599 | * back to it's current level. | |
600 | */ | |
601 | static void requeue_and_update_tick(struct mq_policy *mq, struct entry *e) | |
602 | { | |
603 | if (updated_this_tick(mq, e)) | |
604 | return; | |
605 | ||
606 | e->hit_count++; | |
607 | mq->hit_count++; | |
608 | check_generation(mq); | |
609 | ||
610 | /* generation adjustment, to stop the counts increasing forever. */ | |
611 | /* FIXME: divide? */ | |
612 | /* e->hit_count -= min(e->hit_count - 1, mq->generation - e->generation); */ | |
613 | e->generation = mq->generation; | |
614 | ||
615 | del(mq, e); | |
616 | push(mq, e); | |
617 | } | |
618 | ||
619 | /* | |
620 | * Demote the least recently used entry from the cache to the pre_cache. | |
621 | * Returns the new cache entry to use, and the old origin block it was | |
622 | * mapped to. | |
623 | * | |
624 | * We drop the hit count on the demoted entry back to 1 to stop it bouncing | |
625 | * straight back into the cache if it's subsequently hit. There are | |
626 | * various options here, and more experimentation would be good: | |
627 | * | |
628 | * - just forget about the demoted entry completely (ie. don't insert it | |
629 | into the pre_cache). | |
630 | * - divide the hit count rather that setting to some hard coded value. | |
631 | * - set the hit count to a hard coded value other than 1, eg, is it better | |
632 | * if it goes in at level 2? | |
633 | */ | |
634 | static dm_cblock_t demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock) | |
635 | { | |
636 | dm_cblock_t result; | |
637 | struct entry *demoted = pop(mq, &mq->cache); | |
638 | ||
639 | BUG_ON(!demoted); | |
640 | result = demoted->cblock; | |
641 | *oblock = demoted->oblock; | |
642 | demoted->in_cache = false; | |
643 | demoted->hit_count = 1; | |
644 | push(mq, demoted); | |
645 | ||
646 | return result; | |
647 | } | |
648 | ||
649 | /* | |
650 | * We modify the basic promotion_threshold depending on the specific io. | |
651 | * | |
652 | * If the origin block has been discarded then there's no cost to copy it | |
653 | * to the cache. | |
654 | * | |
655 | * We bias towards reads, since they can be demoted at no cost if they | |
656 | * haven't been dirtied. | |
657 | */ | |
658 | #define DISCARDED_PROMOTE_THRESHOLD 1 | |
659 | #define READ_PROMOTE_THRESHOLD 4 | |
660 | #define WRITE_PROMOTE_THRESHOLD 8 | |
661 | ||
662 | static unsigned adjusted_promote_threshold(struct mq_policy *mq, | |
663 | bool discarded_oblock, int data_dir) | |
664 | { | |
665 | if (discarded_oblock && any_free_cblocks(mq) && data_dir == WRITE) | |
666 | /* | |
667 | * We don't need to do any copying at all, so give this a | |
668 | * very low threshold. In practice this only triggers | |
669 | * during initial population after a format. | |
670 | */ | |
671 | return DISCARDED_PROMOTE_THRESHOLD; | |
672 | ||
673 | return data_dir == READ ? | |
674 | (mq->promote_threshold + READ_PROMOTE_THRESHOLD) : | |
675 | (mq->promote_threshold + WRITE_PROMOTE_THRESHOLD); | |
676 | } | |
677 | ||
678 | static bool should_promote(struct mq_policy *mq, struct entry *e, | |
679 | bool discarded_oblock, int data_dir) | |
680 | { | |
681 | return e->hit_count >= | |
682 | adjusted_promote_threshold(mq, discarded_oblock, data_dir); | |
683 | } | |
684 | ||
685 | static int cache_entry_found(struct mq_policy *mq, | |
686 | struct entry *e, | |
687 | struct policy_result *result) | |
688 | { | |
689 | requeue_and_update_tick(mq, e); | |
690 | ||
691 | if (e->in_cache) { | |
692 | result->op = POLICY_HIT; | |
693 | result->cblock = e->cblock; | |
694 | } | |
695 | ||
696 | return 0; | |
697 | } | |
698 | ||
699 | /* | |
700 | * Moves and entry from the pre_cache to the cache. The main work is | |
701 | * finding which cache block to use. | |
702 | */ | |
703 | static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e, | |
704 | struct policy_result *result) | |
705 | { | |
706 | dm_cblock_t cblock; | |
707 | ||
708 | if (find_free_cblock(mq, &cblock) == -ENOSPC) { | |
709 | result->op = POLICY_REPLACE; | |
710 | cblock = demote_cblock(mq, &result->old_oblock); | |
711 | } else | |
712 | result->op = POLICY_NEW; | |
713 | ||
714 | result->cblock = e->cblock = cblock; | |
715 | ||
716 | del(mq, e); | |
717 | e->in_cache = true; | |
718 | push(mq, e); | |
719 | ||
720 | return 0; | |
721 | } | |
722 | ||
723 | static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e, | |
724 | bool can_migrate, bool discarded_oblock, | |
725 | int data_dir, struct policy_result *result) | |
726 | { | |
727 | int r = 0; | |
728 | bool updated = updated_this_tick(mq, e); | |
729 | ||
730 | requeue_and_update_tick(mq, e); | |
731 | ||
732 | if ((!discarded_oblock && updated) || | |
733 | !should_promote(mq, e, discarded_oblock, data_dir)) | |
734 | result->op = POLICY_MISS; | |
735 | else if (!can_migrate) | |
736 | r = -EWOULDBLOCK; | |
737 | else | |
738 | r = pre_cache_to_cache(mq, e, result); | |
739 | ||
740 | return r; | |
741 | } | |
742 | ||
743 | static void insert_in_pre_cache(struct mq_policy *mq, | |
744 | dm_oblock_t oblock) | |
745 | { | |
746 | struct entry *e = alloc_entry(mq); | |
747 | ||
748 | if (!e) | |
749 | /* | |
750 | * There's no spare entry structure, so we grab the least | |
751 | * used one from the pre_cache. | |
752 | */ | |
753 | e = pop(mq, &mq->pre_cache); | |
754 | ||
755 | if (unlikely(!e)) { | |
756 | DMWARN("couldn't pop from pre cache"); | |
757 | return; | |
758 | } | |
759 | ||
760 | e->in_cache = false; | |
761 | e->oblock = oblock; | |
762 | e->hit_count = 1; | |
763 | e->generation = mq->generation; | |
764 | push(mq, e); | |
765 | } | |
766 | ||
767 | static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock, | |
768 | struct policy_result *result) | |
769 | { | |
770 | struct entry *e; | |
771 | dm_cblock_t cblock; | |
772 | ||
773 | if (find_free_cblock(mq, &cblock) == -ENOSPC) { | |
774 | result->op = POLICY_MISS; | |
775 | insert_in_pre_cache(mq, oblock); | |
776 | return; | |
777 | } | |
778 | ||
779 | e = alloc_entry(mq); | |
780 | if (unlikely(!e)) { | |
781 | result->op = POLICY_MISS; | |
782 | return; | |
783 | } | |
784 | ||
785 | e->oblock = oblock; | |
786 | e->cblock = cblock; | |
787 | e->in_cache = true; | |
788 | e->hit_count = 1; | |
789 | e->generation = mq->generation; | |
790 | push(mq, e); | |
791 | ||
792 | result->op = POLICY_NEW; | |
793 | result->cblock = e->cblock; | |
794 | } | |
795 | ||
796 | static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock, | |
797 | bool can_migrate, bool discarded_oblock, | |
798 | int data_dir, struct policy_result *result) | |
799 | { | |
800 | if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) == 1) { | |
801 | if (can_migrate) | |
802 | insert_in_cache(mq, oblock, result); | |
803 | else | |
804 | return -EWOULDBLOCK; | |
805 | } else { | |
806 | insert_in_pre_cache(mq, oblock); | |
807 | result->op = POLICY_MISS; | |
808 | } | |
809 | ||
810 | return 0; | |
811 | } | |
812 | ||
813 | /* | |
814 | * Looks the oblock up in the hash table, then decides whether to put in | |
815 | * pre_cache, or cache etc. | |
816 | */ | |
817 | static int map(struct mq_policy *mq, dm_oblock_t oblock, | |
818 | bool can_migrate, bool discarded_oblock, | |
819 | int data_dir, struct policy_result *result) | |
820 | { | |
821 | int r = 0; | |
822 | struct entry *e = hash_lookup(mq, oblock); | |
823 | ||
824 | if (e && e->in_cache) | |
825 | r = cache_entry_found(mq, e, result); | |
826 | else if (iot_pattern(&mq->tracker) == PATTERN_SEQUENTIAL) | |
827 | result->op = POLICY_MISS; | |
828 | else if (e) | |
829 | r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock, | |
830 | data_dir, result); | |
831 | else | |
832 | r = no_entry_found(mq, oblock, can_migrate, discarded_oblock, | |
833 | data_dir, result); | |
834 | ||
835 | if (r == -EWOULDBLOCK) | |
836 | result->op = POLICY_MISS; | |
837 | ||
838 | return r; | |
839 | } | |
840 | ||
841 | /*----------------------------------------------------------------*/ | |
842 | ||
843 | /* | |
844 | * Public interface, via the policy struct. See dm-cache-policy.h for a | |
845 | * description of these. | |
846 | */ | |
847 | ||
848 | static struct mq_policy *to_mq_policy(struct dm_cache_policy *p) | |
849 | { | |
850 | return container_of(p, struct mq_policy, policy); | |
851 | } | |
852 | ||
853 | static void mq_destroy(struct dm_cache_policy *p) | |
854 | { | |
855 | struct mq_policy *mq = to_mq_policy(p); | |
856 | ||
857 | free_bitset(mq->allocation_bitset); | |
858 | kfree(mq->table); | |
859 | free_entries(mq); | |
860 | kfree(mq); | |
861 | } | |
862 | ||
863 | static void copy_tick(struct mq_policy *mq) | |
864 | { | |
865 | unsigned long flags; | |
866 | ||
867 | spin_lock_irqsave(&mq->tick_lock, flags); | |
868 | mq->tick = mq->tick_protected; | |
869 | spin_unlock_irqrestore(&mq->tick_lock, flags); | |
870 | } | |
871 | ||
872 | static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock, | |
873 | bool can_block, bool can_migrate, bool discarded_oblock, | |
874 | struct bio *bio, struct policy_result *result) | |
875 | { | |
876 | int r; | |
877 | struct mq_policy *mq = to_mq_policy(p); | |
878 | ||
879 | result->op = POLICY_MISS; | |
880 | ||
881 | if (can_block) | |
882 | mutex_lock(&mq->lock); | |
883 | else if (!mutex_trylock(&mq->lock)) | |
884 | return -EWOULDBLOCK; | |
885 | ||
886 | copy_tick(mq); | |
887 | ||
888 | iot_examine_bio(&mq->tracker, bio); | |
889 | r = map(mq, oblock, can_migrate, discarded_oblock, | |
890 | bio_data_dir(bio), result); | |
891 | ||
892 | mutex_unlock(&mq->lock); | |
893 | ||
894 | return r; | |
895 | } | |
896 | ||
897 | static int mq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock) | |
898 | { | |
899 | int r; | |
900 | struct mq_policy *mq = to_mq_policy(p); | |
901 | struct entry *e; | |
902 | ||
903 | if (!mutex_trylock(&mq->lock)) | |
904 | return -EWOULDBLOCK; | |
905 | ||
906 | e = hash_lookup(mq, oblock); | |
907 | if (e && e->in_cache) { | |
908 | *cblock = e->cblock; | |
909 | r = 0; | |
910 | } else | |
911 | r = -ENOENT; | |
912 | ||
913 | mutex_unlock(&mq->lock); | |
914 | ||
915 | return r; | |
916 | } | |
917 | ||
918 | static int mq_load_mapping(struct dm_cache_policy *p, | |
919 | dm_oblock_t oblock, dm_cblock_t cblock, | |
920 | uint32_t hint, bool hint_valid) | |
921 | { | |
922 | struct mq_policy *mq = to_mq_policy(p); | |
923 | struct entry *e; | |
924 | ||
925 | e = alloc_entry(mq); | |
926 | if (!e) | |
927 | return -ENOMEM; | |
928 | ||
929 | e->cblock = cblock; | |
930 | e->oblock = oblock; | |
931 | e->in_cache = true; | |
932 | e->hit_count = hint_valid ? hint : 1; | |
933 | e->generation = mq->generation; | |
934 | push(mq, e); | |
935 | ||
936 | return 0; | |
937 | } | |
938 | ||
939 | static int mq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn, | |
940 | void *context) | |
941 | { | |
942 | struct mq_policy *mq = to_mq_policy(p); | |
943 | int r = 0; | |
944 | struct entry *e; | |
945 | unsigned level; | |
946 | ||
947 | mutex_lock(&mq->lock); | |
948 | ||
949 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
950 | list_for_each_entry(e, &mq->cache.qs[level], list) { | |
951 | r = fn(context, e->cblock, e->oblock, e->hit_count); | |
952 | if (r) | |
953 | goto out; | |
954 | } | |
955 | ||
956 | out: | |
957 | mutex_unlock(&mq->lock); | |
958 | ||
959 | return r; | |
960 | } | |
961 | ||
b936bf8b | 962 | static void mq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock) |
f2836352 | 963 | { |
b936bf8b GU |
964 | struct mq_policy *mq = to_mq_policy(p); |
965 | struct entry *e; | |
966 | ||
967 | mutex_lock(&mq->lock); | |
968 | ||
969 | e = hash_lookup(mq, oblock); | |
f2836352 JT |
970 | |
971 | BUG_ON(!e || !e->in_cache); | |
972 | ||
973 | del(mq, e); | |
974 | e->in_cache = false; | |
975 | push(mq, e); | |
f2836352 | 976 | |
f2836352 JT |
977 | mutex_unlock(&mq->lock); |
978 | } | |
979 | ||
980 | static void force_mapping(struct mq_policy *mq, | |
981 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) | |
982 | { | |
983 | struct entry *e = hash_lookup(mq, current_oblock); | |
984 | ||
985 | BUG_ON(!e || !e->in_cache); | |
986 | ||
987 | del(mq, e); | |
988 | e->oblock = new_oblock; | |
989 | push(mq, e); | |
990 | } | |
991 | ||
992 | static void mq_force_mapping(struct dm_cache_policy *p, | |
993 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) | |
994 | { | |
995 | struct mq_policy *mq = to_mq_policy(p); | |
996 | ||
997 | mutex_lock(&mq->lock); | |
998 | force_mapping(mq, current_oblock, new_oblock); | |
999 | mutex_unlock(&mq->lock); | |
1000 | } | |
1001 | ||
1002 | static dm_cblock_t mq_residency(struct dm_cache_policy *p) | |
1003 | { | |
1004 | struct mq_policy *mq = to_mq_policy(p); | |
1005 | ||
1006 | /* FIXME: lock mutex, not sure we can block here */ | |
1007 | return to_cblock(mq->nr_cblocks_allocated); | |
1008 | } | |
1009 | ||
1010 | static void mq_tick(struct dm_cache_policy *p) | |
1011 | { | |
1012 | struct mq_policy *mq = to_mq_policy(p); | |
1013 | unsigned long flags; | |
1014 | ||
1015 | spin_lock_irqsave(&mq->tick_lock, flags); | |
1016 | mq->tick_protected++; | |
1017 | spin_unlock_irqrestore(&mq->tick_lock, flags); | |
1018 | } | |
1019 | ||
1020 | static int mq_set_config_value(struct dm_cache_policy *p, | |
1021 | const char *key, const char *value) | |
1022 | { | |
1023 | struct mq_policy *mq = to_mq_policy(p); | |
1024 | enum io_pattern pattern; | |
1025 | unsigned long tmp; | |
1026 | ||
1027 | if (!strcasecmp(key, "random_threshold")) | |
1028 | pattern = PATTERN_RANDOM; | |
1029 | else if (!strcasecmp(key, "sequential_threshold")) | |
1030 | pattern = PATTERN_SEQUENTIAL; | |
1031 | else | |
1032 | return -EINVAL; | |
1033 | ||
1034 | if (kstrtoul(value, 10, &tmp)) | |
1035 | return -EINVAL; | |
1036 | ||
1037 | mq->tracker.thresholds[pattern] = tmp; | |
1038 | ||
1039 | return 0; | |
1040 | } | |
1041 | ||
1042 | static int mq_emit_config_values(struct dm_cache_policy *p, char *result, unsigned maxlen) | |
1043 | { | |
1044 | ssize_t sz = 0; | |
1045 | struct mq_policy *mq = to_mq_policy(p); | |
1046 | ||
1047 | DMEMIT("4 random_threshold %u sequential_threshold %u", | |
1048 | mq->tracker.thresholds[PATTERN_RANDOM], | |
1049 | mq->tracker.thresholds[PATTERN_SEQUENTIAL]); | |
1050 | ||
1051 | return 0; | |
1052 | } | |
1053 | ||
1054 | /* Init the policy plugin interface function pointers. */ | |
1055 | static void init_policy_functions(struct mq_policy *mq) | |
1056 | { | |
1057 | mq->policy.destroy = mq_destroy; | |
1058 | mq->policy.map = mq_map; | |
1059 | mq->policy.lookup = mq_lookup; | |
1060 | mq->policy.load_mapping = mq_load_mapping; | |
1061 | mq->policy.walk_mappings = mq_walk_mappings; | |
1062 | mq->policy.remove_mapping = mq_remove_mapping; | |
1063 | mq->policy.writeback_work = NULL; | |
1064 | mq->policy.force_mapping = mq_force_mapping; | |
1065 | mq->policy.residency = mq_residency; | |
1066 | mq->policy.tick = mq_tick; | |
1067 | mq->policy.emit_config_values = mq_emit_config_values; | |
1068 | mq->policy.set_config_value = mq_set_config_value; | |
1069 | } | |
1070 | ||
1071 | static struct dm_cache_policy *mq_create(dm_cblock_t cache_size, | |
1072 | sector_t origin_size, | |
1073 | sector_t cache_block_size) | |
1074 | { | |
1075 | int r; | |
1076 | struct mq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL); | |
1077 | ||
1078 | if (!mq) | |
1079 | return NULL; | |
1080 | ||
1081 | init_policy_functions(mq); | |
1082 | iot_init(&mq->tracker, SEQUENTIAL_THRESHOLD_DEFAULT, RANDOM_THRESHOLD_DEFAULT); | |
1083 | ||
1084 | mq->cache_size = cache_size; | |
1085 | mq->tick_protected = 0; | |
1086 | mq->tick = 0; | |
1087 | mq->hit_count = 0; | |
1088 | mq->generation = 0; | |
1089 | mq->promote_threshold = 0; | |
1090 | mutex_init(&mq->lock); | |
1091 | spin_lock_init(&mq->tick_lock); | |
1092 | mq->find_free_nr_words = dm_div_up(from_cblock(mq->cache_size), BITS_PER_LONG); | |
1093 | mq->find_free_last_word = 0; | |
1094 | ||
1095 | queue_init(&mq->pre_cache); | |
1096 | queue_init(&mq->cache); | |
1097 | mq->generation_period = max((unsigned) from_cblock(cache_size), 1024U); | |
1098 | ||
1099 | mq->nr_entries = 2 * from_cblock(cache_size); | |
1100 | r = alloc_entries(mq, mq->nr_entries); | |
1101 | if (r) | |
1102 | goto bad_cache_alloc; | |
1103 | ||
1104 | mq->nr_entries_allocated = 0; | |
1105 | mq->nr_cblocks_allocated = 0; | |
1106 | ||
1107 | mq->nr_buckets = next_power(from_cblock(cache_size) / 2, 16); | |
1108 | mq->hash_bits = ffs(mq->nr_buckets) - 1; | |
1109 | mq->table = kzalloc(sizeof(*mq->table) * mq->nr_buckets, GFP_KERNEL); | |
1110 | if (!mq->table) | |
1111 | goto bad_alloc_table; | |
1112 | ||
1113 | mq->allocation_bitset = alloc_bitset(from_cblock(cache_size)); | |
1114 | if (!mq->allocation_bitset) | |
1115 | goto bad_alloc_bitset; | |
1116 | ||
1117 | return &mq->policy; | |
1118 | ||
1119 | bad_alloc_bitset: | |
1120 | kfree(mq->table); | |
1121 | bad_alloc_table: | |
1122 | free_entries(mq); | |
1123 | bad_cache_alloc: | |
1124 | kfree(mq); | |
1125 | ||
1126 | return NULL; | |
1127 | } | |
1128 | ||
1129 | /*----------------------------------------------------------------*/ | |
1130 | ||
1131 | static struct dm_cache_policy_type mq_policy_type = { | |
1132 | .name = "mq", | |
4e7f506f | 1133 | .version = {1, 0, 0}, |
f2836352 JT |
1134 | .hint_size = 4, |
1135 | .owner = THIS_MODULE, | |
1136 | .create = mq_create | |
1137 | }; | |
1138 | ||
1139 | static struct dm_cache_policy_type default_policy_type = { | |
1140 | .name = "default", | |
4e7f506f | 1141 | .version = {1, 0, 0}, |
f2836352 JT |
1142 | .hint_size = 4, |
1143 | .owner = THIS_MODULE, | |
1144 | .create = mq_create | |
1145 | }; | |
1146 | ||
1147 | static int __init mq_init(void) | |
1148 | { | |
1149 | int r; | |
1150 | ||
1151 | mq_entry_cache = kmem_cache_create("dm_mq_policy_cache_entry", | |
1152 | sizeof(struct entry), | |
1153 | __alignof__(struct entry), | |
1154 | 0, NULL); | |
1155 | if (!mq_entry_cache) | |
1156 | goto bad; | |
1157 | ||
1158 | r = dm_cache_policy_register(&mq_policy_type); | |
1159 | if (r) { | |
1160 | DMERR("register failed %d", r); | |
1161 | goto bad_register_mq; | |
1162 | } | |
1163 | ||
1164 | r = dm_cache_policy_register(&default_policy_type); | |
1165 | if (!r) { | |
4e7f506f MS |
1166 | DMINFO("version %u.%u.%u loaded", |
1167 | mq_policy_type.version[0], | |
1168 | mq_policy_type.version[1], | |
1169 | mq_policy_type.version[2]); | |
f2836352 JT |
1170 | return 0; |
1171 | } | |
1172 | ||
1173 | DMERR("register failed (as default) %d", r); | |
1174 | ||
1175 | dm_cache_policy_unregister(&mq_policy_type); | |
1176 | bad_register_mq: | |
1177 | kmem_cache_destroy(mq_entry_cache); | |
1178 | bad: | |
1179 | return -ENOMEM; | |
1180 | } | |
1181 | ||
1182 | static void __exit mq_exit(void) | |
1183 | { | |
1184 | dm_cache_policy_unregister(&mq_policy_type); | |
1185 | dm_cache_policy_unregister(&default_policy_type); | |
1186 | ||
1187 | kmem_cache_destroy(mq_entry_cache); | |
1188 | } | |
1189 | ||
1190 | module_init(mq_init); | |
1191 | module_exit(mq_exit); | |
1192 | ||
1193 | MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>"); | |
1194 | MODULE_LICENSE("GPL"); | |
1195 | MODULE_DESCRIPTION("mq cache policy"); | |
1196 | ||
1197 | MODULE_ALIAS("dm-cache-default"); |