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> | |
e65ff870 | 11 | #include <linux/jiffies.h> |
f2836352 JT |
12 | #include <linux/module.h> |
13 | #include <linux/mutex.h> | |
14 | #include <linux/slab.h> | |
15 | #include <linux/vmalloc.h> | |
16 | ||
17 | #define DM_MSG_PREFIX "cache-policy-mq" | |
f2836352 JT |
18 | |
19 | static struct kmem_cache *mq_entry_cache; | |
20 | ||
21 | /*----------------------------------------------------------------*/ | |
22 | ||
23 | static unsigned next_power(unsigned n, unsigned min) | |
24 | { | |
25 | return roundup_pow_of_two(max(n, min)); | |
26 | } | |
27 | ||
28 | /*----------------------------------------------------------------*/ | |
29 | ||
f2836352 JT |
30 | /* |
31 | * Large, sequential ios are probably better left on the origin device since | |
32 | * spindles tend to have good bandwidth. | |
33 | * | |
34 | * The io_tracker tries to spot when the io is in one of these sequential | |
35 | * modes. | |
36 | * | |
37 | * Two thresholds to switch between random and sequential io mode are defaulting | |
38 | * as follows and can be adjusted via the constructor and message interfaces. | |
39 | */ | |
40 | #define RANDOM_THRESHOLD_DEFAULT 4 | |
41 | #define SEQUENTIAL_THRESHOLD_DEFAULT 512 | |
42 | ||
43 | enum io_pattern { | |
44 | PATTERN_SEQUENTIAL, | |
45 | PATTERN_RANDOM | |
46 | }; | |
47 | ||
48 | struct io_tracker { | |
49 | enum io_pattern pattern; | |
50 | ||
51 | unsigned nr_seq_samples; | |
52 | unsigned nr_rand_samples; | |
53 | unsigned thresholds[2]; | |
54 | ||
55 | dm_oblock_t last_end_oblock; | |
56 | }; | |
57 | ||
58 | static void iot_init(struct io_tracker *t, | |
59 | int sequential_threshold, int random_threshold) | |
60 | { | |
61 | t->pattern = PATTERN_RANDOM; | |
62 | t->nr_seq_samples = 0; | |
63 | t->nr_rand_samples = 0; | |
64 | t->last_end_oblock = 0; | |
65 | t->thresholds[PATTERN_RANDOM] = random_threshold; | |
66 | t->thresholds[PATTERN_SEQUENTIAL] = sequential_threshold; | |
67 | } | |
68 | ||
69 | static enum io_pattern iot_pattern(struct io_tracker *t) | |
70 | { | |
71 | return t->pattern; | |
72 | } | |
73 | ||
74 | static void iot_update_stats(struct io_tracker *t, struct bio *bio) | |
75 | { | |
4f024f37 | 76 | if (bio->bi_iter.bi_sector == from_oblock(t->last_end_oblock) + 1) |
f2836352 JT |
77 | t->nr_seq_samples++; |
78 | else { | |
79 | /* | |
80 | * Just one non-sequential IO is enough to reset the | |
81 | * counters. | |
82 | */ | |
83 | if (t->nr_seq_samples) { | |
84 | t->nr_seq_samples = 0; | |
85 | t->nr_rand_samples = 0; | |
86 | } | |
87 | ||
88 | t->nr_rand_samples++; | |
89 | } | |
90 | ||
4f024f37 | 91 | t->last_end_oblock = to_oblock(bio_end_sector(bio) - 1); |
f2836352 JT |
92 | } |
93 | ||
94 | static void iot_check_for_pattern_switch(struct io_tracker *t) | |
95 | { | |
96 | switch (t->pattern) { | |
97 | case PATTERN_SEQUENTIAL: | |
98 | if (t->nr_rand_samples >= t->thresholds[PATTERN_RANDOM]) { | |
99 | t->pattern = PATTERN_RANDOM; | |
100 | t->nr_seq_samples = t->nr_rand_samples = 0; | |
101 | } | |
102 | break; | |
103 | ||
104 | case PATTERN_RANDOM: | |
105 | if (t->nr_seq_samples >= t->thresholds[PATTERN_SEQUENTIAL]) { | |
106 | t->pattern = PATTERN_SEQUENTIAL; | |
107 | t->nr_seq_samples = t->nr_rand_samples = 0; | |
108 | } | |
109 | break; | |
110 | } | |
111 | } | |
112 | ||
113 | static void iot_examine_bio(struct io_tracker *t, struct bio *bio) | |
114 | { | |
115 | iot_update_stats(t, bio); | |
116 | iot_check_for_pattern_switch(t); | |
117 | } | |
118 | ||
119 | /*----------------------------------------------------------------*/ | |
120 | ||
121 | ||
122 | /* | |
123 | * This queue is divided up into different levels. Allowing us to push | |
124 | * entries to the back of any of the levels. Think of it as a partially | |
125 | * sorted queue. | |
126 | */ | |
127 | #define NR_QUEUE_LEVELS 16u | |
3e45c91e | 128 | #define NR_SENTINELS NR_QUEUE_LEVELS * 3 |
f2836352 | 129 | |
e65ff870 JT |
130 | #define WRITEBACK_PERIOD HZ |
131 | ||
f2836352 | 132 | struct queue { |
75da39bf | 133 | unsigned nr_elts; |
e65ff870 JT |
134 | bool current_writeback_sentinels; |
135 | unsigned long next_writeback; | |
f2836352 | 136 | struct list_head qs[NR_QUEUE_LEVELS]; |
3e45c91e | 137 | struct list_head sentinels[NR_SENTINELS]; |
f2836352 JT |
138 | }; |
139 | ||
140 | static void queue_init(struct queue *q) | |
141 | { | |
142 | unsigned i; | |
143 | ||
75da39bf | 144 | q->nr_elts = 0; |
e65ff870 JT |
145 | q->current_writeback_sentinels = false; |
146 | q->next_writeback = 0; | |
3e45c91e | 147 | for (i = 0; i < NR_QUEUE_LEVELS; i++) { |
f2836352 | 148 | INIT_LIST_HEAD(q->qs + i); |
3e45c91e | 149 | INIT_LIST_HEAD(q->sentinels + i); |
e65ff870 JT |
150 | INIT_LIST_HEAD(q->sentinels + NR_QUEUE_LEVELS + i); |
151 | INIT_LIST_HEAD(q->sentinels + (2 * NR_QUEUE_LEVELS) + i); | |
3e45c91e | 152 | } |
f2836352 JT |
153 | } |
154 | ||
e65ff870 JT |
155 | static unsigned queue_size(struct queue *q) |
156 | { | |
157 | return q->nr_elts; | |
158 | } | |
159 | ||
c86c3070 JT |
160 | static bool queue_empty(struct queue *q) |
161 | { | |
75da39bf | 162 | return q->nr_elts == 0; |
c86c3070 JT |
163 | } |
164 | ||
f2836352 JT |
165 | /* |
166 | * Insert an entry to the back of the given level. | |
167 | */ | |
168 | static void queue_push(struct queue *q, unsigned level, struct list_head *elt) | |
169 | { | |
75da39bf | 170 | q->nr_elts++; |
f2836352 JT |
171 | list_add_tail(elt, q->qs + level); |
172 | } | |
173 | ||
75da39bf | 174 | static void queue_remove(struct queue *q, struct list_head *elt) |
f2836352 | 175 | { |
75da39bf | 176 | q->nr_elts--; |
f2836352 JT |
177 | list_del(elt); |
178 | } | |
179 | ||
3e45c91e JT |
180 | static bool is_sentinel(struct queue *q, struct list_head *h) |
181 | { | |
182 | return (h >= q->sentinels) && (h < (q->sentinels + NR_SENTINELS)); | |
183 | } | |
184 | ||
f2836352 JT |
185 | /* |
186 | * Gives us the oldest entry of the lowest popoulated level. If the first | |
187 | * level is emptied then we shift down one level. | |
188 | */ | |
b155aa0e | 189 | static struct list_head *queue_peek(struct queue *q) |
f2836352 JT |
190 | { |
191 | unsigned level; | |
3e45c91e | 192 | struct list_head *h; |
f2836352 JT |
193 | |
194 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
3e45c91e JT |
195 | list_for_each(h, q->qs + level) |
196 | if (!is_sentinel(q, h)) | |
197 | return h; | |
f2836352 | 198 | |
b155aa0e JT |
199 | return NULL; |
200 | } | |
f2836352 | 201 | |
b155aa0e JT |
202 | static struct list_head *queue_pop(struct queue *q) |
203 | { | |
204 | struct list_head *r = queue_peek(q); | |
f2836352 | 205 | |
b155aa0e | 206 | if (r) { |
75da39bf | 207 | q->nr_elts--; |
b155aa0e | 208 | list_del(r); |
b155aa0e JT |
209 | } |
210 | ||
211 | return r; | |
f2836352 JT |
212 | } |
213 | ||
e65ff870 JT |
214 | /* |
215 | * Pops an entry from a level that is not past a sentinel. | |
216 | */ | |
217 | static struct list_head *queue_pop_old(struct queue *q) | |
218 | { | |
219 | unsigned level; | |
220 | struct list_head *h; | |
221 | ||
222 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
223 | list_for_each(h, q->qs + level) { | |
224 | if (is_sentinel(q, h)) | |
225 | break; | |
226 | ||
227 | q->nr_elts--; | |
228 | list_del(h); | |
229 | return h; | |
230 | } | |
231 | ||
232 | return NULL; | |
233 | } | |
234 | ||
f2836352 JT |
235 | static struct list_head *list_pop(struct list_head *lh) |
236 | { | |
237 | struct list_head *r = lh->next; | |
238 | ||
239 | BUG_ON(!r); | |
240 | list_del_init(r); | |
241 | ||
242 | return r; | |
243 | } | |
244 | ||
e65ff870 JT |
245 | static struct list_head *writeback_sentinel(struct queue *q, unsigned level) |
246 | { | |
247 | if (q->current_writeback_sentinels) | |
248 | return q->sentinels + NR_QUEUE_LEVELS + level; | |
249 | else | |
250 | return q->sentinels + 2 * NR_QUEUE_LEVELS + level; | |
251 | } | |
252 | ||
253 | static void queue_update_writeback_sentinels(struct queue *q) | |
254 | { | |
255 | unsigned i; | |
256 | struct list_head *h; | |
257 | ||
258 | if (time_after(jiffies, q->next_writeback)) { | |
259 | for (i = 0; i < NR_QUEUE_LEVELS; i++) { | |
260 | h = writeback_sentinel(q, i); | |
261 | list_del(h); | |
262 | list_add_tail(h, q->qs + i); | |
263 | } | |
264 | ||
265 | q->next_writeback = jiffies + WRITEBACK_PERIOD; | |
266 | q->current_writeback_sentinels = !q->current_writeback_sentinels; | |
267 | } | |
268 | } | |
269 | ||
3e45c91e JT |
270 | /* |
271 | * Sometimes we want to iterate through entries that have been pushed since | |
272 | * a certain event. We use sentinel entries on the queues to delimit these | |
273 | * 'tick' events. | |
274 | */ | |
275 | static void queue_tick(struct queue *q) | |
276 | { | |
277 | unsigned i; | |
278 | ||
279 | for (i = 0; i < NR_QUEUE_LEVELS; i++) { | |
280 | list_del(q->sentinels + i); | |
281 | list_add_tail(q->sentinels + i, q->qs + i); | |
282 | } | |
283 | } | |
284 | ||
285 | typedef void (*iter_fn)(struct list_head *, void *); | |
286 | static void queue_iterate_tick(struct queue *q, iter_fn fn, void *context) | |
287 | { | |
288 | unsigned i; | |
289 | struct list_head *h; | |
290 | ||
291 | for (i = 0; i < NR_QUEUE_LEVELS; i++) { | |
292 | list_for_each_prev(h, q->qs + i) { | |
293 | if (is_sentinel(q, h)) | |
294 | break; | |
295 | ||
296 | fn(h, context); | |
297 | } | |
298 | } | |
299 | } | |
300 | ||
f2836352 JT |
301 | /*----------------------------------------------------------------*/ |
302 | ||
303 | /* | |
304 | * Describes a cache entry. Used in both the cache and the pre_cache. | |
305 | */ | |
306 | struct entry { | |
307 | struct hlist_node hlist; | |
308 | struct list_head list; | |
309 | dm_oblock_t oblock; | |
f2836352 JT |
310 | |
311 | /* | |
312 | * FIXME: pack these better | |
313 | */ | |
01911c19 | 314 | bool dirty:1; |
f2836352 | 315 | unsigned hit_count; |
f2836352 JT |
316 | }; |
317 | ||
633618e3 JT |
318 | /* |
319 | * Rather than storing the cblock in an entry, we allocate all entries in | |
320 | * an array, and infer the cblock from the entry position. | |
321 | * | |
322 | * Free entries are linked together into a list. | |
323 | */ | |
324 | struct entry_pool { | |
325 | struct entry *entries, *entries_end; | |
326 | struct list_head free; | |
327 | unsigned nr_allocated; | |
328 | }; | |
329 | ||
330 | static int epool_init(struct entry_pool *ep, unsigned nr_entries) | |
331 | { | |
332 | unsigned i; | |
333 | ||
334 | ep->entries = vzalloc(sizeof(struct entry) * nr_entries); | |
335 | if (!ep->entries) | |
336 | return -ENOMEM; | |
337 | ||
338 | ep->entries_end = ep->entries + nr_entries; | |
339 | ||
340 | INIT_LIST_HEAD(&ep->free); | |
341 | for (i = 0; i < nr_entries; i++) | |
342 | list_add(&ep->entries[i].list, &ep->free); | |
343 | ||
344 | ep->nr_allocated = 0; | |
345 | ||
346 | return 0; | |
347 | } | |
348 | ||
349 | static void epool_exit(struct entry_pool *ep) | |
350 | { | |
351 | vfree(ep->entries); | |
352 | } | |
353 | ||
354 | static struct entry *alloc_entry(struct entry_pool *ep) | |
355 | { | |
356 | struct entry *e; | |
357 | ||
358 | if (list_empty(&ep->free)) | |
359 | return NULL; | |
360 | ||
361 | e = list_entry(list_pop(&ep->free), struct entry, list); | |
362 | INIT_LIST_HEAD(&e->list); | |
363 | INIT_HLIST_NODE(&e->hlist); | |
364 | ep->nr_allocated++; | |
365 | ||
366 | return e; | |
367 | } | |
368 | ||
369 | /* | |
370 | * This assumes the cblock hasn't already been allocated. | |
371 | */ | |
372 | static struct entry *alloc_particular_entry(struct entry_pool *ep, dm_cblock_t cblock) | |
373 | { | |
374 | struct entry *e = ep->entries + from_cblock(cblock); | |
633618e3 | 375 | |
b8158051 | 376 | list_del_init(&e->list); |
633618e3 JT |
377 | INIT_HLIST_NODE(&e->hlist); |
378 | ep->nr_allocated++; | |
379 | ||
380 | return e; | |
381 | } | |
382 | ||
383 | static void free_entry(struct entry_pool *ep, struct entry *e) | |
384 | { | |
385 | BUG_ON(!ep->nr_allocated); | |
386 | ep->nr_allocated--; | |
387 | INIT_HLIST_NODE(&e->hlist); | |
388 | list_add(&e->list, &ep->free); | |
389 | } | |
390 | ||
532906aa JT |
391 | /* |
392 | * Returns NULL if the entry is free. | |
393 | */ | |
394 | static struct entry *epool_find(struct entry_pool *ep, dm_cblock_t cblock) | |
395 | { | |
396 | struct entry *e = ep->entries + from_cblock(cblock); | |
7b6b2bc9 | 397 | return !hlist_unhashed(&e->hlist) ? e : NULL; |
532906aa JT |
398 | } |
399 | ||
633618e3 JT |
400 | static bool epool_empty(struct entry_pool *ep) |
401 | { | |
402 | return list_empty(&ep->free); | |
403 | } | |
404 | ||
405 | static bool in_pool(struct entry_pool *ep, struct entry *e) | |
406 | { | |
407 | return e >= ep->entries && e < ep->entries_end; | |
408 | } | |
409 | ||
410 | static dm_cblock_t infer_cblock(struct entry_pool *ep, struct entry *e) | |
411 | { | |
412 | return to_cblock(e - ep->entries); | |
413 | } | |
414 | ||
415 | /*----------------------------------------------------------------*/ | |
416 | ||
f2836352 JT |
417 | struct mq_policy { |
418 | struct dm_cache_policy policy; | |
419 | ||
420 | /* protects everything */ | |
421 | struct mutex lock; | |
422 | dm_cblock_t cache_size; | |
423 | struct io_tracker tracker; | |
424 | ||
633618e3 JT |
425 | /* |
426 | * Entries come from two pools, one of pre-cache entries, and one | |
427 | * for the cache proper. | |
428 | */ | |
429 | struct entry_pool pre_cache_pool; | |
430 | struct entry_pool cache_pool; | |
431 | ||
f2836352 | 432 | /* |
01911c19 JT |
433 | * We maintain three queues of entries. The cache proper, |
434 | * consisting of a clean and dirty queue, contains the currently | |
435 | * active mappings. Whereas the pre_cache tracks blocks that | |
436 | * are being hit frequently and potential candidates for promotion | |
437 | * to the cache. | |
f2836352 JT |
438 | */ |
439 | struct queue pre_cache; | |
01911c19 JT |
440 | struct queue cache_clean; |
441 | struct queue cache_dirty; | |
f2836352 JT |
442 | |
443 | /* | |
444 | * Keeps track of time, incremented by the core. We use this to | |
445 | * avoid attributing multiple hits within the same tick. | |
446 | * | |
447 | * Access to tick_protected should be done with the spin lock held. | |
448 | * It's copied to tick at the start of the map function (within the | |
449 | * mutex). | |
450 | */ | |
451 | spinlock_t tick_lock; | |
452 | unsigned tick_protected; | |
453 | unsigned tick; | |
454 | ||
455 | /* | |
456 | * A count of the number of times the map function has been called | |
457 | * and found an entry in the pre_cache or cache. Currently used to | |
458 | * calculate the generation. | |
459 | */ | |
460 | unsigned hit_count; | |
461 | ||
462 | /* | |
463 | * A generation is a longish period that is used to trigger some | |
464 | * book keeping effects. eg, decrementing hit counts on entries. | |
465 | * This is needed to allow the cache to evolve as io patterns | |
466 | * change. | |
467 | */ | |
468 | unsigned generation; | |
469 | unsigned generation_period; /* in lookups (will probably change) */ | |
470 | ||
78e03d69 JT |
471 | unsigned discard_promote_adjustment; |
472 | unsigned read_promote_adjustment; | |
473 | unsigned write_promote_adjustment; | |
474 | ||
f2836352 JT |
475 | /* |
476 | * The hash table allows us to quickly find an entry by origin | |
477 | * block. Both pre_cache and cache entries are in here. | |
478 | */ | |
479 | unsigned nr_buckets; | |
480 | dm_block_t hash_bits; | |
481 | struct hlist_head *table; | |
482 | }; | |
483 | ||
78e03d69 JT |
484 | #define DEFAULT_DISCARD_PROMOTE_ADJUSTMENT 1 |
485 | #define DEFAULT_READ_PROMOTE_ADJUSTMENT 4 | |
486 | #define DEFAULT_WRITE_PROMOTE_ADJUSTMENT 8 | |
b155aa0e | 487 | #define DISCOURAGE_DEMOTING_DIRTY_THRESHOLD 128 |
78e03d69 | 488 | |
f2836352 JT |
489 | /*----------------------------------------------------------------*/ |
490 | ||
491 | /* | |
492 | * Simple hash table implementation. Should replace with the standard hash | |
493 | * table that's making its way upstream. | |
494 | */ | |
495 | static void hash_insert(struct mq_policy *mq, struct entry *e) | |
496 | { | |
497 | unsigned h = hash_64(from_oblock(e->oblock), mq->hash_bits); | |
498 | ||
499 | hlist_add_head(&e->hlist, mq->table + h); | |
500 | } | |
501 | ||
502 | static struct entry *hash_lookup(struct mq_policy *mq, dm_oblock_t oblock) | |
503 | { | |
504 | unsigned h = hash_64(from_oblock(oblock), mq->hash_bits); | |
505 | struct hlist_head *bucket = mq->table + h; | |
506 | struct entry *e; | |
507 | ||
508 | hlist_for_each_entry(e, bucket, hlist) | |
509 | if (e->oblock == oblock) { | |
510 | hlist_del(&e->hlist); | |
511 | hlist_add_head(&e->hlist, bucket); | |
512 | return e; | |
513 | } | |
514 | ||
515 | return NULL; | |
516 | } | |
517 | ||
518 | static void hash_remove(struct entry *e) | |
519 | { | |
520 | hlist_del(&e->hlist); | |
521 | } | |
522 | ||
523 | /*----------------------------------------------------------------*/ | |
524 | ||
f2836352 JT |
525 | static bool any_free_cblocks(struct mq_policy *mq) |
526 | { | |
633618e3 | 527 | return !epool_empty(&mq->cache_pool); |
f2836352 JT |
528 | } |
529 | ||
c86c3070 JT |
530 | static bool any_clean_cblocks(struct mq_policy *mq) |
531 | { | |
532 | return !queue_empty(&mq->cache_clean); | |
533 | } | |
534 | ||
f2836352 JT |
535 | /*----------------------------------------------------------------*/ |
536 | ||
537 | /* | |
538 | * Now we get to the meat of the policy. This section deals with deciding | |
539 | * when to to add entries to the pre_cache and cache, and move between | |
540 | * them. | |
541 | */ | |
542 | ||
543 | /* | |
544 | * The queue level is based on the log2 of the hit count. | |
545 | */ | |
546 | static unsigned queue_level(struct entry *e) | |
547 | { | |
548 | return min((unsigned) ilog2(e->hit_count), NR_QUEUE_LEVELS - 1u); | |
549 | } | |
550 | ||
633618e3 JT |
551 | static bool in_cache(struct mq_policy *mq, struct entry *e) |
552 | { | |
553 | return in_pool(&mq->cache_pool, e); | |
554 | } | |
555 | ||
f2836352 JT |
556 | /* |
557 | * Inserts the entry into the pre_cache or the cache. Ensures the cache | |
633618e3 JT |
558 | * block is marked as allocated if necc. Inserts into the hash table. |
559 | * Sets the tick which records when the entry was last moved about. | |
f2836352 JT |
560 | */ |
561 | static void push(struct mq_policy *mq, struct entry *e) | |
562 | { | |
f2836352 JT |
563 | hash_insert(mq, e); |
564 | ||
633618e3 | 565 | if (in_cache(mq, e)) |
01911c19 JT |
566 | queue_push(e->dirty ? &mq->cache_dirty : &mq->cache_clean, |
567 | queue_level(e), &e->list); | |
633618e3 | 568 | else |
f2836352 JT |
569 | queue_push(&mq->pre_cache, queue_level(e), &e->list); |
570 | } | |
571 | ||
572 | /* | |
573 | * Removes an entry from pre_cache or cache. Removes from the hash table. | |
f2836352 JT |
574 | */ |
575 | static void del(struct mq_policy *mq, struct entry *e) | |
576 | { | |
75da39bf JT |
577 | if (in_cache(mq, e)) |
578 | queue_remove(e->dirty ? &mq->cache_dirty : &mq->cache_clean, &e->list); | |
579 | else | |
580 | queue_remove(&mq->pre_cache, &e->list); | |
581 | ||
f2836352 | 582 | hash_remove(e); |
f2836352 JT |
583 | } |
584 | ||
585 | /* | |
586 | * Like del, except it removes the first entry in the queue (ie. the least | |
587 | * recently used). | |
588 | */ | |
589 | static struct entry *pop(struct mq_policy *mq, struct queue *q) | |
590 | { | |
0184b44e JT |
591 | struct entry *e; |
592 | struct list_head *h = queue_pop(q); | |
f2836352 | 593 | |
0184b44e JT |
594 | if (!h) |
595 | return NULL; | |
f2836352 | 596 | |
0184b44e JT |
597 | e = container_of(h, struct entry, list); |
598 | hash_remove(e); | |
f2836352 JT |
599 | |
600 | return e; | |
601 | } | |
602 | ||
e65ff870 JT |
603 | static struct entry *pop_old(struct mq_policy *mq, struct queue *q) |
604 | { | |
605 | struct entry *e; | |
606 | struct list_head *h = queue_pop_old(q); | |
607 | ||
608 | if (!h) | |
609 | return NULL; | |
610 | ||
611 | e = container_of(h, struct entry, list); | |
612 | hash_remove(e); | |
613 | ||
614 | return e; | |
615 | } | |
616 | ||
b155aa0e JT |
617 | static struct entry *peek(struct queue *q) |
618 | { | |
619 | struct list_head *h = queue_peek(q); | |
620 | return h ? container_of(h, struct entry, list) : NULL; | |
621 | } | |
622 | ||
f2836352 JT |
623 | /* |
624 | * The promotion threshold is adjusted every generation. As are the counts | |
625 | * of the entries. | |
626 | * | |
627 | * At the moment the threshold is taken by averaging the hit counts of some | |
01911c19 JT |
628 | * of the entries in the cache (the first 20 entries across all levels in |
629 | * ascending order, giving preference to the clean entries at each level). | |
f2836352 JT |
630 | * |
631 | * We can be much cleverer than this though. For example, each promotion | |
632 | * could bump up the threshold helping to prevent churn. Much more to do | |
633 | * here. | |
634 | */ | |
635 | ||
636 | #define MAX_TO_AVERAGE 20 | |
637 | ||
638 | static void check_generation(struct mq_policy *mq) | |
639 | { | |
640 | unsigned total = 0, nr = 0, count = 0, level; | |
641 | struct list_head *head; | |
642 | struct entry *e; | |
643 | ||
633618e3 | 644 | if ((mq->hit_count >= mq->generation_period) && (epool_empty(&mq->cache_pool))) { |
f2836352 JT |
645 | mq->hit_count = 0; |
646 | mq->generation++; | |
647 | ||
648 | for (level = 0; level < NR_QUEUE_LEVELS && count < MAX_TO_AVERAGE; level++) { | |
01911c19 JT |
649 | head = mq->cache_clean.qs + level; |
650 | list_for_each_entry(e, head, list) { | |
651 | nr++; | |
652 | total += e->hit_count; | |
653 | ||
654 | if (++count >= MAX_TO_AVERAGE) | |
655 | break; | |
656 | } | |
657 | ||
658 | head = mq->cache_dirty.qs + level; | |
f2836352 JT |
659 | list_for_each_entry(e, head, list) { |
660 | nr++; | |
661 | total += e->hit_count; | |
662 | ||
663 | if (++count >= MAX_TO_AVERAGE) | |
664 | break; | |
665 | } | |
666 | } | |
f2836352 JT |
667 | } |
668 | } | |
669 | ||
670 | /* | |
671 | * Whenever we use an entry we bump up it's hit counter, and push it to the | |
672 | * back to it's current level. | |
673 | */ | |
3e45c91e | 674 | static void requeue(struct mq_policy *mq, struct entry *e) |
f2836352 | 675 | { |
f2836352 | 676 | check_generation(mq); |
f2836352 JT |
677 | del(mq, e); |
678 | push(mq, e); | |
679 | } | |
680 | ||
681 | /* | |
682 | * Demote the least recently used entry from the cache to the pre_cache. | |
683 | * Returns the new cache entry to use, and the old origin block it was | |
684 | * mapped to. | |
685 | * | |
686 | * We drop the hit count on the demoted entry back to 1 to stop it bouncing | |
687 | * straight back into the cache if it's subsequently hit. There are | |
688 | * various options here, and more experimentation would be good: | |
689 | * | |
690 | * - just forget about the demoted entry completely (ie. don't insert it | |
691 | into the pre_cache). | |
692 | * - divide the hit count rather that setting to some hard coded value. | |
693 | * - set the hit count to a hard coded value other than 1, eg, is it better | |
694 | * if it goes in at level 2? | |
695 | */ | |
633618e3 | 696 | static int demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock) |
f2836352 | 697 | { |
01911c19 | 698 | struct entry *demoted = pop(mq, &mq->cache_clean); |
f2836352 | 699 | |
01911c19 JT |
700 | if (!demoted) |
701 | /* | |
702 | * We could get a block from mq->cache_dirty, but that | |
703 | * would add extra latency to the triggering bio as it | |
704 | * waits for the writeback. Better to not promote this | |
705 | * time and hope there's a clean block next time this block | |
706 | * is hit. | |
707 | */ | |
708 | return -ENOSPC; | |
709 | ||
f2836352 | 710 | *oblock = demoted->oblock; |
633618e3 JT |
711 | free_entry(&mq->cache_pool, demoted); |
712 | ||
713 | /* | |
714 | * We used to put the demoted block into the pre-cache, but I think | |
715 | * it's simpler to just let it work it's way up from zero again. | |
716 | * Stops blocks flickering in and out of the cache. | |
717 | */ | |
f2836352 | 718 | |
01911c19 | 719 | return 0; |
f2836352 JT |
720 | } |
721 | ||
b155aa0e JT |
722 | /* |
723 | * Entries in the pre_cache whose hit count passes the promotion | |
724 | * threshold move to the cache proper. Working out the correct | |
725 | * value for the promotion_threshold is crucial to this policy. | |
726 | */ | |
727 | static unsigned promote_threshold(struct mq_policy *mq) | |
728 | { | |
729 | struct entry *e; | |
730 | ||
731 | if (any_free_cblocks(mq)) | |
732 | return 0; | |
733 | ||
734 | e = peek(&mq->cache_clean); | |
735 | if (e) | |
736 | return e->hit_count; | |
737 | ||
738 | e = peek(&mq->cache_dirty); | |
739 | if (e) | |
740 | return e->hit_count + DISCOURAGE_DEMOTING_DIRTY_THRESHOLD; | |
741 | ||
742 | /* This should never happen */ | |
743 | return 0; | |
744 | } | |
745 | ||
f2836352 JT |
746 | /* |
747 | * We modify the basic promotion_threshold depending on the specific io. | |
748 | * | |
749 | * If the origin block has been discarded then there's no cost to copy it | |
750 | * to the cache. | |
751 | * | |
752 | * We bias towards reads, since they can be demoted at no cost if they | |
753 | * haven't been dirtied. | |
754 | */ | |
f2836352 JT |
755 | static unsigned adjusted_promote_threshold(struct mq_policy *mq, |
756 | bool discarded_oblock, int data_dir) | |
757 | { | |
c86c3070 | 758 | if (data_dir == READ) |
b155aa0e | 759 | return promote_threshold(mq) + mq->read_promote_adjustment; |
c86c3070 JT |
760 | |
761 | if (discarded_oblock && (any_free_cblocks(mq) || any_clean_cblocks(mq))) { | |
f2836352 JT |
762 | /* |
763 | * We don't need to do any copying at all, so give this a | |
c86c3070 | 764 | * very low threshold. |
f2836352 | 765 | */ |
78e03d69 | 766 | return mq->discard_promote_adjustment; |
c86c3070 | 767 | } |
f2836352 | 768 | |
b155aa0e | 769 | return promote_threshold(mq) + mq->write_promote_adjustment; |
f2836352 JT |
770 | } |
771 | ||
772 | static bool should_promote(struct mq_policy *mq, struct entry *e, | |
773 | bool discarded_oblock, int data_dir) | |
774 | { | |
775 | return e->hit_count >= | |
776 | adjusted_promote_threshold(mq, discarded_oblock, data_dir); | |
777 | } | |
778 | ||
779 | static int cache_entry_found(struct mq_policy *mq, | |
780 | struct entry *e, | |
781 | struct policy_result *result) | |
782 | { | |
3e45c91e | 783 | requeue(mq, e); |
f2836352 | 784 | |
633618e3 | 785 | if (in_cache(mq, e)) { |
f2836352 | 786 | result->op = POLICY_HIT; |
633618e3 | 787 | result->cblock = infer_cblock(&mq->cache_pool, e); |
f2836352 JT |
788 | } |
789 | ||
790 | return 0; | |
791 | } | |
792 | ||
793 | /* | |
0184b44e | 794 | * Moves an entry from the pre_cache to the cache. The main work is |
f2836352 JT |
795 | * finding which cache block to use. |
796 | */ | |
797 | static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e, | |
798 | struct policy_result *result) | |
799 | { | |
01911c19 | 800 | int r; |
633618e3 | 801 | struct entry *new_e; |
f2836352 | 802 | |
633618e3 JT |
803 | /* Ensure there's a free cblock in the cache */ |
804 | if (epool_empty(&mq->cache_pool)) { | |
f2836352 | 805 | result->op = POLICY_REPLACE; |
633618e3 | 806 | r = demote_cblock(mq, &result->old_oblock); |
01911c19 JT |
807 | if (r) { |
808 | result->op = POLICY_MISS; | |
809 | return 0; | |
810 | } | |
f2836352 JT |
811 | } else |
812 | result->op = POLICY_NEW; | |
813 | ||
633618e3 JT |
814 | new_e = alloc_entry(&mq->cache_pool); |
815 | BUG_ON(!new_e); | |
816 | ||
817 | new_e->oblock = e->oblock; | |
818 | new_e->dirty = false; | |
819 | new_e->hit_count = e->hit_count; | |
f2836352 JT |
820 | |
821 | del(mq, e); | |
633618e3 JT |
822 | free_entry(&mq->pre_cache_pool, e); |
823 | push(mq, new_e); | |
824 | ||
825 | result->cblock = infer_cblock(&mq->cache_pool, new_e); | |
f2836352 JT |
826 | |
827 | return 0; | |
828 | } | |
829 | ||
830 | static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e, | |
831 | bool can_migrate, bool discarded_oblock, | |
832 | int data_dir, struct policy_result *result) | |
833 | { | |
834 | int r = 0; | |
f2836352 | 835 | |
3e45c91e JT |
836 | if (!should_promote(mq, e, discarded_oblock, data_dir)) { |
837 | requeue(mq, e); | |
f2836352 | 838 | result->op = POLICY_MISS; |
af95e7a6 JT |
839 | |
840 | } else if (!can_migrate) | |
f2836352 | 841 | r = -EWOULDBLOCK; |
af95e7a6 JT |
842 | |
843 | else { | |
3e45c91e | 844 | requeue(mq, e); |
f2836352 | 845 | r = pre_cache_to_cache(mq, e, result); |
af95e7a6 | 846 | } |
f2836352 JT |
847 | |
848 | return r; | |
849 | } | |
850 | ||
851 | static void insert_in_pre_cache(struct mq_policy *mq, | |
852 | dm_oblock_t oblock) | |
853 | { | |
633618e3 | 854 | struct entry *e = alloc_entry(&mq->pre_cache_pool); |
f2836352 JT |
855 | |
856 | if (!e) | |
857 | /* | |
858 | * There's no spare entry structure, so we grab the least | |
859 | * used one from the pre_cache. | |
860 | */ | |
861 | e = pop(mq, &mq->pre_cache); | |
862 | ||
863 | if (unlikely(!e)) { | |
864 | DMWARN("couldn't pop from pre cache"); | |
865 | return; | |
866 | } | |
867 | ||
633618e3 JT |
868 | e->dirty = false; |
869 | e->oblock = oblock; | |
870 | e->hit_count = 1; | |
633618e3 | 871 | push(mq, e); |
f2836352 JT |
872 | } |
873 | ||
874 | static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock, | |
875 | struct policy_result *result) | |
876 | { | |
c86c3070 | 877 | int r; |
f2836352 | 878 | struct entry *e; |
f2836352 | 879 | |
633618e3 JT |
880 | if (epool_empty(&mq->cache_pool)) { |
881 | result->op = POLICY_REPLACE; | |
882 | r = demote_cblock(mq, &result->old_oblock); | |
c86c3070 JT |
883 | if (unlikely(r)) { |
884 | result->op = POLICY_MISS; | |
885 | insert_in_pre_cache(mq, oblock); | |
886 | return; | |
887 | } | |
f2836352 | 888 | |
c86c3070 JT |
889 | /* |
890 | * This will always succeed, since we've just demoted. | |
891 | */ | |
633618e3 JT |
892 | e = alloc_entry(&mq->cache_pool); |
893 | BUG_ON(!e); | |
c86c3070 JT |
894 | |
895 | } else { | |
633618e3 | 896 | e = alloc_entry(&mq->cache_pool); |
c86c3070 | 897 | result->op = POLICY_NEW; |
f2836352 JT |
898 | } |
899 | ||
900 | e->oblock = oblock; | |
01911c19 | 901 | e->dirty = false; |
f2836352 | 902 | e->hit_count = 1; |
f2836352 JT |
903 | push(mq, e); |
904 | ||
633618e3 | 905 | result->cblock = infer_cblock(&mq->cache_pool, e); |
f2836352 JT |
906 | } |
907 | ||
908 | static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock, | |
909 | bool can_migrate, bool discarded_oblock, | |
910 | int data_dir, struct policy_result *result) | |
911 | { | |
78e03d69 | 912 | if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) <= 1) { |
f2836352 JT |
913 | if (can_migrate) |
914 | insert_in_cache(mq, oblock, result); | |
915 | else | |
916 | return -EWOULDBLOCK; | |
917 | } else { | |
918 | insert_in_pre_cache(mq, oblock); | |
919 | result->op = POLICY_MISS; | |
920 | } | |
921 | ||
922 | return 0; | |
923 | } | |
924 | ||
925 | /* | |
926 | * Looks the oblock up in the hash table, then decides whether to put in | |
927 | * pre_cache, or cache etc. | |
928 | */ | |
929 | static int map(struct mq_policy *mq, dm_oblock_t oblock, | |
930 | bool can_migrate, bool discarded_oblock, | |
931 | int data_dir, struct policy_result *result) | |
932 | { | |
933 | int r = 0; | |
934 | struct entry *e = hash_lookup(mq, oblock); | |
935 | ||
633618e3 | 936 | if (e && in_cache(mq, e)) |
f2836352 | 937 | r = cache_entry_found(mq, e, result); |
633618e3 | 938 | |
f1afb36a MS |
939 | else if (mq->tracker.thresholds[PATTERN_SEQUENTIAL] && |
940 | iot_pattern(&mq->tracker) == PATTERN_SEQUENTIAL) | |
f2836352 | 941 | result->op = POLICY_MISS; |
633618e3 | 942 | |
f2836352 JT |
943 | else if (e) |
944 | r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock, | |
945 | data_dir, result); | |
633618e3 | 946 | |
f2836352 JT |
947 | else |
948 | r = no_entry_found(mq, oblock, can_migrate, discarded_oblock, | |
949 | data_dir, result); | |
950 | ||
951 | if (r == -EWOULDBLOCK) | |
952 | result->op = POLICY_MISS; | |
953 | ||
954 | return r; | |
955 | } | |
956 | ||
957 | /*----------------------------------------------------------------*/ | |
958 | ||
959 | /* | |
960 | * Public interface, via the policy struct. See dm-cache-policy.h for a | |
961 | * description of these. | |
962 | */ | |
963 | ||
964 | static struct mq_policy *to_mq_policy(struct dm_cache_policy *p) | |
965 | { | |
966 | return container_of(p, struct mq_policy, policy); | |
967 | } | |
968 | ||
969 | static void mq_destroy(struct dm_cache_policy *p) | |
970 | { | |
971 | struct mq_policy *mq = to_mq_policy(p); | |
972 | ||
14f398ca | 973 | vfree(mq->table); |
633618e3 JT |
974 | epool_exit(&mq->cache_pool); |
975 | epool_exit(&mq->pre_cache_pool); | |
f2836352 JT |
976 | kfree(mq); |
977 | } | |
978 | ||
3e45c91e JT |
979 | static void update_pre_cache_hits(struct list_head *h, void *context) |
980 | { | |
981 | struct entry *e = container_of(h, struct entry, list); | |
982 | e->hit_count++; | |
983 | } | |
984 | ||
985 | static void update_cache_hits(struct list_head *h, void *context) | |
986 | { | |
987 | struct mq_policy *mq = context; | |
988 | struct entry *e = container_of(h, struct entry, list); | |
989 | e->hit_count++; | |
990 | mq->hit_count++; | |
991 | } | |
992 | ||
f2836352 JT |
993 | static void copy_tick(struct mq_policy *mq) |
994 | { | |
3e45c91e | 995 | unsigned long flags, tick; |
f2836352 JT |
996 | |
997 | spin_lock_irqsave(&mq->tick_lock, flags); | |
3e45c91e JT |
998 | tick = mq->tick_protected; |
999 | if (tick != mq->tick) { | |
1000 | queue_iterate_tick(&mq->pre_cache, update_pre_cache_hits, mq); | |
1001 | queue_iterate_tick(&mq->cache_dirty, update_cache_hits, mq); | |
1002 | queue_iterate_tick(&mq->cache_clean, update_cache_hits, mq); | |
1003 | mq->tick = tick; | |
1004 | } | |
1005 | ||
1006 | queue_tick(&mq->pre_cache); | |
1007 | queue_tick(&mq->cache_dirty); | |
1008 | queue_tick(&mq->cache_clean); | |
e65ff870 | 1009 | queue_update_writeback_sentinels(&mq->cache_dirty); |
f2836352 JT |
1010 | spin_unlock_irqrestore(&mq->tick_lock, flags); |
1011 | } | |
1012 | ||
1013 | static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock, | |
1014 | bool can_block, bool can_migrate, bool discarded_oblock, | |
1015 | struct bio *bio, struct policy_result *result) | |
1016 | { | |
1017 | int r; | |
1018 | struct mq_policy *mq = to_mq_policy(p); | |
1019 | ||
1020 | result->op = POLICY_MISS; | |
1021 | ||
1022 | if (can_block) | |
1023 | mutex_lock(&mq->lock); | |
1024 | else if (!mutex_trylock(&mq->lock)) | |
1025 | return -EWOULDBLOCK; | |
1026 | ||
1027 | copy_tick(mq); | |
1028 | ||
1029 | iot_examine_bio(&mq->tracker, bio); | |
1030 | r = map(mq, oblock, can_migrate, discarded_oblock, | |
1031 | bio_data_dir(bio), result); | |
1032 | ||
1033 | mutex_unlock(&mq->lock); | |
1034 | ||
1035 | return r; | |
1036 | } | |
1037 | ||
1038 | static int mq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock) | |
1039 | { | |
1040 | int r; | |
1041 | struct mq_policy *mq = to_mq_policy(p); | |
1042 | struct entry *e; | |
1043 | ||
1044 | if (!mutex_trylock(&mq->lock)) | |
1045 | return -EWOULDBLOCK; | |
1046 | ||
1047 | e = hash_lookup(mq, oblock); | |
633618e3 JT |
1048 | if (e && in_cache(mq, e)) { |
1049 | *cblock = infer_cblock(&mq->cache_pool, e); | |
f2836352 JT |
1050 | r = 0; |
1051 | } else | |
1052 | r = -ENOENT; | |
1053 | ||
1054 | mutex_unlock(&mq->lock); | |
1055 | ||
1056 | return r; | |
1057 | } | |
1058 | ||
633618e3 | 1059 | static void __mq_set_clear_dirty(struct mq_policy *mq, dm_oblock_t oblock, bool set) |
01911c19 | 1060 | { |
01911c19 JT |
1061 | struct entry *e; |
1062 | ||
01911c19 | 1063 | e = hash_lookup(mq, oblock); |
633618e3 | 1064 | BUG_ON(!e || !in_cache(mq, e)); |
01911c19 | 1065 | |
633618e3 JT |
1066 | del(mq, e); |
1067 | e->dirty = set; | |
1068 | push(mq, e); | |
01911c19 JT |
1069 | } |
1070 | ||
1071 | static void mq_set_dirty(struct dm_cache_policy *p, dm_oblock_t oblock) | |
1072 | { | |
633618e3 JT |
1073 | struct mq_policy *mq = to_mq_policy(p); |
1074 | ||
1075 | mutex_lock(&mq->lock); | |
1076 | __mq_set_clear_dirty(mq, oblock, true); | |
1077 | mutex_unlock(&mq->lock); | |
01911c19 JT |
1078 | } |
1079 | ||
1080 | static void mq_clear_dirty(struct dm_cache_policy *p, dm_oblock_t oblock) | |
1081 | { | |
633618e3 JT |
1082 | struct mq_policy *mq = to_mq_policy(p); |
1083 | ||
1084 | mutex_lock(&mq->lock); | |
1085 | __mq_set_clear_dirty(mq, oblock, false); | |
1086 | mutex_unlock(&mq->lock); | |
01911c19 JT |
1087 | } |
1088 | ||
f2836352 JT |
1089 | static int mq_load_mapping(struct dm_cache_policy *p, |
1090 | dm_oblock_t oblock, dm_cblock_t cblock, | |
1091 | uint32_t hint, bool hint_valid) | |
1092 | { | |
1093 | struct mq_policy *mq = to_mq_policy(p); | |
1094 | struct entry *e; | |
1095 | ||
633618e3 | 1096 | e = alloc_particular_entry(&mq->cache_pool, cblock); |
f2836352 | 1097 | e->oblock = oblock; |
01911c19 | 1098 | e->dirty = false; /* this gets corrected in a minute */ |
f2836352 | 1099 | e->hit_count = hint_valid ? hint : 1; |
f2836352 JT |
1100 | push(mq, e); |
1101 | ||
1102 | return 0; | |
1103 | } | |
1104 | ||
633618e3 JT |
1105 | static int mq_save_hints(struct mq_policy *mq, struct queue *q, |
1106 | policy_walk_fn fn, void *context) | |
1107 | { | |
1108 | int r; | |
1109 | unsigned level; | |
3e45c91e | 1110 | struct list_head *h; |
633618e3 JT |
1111 | struct entry *e; |
1112 | ||
1113 | for (level = 0; level < NR_QUEUE_LEVELS; level++) | |
3e45c91e JT |
1114 | list_for_each(h, q->qs + level) { |
1115 | if (is_sentinel(q, h)) | |
1116 | continue; | |
1117 | ||
1118 | e = container_of(h, struct entry, list); | |
633618e3 JT |
1119 | r = fn(context, infer_cblock(&mq->cache_pool, e), |
1120 | e->oblock, e->hit_count); | |
1121 | if (r) | |
1122 | return r; | |
1123 | } | |
1124 | ||
1125 | return 0; | |
1126 | } | |
1127 | ||
f2836352 JT |
1128 | static int mq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn, |
1129 | void *context) | |
1130 | { | |
1131 | struct mq_policy *mq = to_mq_policy(p); | |
1132 | int r = 0; | |
f2836352 JT |
1133 | |
1134 | mutex_lock(&mq->lock); | |
1135 | ||
633618e3 JT |
1136 | r = mq_save_hints(mq, &mq->cache_clean, fn, context); |
1137 | if (!r) | |
1138 | r = mq_save_hints(mq, &mq->cache_dirty, fn, context); | |
f2836352 | 1139 | |
f2836352 JT |
1140 | mutex_unlock(&mq->lock); |
1141 | ||
1142 | return r; | |
1143 | } | |
1144 | ||
633618e3 | 1145 | static void __remove_mapping(struct mq_policy *mq, dm_oblock_t oblock) |
f2836352 | 1146 | { |
b936bf8b GU |
1147 | struct entry *e; |
1148 | ||
b936bf8b | 1149 | e = hash_lookup(mq, oblock); |
633618e3 | 1150 | BUG_ON(!e || !in_cache(mq, e)); |
f2836352 JT |
1151 | |
1152 | del(mq, e); | |
633618e3 JT |
1153 | free_entry(&mq->cache_pool, e); |
1154 | } | |
1155 | ||
1156 | static void mq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock) | |
1157 | { | |
1158 | struct mq_policy *mq = to_mq_policy(p); | |
f2836352 | 1159 | |
633618e3 JT |
1160 | mutex_lock(&mq->lock); |
1161 | __remove_mapping(mq, oblock); | |
f2836352 JT |
1162 | mutex_unlock(&mq->lock); |
1163 | } | |
1164 | ||
532906aa JT |
1165 | static int __remove_cblock(struct mq_policy *mq, dm_cblock_t cblock) |
1166 | { | |
1167 | struct entry *e = epool_find(&mq->cache_pool, cblock); | |
1168 | ||
1169 | if (!e) | |
1170 | return -ENODATA; | |
1171 | ||
1172 | del(mq, e); | |
1173 | free_entry(&mq->cache_pool, e); | |
1174 | ||
1175 | return 0; | |
1176 | } | |
1177 | ||
1178 | static int mq_remove_cblock(struct dm_cache_policy *p, dm_cblock_t cblock) | |
1179 | { | |
1180 | int r; | |
1181 | struct mq_policy *mq = to_mq_policy(p); | |
1182 | ||
1183 | mutex_lock(&mq->lock); | |
1184 | r = __remove_cblock(mq, cblock); | |
1185 | mutex_unlock(&mq->lock); | |
1186 | ||
1187 | return r; | |
1188 | } | |
1189 | ||
e65ff870 JT |
1190 | #define CLEAN_TARGET_PERCENTAGE 25 |
1191 | ||
1192 | static bool clean_target_met(struct mq_policy *mq) | |
1193 | { | |
1194 | /* | |
1195 | * Cache entries may not be populated. So we're cannot rely on the | |
1196 | * size of the clean queue. | |
1197 | */ | |
1198 | unsigned nr_clean = from_cblock(mq->cache_size) - queue_size(&mq->cache_dirty); | |
1199 | unsigned target = from_cblock(mq->cache_size) * CLEAN_TARGET_PERCENTAGE / 100; | |
1200 | ||
1201 | return nr_clean >= target; | |
1202 | } | |
1203 | ||
01911c19 JT |
1204 | static int __mq_writeback_work(struct mq_policy *mq, dm_oblock_t *oblock, |
1205 | dm_cblock_t *cblock) | |
1206 | { | |
e65ff870 JT |
1207 | struct entry *e = pop_old(mq, &mq->cache_dirty); |
1208 | ||
1209 | if (!e && !clean_target_met(mq)) | |
1210 | e = pop(mq, &mq->cache_dirty); | |
01911c19 JT |
1211 | |
1212 | if (!e) | |
1213 | return -ENODATA; | |
1214 | ||
1215 | *oblock = e->oblock; | |
633618e3 | 1216 | *cblock = infer_cblock(&mq->cache_pool, e); |
01911c19 JT |
1217 | e->dirty = false; |
1218 | push(mq, e); | |
1219 | ||
1220 | return 0; | |
1221 | } | |
1222 | ||
1223 | static int mq_writeback_work(struct dm_cache_policy *p, dm_oblock_t *oblock, | |
1224 | dm_cblock_t *cblock) | |
1225 | { | |
1226 | int r; | |
1227 | struct mq_policy *mq = to_mq_policy(p); | |
1228 | ||
1229 | mutex_lock(&mq->lock); | |
1230 | r = __mq_writeback_work(mq, oblock, cblock); | |
1231 | mutex_unlock(&mq->lock); | |
1232 | ||
1233 | return r; | |
1234 | } | |
1235 | ||
633618e3 JT |
1236 | static void __force_mapping(struct mq_policy *mq, |
1237 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) | |
f2836352 JT |
1238 | { |
1239 | struct entry *e = hash_lookup(mq, current_oblock); | |
1240 | ||
633618e3 JT |
1241 | if (e && in_cache(mq, e)) { |
1242 | del(mq, e); | |
1243 | e->oblock = new_oblock; | |
1244 | e->dirty = true; | |
1245 | push(mq, e); | |
1246 | } | |
f2836352 JT |
1247 | } |
1248 | ||
1249 | static void mq_force_mapping(struct dm_cache_policy *p, | |
1250 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) | |
1251 | { | |
1252 | struct mq_policy *mq = to_mq_policy(p); | |
1253 | ||
1254 | mutex_lock(&mq->lock); | |
633618e3 | 1255 | __force_mapping(mq, current_oblock, new_oblock); |
f2836352 JT |
1256 | mutex_unlock(&mq->lock); |
1257 | } | |
1258 | ||
1259 | static dm_cblock_t mq_residency(struct dm_cache_policy *p) | |
1260 | { | |
99ba2ae4 | 1261 | dm_cblock_t r; |
f2836352 JT |
1262 | struct mq_policy *mq = to_mq_policy(p); |
1263 | ||
99ba2ae4 | 1264 | mutex_lock(&mq->lock); |
633618e3 | 1265 | r = to_cblock(mq->cache_pool.nr_allocated); |
99ba2ae4 JT |
1266 | mutex_unlock(&mq->lock); |
1267 | ||
1268 | return r; | |
f2836352 JT |
1269 | } |
1270 | ||
1271 | static void mq_tick(struct dm_cache_policy *p) | |
1272 | { | |
1273 | struct mq_policy *mq = to_mq_policy(p); | |
1274 | unsigned long flags; | |
1275 | ||
1276 | spin_lock_irqsave(&mq->tick_lock, flags); | |
1277 | mq->tick_protected++; | |
1278 | spin_unlock_irqrestore(&mq->tick_lock, flags); | |
1279 | } | |
1280 | ||
1281 | static int mq_set_config_value(struct dm_cache_policy *p, | |
1282 | const char *key, const char *value) | |
1283 | { | |
1284 | struct mq_policy *mq = to_mq_policy(p); | |
f2836352 JT |
1285 | unsigned long tmp; |
1286 | ||
f2836352 JT |
1287 | if (kstrtoul(value, 10, &tmp)) |
1288 | return -EINVAL; | |
1289 | ||
78e03d69 JT |
1290 | if (!strcasecmp(key, "random_threshold")) { |
1291 | mq->tracker.thresholds[PATTERN_RANDOM] = tmp; | |
1292 | ||
1293 | } else if (!strcasecmp(key, "sequential_threshold")) { | |
1294 | mq->tracker.thresholds[PATTERN_SEQUENTIAL] = tmp; | |
1295 | ||
1296 | } else if (!strcasecmp(key, "discard_promote_adjustment")) | |
1297 | mq->discard_promote_adjustment = tmp; | |
1298 | ||
1299 | else if (!strcasecmp(key, "read_promote_adjustment")) | |
1300 | mq->read_promote_adjustment = tmp; | |
1301 | ||
1302 | else if (!strcasecmp(key, "write_promote_adjustment")) | |
1303 | mq->write_promote_adjustment = tmp; | |
1304 | ||
1305 | else | |
1306 | return -EINVAL; | |
f2836352 JT |
1307 | |
1308 | return 0; | |
1309 | } | |
1310 | ||
1311 | static int mq_emit_config_values(struct dm_cache_policy *p, char *result, unsigned maxlen) | |
1312 | { | |
1313 | ssize_t sz = 0; | |
1314 | struct mq_policy *mq = to_mq_policy(p); | |
1315 | ||
78e03d69 JT |
1316 | DMEMIT("10 random_threshold %u " |
1317 | "sequential_threshold %u " | |
1318 | "discard_promote_adjustment %u " | |
1319 | "read_promote_adjustment %u " | |
1320 | "write_promote_adjustment %u", | |
f2836352 | 1321 | mq->tracker.thresholds[PATTERN_RANDOM], |
78e03d69 JT |
1322 | mq->tracker.thresholds[PATTERN_SEQUENTIAL], |
1323 | mq->discard_promote_adjustment, | |
1324 | mq->read_promote_adjustment, | |
1325 | mq->write_promote_adjustment); | |
f2836352 JT |
1326 | |
1327 | return 0; | |
1328 | } | |
1329 | ||
1330 | /* Init the policy plugin interface function pointers. */ | |
1331 | static void init_policy_functions(struct mq_policy *mq) | |
1332 | { | |
1333 | mq->policy.destroy = mq_destroy; | |
1334 | mq->policy.map = mq_map; | |
1335 | mq->policy.lookup = mq_lookup; | |
01911c19 JT |
1336 | mq->policy.set_dirty = mq_set_dirty; |
1337 | mq->policy.clear_dirty = mq_clear_dirty; | |
f2836352 JT |
1338 | mq->policy.load_mapping = mq_load_mapping; |
1339 | mq->policy.walk_mappings = mq_walk_mappings; | |
1340 | mq->policy.remove_mapping = mq_remove_mapping; | |
532906aa | 1341 | mq->policy.remove_cblock = mq_remove_cblock; |
01911c19 | 1342 | mq->policy.writeback_work = mq_writeback_work; |
f2836352 JT |
1343 | mq->policy.force_mapping = mq_force_mapping; |
1344 | mq->policy.residency = mq_residency; | |
1345 | mq->policy.tick = mq_tick; | |
1346 | mq->policy.emit_config_values = mq_emit_config_values; | |
1347 | mq->policy.set_config_value = mq_set_config_value; | |
1348 | } | |
1349 | ||
1350 | static struct dm_cache_policy *mq_create(dm_cblock_t cache_size, | |
1351 | sector_t origin_size, | |
1352 | sector_t cache_block_size) | |
1353 | { | |
f2836352 JT |
1354 | struct mq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL); |
1355 | ||
1356 | if (!mq) | |
1357 | return NULL; | |
1358 | ||
1359 | init_policy_functions(mq); | |
1360 | iot_init(&mq->tracker, SEQUENTIAL_THRESHOLD_DEFAULT, RANDOM_THRESHOLD_DEFAULT); | |
f2836352 | 1361 | mq->cache_size = cache_size; |
633618e3 JT |
1362 | |
1363 | if (epool_init(&mq->pre_cache_pool, from_cblock(cache_size))) { | |
1364 | DMERR("couldn't initialize pool of pre-cache entries"); | |
1365 | goto bad_pre_cache_init; | |
1366 | } | |
1367 | ||
1368 | if (epool_init(&mq->cache_pool, from_cblock(cache_size))) { | |
1369 | DMERR("couldn't initialize pool of cache entries"); | |
1370 | goto bad_cache_init; | |
1371 | } | |
1372 | ||
f2836352 JT |
1373 | mq->tick_protected = 0; |
1374 | mq->tick = 0; | |
1375 | mq->hit_count = 0; | |
1376 | mq->generation = 0; | |
78e03d69 JT |
1377 | mq->discard_promote_adjustment = DEFAULT_DISCARD_PROMOTE_ADJUSTMENT; |
1378 | mq->read_promote_adjustment = DEFAULT_READ_PROMOTE_ADJUSTMENT; | |
1379 | mq->write_promote_adjustment = DEFAULT_WRITE_PROMOTE_ADJUSTMENT; | |
f2836352 JT |
1380 | mutex_init(&mq->lock); |
1381 | spin_lock_init(&mq->tick_lock); | |
f2836352 JT |
1382 | |
1383 | queue_init(&mq->pre_cache); | |
01911c19 JT |
1384 | queue_init(&mq->cache_clean); |
1385 | queue_init(&mq->cache_dirty); | |
1386 | ||
f2836352 JT |
1387 | mq->generation_period = max((unsigned) from_cblock(cache_size), 1024U); |
1388 | ||
f2836352 JT |
1389 | mq->nr_buckets = next_power(from_cblock(cache_size) / 2, 16); |
1390 | mq->hash_bits = ffs(mq->nr_buckets) - 1; | |
14f398ca | 1391 | mq->table = vzalloc(sizeof(*mq->table) * mq->nr_buckets); |
f2836352 JT |
1392 | if (!mq->table) |
1393 | goto bad_alloc_table; | |
1394 | ||
f2836352 JT |
1395 | return &mq->policy; |
1396 | ||
f2836352 | 1397 | bad_alloc_table: |
633618e3 JT |
1398 | epool_exit(&mq->cache_pool); |
1399 | bad_cache_init: | |
1400 | epool_exit(&mq->pre_cache_pool); | |
1401 | bad_pre_cache_init: | |
f2836352 JT |
1402 | kfree(mq); |
1403 | ||
1404 | return NULL; | |
1405 | } | |
1406 | ||
1407 | /*----------------------------------------------------------------*/ | |
1408 | ||
1409 | static struct dm_cache_policy_type mq_policy_type = { | |
1410 | .name = "mq", | |
f1afb36a | 1411 | .version = {1, 3, 0}, |
f2836352 JT |
1412 | .hint_size = 4, |
1413 | .owner = THIS_MODULE, | |
1414 | .create = mq_create | |
1415 | }; | |
1416 | ||
1417 | static struct dm_cache_policy_type default_policy_type = { | |
1418 | .name = "default", | |
f1afb36a | 1419 | .version = {1, 3, 0}, |
f2836352 JT |
1420 | .hint_size = 4, |
1421 | .owner = THIS_MODULE, | |
2e68c4e6 MS |
1422 | .create = mq_create, |
1423 | .real = &mq_policy_type | |
f2836352 JT |
1424 | }; |
1425 | ||
1426 | static int __init mq_init(void) | |
1427 | { | |
1428 | int r; | |
1429 | ||
1430 | mq_entry_cache = kmem_cache_create("dm_mq_policy_cache_entry", | |
1431 | sizeof(struct entry), | |
1432 | __alignof__(struct entry), | |
1433 | 0, NULL); | |
1434 | if (!mq_entry_cache) | |
1435 | goto bad; | |
1436 | ||
1437 | r = dm_cache_policy_register(&mq_policy_type); | |
1438 | if (r) { | |
1439 | DMERR("register failed %d", r); | |
1440 | goto bad_register_mq; | |
1441 | } | |
1442 | ||
1443 | r = dm_cache_policy_register(&default_policy_type); | |
1444 | if (!r) { | |
4e7f506f MS |
1445 | DMINFO("version %u.%u.%u loaded", |
1446 | mq_policy_type.version[0], | |
1447 | mq_policy_type.version[1], | |
1448 | mq_policy_type.version[2]); | |
f2836352 JT |
1449 | return 0; |
1450 | } | |
1451 | ||
1452 | DMERR("register failed (as default) %d", r); | |
1453 | ||
1454 | dm_cache_policy_unregister(&mq_policy_type); | |
1455 | bad_register_mq: | |
1456 | kmem_cache_destroy(mq_entry_cache); | |
1457 | bad: | |
1458 | return -ENOMEM; | |
1459 | } | |
1460 | ||
1461 | static void __exit mq_exit(void) | |
1462 | { | |
1463 | dm_cache_policy_unregister(&mq_policy_type); | |
1464 | dm_cache_policy_unregister(&default_policy_type); | |
1465 | ||
1466 | kmem_cache_destroy(mq_entry_cache); | |
1467 | } | |
1468 | ||
1469 | module_init(mq_init); | |
1470 | module_exit(mq_exit); | |
1471 | ||
1472 | MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>"); | |
1473 | MODULE_LICENSE("GPL"); | |
1474 | MODULE_DESCRIPTION("mq cache policy"); | |
1475 | ||
1476 | MODULE_ALIAS("dm-cache-default"); |