Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 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" |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 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 | /* |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 154 | * Checks to see if the queue is empty. |
| 155 | * FIXME: reduce cpu usage. |
| 156 | */ |
| 157 | static bool queue_empty(struct queue *q) |
| 158 | { |
| 159 | unsigned i; |
| 160 | |
| 161 | for (i = 0; i < NR_QUEUE_LEVELS; i++) |
| 162 | if (!list_empty(q->qs + i)) |
| 163 | return false; |
| 164 | |
| 165 | return true; |
| 166 | } |
| 167 | |
| 168 | /* |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 169 | * Insert an entry to the back of the given level. |
| 170 | */ |
| 171 | static void queue_push(struct queue *q, unsigned level, struct list_head *elt) |
| 172 | { |
| 173 | list_add_tail(elt, q->qs + level); |
| 174 | } |
| 175 | |
| 176 | static void queue_remove(struct list_head *elt) |
| 177 | { |
| 178 | list_del(elt); |
| 179 | } |
| 180 | |
| 181 | /* |
| 182 | * Shifts all regions down one level. This has no effect on the order of |
| 183 | * the queue. |
| 184 | */ |
| 185 | static void queue_shift_down(struct queue *q) |
| 186 | { |
| 187 | unsigned level; |
| 188 | |
| 189 | for (level = 1; level < NR_QUEUE_LEVELS; level++) |
| 190 | list_splice_init(q->qs + level, q->qs + level - 1); |
| 191 | } |
| 192 | |
| 193 | /* |
| 194 | * Gives us the oldest entry of the lowest popoulated level. If the first |
| 195 | * level is emptied then we shift down one level. |
| 196 | */ |
| 197 | static struct list_head *queue_pop(struct queue *q) |
| 198 | { |
| 199 | unsigned level; |
| 200 | struct list_head *r; |
| 201 | |
| 202 | for (level = 0; level < NR_QUEUE_LEVELS; level++) |
| 203 | if (!list_empty(q->qs + level)) { |
| 204 | r = q->qs[level].next; |
| 205 | list_del(r); |
| 206 | |
| 207 | /* have we just emptied the bottom level? */ |
| 208 | if (level == 0 && list_empty(q->qs)) |
| 209 | queue_shift_down(q); |
| 210 | |
| 211 | return r; |
| 212 | } |
| 213 | |
| 214 | return NULL; |
| 215 | } |
| 216 | |
| 217 | static struct list_head *list_pop(struct list_head *lh) |
| 218 | { |
| 219 | struct list_head *r = lh->next; |
| 220 | |
| 221 | BUG_ON(!r); |
| 222 | list_del_init(r); |
| 223 | |
| 224 | return r; |
| 225 | } |
| 226 | |
| 227 | /*----------------------------------------------------------------*/ |
| 228 | |
| 229 | /* |
| 230 | * Describes a cache entry. Used in both the cache and the pre_cache. |
| 231 | */ |
| 232 | struct entry { |
| 233 | struct hlist_node hlist; |
| 234 | struct list_head list; |
| 235 | dm_oblock_t oblock; |
| 236 | dm_cblock_t cblock; /* valid iff in_cache */ |
| 237 | |
| 238 | /* |
| 239 | * FIXME: pack these better |
| 240 | */ |
| 241 | bool in_cache:1; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 242 | bool dirty:1; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 243 | unsigned hit_count; |
| 244 | unsigned generation; |
| 245 | unsigned tick; |
| 246 | }; |
| 247 | |
| 248 | struct mq_policy { |
| 249 | struct dm_cache_policy policy; |
| 250 | |
| 251 | /* protects everything */ |
| 252 | struct mutex lock; |
| 253 | dm_cblock_t cache_size; |
| 254 | struct io_tracker tracker; |
| 255 | |
| 256 | /* |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 257 | * We maintain three queues of entries. The cache proper, |
| 258 | * consisting of a clean and dirty queue, contains the currently |
| 259 | * active mappings. Whereas the pre_cache tracks blocks that |
| 260 | * are being hit frequently and potential candidates for promotion |
| 261 | * to the cache. |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 262 | */ |
| 263 | struct queue pre_cache; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 264 | struct queue cache_clean; |
| 265 | struct queue cache_dirty; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 266 | |
| 267 | /* |
| 268 | * Keeps track of time, incremented by the core. We use this to |
| 269 | * avoid attributing multiple hits within the same tick. |
| 270 | * |
| 271 | * Access to tick_protected should be done with the spin lock held. |
| 272 | * It's copied to tick at the start of the map function (within the |
| 273 | * mutex). |
| 274 | */ |
| 275 | spinlock_t tick_lock; |
| 276 | unsigned tick_protected; |
| 277 | unsigned tick; |
| 278 | |
| 279 | /* |
| 280 | * A count of the number of times the map function has been called |
| 281 | * and found an entry in the pre_cache or cache. Currently used to |
| 282 | * calculate the generation. |
| 283 | */ |
| 284 | unsigned hit_count; |
| 285 | |
| 286 | /* |
| 287 | * A generation is a longish period that is used to trigger some |
| 288 | * book keeping effects. eg, decrementing hit counts on entries. |
| 289 | * This is needed to allow the cache to evolve as io patterns |
| 290 | * change. |
| 291 | */ |
| 292 | unsigned generation; |
| 293 | unsigned generation_period; /* in lookups (will probably change) */ |
| 294 | |
| 295 | /* |
| 296 | * Entries in the pre_cache whose hit count passes the promotion |
| 297 | * threshold move to the cache proper. Working out the correct |
| 298 | * value for the promotion_threshold is crucial to this policy. |
| 299 | */ |
| 300 | unsigned promote_threshold; |
| 301 | |
| 302 | /* |
| 303 | * We need cache_size entries for the cache, and choose to have |
| 304 | * cache_size entries for the pre_cache too. One motivation for |
| 305 | * using the same size is to make the hit counts directly |
| 306 | * comparable between pre_cache and cache. |
| 307 | */ |
| 308 | unsigned nr_entries; |
| 309 | unsigned nr_entries_allocated; |
| 310 | struct list_head free; |
| 311 | |
| 312 | /* |
| 313 | * Cache blocks may be unallocated. We store this info in a |
| 314 | * bitset. |
| 315 | */ |
| 316 | unsigned long *allocation_bitset; |
| 317 | unsigned nr_cblocks_allocated; |
| 318 | unsigned find_free_nr_words; |
| 319 | unsigned find_free_last_word; |
| 320 | |
| 321 | /* |
| 322 | * The hash table allows us to quickly find an entry by origin |
| 323 | * block. Both pre_cache and cache entries are in here. |
| 324 | */ |
| 325 | unsigned nr_buckets; |
| 326 | dm_block_t hash_bits; |
| 327 | struct hlist_head *table; |
| 328 | }; |
| 329 | |
| 330 | /*----------------------------------------------------------------*/ |
| 331 | /* Free/alloc mq cache entry structures. */ |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 332 | static void concat_queue(struct list_head *lh, struct queue *q) |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 333 | { |
| 334 | unsigned level; |
| 335 | |
| 336 | for (level = 0; level < NR_QUEUE_LEVELS; level++) |
| 337 | list_splice(q->qs + level, lh); |
| 338 | } |
| 339 | |
| 340 | static void free_entries(struct mq_policy *mq) |
| 341 | { |
| 342 | struct entry *e, *tmp; |
| 343 | |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 344 | concat_queue(&mq->free, &mq->pre_cache); |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 345 | concat_queue(&mq->free, &mq->cache_clean); |
| 346 | concat_queue(&mq->free, &mq->cache_dirty); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 347 | |
| 348 | list_for_each_entry_safe(e, tmp, &mq->free, list) |
| 349 | kmem_cache_free(mq_entry_cache, e); |
| 350 | } |
| 351 | |
| 352 | static int alloc_entries(struct mq_policy *mq, unsigned elts) |
| 353 | { |
| 354 | unsigned u = mq->nr_entries; |
| 355 | |
| 356 | INIT_LIST_HEAD(&mq->free); |
| 357 | mq->nr_entries_allocated = 0; |
| 358 | |
| 359 | while (u--) { |
| 360 | struct entry *e = kmem_cache_zalloc(mq_entry_cache, GFP_KERNEL); |
| 361 | |
| 362 | if (!e) { |
| 363 | free_entries(mq); |
| 364 | return -ENOMEM; |
| 365 | } |
| 366 | |
| 367 | |
| 368 | list_add(&e->list, &mq->free); |
| 369 | } |
| 370 | |
| 371 | return 0; |
| 372 | } |
| 373 | |
| 374 | /*----------------------------------------------------------------*/ |
| 375 | |
| 376 | /* |
| 377 | * Simple hash table implementation. Should replace with the standard hash |
| 378 | * table that's making its way upstream. |
| 379 | */ |
| 380 | static void hash_insert(struct mq_policy *mq, struct entry *e) |
| 381 | { |
| 382 | unsigned h = hash_64(from_oblock(e->oblock), mq->hash_bits); |
| 383 | |
| 384 | hlist_add_head(&e->hlist, mq->table + h); |
| 385 | } |
| 386 | |
| 387 | static struct entry *hash_lookup(struct mq_policy *mq, dm_oblock_t oblock) |
| 388 | { |
| 389 | unsigned h = hash_64(from_oblock(oblock), mq->hash_bits); |
| 390 | struct hlist_head *bucket = mq->table + h; |
| 391 | struct entry *e; |
| 392 | |
| 393 | hlist_for_each_entry(e, bucket, hlist) |
| 394 | if (e->oblock == oblock) { |
| 395 | hlist_del(&e->hlist); |
| 396 | hlist_add_head(&e->hlist, bucket); |
| 397 | return e; |
| 398 | } |
| 399 | |
| 400 | return NULL; |
| 401 | } |
| 402 | |
| 403 | static void hash_remove(struct entry *e) |
| 404 | { |
| 405 | hlist_del(&e->hlist); |
| 406 | } |
| 407 | |
| 408 | /*----------------------------------------------------------------*/ |
| 409 | |
| 410 | /* |
| 411 | * Allocates a new entry structure. The memory is allocated in one lump, |
| 412 | * so we just handing it out here. Returns NULL if all entries have |
| 413 | * already been allocated. Cannot fail otherwise. |
| 414 | */ |
| 415 | static struct entry *alloc_entry(struct mq_policy *mq) |
| 416 | { |
| 417 | struct entry *e; |
| 418 | |
| 419 | if (mq->nr_entries_allocated >= mq->nr_entries) { |
| 420 | BUG_ON(!list_empty(&mq->free)); |
| 421 | return NULL; |
| 422 | } |
| 423 | |
| 424 | e = list_entry(list_pop(&mq->free), struct entry, list); |
| 425 | INIT_LIST_HEAD(&e->list); |
| 426 | INIT_HLIST_NODE(&e->hlist); |
| 427 | |
| 428 | mq->nr_entries_allocated++; |
| 429 | return e; |
| 430 | } |
| 431 | |
| 432 | /*----------------------------------------------------------------*/ |
| 433 | |
| 434 | /* |
| 435 | * Mark cache blocks allocated or not in the bitset. |
| 436 | */ |
| 437 | static void alloc_cblock(struct mq_policy *mq, dm_cblock_t cblock) |
| 438 | { |
| 439 | BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size)); |
| 440 | BUG_ON(test_bit(from_cblock(cblock), mq->allocation_bitset)); |
| 441 | |
| 442 | set_bit(from_cblock(cblock), mq->allocation_bitset); |
| 443 | mq->nr_cblocks_allocated++; |
| 444 | } |
| 445 | |
| 446 | static void free_cblock(struct mq_policy *mq, dm_cblock_t cblock) |
| 447 | { |
| 448 | BUG_ON(from_cblock(cblock) > from_cblock(mq->cache_size)); |
| 449 | BUG_ON(!test_bit(from_cblock(cblock), mq->allocation_bitset)); |
| 450 | |
| 451 | clear_bit(from_cblock(cblock), mq->allocation_bitset); |
| 452 | mq->nr_cblocks_allocated--; |
| 453 | } |
| 454 | |
| 455 | static bool any_free_cblocks(struct mq_policy *mq) |
| 456 | { |
| 457 | return mq->nr_cblocks_allocated < from_cblock(mq->cache_size); |
| 458 | } |
| 459 | |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 460 | static bool any_clean_cblocks(struct mq_policy *mq) |
| 461 | { |
| 462 | return !queue_empty(&mq->cache_clean); |
| 463 | } |
| 464 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 465 | /* |
| 466 | * Fills result out with a cache block that isn't in use, or return |
| 467 | * -ENOSPC. This does _not_ mark the cblock as allocated, the caller is |
| 468 | * reponsible for that. |
| 469 | */ |
| 470 | static int __find_free_cblock(struct mq_policy *mq, unsigned begin, unsigned end, |
| 471 | dm_cblock_t *result, unsigned *last_word) |
| 472 | { |
| 473 | int r = -ENOSPC; |
| 474 | unsigned w; |
| 475 | |
| 476 | for (w = begin; w < end; w++) { |
| 477 | /* |
| 478 | * ffz is undefined if no zero exists |
| 479 | */ |
| 480 | if (mq->allocation_bitset[w] != ~0UL) { |
| 481 | *last_word = w; |
| 482 | *result = to_cblock((w * BITS_PER_LONG) + ffz(mq->allocation_bitset[w])); |
| 483 | if (from_cblock(*result) < from_cblock(mq->cache_size)) |
| 484 | r = 0; |
| 485 | |
| 486 | break; |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | return r; |
| 491 | } |
| 492 | |
| 493 | static int find_free_cblock(struct mq_policy *mq, dm_cblock_t *result) |
| 494 | { |
| 495 | int r; |
| 496 | |
| 497 | if (!any_free_cblocks(mq)) |
| 498 | return -ENOSPC; |
| 499 | |
| 500 | r = __find_free_cblock(mq, mq->find_free_last_word, mq->find_free_nr_words, result, &mq->find_free_last_word); |
| 501 | if (r == -ENOSPC && mq->find_free_last_word) |
| 502 | r = __find_free_cblock(mq, 0, mq->find_free_last_word, result, &mq->find_free_last_word); |
| 503 | |
| 504 | return r; |
| 505 | } |
| 506 | |
| 507 | /*----------------------------------------------------------------*/ |
| 508 | |
| 509 | /* |
| 510 | * Now we get to the meat of the policy. This section deals with deciding |
| 511 | * when to to add entries to the pre_cache and cache, and move between |
| 512 | * them. |
| 513 | */ |
| 514 | |
| 515 | /* |
| 516 | * The queue level is based on the log2 of the hit count. |
| 517 | */ |
| 518 | static unsigned queue_level(struct entry *e) |
| 519 | { |
| 520 | return min((unsigned) ilog2(e->hit_count), NR_QUEUE_LEVELS - 1u); |
| 521 | } |
| 522 | |
| 523 | /* |
| 524 | * Inserts the entry into the pre_cache or the cache. Ensures the cache |
| 525 | * block is marked as allocated if necc. Inserts into the hash table. Sets the |
| 526 | * tick which records when the entry was last moved about. |
| 527 | */ |
| 528 | static void push(struct mq_policy *mq, struct entry *e) |
| 529 | { |
| 530 | e->tick = mq->tick; |
| 531 | hash_insert(mq, e); |
| 532 | |
| 533 | if (e->in_cache) { |
| 534 | alloc_cblock(mq, e->cblock); |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 535 | queue_push(e->dirty ? &mq->cache_dirty : &mq->cache_clean, |
| 536 | queue_level(e), &e->list); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 537 | } else |
| 538 | queue_push(&mq->pre_cache, queue_level(e), &e->list); |
| 539 | } |
| 540 | |
| 541 | /* |
| 542 | * Removes an entry from pre_cache or cache. Removes from the hash table. |
| 543 | * Frees off the cache block if necc. |
| 544 | */ |
| 545 | static void del(struct mq_policy *mq, struct entry *e) |
| 546 | { |
| 547 | queue_remove(&e->list); |
| 548 | hash_remove(e); |
| 549 | if (e->in_cache) |
| 550 | free_cblock(mq, e->cblock); |
| 551 | } |
| 552 | |
| 553 | /* |
| 554 | * Like del, except it removes the first entry in the queue (ie. the least |
| 555 | * recently used). |
| 556 | */ |
| 557 | static struct entry *pop(struct mq_policy *mq, struct queue *q) |
| 558 | { |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 559 | struct entry *e; |
| 560 | struct list_head *h = queue_pop(q); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 561 | |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 562 | if (!h) |
| 563 | return NULL; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 564 | |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 565 | e = container_of(h, struct entry, list); |
| 566 | hash_remove(e); |
| 567 | if (e->in_cache) |
| 568 | free_cblock(mq, e->cblock); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 569 | |
| 570 | return e; |
| 571 | } |
| 572 | |
| 573 | /* |
| 574 | * Has this entry already been updated? |
| 575 | */ |
| 576 | static bool updated_this_tick(struct mq_policy *mq, struct entry *e) |
| 577 | { |
| 578 | return mq->tick == e->tick; |
| 579 | } |
| 580 | |
| 581 | /* |
| 582 | * The promotion threshold is adjusted every generation. As are the counts |
| 583 | * of the entries. |
| 584 | * |
| 585 | * At the moment the threshold is taken by averaging the hit counts of some |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 586 | * of the entries in the cache (the first 20 entries across all levels in |
| 587 | * ascending order, giving preference to the clean entries at each level). |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 588 | * |
| 589 | * We can be much cleverer than this though. For example, each promotion |
| 590 | * could bump up the threshold helping to prevent churn. Much more to do |
| 591 | * here. |
| 592 | */ |
| 593 | |
| 594 | #define MAX_TO_AVERAGE 20 |
| 595 | |
| 596 | static void check_generation(struct mq_policy *mq) |
| 597 | { |
| 598 | unsigned total = 0, nr = 0, count = 0, level; |
| 599 | struct list_head *head; |
| 600 | struct entry *e; |
| 601 | |
| 602 | if ((mq->hit_count >= mq->generation_period) && |
| 603 | (mq->nr_cblocks_allocated == from_cblock(mq->cache_size))) { |
| 604 | |
| 605 | mq->hit_count = 0; |
| 606 | mq->generation++; |
| 607 | |
| 608 | for (level = 0; level < NR_QUEUE_LEVELS && count < MAX_TO_AVERAGE; level++) { |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 609 | head = mq->cache_clean.qs + level; |
| 610 | list_for_each_entry(e, head, list) { |
| 611 | nr++; |
| 612 | total += e->hit_count; |
| 613 | |
| 614 | if (++count >= MAX_TO_AVERAGE) |
| 615 | break; |
| 616 | } |
| 617 | |
| 618 | head = mq->cache_dirty.qs + level; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 619 | list_for_each_entry(e, head, list) { |
| 620 | nr++; |
| 621 | total += e->hit_count; |
| 622 | |
| 623 | if (++count >= MAX_TO_AVERAGE) |
| 624 | break; |
| 625 | } |
| 626 | } |
| 627 | |
| 628 | mq->promote_threshold = nr ? total / nr : 1; |
| 629 | if (mq->promote_threshold * nr < total) |
| 630 | mq->promote_threshold++; |
| 631 | } |
| 632 | } |
| 633 | |
| 634 | /* |
| 635 | * Whenever we use an entry we bump up it's hit counter, and push it to the |
| 636 | * back to it's current level. |
| 637 | */ |
| 638 | static void requeue_and_update_tick(struct mq_policy *mq, struct entry *e) |
| 639 | { |
| 640 | if (updated_this_tick(mq, e)) |
| 641 | return; |
| 642 | |
| 643 | e->hit_count++; |
| 644 | mq->hit_count++; |
| 645 | check_generation(mq); |
| 646 | |
| 647 | /* generation adjustment, to stop the counts increasing forever. */ |
| 648 | /* FIXME: divide? */ |
| 649 | /* e->hit_count -= min(e->hit_count - 1, mq->generation - e->generation); */ |
| 650 | e->generation = mq->generation; |
| 651 | |
| 652 | del(mq, e); |
| 653 | push(mq, e); |
| 654 | } |
| 655 | |
| 656 | /* |
| 657 | * Demote the least recently used entry from the cache to the pre_cache. |
| 658 | * Returns the new cache entry to use, and the old origin block it was |
| 659 | * mapped to. |
| 660 | * |
| 661 | * We drop the hit count on the demoted entry back to 1 to stop it bouncing |
| 662 | * straight back into the cache if it's subsequently hit. There are |
| 663 | * various options here, and more experimentation would be good: |
| 664 | * |
| 665 | * - just forget about the demoted entry completely (ie. don't insert it |
| 666 | into the pre_cache). |
| 667 | * - divide the hit count rather that setting to some hard coded value. |
| 668 | * - set the hit count to a hard coded value other than 1, eg, is it better |
| 669 | * if it goes in at level 2? |
| 670 | */ |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 671 | static int demote_cblock(struct mq_policy *mq, dm_oblock_t *oblock, dm_cblock_t *cblock) |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 672 | { |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 673 | struct entry *demoted = pop(mq, &mq->cache_clean); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 674 | |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 675 | if (!demoted) |
| 676 | /* |
| 677 | * We could get a block from mq->cache_dirty, but that |
| 678 | * would add extra latency to the triggering bio as it |
| 679 | * waits for the writeback. Better to not promote this |
| 680 | * time and hope there's a clean block next time this block |
| 681 | * is hit. |
| 682 | */ |
| 683 | return -ENOSPC; |
| 684 | |
| 685 | *cblock = demoted->cblock; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 686 | *oblock = demoted->oblock; |
| 687 | demoted->in_cache = false; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 688 | demoted->dirty = false; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 689 | demoted->hit_count = 1; |
| 690 | push(mq, demoted); |
| 691 | |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 692 | return 0; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 693 | } |
| 694 | |
| 695 | /* |
| 696 | * We modify the basic promotion_threshold depending on the specific io. |
| 697 | * |
| 698 | * If the origin block has been discarded then there's no cost to copy it |
| 699 | * to the cache. |
| 700 | * |
| 701 | * We bias towards reads, since they can be demoted at no cost if they |
| 702 | * haven't been dirtied. |
| 703 | */ |
| 704 | #define DISCARDED_PROMOTE_THRESHOLD 1 |
| 705 | #define READ_PROMOTE_THRESHOLD 4 |
| 706 | #define WRITE_PROMOTE_THRESHOLD 8 |
| 707 | |
| 708 | static unsigned adjusted_promote_threshold(struct mq_policy *mq, |
| 709 | bool discarded_oblock, int data_dir) |
| 710 | { |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 711 | if (data_dir == READ) |
| 712 | return mq->promote_threshold + READ_PROMOTE_THRESHOLD; |
| 713 | |
| 714 | if (discarded_oblock && (any_free_cblocks(mq) || any_clean_cblocks(mq))) { |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 715 | /* |
| 716 | * We don't need to do any copying at all, so give this a |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 717 | * very low threshold. |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 718 | */ |
| 719 | return DISCARDED_PROMOTE_THRESHOLD; |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 720 | } |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 721 | |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 722 | return mq->promote_threshold + WRITE_PROMOTE_THRESHOLD; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 723 | } |
| 724 | |
| 725 | static bool should_promote(struct mq_policy *mq, struct entry *e, |
| 726 | bool discarded_oblock, int data_dir) |
| 727 | { |
| 728 | return e->hit_count >= |
| 729 | adjusted_promote_threshold(mq, discarded_oblock, data_dir); |
| 730 | } |
| 731 | |
| 732 | static int cache_entry_found(struct mq_policy *mq, |
| 733 | struct entry *e, |
| 734 | struct policy_result *result) |
| 735 | { |
| 736 | requeue_and_update_tick(mq, e); |
| 737 | |
| 738 | if (e->in_cache) { |
| 739 | result->op = POLICY_HIT; |
| 740 | result->cblock = e->cblock; |
| 741 | } |
| 742 | |
| 743 | return 0; |
| 744 | } |
| 745 | |
| 746 | /* |
Joe Thornber | 0184b44 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 747 | * Moves an entry from the pre_cache to the cache. The main work is |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 748 | * finding which cache block to use. |
| 749 | */ |
| 750 | static int pre_cache_to_cache(struct mq_policy *mq, struct entry *e, |
| 751 | struct policy_result *result) |
| 752 | { |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 753 | int r; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 754 | dm_cblock_t cblock; |
| 755 | |
| 756 | if (find_free_cblock(mq, &cblock) == -ENOSPC) { |
| 757 | result->op = POLICY_REPLACE; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 758 | r = demote_cblock(mq, &result->old_oblock, &cblock); |
| 759 | if (r) { |
| 760 | result->op = POLICY_MISS; |
| 761 | return 0; |
| 762 | } |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 763 | } else |
| 764 | result->op = POLICY_NEW; |
| 765 | |
| 766 | result->cblock = e->cblock = cblock; |
| 767 | |
| 768 | del(mq, e); |
| 769 | e->in_cache = true; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 770 | e->dirty = false; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 771 | push(mq, e); |
| 772 | |
| 773 | return 0; |
| 774 | } |
| 775 | |
| 776 | static int pre_cache_entry_found(struct mq_policy *mq, struct entry *e, |
| 777 | bool can_migrate, bool discarded_oblock, |
| 778 | int data_dir, struct policy_result *result) |
| 779 | { |
| 780 | int r = 0; |
| 781 | bool updated = updated_this_tick(mq, e); |
| 782 | |
| 783 | requeue_and_update_tick(mq, e); |
| 784 | |
| 785 | if ((!discarded_oblock && updated) || |
| 786 | !should_promote(mq, e, discarded_oblock, data_dir)) |
| 787 | result->op = POLICY_MISS; |
| 788 | else if (!can_migrate) |
| 789 | r = -EWOULDBLOCK; |
| 790 | else |
| 791 | r = pre_cache_to_cache(mq, e, result); |
| 792 | |
| 793 | return r; |
| 794 | } |
| 795 | |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 796 | static void insert_entry_in_pre_cache(struct mq_policy *mq, |
| 797 | struct entry *e, dm_oblock_t oblock) |
| 798 | { |
| 799 | e->in_cache = false; |
| 800 | e->dirty = false; |
| 801 | e->oblock = oblock; |
| 802 | e->hit_count = 1; |
| 803 | e->generation = mq->generation; |
| 804 | push(mq, e); |
| 805 | } |
| 806 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 807 | static void insert_in_pre_cache(struct mq_policy *mq, |
| 808 | dm_oblock_t oblock) |
| 809 | { |
| 810 | struct entry *e = alloc_entry(mq); |
| 811 | |
| 812 | if (!e) |
| 813 | /* |
| 814 | * There's no spare entry structure, so we grab the least |
| 815 | * used one from the pre_cache. |
| 816 | */ |
| 817 | e = pop(mq, &mq->pre_cache); |
| 818 | |
| 819 | if (unlikely(!e)) { |
| 820 | DMWARN("couldn't pop from pre cache"); |
| 821 | return; |
| 822 | } |
| 823 | |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 824 | insert_entry_in_pre_cache(mq, e, oblock); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 825 | } |
| 826 | |
| 827 | static void insert_in_cache(struct mq_policy *mq, dm_oblock_t oblock, |
| 828 | struct policy_result *result) |
| 829 | { |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 830 | int r; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 831 | struct entry *e; |
| 832 | dm_cblock_t cblock; |
| 833 | |
| 834 | if (find_free_cblock(mq, &cblock) == -ENOSPC) { |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 835 | r = demote_cblock(mq, &result->old_oblock, &cblock); |
| 836 | if (unlikely(r)) { |
| 837 | result->op = POLICY_MISS; |
| 838 | insert_in_pre_cache(mq, oblock); |
| 839 | return; |
| 840 | } |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 841 | |
Joe Thornber | c86c307 | 2013-10-24 14:10:28 -0400 | [diff] [blame^] | 842 | /* |
| 843 | * This will always succeed, since we've just demoted. |
| 844 | */ |
| 845 | e = pop(mq, &mq->pre_cache); |
| 846 | result->op = POLICY_REPLACE; |
| 847 | |
| 848 | } else { |
| 849 | e = alloc_entry(mq); |
| 850 | if (unlikely(!e)) |
| 851 | e = pop(mq, &mq->pre_cache); |
| 852 | |
| 853 | if (unlikely(!e)) { |
| 854 | result->op = POLICY_MISS; |
| 855 | return; |
| 856 | } |
| 857 | |
| 858 | result->op = POLICY_NEW; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 859 | } |
| 860 | |
| 861 | e->oblock = oblock; |
| 862 | e->cblock = cblock; |
| 863 | e->in_cache = true; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 864 | e->dirty = false; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 865 | e->hit_count = 1; |
| 866 | e->generation = mq->generation; |
| 867 | push(mq, e); |
| 868 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 869 | result->cblock = e->cblock; |
| 870 | } |
| 871 | |
| 872 | static int no_entry_found(struct mq_policy *mq, dm_oblock_t oblock, |
| 873 | bool can_migrate, bool discarded_oblock, |
| 874 | int data_dir, struct policy_result *result) |
| 875 | { |
| 876 | if (adjusted_promote_threshold(mq, discarded_oblock, data_dir) == 1) { |
| 877 | if (can_migrate) |
| 878 | insert_in_cache(mq, oblock, result); |
| 879 | else |
| 880 | return -EWOULDBLOCK; |
| 881 | } else { |
| 882 | insert_in_pre_cache(mq, oblock); |
| 883 | result->op = POLICY_MISS; |
| 884 | } |
| 885 | |
| 886 | return 0; |
| 887 | } |
| 888 | |
| 889 | /* |
| 890 | * Looks the oblock up in the hash table, then decides whether to put in |
| 891 | * pre_cache, or cache etc. |
| 892 | */ |
| 893 | static int map(struct mq_policy *mq, dm_oblock_t oblock, |
| 894 | bool can_migrate, bool discarded_oblock, |
| 895 | int data_dir, struct policy_result *result) |
| 896 | { |
| 897 | int r = 0; |
| 898 | struct entry *e = hash_lookup(mq, oblock); |
| 899 | |
| 900 | if (e && e->in_cache) |
| 901 | r = cache_entry_found(mq, e, result); |
| 902 | else if (iot_pattern(&mq->tracker) == PATTERN_SEQUENTIAL) |
| 903 | result->op = POLICY_MISS; |
| 904 | else if (e) |
| 905 | r = pre_cache_entry_found(mq, e, can_migrate, discarded_oblock, |
| 906 | data_dir, result); |
| 907 | else |
| 908 | r = no_entry_found(mq, oblock, can_migrate, discarded_oblock, |
| 909 | data_dir, result); |
| 910 | |
| 911 | if (r == -EWOULDBLOCK) |
| 912 | result->op = POLICY_MISS; |
| 913 | |
| 914 | return r; |
| 915 | } |
| 916 | |
| 917 | /*----------------------------------------------------------------*/ |
| 918 | |
| 919 | /* |
| 920 | * Public interface, via the policy struct. See dm-cache-policy.h for a |
| 921 | * description of these. |
| 922 | */ |
| 923 | |
| 924 | static struct mq_policy *to_mq_policy(struct dm_cache_policy *p) |
| 925 | { |
| 926 | return container_of(p, struct mq_policy, policy); |
| 927 | } |
| 928 | |
| 929 | static void mq_destroy(struct dm_cache_policy *p) |
| 930 | { |
| 931 | struct mq_policy *mq = to_mq_policy(p); |
| 932 | |
| 933 | free_bitset(mq->allocation_bitset); |
| 934 | kfree(mq->table); |
| 935 | free_entries(mq); |
| 936 | kfree(mq); |
| 937 | } |
| 938 | |
| 939 | static void copy_tick(struct mq_policy *mq) |
| 940 | { |
| 941 | unsigned long flags; |
| 942 | |
| 943 | spin_lock_irqsave(&mq->tick_lock, flags); |
| 944 | mq->tick = mq->tick_protected; |
| 945 | spin_unlock_irqrestore(&mq->tick_lock, flags); |
| 946 | } |
| 947 | |
| 948 | static int mq_map(struct dm_cache_policy *p, dm_oblock_t oblock, |
| 949 | bool can_block, bool can_migrate, bool discarded_oblock, |
| 950 | struct bio *bio, struct policy_result *result) |
| 951 | { |
| 952 | int r; |
| 953 | struct mq_policy *mq = to_mq_policy(p); |
| 954 | |
| 955 | result->op = POLICY_MISS; |
| 956 | |
| 957 | if (can_block) |
| 958 | mutex_lock(&mq->lock); |
| 959 | else if (!mutex_trylock(&mq->lock)) |
| 960 | return -EWOULDBLOCK; |
| 961 | |
| 962 | copy_tick(mq); |
| 963 | |
| 964 | iot_examine_bio(&mq->tracker, bio); |
| 965 | r = map(mq, oblock, can_migrate, discarded_oblock, |
| 966 | bio_data_dir(bio), result); |
| 967 | |
| 968 | mutex_unlock(&mq->lock); |
| 969 | |
| 970 | return r; |
| 971 | } |
| 972 | |
| 973 | static int mq_lookup(struct dm_cache_policy *p, dm_oblock_t oblock, dm_cblock_t *cblock) |
| 974 | { |
| 975 | int r; |
| 976 | struct mq_policy *mq = to_mq_policy(p); |
| 977 | struct entry *e; |
| 978 | |
| 979 | if (!mutex_trylock(&mq->lock)) |
| 980 | return -EWOULDBLOCK; |
| 981 | |
| 982 | e = hash_lookup(mq, oblock); |
| 983 | if (e && e->in_cache) { |
| 984 | *cblock = e->cblock; |
| 985 | r = 0; |
| 986 | } else |
| 987 | r = -ENOENT; |
| 988 | |
| 989 | mutex_unlock(&mq->lock); |
| 990 | |
| 991 | return r; |
| 992 | } |
| 993 | |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 994 | /* |
| 995 | * FIXME: __mq_set_clear_dirty can block due to mutex. |
| 996 | * Ideally a policy should not block in functions called |
| 997 | * from the map() function. Explore using RCU. |
| 998 | */ |
| 999 | static void __mq_set_clear_dirty(struct dm_cache_policy *p, dm_oblock_t oblock, bool set) |
| 1000 | { |
| 1001 | struct mq_policy *mq = to_mq_policy(p); |
| 1002 | struct entry *e; |
| 1003 | |
| 1004 | mutex_lock(&mq->lock); |
| 1005 | e = hash_lookup(mq, oblock); |
| 1006 | if (!e) |
| 1007 | DMWARN("__mq_set_clear_dirty called for a block that isn't in the cache"); |
| 1008 | else { |
| 1009 | BUG_ON(!e->in_cache); |
| 1010 | |
| 1011 | del(mq, e); |
| 1012 | e->dirty = set; |
| 1013 | push(mq, e); |
| 1014 | } |
| 1015 | mutex_unlock(&mq->lock); |
| 1016 | } |
| 1017 | |
| 1018 | static void mq_set_dirty(struct dm_cache_policy *p, dm_oblock_t oblock) |
| 1019 | { |
| 1020 | __mq_set_clear_dirty(p, oblock, true); |
| 1021 | } |
| 1022 | |
| 1023 | static void mq_clear_dirty(struct dm_cache_policy *p, dm_oblock_t oblock) |
| 1024 | { |
| 1025 | __mq_set_clear_dirty(p, oblock, false); |
| 1026 | } |
| 1027 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1028 | static int mq_load_mapping(struct dm_cache_policy *p, |
| 1029 | dm_oblock_t oblock, dm_cblock_t cblock, |
| 1030 | uint32_t hint, bool hint_valid) |
| 1031 | { |
| 1032 | struct mq_policy *mq = to_mq_policy(p); |
| 1033 | struct entry *e; |
| 1034 | |
| 1035 | e = alloc_entry(mq); |
| 1036 | if (!e) |
| 1037 | return -ENOMEM; |
| 1038 | |
| 1039 | e->cblock = cblock; |
| 1040 | e->oblock = oblock; |
| 1041 | e->in_cache = true; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1042 | e->dirty = false; /* this gets corrected in a minute */ |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1043 | e->hit_count = hint_valid ? hint : 1; |
| 1044 | e->generation = mq->generation; |
| 1045 | push(mq, e); |
| 1046 | |
| 1047 | return 0; |
| 1048 | } |
| 1049 | |
| 1050 | static int mq_walk_mappings(struct dm_cache_policy *p, policy_walk_fn fn, |
| 1051 | void *context) |
| 1052 | { |
| 1053 | struct mq_policy *mq = to_mq_policy(p); |
| 1054 | int r = 0; |
| 1055 | struct entry *e; |
| 1056 | unsigned level; |
| 1057 | |
| 1058 | mutex_lock(&mq->lock); |
| 1059 | |
| 1060 | for (level = 0; level < NR_QUEUE_LEVELS; level++) |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1061 | list_for_each_entry(e, &mq->cache_clean.qs[level], list) { |
| 1062 | r = fn(context, e->cblock, e->oblock, e->hit_count); |
| 1063 | if (r) |
| 1064 | goto out; |
| 1065 | } |
| 1066 | |
| 1067 | for (level = 0; level < NR_QUEUE_LEVELS; level++) |
| 1068 | list_for_each_entry(e, &mq->cache_dirty.qs[level], list) { |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1069 | r = fn(context, e->cblock, e->oblock, e->hit_count); |
| 1070 | if (r) |
| 1071 | goto out; |
| 1072 | } |
| 1073 | |
| 1074 | out: |
| 1075 | mutex_unlock(&mq->lock); |
| 1076 | |
| 1077 | return r; |
| 1078 | } |
| 1079 | |
Geert Uytterhoeven | b936bf8 | 2013-07-26 09:57:31 +0200 | [diff] [blame] | 1080 | static void mq_remove_mapping(struct dm_cache_policy *p, dm_oblock_t oblock) |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1081 | { |
Geert Uytterhoeven | b936bf8 | 2013-07-26 09:57:31 +0200 | [diff] [blame] | 1082 | struct mq_policy *mq = to_mq_policy(p); |
| 1083 | struct entry *e; |
| 1084 | |
| 1085 | mutex_lock(&mq->lock); |
| 1086 | |
| 1087 | e = hash_lookup(mq, oblock); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1088 | |
| 1089 | BUG_ON(!e || !e->in_cache); |
| 1090 | |
| 1091 | del(mq, e); |
| 1092 | e->in_cache = false; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1093 | e->dirty = false; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1094 | push(mq, e); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1095 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1096 | mutex_unlock(&mq->lock); |
| 1097 | } |
| 1098 | |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1099 | static int __mq_writeback_work(struct mq_policy *mq, dm_oblock_t *oblock, |
| 1100 | dm_cblock_t *cblock) |
| 1101 | { |
| 1102 | struct entry *e = pop(mq, &mq->cache_dirty); |
| 1103 | |
| 1104 | if (!e) |
| 1105 | return -ENODATA; |
| 1106 | |
| 1107 | *oblock = e->oblock; |
| 1108 | *cblock = e->cblock; |
| 1109 | e->dirty = false; |
| 1110 | push(mq, e); |
| 1111 | |
| 1112 | return 0; |
| 1113 | } |
| 1114 | |
| 1115 | static int mq_writeback_work(struct dm_cache_policy *p, dm_oblock_t *oblock, |
| 1116 | dm_cblock_t *cblock) |
| 1117 | { |
| 1118 | int r; |
| 1119 | struct mq_policy *mq = to_mq_policy(p); |
| 1120 | |
| 1121 | mutex_lock(&mq->lock); |
| 1122 | r = __mq_writeback_work(mq, oblock, cblock); |
| 1123 | mutex_unlock(&mq->lock); |
| 1124 | |
| 1125 | return r; |
| 1126 | } |
| 1127 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1128 | static void force_mapping(struct mq_policy *mq, |
| 1129 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) |
| 1130 | { |
| 1131 | struct entry *e = hash_lookup(mq, current_oblock); |
| 1132 | |
| 1133 | BUG_ON(!e || !e->in_cache); |
| 1134 | |
| 1135 | del(mq, e); |
| 1136 | e->oblock = new_oblock; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1137 | e->dirty = true; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1138 | push(mq, e); |
| 1139 | } |
| 1140 | |
| 1141 | static void mq_force_mapping(struct dm_cache_policy *p, |
| 1142 | dm_oblock_t current_oblock, dm_oblock_t new_oblock) |
| 1143 | { |
| 1144 | struct mq_policy *mq = to_mq_policy(p); |
| 1145 | |
| 1146 | mutex_lock(&mq->lock); |
| 1147 | force_mapping(mq, current_oblock, new_oblock); |
| 1148 | mutex_unlock(&mq->lock); |
| 1149 | } |
| 1150 | |
| 1151 | static dm_cblock_t mq_residency(struct dm_cache_policy *p) |
| 1152 | { |
Joe Thornber | 99ba2ae | 2013-10-21 11:44:57 +0100 | [diff] [blame] | 1153 | dm_cblock_t r; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1154 | struct mq_policy *mq = to_mq_policy(p); |
| 1155 | |
Joe Thornber | 99ba2ae | 2013-10-21 11:44:57 +0100 | [diff] [blame] | 1156 | mutex_lock(&mq->lock); |
| 1157 | r = to_cblock(mq->nr_cblocks_allocated); |
| 1158 | mutex_unlock(&mq->lock); |
| 1159 | |
| 1160 | return r; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1161 | } |
| 1162 | |
| 1163 | static void mq_tick(struct dm_cache_policy *p) |
| 1164 | { |
| 1165 | struct mq_policy *mq = to_mq_policy(p); |
| 1166 | unsigned long flags; |
| 1167 | |
| 1168 | spin_lock_irqsave(&mq->tick_lock, flags); |
| 1169 | mq->tick_protected++; |
| 1170 | spin_unlock_irqrestore(&mq->tick_lock, flags); |
| 1171 | } |
| 1172 | |
| 1173 | static int mq_set_config_value(struct dm_cache_policy *p, |
| 1174 | const char *key, const char *value) |
| 1175 | { |
| 1176 | struct mq_policy *mq = to_mq_policy(p); |
| 1177 | enum io_pattern pattern; |
| 1178 | unsigned long tmp; |
| 1179 | |
| 1180 | if (!strcasecmp(key, "random_threshold")) |
| 1181 | pattern = PATTERN_RANDOM; |
| 1182 | else if (!strcasecmp(key, "sequential_threshold")) |
| 1183 | pattern = PATTERN_SEQUENTIAL; |
| 1184 | else |
| 1185 | return -EINVAL; |
| 1186 | |
| 1187 | if (kstrtoul(value, 10, &tmp)) |
| 1188 | return -EINVAL; |
| 1189 | |
| 1190 | mq->tracker.thresholds[pattern] = tmp; |
| 1191 | |
| 1192 | return 0; |
| 1193 | } |
| 1194 | |
| 1195 | static int mq_emit_config_values(struct dm_cache_policy *p, char *result, unsigned maxlen) |
| 1196 | { |
| 1197 | ssize_t sz = 0; |
| 1198 | struct mq_policy *mq = to_mq_policy(p); |
| 1199 | |
| 1200 | DMEMIT("4 random_threshold %u sequential_threshold %u", |
| 1201 | mq->tracker.thresholds[PATTERN_RANDOM], |
| 1202 | mq->tracker.thresholds[PATTERN_SEQUENTIAL]); |
| 1203 | |
| 1204 | return 0; |
| 1205 | } |
| 1206 | |
| 1207 | /* Init the policy plugin interface function pointers. */ |
| 1208 | static void init_policy_functions(struct mq_policy *mq) |
| 1209 | { |
| 1210 | mq->policy.destroy = mq_destroy; |
| 1211 | mq->policy.map = mq_map; |
| 1212 | mq->policy.lookup = mq_lookup; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1213 | mq->policy.set_dirty = mq_set_dirty; |
| 1214 | mq->policy.clear_dirty = mq_clear_dirty; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1215 | mq->policy.load_mapping = mq_load_mapping; |
| 1216 | mq->policy.walk_mappings = mq_walk_mappings; |
| 1217 | mq->policy.remove_mapping = mq_remove_mapping; |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1218 | mq->policy.writeback_work = mq_writeback_work; |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1219 | mq->policy.force_mapping = mq_force_mapping; |
| 1220 | mq->policy.residency = mq_residency; |
| 1221 | mq->policy.tick = mq_tick; |
| 1222 | mq->policy.emit_config_values = mq_emit_config_values; |
| 1223 | mq->policy.set_config_value = mq_set_config_value; |
| 1224 | } |
| 1225 | |
| 1226 | static struct dm_cache_policy *mq_create(dm_cblock_t cache_size, |
| 1227 | sector_t origin_size, |
| 1228 | sector_t cache_block_size) |
| 1229 | { |
| 1230 | int r; |
| 1231 | struct mq_policy *mq = kzalloc(sizeof(*mq), GFP_KERNEL); |
| 1232 | |
| 1233 | if (!mq) |
| 1234 | return NULL; |
| 1235 | |
| 1236 | init_policy_functions(mq); |
| 1237 | iot_init(&mq->tracker, SEQUENTIAL_THRESHOLD_DEFAULT, RANDOM_THRESHOLD_DEFAULT); |
| 1238 | |
| 1239 | mq->cache_size = cache_size; |
| 1240 | mq->tick_protected = 0; |
| 1241 | mq->tick = 0; |
| 1242 | mq->hit_count = 0; |
| 1243 | mq->generation = 0; |
| 1244 | mq->promote_threshold = 0; |
| 1245 | mutex_init(&mq->lock); |
| 1246 | spin_lock_init(&mq->tick_lock); |
| 1247 | mq->find_free_nr_words = dm_div_up(from_cblock(mq->cache_size), BITS_PER_LONG); |
| 1248 | mq->find_free_last_word = 0; |
| 1249 | |
| 1250 | queue_init(&mq->pre_cache); |
Joe Thornber | 01911c1 | 2013-10-24 14:10:28 -0400 | [diff] [blame] | 1251 | queue_init(&mq->cache_clean); |
| 1252 | queue_init(&mq->cache_dirty); |
| 1253 | |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1254 | mq->generation_period = max((unsigned) from_cblock(cache_size), 1024U); |
| 1255 | |
| 1256 | mq->nr_entries = 2 * from_cblock(cache_size); |
| 1257 | r = alloc_entries(mq, mq->nr_entries); |
| 1258 | if (r) |
| 1259 | goto bad_cache_alloc; |
| 1260 | |
| 1261 | mq->nr_entries_allocated = 0; |
| 1262 | mq->nr_cblocks_allocated = 0; |
| 1263 | |
| 1264 | mq->nr_buckets = next_power(from_cblock(cache_size) / 2, 16); |
| 1265 | mq->hash_bits = ffs(mq->nr_buckets) - 1; |
| 1266 | mq->table = kzalloc(sizeof(*mq->table) * mq->nr_buckets, GFP_KERNEL); |
| 1267 | if (!mq->table) |
| 1268 | goto bad_alloc_table; |
| 1269 | |
| 1270 | mq->allocation_bitset = alloc_bitset(from_cblock(cache_size)); |
| 1271 | if (!mq->allocation_bitset) |
| 1272 | goto bad_alloc_bitset; |
| 1273 | |
| 1274 | return &mq->policy; |
| 1275 | |
| 1276 | bad_alloc_bitset: |
| 1277 | kfree(mq->table); |
| 1278 | bad_alloc_table: |
| 1279 | free_entries(mq); |
| 1280 | bad_cache_alloc: |
| 1281 | kfree(mq); |
| 1282 | |
| 1283 | return NULL; |
| 1284 | } |
| 1285 | |
| 1286 | /*----------------------------------------------------------------*/ |
| 1287 | |
| 1288 | static struct dm_cache_policy_type mq_policy_type = { |
| 1289 | .name = "mq", |
Mike Snitzer | 4e7f506 | 2013-03-20 17:21:27 +0000 | [diff] [blame] | 1290 | .version = {1, 0, 0}, |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1291 | .hint_size = 4, |
| 1292 | .owner = THIS_MODULE, |
| 1293 | .create = mq_create |
| 1294 | }; |
| 1295 | |
| 1296 | static struct dm_cache_policy_type default_policy_type = { |
| 1297 | .name = "default", |
Mike Snitzer | 4e7f506 | 2013-03-20 17:21:27 +0000 | [diff] [blame] | 1298 | .version = {1, 0, 0}, |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1299 | .hint_size = 4, |
| 1300 | .owner = THIS_MODULE, |
| 1301 | .create = mq_create |
| 1302 | }; |
| 1303 | |
| 1304 | static int __init mq_init(void) |
| 1305 | { |
| 1306 | int r; |
| 1307 | |
| 1308 | mq_entry_cache = kmem_cache_create("dm_mq_policy_cache_entry", |
| 1309 | sizeof(struct entry), |
| 1310 | __alignof__(struct entry), |
| 1311 | 0, NULL); |
| 1312 | if (!mq_entry_cache) |
| 1313 | goto bad; |
| 1314 | |
| 1315 | r = dm_cache_policy_register(&mq_policy_type); |
| 1316 | if (r) { |
| 1317 | DMERR("register failed %d", r); |
| 1318 | goto bad_register_mq; |
| 1319 | } |
| 1320 | |
| 1321 | r = dm_cache_policy_register(&default_policy_type); |
| 1322 | if (!r) { |
Mike Snitzer | 4e7f506 | 2013-03-20 17:21:27 +0000 | [diff] [blame] | 1323 | DMINFO("version %u.%u.%u loaded", |
| 1324 | mq_policy_type.version[0], |
| 1325 | mq_policy_type.version[1], |
| 1326 | mq_policy_type.version[2]); |
Joe Thornber | f283635 | 2013-03-01 22:45:51 +0000 | [diff] [blame] | 1327 | return 0; |
| 1328 | } |
| 1329 | |
| 1330 | DMERR("register failed (as default) %d", r); |
| 1331 | |
| 1332 | dm_cache_policy_unregister(&mq_policy_type); |
| 1333 | bad_register_mq: |
| 1334 | kmem_cache_destroy(mq_entry_cache); |
| 1335 | bad: |
| 1336 | return -ENOMEM; |
| 1337 | } |
| 1338 | |
| 1339 | static void __exit mq_exit(void) |
| 1340 | { |
| 1341 | dm_cache_policy_unregister(&mq_policy_type); |
| 1342 | dm_cache_policy_unregister(&default_policy_type); |
| 1343 | |
| 1344 | kmem_cache_destroy(mq_entry_cache); |
| 1345 | } |
| 1346 | |
| 1347 | module_init(mq_init); |
| 1348 | module_exit(mq_exit); |
| 1349 | |
| 1350 | MODULE_AUTHOR("Joe Thornber <dm-devel@redhat.com>"); |
| 1351 | MODULE_LICENSE("GPL"); |
| 1352 | MODULE_DESCRIPTION("mq cache policy"); |
| 1353 | |
| 1354 | MODULE_ALIAS("dm-cache-default"); |