1 /* SPDX-License-Identifier: GPL-2.0 */ 2 #ifndef _BCACHE_BSET_H 3 #define _BCACHE_BSET_H 4 5 #include <linux/bcache.h> 6 #include <linux/kernel.h> 7 #include <linux/types.h> 8 9 #include "util.h" /* for time_stats */ 10 11 /* 12 * BKEYS: 13 * 14 * A bkey contains a key, a size field, a variable number of pointers, and some 15 * ancillary flag bits. 16 * 17 * We use two different functions for validating bkeys, bch_ptr_invalid and 18 * bch_ptr_bad(). 19 * 20 * bch_ptr_invalid() primarily filters out keys and pointers that would be 21 * invalid due to some sort of bug, whereas bch_ptr_bad() filters out keys and 22 * pointer that occur in normal practice but don't point to real data. 23 * 24 * The one exception to the rule that ptr_invalid() filters out invalid keys is 25 * that it also filters out keys of size 0 - these are keys that have been 26 * completely overwritten. It'd be safe to delete these in memory while leaving 27 * them on disk, just unnecessary work - so we filter them out when resorting 28 * instead. 29 * 30 * We can't filter out stale keys when we're resorting, because garbage 31 * collection needs to find them to ensure bucket gens don't wrap around - 32 * unless we're rewriting the btree node those stale keys still exist on disk. 33 * 34 * We also implement functions here for removing some number of sectors from the 35 * front or the back of a bkey - this is mainly used for fixing overlapping 36 * extents, by removing the overlapping sectors from the older key. 37 * 38 * BSETS: 39 * 40 * A bset is an array of bkeys laid out contiguously in memory in sorted order, 41 * along with a header. A btree node is made up of a number of these, written at 42 * different times. 43 * 44 * There could be many of them on disk, but we never allow there to be more than 45 * 4 in memory - we lazily resort as needed. 46 * 47 * We implement code here for creating and maintaining auxiliary search trees 48 * (described below) for searching an individial bset, and on top of that we 49 * implement a btree iterator. 50 * 51 * BTREE ITERATOR: 52 * 53 * Most of the code in bcache doesn't care about an individual bset - it needs 54 * to search entire btree nodes and iterate over them in sorted order. 55 * 56 * The btree iterator code serves both functions; it iterates through the keys 57 * in a btree node in sorted order, starting from either keys after a specific 58 * point (if you pass it a search key) or the start of the btree node. 59 * 60 * AUXILIARY SEARCH TREES: 61 * 62 * Since keys are variable length, we can't use a binary search on a bset - we 63 * wouldn't be able to find the start of the next key. But binary searches are 64 * slow anyways, due to terrible cache behaviour; bcache originally used binary 65 * searches and that code topped out at under 50k lookups/second. 66 * 67 * So we need to construct some sort of lookup table. Since we only insert keys 68 * into the last (unwritten) set, most of the keys within a given btree node are 69 * usually in sets that are mostly constant. We use two different types of 70 * lookup tables to take advantage of this. 71 * 72 * Both lookup tables share in common that they don't index every key in the 73 * set; they index one key every BSET_CACHELINE bytes, and then a linear search 74 * is used for the rest. 75 * 76 * For sets that have been written to disk and are no longer being inserted 77 * into, we construct a binary search tree in an array - traversing a binary 78 * search tree in an array gives excellent locality of reference and is very 79 * fast, since both children of any node are adjacent to each other in memory 80 * (and their grandchildren, and great grandchildren...) - this means 81 * prefetching can be used to great effect. 82 * 83 * It's quite useful performance wise to keep these nodes small - not just 84 * because they're more likely to be in L2, but also because we can prefetch 85 * more nodes on a single cacheline and thus prefetch more iterations in advance 86 * when traversing this tree. 87 * 88 * Nodes in the auxiliary search tree must contain both a key to compare against 89 * (we don't want to fetch the key from the set, that would defeat the purpose), 90 * and a pointer to the key. We use a few tricks to compress both of these. 91 * 92 * To compress the pointer, we take advantage of the fact that one node in the 93 * search tree corresponds to precisely BSET_CACHELINE bytes in the set. We have 94 * a function (to_inorder()) that takes the index of a node in a binary tree and 95 * returns what its index would be in an inorder traversal, so we only have to 96 * store the low bits of the offset. 97 * 98 * The key is 84 bits (KEY_DEV + key->key, the offset on the device). To 99 * compress that, we take advantage of the fact that when we're traversing the 100 * search tree at every iteration we know that both our search key and the key 101 * we're looking for lie within some range - bounded by our previous 102 * comparisons. (We special case the start of a search so that this is true even 103 * at the root of the tree). 104 * 105 * So we know the key we're looking for is between a and b, and a and b don't 106 * differ higher than bit 50, we don't need to check anything higher than bit 107 * 50. 108 * 109 * We don't usually need the rest of the bits, either; we only need enough bits 110 * to partition the key range we're currently checking. Consider key n - the 111 * key our auxiliary search tree node corresponds to, and key p, the key 112 * immediately preceding n. The lowest bit we need to store in the auxiliary 113 * search tree is the highest bit that differs between n and p. 114 * 115 * Note that this could be bit 0 - we might sometimes need all 80 bits to do the 116 * comparison. But we'd really like our nodes in the auxiliary search tree to be 117 * of fixed size. 118 * 119 * The solution is to make them fixed size, and when we're constructing a node 120 * check if p and n differed in the bits we needed them to. If they don't we 121 * flag that node, and when doing lookups we fallback to comparing against the 122 * real key. As long as this doesn't happen to often (and it seems to reliably 123 * happen a bit less than 1% of the time), we win - even on failures, that key 124 * is then more likely to be in cache than if we were doing binary searches all 125 * the way, since we're touching so much less memory. 126 * 127 * The keys in the auxiliary search tree are stored in (software) floating 128 * point, with an exponent and a mantissa. The exponent needs to be big enough 129 * to address all the bits in the original key, but the number of bits in the 130 * mantissa is somewhat arbitrary; more bits just gets us fewer failures. 131 * 132 * We need 7 bits for the exponent and 3 bits for the key's offset (since keys 133 * are 8 byte aligned); using 22 bits for the mantissa means a node is 4 bytes. 134 * We need one node per 128 bytes in the btree node, which means the auxiliary 135 * search trees take up 3% as much memory as the btree itself. 136 * 137 * Constructing these auxiliary search trees is moderately expensive, and we 138 * don't want to be constantly rebuilding the search tree for the last set 139 * whenever we insert another key into it. For the unwritten set, we use a much 140 * simpler lookup table - it's just a flat array, so index i in the lookup table 141 * corresponds to the i range of BSET_CACHELINE bytes in the set. Indexing 142 * within each byte range works the same as with the auxiliary search trees. 143 * 144 * These are much easier to keep up to date when we insert a key - we do it 145 * somewhat lazily; when we shift a key up we usually just increment the pointer 146 * to it, only when it would overflow do we go to the trouble of finding the 147 * first key in that range of bytes again. 148 */ 149 150 struct btree_keys; 151 struct btree_iter; 152 struct btree_iter_set; 153 struct bkey_float; 154 155 #define MAX_BSETS 4U 156 157 struct bset_tree { 158 /* 159 * We construct a binary tree in an array as if the array 160 * started at 1, so that things line up on the same cachelines 161 * better: see comments in bset.c at cacheline_to_bkey() for 162 * details 163 */ 164 165 /* size of the binary tree and prev array */ 166 unsigned size; 167 168 /* function of size - precalculated for to_inorder() */ 169 unsigned extra; 170 171 /* copy of the last key in the set */ 172 struct bkey end; 173 struct bkey_float *tree; 174 175 /* 176 * The nodes in the bset tree point to specific keys - this 177 * array holds the sizes of the previous key. 178 * 179 * Conceptually it's a member of struct bkey_float, but we want 180 * to keep bkey_float to 4 bytes and prev isn't used in the fast 181 * path. 182 */ 183 uint8_t *prev; 184 185 /* The actual btree node, with pointers to each sorted set */ 186 struct bset *data; 187 }; 188 189 struct btree_keys_ops { 190 bool (*sort_cmp)(struct btree_iter_set, 191 struct btree_iter_set); 192 struct bkey *(*sort_fixup)(struct btree_iter *, struct bkey *); 193 bool (*insert_fixup)(struct btree_keys *, struct bkey *, 194 struct btree_iter *, struct bkey *); 195 bool (*key_invalid)(struct btree_keys *, 196 const struct bkey *); 197 bool (*key_bad)(struct btree_keys *, const struct bkey *); 198 bool (*key_merge)(struct btree_keys *, 199 struct bkey *, struct bkey *); 200 void (*key_to_text)(char *, size_t, const struct bkey *); 201 void (*key_dump)(struct btree_keys *, const struct bkey *); 202 203 /* 204 * Only used for deciding whether to use START_KEY(k) or just the key 205 * itself in a couple places 206 */ 207 bool is_extents; 208 }; 209 210 struct btree_keys { 211 const struct btree_keys_ops *ops; 212 uint8_t page_order; 213 uint8_t nsets; 214 unsigned last_set_unwritten:1; 215 bool *expensive_debug_checks; 216 217 /* 218 * Sets of sorted keys - the real btree node - plus a binary search tree 219 * 220 * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point 221 * to the memory we have allocated for this btree node. Additionally, 222 * set[0]->data points to the entire btree node as it exists on disk. 223 */ 224 struct bset_tree set[MAX_BSETS]; 225 }; 226 227 static inline struct bset_tree *bset_tree_last(struct btree_keys *b) 228 { 229 return b->set + b->nsets; 230 } 231 232 static inline bool bset_written(struct btree_keys *b, struct bset_tree *t) 233 { 234 return t <= b->set + b->nsets - b->last_set_unwritten; 235 } 236 237 static inline bool bkey_written(struct btree_keys *b, struct bkey *k) 238 { 239 return !b->last_set_unwritten || k < b->set[b->nsets].data->start; 240 } 241 242 static inline unsigned bset_byte_offset(struct btree_keys *b, struct bset *i) 243 { 244 return ((size_t) i) - ((size_t) b->set->data); 245 } 246 247 static inline unsigned bset_sector_offset(struct btree_keys *b, struct bset *i) 248 { 249 return bset_byte_offset(b, i) >> 9; 250 } 251 252 #define __set_bytes(i, k) (sizeof(*(i)) + (k) * sizeof(uint64_t)) 253 #define set_bytes(i) __set_bytes(i, i->keys) 254 255 #define __set_blocks(i, k, block_bytes) \ 256 DIV_ROUND_UP(__set_bytes(i, k), block_bytes) 257 #define set_blocks(i, block_bytes) \ 258 __set_blocks(i, (i)->keys, block_bytes) 259 260 static inline size_t bch_btree_keys_u64s_remaining(struct btree_keys *b) 261 { 262 struct bset_tree *t = bset_tree_last(b); 263 264 BUG_ON((PAGE_SIZE << b->page_order) < 265 (bset_byte_offset(b, t->data) + set_bytes(t->data))); 266 267 if (!b->last_set_unwritten) 268 return 0; 269 270 return ((PAGE_SIZE << b->page_order) - 271 (bset_byte_offset(b, t->data) + set_bytes(t->data))) / 272 sizeof(u64); 273 } 274 275 static inline struct bset *bset_next_set(struct btree_keys *b, 276 unsigned block_bytes) 277 { 278 struct bset *i = bset_tree_last(b)->data; 279 280 return ((void *) i) + roundup(set_bytes(i), block_bytes); 281 } 282 283 void bch_btree_keys_free(struct btree_keys *); 284 int bch_btree_keys_alloc(struct btree_keys *, unsigned, gfp_t); 285 void bch_btree_keys_init(struct btree_keys *, const struct btree_keys_ops *, 286 bool *); 287 288 void bch_bset_init_next(struct btree_keys *, struct bset *, uint64_t); 289 void bch_bset_build_written_tree(struct btree_keys *); 290 void bch_bset_fix_invalidated_key(struct btree_keys *, struct bkey *); 291 bool bch_bkey_try_merge(struct btree_keys *, struct bkey *, struct bkey *); 292 void bch_bset_insert(struct btree_keys *, struct bkey *, struct bkey *); 293 unsigned bch_btree_insert_key(struct btree_keys *, struct bkey *, 294 struct bkey *); 295 296 enum { 297 BTREE_INSERT_STATUS_NO_INSERT = 0, 298 BTREE_INSERT_STATUS_INSERT, 299 BTREE_INSERT_STATUS_BACK_MERGE, 300 BTREE_INSERT_STATUS_OVERWROTE, 301 BTREE_INSERT_STATUS_FRONT_MERGE, 302 }; 303 304 /* Btree key iteration */ 305 306 struct btree_iter { 307 size_t size, used; 308 #ifdef CONFIG_BCACHE_DEBUG 309 struct btree_keys *b; 310 #endif 311 struct btree_iter_set { 312 struct bkey *k, *end; 313 } data[MAX_BSETS]; 314 }; 315 316 typedef bool (*ptr_filter_fn)(struct btree_keys *, const struct bkey *); 317 318 struct bkey *bch_btree_iter_next(struct btree_iter *); 319 struct bkey *bch_btree_iter_next_filter(struct btree_iter *, 320 struct btree_keys *, ptr_filter_fn); 321 322 void bch_btree_iter_push(struct btree_iter *, struct bkey *, struct bkey *); 323 struct bkey *bch_btree_iter_init(struct btree_keys *, struct btree_iter *, 324 struct bkey *); 325 326 struct bkey *__bch_bset_search(struct btree_keys *, struct bset_tree *, 327 const struct bkey *); 328 329 /* 330 * Returns the first key that is strictly greater than search 331 */ 332 static inline struct bkey *bch_bset_search(struct btree_keys *b, 333 struct bset_tree *t, 334 const struct bkey *search) 335 { 336 return search ? __bch_bset_search(b, t, search) : t->data->start; 337 } 338 339 #define for_each_key_filter(b, k, iter, filter) \ 340 for (bch_btree_iter_init((b), (iter), NULL); \ 341 ((k) = bch_btree_iter_next_filter((iter), (b), filter));) 342 343 #define for_each_key(b, k, iter) \ 344 for (bch_btree_iter_init((b), (iter), NULL); \ 345 ((k) = bch_btree_iter_next(iter));) 346 347 /* Sorting */ 348 349 struct bset_sort_state { 350 mempool_t pool; 351 352 unsigned page_order; 353 unsigned crit_factor; 354 355 struct time_stats time; 356 }; 357 358 void bch_bset_sort_state_free(struct bset_sort_state *); 359 int bch_bset_sort_state_init(struct bset_sort_state *, unsigned); 360 void bch_btree_sort_lazy(struct btree_keys *, struct bset_sort_state *); 361 void bch_btree_sort_into(struct btree_keys *, struct btree_keys *, 362 struct bset_sort_state *); 363 void bch_btree_sort_and_fix_extents(struct btree_keys *, struct btree_iter *, 364 struct bset_sort_state *); 365 void bch_btree_sort_partial(struct btree_keys *, unsigned, 366 struct bset_sort_state *); 367 368 static inline void bch_btree_sort(struct btree_keys *b, 369 struct bset_sort_state *state) 370 { 371 bch_btree_sort_partial(b, 0, state); 372 } 373 374 struct bset_stats { 375 size_t sets_written, sets_unwritten; 376 size_t bytes_written, bytes_unwritten; 377 size_t floats, failed; 378 }; 379 380 void bch_btree_keys_stats(struct btree_keys *, struct bset_stats *); 381 382 /* Bkey utility code */ 383 384 #define bset_bkey_last(i) bkey_idx((struct bkey *) (i)->d, (i)->keys) 385 386 static inline struct bkey *bset_bkey_idx(struct bset *i, unsigned idx) 387 { 388 return bkey_idx(i->start, idx); 389 } 390 391 static inline void bkey_init(struct bkey *k) 392 { 393 *k = ZERO_KEY; 394 } 395 396 static __always_inline int64_t bkey_cmp(const struct bkey *l, 397 const struct bkey *r) 398 { 399 return unlikely(KEY_INODE(l) != KEY_INODE(r)) 400 ? (int64_t) KEY_INODE(l) - (int64_t) KEY_INODE(r) 401 : (int64_t) KEY_OFFSET(l) - (int64_t) KEY_OFFSET(r); 402 } 403 404 void bch_bkey_copy_single_ptr(struct bkey *, const struct bkey *, 405 unsigned); 406 bool __bch_cut_front(const struct bkey *, struct bkey *); 407 bool __bch_cut_back(const struct bkey *, struct bkey *); 408 409 static inline bool bch_cut_front(const struct bkey *where, struct bkey *k) 410 { 411 BUG_ON(bkey_cmp(where, k) > 0); 412 return __bch_cut_front(where, k); 413 } 414 415 static inline bool bch_cut_back(const struct bkey *where, struct bkey *k) 416 { 417 BUG_ON(bkey_cmp(where, &START_KEY(k)) < 0); 418 return __bch_cut_back(where, k); 419 } 420 421 #define PRECEDING_KEY(_k) \ 422 ({ \ 423 struct bkey *_ret = NULL; \ 424 \ 425 if (KEY_INODE(_k) || KEY_OFFSET(_k)) { \ 426 _ret = &KEY(KEY_INODE(_k), KEY_OFFSET(_k), 0); \ 427 \ 428 if (!_ret->low) \ 429 _ret->high--; \ 430 _ret->low--; \ 431 } \ 432 \ 433 _ret; \ 434 }) 435 436 static inline bool bch_ptr_invalid(struct btree_keys *b, const struct bkey *k) 437 { 438 return b->ops->key_invalid(b, k); 439 } 440 441 static inline bool bch_ptr_bad(struct btree_keys *b, const struct bkey *k) 442 { 443 return b->ops->key_bad(b, k); 444 } 445 446 static inline void bch_bkey_to_text(struct btree_keys *b, char *buf, 447 size_t size, const struct bkey *k) 448 { 449 return b->ops->key_to_text(buf, size, k); 450 } 451 452 static inline bool bch_bkey_equal_header(const struct bkey *l, 453 const struct bkey *r) 454 { 455 return (KEY_DIRTY(l) == KEY_DIRTY(r) && 456 KEY_PTRS(l) == KEY_PTRS(r) && 457 KEY_CSUM(l) == KEY_CSUM(r)); 458 } 459 460 /* Keylists */ 461 462 struct keylist { 463 union { 464 struct bkey *keys; 465 uint64_t *keys_p; 466 }; 467 union { 468 struct bkey *top; 469 uint64_t *top_p; 470 }; 471 472 /* Enough room for btree_split's keys without realloc */ 473 #define KEYLIST_INLINE 16 474 uint64_t inline_keys[KEYLIST_INLINE]; 475 }; 476 477 static inline void bch_keylist_init(struct keylist *l) 478 { 479 l->top_p = l->keys_p = l->inline_keys; 480 } 481 482 static inline void bch_keylist_init_single(struct keylist *l, struct bkey *k) 483 { 484 l->keys = k; 485 l->top = bkey_next(k); 486 } 487 488 static inline void bch_keylist_push(struct keylist *l) 489 { 490 l->top = bkey_next(l->top); 491 } 492 493 static inline void bch_keylist_add(struct keylist *l, struct bkey *k) 494 { 495 bkey_copy(l->top, k); 496 bch_keylist_push(l); 497 } 498 499 static inline bool bch_keylist_empty(struct keylist *l) 500 { 501 return l->top == l->keys; 502 } 503 504 static inline void bch_keylist_reset(struct keylist *l) 505 { 506 l->top = l->keys; 507 } 508 509 static inline void bch_keylist_free(struct keylist *l) 510 { 511 if (l->keys_p != l->inline_keys) 512 kfree(l->keys_p); 513 } 514 515 static inline size_t bch_keylist_nkeys(struct keylist *l) 516 { 517 return l->top_p - l->keys_p; 518 } 519 520 static inline size_t bch_keylist_bytes(struct keylist *l) 521 { 522 return bch_keylist_nkeys(l) * sizeof(uint64_t); 523 } 524 525 struct bkey *bch_keylist_pop(struct keylist *); 526 void bch_keylist_pop_front(struct keylist *); 527 int __bch_keylist_realloc(struct keylist *, unsigned); 528 529 /* Debug stuff */ 530 531 #ifdef CONFIG_BCACHE_DEBUG 532 533 int __bch_count_data(struct btree_keys *); 534 void __printf(2, 3) __bch_check_keys(struct btree_keys *, const char *, ...); 535 void bch_dump_bset(struct btree_keys *, struct bset *, unsigned); 536 void bch_dump_bucket(struct btree_keys *); 537 538 #else 539 540 static inline int __bch_count_data(struct btree_keys *b) { return -1; } 541 static inline void __printf(2, 3) 542 __bch_check_keys(struct btree_keys *b, const char *fmt, ...) {} 543 static inline void bch_dump_bucket(struct btree_keys *b) {} 544 void bch_dump_bset(struct btree_keys *, struct bset *, unsigned); 545 546 #endif 547 548 static inline bool btree_keys_expensive_checks(struct btree_keys *b) 549 { 550 #ifdef CONFIG_BCACHE_DEBUG 551 return *b->expensive_debug_checks; 552 #else 553 return false; 554 #endif 555 } 556 557 static inline int bch_count_data(struct btree_keys *b) 558 { 559 return btree_keys_expensive_checks(b) ? __bch_count_data(b) : -1; 560 } 561 562 #define bch_check_keys(b, ...) \ 563 do { \ 564 if (btree_keys_expensive_checks(b)) \ 565 __bch_check_keys(b, __VA_ARGS__); \ 566 } while (0) 567 568 #endif 569