Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1 | /* |
| 2 | * Code for working with individual keys, and sorted sets of keys with in a |
| 3 | * btree node |
| 4 | * |
| 5 | * Copyright 2012 Google, Inc. |
| 6 | */ |
| 7 | |
| 8 | #include "bcache.h" |
| 9 | #include "btree.h" |
| 10 | #include "debug.h" |
| 11 | |
| 12 | #include <linux/random.h> |
Geert Uytterhoeven | cd953ed | 2013-03-27 18:56:28 +0100 | [diff] [blame] | 13 | #include <linux/prefetch.h> |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 14 | |
| 15 | /* Keylists */ |
| 16 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 17 | int bch_keylist_realloc(struct keylist *l, int nptrs, struct cache_set *c) |
| 18 | { |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 19 | size_t oldsize = bch_keylist_nkeys(l); |
| 20 | size_t newsize = oldsize + 2 + nptrs; |
| 21 | uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p; |
| 22 | uint64_t *new_keys; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 23 | |
| 24 | /* The journalling code doesn't handle the case where the keys to insert |
| 25 | * is bigger than an empty write: If we just return -ENOMEM here, |
| 26 | * bio_insert() and bio_invalidate() will insert the keys created so far |
| 27 | * and finish the rest when the keylist is empty. |
| 28 | */ |
| 29 | if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset)) |
| 30 | return -ENOMEM; |
| 31 | |
| 32 | newsize = roundup_pow_of_two(newsize); |
| 33 | |
| 34 | if (newsize <= KEYLIST_INLINE || |
| 35 | roundup_pow_of_two(oldsize) == newsize) |
| 36 | return 0; |
| 37 | |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 38 | new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 39 | |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 40 | if (!new_keys) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 41 | return -ENOMEM; |
| 42 | |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 43 | if (!old_keys) |
| 44 | memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 45 | |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 46 | l->keys_p = new_keys; |
| 47 | l->top_p = new_keys + oldsize; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 48 | |
| 49 | return 0; |
| 50 | } |
| 51 | |
| 52 | struct bkey *bch_keylist_pop(struct keylist *l) |
| 53 | { |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 54 | struct bkey *k = l->keys; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 55 | |
| 56 | if (k == l->top) |
| 57 | return NULL; |
| 58 | |
| 59 | while (bkey_next(k) != l->top) |
| 60 | k = bkey_next(k); |
| 61 | |
| 62 | return l->top = k; |
| 63 | } |
| 64 | |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 65 | void bch_keylist_pop_front(struct keylist *l) |
| 66 | { |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 67 | l->top_p -= bkey_u64s(l->keys); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 68 | |
Kent Overstreet | c2f95ae | 2013-07-24 17:24:25 -0700 | [diff] [blame] | 69 | memmove(l->keys, |
| 70 | bkey_next(l->keys), |
| 71 | bch_keylist_bytes(l)); |
Kent Overstreet | 26c949f | 2013-09-10 18:41:15 -0700 | [diff] [blame] | 72 | } |
| 73 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 74 | /* Pointer validation */ |
| 75 | |
| 76 | bool __bch_ptr_invalid(struct cache_set *c, int level, const struct bkey *k) |
| 77 | { |
| 78 | unsigned i; |
Kent Overstreet | 85b1492 | 2013-05-14 20:33:16 -0700 | [diff] [blame] | 79 | char buf[80]; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 80 | |
| 81 | if (level && (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))) |
| 82 | goto bad; |
| 83 | |
| 84 | if (!level && KEY_SIZE(k) > KEY_OFFSET(k)) |
| 85 | goto bad; |
| 86 | |
| 87 | if (!KEY_SIZE(k)) |
| 88 | return true; |
| 89 | |
| 90 | for (i = 0; i < KEY_PTRS(k); i++) |
| 91 | if (ptr_available(c, k, i)) { |
| 92 | struct cache *ca = PTR_CACHE(c, k, i); |
| 93 | size_t bucket = PTR_BUCKET_NR(c, k, i); |
| 94 | size_t r = bucket_remainder(c, PTR_OFFSET(k, i)); |
| 95 | |
| 96 | if (KEY_SIZE(k) + r > c->sb.bucket_size || |
| 97 | bucket < ca->sb.first_bucket || |
| 98 | bucket >= ca->sb.nbuckets) |
| 99 | goto bad; |
| 100 | } |
| 101 | |
| 102 | return false; |
| 103 | bad: |
Kent Overstreet | 85b1492 | 2013-05-14 20:33:16 -0700 | [diff] [blame] | 104 | bch_bkey_to_text(buf, sizeof(buf), k); |
| 105 | cache_bug(c, "spotted bad key %s: %s", buf, bch_ptr_status(c, k)); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 106 | return true; |
| 107 | } |
| 108 | |
| 109 | bool bch_ptr_bad(struct btree *b, const struct bkey *k) |
| 110 | { |
| 111 | struct bucket *g; |
| 112 | unsigned i, stale; |
| 113 | |
| 114 | if (!bkey_cmp(k, &ZERO_KEY) || |
| 115 | !KEY_PTRS(k) || |
| 116 | bch_ptr_invalid(b, k)) |
| 117 | return true; |
| 118 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 119 | for (i = 0; i < KEY_PTRS(k); i++) { |
| 120 | if (!ptr_available(b->c, k, i)) |
| 121 | return true; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 122 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 123 | g = PTR_BUCKET(b->c, k, i); |
| 124 | stale = ptr_stale(b->c, k, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 125 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 126 | btree_bug_on(stale > 96, b, |
| 127 | "key too stale: %i, need_gc %u", |
| 128 | stale, b->c->need_gc); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 129 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 130 | btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k), |
| 131 | b, "stale dirty pointer"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 132 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 133 | if (stale) |
| 134 | return true; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 135 | |
| 136 | #ifdef CONFIG_BCACHE_EDEBUG |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 137 | if (!mutex_trylock(&b->c->bucket_lock)) |
| 138 | continue; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 139 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 140 | if (b->level) { |
| 141 | if (KEY_DIRTY(k) || |
| 142 | g->prio != BTREE_PRIO || |
| 143 | (b->c->gc_mark_valid && |
| 144 | GC_MARK(g) != GC_MARK_METADATA)) |
| 145 | goto bug; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 146 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 147 | } else { |
| 148 | if (g->prio == BTREE_PRIO) |
| 149 | goto bug; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 150 | |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 151 | if (KEY_DIRTY(k) && |
| 152 | b->c->gc_mark_valid && |
| 153 | GC_MARK(g) != GC_MARK_DIRTY) |
| 154 | goto bug; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 155 | } |
Kent Overstreet | e58ff15 | 2013-07-24 18:14:44 -0700 | [diff] [blame^] | 156 | mutex_unlock(&b->c->bucket_lock); |
| 157 | #endif |
| 158 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 159 | |
| 160 | return false; |
| 161 | #ifdef CONFIG_BCACHE_EDEBUG |
| 162 | bug: |
| 163 | mutex_unlock(&b->c->bucket_lock); |
Kent Overstreet | 85b1492 | 2013-05-14 20:33:16 -0700 | [diff] [blame] | 164 | |
| 165 | { |
| 166 | char buf[80]; |
| 167 | |
| 168 | bch_bkey_to_text(buf, sizeof(buf), k); |
| 169 | btree_bug(b, |
Kent Overstreet | b1a67b0 | 2013-03-25 11:46:44 -0700 | [diff] [blame] | 170 | "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i", |
Kent Overstreet | 85b1492 | 2013-05-14 20:33:16 -0700 | [diff] [blame] | 171 | buf, PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin), |
| 172 | g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen); |
| 173 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 174 | return true; |
| 175 | #endif |
| 176 | } |
| 177 | |
| 178 | /* Key/pointer manipulation */ |
| 179 | |
| 180 | void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src, |
| 181 | unsigned i) |
| 182 | { |
| 183 | BUG_ON(i > KEY_PTRS(src)); |
| 184 | |
| 185 | /* Only copy the header, key, and one pointer. */ |
| 186 | memcpy(dest, src, 2 * sizeof(uint64_t)); |
| 187 | dest->ptr[0] = src->ptr[i]; |
| 188 | SET_KEY_PTRS(dest, 1); |
| 189 | /* We didn't copy the checksum so clear that bit. */ |
| 190 | SET_KEY_CSUM(dest, 0); |
| 191 | } |
| 192 | |
| 193 | bool __bch_cut_front(const struct bkey *where, struct bkey *k) |
| 194 | { |
| 195 | unsigned i, len = 0; |
| 196 | |
| 197 | if (bkey_cmp(where, &START_KEY(k)) <= 0) |
| 198 | return false; |
| 199 | |
| 200 | if (bkey_cmp(where, k) < 0) |
| 201 | len = KEY_OFFSET(k) - KEY_OFFSET(where); |
| 202 | else |
| 203 | bkey_copy_key(k, where); |
| 204 | |
| 205 | for (i = 0; i < KEY_PTRS(k); i++) |
| 206 | SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len); |
| 207 | |
| 208 | BUG_ON(len > KEY_SIZE(k)); |
| 209 | SET_KEY_SIZE(k, len); |
| 210 | return true; |
| 211 | } |
| 212 | |
| 213 | bool __bch_cut_back(const struct bkey *where, struct bkey *k) |
| 214 | { |
| 215 | unsigned len = 0; |
| 216 | |
| 217 | if (bkey_cmp(where, k) >= 0) |
| 218 | return false; |
| 219 | |
| 220 | BUG_ON(KEY_INODE(where) != KEY_INODE(k)); |
| 221 | |
| 222 | if (bkey_cmp(where, &START_KEY(k)) > 0) |
| 223 | len = KEY_OFFSET(where) - KEY_START(k); |
| 224 | |
| 225 | bkey_copy_key(k, where); |
| 226 | |
| 227 | BUG_ON(len > KEY_SIZE(k)); |
| 228 | SET_KEY_SIZE(k, len); |
| 229 | return true; |
| 230 | } |
| 231 | |
| 232 | static uint64_t merge_chksums(struct bkey *l, struct bkey *r) |
| 233 | { |
| 234 | return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) & |
| 235 | ~((uint64_t)1 << 63); |
| 236 | } |
| 237 | |
| 238 | /* Tries to merge l and r: l should be lower than r |
| 239 | * Returns true if we were able to merge. If we did merge, l will be the merged |
| 240 | * key, r will be untouched. |
| 241 | */ |
| 242 | bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r) |
| 243 | { |
| 244 | unsigned i; |
| 245 | |
| 246 | if (key_merging_disabled(b->c)) |
| 247 | return false; |
| 248 | |
| 249 | if (KEY_PTRS(l) != KEY_PTRS(r) || |
| 250 | KEY_DIRTY(l) != KEY_DIRTY(r) || |
| 251 | bkey_cmp(l, &START_KEY(r))) |
| 252 | return false; |
| 253 | |
| 254 | for (i = 0; i < KEY_PTRS(l); i++) |
| 255 | if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] || |
| 256 | PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i)) |
| 257 | return false; |
| 258 | |
| 259 | /* Keys with no pointers aren't restricted to one bucket and could |
| 260 | * overflow KEY_SIZE |
| 261 | */ |
| 262 | if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) { |
| 263 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l)); |
| 264 | SET_KEY_SIZE(l, USHRT_MAX); |
| 265 | |
| 266 | bch_cut_front(l, r); |
| 267 | return false; |
| 268 | } |
| 269 | |
| 270 | if (KEY_CSUM(l)) { |
| 271 | if (KEY_CSUM(r)) |
| 272 | l->ptr[KEY_PTRS(l)] = merge_chksums(l, r); |
| 273 | else |
| 274 | SET_KEY_CSUM(l, 0); |
| 275 | } |
| 276 | |
| 277 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r)); |
| 278 | SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r)); |
| 279 | |
| 280 | return true; |
| 281 | } |
| 282 | |
| 283 | /* Binary tree stuff for auxiliary search trees */ |
| 284 | |
| 285 | static unsigned inorder_next(unsigned j, unsigned size) |
| 286 | { |
| 287 | if (j * 2 + 1 < size) { |
| 288 | j = j * 2 + 1; |
| 289 | |
| 290 | while (j * 2 < size) |
| 291 | j *= 2; |
| 292 | } else |
| 293 | j >>= ffz(j) + 1; |
| 294 | |
| 295 | return j; |
| 296 | } |
| 297 | |
| 298 | static unsigned inorder_prev(unsigned j, unsigned size) |
| 299 | { |
| 300 | if (j * 2 < size) { |
| 301 | j = j * 2; |
| 302 | |
| 303 | while (j * 2 + 1 < size) |
| 304 | j = j * 2 + 1; |
| 305 | } else |
| 306 | j >>= ffs(j); |
| 307 | |
| 308 | return j; |
| 309 | } |
| 310 | |
| 311 | /* I have no idea why this code works... and I'm the one who wrote it |
| 312 | * |
| 313 | * However, I do know what it does: |
| 314 | * Given a binary tree constructed in an array (i.e. how you normally implement |
| 315 | * a heap), it converts a node in the tree - referenced by array index - to the |
| 316 | * index it would have if you did an inorder traversal. |
| 317 | * |
| 318 | * Also tested for every j, size up to size somewhere around 6 million. |
| 319 | * |
| 320 | * The binary tree starts at array index 1, not 0 |
| 321 | * extra is a function of size: |
| 322 | * extra = (size - rounddown_pow_of_two(size - 1)) << 1; |
| 323 | */ |
| 324 | static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra) |
| 325 | { |
| 326 | unsigned b = fls(j); |
| 327 | unsigned shift = fls(size - 1) - b; |
| 328 | |
| 329 | j ^= 1U << (b - 1); |
| 330 | j <<= 1; |
| 331 | j |= 1; |
| 332 | j <<= shift; |
| 333 | |
| 334 | if (j > extra) |
| 335 | j -= (j - extra) >> 1; |
| 336 | |
| 337 | return j; |
| 338 | } |
| 339 | |
| 340 | static unsigned to_inorder(unsigned j, struct bset_tree *t) |
| 341 | { |
| 342 | return __to_inorder(j, t->size, t->extra); |
| 343 | } |
| 344 | |
| 345 | static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra) |
| 346 | { |
| 347 | unsigned shift; |
| 348 | |
| 349 | if (j > extra) |
| 350 | j += j - extra; |
| 351 | |
| 352 | shift = ffs(j); |
| 353 | |
| 354 | j >>= shift; |
| 355 | j |= roundup_pow_of_two(size) >> shift; |
| 356 | |
| 357 | return j; |
| 358 | } |
| 359 | |
| 360 | static unsigned inorder_to_tree(unsigned j, struct bset_tree *t) |
| 361 | { |
| 362 | return __inorder_to_tree(j, t->size, t->extra); |
| 363 | } |
| 364 | |
| 365 | #if 0 |
| 366 | void inorder_test(void) |
| 367 | { |
| 368 | unsigned long done = 0; |
| 369 | ktime_t start = ktime_get(); |
| 370 | |
| 371 | for (unsigned size = 2; |
| 372 | size < 65536000; |
| 373 | size++) { |
| 374 | unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1; |
| 375 | unsigned i = 1, j = rounddown_pow_of_two(size - 1); |
| 376 | |
| 377 | if (!(size % 4096)) |
| 378 | printk(KERN_NOTICE "loop %u, %llu per us\n", size, |
| 379 | done / ktime_us_delta(ktime_get(), start)); |
| 380 | |
| 381 | while (1) { |
| 382 | if (__inorder_to_tree(i, size, extra) != j) |
| 383 | panic("size %10u j %10u i %10u", size, j, i); |
| 384 | |
| 385 | if (__to_inorder(j, size, extra) != i) |
| 386 | panic("size %10u j %10u i %10u", size, j, i); |
| 387 | |
| 388 | if (j == rounddown_pow_of_two(size) - 1) |
| 389 | break; |
| 390 | |
| 391 | BUG_ON(inorder_prev(inorder_next(j, size), size) != j); |
| 392 | |
| 393 | j = inorder_next(j, size); |
| 394 | i++; |
| 395 | } |
| 396 | |
| 397 | done += size - 1; |
| 398 | } |
| 399 | } |
| 400 | #endif |
| 401 | |
| 402 | /* |
Phil Viana | 48a7302 | 2013-06-03 09:51:42 -0300 | [diff] [blame] | 403 | * Cacheline/offset <-> bkey pointer arithmetic: |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 404 | * |
| 405 | * t->tree is a binary search tree in an array; each node corresponds to a key |
| 406 | * in one cacheline in t->set (BSET_CACHELINE bytes). |
| 407 | * |
| 408 | * This means we don't have to store the full index of the key that a node in |
| 409 | * the binary tree points to; to_inorder() gives us the cacheline, and then |
| 410 | * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes. |
| 411 | * |
Phil Viana | 48a7302 | 2013-06-03 09:51:42 -0300 | [diff] [blame] | 412 | * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 413 | * make this work. |
| 414 | * |
| 415 | * To construct the bfloat for an arbitrary key we need to know what the key |
| 416 | * immediately preceding it is: we have to check if the two keys differ in the |
| 417 | * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size |
| 418 | * of the previous key so we can walk backwards to it from t->tree[j]'s key. |
| 419 | */ |
| 420 | |
| 421 | static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline, |
| 422 | unsigned offset) |
| 423 | { |
| 424 | return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8; |
| 425 | } |
| 426 | |
| 427 | static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k) |
| 428 | { |
| 429 | return ((void *) k - (void *) t->data) / BSET_CACHELINE; |
| 430 | } |
| 431 | |
| 432 | static unsigned bkey_to_cacheline_offset(struct bkey *k) |
| 433 | { |
| 434 | return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t); |
| 435 | } |
| 436 | |
| 437 | static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j) |
| 438 | { |
| 439 | return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m); |
| 440 | } |
| 441 | |
| 442 | static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j) |
| 443 | { |
| 444 | return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]); |
| 445 | } |
| 446 | |
| 447 | /* |
| 448 | * For the write set - the one we're currently inserting keys into - we don't |
| 449 | * maintain a full search tree, we just keep a simple lookup table in t->prev. |
| 450 | */ |
| 451 | static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline) |
| 452 | { |
| 453 | return cacheline_to_bkey(t, cacheline, t->prev[cacheline]); |
| 454 | } |
| 455 | |
| 456 | static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift) |
| 457 | { |
| 458 | #ifdef CONFIG_X86_64 |
| 459 | asm("shrd %[shift],%[high],%[low]" |
| 460 | : [low] "+Rm" (low) |
| 461 | : [high] "R" (high), |
| 462 | [shift] "ci" (shift) |
| 463 | : "cc"); |
| 464 | #else |
| 465 | low >>= shift; |
| 466 | low |= (high << 1) << (63U - shift); |
| 467 | #endif |
| 468 | return low; |
| 469 | } |
| 470 | |
| 471 | static inline unsigned bfloat_mantissa(const struct bkey *k, |
| 472 | struct bkey_float *f) |
| 473 | { |
| 474 | const uint64_t *p = &k->low - (f->exponent >> 6); |
| 475 | return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK; |
| 476 | } |
| 477 | |
| 478 | static void make_bfloat(struct bset_tree *t, unsigned j) |
| 479 | { |
| 480 | struct bkey_float *f = &t->tree[j]; |
| 481 | struct bkey *m = tree_to_bkey(t, j); |
| 482 | struct bkey *p = tree_to_prev_bkey(t, j); |
| 483 | |
| 484 | struct bkey *l = is_power_of_2(j) |
| 485 | ? t->data->start |
| 486 | : tree_to_prev_bkey(t, j >> ffs(j)); |
| 487 | |
| 488 | struct bkey *r = is_power_of_2(j + 1) |
| 489 | ? node(t->data, t->data->keys - bkey_u64s(&t->end)) |
| 490 | : tree_to_bkey(t, j >> (ffz(j) + 1)); |
| 491 | |
| 492 | BUG_ON(m < l || m > r); |
| 493 | BUG_ON(bkey_next(p) != m); |
| 494 | |
| 495 | if (KEY_INODE(l) != KEY_INODE(r)) |
| 496 | f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64; |
| 497 | else |
| 498 | f->exponent = fls64(r->low ^ l->low); |
| 499 | |
| 500 | f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0); |
| 501 | |
| 502 | /* |
| 503 | * Setting f->exponent = 127 flags this node as failed, and causes the |
| 504 | * lookup code to fall back to comparing against the original key. |
| 505 | */ |
| 506 | |
| 507 | if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f)) |
| 508 | f->mantissa = bfloat_mantissa(m, f) - 1; |
| 509 | else |
| 510 | f->exponent = 127; |
| 511 | } |
| 512 | |
| 513 | static void bset_alloc_tree(struct btree *b, struct bset_tree *t) |
| 514 | { |
| 515 | if (t != b->sets) { |
| 516 | unsigned j = roundup(t[-1].size, |
| 517 | 64 / sizeof(struct bkey_float)); |
| 518 | |
| 519 | t->tree = t[-1].tree + j; |
| 520 | t->prev = t[-1].prev + j; |
| 521 | } |
| 522 | |
| 523 | while (t < b->sets + MAX_BSETS) |
| 524 | t++->size = 0; |
| 525 | } |
| 526 | |
| 527 | static void bset_build_unwritten_tree(struct btree *b) |
| 528 | { |
| 529 | struct bset_tree *t = b->sets + b->nsets; |
| 530 | |
| 531 | bset_alloc_tree(b, t); |
| 532 | |
| 533 | if (t->tree != b->sets->tree + bset_tree_space(b)) { |
| 534 | t->prev[0] = bkey_to_cacheline_offset(t->data->start); |
| 535 | t->size = 1; |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | static void bset_build_written_tree(struct btree *b) |
| 540 | { |
| 541 | struct bset_tree *t = b->sets + b->nsets; |
| 542 | struct bkey *k = t->data->start; |
| 543 | unsigned j, cacheline = 1; |
| 544 | |
| 545 | bset_alloc_tree(b, t); |
| 546 | |
| 547 | t->size = min_t(unsigned, |
| 548 | bkey_to_cacheline(t, end(t->data)), |
| 549 | b->sets->tree + bset_tree_space(b) - t->tree); |
| 550 | |
| 551 | if (t->size < 2) { |
| 552 | t->size = 0; |
| 553 | return; |
| 554 | } |
| 555 | |
| 556 | t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1; |
| 557 | |
| 558 | /* First we figure out where the first key in each cacheline is */ |
| 559 | for (j = inorder_next(0, t->size); |
| 560 | j; |
| 561 | j = inorder_next(j, t->size)) { |
| 562 | while (bkey_to_cacheline(t, k) != cacheline) |
| 563 | k = bkey_next(k); |
| 564 | |
| 565 | t->prev[j] = bkey_u64s(k); |
| 566 | k = bkey_next(k); |
| 567 | cacheline++; |
| 568 | t->tree[j].m = bkey_to_cacheline_offset(k); |
| 569 | } |
| 570 | |
| 571 | while (bkey_next(k) != end(t->data)) |
| 572 | k = bkey_next(k); |
| 573 | |
| 574 | t->end = *k; |
| 575 | |
| 576 | /* Then we build the tree */ |
| 577 | for (j = inorder_next(0, t->size); |
| 578 | j; |
| 579 | j = inorder_next(j, t->size)) |
| 580 | make_bfloat(t, j); |
| 581 | } |
| 582 | |
| 583 | void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k) |
| 584 | { |
| 585 | struct bset_tree *t; |
| 586 | unsigned inorder, j = 1; |
| 587 | |
| 588 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) |
| 589 | if (k < end(t->data)) |
| 590 | goto found_set; |
| 591 | |
| 592 | BUG(); |
| 593 | found_set: |
| 594 | if (!t->size || !bset_written(b, t)) |
| 595 | return; |
| 596 | |
| 597 | inorder = bkey_to_cacheline(t, k); |
| 598 | |
| 599 | if (k == t->data->start) |
| 600 | goto fix_left; |
| 601 | |
| 602 | if (bkey_next(k) == end(t->data)) { |
| 603 | t->end = *k; |
| 604 | goto fix_right; |
| 605 | } |
| 606 | |
| 607 | j = inorder_to_tree(inorder, t); |
| 608 | |
| 609 | if (j && |
| 610 | j < t->size && |
| 611 | k == tree_to_bkey(t, j)) |
| 612 | fix_left: do { |
| 613 | make_bfloat(t, j); |
| 614 | j = j * 2; |
| 615 | } while (j < t->size); |
| 616 | |
| 617 | j = inorder_to_tree(inorder + 1, t); |
| 618 | |
| 619 | if (j && |
| 620 | j < t->size && |
| 621 | k == tree_to_prev_bkey(t, j)) |
| 622 | fix_right: do { |
| 623 | make_bfloat(t, j); |
| 624 | j = j * 2 + 1; |
| 625 | } while (j < t->size); |
| 626 | } |
| 627 | |
| 628 | void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k) |
| 629 | { |
| 630 | struct bset_tree *t = &b->sets[b->nsets]; |
| 631 | unsigned shift = bkey_u64s(k); |
| 632 | unsigned j = bkey_to_cacheline(t, k); |
| 633 | |
| 634 | /* We're getting called from btree_split() or btree_gc, just bail out */ |
| 635 | if (!t->size) |
| 636 | return; |
| 637 | |
| 638 | /* k is the key we just inserted; we need to find the entry in the |
| 639 | * lookup table for the first key that is strictly greater than k: |
| 640 | * it's either k's cacheline or the next one |
| 641 | */ |
| 642 | if (j < t->size && |
| 643 | table_to_bkey(t, j) <= k) |
| 644 | j++; |
| 645 | |
| 646 | /* Adjust all the lookup table entries, and find a new key for any that |
| 647 | * have gotten too big |
| 648 | */ |
| 649 | for (; j < t->size; j++) { |
| 650 | t->prev[j] += shift; |
| 651 | |
| 652 | if (t->prev[j] > 7) { |
| 653 | k = table_to_bkey(t, j - 1); |
| 654 | |
| 655 | while (k < cacheline_to_bkey(t, j, 0)) |
| 656 | k = bkey_next(k); |
| 657 | |
| 658 | t->prev[j] = bkey_to_cacheline_offset(k); |
| 659 | } |
| 660 | } |
| 661 | |
| 662 | if (t->size == b->sets->tree + bset_tree_space(b) - t->tree) |
| 663 | return; |
| 664 | |
| 665 | /* Possibly add a new entry to the end of the lookup table */ |
| 666 | |
| 667 | for (k = table_to_bkey(t, t->size - 1); |
| 668 | k != end(t->data); |
| 669 | k = bkey_next(k)) |
| 670 | if (t->size == bkey_to_cacheline(t, k)) { |
| 671 | t->prev[t->size] = bkey_to_cacheline_offset(k); |
| 672 | t->size++; |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | void bch_bset_init_next(struct btree *b) |
| 677 | { |
| 678 | struct bset *i = write_block(b); |
| 679 | |
| 680 | if (i != b->sets[0].data) { |
| 681 | b->sets[++b->nsets].data = i; |
| 682 | i->seq = b->sets[0].data->seq; |
| 683 | } else |
| 684 | get_random_bytes(&i->seq, sizeof(uint64_t)); |
| 685 | |
Kent Overstreet | 81ab419 | 2013-10-31 15:46:42 -0700 | [diff] [blame] | 686 | i->magic = bset_magic(&b->c->sb); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 687 | i->version = 0; |
| 688 | i->keys = 0; |
| 689 | |
| 690 | bset_build_unwritten_tree(b); |
| 691 | } |
| 692 | |
| 693 | struct bset_search_iter { |
| 694 | struct bkey *l, *r; |
| 695 | }; |
| 696 | |
| 697 | static struct bset_search_iter bset_search_write_set(struct btree *b, |
| 698 | struct bset_tree *t, |
| 699 | const struct bkey *search) |
| 700 | { |
| 701 | unsigned li = 0, ri = t->size; |
| 702 | |
| 703 | BUG_ON(!b->nsets && |
| 704 | t->size < bkey_to_cacheline(t, end(t->data))); |
| 705 | |
| 706 | while (li + 1 != ri) { |
| 707 | unsigned m = (li + ri) >> 1; |
| 708 | |
| 709 | if (bkey_cmp(table_to_bkey(t, m), search) > 0) |
| 710 | ri = m; |
| 711 | else |
| 712 | li = m; |
| 713 | } |
| 714 | |
| 715 | return (struct bset_search_iter) { |
| 716 | table_to_bkey(t, li), |
| 717 | ri < t->size ? table_to_bkey(t, ri) : end(t->data) |
| 718 | }; |
| 719 | } |
| 720 | |
| 721 | static struct bset_search_iter bset_search_tree(struct btree *b, |
| 722 | struct bset_tree *t, |
| 723 | const struct bkey *search) |
| 724 | { |
| 725 | struct bkey *l, *r; |
| 726 | struct bkey_float *f; |
| 727 | unsigned inorder, j, n = 1; |
| 728 | |
| 729 | do { |
| 730 | unsigned p = n << 4; |
| 731 | p &= ((int) (p - t->size)) >> 31; |
| 732 | |
| 733 | prefetch(&t->tree[p]); |
| 734 | |
| 735 | j = n; |
| 736 | f = &t->tree[j]; |
| 737 | |
| 738 | /* |
| 739 | * n = (f->mantissa > bfloat_mantissa()) |
| 740 | * ? j * 2 |
| 741 | * : j * 2 + 1; |
| 742 | * |
| 743 | * We need to subtract 1 from f->mantissa for the sign bit trick |
| 744 | * to work - that's done in make_bfloat() |
| 745 | */ |
| 746 | if (likely(f->exponent != 127)) |
| 747 | n = j * 2 + (((unsigned) |
| 748 | (f->mantissa - |
| 749 | bfloat_mantissa(search, f))) >> 31); |
| 750 | else |
| 751 | n = (bkey_cmp(tree_to_bkey(t, j), search) > 0) |
| 752 | ? j * 2 |
| 753 | : j * 2 + 1; |
| 754 | } while (n < t->size); |
| 755 | |
| 756 | inorder = to_inorder(j, t); |
| 757 | |
| 758 | /* |
| 759 | * n would have been the node we recursed to - the low bit tells us if |
| 760 | * we recursed left or recursed right. |
| 761 | */ |
| 762 | if (n & 1) { |
| 763 | l = cacheline_to_bkey(t, inorder, f->m); |
| 764 | |
| 765 | if (++inorder != t->size) { |
| 766 | f = &t->tree[inorder_next(j, t->size)]; |
| 767 | r = cacheline_to_bkey(t, inorder, f->m); |
| 768 | } else |
| 769 | r = end(t->data); |
| 770 | } else { |
| 771 | r = cacheline_to_bkey(t, inorder, f->m); |
| 772 | |
| 773 | if (--inorder) { |
| 774 | f = &t->tree[inorder_prev(j, t->size)]; |
| 775 | l = cacheline_to_bkey(t, inorder, f->m); |
| 776 | } else |
| 777 | l = t->data->start; |
| 778 | } |
| 779 | |
| 780 | return (struct bset_search_iter) {l, r}; |
| 781 | } |
| 782 | |
| 783 | struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t, |
| 784 | const struct bkey *search) |
| 785 | { |
| 786 | struct bset_search_iter i; |
| 787 | |
| 788 | /* |
| 789 | * First, we search for a cacheline, then lastly we do a linear search |
| 790 | * within that cacheline. |
| 791 | * |
| 792 | * To search for the cacheline, there's three different possibilities: |
| 793 | * * The set is too small to have a search tree, so we just do a linear |
| 794 | * search over the whole set. |
| 795 | * * The set is the one we're currently inserting into; keeping a full |
| 796 | * auxiliary search tree up to date would be too expensive, so we |
| 797 | * use a much simpler lookup table to do a binary search - |
| 798 | * bset_search_write_set(). |
| 799 | * * Or we use the auxiliary search tree we constructed earlier - |
| 800 | * bset_search_tree() |
| 801 | */ |
| 802 | |
| 803 | if (unlikely(!t->size)) { |
| 804 | i.l = t->data->start; |
| 805 | i.r = end(t->data); |
| 806 | } else if (bset_written(b, t)) { |
| 807 | /* |
| 808 | * Each node in the auxiliary search tree covers a certain range |
| 809 | * of bits, and keys above and below the set it covers might |
| 810 | * differ outside those bits - so we have to special case the |
| 811 | * start and end - handle that here: |
| 812 | */ |
| 813 | |
| 814 | if (unlikely(bkey_cmp(search, &t->end) >= 0)) |
| 815 | return end(t->data); |
| 816 | |
| 817 | if (unlikely(bkey_cmp(search, t->data->start) < 0)) |
| 818 | return t->data->start; |
| 819 | |
| 820 | i = bset_search_tree(b, t, search); |
| 821 | } else |
| 822 | i = bset_search_write_set(b, t, search); |
| 823 | |
| 824 | #ifdef CONFIG_BCACHE_EDEBUG |
| 825 | BUG_ON(bset_written(b, t) && |
| 826 | i.l != t->data->start && |
| 827 | bkey_cmp(tree_to_prev_bkey(t, |
| 828 | inorder_to_tree(bkey_to_cacheline(t, i.l), t)), |
| 829 | search) > 0); |
| 830 | |
| 831 | BUG_ON(i.r != end(t->data) && |
| 832 | bkey_cmp(i.r, search) <= 0); |
| 833 | #endif |
| 834 | |
| 835 | while (likely(i.l != i.r) && |
| 836 | bkey_cmp(i.l, search) <= 0) |
| 837 | i.l = bkey_next(i.l); |
| 838 | |
| 839 | return i.l; |
| 840 | } |
| 841 | |
| 842 | /* Btree iterator */ |
| 843 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 844 | /* |
| 845 | * Returns true if l > r - unless l == r, in which case returns true if l is |
| 846 | * older than r. |
| 847 | * |
| 848 | * Necessary for btree_sort_fixup() - if there are multiple keys that compare |
| 849 | * equal in different sets, we have to process them newest to oldest. |
| 850 | */ |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 851 | static inline bool btree_iter_cmp(struct btree_iter_set l, |
| 852 | struct btree_iter_set r) |
| 853 | { |
| 854 | int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k)); |
| 855 | |
| 856 | return c ? c > 0 : l.k < r.k; |
| 857 | } |
| 858 | |
| 859 | static inline bool btree_iter_end(struct btree_iter *iter) |
| 860 | { |
| 861 | return !iter->used; |
| 862 | } |
| 863 | |
| 864 | void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k, |
| 865 | struct bkey *end) |
| 866 | { |
| 867 | if (k != end) |
| 868 | BUG_ON(!heap_add(iter, |
| 869 | ((struct btree_iter_set) { k, end }), |
| 870 | btree_iter_cmp)); |
| 871 | } |
| 872 | |
| 873 | struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter, |
| 874 | struct bkey *search, struct bset_tree *start) |
| 875 | { |
| 876 | struct bkey *ret = NULL; |
| 877 | iter->size = ARRAY_SIZE(iter->data); |
| 878 | iter->used = 0; |
| 879 | |
| 880 | for (; start <= &b->sets[b->nsets]; start++) { |
| 881 | ret = bch_bset_search(b, start, search); |
| 882 | bch_btree_iter_push(iter, ret, end(start->data)); |
| 883 | } |
| 884 | |
| 885 | return ret; |
| 886 | } |
| 887 | |
| 888 | struct bkey *bch_btree_iter_next(struct btree_iter *iter) |
| 889 | { |
| 890 | struct btree_iter_set unused; |
| 891 | struct bkey *ret = NULL; |
| 892 | |
| 893 | if (!btree_iter_end(iter)) { |
| 894 | ret = iter->data->k; |
| 895 | iter->data->k = bkey_next(iter->data->k); |
| 896 | |
| 897 | if (iter->data->k > iter->data->end) { |
Kent Overstreet | cc0f4ea | 2013-03-27 12:47:45 -0700 | [diff] [blame] | 898 | WARN_ONCE(1, "bset was corrupt!\n"); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 899 | iter->data->k = iter->data->end; |
| 900 | } |
| 901 | |
| 902 | if (iter->data->k == iter->data->end) |
| 903 | heap_pop(iter, unused, btree_iter_cmp); |
| 904 | else |
| 905 | heap_sift(iter, 0, btree_iter_cmp); |
| 906 | } |
| 907 | |
| 908 | return ret; |
| 909 | } |
| 910 | |
| 911 | struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter, |
| 912 | struct btree *b, ptr_filter_fn fn) |
| 913 | { |
| 914 | struct bkey *ret; |
| 915 | |
| 916 | do { |
| 917 | ret = bch_btree_iter_next(iter); |
| 918 | } while (ret && fn(b, ret)); |
| 919 | |
| 920 | return ret; |
| 921 | } |
| 922 | |
| 923 | struct bkey *bch_next_recurse_key(struct btree *b, struct bkey *search) |
| 924 | { |
| 925 | struct btree_iter iter; |
| 926 | |
| 927 | bch_btree_iter_init(b, &iter, search); |
| 928 | return bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); |
| 929 | } |
| 930 | |
| 931 | /* Mergesort */ |
| 932 | |
Kent Overstreet | 8478643 | 2013-09-23 23:17:35 -0700 | [diff] [blame] | 933 | static void sort_key_next(struct btree_iter *iter, |
| 934 | struct btree_iter_set *i) |
| 935 | { |
| 936 | i->k = bkey_next(i->k); |
| 937 | |
| 938 | if (i->k == i->end) |
| 939 | *i = iter->data[--iter->used]; |
| 940 | } |
| 941 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 942 | static void btree_sort_fixup(struct btree_iter *iter) |
| 943 | { |
| 944 | while (iter->used > 1) { |
| 945 | struct btree_iter_set *top = iter->data, *i = top + 1; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 946 | |
| 947 | if (iter->used > 2 && |
| 948 | btree_iter_cmp(i[0], i[1])) |
| 949 | i++; |
| 950 | |
Kent Overstreet | 8478643 | 2013-09-23 23:17:35 -0700 | [diff] [blame] | 951 | if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 952 | break; |
| 953 | |
Kent Overstreet | 8478643 | 2013-09-23 23:17:35 -0700 | [diff] [blame] | 954 | if (!KEY_SIZE(i->k)) { |
| 955 | sort_key_next(iter, i); |
| 956 | heap_sift(iter, i - top, btree_iter_cmp); |
| 957 | continue; |
| 958 | } |
| 959 | |
| 960 | if (top->k > i->k) { |
| 961 | if (bkey_cmp(top->k, i->k) >= 0) |
| 962 | sort_key_next(iter, i); |
| 963 | else |
| 964 | bch_cut_front(top->k, i->k); |
| 965 | |
| 966 | heap_sift(iter, i - top, btree_iter_cmp); |
| 967 | } else { |
| 968 | /* can't happen because of comparison func */ |
| 969 | BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k))); |
| 970 | bch_cut_back(&START_KEY(i->k), top->k); |
| 971 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 972 | } |
| 973 | } |
| 974 | |
| 975 | static void btree_mergesort(struct btree *b, struct bset *out, |
| 976 | struct btree_iter *iter, |
| 977 | bool fixup, bool remove_stale) |
| 978 | { |
| 979 | struct bkey *k, *last = NULL; |
| 980 | bool (*bad)(struct btree *, const struct bkey *) = remove_stale |
| 981 | ? bch_ptr_bad |
| 982 | : bch_ptr_invalid; |
| 983 | |
| 984 | while (!btree_iter_end(iter)) { |
| 985 | if (fixup && !b->level) |
| 986 | btree_sort_fixup(iter); |
| 987 | |
| 988 | k = bch_btree_iter_next(iter); |
| 989 | if (bad(b, k)) |
| 990 | continue; |
| 991 | |
| 992 | if (!last) { |
| 993 | last = out->start; |
| 994 | bkey_copy(last, k); |
| 995 | } else if (b->level || |
| 996 | !bch_bkey_try_merge(b, last, k)) { |
| 997 | last = bkey_next(last); |
| 998 | bkey_copy(last, k); |
| 999 | } |
| 1000 | } |
| 1001 | |
| 1002 | out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0; |
| 1003 | |
| 1004 | pr_debug("sorted %i keys", out->keys); |
| 1005 | bch_check_key_order(b, out); |
| 1006 | } |
| 1007 | |
| 1008 | static void __btree_sort(struct btree *b, struct btree_iter *iter, |
| 1009 | unsigned start, unsigned order, bool fixup) |
| 1010 | { |
| 1011 | uint64_t start_time; |
| 1012 | bool remove_stale = !b->written; |
| 1013 | struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO, |
| 1014 | order); |
| 1015 | if (!out) { |
| 1016 | mutex_lock(&b->c->sort_lock); |
| 1017 | out = b->c->sort; |
| 1018 | order = ilog2(bucket_pages(b->c)); |
| 1019 | } |
| 1020 | |
| 1021 | start_time = local_clock(); |
| 1022 | |
| 1023 | btree_mergesort(b, out, iter, fixup, remove_stale); |
| 1024 | b->nsets = start; |
| 1025 | |
| 1026 | if (!fixup && !start && b->written) |
| 1027 | bch_btree_verify(b, out); |
| 1028 | |
| 1029 | if (!start && order == b->page_order) { |
| 1030 | /* |
| 1031 | * Our temporary buffer is the same size as the btree node's |
| 1032 | * buffer, we can just swap buffers instead of doing a big |
| 1033 | * memcpy() |
| 1034 | */ |
| 1035 | |
Kent Overstreet | 81ab419 | 2013-10-31 15:46:42 -0700 | [diff] [blame] | 1036 | out->magic = bset_magic(&b->c->sb); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1037 | out->seq = b->sets[0].data->seq; |
| 1038 | out->version = b->sets[0].data->version; |
| 1039 | swap(out, b->sets[0].data); |
| 1040 | |
| 1041 | if (b->c->sort == b->sets[0].data) |
| 1042 | b->c->sort = out; |
| 1043 | } else { |
| 1044 | b->sets[start].data->keys = out->keys; |
| 1045 | memcpy(b->sets[start].data->start, out->start, |
| 1046 | (void *) end(out) - (void *) out->start); |
| 1047 | } |
| 1048 | |
| 1049 | if (out == b->c->sort) |
| 1050 | mutex_unlock(&b->c->sort_lock); |
| 1051 | else |
| 1052 | free_pages((unsigned long) out, order); |
| 1053 | |
| 1054 | if (b->written) |
| 1055 | bset_build_written_tree(b); |
| 1056 | |
| 1057 | if (!start) { |
| 1058 | spin_lock(&b->c->sort_time_lock); |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 1059 | bch_time_stats_update(&b->c->sort_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1060 | spin_unlock(&b->c->sort_time_lock); |
| 1061 | } |
| 1062 | } |
| 1063 | |
| 1064 | void bch_btree_sort_partial(struct btree *b, unsigned start) |
| 1065 | { |
| 1066 | size_t oldsize = 0, order = b->page_order, keys = 0; |
| 1067 | struct btree_iter iter; |
| 1068 | __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]); |
| 1069 | |
| 1070 | BUG_ON(b->sets[b->nsets].data == write_block(b) && |
| 1071 | (b->sets[b->nsets].size || b->nsets)); |
| 1072 | |
| 1073 | if (b->written) |
| 1074 | oldsize = bch_count_data(b); |
| 1075 | |
| 1076 | if (start) { |
| 1077 | unsigned i; |
| 1078 | |
| 1079 | for (i = start; i <= b->nsets; i++) |
| 1080 | keys += b->sets[i].data->keys; |
| 1081 | |
Kent Overstreet | b1a67b0 | 2013-03-25 11:46:44 -0700 | [diff] [blame] | 1082 | order = roundup_pow_of_two(__set_bytes(b->sets->data, |
| 1083 | keys)) / PAGE_SIZE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1084 | if (order) |
| 1085 | order = ilog2(order); |
| 1086 | } |
| 1087 | |
| 1088 | __btree_sort(b, &iter, start, order, false); |
| 1089 | |
| 1090 | EBUG_ON(b->written && bch_count_data(b) != oldsize); |
| 1091 | } |
| 1092 | |
| 1093 | void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter) |
| 1094 | { |
| 1095 | BUG_ON(!b->written); |
| 1096 | __btree_sort(b, iter, 0, b->page_order, true); |
| 1097 | } |
| 1098 | |
| 1099 | void bch_btree_sort_into(struct btree *b, struct btree *new) |
| 1100 | { |
| 1101 | uint64_t start_time = local_clock(); |
| 1102 | |
| 1103 | struct btree_iter iter; |
| 1104 | bch_btree_iter_init(b, &iter, NULL); |
| 1105 | |
| 1106 | btree_mergesort(b, new->sets->data, &iter, false, true); |
| 1107 | |
| 1108 | spin_lock(&b->c->sort_time_lock); |
Kent Overstreet | 169ef1c | 2013-03-28 12:50:55 -0600 | [diff] [blame] | 1109 | bch_time_stats_update(&b->c->sort_time, start_time); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1110 | spin_unlock(&b->c->sort_time_lock); |
| 1111 | |
| 1112 | bkey_copy_key(&new->key, &b->key); |
| 1113 | new->sets->size = 0; |
| 1114 | } |
| 1115 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1116 | #define SORT_CRIT (4096 / sizeof(uint64_t)) |
| 1117 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1118 | void bch_btree_sort_lazy(struct btree *b) |
| 1119 | { |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1120 | unsigned crit = SORT_CRIT; |
| 1121 | int i; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1122 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1123 | /* Don't sort if nothing to do */ |
| 1124 | if (!b->nsets) |
| 1125 | goto out; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1126 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1127 | /* If not a leaf node, always sort */ |
| 1128 | if (b->level) { |
| 1129 | bch_btree_sort(b); |
| 1130 | return; |
| 1131 | } |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1132 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1133 | for (i = b->nsets - 1; i >= 0; --i) { |
| 1134 | crit *= b->c->sort_crit_factor; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1135 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1136 | if (b->sets[i].data->keys < crit) { |
| 1137 | bch_btree_sort_partial(b, i); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1138 | return; |
| 1139 | } |
| 1140 | } |
| 1141 | |
Kent Overstreet | 6ded34d | 2013-05-11 15:59:37 -0700 | [diff] [blame] | 1142 | /* Sort if we'd overflow */ |
| 1143 | if (b->nsets + 1 == MAX_BSETS) { |
| 1144 | bch_btree_sort(b); |
| 1145 | return; |
| 1146 | } |
| 1147 | |
| 1148 | out: |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1149 | bset_build_written_tree(b); |
| 1150 | } |
| 1151 | |
| 1152 | /* Sysfs stuff */ |
| 1153 | |
| 1154 | struct bset_stats { |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 1155 | struct btree_op op; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1156 | size_t nodes; |
| 1157 | size_t sets_written, sets_unwritten; |
| 1158 | size_t bytes_written, bytes_unwritten; |
| 1159 | size_t floats, failed; |
| 1160 | }; |
| 1161 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 1162 | static int btree_bset_stats(struct btree_op *op, struct btree *b) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1163 | { |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 1164 | struct bset_stats *stats = container_of(op, struct bset_stats, op); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1165 | unsigned i; |
| 1166 | |
| 1167 | stats->nodes++; |
| 1168 | |
| 1169 | for (i = 0; i <= b->nsets; i++) { |
| 1170 | struct bset_tree *t = &b->sets[i]; |
| 1171 | size_t bytes = t->data->keys * sizeof(uint64_t); |
| 1172 | size_t j; |
| 1173 | |
| 1174 | if (bset_written(b, t)) { |
| 1175 | stats->sets_written++; |
| 1176 | stats->bytes_written += bytes; |
| 1177 | |
| 1178 | stats->floats += t->size - 1; |
| 1179 | |
| 1180 | for (j = 1; j < t->size; j++) |
| 1181 | if (t->tree[j].exponent == 127) |
| 1182 | stats->failed++; |
| 1183 | } else { |
| 1184 | stats->sets_unwritten++; |
| 1185 | stats->bytes_unwritten += bytes; |
| 1186 | } |
| 1187 | } |
| 1188 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 1189 | return MAP_CONTINUE; |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1190 | } |
| 1191 | |
| 1192 | int bch_bset_print_stats(struct cache_set *c, char *buf) |
| 1193 | { |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1194 | struct bset_stats t; |
| 1195 | int ret; |
| 1196 | |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1197 | memset(&t, 0, sizeof(struct bset_stats)); |
Kent Overstreet | b54d693 | 2013-07-24 18:04:18 -0700 | [diff] [blame] | 1198 | bch_btree_op_init(&t.op, -1); |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1199 | |
Kent Overstreet | 48dad8b | 2013-09-10 18:48:51 -0700 | [diff] [blame] | 1200 | ret = bch_btree_map_nodes(&t.op, c, &ZERO_KEY, btree_bset_stats); |
| 1201 | if (ret < 0) |
Kent Overstreet | cafe563 | 2013-03-23 16:11:31 -0700 | [diff] [blame] | 1202 | return ret; |
| 1203 | |
| 1204 | return snprintf(buf, PAGE_SIZE, |
| 1205 | "btree nodes: %zu\n" |
| 1206 | "written sets: %zu\n" |
| 1207 | "unwritten sets: %zu\n" |
| 1208 | "written key bytes: %zu\n" |
| 1209 | "unwritten key bytes: %zu\n" |
| 1210 | "floats: %zu\n" |
| 1211 | "failed: %zu\n", |
| 1212 | t.nodes, |
| 1213 | t.sets_written, t.sets_unwritten, |
| 1214 | t.bytes_written, t.bytes_unwritten, |
| 1215 | t.floats, t.failed); |
| 1216 | } |