]>
Commit | Line | Data |
---|---|---|
cafe5635 KO |
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> | |
cd953ed0 | 13 | #include <linux/prefetch.h> |
cafe5635 KO |
14 | |
15 | /* Keylists */ | |
16 | ||
17 | void bch_keylist_copy(struct keylist *dest, struct keylist *src) | |
18 | { | |
19 | *dest = *src; | |
20 | ||
21 | if (src->list == src->d) { | |
22 | size_t n = (uint64_t *) src->top - src->d; | |
23 | dest->top = (struct bkey *) &dest->d[n]; | |
24 | dest->list = dest->d; | |
25 | } | |
26 | } | |
27 | ||
28 | int bch_keylist_realloc(struct keylist *l, int nptrs, struct cache_set *c) | |
29 | { | |
30 | unsigned oldsize = (uint64_t *) l->top - l->list; | |
31 | unsigned newsize = oldsize + 2 + nptrs; | |
32 | uint64_t *new; | |
33 | ||
34 | /* The journalling code doesn't handle the case where the keys to insert | |
35 | * is bigger than an empty write: If we just return -ENOMEM here, | |
36 | * bio_insert() and bio_invalidate() will insert the keys created so far | |
37 | * and finish the rest when the keylist is empty. | |
38 | */ | |
39 | if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset)) | |
40 | return -ENOMEM; | |
41 | ||
42 | newsize = roundup_pow_of_two(newsize); | |
43 | ||
44 | if (newsize <= KEYLIST_INLINE || | |
45 | roundup_pow_of_two(oldsize) == newsize) | |
46 | return 0; | |
47 | ||
48 | new = krealloc(l->list == l->d ? NULL : l->list, | |
49 | sizeof(uint64_t) * newsize, GFP_NOIO); | |
50 | ||
51 | if (!new) | |
52 | return -ENOMEM; | |
53 | ||
54 | if (l->list == l->d) | |
55 | memcpy(new, l->list, sizeof(uint64_t) * KEYLIST_INLINE); | |
56 | ||
57 | l->list = new; | |
58 | l->top = (struct bkey *) (&l->list[oldsize]); | |
59 | ||
60 | return 0; | |
61 | } | |
62 | ||
63 | struct bkey *bch_keylist_pop(struct keylist *l) | |
64 | { | |
65 | struct bkey *k = l->bottom; | |
66 | ||
67 | if (k == l->top) | |
68 | return NULL; | |
69 | ||
70 | while (bkey_next(k) != l->top) | |
71 | k = bkey_next(k); | |
72 | ||
73 | return l->top = k; | |
74 | } | |
75 | ||
76 | /* Pointer validation */ | |
77 | ||
78 | bool __bch_ptr_invalid(struct cache_set *c, int level, const struct bkey *k) | |
79 | { | |
80 | unsigned i; | |
81 | ||
82 | if (level && (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))) | |
83 | goto bad; | |
84 | ||
85 | if (!level && KEY_SIZE(k) > KEY_OFFSET(k)) | |
86 | goto bad; | |
87 | ||
88 | if (!KEY_SIZE(k)) | |
89 | return true; | |
90 | ||
91 | for (i = 0; i < KEY_PTRS(k); i++) | |
92 | if (ptr_available(c, k, i)) { | |
93 | struct cache *ca = PTR_CACHE(c, k, i); | |
94 | size_t bucket = PTR_BUCKET_NR(c, k, i); | |
95 | size_t r = bucket_remainder(c, PTR_OFFSET(k, i)); | |
96 | ||
97 | if (KEY_SIZE(k) + r > c->sb.bucket_size || | |
98 | bucket < ca->sb.first_bucket || | |
99 | bucket >= ca->sb.nbuckets) | |
100 | goto bad; | |
101 | } | |
102 | ||
103 | return false; | |
104 | bad: | |
105 | cache_bug(c, "spotted bad key %s: %s", pkey(k), bch_ptr_status(c, k)); | |
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 | ||
119 | if (KEY_PTRS(k) && PTR_DEV(k, 0) == PTR_CHECK_DEV) | |
120 | return true; | |
121 | ||
122 | for (i = 0; i < KEY_PTRS(k); i++) | |
123 | if (ptr_available(b->c, k, i)) { | |
124 | g = PTR_BUCKET(b->c, k, i); | |
125 | stale = ptr_stale(b->c, k, i); | |
126 | ||
127 | btree_bug_on(stale > 96, b, | |
128 | "key too stale: %i, need_gc %u", | |
129 | stale, b->c->need_gc); | |
130 | ||
131 | btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k), | |
132 | b, "stale dirty pointer"); | |
133 | ||
134 | if (stale) | |
135 | return true; | |
136 | ||
137 | #ifdef CONFIG_BCACHE_EDEBUG | |
138 | if (!mutex_trylock(&b->c->bucket_lock)) | |
139 | continue; | |
140 | ||
141 | if (b->level) { | |
142 | if (KEY_DIRTY(k) || | |
143 | g->prio != BTREE_PRIO || | |
144 | (b->c->gc_mark_valid && | |
145 | GC_MARK(g) != GC_MARK_METADATA)) | |
146 | goto bug; | |
147 | ||
148 | } else { | |
149 | if (g->prio == BTREE_PRIO) | |
150 | goto bug; | |
151 | ||
152 | if (KEY_DIRTY(k) && | |
153 | b->c->gc_mark_valid && | |
154 | GC_MARK(g) != GC_MARK_DIRTY) | |
155 | goto bug; | |
156 | } | |
157 | mutex_unlock(&b->c->bucket_lock); | |
158 | #endif | |
159 | } | |
160 | ||
161 | return false; | |
162 | #ifdef CONFIG_BCACHE_EDEBUG | |
163 | bug: | |
164 | mutex_unlock(&b->c->bucket_lock); | |
b1a67b0f KO |
165 | btree_bug(b, |
166 | "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i", | |
167 | pkey(k), PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin), | |
cafe5635 KO |
168 | g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen); |
169 | return true; | |
170 | #endif | |
171 | } | |
172 | ||
173 | /* Key/pointer manipulation */ | |
174 | ||
175 | void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src, | |
176 | unsigned i) | |
177 | { | |
178 | BUG_ON(i > KEY_PTRS(src)); | |
179 | ||
180 | /* Only copy the header, key, and one pointer. */ | |
181 | memcpy(dest, src, 2 * sizeof(uint64_t)); | |
182 | dest->ptr[0] = src->ptr[i]; | |
183 | SET_KEY_PTRS(dest, 1); | |
184 | /* We didn't copy the checksum so clear that bit. */ | |
185 | SET_KEY_CSUM(dest, 0); | |
186 | } | |
187 | ||
188 | bool __bch_cut_front(const struct bkey *where, struct bkey *k) | |
189 | { | |
190 | unsigned i, len = 0; | |
191 | ||
192 | if (bkey_cmp(where, &START_KEY(k)) <= 0) | |
193 | return false; | |
194 | ||
195 | if (bkey_cmp(where, k) < 0) | |
196 | len = KEY_OFFSET(k) - KEY_OFFSET(where); | |
197 | else | |
198 | bkey_copy_key(k, where); | |
199 | ||
200 | for (i = 0; i < KEY_PTRS(k); i++) | |
201 | SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len); | |
202 | ||
203 | BUG_ON(len > KEY_SIZE(k)); | |
204 | SET_KEY_SIZE(k, len); | |
205 | return true; | |
206 | } | |
207 | ||
208 | bool __bch_cut_back(const struct bkey *where, struct bkey *k) | |
209 | { | |
210 | unsigned len = 0; | |
211 | ||
212 | if (bkey_cmp(where, k) >= 0) | |
213 | return false; | |
214 | ||
215 | BUG_ON(KEY_INODE(where) != KEY_INODE(k)); | |
216 | ||
217 | if (bkey_cmp(where, &START_KEY(k)) > 0) | |
218 | len = KEY_OFFSET(where) - KEY_START(k); | |
219 | ||
220 | bkey_copy_key(k, where); | |
221 | ||
222 | BUG_ON(len > KEY_SIZE(k)); | |
223 | SET_KEY_SIZE(k, len); | |
224 | return true; | |
225 | } | |
226 | ||
227 | static uint64_t merge_chksums(struct bkey *l, struct bkey *r) | |
228 | { | |
229 | return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) & | |
230 | ~((uint64_t)1 << 63); | |
231 | } | |
232 | ||
233 | /* Tries to merge l and r: l should be lower than r | |
234 | * Returns true if we were able to merge. If we did merge, l will be the merged | |
235 | * key, r will be untouched. | |
236 | */ | |
237 | bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r) | |
238 | { | |
239 | unsigned i; | |
240 | ||
241 | if (key_merging_disabled(b->c)) | |
242 | return false; | |
243 | ||
244 | if (KEY_PTRS(l) != KEY_PTRS(r) || | |
245 | KEY_DIRTY(l) != KEY_DIRTY(r) || | |
246 | bkey_cmp(l, &START_KEY(r))) | |
247 | return false; | |
248 | ||
249 | for (i = 0; i < KEY_PTRS(l); i++) | |
250 | if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] || | |
251 | PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i)) | |
252 | return false; | |
253 | ||
254 | /* Keys with no pointers aren't restricted to one bucket and could | |
255 | * overflow KEY_SIZE | |
256 | */ | |
257 | if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) { | |
258 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l)); | |
259 | SET_KEY_SIZE(l, USHRT_MAX); | |
260 | ||
261 | bch_cut_front(l, r); | |
262 | return false; | |
263 | } | |
264 | ||
265 | if (KEY_CSUM(l)) { | |
266 | if (KEY_CSUM(r)) | |
267 | l->ptr[KEY_PTRS(l)] = merge_chksums(l, r); | |
268 | else | |
269 | SET_KEY_CSUM(l, 0); | |
270 | } | |
271 | ||
272 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r)); | |
273 | SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r)); | |
274 | ||
275 | return true; | |
276 | } | |
277 | ||
278 | /* Binary tree stuff for auxiliary search trees */ | |
279 | ||
280 | static unsigned inorder_next(unsigned j, unsigned size) | |
281 | { | |
282 | if (j * 2 + 1 < size) { | |
283 | j = j * 2 + 1; | |
284 | ||
285 | while (j * 2 < size) | |
286 | j *= 2; | |
287 | } else | |
288 | j >>= ffz(j) + 1; | |
289 | ||
290 | return j; | |
291 | } | |
292 | ||
293 | static unsigned inorder_prev(unsigned j, unsigned size) | |
294 | { | |
295 | if (j * 2 < size) { | |
296 | j = j * 2; | |
297 | ||
298 | while (j * 2 + 1 < size) | |
299 | j = j * 2 + 1; | |
300 | } else | |
301 | j >>= ffs(j); | |
302 | ||
303 | return j; | |
304 | } | |
305 | ||
306 | /* I have no idea why this code works... and I'm the one who wrote it | |
307 | * | |
308 | * However, I do know what it does: | |
309 | * Given a binary tree constructed in an array (i.e. how you normally implement | |
310 | * a heap), it converts a node in the tree - referenced by array index - to the | |
311 | * index it would have if you did an inorder traversal. | |
312 | * | |
313 | * Also tested for every j, size up to size somewhere around 6 million. | |
314 | * | |
315 | * The binary tree starts at array index 1, not 0 | |
316 | * extra is a function of size: | |
317 | * extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
318 | */ | |
319 | static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra) | |
320 | { | |
321 | unsigned b = fls(j); | |
322 | unsigned shift = fls(size - 1) - b; | |
323 | ||
324 | j ^= 1U << (b - 1); | |
325 | j <<= 1; | |
326 | j |= 1; | |
327 | j <<= shift; | |
328 | ||
329 | if (j > extra) | |
330 | j -= (j - extra) >> 1; | |
331 | ||
332 | return j; | |
333 | } | |
334 | ||
335 | static unsigned to_inorder(unsigned j, struct bset_tree *t) | |
336 | { | |
337 | return __to_inorder(j, t->size, t->extra); | |
338 | } | |
339 | ||
340 | static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra) | |
341 | { | |
342 | unsigned shift; | |
343 | ||
344 | if (j > extra) | |
345 | j += j - extra; | |
346 | ||
347 | shift = ffs(j); | |
348 | ||
349 | j >>= shift; | |
350 | j |= roundup_pow_of_two(size) >> shift; | |
351 | ||
352 | return j; | |
353 | } | |
354 | ||
355 | static unsigned inorder_to_tree(unsigned j, struct bset_tree *t) | |
356 | { | |
357 | return __inorder_to_tree(j, t->size, t->extra); | |
358 | } | |
359 | ||
360 | #if 0 | |
361 | void inorder_test(void) | |
362 | { | |
363 | unsigned long done = 0; | |
364 | ktime_t start = ktime_get(); | |
365 | ||
366 | for (unsigned size = 2; | |
367 | size < 65536000; | |
368 | size++) { | |
369 | unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
370 | unsigned i = 1, j = rounddown_pow_of_two(size - 1); | |
371 | ||
372 | if (!(size % 4096)) | |
373 | printk(KERN_NOTICE "loop %u, %llu per us\n", size, | |
374 | done / ktime_us_delta(ktime_get(), start)); | |
375 | ||
376 | while (1) { | |
377 | if (__inorder_to_tree(i, size, extra) != j) | |
378 | panic("size %10u j %10u i %10u", size, j, i); | |
379 | ||
380 | if (__to_inorder(j, size, extra) != i) | |
381 | panic("size %10u j %10u i %10u", size, j, i); | |
382 | ||
383 | if (j == rounddown_pow_of_two(size) - 1) | |
384 | break; | |
385 | ||
386 | BUG_ON(inorder_prev(inorder_next(j, size), size) != j); | |
387 | ||
388 | j = inorder_next(j, size); | |
389 | i++; | |
390 | } | |
391 | ||
392 | done += size - 1; | |
393 | } | |
394 | } | |
395 | #endif | |
396 | ||
397 | /* | |
398 | * Cacheline/offset <-> bkey pointer arithmatic: | |
399 | * | |
400 | * t->tree is a binary search tree in an array; each node corresponds to a key | |
401 | * in one cacheline in t->set (BSET_CACHELINE bytes). | |
402 | * | |
403 | * This means we don't have to store the full index of the key that a node in | |
404 | * the binary tree points to; to_inorder() gives us the cacheline, and then | |
405 | * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes. | |
406 | * | |
407 | * cacheline_to_bkey() and friends abstract out all the pointer arithmatic to | |
408 | * make this work. | |
409 | * | |
410 | * To construct the bfloat for an arbitrary key we need to know what the key | |
411 | * immediately preceding it is: we have to check if the two keys differ in the | |
412 | * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size | |
413 | * of the previous key so we can walk backwards to it from t->tree[j]'s key. | |
414 | */ | |
415 | ||
416 | static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline, | |
417 | unsigned offset) | |
418 | { | |
419 | return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8; | |
420 | } | |
421 | ||
422 | static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k) | |
423 | { | |
424 | return ((void *) k - (void *) t->data) / BSET_CACHELINE; | |
425 | } | |
426 | ||
427 | static unsigned bkey_to_cacheline_offset(struct bkey *k) | |
428 | { | |
429 | return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t); | |
430 | } | |
431 | ||
432 | static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j) | |
433 | { | |
434 | return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m); | |
435 | } | |
436 | ||
437 | static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j) | |
438 | { | |
439 | return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]); | |
440 | } | |
441 | ||
442 | /* | |
443 | * For the write set - the one we're currently inserting keys into - we don't | |
444 | * maintain a full search tree, we just keep a simple lookup table in t->prev. | |
445 | */ | |
446 | static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline) | |
447 | { | |
448 | return cacheline_to_bkey(t, cacheline, t->prev[cacheline]); | |
449 | } | |
450 | ||
451 | static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift) | |
452 | { | |
453 | #ifdef CONFIG_X86_64 | |
454 | asm("shrd %[shift],%[high],%[low]" | |
455 | : [low] "+Rm" (low) | |
456 | : [high] "R" (high), | |
457 | [shift] "ci" (shift) | |
458 | : "cc"); | |
459 | #else | |
460 | low >>= shift; | |
461 | low |= (high << 1) << (63U - shift); | |
462 | #endif | |
463 | return low; | |
464 | } | |
465 | ||
466 | static inline unsigned bfloat_mantissa(const struct bkey *k, | |
467 | struct bkey_float *f) | |
468 | { | |
469 | const uint64_t *p = &k->low - (f->exponent >> 6); | |
470 | return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK; | |
471 | } | |
472 | ||
473 | static void make_bfloat(struct bset_tree *t, unsigned j) | |
474 | { | |
475 | struct bkey_float *f = &t->tree[j]; | |
476 | struct bkey *m = tree_to_bkey(t, j); | |
477 | struct bkey *p = tree_to_prev_bkey(t, j); | |
478 | ||
479 | struct bkey *l = is_power_of_2(j) | |
480 | ? t->data->start | |
481 | : tree_to_prev_bkey(t, j >> ffs(j)); | |
482 | ||
483 | struct bkey *r = is_power_of_2(j + 1) | |
484 | ? node(t->data, t->data->keys - bkey_u64s(&t->end)) | |
485 | : tree_to_bkey(t, j >> (ffz(j) + 1)); | |
486 | ||
487 | BUG_ON(m < l || m > r); | |
488 | BUG_ON(bkey_next(p) != m); | |
489 | ||
490 | if (KEY_INODE(l) != KEY_INODE(r)) | |
491 | f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64; | |
492 | else | |
493 | f->exponent = fls64(r->low ^ l->low); | |
494 | ||
495 | f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0); | |
496 | ||
497 | /* | |
498 | * Setting f->exponent = 127 flags this node as failed, and causes the | |
499 | * lookup code to fall back to comparing against the original key. | |
500 | */ | |
501 | ||
502 | if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f)) | |
503 | f->mantissa = bfloat_mantissa(m, f) - 1; | |
504 | else | |
505 | f->exponent = 127; | |
506 | } | |
507 | ||
508 | static void bset_alloc_tree(struct btree *b, struct bset_tree *t) | |
509 | { | |
510 | if (t != b->sets) { | |
511 | unsigned j = roundup(t[-1].size, | |
512 | 64 / sizeof(struct bkey_float)); | |
513 | ||
514 | t->tree = t[-1].tree + j; | |
515 | t->prev = t[-1].prev + j; | |
516 | } | |
517 | ||
518 | while (t < b->sets + MAX_BSETS) | |
519 | t++->size = 0; | |
520 | } | |
521 | ||
522 | static void bset_build_unwritten_tree(struct btree *b) | |
523 | { | |
524 | struct bset_tree *t = b->sets + b->nsets; | |
525 | ||
526 | bset_alloc_tree(b, t); | |
527 | ||
528 | if (t->tree != b->sets->tree + bset_tree_space(b)) { | |
529 | t->prev[0] = bkey_to_cacheline_offset(t->data->start); | |
530 | t->size = 1; | |
531 | } | |
532 | } | |
533 | ||
534 | static void bset_build_written_tree(struct btree *b) | |
535 | { | |
536 | struct bset_tree *t = b->sets + b->nsets; | |
537 | struct bkey *k = t->data->start; | |
538 | unsigned j, cacheline = 1; | |
539 | ||
540 | bset_alloc_tree(b, t); | |
541 | ||
542 | t->size = min_t(unsigned, | |
543 | bkey_to_cacheline(t, end(t->data)), | |
544 | b->sets->tree + bset_tree_space(b) - t->tree); | |
545 | ||
546 | if (t->size < 2) { | |
547 | t->size = 0; | |
548 | return; | |
549 | } | |
550 | ||
551 | t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1; | |
552 | ||
553 | /* First we figure out where the first key in each cacheline is */ | |
554 | for (j = inorder_next(0, t->size); | |
555 | j; | |
556 | j = inorder_next(j, t->size)) { | |
557 | while (bkey_to_cacheline(t, k) != cacheline) | |
558 | k = bkey_next(k); | |
559 | ||
560 | t->prev[j] = bkey_u64s(k); | |
561 | k = bkey_next(k); | |
562 | cacheline++; | |
563 | t->tree[j].m = bkey_to_cacheline_offset(k); | |
564 | } | |
565 | ||
566 | while (bkey_next(k) != end(t->data)) | |
567 | k = bkey_next(k); | |
568 | ||
569 | t->end = *k; | |
570 | ||
571 | /* Then we build the tree */ | |
572 | for (j = inorder_next(0, t->size); | |
573 | j; | |
574 | j = inorder_next(j, t->size)) | |
575 | make_bfloat(t, j); | |
576 | } | |
577 | ||
578 | void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k) | |
579 | { | |
580 | struct bset_tree *t; | |
581 | unsigned inorder, j = 1; | |
582 | ||
583 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
584 | if (k < end(t->data)) | |
585 | goto found_set; | |
586 | ||
587 | BUG(); | |
588 | found_set: | |
589 | if (!t->size || !bset_written(b, t)) | |
590 | return; | |
591 | ||
592 | inorder = bkey_to_cacheline(t, k); | |
593 | ||
594 | if (k == t->data->start) | |
595 | goto fix_left; | |
596 | ||
597 | if (bkey_next(k) == end(t->data)) { | |
598 | t->end = *k; | |
599 | goto fix_right; | |
600 | } | |
601 | ||
602 | j = inorder_to_tree(inorder, t); | |
603 | ||
604 | if (j && | |
605 | j < t->size && | |
606 | k == tree_to_bkey(t, j)) | |
607 | fix_left: do { | |
608 | make_bfloat(t, j); | |
609 | j = j * 2; | |
610 | } while (j < t->size); | |
611 | ||
612 | j = inorder_to_tree(inorder + 1, t); | |
613 | ||
614 | if (j && | |
615 | j < t->size && | |
616 | k == tree_to_prev_bkey(t, j)) | |
617 | fix_right: do { | |
618 | make_bfloat(t, j); | |
619 | j = j * 2 + 1; | |
620 | } while (j < t->size); | |
621 | } | |
622 | ||
623 | void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k) | |
624 | { | |
625 | struct bset_tree *t = &b->sets[b->nsets]; | |
626 | unsigned shift = bkey_u64s(k); | |
627 | unsigned j = bkey_to_cacheline(t, k); | |
628 | ||
629 | /* We're getting called from btree_split() or btree_gc, just bail out */ | |
630 | if (!t->size) | |
631 | return; | |
632 | ||
633 | /* k is the key we just inserted; we need to find the entry in the | |
634 | * lookup table for the first key that is strictly greater than k: | |
635 | * it's either k's cacheline or the next one | |
636 | */ | |
637 | if (j < t->size && | |
638 | table_to_bkey(t, j) <= k) | |
639 | j++; | |
640 | ||
641 | /* Adjust all the lookup table entries, and find a new key for any that | |
642 | * have gotten too big | |
643 | */ | |
644 | for (; j < t->size; j++) { | |
645 | t->prev[j] += shift; | |
646 | ||
647 | if (t->prev[j] > 7) { | |
648 | k = table_to_bkey(t, j - 1); | |
649 | ||
650 | while (k < cacheline_to_bkey(t, j, 0)) | |
651 | k = bkey_next(k); | |
652 | ||
653 | t->prev[j] = bkey_to_cacheline_offset(k); | |
654 | } | |
655 | } | |
656 | ||
657 | if (t->size == b->sets->tree + bset_tree_space(b) - t->tree) | |
658 | return; | |
659 | ||
660 | /* Possibly add a new entry to the end of the lookup table */ | |
661 | ||
662 | for (k = table_to_bkey(t, t->size - 1); | |
663 | k != end(t->data); | |
664 | k = bkey_next(k)) | |
665 | if (t->size == bkey_to_cacheline(t, k)) { | |
666 | t->prev[t->size] = bkey_to_cacheline_offset(k); | |
667 | t->size++; | |
668 | } | |
669 | } | |
670 | ||
671 | void bch_bset_init_next(struct btree *b) | |
672 | { | |
673 | struct bset *i = write_block(b); | |
674 | ||
675 | if (i != b->sets[0].data) { | |
676 | b->sets[++b->nsets].data = i; | |
677 | i->seq = b->sets[0].data->seq; | |
678 | } else | |
679 | get_random_bytes(&i->seq, sizeof(uint64_t)); | |
680 | ||
681 | i->magic = bset_magic(b->c); | |
682 | i->version = 0; | |
683 | i->keys = 0; | |
684 | ||
685 | bset_build_unwritten_tree(b); | |
686 | } | |
687 | ||
688 | struct bset_search_iter { | |
689 | struct bkey *l, *r; | |
690 | }; | |
691 | ||
692 | static struct bset_search_iter bset_search_write_set(struct btree *b, | |
693 | struct bset_tree *t, | |
694 | const struct bkey *search) | |
695 | { | |
696 | unsigned li = 0, ri = t->size; | |
697 | ||
698 | BUG_ON(!b->nsets && | |
699 | t->size < bkey_to_cacheline(t, end(t->data))); | |
700 | ||
701 | while (li + 1 != ri) { | |
702 | unsigned m = (li + ri) >> 1; | |
703 | ||
704 | if (bkey_cmp(table_to_bkey(t, m), search) > 0) | |
705 | ri = m; | |
706 | else | |
707 | li = m; | |
708 | } | |
709 | ||
710 | return (struct bset_search_iter) { | |
711 | table_to_bkey(t, li), | |
712 | ri < t->size ? table_to_bkey(t, ri) : end(t->data) | |
713 | }; | |
714 | } | |
715 | ||
716 | static struct bset_search_iter bset_search_tree(struct btree *b, | |
717 | struct bset_tree *t, | |
718 | const struct bkey *search) | |
719 | { | |
720 | struct bkey *l, *r; | |
721 | struct bkey_float *f; | |
722 | unsigned inorder, j, n = 1; | |
723 | ||
724 | do { | |
725 | unsigned p = n << 4; | |
726 | p &= ((int) (p - t->size)) >> 31; | |
727 | ||
728 | prefetch(&t->tree[p]); | |
729 | ||
730 | j = n; | |
731 | f = &t->tree[j]; | |
732 | ||
733 | /* | |
734 | * n = (f->mantissa > bfloat_mantissa()) | |
735 | * ? j * 2 | |
736 | * : j * 2 + 1; | |
737 | * | |
738 | * We need to subtract 1 from f->mantissa for the sign bit trick | |
739 | * to work - that's done in make_bfloat() | |
740 | */ | |
741 | if (likely(f->exponent != 127)) | |
742 | n = j * 2 + (((unsigned) | |
743 | (f->mantissa - | |
744 | bfloat_mantissa(search, f))) >> 31); | |
745 | else | |
746 | n = (bkey_cmp(tree_to_bkey(t, j), search) > 0) | |
747 | ? j * 2 | |
748 | : j * 2 + 1; | |
749 | } while (n < t->size); | |
750 | ||
751 | inorder = to_inorder(j, t); | |
752 | ||
753 | /* | |
754 | * n would have been the node we recursed to - the low bit tells us if | |
755 | * we recursed left or recursed right. | |
756 | */ | |
757 | if (n & 1) { | |
758 | l = cacheline_to_bkey(t, inorder, f->m); | |
759 | ||
760 | if (++inorder != t->size) { | |
761 | f = &t->tree[inorder_next(j, t->size)]; | |
762 | r = cacheline_to_bkey(t, inorder, f->m); | |
763 | } else | |
764 | r = end(t->data); | |
765 | } else { | |
766 | r = cacheline_to_bkey(t, inorder, f->m); | |
767 | ||
768 | if (--inorder) { | |
769 | f = &t->tree[inorder_prev(j, t->size)]; | |
770 | l = cacheline_to_bkey(t, inorder, f->m); | |
771 | } else | |
772 | l = t->data->start; | |
773 | } | |
774 | ||
775 | return (struct bset_search_iter) {l, r}; | |
776 | } | |
777 | ||
778 | struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t, | |
779 | const struct bkey *search) | |
780 | { | |
781 | struct bset_search_iter i; | |
782 | ||
783 | /* | |
784 | * First, we search for a cacheline, then lastly we do a linear search | |
785 | * within that cacheline. | |
786 | * | |
787 | * To search for the cacheline, there's three different possibilities: | |
788 | * * The set is too small to have a search tree, so we just do a linear | |
789 | * search over the whole set. | |
790 | * * The set is the one we're currently inserting into; keeping a full | |
791 | * auxiliary search tree up to date would be too expensive, so we | |
792 | * use a much simpler lookup table to do a binary search - | |
793 | * bset_search_write_set(). | |
794 | * * Or we use the auxiliary search tree we constructed earlier - | |
795 | * bset_search_tree() | |
796 | */ | |
797 | ||
798 | if (unlikely(!t->size)) { | |
799 | i.l = t->data->start; | |
800 | i.r = end(t->data); | |
801 | } else if (bset_written(b, t)) { | |
802 | /* | |
803 | * Each node in the auxiliary search tree covers a certain range | |
804 | * of bits, and keys above and below the set it covers might | |
805 | * differ outside those bits - so we have to special case the | |
806 | * start and end - handle that here: | |
807 | */ | |
808 | ||
809 | if (unlikely(bkey_cmp(search, &t->end) >= 0)) | |
810 | return end(t->data); | |
811 | ||
812 | if (unlikely(bkey_cmp(search, t->data->start) < 0)) | |
813 | return t->data->start; | |
814 | ||
815 | i = bset_search_tree(b, t, search); | |
816 | } else | |
817 | i = bset_search_write_set(b, t, search); | |
818 | ||
819 | #ifdef CONFIG_BCACHE_EDEBUG | |
820 | BUG_ON(bset_written(b, t) && | |
821 | i.l != t->data->start && | |
822 | bkey_cmp(tree_to_prev_bkey(t, | |
823 | inorder_to_tree(bkey_to_cacheline(t, i.l), t)), | |
824 | search) > 0); | |
825 | ||
826 | BUG_ON(i.r != end(t->data) && | |
827 | bkey_cmp(i.r, search) <= 0); | |
828 | #endif | |
829 | ||
830 | while (likely(i.l != i.r) && | |
831 | bkey_cmp(i.l, search) <= 0) | |
832 | i.l = bkey_next(i.l); | |
833 | ||
834 | return i.l; | |
835 | } | |
836 | ||
837 | /* Btree iterator */ | |
838 | ||
839 | static inline bool btree_iter_cmp(struct btree_iter_set l, | |
840 | struct btree_iter_set r) | |
841 | { | |
842 | int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k)); | |
843 | ||
844 | return c ? c > 0 : l.k < r.k; | |
845 | } | |
846 | ||
847 | static inline bool btree_iter_end(struct btree_iter *iter) | |
848 | { | |
849 | return !iter->used; | |
850 | } | |
851 | ||
852 | void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k, | |
853 | struct bkey *end) | |
854 | { | |
855 | if (k != end) | |
856 | BUG_ON(!heap_add(iter, | |
857 | ((struct btree_iter_set) { k, end }), | |
858 | btree_iter_cmp)); | |
859 | } | |
860 | ||
861 | struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter, | |
862 | struct bkey *search, struct bset_tree *start) | |
863 | { | |
864 | struct bkey *ret = NULL; | |
865 | iter->size = ARRAY_SIZE(iter->data); | |
866 | iter->used = 0; | |
867 | ||
868 | for (; start <= &b->sets[b->nsets]; start++) { | |
869 | ret = bch_bset_search(b, start, search); | |
870 | bch_btree_iter_push(iter, ret, end(start->data)); | |
871 | } | |
872 | ||
873 | return ret; | |
874 | } | |
875 | ||
876 | struct bkey *bch_btree_iter_next(struct btree_iter *iter) | |
877 | { | |
878 | struct btree_iter_set unused; | |
879 | struct bkey *ret = NULL; | |
880 | ||
881 | if (!btree_iter_end(iter)) { | |
882 | ret = iter->data->k; | |
883 | iter->data->k = bkey_next(iter->data->k); | |
884 | ||
885 | if (iter->data->k > iter->data->end) { | |
886 | __WARN(); | |
887 | iter->data->k = iter->data->end; | |
888 | } | |
889 | ||
890 | if (iter->data->k == iter->data->end) | |
891 | heap_pop(iter, unused, btree_iter_cmp); | |
892 | else | |
893 | heap_sift(iter, 0, btree_iter_cmp); | |
894 | } | |
895 | ||
896 | return ret; | |
897 | } | |
898 | ||
899 | struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter, | |
900 | struct btree *b, ptr_filter_fn fn) | |
901 | { | |
902 | struct bkey *ret; | |
903 | ||
904 | do { | |
905 | ret = bch_btree_iter_next(iter); | |
906 | } while (ret && fn(b, ret)); | |
907 | ||
908 | return ret; | |
909 | } | |
910 | ||
911 | struct bkey *bch_next_recurse_key(struct btree *b, struct bkey *search) | |
912 | { | |
913 | struct btree_iter iter; | |
914 | ||
915 | bch_btree_iter_init(b, &iter, search); | |
916 | return bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); | |
917 | } | |
918 | ||
919 | /* Mergesort */ | |
920 | ||
921 | static void btree_sort_fixup(struct btree_iter *iter) | |
922 | { | |
923 | while (iter->used > 1) { | |
924 | struct btree_iter_set *top = iter->data, *i = top + 1; | |
925 | struct bkey *k; | |
926 | ||
927 | if (iter->used > 2 && | |
928 | btree_iter_cmp(i[0], i[1])) | |
929 | i++; | |
930 | ||
931 | for (k = i->k; | |
932 | k != i->end && bkey_cmp(top->k, &START_KEY(k)) > 0; | |
933 | k = bkey_next(k)) | |
934 | if (top->k > i->k) | |
935 | __bch_cut_front(top->k, k); | |
936 | else if (KEY_SIZE(k)) | |
937 | bch_cut_back(&START_KEY(k), top->k); | |
938 | ||
939 | if (top->k < i->k || k == i->k) | |
940 | break; | |
941 | ||
942 | heap_sift(iter, i - top, btree_iter_cmp); | |
943 | } | |
944 | } | |
945 | ||
946 | static void btree_mergesort(struct btree *b, struct bset *out, | |
947 | struct btree_iter *iter, | |
948 | bool fixup, bool remove_stale) | |
949 | { | |
950 | struct bkey *k, *last = NULL; | |
951 | bool (*bad)(struct btree *, const struct bkey *) = remove_stale | |
952 | ? bch_ptr_bad | |
953 | : bch_ptr_invalid; | |
954 | ||
955 | while (!btree_iter_end(iter)) { | |
956 | if (fixup && !b->level) | |
957 | btree_sort_fixup(iter); | |
958 | ||
959 | k = bch_btree_iter_next(iter); | |
960 | if (bad(b, k)) | |
961 | continue; | |
962 | ||
963 | if (!last) { | |
964 | last = out->start; | |
965 | bkey_copy(last, k); | |
966 | } else if (b->level || | |
967 | !bch_bkey_try_merge(b, last, k)) { | |
968 | last = bkey_next(last); | |
969 | bkey_copy(last, k); | |
970 | } | |
971 | } | |
972 | ||
973 | out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0; | |
974 | ||
975 | pr_debug("sorted %i keys", out->keys); | |
976 | bch_check_key_order(b, out); | |
977 | } | |
978 | ||
979 | static void __btree_sort(struct btree *b, struct btree_iter *iter, | |
980 | unsigned start, unsigned order, bool fixup) | |
981 | { | |
982 | uint64_t start_time; | |
983 | bool remove_stale = !b->written; | |
984 | struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO, | |
985 | order); | |
986 | if (!out) { | |
987 | mutex_lock(&b->c->sort_lock); | |
988 | out = b->c->sort; | |
989 | order = ilog2(bucket_pages(b->c)); | |
990 | } | |
991 | ||
992 | start_time = local_clock(); | |
993 | ||
994 | btree_mergesort(b, out, iter, fixup, remove_stale); | |
995 | b->nsets = start; | |
996 | ||
997 | if (!fixup && !start && b->written) | |
998 | bch_btree_verify(b, out); | |
999 | ||
1000 | if (!start && order == b->page_order) { | |
1001 | /* | |
1002 | * Our temporary buffer is the same size as the btree node's | |
1003 | * buffer, we can just swap buffers instead of doing a big | |
1004 | * memcpy() | |
1005 | */ | |
1006 | ||
1007 | out->magic = bset_magic(b->c); | |
1008 | out->seq = b->sets[0].data->seq; | |
1009 | out->version = b->sets[0].data->version; | |
1010 | swap(out, b->sets[0].data); | |
1011 | ||
1012 | if (b->c->sort == b->sets[0].data) | |
1013 | b->c->sort = out; | |
1014 | } else { | |
1015 | b->sets[start].data->keys = out->keys; | |
1016 | memcpy(b->sets[start].data->start, out->start, | |
1017 | (void *) end(out) - (void *) out->start); | |
1018 | } | |
1019 | ||
1020 | if (out == b->c->sort) | |
1021 | mutex_unlock(&b->c->sort_lock); | |
1022 | else | |
1023 | free_pages((unsigned long) out, order); | |
1024 | ||
1025 | if (b->written) | |
1026 | bset_build_written_tree(b); | |
1027 | ||
1028 | if (!start) { | |
1029 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1030 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1031 | spin_unlock(&b->c->sort_time_lock); |
1032 | } | |
1033 | } | |
1034 | ||
1035 | void bch_btree_sort_partial(struct btree *b, unsigned start) | |
1036 | { | |
1037 | size_t oldsize = 0, order = b->page_order, keys = 0; | |
1038 | struct btree_iter iter; | |
1039 | __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]); | |
1040 | ||
1041 | BUG_ON(b->sets[b->nsets].data == write_block(b) && | |
1042 | (b->sets[b->nsets].size || b->nsets)); | |
1043 | ||
1044 | if (b->written) | |
1045 | oldsize = bch_count_data(b); | |
1046 | ||
1047 | if (start) { | |
1048 | unsigned i; | |
1049 | ||
1050 | for (i = start; i <= b->nsets; i++) | |
1051 | keys += b->sets[i].data->keys; | |
1052 | ||
b1a67b0f KO |
1053 | order = roundup_pow_of_two(__set_bytes(b->sets->data, |
1054 | keys)) / PAGE_SIZE; | |
cafe5635 KO |
1055 | if (order) |
1056 | order = ilog2(order); | |
1057 | } | |
1058 | ||
1059 | __btree_sort(b, &iter, start, order, false); | |
1060 | ||
1061 | EBUG_ON(b->written && bch_count_data(b) != oldsize); | |
1062 | } | |
1063 | ||
1064 | void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter) | |
1065 | { | |
1066 | BUG_ON(!b->written); | |
1067 | __btree_sort(b, iter, 0, b->page_order, true); | |
1068 | } | |
1069 | ||
1070 | void bch_btree_sort_into(struct btree *b, struct btree *new) | |
1071 | { | |
1072 | uint64_t start_time = local_clock(); | |
1073 | ||
1074 | struct btree_iter iter; | |
1075 | bch_btree_iter_init(b, &iter, NULL); | |
1076 | ||
1077 | btree_mergesort(b, new->sets->data, &iter, false, true); | |
1078 | ||
1079 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1080 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1081 | spin_unlock(&b->c->sort_time_lock); |
1082 | ||
1083 | bkey_copy_key(&new->key, &b->key); | |
1084 | new->sets->size = 0; | |
1085 | } | |
1086 | ||
1087 | void bch_btree_sort_lazy(struct btree *b) | |
1088 | { | |
1089 | if (b->nsets) { | |
1090 | unsigned i, j, keys = 0, total; | |
1091 | ||
1092 | for (i = 0; i <= b->nsets; i++) | |
1093 | keys += b->sets[i].data->keys; | |
1094 | ||
1095 | total = keys; | |
1096 | ||
1097 | for (j = 0; j < b->nsets; j++) { | |
1098 | if (keys * 2 < total || | |
1099 | keys < 1000) { | |
1100 | bch_btree_sort_partial(b, j); | |
1101 | return; | |
1102 | } | |
1103 | ||
1104 | keys -= b->sets[j].data->keys; | |
1105 | } | |
1106 | ||
1107 | /* Must sort if b->nsets == 3 or we'll overflow */ | |
1108 | if (b->nsets >= (MAX_BSETS - 1) - b->level) { | |
1109 | bch_btree_sort(b); | |
1110 | return; | |
1111 | } | |
1112 | } | |
1113 | ||
1114 | bset_build_written_tree(b); | |
1115 | } | |
1116 | ||
1117 | /* Sysfs stuff */ | |
1118 | ||
1119 | struct bset_stats { | |
1120 | size_t nodes; | |
1121 | size_t sets_written, sets_unwritten; | |
1122 | size_t bytes_written, bytes_unwritten; | |
1123 | size_t floats, failed; | |
1124 | }; | |
1125 | ||
1126 | static int bch_btree_bset_stats(struct btree *b, struct btree_op *op, | |
1127 | struct bset_stats *stats) | |
1128 | { | |
1129 | struct bkey *k; | |
1130 | unsigned i; | |
1131 | ||
1132 | stats->nodes++; | |
1133 | ||
1134 | for (i = 0; i <= b->nsets; i++) { | |
1135 | struct bset_tree *t = &b->sets[i]; | |
1136 | size_t bytes = t->data->keys * sizeof(uint64_t); | |
1137 | size_t j; | |
1138 | ||
1139 | if (bset_written(b, t)) { | |
1140 | stats->sets_written++; | |
1141 | stats->bytes_written += bytes; | |
1142 | ||
1143 | stats->floats += t->size - 1; | |
1144 | ||
1145 | for (j = 1; j < t->size; j++) | |
1146 | if (t->tree[j].exponent == 127) | |
1147 | stats->failed++; | |
1148 | } else { | |
1149 | stats->sets_unwritten++; | |
1150 | stats->bytes_unwritten += bytes; | |
1151 | } | |
1152 | } | |
1153 | ||
1154 | if (b->level) { | |
1155 | struct btree_iter iter; | |
1156 | ||
1157 | for_each_key_filter(b, k, &iter, bch_ptr_bad) { | |
1158 | int ret = btree(bset_stats, k, b, op, stats); | |
1159 | if (ret) | |
1160 | return ret; | |
1161 | } | |
1162 | } | |
1163 | ||
1164 | return 0; | |
1165 | } | |
1166 | ||
1167 | int bch_bset_print_stats(struct cache_set *c, char *buf) | |
1168 | { | |
1169 | struct btree_op op; | |
1170 | struct bset_stats t; | |
1171 | int ret; | |
1172 | ||
1173 | bch_btree_op_init_stack(&op); | |
1174 | memset(&t, 0, sizeof(struct bset_stats)); | |
1175 | ||
1176 | ret = btree_root(bset_stats, c, &op, &t); | |
1177 | if (ret) | |
1178 | return ret; | |
1179 | ||
1180 | return snprintf(buf, PAGE_SIZE, | |
1181 | "btree nodes: %zu\n" | |
1182 | "written sets: %zu\n" | |
1183 | "unwritten sets: %zu\n" | |
1184 | "written key bytes: %zu\n" | |
1185 | "unwritten key bytes: %zu\n" | |
1186 | "floats: %zu\n" | |
1187 | "failed: %zu\n", | |
1188 | t.nodes, | |
1189 | t.sets_written, t.sets_unwritten, | |
1190 | t.bytes_written, t.bytes_unwritten, | |
1191 | t.floats, t.failed); | |
1192 | } |