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; | |
85b1492e | 81 | char buf[80]; |
cafe5635 KO |
82 | |
83 | if (level && (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))) | |
84 | goto bad; | |
85 | ||
86 | if (!level && KEY_SIZE(k) > KEY_OFFSET(k)) | |
87 | goto bad; | |
88 | ||
89 | if (!KEY_SIZE(k)) | |
90 | return true; | |
91 | ||
92 | for (i = 0; i < KEY_PTRS(k); i++) | |
93 | if (ptr_available(c, k, i)) { | |
94 | struct cache *ca = PTR_CACHE(c, k, i); | |
95 | size_t bucket = PTR_BUCKET_NR(c, k, i); | |
96 | size_t r = bucket_remainder(c, PTR_OFFSET(k, i)); | |
97 | ||
98 | if (KEY_SIZE(k) + r > c->sb.bucket_size || | |
99 | bucket < ca->sb.first_bucket || | |
100 | bucket >= ca->sb.nbuckets) | |
101 | goto bad; | |
102 | } | |
103 | ||
104 | return false; | |
105 | bad: | |
85b1492e KO |
106 | bch_bkey_to_text(buf, sizeof(buf), k); |
107 | cache_bug(c, "spotted bad key %s: %s", buf, bch_ptr_status(c, k)); | |
cafe5635 KO |
108 | return true; |
109 | } | |
110 | ||
111 | bool bch_ptr_bad(struct btree *b, const struct bkey *k) | |
112 | { | |
113 | struct bucket *g; | |
114 | unsigned i, stale; | |
115 | ||
116 | if (!bkey_cmp(k, &ZERO_KEY) || | |
117 | !KEY_PTRS(k) || | |
118 | bch_ptr_invalid(b, k)) | |
119 | return true; | |
120 | ||
121 | if (KEY_PTRS(k) && PTR_DEV(k, 0) == PTR_CHECK_DEV) | |
122 | return true; | |
123 | ||
124 | for (i = 0; i < KEY_PTRS(k); i++) | |
125 | if (ptr_available(b->c, k, i)) { | |
126 | g = PTR_BUCKET(b->c, k, i); | |
127 | stale = ptr_stale(b->c, k, i); | |
128 | ||
129 | btree_bug_on(stale > 96, b, | |
130 | "key too stale: %i, need_gc %u", | |
131 | stale, b->c->need_gc); | |
132 | ||
133 | btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k), | |
134 | b, "stale dirty pointer"); | |
135 | ||
136 | if (stale) | |
137 | return true; | |
138 | ||
139 | #ifdef CONFIG_BCACHE_EDEBUG | |
140 | if (!mutex_trylock(&b->c->bucket_lock)) | |
141 | continue; | |
142 | ||
143 | if (b->level) { | |
144 | if (KEY_DIRTY(k) || | |
145 | g->prio != BTREE_PRIO || | |
146 | (b->c->gc_mark_valid && | |
147 | GC_MARK(g) != GC_MARK_METADATA)) | |
148 | goto bug; | |
149 | ||
150 | } else { | |
151 | if (g->prio == BTREE_PRIO) | |
152 | goto bug; | |
153 | ||
154 | if (KEY_DIRTY(k) && | |
155 | b->c->gc_mark_valid && | |
156 | GC_MARK(g) != GC_MARK_DIRTY) | |
157 | goto bug; | |
158 | } | |
159 | mutex_unlock(&b->c->bucket_lock); | |
160 | #endif | |
161 | } | |
162 | ||
163 | return false; | |
164 | #ifdef CONFIG_BCACHE_EDEBUG | |
165 | bug: | |
166 | mutex_unlock(&b->c->bucket_lock); | |
85b1492e KO |
167 | |
168 | { | |
169 | char buf[80]; | |
170 | ||
171 | bch_bkey_to_text(buf, sizeof(buf), k); | |
172 | btree_bug(b, | |
b1a67b0f | 173 | "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i", |
85b1492e KO |
174 | buf, PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin), |
175 | g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen); | |
176 | } | |
cafe5635 KO |
177 | return true; |
178 | #endif | |
179 | } | |
180 | ||
181 | /* Key/pointer manipulation */ | |
182 | ||
183 | void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src, | |
184 | unsigned i) | |
185 | { | |
186 | BUG_ON(i > KEY_PTRS(src)); | |
187 | ||
188 | /* Only copy the header, key, and one pointer. */ | |
189 | memcpy(dest, src, 2 * sizeof(uint64_t)); | |
190 | dest->ptr[0] = src->ptr[i]; | |
191 | SET_KEY_PTRS(dest, 1); | |
192 | /* We didn't copy the checksum so clear that bit. */ | |
193 | SET_KEY_CSUM(dest, 0); | |
194 | } | |
195 | ||
196 | bool __bch_cut_front(const struct bkey *where, struct bkey *k) | |
197 | { | |
198 | unsigned i, len = 0; | |
199 | ||
200 | if (bkey_cmp(where, &START_KEY(k)) <= 0) | |
201 | return false; | |
202 | ||
203 | if (bkey_cmp(where, k) < 0) | |
204 | len = KEY_OFFSET(k) - KEY_OFFSET(where); | |
205 | else | |
206 | bkey_copy_key(k, where); | |
207 | ||
208 | for (i = 0; i < KEY_PTRS(k); i++) | |
209 | SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len); | |
210 | ||
211 | BUG_ON(len > KEY_SIZE(k)); | |
212 | SET_KEY_SIZE(k, len); | |
213 | return true; | |
214 | } | |
215 | ||
216 | bool __bch_cut_back(const struct bkey *where, struct bkey *k) | |
217 | { | |
218 | unsigned len = 0; | |
219 | ||
220 | if (bkey_cmp(where, k) >= 0) | |
221 | return false; | |
222 | ||
223 | BUG_ON(KEY_INODE(where) != KEY_INODE(k)); | |
224 | ||
225 | if (bkey_cmp(where, &START_KEY(k)) > 0) | |
226 | len = KEY_OFFSET(where) - KEY_START(k); | |
227 | ||
228 | bkey_copy_key(k, where); | |
229 | ||
230 | BUG_ON(len > KEY_SIZE(k)); | |
231 | SET_KEY_SIZE(k, len); | |
232 | return true; | |
233 | } | |
234 | ||
235 | static uint64_t merge_chksums(struct bkey *l, struct bkey *r) | |
236 | { | |
237 | return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) & | |
238 | ~((uint64_t)1 << 63); | |
239 | } | |
240 | ||
241 | /* Tries to merge l and r: l should be lower than r | |
242 | * Returns true if we were able to merge. If we did merge, l will be the merged | |
243 | * key, r will be untouched. | |
244 | */ | |
245 | bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r) | |
246 | { | |
247 | unsigned i; | |
248 | ||
249 | if (key_merging_disabled(b->c)) | |
250 | return false; | |
251 | ||
252 | if (KEY_PTRS(l) != KEY_PTRS(r) || | |
253 | KEY_DIRTY(l) != KEY_DIRTY(r) || | |
254 | bkey_cmp(l, &START_KEY(r))) | |
255 | return false; | |
256 | ||
257 | for (i = 0; i < KEY_PTRS(l); i++) | |
258 | if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] || | |
259 | PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i)) | |
260 | return false; | |
261 | ||
262 | /* Keys with no pointers aren't restricted to one bucket and could | |
263 | * overflow KEY_SIZE | |
264 | */ | |
265 | if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) { | |
266 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l)); | |
267 | SET_KEY_SIZE(l, USHRT_MAX); | |
268 | ||
269 | bch_cut_front(l, r); | |
270 | return false; | |
271 | } | |
272 | ||
273 | if (KEY_CSUM(l)) { | |
274 | if (KEY_CSUM(r)) | |
275 | l->ptr[KEY_PTRS(l)] = merge_chksums(l, r); | |
276 | else | |
277 | SET_KEY_CSUM(l, 0); | |
278 | } | |
279 | ||
280 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r)); | |
281 | SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r)); | |
282 | ||
283 | return true; | |
284 | } | |
285 | ||
286 | /* Binary tree stuff for auxiliary search trees */ | |
287 | ||
288 | static unsigned inorder_next(unsigned j, unsigned size) | |
289 | { | |
290 | if (j * 2 + 1 < size) { | |
291 | j = j * 2 + 1; | |
292 | ||
293 | while (j * 2 < size) | |
294 | j *= 2; | |
295 | } else | |
296 | j >>= ffz(j) + 1; | |
297 | ||
298 | return j; | |
299 | } | |
300 | ||
301 | static unsigned inorder_prev(unsigned j, unsigned size) | |
302 | { | |
303 | if (j * 2 < size) { | |
304 | j = j * 2; | |
305 | ||
306 | while (j * 2 + 1 < size) | |
307 | j = j * 2 + 1; | |
308 | } else | |
309 | j >>= ffs(j); | |
310 | ||
311 | return j; | |
312 | } | |
313 | ||
314 | /* I have no idea why this code works... and I'm the one who wrote it | |
315 | * | |
316 | * However, I do know what it does: | |
317 | * Given a binary tree constructed in an array (i.e. how you normally implement | |
318 | * a heap), it converts a node in the tree - referenced by array index - to the | |
319 | * index it would have if you did an inorder traversal. | |
320 | * | |
321 | * Also tested for every j, size up to size somewhere around 6 million. | |
322 | * | |
323 | * The binary tree starts at array index 1, not 0 | |
324 | * extra is a function of size: | |
325 | * extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
326 | */ | |
327 | static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra) | |
328 | { | |
329 | unsigned b = fls(j); | |
330 | unsigned shift = fls(size - 1) - b; | |
331 | ||
332 | j ^= 1U << (b - 1); | |
333 | j <<= 1; | |
334 | j |= 1; | |
335 | j <<= shift; | |
336 | ||
337 | if (j > extra) | |
338 | j -= (j - extra) >> 1; | |
339 | ||
340 | return j; | |
341 | } | |
342 | ||
343 | static unsigned to_inorder(unsigned j, struct bset_tree *t) | |
344 | { | |
345 | return __to_inorder(j, t->size, t->extra); | |
346 | } | |
347 | ||
348 | static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra) | |
349 | { | |
350 | unsigned shift; | |
351 | ||
352 | if (j > extra) | |
353 | j += j - extra; | |
354 | ||
355 | shift = ffs(j); | |
356 | ||
357 | j >>= shift; | |
358 | j |= roundup_pow_of_two(size) >> shift; | |
359 | ||
360 | return j; | |
361 | } | |
362 | ||
363 | static unsigned inorder_to_tree(unsigned j, struct bset_tree *t) | |
364 | { | |
365 | return __inorder_to_tree(j, t->size, t->extra); | |
366 | } | |
367 | ||
368 | #if 0 | |
369 | void inorder_test(void) | |
370 | { | |
371 | unsigned long done = 0; | |
372 | ktime_t start = ktime_get(); | |
373 | ||
374 | for (unsigned size = 2; | |
375 | size < 65536000; | |
376 | size++) { | |
377 | unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
378 | unsigned i = 1, j = rounddown_pow_of_two(size - 1); | |
379 | ||
380 | if (!(size % 4096)) | |
381 | printk(KERN_NOTICE "loop %u, %llu per us\n", size, | |
382 | done / ktime_us_delta(ktime_get(), start)); | |
383 | ||
384 | while (1) { | |
385 | if (__inorder_to_tree(i, size, extra) != j) | |
386 | panic("size %10u j %10u i %10u", size, j, i); | |
387 | ||
388 | if (__to_inorder(j, size, extra) != i) | |
389 | panic("size %10u j %10u i %10u", size, j, i); | |
390 | ||
391 | if (j == rounddown_pow_of_two(size) - 1) | |
392 | break; | |
393 | ||
394 | BUG_ON(inorder_prev(inorder_next(j, size), size) != j); | |
395 | ||
396 | j = inorder_next(j, size); | |
397 | i++; | |
398 | } | |
399 | ||
400 | done += size - 1; | |
401 | } | |
402 | } | |
403 | #endif | |
404 | ||
405 | /* | |
406 | * Cacheline/offset <-> bkey pointer arithmatic: | |
407 | * | |
408 | * t->tree is a binary search tree in an array; each node corresponds to a key | |
409 | * in one cacheline in t->set (BSET_CACHELINE bytes). | |
410 | * | |
411 | * This means we don't have to store the full index of the key that a node in | |
412 | * the binary tree points to; to_inorder() gives us the cacheline, and then | |
413 | * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes. | |
414 | * | |
415 | * cacheline_to_bkey() and friends abstract out all the pointer arithmatic to | |
416 | * make this work. | |
417 | * | |
418 | * To construct the bfloat for an arbitrary key we need to know what the key | |
419 | * immediately preceding it is: we have to check if the two keys differ in the | |
420 | * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size | |
421 | * of the previous key so we can walk backwards to it from t->tree[j]'s key. | |
422 | */ | |
423 | ||
424 | static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline, | |
425 | unsigned offset) | |
426 | { | |
427 | return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8; | |
428 | } | |
429 | ||
430 | static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k) | |
431 | { | |
432 | return ((void *) k - (void *) t->data) / BSET_CACHELINE; | |
433 | } | |
434 | ||
435 | static unsigned bkey_to_cacheline_offset(struct bkey *k) | |
436 | { | |
437 | return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t); | |
438 | } | |
439 | ||
440 | static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j) | |
441 | { | |
442 | return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m); | |
443 | } | |
444 | ||
445 | static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j) | |
446 | { | |
447 | return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]); | |
448 | } | |
449 | ||
450 | /* | |
451 | * For the write set - the one we're currently inserting keys into - we don't | |
452 | * maintain a full search tree, we just keep a simple lookup table in t->prev. | |
453 | */ | |
454 | static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline) | |
455 | { | |
456 | return cacheline_to_bkey(t, cacheline, t->prev[cacheline]); | |
457 | } | |
458 | ||
459 | static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift) | |
460 | { | |
461 | #ifdef CONFIG_X86_64 | |
462 | asm("shrd %[shift],%[high],%[low]" | |
463 | : [low] "+Rm" (low) | |
464 | : [high] "R" (high), | |
465 | [shift] "ci" (shift) | |
466 | : "cc"); | |
467 | #else | |
468 | low >>= shift; | |
469 | low |= (high << 1) << (63U - shift); | |
470 | #endif | |
471 | return low; | |
472 | } | |
473 | ||
474 | static inline unsigned bfloat_mantissa(const struct bkey *k, | |
475 | struct bkey_float *f) | |
476 | { | |
477 | const uint64_t *p = &k->low - (f->exponent >> 6); | |
478 | return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK; | |
479 | } | |
480 | ||
481 | static void make_bfloat(struct bset_tree *t, unsigned j) | |
482 | { | |
483 | struct bkey_float *f = &t->tree[j]; | |
484 | struct bkey *m = tree_to_bkey(t, j); | |
485 | struct bkey *p = tree_to_prev_bkey(t, j); | |
486 | ||
487 | struct bkey *l = is_power_of_2(j) | |
488 | ? t->data->start | |
489 | : tree_to_prev_bkey(t, j >> ffs(j)); | |
490 | ||
491 | struct bkey *r = is_power_of_2(j + 1) | |
492 | ? node(t->data, t->data->keys - bkey_u64s(&t->end)) | |
493 | : tree_to_bkey(t, j >> (ffz(j) + 1)); | |
494 | ||
495 | BUG_ON(m < l || m > r); | |
496 | BUG_ON(bkey_next(p) != m); | |
497 | ||
498 | if (KEY_INODE(l) != KEY_INODE(r)) | |
499 | f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64; | |
500 | else | |
501 | f->exponent = fls64(r->low ^ l->low); | |
502 | ||
503 | f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0); | |
504 | ||
505 | /* | |
506 | * Setting f->exponent = 127 flags this node as failed, and causes the | |
507 | * lookup code to fall back to comparing against the original key. | |
508 | */ | |
509 | ||
510 | if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f)) | |
511 | f->mantissa = bfloat_mantissa(m, f) - 1; | |
512 | else | |
513 | f->exponent = 127; | |
514 | } | |
515 | ||
516 | static void bset_alloc_tree(struct btree *b, struct bset_tree *t) | |
517 | { | |
518 | if (t != b->sets) { | |
519 | unsigned j = roundup(t[-1].size, | |
520 | 64 / sizeof(struct bkey_float)); | |
521 | ||
522 | t->tree = t[-1].tree + j; | |
523 | t->prev = t[-1].prev + j; | |
524 | } | |
525 | ||
526 | while (t < b->sets + MAX_BSETS) | |
527 | t++->size = 0; | |
528 | } | |
529 | ||
530 | static void bset_build_unwritten_tree(struct btree *b) | |
531 | { | |
532 | struct bset_tree *t = b->sets + b->nsets; | |
533 | ||
534 | bset_alloc_tree(b, t); | |
535 | ||
536 | if (t->tree != b->sets->tree + bset_tree_space(b)) { | |
537 | t->prev[0] = bkey_to_cacheline_offset(t->data->start); | |
538 | t->size = 1; | |
539 | } | |
540 | } | |
541 | ||
542 | static void bset_build_written_tree(struct btree *b) | |
543 | { | |
544 | struct bset_tree *t = b->sets + b->nsets; | |
545 | struct bkey *k = t->data->start; | |
546 | unsigned j, cacheline = 1; | |
547 | ||
548 | bset_alloc_tree(b, t); | |
549 | ||
550 | t->size = min_t(unsigned, | |
551 | bkey_to_cacheline(t, end(t->data)), | |
552 | b->sets->tree + bset_tree_space(b) - t->tree); | |
553 | ||
554 | if (t->size < 2) { | |
555 | t->size = 0; | |
556 | return; | |
557 | } | |
558 | ||
559 | t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1; | |
560 | ||
561 | /* First we figure out where the first key in each cacheline is */ | |
562 | for (j = inorder_next(0, t->size); | |
563 | j; | |
564 | j = inorder_next(j, t->size)) { | |
565 | while (bkey_to_cacheline(t, k) != cacheline) | |
566 | k = bkey_next(k); | |
567 | ||
568 | t->prev[j] = bkey_u64s(k); | |
569 | k = bkey_next(k); | |
570 | cacheline++; | |
571 | t->tree[j].m = bkey_to_cacheline_offset(k); | |
572 | } | |
573 | ||
574 | while (bkey_next(k) != end(t->data)) | |
575 | k = bkey_next(k); | |
576 | ||
577 | t->end = *k; | |
578 | ||
579 | /* Then we build the tree */ | |
580 | for (j = inorder_next(0, t->size); | |
581 | j; | |
582 | j = inorder_next(j, t->size)) | |
583 | make_bfloat(t, j); | |
584 | } | |
585 | ||
586 | void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k) | |
587 | { | |
588 | struct bset_tree *t; | |
589 | unsigned inorder, j = 1; | |
590 | ||
591 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
592 | if (k < end(t->data)) | |
593 | goto found_set; | |
594 | ||
595 | BUG(); | |
596 | found_set: | |
597 | if (!t->size || !bset_written(b, t)) | |
598 | return; | |
599 | ||
600 | inorder = bkey_to_cacheline(t, k); | |
601 | ||
602 | if (k == t->data->start) | |
603 | goto fix_left; | |
604 | ||
605 | if (bkey_next(k) == end(t->data)) { | |
606 | t->end = *k; | |
607 | goto fix_right; | |
608 | } | |
609 | ||
610 | j = inorder_to_tree(inorder, t); | |
611 | ||
612 | if (j && | |
613 | j < t->size && | |
614 | k == tree_to_bkey(t, j)) | |
615 | fix_left: do { | |
616 | make_bfloat(t, j); | |
617 | j = j * 2; | |
618 | } while (j < t->size); | |
619 | ||
620 | j = inorder_to_tree(inorder + 1, t); | |
621 | ||
622 | if (j && | |
623 | j < t->size && | |
624 | k == tree_to_prev_bkey(t, j)) | |
625 | fix_right: do { | |
626 | make_bfloat(t, j); | |
627 | j = j * 2 + 1; | |
628 | } while (j < t->size); | |
629 | } | |
630 | ||
631 | void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k) | |
632 | { | |
633 | struct bset_tree *t = &b->sets[b->nsets]; | |
634 | unsigned shift = bkey_u64s(k); | |
635 | unsigned j = bkey_to_cacheline(t, k); | |
636 | ||
637 | /* We're getting called from btree_split() or btree_gc, just bail out */ | |
638 | if (!t->size) | |
639 | return; | |
640 | ||
641 | /* k is the key we just inserted; we need to find the entry in the | |
642 | * lookup table for the first key that is strictly greater than k: | |
643 | * it's either k's cacheline or the next one | |
644 | */ | |
645 | if (j < t->size && | |
646 | table_to_bkey(t, j) <= k) | |
647 | j++; | |
648 | ||
649 | /* Adjust all the lookup table entries, and find a new key for any that | |
650 | * have gotten too big | |
651 | */ | |
652 | for (; j < t->size; j++) { | |
653 | t->prev[j] += shift; | |
654 | ||
655 | if (t->prev[j] > 7) { | |
656 | k = table_to_bkey(t, j - 1); | |
657 | ||
658 | while (k < cacheline_to_bkey(t, j, 0)) | |
659 | k = bkey_next(k); | |
660 | ||
661 | t->prev[j] = bkey_to_cacheline_offset(k); | |
662 | } | |
663 | } | |
664 | ||
665 | if (t->size == b->sets->tree + bset_tree_space(b) - t->tree) | |
666 | return; | |
667 | ||
668 | /* Possibly add a new entry to the end of the lookup table */ | |
669 | ||
670 | for (k = table_to_bkey(t, t->size - 1); | |
671 | k != end(t->data); | |
672 | k = bkey_next(k)) | |
673 | if (t->size == bkey_to_cacheline(t, k)) { | |
674 | t->prev[t->size] = bkey_to_cacheline_offset(k); | |
675 | t->size++; | |
676 | } | |
677 | } | |
678 | ||
679 | void bch_bset_init_next(struct btree *b) | |
680 | { | |
681 | struct bset *i = write_block(b); | |
682 | ||
683 | if (i != b->sets[0].data) { | |
684 | b->sets[++b->nsets].data = i; | |
685 | i->seq = b->sets[0].data->seq; | |
686 | } else | |
687 | get_random_bytes(&i->seq, sizeof(uint64_t)); | |
688 | ||
689 | i->magic = bset_magic(b->c); | |
690 | i->version = 0; | |
691 | i->keys = 0; | |
692 | ||
693 | bset_build_unwritten_tree(b); | |
694 | } | |
695 | ||
696 | struct bset_search_iter { | |
697 | struct bkey *l, *r; | |
698 | }; | |
699 | ||
700 | static struct bset_search_iter bset_search_write_set(struct btree *b, | |
701 | struct bset_tree *t, | |
702 | const struct bkey *search) | |
703 | { | |
704 | unsigned li = 0, ri = t->size; | |
705 | ||
706 | BUG_ON(!b->nsets && | |
707 | t->size < bkey_to_cacheline(t, end(t->data))); | |
708 | ||
709 | while (li + 1 != ri) { | |
710 | unsigned m = (li + ri) >> 1; | |
711 | ||
712 | if (bkey_cmp(table_to_bkey(t, m), search) > 0) | |
713 | ri = m; | |
714 | else | |
715 | li = m; | |
716 | } | |
717 | ||
718 | return (struct bset_search_iter) { | |
719 | table_to_bkey(t, li), | |
720 | ri < t->size ? table_to_bkey(t, ri) : end(t->data) | |
721 | }; | |
722 | } | |
723 | ||
724 | static struct bset_search_iter bset_search_tree(struct btree *b, | |
725 | struct bset_tree *t, | |
726 | const struct bkey *search) | |
727 | { | |
728 | struct bkey *l, *r; | |
729 | struct bkey_float *f; | |
730 | unsigned inorder, j, n = 1; | |
731 | ||
732 | do { | |
733 | unsigned p = n << 4; | |
734 | p &= ((int) (p - t->size)) >> 31; | |
735 | ||
736 | prefetch(&t->tree[p]); | |
737 | ||
738 | j = n; | |
739 | f = &t->tree[j]; | |
740 | ||
741 | /* | |
742 | * n = (f->mantissa > bfloat_mantissa()) | |
743 | * ? j * 2 | |
744 | * : j * 2 + 1; | |
745 | * | |
746 | * We need to subtract 1 from f->mantissa for the sign bit trick | |
747 | * to work - that's done in make_bfloat() | |
748 | */ | |
749 | if (likely(f->exponent != 127)) | |
750 | n = j * 2 + (((unsigned) | |
751 | (f->mantissa - | |
752 | bfloat_mantissa(search, f))) >> 31); | |
753 | else | |
754 | n = (bkey_cmp(tree_to_bkey(t, j), search) > 0) | |
755 | ? j * 2 | |
756 | : j * 2 + 1; | |
757 | } while (n < t->size); | |
758 | ||
759 | inorder = to_inorder(j, t); | |
760 | ||
761 | /* | |
762 | * n would have been the node we recursed to - the low bit tells us if | |
763 | * we recursed left or recursed right. | |
764 | */ | |
765 | if (n & 1) { | |
766 | l = cacheline_to_bkey(t, inorder, f->m); | |
767 | ||
768 | if (++inorder != t->size) { | |
769 | f = &t->tree[inorder_next(j, t->size)]; | |
770 | r = cacheline_to_bkey(t, inorder, f->m); | |
771 | } else | |
772 | r = end(t->data); | |
773 | } else { | |
774 | r = cacheline_to_bkey(t, inorder, f->m); | |
775 | ||
776 | if (--inorder) { | |
777 | f = &t->tree[inorder_prev(j, t->size)]; | |
778 | l = cacheline_to_bkey(t, inorder, f->m); | |
779 | } else | |
780 | l = t->data->start; | |
781 | } | |
782 | ||
783 | return (struct bset_search_iter) {l, r}; | |
784 | } | |
785 | ||
786 | struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t, | |
787 | const struct bkey *search) | |
788 | { | |
789 | struct bset_search_iter i; | |
790 | ||
791 | /* | |
792 | * First, we search for a cacheline, then lastly we do a linear search | |
793 | * within that cacheline. | |
794 | * | |
795 | * To search for the cacheline, there's three different possibilities: | |
796 | * * The set is too small to have a search tree, so we just do a linear | |
797 | * search over the whole set. | |
798 | * * The set is the one we're currently inserting into; keeping a full | |
799 | * auxiliary search tree up to date would be too expensive, so we | |
800 | * use a much simpler lookup table to do a binary search - | |
801 | * bset_search_write_set(). | |
802 | * * Or we use the auxiliary search tree we constructed earlier - | |
803 | * bset_search_tree() | |
804 | */ | |
805 | ||
806 | if (unlikely(!t->size)) { | |
807 | i.l = t->data->start; | |
808 | i.r = end(t->data); | |
809 | } else if (bset_written(b, t)) { | |
810 | /* | |
811 | * Each node in the auxiliary search tree covers a certain range | |
812 | * of bits, and keys above and below the set it covers might | |
813 | * differ outside those bits - so we have to special case the | |
814 | * start and end - handle that here: | |
815 | */ | |
816 | ||
817 | if (unlikely(bkey_cmp(search, &t->end) >= 0)) | |
818 | return end(t->data); | |
819 | ||
820 | if (unlikely(bkey_cmp(search, t->data->start) < 0)) | |
821 | return t->data->start; | |
822 | ||
823 | i = bset_search_tree(b, t, search); | |
824 | } else | |
825 | i = bset_search_write_set(b, t, search); | |
826 | ||
827 | #ifdef CONFIG_BCACHE_EDEBUG | |
828 | BUG_ON(bset_written(b, t) && | |
829 | i.l != t->data->start && | |
830 | bkey_cmp(tree_to_prev_bkey(t, | |
831 | inorder_to_tree(bkey_to_cacheline(t, i.l), t)), | |
832 | search) > 0); | |
833 | ||
834 | BUG_ON(i.r != end(t->data) && | |
835 | bkey_cmp(i.r, search) <= 0); | |
836 | #endif | |
837 | ||
838 | while (likely(i.l != i.r) && | |
839 | bkey_cmp(i.l, search) <= 0) | |
840 | i.l = bkey_next(i.l); | |
841 | ||
842 | return i.l; | |
843 | } | |
844 | ||
845 | /* Btree iterator */ | |
846 | ||
847 | static inline bool btree_iter_cmp(struct btree_iter_set l, | |
848 | struct btree_iter_set r) | |
849 | { | |
850 | int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k)); | |
851 | ||
852 | return c ? c > 0 : l.k < r.k; | |
853 | } | |
854 | ||
855 | static inline bool btree_iter_end(struct btree_iter *iter) | |
856 | { | |
857 | return !iter->used; | |
858 | } | |
859 | ||
860 | void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k, | |
861 | struct bkey *end) | |
862 | { | |
863 | if (k != end) | |
864 | BUG_ON(!heap_add(iter, | |
865 | ((struct btree_iter_set) { k, end }), | |
866 | btree_iter_cmp)); | |
867 | } | |
868 | ||
869 | struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter, | |
870 | struct bkey *search, struct bset_tree *start) | |
871 | { | |
872 | struct bkey *ret = NULL; | |
873 | iter->size = ARRAY_SIZE(iter->data); | |
874 | iter->used = 0; | |
875 | ||
876 | for (; start <= &b->sets[b->nsets]; start++) { | |
877 | ret = bch_bset_search(b, start, search); | |
878 | bch_btree_iter_push(iter, ret, end(start->data)); | |
879 | } | |
880 | ||
881 | return ret; | |
882 | } | |
883 | ||
884 | struct bkey *bch_btree_iter_next(struct btree_iter *iter) | |
885 | { | |
886 | struct btree_iter_set unused; | |
887 | struct bkey *ret = NULL; | |
888 | ||
889 | if (!btree_iter_end(iter)) { | |
890 | ret = iter->data->k; | |
891 | iter->data->k = bkey_next(iter->data->k); | |
892 | ||
893 | if (iter->data->k > iter->data->end) { | |
cc0f4eaa | 894 | WARN_ONCE(1, "bset was corrupt!\n"); |
cafe5635 KO |
895 | iter->data->k = iter->data->end; |
896 | } | |
897 | ||
898 | if (iter->data->k == iter->data->end) | |
899 | heap_pop(iter, unused, btree_iter_cmp); | |
900 | else | |
901 | heap_sift(iter, 0, btree_iter_cmp); | |
902 | } | |
903 | ||
904 | return ret; | |
905 | } | |
906 | ||
907 | struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter, | |
908 | struct btree *b, ptr_filter_fn fn) | |
909 | { | |
910 | struct bkey *ret; | |
911 | ||
912 | do { | |
913 | ret = bch_btree_iter_next(iter); | |
914 | } while (ret && fn(b, ret)); | |
915 | ||
916 | return ret; | |
917 | } | |
918 | ||
919 | struct bkey *bch_next_recurse_key(struct btree *b, struct bkey *search) | |
920 | { | |
921 | struct btree_iter iter; | |
922 | ||
923 | bch_btree_iter_init(b, &iter, search); | |
924 | return bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); | |
925 | } | |
926 | ||
927 | /* Mergesort */ | |
928 | ||
929 | static void btree_sort_fixup(struct btree_iter *iter) | |
930 | { | |
931 | while (iter->used > 1) { | |
932 | struct btree_iter_set *top = iter->data, *i = top + 1; | |
933 | struct bkey *k; | |
934 | ||
935 | if (iter->used > 2 && | |
936 | btree_iter_cmp(i[0], i[1])) | |
937 | i++; | |
938 | ||
939 | for (k = i->k; | |
940 | k != i->end && bkey_cmp(top->k, &START_KEY(k)) > 0; | |
941 | k = bkey_next(k)) | |
942 | if (top->k > i->k) | |
943 | __bch_cut_front(top->k, k); | |
944 | else if (KEY_SIZE(k)) | |
945 | bch_cut_back(&START_KEY(k), top->k); | |
946 | ||
947 | if (top->k < i->k || k == i->k) | |
948 | break; | |
949 | ||
950 | heap_sift(iter, i - top, btree_iter_cmp); | |
951 | } | |
952 | } | |
953 | ||
954 | static void btree_mergesort(struct btree *b, struct bset *out, | |
955 | struct btree_iter *iter, | |
956 | bool fixup, bool remove_stale) | |
957 | { | |
958 | struct bkey *k, *last = NULL; | |
959 | bool (*bad)(struct btree *, const struct bkey *) = remove_stale | |
960 | ? bch_ptr_bad | |
961 | : bch_ptr_invalid; | |
962 | ||
963 | while (!btree_iter_end(iter)) { | |
964 | if (fixup && !b->level) | |
965 | btree_sort_fixup(iter); | |
966 | ||
967 | k = bch_btree_iter_next(iter); | |
968 | if (bad(b, k)) | |
969 | continue; | |
970 | ||
971 | if (!last) { | |
972 | last = out->start; | |
973 | bkey_copy(last, k); | |
974 | } else if (b->level || | |
975 | !bch_bkey_try_merge(b, last, k)) { | |
976 | last = bkey_next(last); | |
977 | bkey_copy(last, k); | |
978 | } | |
979 | } | |
980 | ||
981 | out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0; | |
982 | ||
983 | pr_debug("sorted %i keys", out->keys); | |
984 | bch_check_key_order(b, out); | |
985 | } | |
986 | ||
987 | static void __btree_sort(struct btree *b, struct btree_iter *iter, | |
988 | unsigned start, unsigned order, bool fixup) | |
989 | { | |
990 | uint64_t start_time; | |
991 | bool remove_stale = !b->written; | |
992 | struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO, | |
993 | order); | |
994 | if (!out) { | |
995 | mutex_lock(&b->c->sort_lock); | |
996 | out = b->c->sort; | |
997 | order = ilog2(bucket_pages(b->c)); | |
998 | } | |
999 | ||
1000 | start_time = local_clock(); | |
1001 | ||
1002 | btree_mergesort(b, out, iter, fixup, remove_stale); | |
1003 | b->nsets = start; | |
1004 | ||
1005 | if (!fixup && !start && b->written) | |
1006 | bch_btree_verify(b, out); | |
1007 | ||
1008 | if (!start && order == b->page_order) { | |
1009 | /* | |
1010 | * Our temporary buffer is the same size as the btree node's | |
1011 | * buffer, we can just swap buffers instead of doing a big | |
1012 | * memcpy() | |
1013 | */ | |
1014 | ||
1015 | out->magic = bset_magic(b->c); | |
1016 | out->seq = b->sets[0].data->seq; | |
1017 | out->version = b->sets[0].data->version; | |
1018 | swap(out, b->sets[0].data); | |
1019 | ||
1020 | if (b->c->sort == b->sets[0].data) | |
1021 | b->c->sort = out; | |
1022 | } else { | |
1023 | b->sets[start].data->keys = out->keys; | |
1024 | memcpy(b->sets[start].data->start, out->start, | |
1025 | (void *) end(out) - (void *) out->start); | |
1026 | } | |
1027 | ||
1028 | if (out == b->c->sort) | |
1029 | mutex_unlock(&b->c->sort_lock); | |
1030 | else | |
1031 | free_pages((unsigned long) out, order); | |
1032 | ||
1033 | if (b->written) | |
1034 | bset_build_written_tree(b); | |
1035 | ||
1036 | if (!start) { | |
1037 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1038 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1039 | spin_unlock(&b->c->sort_time_lock); |
1040 | } | |
1041 | } | |
1042 | ||
1043 | void bch_btree_sort_partial(struct btree *b, unsigned start) | |
1044 | { | |
1045 | size_t oldsize = 0, order = b->page_order, keys = 0; | |
1046 | struct btree_iter iter; | |
1047 | __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]); | |
1048 | ||
1049 | BUG_ON(b->sets[b->nsets].data == write_block(b) && | |
1050 | (b->sets[b->nsets].size || b->nsets)); | |
1051 | ||
1052 | if (b->written) | |
1053 | oldsize = bch_count_data(b); | |
1054 | ||
1055 | if (start) { | |
1056 | unsigned i; | |
1057 | ||
1058 | for (i = start; i <= b->nsets; i++) | |
1059 | keys += b->sets[i].data->keys; | |
1060 | ||
b1a67b0f KO |
1061 | order = roundup_pow_of_two(__set_bytes(b->sets->data, |
1062 | keys)) / PAGE_SIZE; | |
cafe5635 KO |
1063 | if (order) |
1064 | order = ilog2(order); | |
1065 | } | |
1066 | ||
1067 | __btree_sort(b, &iter, start, order, false); | |
1068 | ||
1069 | EBUG_ON(b->written && bch_count_data(b) != oldsize); | |
1070 | } | |
1071 | ||
1072 | void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter) | |
1073 | { | |
1074 | BUG_ON(!b->written); | |
1075 | __btree_sort(b, iter, 0, b->page_order, true); | |
1076 | } | |
1077 | ||
1078 | void bch_btree_sort_into(struct btree *b, struct btree *new) | |
1079 | { | |
1080 | uint64_t start_time = local_clock(); | |
1081 | ||
1082 | struct btree_iter iter; | |
1083 | bch_btree_iter_init(b, &iter, NULL); | |
1084 | ||
1085 | btree_mergesort(b, new->sets->data, &iter, false, true); | |
1086 | ||
1087 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1088 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1089 | spin_unlock(&b->c->sort_time_lock); |
1090 | ||
1091 | bkey_copy_key(&new->key, &b->key); | |
1092 | new->sets->size = 0; | |
1093 | } | |
1094 | ||
1095 | void bch_btree_sort_lazy(struct btree *b) | |
1096 | { | |
1097 | if (b->nsets) { | |
1098 | unsigned i, j, keys = 0, total; | |
1099 | ||
1100 | for (i = 0; i <= b->nsets; i++) | |
1101 | keys += b->sets[i].data->keys; | |
1102 | ||
1103 | total = keys; | |
1104 | ||
1105 | for (j = 0; j < b->nsets; j++) { | |
1106 | if (keys * 2 < total || | |
1107 | keys < 1000) { | |
1108 | bch_btree_sort_partial(b, j); | |
1109 | return; | |
1110 | } | |
1111 | ||
1112 | keys -= b->sets[j].data->keys; | |
1113 | } | |
1114 | ||
1115 | /* Must sort if b->nsets == 3 or we'll overflow */ | |
1116 | if (b->nsets >= (MAX_BSETS - 1) - b->level) { | |
1117 | bch_btree_sort(b); | |
1118 | return; | |
1119 | } | |
1120 | } | |
1121 | ||
1122 | bset_build_written_tree(b); | |
1123 | } | |
1124 | ||
1125 | /* Sysfs stuff */ | |
1126 | ||
1127 | struct bset_stats { | |
1128 | size_t nodes; | |
1129 | size_t sets_written, sets_unwritten; | |
1130 | size_t bytes_written, bytes_unwritten; | |
1131 | size_t floats, failed; | |
1132 | }; | |
1133 | ||
1134 | static int bch_btree_bset_stats(struct btree *b, struct btree_op *op, | |
1135 | struct bset_stats *stats) | |
1136 | { | |
1137 | struct bkey *k; | |
1138 | unsigned i; | |
1139 | ||
1140 | stats->nodes++; | |
1141 | ||
1142 | for (i = 0; i <= b->nsets; i++) { | |
1143 | struct bset_tree *t = &b->sets[i]; | |
1144 | size_t bytes = t->data->keys * sizeof(uint64_t); | |
1145 | size_t j; | |
1146 | ||
1147 | if (bset_written(b, t)) { | |
1148 | stats->sets_written++; | |
1149 | stats->bytes_written += bytes; | |
1150 | ||
1151 | stats->floats += t->size - 1; | |
1152 | ||
1153 | for (j = 1; j < t->size; j++) | |
1154 | if (t->tree[j].exponent == 127) | |
1155 | stats->failed++; | |
1156 | } else { | |
1157 | stats->sets_unwritten++; | |
1158 | stats->bytes_unwritten += bytes; | |
1159 | } | |
1160 | } | |
1161 | ||
1162 | if (b->level) { | |
1163 | struct btree_iter iter; | |
1164 | ||
1165 | for_each_key_filter(b, k, &iter, bch_ptr_bad) { | |
1166 | int ret = btree(bset_stats, k, b, op, stats); | |
1167 | if (ret) | |
1168 | return ret; | |
1169 | } | |
1170 | } | |
1171 | ||
1172 | return 0; | |
1173 | } | |
1174 | ||
1175 | int bch_bset_print_stats(struct cache_set *c, char *buf) | |
1176 | { | |
1177 | struct btree_op op; | |
1178 | struct bset_stats t; | |
1179 | int ret; | |
1180 | ||
1181 | bch_btree_op_init_stack(&op); | |
1182 | memset(&t, 0, sizeof(struct bset_stats)); | |
1183 | ||
1184 | ret = btree_root(bset_stats, c, &op, &t); | |
1185 | if (ret) | |
1186 | return ret; | |
1187 | ||
1188 | return snprintf(buf, PAGE_SIZE, | |
1189 | "btree nodes: %zu\n" | |
1190 | "written sets: %zu\n" | |
1191 | "unwritten sets: %zu\n" | |
1192 | "written key bytes: %zu\n" | |
1193 | "unwritten key bytes: %zu\n" | |
1194 | "floats: %zu\n" | |
1195 | "failed: %zu\n", | |
1196 | t.nodes, | |
1197 | t.sets_written, t.sets_unwritten, | |
1198 | t.bytes_written, t.bytes_unwritten, | |
1199 | t.floats, t.failed); | |
1200 | } |