]>
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 | ||
cafe5635 KO |
17 | int bch_keylist_realloc(struct keylist *l, int nptrs, struct cache_set *c) |
18 | { | |
c2f95ae2 KO |
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; | |
cafe5635 KO |
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 | ||
c2f95ae2 | 38 | new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO); |
cafe5635 | 39 | |
c2f95ae2 | 40 | if (!new_keys) |
cafe5635 KO |
41 | return -ENOMEM; |
42 | ||
c2f95ae2 KO |
43 | if (!old_keys) |
44 | memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize); | |
cafe5635 | 45 | |
c2f95ae2 KO |
46 | l->keys_p = new_keys; |
47 | l->top_p = new_keys + oldsize; | |
cafe5635 KO |
48 | |
49 | return 0; | |
50 | } | |
51 | ||
52 | struct bkey *bch_keylist_pop(struct keylist *l) | |
53 | { | |
c2f95ae2 | 54 | struct bkey *k = l->keys; |
cafe5635 KO |
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 | ||
26c949f8 KO |
65 | void bch_keylist_pop_front(struct keylist *l) |
66 | { | |
c2f95ae2 | 67 | l->top_p -= bkey_u64s(l->keys); |
26c949f8 | 68 | |
c2f95ae2 KO |
69 | memmove(l->keys, |
70 | bkey_next(l->keys), | |
71 | bch_keylist_bytes(l)); | |
26c949f8 KO |
72 | } |
73 | ||
cafe5635 KO |
74 | /* Pointer validation */ |
75 | ||
d5cc66e9 | 76 | static bool __ptr_invalid(struct cache_set *c, const struct bkey *k) |
cafe5635 KO |
77 | { |
78 | unsigned i; | |
cafe5635 KO |
79 | |
80 | for (i = 0; i < KEY_PTRS(k); i++) | |
81 | if (ptr_available(c, k, i)) { | |
82 | struct cache *ca = PTR_CACHE(c, k, i); | |
83 | size_t bucket = PTR_BUCKET_NR(c, k, i); | |
84 | size_t r = bucket_remainder(c, PTR_OFFSET(k, i)); | |
85 | ||
86 | if (KEY_SIZE(k) + r > c->sb.bucket_size || | |
87 | bucket < ca->sb.first_bucket || | |
88 | bucket >= ca->sb.nbuckets) | |
d5cc66e9 | 89 | return true; |
cafe5635 KO |
90 | } |
91 | ||
d5cc66e9 KO |
92 | return false; |
93 | } | |
94 | ||
95 | bool bch_btree_ptr_invalid(struct cache_set *c, const struct bkey *k) | |
96 | { | |
97 | char buf[80]; | |
98 | ||
99 | if (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k)) | |
100 | goto bad; | |
101 | ||
102 | if (__ptr_invalid(c, k)) | |
103 | goto bad; | |
104 | ||
105 | return false; | |
106 | bad: | |
107 | bch_bkey_to_text(buf, sizeof(buf), k); | |
108 | cache_bug(c, "spotted btree ptr %s: %s", buf, bch_ptr_status(c, k)); | |
109 | return true; | |
110 | } | |
111 | ||
112 | bool bch_extent_ptr_invalid(struct cache_set *c, const struct bkey *k) | |
113 | { | |
114 | char buf[80]; | |
115 | ||
116 | if (!KEY_SIZE(k)) | |
117 | return true; | |
118 | ||
119 | if (KEY_SIZE(k) > KEY_OFFSET(k)) | |
120 | goto bad; | |
121 | ||
122 | if (__ptr_invalid(c, k)) | |
123 | goto bad; | |
124 | ||
cafe5635 KO |
125 | return false; |
126 | bad: | |
85b1492e | 127 | bch_bkey_to_text(buf, sizeof(buf), k); |
d5cc66e9 | 128 | cache_bug(c, "spotted extent %s: %s", buf, bch_ptr_status(c, k)); |
cafe5635 KO |
129 | return true; |
130 | } | |
131 | ||
280481d0 KO |
132 | static bool ptr_bad_expensive_checks(struct btree *b, const struct bkey *k, |
133 | unsigned ptr) | |
134 | { | |
135 | struct bucket *g = PTR_BUCKET(b->c, k, ptr); | |
136 | char buf[80]; | |
137 | ||
138 | if (mutex_trylock(&b->c->bucket_lock)) { | |
139 | if (b->level) { | |
140 | if (KEY_DIRTY(k) || | |
141 | g->prio != BTREE_PRIO || | |
142 | (b->c->gc_mark_valid && | |
143 | GC_MARK(g) != GC_MARK_METADATA)) | |
144 | goto err; | |
145 | ||
146 | } else { | |
147 | if (g->prio == BTREE_PRIO) | |
148 | goto err; | |
149 | ||
150 | if (KEY_DIRTY(k) && | |
151 | b->c->gc_mark_valid && | |
152 | GC_MARK(g) != GC_MARK_DIRTY) | |
153 | goto err; | |
154 | } | |
155 | mutex_unlock(&b->c->bucket_lock); | |
156 | } | |
157 | ||
158 | return false; | |
159 | err: | |
160 | mutex_unlock(&b->c->bucket_lock); | |
161 | bch_bkey_to_text(buf, sizeof(buf), k); | |
162 | btree_bug(b, | |
163 | "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i", | |
164 | buf, PTR_BUCKET_NR(b->c, k, ptr), atomic_read(&g->pin), | |
165 | g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen); | |
166 | return true; | |
167 | } | |
168 | ||
cafe5635 KO |
169 | bool bch_ptr_bad(struct btree *b, const struct bkey *k) |
170 | { | |
171 | struct bucket *g; | |
172 | unsigned i, stale; | |
173 | ||
174 | if (!bkey_cmp(k, &ZERO_KEY) || | |
175 | !KEY_PTRS(k) || | |
176 | bch_ptr_invalid(b, k)) | |
177 | return true; | |
178 | ||
e58ff155 KO |
179 | for (i = 0; i < KEY_PTRS(k); i++) { |
180 | if (!ptr_available(b->c, k, i)) | |
181 | return true; | |
cafe5635 | 182 | |
e58ff155 KO |
183 | g = PTR_BUCKET(b->c, k, i); |
184 | stale = ptr_stale(b->c, k, i); | |
cafe5635 | 185 | |
e58ff155 KO |
186 | btree_bug_on(stale > 96, b, |
187 | "key too stale: %i, need_gc %u", | |
188 | stale, b->c->need_gc); | |
cafe5635 | 189 | |
e58ff155 KO |
190 | btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k), |
191 | b, "stale dirty pointer"); | |
cafe5635 | 192 | |
e58ff155 KO |
193 | if (stale) |
194 | return true; | |
cafe5635 | 195 | |
280481d0 KO |
196 | if (expensive_debug_checks(b->c) && |
197 | ptr_bad_expensive_checks(b, k, i)) | |
198 | return true; | |
e58ff155 | 199 | } |
cafe5635 KO |
200 | |
201 | return false; | |
cafe5635 KO |
202 | } |
203 | ||
204 | /* Key/pointer manipulation */ | |
205 | ||
206 | void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src, | |
207 | unsigned i) | |
208 | { | |
209 | BUG_ON(i > KEY_PTRS(src)); | |
210 | ||
211 | /* Only copy the header, key, and one pointer. */ | |
212 | memcpy(dest, src, 2 * sizeof(uint64_t)); | |
213 | dest->ptr[0] = src->ptr[i]; | |
214 | SET_KEY_PTRS(dest, 1); | |
215 | /* We didn't copy the checksum so clear that bit. */ | |
216 | SET_KEY_CSUM(dest, 0); | |
217 | } | |
218 | ||
219 | bool __bch_cut_front(const struct bkey *where, struct bkey *k) | |
220 | { | |
221 | unsigned i, len = 0; | |
222 | ||
223 | if (bkey_cmp(where, &START_KEY(k)) <= 0) | |
224 | return false; | |
225 | ||
226 | if (bkey_cmp(where, k) < 0) | |
227 | len = KEY_OFFSET(k) - KEY_OFFSET(where); | |
228 | else | |
229 | bkey_copy_key(k, where); | |
230 | ||
231 | for (i = 0; i < KEY_PTRS(k); i++) | |
232 | SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len); | |
233 | ||
234 | BUG_ON(len > KEY_SIZE(k)); | |
235 | SET_KEY_SIZE(k, len); | |
236 | return true; | |
237 | } | |
238 | ||
239 | bool __bch_cut_back(const struct bkey *where, struct bkey *k) | |
240 | { | |
241 | unsigned len = 0; | |
242 | ||
243 | if (bkey_cmp(where, k) >= 0) | |
244 | return false; | |
245 | ||
246 | BUG_ON(KEY_INODE(where) != KEY_INODE(k)); | |
247 | ||
248 | if (bkey_cmp(where, &START_KEY(k)) > 0) | |
249 | len = KEY_OFFSET(where) - KEY_START(k); | |
250 | ||
251 | bkey_copy_key(k, where); | |
252 | ||
253 | BUG_ON(len > KEY_SIZE(k)); | |
254 | SET_KEY_SIZE(k, len); | |
255 | return true; | |
256 | } | |
257 | ||
258 | static uint64_t merge_chksums(struct bkey *l, struct bkey *r) | |
259 | { | |
260 | return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) & | |
261 | ~((uint64_t)1 << 63); | |
262 | } | |
263 | ||
264 | /* Tries to merge l and r: l should be lower than r | |
265 | * Returns true if we were able to merge. If we did merge, l will be the merged | |
266 | * key, r will be untouched. | |
267 | */ | |
268 | bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r) | |
269 | { | |
270 | unsigned i; | |
271 | ||
272 | if (key_merging_disabled(b->c)) | |
273 | return false; | |
274 | ||
275 | if (KEY_PTRS(l) != KEY_PTRS(r) || | |
276 | KEY_DIRTY(l) != KEY_DIRTY(r) || | |
277 | bkey_cmp(l, &START_KEY(r))) | |
278 | return false; | |
279 | ||
280 | for (i = 0; i < KEY_PTRS(l); i++) | |
281 | if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] || | |
282 | PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i)) | |
283 | return false; | |
284 | ||
285 | /* Keys with no pointers aren't restricted to one bucket and could | |
286 | * overflow KEY_SIZE | |
287 | */ | |
288 | if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) { | |
289 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l)); | |
290 | SET_KEY_SIZE(l, USHRT_MAX); | |
291 | ||
292 | bch_cut_front(l, r); | |
293 | return false; | |
294 | } | |
295 | ||
296 | if (KEY_CSUM(l)) { | |
297 | if (KEY_CSUM(r)) | |
298 | l->ptr[KEY_PTRS(l)] = merge_chksums(l, r); | |
299 | else | |
300 | SET_KEY_CSUM(l, 0); | |
301 | } | |
302 | ||
303 | SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r)); | |
304 | SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r)); | |
305 | ||
306 | return true; | |
307 | } | |
308 | ||
309 | /* Binary tree stuff for auxiliary search trees */ | |
310 | ||
311 | static unsigned inorder_next(unsigned j, unsigned size) | |
312 | { | |
313 | if (j * 2 + 1 < size) { | |
314 | j = j * 2 + 1; | |
315 | ||
316 | while (j * 2 < size) | |
317 | j *= 2; | |
318 | } else | |
319 | j >>= ffz(j) + 1; | |
320 | ||
321 | return j; | |
322 | } | |
323 | ||
324 | static unsigned inorder_prev(unsigned j, unsigned size) | |
325 | { | |
326 | if (j * 2 < size) { | |
327 | j = j * 2; | |
328 | ||
329 | while (j * 2 + 1 < size) | |
330 | j = j * 2 + 1; | |
331 | } else | |
332 | j >>= ffs(j); | |
333 | ||
334 | return j; | |
335 | } | |
336 | ||
337 | /* I have no idea why this code works... and I'm the one who wrote it | |
338 | * | |
339 | * However, I do know what it does: | |
340 | * Given a binary tree constructed in an array (i.e. how you normally implement | |
341 | * a heap), it converts a node in the tree - referenced by array index - to the | |
342 | * index it would have if you did an inorder traversal. | |
343 | * | |
344 | * Also tested for every j, size up to size somewhere around 6 million. | |
345 | * | |
346 | * The binary tree starts at array index 1, not 0 | |
347 | * extra is a function of size: | |
348 | * extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
349 | */ | |
350 | static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra) | |
351 | { | |
352 | unsigned b = fls(j); | |
353 | unsigned shift = fls(size - 1) - b; | |
354 | ||
355 | j ^= 1U << (b - 1); | |
356 | j <<= 1; | |
357 | j |= 1; | |
358 | j <<= shift; | |
359 | ||
360 | if (j > extra) | |
361 | j -= (j - extra) >> 1; | |
362 | ||
363 | return j; | |
364 | } | |
365 | ||
366 | static unsigned to_inorder(unsigned j, struct bset_tree *t) | |
367 | { | |
368 | return __to_inorder(j, t->size, t->extra); | |
369 | } | |
370 | ||
371 | static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra) | |
372 | { | |
373 | unsigned shift; | |
374 | ||
375 | if (j > extra) | |
376 | j += j - extra; | |
377 | ||
378 | shift = ffs(j); | |
379 | ||
380 | j >>= shift; | |
381 | j |= roundup_pow_of_two(size) >> shift; | |
382 | ||
383 | return j; | |
384 | } | |
385 | ||
386 | static unsigned inorder_to_tree(unsigned j, struct bset_tree *t) | |
387 | { | |
388 | return __inorder_to_tree(j, t->size, t->extra); | |
389 | } | |
390 | ||
391 | #if 0 | |
392 | void inorder_test(void) | |
393 | { | |
394 | unsigned long done = 0; | |
395 | ktime_t start = ktime_get(); | |
396 | ||
397 | for (unsigned size = 2; | |
398 | size < 65536000; | |
399 | size++) { | |
400 | unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1; | |
401 | unsigned i = 1, j = rounddown_pow_of_two(size - 1); | |
402 | ||
403 | if (!(size % 4096)) | |
404 | printk(KERN_NOTICE "loop %u, %llu per us\n", size, | |
405 | done / ktime_us_delta(ktime_get(), start)); | |
406 | ||
407 | while (1) { | |
408 | if (__inorder_to_tree(i, size, extra) != j) | |
409 | panic("size %10u j %10u i %10u", size, j, i); | |
410 | ||
411 | if (__to_inorder(j, size, extra) != i) | |
412 | panic("size %10u j %10u i %10u", size, j, i); | |
413 | ||
414 | if (j == rounddown_pow_of_two(size) - 1) | |
415 | break; | |
416 | ||
417 | BUG_ON(inorder_prev(inorder_next(j, size), size) != j); | |
418 | ||
419 | j = inorder_next(j, size); | |
420 | i++; | |
421 | } | |
422 | ||
423 | done += size - 1; | |
424 | } | |
425 | } | |
426 | #endif | |
427 | ||
428 | /* | |
48a73025 | 429 | * Cacheline/offset <-> bkey pointer arithmetic: |
cafe5635 KO |
430 | * |
431 | * t->tree is a binary search tree in an array; each node corresponds to a key | |
432 | * in one cacheline in t->set (BSET_CACHELINE bytes). | |
433 | * | |
434 | * This means we don't have to store the full index of the key that a node in | |
435 | * the binary tree points to; to_inorder() gives us the cacheline, and then | |
436 | * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes. | |
437 | * | |
48a73025 | 438 | * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to |
cafe5635 KO |
439 | * make this work. |
440 | * | |
441 | * To construct the bfloat for an arbitrary key we need to know what the key | |
442 | * immediately preceding it is: we have to check if the two keys differ in the | |
443 | * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size | |
444 | * of the previous key so we can walk backwards to it from t->tree[j]'s key. | |
445 | */ | |
446 | ||
447 | static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline, | |
448 | unsigned offset) | |
449 | { | |
450 | return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8; | |
451 | } | |
452 | ||
453 | static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k) | |
454 | { | |
455 | return ((void *) k - (void *) t->data) / BSET_CACHELINE; | |
456 | } | |
457 | ||
458 | static unsigned bkey_to_cacheline_offset(struct bkey *k) | |
459 | { | |
460 | return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t); | |
461 | } | |
462 | ||
463 | static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j) | |
464 | { | |
465 | return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m); | |
466 | } | |
467 | ||
468 | static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j) | |
469 | { | |
470 | return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]); | |
471 | } | |
472 | ||
473 | /* | |
474 | * For the write set - the one we're currently inserting keys into - we don't | |
475 | * maintain a full search tree, we just keep a simple lookup table in t->prev. | |
476 | */ | |
477 | static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline) | |
478 | { | |
479 | return cacheline_to_bkey(t, cacheline, t->prev[cacheline]); | |
480 | } | |
481 | ||
482 | static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift) | |
483 | { | |
484 | #ifdef CONFIG_X86_64 | |
485 | asm("shrd %[shift],%[high],%[low]" | |
486 | : [low] "+Rm" (low) | |
487 | : [high] "R" (high), | |
488 | [shift] "ci" (shift) | |
489 | : "cc"); | |
490 | #else | |
491 | low >>= shift; | |
492 | low |= (high << 1) << (63U - shift); | |
493 | #endif | |
494 | return low; | |
495 | } | |
496 | ||
497 | static inline unsigned bfloat_mantissa(const struct bkey *k, | |
498 | struct bkey_float *f) | |
499 | { | |
500 | const uint64_t *p = &k->low - (f->exponent >> 6); | |
501 | return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK; | |
502 | } | |
503 | ||
504 | static void make_bfloat(struct bset_tree *t, unsigned j) | |
505 | { | |
506 | struct bkey_float *f = &t->tree[j]; | |
507 | struct bkey *m = tree_to_bkey(t, j); | |
508 | struct bkey *p = tree_to_prev_bkey(t, j); | |
509 | ||
510 | struct bkey *l = is_power_of_2(j) | |
511 | ? t->data->start | |
512 | : tree_to_prev_bkey(t, j >> ffs(j)); | |
513 | ||
514 | struct bkey *r = is_power_of_2(j + 1) | |
515 | ? node(t->data, t->data->keys - bkey_u64s(&t->end)) | |
516 | : tree_to_bkey(t, j >> (ffz(j) + 1)); | |
517 | ||
518 | BUG_ON(m < l || m > r); | |
519 | BUG_ON(bkey_next(p) != m); | |
520 | ||
521 | if (KEY_INODE(l) != KEY_INODE(r)) | |
522 | f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64; | |
523 | else | |
524 | f->exponent = fls64(r->low ^ l->low); | |
525 | ||
526 | f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0); | |
527 | ||
528 | /* | |
529 | * Setting f->exponent = 127 flags this node as failed, and causes the | |
530 | * lookup code to fall back to comparing against the original key. | |
531 | */ | |
532 | ||
533 | if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f)) | |
534 | f->mantissa = bfloat_mantissa(m, f) - 1; | |
535 | else | |
536 | f->exponent = 127; | |
537 | } | |
538 | ||
539 | static void bset_alloc_tree(struct btree *b, struct bset_tree *t) | |
540 | { | |
541 | if (t != b->sets) { | |
542 | unsigned j = roundup(t[-1].size, | |
543 | 64 / sizeof(struct bkey_float)); | |
544 | ||
545 | t->tree = t[-1].tree + j; | |
546 | t->prev = t[-1].prev + j; | |
547 | } | |
548 | ||
549 | while (t < b->sets + MAX_BSETS) | |
550 | t++->size = 0; | |
551 | } | |
552 | ||
553 | static void bset_build_unwritten_tree(struct btree *b) | |
554 | { | |
555 | struct bset_tree *t = b->sets + b->nsets; | |
556 | ||
557 | bset_alloc_tree(b, t); | |
558 | ||
559 | if (t->tree != b->sets->tree + bset_tree_space(b)) { | |
560 | t->prev[0] = bkey_to_cacheline_offset(t->data->start); | |
561 | t->size = 1; | |
562 | } | |
563 | } | |
564 | ||
565 | static void bset_build_written_tree(struct btree *b) | |
566 | { | |
567 | struct bset_tree *t = b->sets + b->nsets; | |
568 | struct bkey *k = t->data->start; | |
569 | unsigned j, cacheline = 1; | |
570 | ||
571 | bset_alloc_tree(b, t); | |
572 | ||
573 | t->size = min_t(unsigned, | |
574 | bkey_to_cacheline(t, end(t->data)), | |
575 | b->sets->tree + bset_tree_space(b) - t->tree); | |
576 | ||
577 | if (t->size < 2) { | |
578 | t->size = 0; | |
579 | return; | |
580 | } | |
581 | ||
582 | t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1; | |
583 | ||
584 | /* First we figure out where the first key in each cacheline is */ | |
585 | for (j = inorder_next(0, t->size); | |
586 | j; | |
587 | j = inorder_next(j, t->size)) { | |
588 | while (bkey_to_cacheline(t, k) != cacheline) | |
589 | k = bkey_next(k); | |
590 | ||
591 | t->prev[j] = bkey_u64s(k); | |
592 | k = bkey_next(k); | |
593 | cacheline++; | |
594 | t->tree[j].m = bkey_to_cacheline_offset(k); | |
595 | } | |
596 | ||
597 | while (bkey_next(k) != end(t->data)) | |
598 | k = bkey_next(k); | |
599 | ||
600 | t->end = *k; | |
601 | ||
602 | /* Then we build the tree */ | |
603 | for (j = inorder_next(0, t->size); | |
604 | j; | |
605 | j = inorder_next(j, t->size)) | |
606 | make_bfloat(t, j); | |
607 | } | |
608 | ||
609 | void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k) | |
610 | { | |
611 | struct bset_tree *t; | |
612 | unsigned inorder, j = 1; | |
613 | ||
614 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
615 | if (k < end(t->data)) | |
616 | goto found_set; | |
617 | ||
618 | BUG(); | |
619 | found_set: | |
620 | if (!t->size || !bset_written(b, t)) | |
621 | return; | |
622 | ||
623 | inorder = bkey_to_cacheline(t, k); | |
624 | ||
625 | if (k == t->data->start) | |
626 | goto fix_left; | |
627 | ||
628 | if (bkey_next(k) == end(t->data)) { | |
629 | t->end = *k; | |
630 | goto fix_right; | |
631 | } | |
632 | ||
633 | j = inorder_to_tree(inorder, t); | |
634 | ||
635 | if (j && | |
636 | j < t->size && | |
637 | k == tree_to_bkey(t, j)) | |
638 | fix_left: do { | |
639 | make_bfloat(t, j); | |
640 | j = j * 2; | |
641 | } while (j < t->size); | |
642 | ||
643 | j = inorder_to_tree(inorder + 1, t); | |
644 | ||
645 | if (j && | |
646 | j < t->size && | |
647 | k == tree_to_prev_bkey(t, j)) | |
648 | fix_right: do { | |
649 | make_bfloat(t, j); | |
650 | j = j * 2 + 1; | |
651 | } while (j < t->size); | |
652 | } | |
653 | ||
654 | void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k) | |
655 | { | |
656 | struct bset_tree *t = &b->sets[b->nsets]; | |
657 | unsigned shift = bkey_u64s(k); | |
658 | unsigned j = bkey_to_cacheline(t, k); | |
659 | ||
660 | /* We're getting called from btree_split() or btree_gc, just bail out */ | |
661 | if (!t->size) | |
662 | return; | |
663 | ||
664 | /* k is the key we just inserted; we need to find the entry in the | |
665 | * lookup table for the first key that is strictly greater than k: | |
666 | * it's either k's cacheline or the next one | |
667 | */ | |
668 | if (j < t->size && | |
669 | table_to_bkey(t, j) <= k) | |
670 | j++; | |
671 | ||
672 | /* Adjust all the lookup table entries, and find a new key for any that | |
673 | * have gotten too big | |
674 | */ | |
675 | for (; j < t->size; j++) { | |
676 | t->prev[j] += shift; | |
677 | ||
678 | if (t->prev[j] > 7) { | |
679 | k = table_to_bkey(t, j - 1); | |
680 | ||
681 | while (k < cacheline_to_bkey(t, j, 0)) | |
682 | k = bkey_next(k); | |
683 | ||
684 | t->prev[j] = bkey_to_cacheline_offset(k); | |
685 | } | |
686 | } | |
687 | ||
688 | if (t->size == b->sets->tree + bset_tree_space(b) - t->tree) | |
689 | return; | |
690 | ||
691 | /* Possibly add a new entry to the end of the lookup table */ | |
692 | ||
693 | for (k = table_to_bkey(t, t->size - 1); | |
694 | k != end(t->data); | |
695 | k = bkey_next(k)) | |
696 | if (t->size == bkey_to_cacheline(t, k)) { | |
697 | t->prev[t->size] = bkey_to_cacheline_offset(k); | |
698 | t->size++; | |
699 | } | |
700 | } | |
701 | ||
702 | void bch_bset_init_next(struct btree *b) | |
703 | { | |
704 | struct bset *i = write_block(b); | |
705 | ||
706 | if (i != b->sets[0].data) { | |
707 | b->sets[++b->nsets].data = i; | |
708 | i->seq = b->sets[0].data->seq; | |
709 | } else | |
710 | get_random_bytes(&i->seq, sizeof(uint64_t)); | |
711 | ||
81ab4190 | 712 | i->magic = bset_magic(&b->c->sb); |
cafe5635 KO |
713 | i->version = 0; |
714 | i->keys = 0; | |
715 | ||
716 | bset_build_unwritten_tree(b); | |
717 | } | |
718 | ||
719 | struct bset_search_iter { | |
720 | struct bkey *l, *r; | |
721 | }; | |
722 | ||
723 | static struct bset_search_iter bset_search_write_set(struct btree *b, | |
724 | struct bset_tree *t, | |
725 | const struct bkey *search) | |
726 | { | |
727 | unsigned li = 0, ri = t->size; | |
728 | ||
729 | BUG_ON(!b->nsets && | |
730 | t->size < bkey_to_cacheline(t, end(t->data))); | |
731 | ||
732 | while (li + 1 != ri) { | |
733 | unsigned m = (li + ri) >> 1; | |
734 | ||
735 | if (bkey_cmp(table_to_bkey(t, m), search) > 0) | |
736 | ri = m; | |
737 | else | |
738 | li = m; | |
739 | } | |
740 | ||
741 | return (struct bset_search_iter) { | |
742 | table_to_bkey(t, li), | |
743 | ri < t->size ? table_to_bkey(t, ri) : end(t->data) | |
744 | }; | |
745 | } | |
746 | ||
747 | static struct bset_search_iter bset_search_tree(struct btree *b, | |
748 | struct bset_tree *t, | |
749 | const struct bkey *search) | |
750 | { | |
751 | struct bkey *l, *r; | |
752 | struct bkey_float *f; | |
753 | unsigned inorder, j, n = 1; | |
754 | ||
755 | do { | |
756 | unsigned p = n << 4; | |
757 | p &= ((int) (p - t->size)) >> 31; | |
758 | ||
759 | prefetch(&t->tree[p]); | |
760 | ||
761 | j = n; | |
762 | f = &t->tree[j]; | |
763 | ||
764 | /* | |
765 | * n = (f->mantissa > bfloat_mantissa()) | |
766 | * ? j * 2 | |
767 | * : j * 2 + 1; | |
768 | * | |
769 | * We need to subtract 1 from f->mantissa for the sign bit trick | |
770 | * to work - that's done in make_bfloat() | |
771 | */ | |
772 | if (likely(f->exponent != 127)) | |
773 | n = j * 2 + (((unsigned) | |
774 | (f->mantissa - | |
775 | bfloat_mantissa(search, f))) >> 31); | |
776 | else | |
777 | n = (bkey_cmp(tree_to_bkey(t, j), search) > 0) | |
778 | ? j * 2 | |
779 | : j * 2 + 1; | |
780 | } while (n < t->size); | |
781 | ||
782 | inorder = to_inorder(j, t); | |
783 | ||
784 | /* | |
785 | * n would have been the node we recursed to - the low bit tells us if | |
786 | * we recursed left or recursed right. | |
787 | */ | |
788 | if (n & 1) { | |
789 | l = cacheline_to_bkey(t, inorder, f->m); | |
790 | ||
791 | if (++inorder != t->size) { | |
792 | f = &t->tree[inorder_next(j, t->size)]; | |
793 | r = cacheline_to_bkey(t, inorder, f->m); | |
794 | } else | |
795 | r = end(t->data); | |
796 | } else { | |
797 | r = cacheline_to_bkey(t, inorder, f->m); | |
798 | ||
799 | if (--inorder) { | |
800 | f = &t->tree[inorder_prev(j, t->size)]; | |
801 | l = cacheline_to_bkey(t, inorder, f->m); | |
802 | } else | |
803 | l = t->data->start; | |
804 | } | |
805 | ||
806 | return (struct bset_search_iter) {l, r}; | |
807 | } | |
808 | ||
809 | struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t, | |
810 | const struct bkey *search) | |
811 | { | |
812 | struct bset_search_iter i; | |
813 | ||
814 | /* | |
815 | * First, we search for a cacheline, then lastly we do a linear search | |
816 | * within that cacheline. | |
817 | * | |
818 | * To search for the cacheline, there's three different possibilities: | |
819 | * * The set is too small to have a search tree, so we just do a linear | |
820 | * search over the whole set. | |
821 | * * The set is the one we're currently inserting into; keeping a full | |
822 | * auxiliary search tree up to date would be too expensive, so we | |
823 | * use a much simpler lookup table to do a binary search - | |
824 | * bset_search_write_set(). | |
825 | * * Or we use the auxiliary search tree we constructed earlier - | |
826 | * bset_search_tree() | |
827 | */ | |
828 | ||
829 | if (unlikely(!t->size)) { | |
830 | i.l = t->data->start; | |
831 | i.r = end(t->data); | |
832 | } else if (bset_written(b, t)) { | |
833 | /* | |
834 | * Each node in the auxiliary search tree covers a certain range | |
835 | * of bits, and keys above and below the set it covers might | |
836 | * differ outside those bits - so we have to special case the | |
837 | * start and end - handle that here: | |
838 | */ | |
839 | ||
840 | if (unlikely(bkey_cmp(search, &t->end) >= 0)) | |
841 | return end(t->data); | |
842 | ||
843 | if (unlikely(bkey_cmp(search, t->data->start) < 0)) | |
844 | return t->data->start; | |
845 | ||
846 | i = bset_search_tree(b, t, search); | |
847 | } else | |
848 | i = bset_search_write_set(b, t, search); | |
849 | ||
280481d0 KO |
850 | if (expensive_debug_checks(b->c)) { |
851 | BUG_ON(bset_written(b, t) && | |
852 | i.l != t->data->start && | |
853 | bkey_cmp(tree_to_prev_bkey(t, | |
854 | inorder_to_tree(bkey_to_cacheline(t, i.l), t)), | |
855 | search) > 0); | |
cafe5635 | 856 | |
280481d0 KO |
857 | BUG_ON(i.r != end(t->data) && |
858 | bkey_cmp(i.r, search) <= 0); | |
859 | } | |
cafe5635 KO |
860 | |
861 | while (likely(i.l != i.r) && | |
862 | bkey_cmp(i.l, search) <= 0) | |
863 | i.l = bkey_next(i.l); | |
864 | ||
865 | return i.l; | |
866 | } | |
867 | ||
868 | /* Btree iterator */ | |
869 | ||
48dad8ba KO |
870 | /* |
871 | * Returns true if l > r - unless l == r, in which case returns true if l is | |
872 | * older than r. | |
873 | * | |
874 | * Necessary for btree_sort_fixup() - if there are multiple keys that compare | |
875 | * equal in different sets, we have to process them newest to oldest. | |
876 | */ | |
cafe5635 KO |
877 | static inline bool btree_iter_cmp(struct btree_iter_set l, |
878 | struct btree_iter_set r) | |
879 | { | |
880 | int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k)); | |
881 | ||
882 | return c ? c > 0 : l.k < r.k; | |
883 | } | |
884 | ||
885 | static inline bool btree_iter_end(struct btree_iter *iter) | |
886 | { | |
887 | return !iter->used; | |
888 | } | |
889 | ||
890 | void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k, | |
891 | struct bkey *end) | |
892 | { | |
893 | if (k != end) | |
894 | BUG_ON(!heap_add(iter, | |
895 | ((struct btree_iter_set) { k, end }), | |
896 | btree_iter_cmp)); | |
897 | } | |
898 | ||
899 | struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter, | |
280481d0 | 900 | struct bkey *search, struct bset_tree *start) |
cafe5635 KO |
901 | { |
902 | struct bkey *ret = NULL; | |
903 | iter->size = ARRAY_SIZE(iter->data); | |
904 | iter->used = 0; | |
905 | ||
280481d0 KO |
906 | #ifdef CONFIG_BCACHE_DEBUG |
907 | iter->b = b; | |
908 | #endif | |
909 | ||
cafe5635 KO |
910 | for (; start <= &b->sets[b->nsets]; start++) { |
911 | ret = bch_bset_search(b, start, search); | |
912 | bch_btree_iter_push(iter, ret, end(start->data)); | |
913 | } | |
914 | ||
915 | return ret; | |
916 | } | |
917 | ||
918 | struct bkey *bch_btree_iter_next(struct btree_iter *iter) | |
919 | { | |
920 | struct btree_iter_set unused; | |
921 | struct bkey *ret = NULL; | |
922 | ||
923 | if (!btree_iter_end(iter)) { | |
280481d0 KO |
924 | bch_btree_iter_next_check(iter); |
925 | ||
cafe5635 KO |
926 | ret = iter->data->k; |
927 | iter->data->k = bkey_next(iter->data->k); | |
928 | ||
929 | if (iter->data->k > iter->data->end) { | |
cc0f4eaa | 930 | WARN_ONCE(1, "bset was corrupt!\n"); |
cafe5635 KO |
931 | iter->data->k = iter->data->end; |
932 | } | |
933 | ||
934 | if (iter->data->k == iter->data->end) | |
935 | heap_pop(iter, unused, btree_iter_cmp); | |
936 | else | |
937 | heap_sift(iter, 0, btree_iter_cmp); | |
938 | } | |
939 | ||
940 | return ret; | |
941 | } | |
942 | ||
943 | struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter, | |
944 | struct btree *b, ptr_filter_fn fn) | |
945 | { | |
946 | struct bkey *ret; | |
947 | ||
948 | do { | |
949 | ret = bch_btree_iter_next(iter); | |
950 | } while (ret && fn(b, ret)); | |
951 | ||
952 | return ret; | |
953 | } | |
954 | ||
cafe5635 KO |
955 | /* Mergesort */ |
956 | ||
84786438 KO |
957 | static void sort_key_next(struct btree_iter *iter, |
958 | struct btree_iter_set *i) | |
959 | { | |
960 | i->k = bkey_next(i->k); | |
961 | ||
962 | if (i->k == i->end) | |
963 | *i = iter->data[--iter->used]; | |
964 | } | |
965 | ||
cafe5635 KO |
966 | static void btree_sort_fixup(struct btree_iter *iter) |
967 | { | |
968 | while (iter->used > 1) { | |
969 | struct btree_iter_set *top = iter->data, *i = top + 1; | |
cafe5635 KO |
970 | |
971 | if (iter->used > 2 && | |
972 | btree_iter_cmp(i[0], i[1])) | |
973 | i++; | |
974 | ||
84786438 | 975 | if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0) |
cafe5635 KO |
976 | break; |
977 | ||
84786438 KO |
978 | if (!KEY_SIZE(i->k)) { |
979 | sort_key_next(iter, i); | |
980 | heap_sift(iter, i - top, btree_iter_cmp); | |
981 | continue; | |
982 | } | |
983 | ||
984 | if (top->k > i->k) { | |
985 | if (bkey_cmp(top->k, i->k) >= 0) | |
986 | sort_key_next(iter, i); | |
987 | else | |
988 | bch_cut_front(top->k, i->k); | |
989 | ||
990 | heap_sift(iter, i - top, btree_iter_cmp); | |
991 | } else { | |
992 | /* can't happen because of comparison func */ | |
993 | BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k))); | |
994 | bch_cut_back(&START_KEY(i->k), top->k); | |
995 | } | |
cafe5635 KO |
996 | } |
997 | } | |
998 | ||
999 | static void btree_mergesort(struct btree *b, struct bset *out, | |
1000 | struct btree_iter *iter, | |
1001 | bool fixup, bool remove_stale) | |
1002 | { | |
1003 | struct bkey *k, *last = NULL; | |
1004 | bool (*bad)(struct btree *, const struct bkey *) = remove_stale | |
1005 | ? bch_ptr_bad | |
1006 | : bch_ptr_invalid; | |
1007 | ||
1008 | while (!btree_iter_end(iter)) { | |
1009 | if (fixup && !b->level) | |
1010 | btree_sort_fixup(iter); | |
1011 | ||
1012 | k = bch_btree_iter_next(iter); | |
1013 | if (bad(b, k)) | |
1014 | continue; | |
1015 | ||
1016 | if (!last) { | |
1017 | last = out->start; | |
1018 | bkey_copy(last, k); | |
1019 | } else if (b->level || | |
1020 | !bch_bkey_try_merge(b, last, k)) { | |
1021 | last = bkey_next(last); | |
1022 | bkey_copy(last, k); | |
1023 | } | |
1024 | } | |
1025 | ||
1026 | out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0; | |
1027 | ||
1028 | pr_debug("sorted %i keys", out->keys); | |
cafe5635 KO |
1029 | } |
1030 | ||
1031 | static void __btree_sort(struct btree *b, struct btree_iter *iter, | |
1032 | unsigned start, unsigned order, bool fixup) | |
1033 | { | |
1034 | uint64_t start_time; | |
1035 | bool remove_stale = !b->written; | |
1036 | struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO, | |
1037 | order); | |
1038 | if (!out) { | |
1039 | mutex_lock(&b->c->sort_lock); | |
1040 | out = b->c->sort; | |
1041 | order = ilog2(bucket_pages(b->c)); | |
1042 | } | |
1043 | ||
1044 | start_time = local_clock(); | |
1045 | ||
1046 | btree_mergesort(b, out, iter, fixup, remove_stale); | |
1047 | b->nsets = start; | |
1048 | ||
1049 | if (!fixup && !start && b->written) | |
1050 | bch_btree_verify(b, out); | |
1051 | ||
1052 | if (!start && order == b->page_order) { | |
1053 | /* | |
1054 | * Our temporary buffer is the same size as the btree node's | |
1055 | * buffer, we can just swap buffers instead of doing a big | |
1056 | * memcpy() | |
1057 | */ | |
1058 | ||
81ab4190 | 1059 | out->magic = bset_magic(&b->c->sb); |
cafe5635 KO |
1060 | out->seq = b->sets[0].data->seq; |
1061 | out->version = b->sets[0].data->version; | |
1062 | swap(out, b->sets[0].data); | |
1063 | ||
1064 | if (b->c->sort == b->sets[0].data) | |
1065 | b->c->sort = out; | |
1066 | } else { | |
1067 | b->sets[start].data->keys = out->keys; | |
1068 | memcpy(b->sets[start].data->start, out->start, | |
1069 | (void *) end(out) - (void *) out->start); | |
1070 | } | |
1071 | ||
1072 | if (out == b->c->sort) | |
1073 | mutex_unlock(&b->c->sort_lock); | |
1074 | else | |
1075 | free_pages((unsigned long) out, order); | |
1076 | ||
1077 | if (b->written) | |
1078 | bset_build_written_tree(b); | |
1079 | ||
1080 | if (!start) { | |
1081 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1082 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1083 | spin_unlock(&b->c->sort_time_lock); |
1084 | } | |
1085 | } | |
1086 | ||
1087 | void bch_btree_sort_partial(struct btree *b, unsigned start) | |
1088 | { | |
280481d0 | 1089 | size_t order = b->page_order, keys = 0; |
cafe5635 | 1090 | struct btree_iter iter; |
280481d0 KO |
1091 | int oldsize = bch_count_data(b); |
1092 | ||
cafe5635 KO |
1093 | __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]); |
1094 | ||
1095 | BUG_ON(b->sets[b->nsets].data == write_block(b) && | |
1096 | (b->sets[b->nsets].size || b->nsets)); | |
1097 | ||
cafe5635 KO |
1098 | |
1099 | if (start) { | |
1100 | unsigned i; | |
1101 | ||
1102 | for (i = start; i <= b->nsets; i++) | |
1103 | keys += b->sets[i].data->keys; | |
1104 | ||
b1a67b0f KO |
1105 | order = roundup_pow_of_two(__set_bytes(b->sets->data, |
1106 | keys)) / PAGE_SIZE; | |
cafe5635 KO |
1107 | if (order) |
1108 | order = ilog2(order); | |
1109 | } | |
1110 | ||
1111 | __btree_sort(b, &iter, start, order, false); | |
1112 | ||
280481d0 | 1113 | EBUG_ON(b->written && oldsize >= 0 && bch_count_data(b) != oldsize); |
cafe5635 KO |
1114 | } |
1115 | ||
1116 | void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter) | |
1117 | { | |
1118 | BUG_ON(!b->written); | |
1119 | __btree_sort(b, iter, 0, b->page_order, true); | |
1120 | } | |
1121 | ||
1122 | void bch_btree_sort_into(struct btree *b, struct btree *new) | |
1123 | { | |
1124 | uint64_t start_time = local_clock(); | |
1125 | ||
1126 | struct btree_iter iter; | |
1127 | bch_btree_iter_init(b, &iter, NULL); | |
1128 | ||
1129 | btree_mergesort(b, new->sets->data, &iter, false, true); | |
1130 | ||
1131 | spin_lock(&b->c->sort_time_lock); | |
169ef1cf | 1132 | bch_time_stats_update(&b->c->sort_time, start_time); |
cafe5635 KO |
1133 | spin_unlock(&b->c->sort_time_lock); |
1134 | ||
1135 | bkey_copy_key(&new->key, &b->key); | |
1136 | new->sets->size = 0; | |
1137 | } | |
1138 | ||
6ded34d1 KO |
1139 | #define SORT_CRIT (4096 / sizeof(uint64_t)) |
1140 | ||
cafe5635 KO |
1141 | void bch_btree_sort_lazy(struct btree *b) |
1142 | { | |
6ded34d1 KO |
1143 | unsigned crit = SORT_CRIT; |
1144 | int i; | |
cafe5635 | 1145 | |
6ded34d1 KO |
1146 | /* Don't sort if nothing to do */ |
1147 | if (!b->nsets) | |
1148 | goto out; | |
cafe5635 | 1149 | |
6ded34d1 KO |
1150 | /* If not a leaf node, always sort */ |
1151 | if (b->level) { | |
1152 | bch_btree_sort(b); | |
1153 | return; | |
1154 | } | |
cafe5635 | 1155 | |
6ded34d1 KO |
1156 | for (i = b->nsets - 1; i >= 0; --i) { |
1157 | crit *= b->c->sort_crit_factor; | |
cafe5635 | 1158 | |
6ded34d1 KO |
1159 | if (b->sets[i].data->keys < crit) { |
1160 | bch_btree_sort_partial(b, i); | |
cafe5635 KO |
1161 | return; |
1162 | } | |
1163 | } | |
1164 | ||
6ded34d1 KO |
1165 | /* Sort if we'd overflow */ |
1166 | if (b->nsets + 1 == MAX_BSETS) { | |
1167 | bch_btree_sort(b); | |
1168 | return; | |
1169 | } | |
1170 | ||
1171 | out: | |
cafe5635 KO |
1172 | bset_build_written_tree(b); |
1173 | } | |
1174 | ||
1175 | /* Sysfs stuff */ | |
1176 | ||
1177 | struct bset_stats { | |
48dad8ba | 1178 | struct btree_op op; |
cafe5635 KO |
1179 | size_t nodes; |
1180 | size_t sets_written, sets_unwritten; | |
1181 | size_t bytes_written, bytes_unwritten; | |
1182 | size_t floats, failed; | |
1183 | }; | |
1184 | ||
48dad8ba | 1185 | static int btree_bset_stats(struct btree_op *op, struct btree *b) |
cafe5635 | 1186 | { |
48dad8ba | 1187 | struct bset_stats *stats = container_of(op, struct bset_stats, op); |
cafe5635 KO |
1188 | unsigned i; |
1189 | ||
1190 | stats->nodes++; | |
1191 | ||
1192 | for (i = 0; i <= b->nsets; i++) { | |
1193 | struct bset_tree *t = &b->sets[i]; | |
1194 | size_t bytes = t->data->keys * sizeof(uint64_t); | |
1195 | size_t j; | |
1196 | ||
1197 | if (bset_written(b, t)) { | |
1198 | stats->sets_written++; | |
1199 | stats->bytes_written += bytes; | |
1200 | ||
1201 | stats->floats += t->size - 1; | |
1202 | ||
1203 | for (j = 1; j < t->size; j++) | |
1204 | if (t->tree[j].exponent == 127) | |
1205 | stats->failed++; | |
1206 | } else { | |
1207 | stats->sets_unwritten++; | |
1208 | stats->bytes_unwritten += bytes; | |
1209 | } | |
1210 | } | |
1211 | ||
48dad8ba | 1212 | return MAP_CONTINUE; |
cafe5635 KO |
1213 | } |
1214 | ||
1215 | int bch_bset_print_stats(struct cache_set *c, char *buf) | |
1216 | { | |
cafe5635 KO |
1217 | struct bset_stats t; |
1218 | int ret; | |
1219 | ||
cafe5635 | 1220 | memset(&t, 0, sizeof(struct bset_stats)); |
b54d6934 | 1221 | bch_btree_op_init(&t.op, -1); |
cafe5635 | 1222 | |
48dad8ba KO |
1223 | ret = bch_btree_map_nodes(&t.op, c, &ZERO_KEY, btree_bset_stats); |
1224 | if (ret < 0) | |
cafe5635 KO |
1225 | return ret; |
1226 | ||
1227 | return snprintf(buf, PAGE_SIZE, | |
1228 | "btree nodes: %zu\n" | |
1229 | "written sets: %zu\n" | |
1230 | "unwritten sets: %zu\n" | |
1231 | "written key bytes: %zu\n" | |
1232 | "unwritten key bytes: %zu\n" | |
1233 | "floats: %zu\n" | |
1234 | "failed: %zu\n", | |
1235 | t.nodes, | |
1236 | t.sets_written, t.sets_unwritten, | |
1237 | t.bytes_written, t.bytes_unwritten, | |
1238 | t.floats, t.failed); | |
1239 | } |