]> git.ipfire.org Git - thirdparty/kernel/linux.git/blame - drivers/md/bcache/bset.c
License cleanup: add SPDX GPL-2.0 license identifier to files with no license
[thirdparty/kernel/linux.git] / drivers / md / bcache / bset.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
cafe5635
KO
2/*
3 * Code for working with individual keys, and sorted sets of keys with in a
4 * btree node
5 *
6 * Copyright 2012 Google, Inc.
7 */
8
89ebb4a2
KO
9#define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__
10
11#include "util.h"
12#include "bset.h"
cafe5635 13
dc9d98d6 14#include <linux/console.h>
e6017571 15#include <linux/sched/clock.h>
cafe5635 16#include <linux/random.h>
cd953ed0 17#include <linux/prefetch.h>
cafe5635 18
dc9d98d6
KO
19#ifdef CONFIG_BCACHE_DEBUG
20
21void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned set)
22{
23 struct bkey *k, *next;
24
25 for (k = i->start; k < bset_bkey_last(i); k = next) {
26 next = bkey_next(k);
27
85cbe1f8
KO
28 printk(KERN_ERR "block %u key %u/%u: ", set,
29 (unsigned) ((u64 *) k - i->d), i->keys);
dc9d98d6
KO
30
31 if (b->ops->key_dump)
32 b->ops->key_dump(b, k);
33 else
34 printk("%llu:%llu\n", KEY_INODE(k), KEY_OFFSET(k));
35
36 if (next < bset_bkey_last(i) &&
37 bkey_cmp(k, b->ops->is_extents ?
38 &START_KEY(next) : next) > 0)
39 printk(KERN_ERR "Key skipped backwards\n");
40 }
41}
42
43void bch_dump_bucket(struct btree_keys *b)
44{
45 unsigned i;
46
47 console_lock();
48 for (i = 0; i <= b->nsets; i++)
49 bch_dump_bset(b, b->set[i].data,
50 bset_sector_offset(b, b->set[i].data));
51 console_unlock();
52}
53
54int __bch_count_data(struct btree_keys *b)
55{
56 unsigned ret = 0;
57 struct btree_iter iter;
58 struct bkey *k;
59
60 if (b->ops->is_extents)
61 for_each_key(b, k, &iter)
62 ret += KEY_SIZE(k);
63 return ret;
64}
65
66void __bch_check_keys(struct btree_keys *b, const char *fmt, ...)
67{
68 va_list args;
69 struct bkey *k, *p = NULL;
70 struct btree_iter iter;
71 const char *err;
72
73 for_each_key(b, k, &iter) {
74 if (b->ops->is_extents) {
75 err = "Keys out of order";
76 if (p && bkey_cmp(&START_KEY(p), &START_KEY(k)) > 0)
77 goto bug;
78
79 if (bch_ptr_invalid(b, k))
80 continue;
81
82 err = "Overlapping keys";
83 if (p && bkey_cmp(p, &START_KEY(k)) > 0)
84 goto bug;
85 } else {
86 if (bch_ptr_bad(b, k))
87 continue;
88
89 err = "Duplicate keys";
90 if (p && !bkey_cmp(p, k))
91 goto bug;
92 }
93 p = k;
94 }
95#if 0
96 err = "Key larger than btree node key";
97 if (p && bkey_cmp(p, &b->key) > 0)
98 goto bug;
99#endif
100 return;
101bug:
102 bch_dump_bucket(b);
103
104 va_start(args, fmt);
105 vprintk(fmt, args);
106 va_end(args);
107
108 panic("bch_check_keys error: %s:\n", err);
109}
110
111static void bch_btree_iter_next_check(struct btree_iter *iter)
112{
113 struct bkey *k = iter->data->k, *next = bkey_next(k);
114
115 if (next < iter->data->end &&
116 bkey_cmp(k, iter->b->ops->is_extents ?
117 &START_KEY(next) : next) > 0) {
118 bch_dump_bucket(iter->b);
119 panic("Key skipped backwards\n");
120 }
121}
122
123#else
124
125static inline void bch_btree_iter_next_check(struct btree_iter *iter) {}
126
127#endif
128
cafe5635
KO
129/* Keylists */
130
085d2a3d 131int __bch_keylist_realloc(struct keylist *l, unsigned u64s)
cafe5635 132{
c2f95ae2 133 size_t oldsize = bch_keylist_nkeys(l);
085d2a3d 134 size_t newsize = oldsize + u64s;
c2f95ae2
KO
135 uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
136 uint64_t *new_keys;
cafe5635 137
cafe5635
KO
138 newsize = roundup_pow_of_two(newsize);
139
140 if (newsize <= KEYLIST_INLINE ||
141 roundup_pow_of_two(oldsize) == newsize)
142 return 0;
143
c2f95ae2 144 new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
cafe5635 145
c2f95ae2 146 if (!new_keys)
cafe5635
KO
147 return -ENOMEM;
148
c2f95ae2
KO
149 if (!old_keys)
150 memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
cafe5635 151
c2f95ae2
KO
152 l->keys_p = new_keys;
153 l->top_p = new_keys + oldsize;
cafe5635
KO
154
155 return 0;
156}
157
158struct bkey *bch_keylist_pop(struct keylist *l)
159{
c2f95ae2 160 struct bkey *k = l->keys;
cafe5635
KO
161
162 if (k == l->top)
163 return NULL;
164
165 while (bkey_next(k) != l->top)
166 k = bkey_next(k);
167
168 return l->top = k;
169}
170
26c949f8
KO
171void bch_keylist_pop_front(struct keylist *l)
172{
c2f95ae2 173 l->top_p -= bkey_u64s(l->keys);
26c949f8 174
c2f95ae2
KO
175 memmove(l->keys,
176 bkey_next(l->keys),
177 bch_keylist_bytes(l));
26c949f8
KO
178}
179
cafe5635
KO
180/* Key/pointer manipulation */
181
182void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
183 unsigned i)
184{
185 BUG_ON(i > KEY_PTRS(src));
186
187 /* Only copy the header, key, and one pointer. */
188 memcpy(dest, src, 2 * sizeof(uint64_t));
189 dest->ptr[0] = src->ptr[i];
190 SET_KEY_PTRS(dest, 1);
191 /* We didn't copy the checksum so clear that bit. */
192 SET_KEY_CSUM(dest, 0);
193}
194
195bool __bch_cut_front(const struct bkey *where, struct bkey *k)
196{
197 unsigned i, len = 0;
198
199 if (bkey_cmp(where, &START_KEY(k)) <= 0)
200 return false;
201
202 if (bkey_cmp(where, k) < 0)
203 len = KEY_OFFSET(k) - KEY_OFFSET(where);
204 else
205 bkey_copy_key(k, where);
206
207 for (i = 0; i < KEY_PTRS(k); i++)
208 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
209
210 BUG_ON(len > KEY_SIZE(k));
211 SET_KEY_SIZE(k, len);
212 return true;
213}
214
215bool __bch_cut_back(const struct bkey *where, struct bkey *k)
216{
217 unsigned len = 0;
218
219 if (bkey_cmp(where, k) >= 0)
220 return false;
221
222 BUG_ON(KEY_INODE(where) != KEY_INODE(k));
223
224 if (bkey_cmp(where, &START_KEY(k)) > 0)
225 len = KEY_OFFSET(where) - KEY_START(k);
226
227 bkey_copy_key(k, where);
228
229 BUG_ON(len > KEY_SIZE(k));
230 SET_KEY_SIZE(k, len);
231 return true;
232}
233
ee811287
KO
234/* Auxiliary search trees */
235
236/* 32 bits total: */
237#define BKEY_MID_BITS 3
238#define BKEY_EXPONENT_BITS 7
239#define BKEY_MANTISSA_BITS (32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS)
240#define BKEY_MANTISSA_MASK ((1 << BKEY_MANTISSA_BITS) - 1)
241
242struct bkey_float {
243 unsigned exponent:BKEY_EXPONENT_BITS;
244 unsigned m:BKEY_MID_BITS;
245 unsigned mantissa:BKEY_MANTISSA_BITS;
246} __packed;
247
248/*
249 * BSET_CACHELINE was originally intended to match the hardware cacheline size -
250 * it used to be 64, but I realized the lookup code would touch slightly less
251 * memory if it was 128.
252 *
253 * It definites the number of bytes (in struct bset) per struct bkey_float in
254 * the auxiliar search tree - when we're done searching the bset_float tree we
255 * have this many bytes left that we do a linear search over.
256 *
257 * Since (after level 5) every level of the bset_tree is on a new cacheline,
258 * we're touching one fewer cacheline in the bset tree in exchange for one more
259 * cacheline in the linear search - but the linear search might stop before it
260 * gets to the second cacheline.
261 */
262
263#define BSET_CACHELINE 128
264
265/* Space required for the btree node keys */
a85e968e 266static inline size_t btree_keys_bytes(struct btree_keys *b)
ee811287
KO
267{
268 return PAGE_SIZE << b->page_order;
269}
270
a85e968e 271static inline size_t btree_keys_cachelines(struct btree_keys *b)
ee811287
KO
272{
273 return btree_keys_bytes(b) / BSET_CACHELINE;
274}
275
276/* Space required for the auxiliary search trees */
a85e968e 277static inline size_t bset_tree_bytes(struct btree_keys *b)
ee811287
KO
278{
279 return btree_keys_cachelines(b) * sizeof(struct bkey_float);
280}
281
282/* Space required for the prev pointers */
a85e968e 283static inline size_t bset_prev_bytes(struct btree_keys *b)
ee811287
KO
284{
285 return btree_keys_cachelines(b) * sizeof(uint8_t);
286}
287
288/* Memory allocation */
289
a85e968e 290void bch_btree_keys_free(struct btree_keys *b)
ee811287 291{
a85e968e 292 struct bset_tree *t = b->set;
ee811287
KO
293
294 if (bset_prev_bytes(b) < PAGE_SIZE)
295 kfree(t->prev);
296 else
297 free_pages((unsigned long) t->prev,
298 get_order(bset_prev_bytes(b)));
299
300 if (bset_tree_bytes(b) < PAGE_SIZE)
301 kfree(t->tree);
302 else
303 free_pages((unsigned long) t->tree,
304 get_order(bset_tree_bytes(b)));
305
306 free_pages((unsigned long) t->data, b->page_order);
307
308 t->prev = NULL;
309 t->tree = NULL;
310 t->data = NULL;
311}
a85e968e 312EXPORT_SYMBOL(bch_btree_keys_free);
ee811287 313
a85e968e 314int bch_btree_keys_alloc(struct btree_keys *b, unsigned page_order, gfp_t gfp)
ee811287 315{
a85e968e 316 struct bset_tree *t = b->set;
ee811287
KO
317
318 BUG_ON(t->data);
319
320 b->page_order = page_order;
321
322 t->data = (void *) __get_free_pages(gfp, b->page_order);
323 if (!t->data)
324 goto err;
325
326 t->tree = bset_tree_bytes(b) < PAGE_SIZE
327 ? kmalloc(bset_tree_bytes(b), gfp)
328 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
329 if (!t->tree)
330 goto err;
331
332 t->prev = bset_prev_bytes(b) < PAGE_SIZE
333 ? kmalloc(bset_prev_bytes(b), gfp)
334 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
335 if (!t->prev)
336 goto err;
337
338 return 0;
339err:
340 bch_btree_keys_free(b);
341 return -ENOMEM;
342}
a85e968e
KO
343EXPORT_SYMBOL(bch_btree_keys_alloc);
344
345void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
346 bool *expensive_debug_checks)
347{
348 unsigned i;
349
350 b->ops = ops;
351 b->expensive_debug_checks = expensive_debug_checks;
352 b->nsets = 0;
353 b->last_set_unwritten = 0;
354
355 /* XXX: shouldn't be needed */
356 for (i = 0; i < MAX_BSETS; i++)
357 b->set[i].size = 0;
358 /*
359 * Second loop starts at 1 because b->keys[0]->data is the memory we
360 * allocated
361 */
362 for (i = 1; i < MAX_BSETS; i++)
363 b->set[i].data = NULL;
364}
365EXPORT_SYMBOL(bch_btree_keys_init);
ee811287 366
cafe5635
KO
367/* Binary tree stuff for auxiliary search trees */
368
369static unsigned inorder_next(unsigned j, unsigned size)
370{
371 if (j * 2 + 1 < size) {
372 j = j * 2 + 1;
373
374 while (j * 2 < size)
375 j *= 2;
376 } else
377 j >>= ffz(j) + 1;
378
379 return j;
380}
381
382static unsigned inorder_prev(unsigned j, unsigned size)
383{
384 if (j * 2 < size) {
385 j = j * 2;
386
387 while (j * 2 + 1 < size)
388 j = j * 2 + 1;
389 } else
390 j >>= ffs(j);
391
392 return j;
393}
394
395/* I have no idea why this code works... and I'm the one who wrote it
396 *
397 * However, I do know what it does:
398 * Given a binary tree constructed in an array (i.e. how you normally implement
399 * a heap), it converts a node in the tree - referenced by array index - to the
400 * index it would have if you did an inorder traversal.
401 *
402 * Also tested for every j, size up to size somewhere around 6 million.
403 *
404 * The binary tree starts at array index 1, not 0
405 * extra is a function of size:
406 * extra = (size - rounddown_pow_of_two(size - 1)) << 1;
407 */
408static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
409{
410 unsigned b = fls(j);
411 unsigned shift = fls(size - 1) - b;
412
413 j ^= 1U << (b - 1);
414 j <<= 1;
415 j |= 1;
416 j <<= shift;
417
418 if (j > extra)
419 j -= (j - extra) >> 1;
420
421 return j;
422}
423
424static unsigned to_inorder(unsigned j, struct bset_tree *t)
425{
426 return __to_inorder(j, t->size, t->extra);
427}
428
429static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
430{
431 unsigned shift;
432
433 if (j > extra)
434 j += j - extra;
435
436 shift = ffs(j);
437
438 j >>= shift;
439 j |= roundup_pow_of_two(size) >> shift;
440
441 return j;
442}
443
444static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
445{
446 return __inorder_to_tree(j, t->size, t->extra);
447}
448
449#if 0
450void inorder_test(void)
451{
452 unsigned long done = 0;
453 ktime_t start = ktime_get();
454
455 for (unsigned size = 2;
456 size < 65536000;
457 size++) {
458 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
459 unsigned i = 1, j = rounddown_pow_of_two(size - 1);
460
461 if (!(size % 4096))
462 printk(KERN_NOTICE "loop %u, %llu per us\n", size,
463 done / ktime_us_delta(ktime_get(), start));
464
465 while (1) {
466 if (__inorder_to_tree(i, size, extra) != j)
467 panic("size %10u j %10u i %10u", size, j, i);
468
469 if (__to_inorder(j, size, extra) != i)
470 panic("size %10u j %10u i %10u", size, j, i);
471
472 if (j == rounddown_pow_of_two(size) - 1)
473 break;
474
475 BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
476
477 j = inorder_next(j, size);
478 i++;
479 }
480
481 done += size - 1;
482 }
483}
484#endif
485
486/*
48a73025 487 * Cacheline/offset <-> bkey pointer arithmetic:
cafe5635
KO
488 *
489 * t->tree is a binary search tree in an array; each node corresponds to a key
490 * in one cacheline in t->set (BSET_CACHELINE bytes).
491 *
492 * This means we don't have to store the full index of the key that a node in
493 * the binary tree points to; to_inorder() gives us the cacheline, and then
494 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
495 *
48a73025 496 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
cafe5635
KO
497 * make this work.
498 *
499 * To construct the bfloat for an arbitrary key we need to know what the key
500 * immediately preceding it is: we have to check if the two keys differ in the
501 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
502 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
503 */
504
505static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
506 unsigned offset)
507{
508 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
509}
510
511static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
512{
513 return ((void *) k - (void *) t->data) / BSET_CACHELINE;
514}
515
9dd6358a
KO
516static unsigned bkey_to_cacheline_offset(struct bset_tree *t,
517 unsigned cacheline,
518 struct bkey *k)
cafe5635 519{
9dd6358a 520 return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0);
cafe5635
KO
521}
522
523static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
524{
525 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
526}
527
528static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
529{
530 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
531}
532
533/*
534 * For the write set - the one we're currently inserting keys into - we don't
535 * maintain a full search tree, we just keep a simple lookup table in t->prev.
536 */
537static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
538{
539 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
540}
541
542static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
543{
cafe5635
KO
544 low >>= shift;
545 low |= (high << 1) << (63U - shift);
cafe5635
KO
546 return low;
547}
548
549static inline unsigned bfloat_mantissa(const struct bkey *k,
550 struct bkey_float *f)
551{
552 const uint64_t *p = &k->low - (f->exponent >> 6);
553 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
554}
555
556static void make_bfloat(struct bset_tree *t, unsigned j)
557{
558 struct bkey_float *f = &t->tree[j];
559 struct bkey *m = tree_to_bkey(t, j);
560 struct bkey *p = tree_to_prev_bkey(t, j);
561
562 struct bkey *l = is_power_of_2(j)
563 ? t->data->start
564 : tree_to_prev_bkey(t, j >> ffs(j));
565
566 struct bkey *r = is_power_of_2(j + 1)
fafff81c 567 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
cafe5635
KO
568 : tree_to_bkey(t, j >> (ffz(j) + 1));
569
570 BUG_ON(m < l || m > r);
571 BUG_ON(bkey_next(p) != m);
572
573 if (KEY_INODE(l) != KEY_INODE(r))
574 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
575 else
576 f->exponent = fls64(r->low ^ l->low);
577
578 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
579
580 /*
581 * Setting f->exponent = 127 flags this node as failed, and causes the
582 * lookup code to fall back to comparing against the original key.
583 */
584
585 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
586 f->mantissa = bfloat_mantissa(m, f) - 1;
587 else
588 f->exponent = 127;
589}
590
a85e968e 591static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t)
cafe5635 592{
a85e968e 593 if (t != b->set) {
cafe5635
KO
594 unsigned j = roundup(t[-1].size,
595 64 / sizeof(struct bkey_float));
596
597 t->tree = t[-1].tree + j;
598 t->prev = t[-1].prev + j;
599 }
600
a85e968e 601 while (t < b->set + MAX_BSETS)
cafe5635
KO
602 t++->size = 0;
603}
604
a85e968e 605static void bch_bset_build_unwritten_tree(struct btree_keys *b)
cafe5635 606{
ee811287 607 struct bset_tree *t = bset_tree_last(b);
cafe5635 608
a85e968e
KO
609 BUG_ON(b->last_set_unwritten);
610 b->last_set_unwritten = 1;
611
cafe5635
KO
612 bset_alloc_tree(b, t);
613
a85e968e 614 if (t->tree != b->set->tree + btree_keys_cachelines(b)) {
9dd6358a 615 t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start);
cafe5635
KO
616 t->size = 1;
617 }
618}
619
a85e968e 620void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic)
ee811287 621{
a85e968e
KO
622 if (i != b->set->data) {
623 b->set[++b->nsets].data = i;
624 i->seq = b->set->data->seq;
ee811287
KO
625 } else
626 get_random_bytes(&i->seq, sizeof(uint64_t));
627
628 i->magic = magic;
629 i->version = 0;
630 i->keys = 0;
631
632 bch_bset_build_unwritten_tree(b);
633}
a85e968e 634EXPORT_SYMBOL(bch_bset_init_next);
ee811287 635
a85e968e 636void bch_bset_build_written_tree(struct btree_keys *b)
cafe5635 637{
ee811287 638 struct bset_tree *t = bset_tree_last(b);
9dd6358a 639 struct bkey *prev = NULL, *k = t->data->start;
cafe5635
KO
640 unsigned j, cacheline = 1;
641
a85e968e
KO
642 b->last_set_unwritten = 0;
643
cafe5635
KO
644 bset_alloc_tree(b, t);
645
646 t->size = min_t(unsigned,
fafff81c 647 bkey_to_cacheline(t, bset_bkey_last(t->data)),
a85e968e 648 b->set->tree + btree_keys_cachelines(b) - t->tree);
cafe5635
KO
649
650 if (t->size < 2) {
651 t->size = 0;
652 return;
653 }
654
655 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
656
657 /* First we figure out where the first key in each cacheline is */
658 for (j = inorder_next(0, t->size);
659 j;
660 j = inorder_next(j, t->size)) {
9dd6358a
KO
661 while (bkey_to_cacheline(t, k) < cacheline)
662 prev = k, k = bkey_next(k);
cafe5635 663
9dd6358a
KO
664 t->prev[j] = bkey_u64s(prev);
665 t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k);
cafe5635
KO
666 }
667
fafff81c 668 while (bkey_next(k) != bset_bkey_last(t->data))
cafe5635
KO
669 k = bkey_next(k);
670
671 t->end = *k;
672
673 /* Then we build the tree */
674 for (j = inorder_next(0, t->size);
675 j;
676 j = inorder_next(j, t->size))
677 make_bfloat(t, j);
678}
a85e968e 679EXPORT_SYMBOL(bch_bset_build_written_tree);
cafe5635 680
829a60b9
KO
681/* Insert */
682
a85e968e 683void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k)
cafe5635
KO
684{
685 struct bset_tree *t;
686 unsigned inorder, j = 1;
687
a85e968e 688 for (t = b->set; t <= bset_tree_last(b); t++)
fafff81c 689 if (k < bset_bkey_last(t->data))
cafe5635
KO
690 goto found_set;
691
692 BUG();
693found_set:
694 if (!t->size || !bset_written(b, t))
695 return;
696
697 inorder = bkey_to_cacheline(t, k);
698
699 if (k == t->data->start)
700 goto fix_left;
701
fafff81c 702 if (bkey_next(k) == bset_bkey_last(t->data)) {
cafe5635
KO
703 t->end = *k;
704 goto fix_right;
705 }
706
707 j = inorder_to_tree(inorder, t);
708
709 if (j &&
710 j < t->size &&
711 k == tree_to_bkey(t, j))
712fix_left: do {
713 make_bfloat(t, j);
714 j = j * 2;
715 } while (j < t->size);
716
717 j = inorder_to_tree(inorder + 1, t);
718
719 if (j &&
720 j < t->size &&
721 k == tree_to_prev_bkey(t, j))
722fix_right: do {
723 make_bfloat(t, j);
724 j = j * 2 + 1;
725 } while (j < t->size);
726}
a85e968e 727EXPORT_SYMBOL(bch_bset_fix_invalidated_key);
cafe5635 728
a85e968e 729static void bch_bset_fix_lookup_table(struct btree_keys *b,
ee811287
KO
730 struct bset_tree *t,
731 struct bkey *k)
cafe5635 732{
cafe5635
KO
733 unsigned shift = bkey_u64s(k);
734 unsigned j = bkey_to_cacheline(t, k);
735
736 /* We're getting called from btree_split() or btree_gc, just bail out */
737 if (!t->size)
738 return;
739
740 /* k is the key we just inserted; we need to find the entry in the
741 * lookup table for the first key that is strictly greater than k:
742 * it's either k's cacheline or the next one
743 */
9dd6358a
KO
744 while (j < t->size &&
745 table_to_bkey(t, j) <= k)
cafe5635
KO
746 j++;
747
748 /* Adjust all the lookup table entries, and find a new key for any that
749 * have gotten too big
750 */
751 for (; j < t->size; j++) {
752 t->prev[j] += shift;
753
754 if (t->prev[j] > 7) {
755 k = table_to_bkey(t, j - 1);
756
757 while (k < cacheline_to_bkey(t, j, 0))
758 k = bkey_next(k);
759
9dd6358a 760 t->prev[j] = bkey_to_cacheline_offset(t, j, k);
cafe5635
KO
761 }
762 }
763
a85e968e 764 if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree)
cafe5635
KO
765 return;
766
767 /* Possibly add a new entry to the end of the lookup table */
768
769 for (k = table_to_bkey(t, t->size - 1);
fafff81c 770 k != bset_bkey_last(t->data);
cafe5635
KO
771 k = bkey_next(k))
772 if (t->size == bkey_to_cacheline(t, k)) {
9dd6358a 773 t->prev[t->size] = bkey_to_cacheline_offset(t, t->size, k);
cafe5635
KO
774 t->size++;
775 }
776}
777
0f49cf3d
NS
778/*
779 * Tries to merge l and r: l should be lower than r
780 * Returns true if we were able to merge. If we did merge, l will be the merged
781 * key, r will be untouched.
782 */
783bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r)
784{
785 if (!b->ops->key_merge)
786 return false;
787
788 /*
789 * Generic header checks
790 * Assumes left and right are in order
791 * Left and right must be exactly aligned
792 */
3bdad1e4
NS
793 if (!bch_bkey_equal_header(l, r) ||
794 bkey_cmp(l, &START_KEY(r)))
0f49cf3d
NS
795 return false;
796
797 return b->ops->key_merge(b, l, r);
798}
799EXPORT_SYMBOL(bch_bkey_try_merge);
800
a85e968e 801void bch_bset_insert(struct btree_keys *b, struct bkey *where,
ee811287 802 struct bkey *insert)
cafe5635 803{
ee811287 804 struct bset_tree *t = bset_tree_last(b);
cafe5635 805
a85e968e 806 BUG_ON(!b->last_set_unwritten);
ee811287
KO
807 BUG_ON(bset_byte_offset(b, t->data) +
808 __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
809 PAGE_SIZE << b->page_order);
cafe5635 810
ee811287
KO
811 memmove((uint64_t *) where + bkey_u64s(insert),
812 where,
813 (void *) bset_bkey_last(t->data) - (void *) where);
cafe5635 814
ee811287
KO
815 t->data->keys += bkey_u64s(insert);
816 bkey_copy(where, insert);
817 bch_bset_fix_lookup_table(b, t, where);
cafe5635 818}
a85e968e 819EXPORT_SYMBOL(bch_bset_insert);
cafe5635 820
829a60b9
KO
821unsigned bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
822 struct bkey *replace_key)
823{
824 unsigned status = BTREE_INSERT_STATUS_NO_INSERT;
825 struct bset *i = bset_tree_last(b)->data;
826 struct bkey *m, *prev = NULL;
827 struct btree_iter iter;
828
829 BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
830
831 m = bch_btree_iter_init(b, &iter, b->ops->is_extents
832 ? PRECEDING_KEY(&START_KEY(k))
833 : PRECEDING_KEY(k));
834
835 if (b->ops->insert_fixup(b, k, &iter, replace_key))
836 return status;
837
838 status = BTREE_INSERT_STATUS_INSERT;
839
840 while (m != bset_bkey_last(i) &&
841 bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0)
842 prev = m, m = bkey_next(m);
843
844 /* prev is in the tree, if we merge we're done */
845 status = BTREE_INSERT_STATUS_BACK_MERGE;
846 if (prev &&
847 bch_bkey_try_merge(b, prev, k))
848 goto merged;
849#if 0
850 status = BTREE_INSERT_STATUS_OVERWROTE;
851 if (m != bset_bkey_last(i) &&
852 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
853 goto copy;
854#endif
855 status = BTREE_INSERT_STATUS_FRONT_MERGE;
856 if (m != bset_bkey_last(i) &&
857 bch_bkey_try_merge(b, k, m))
858 goto copy;
859
860 bch_bset_insert(b, m, k);
861copy: bkey_copy(m, k);
862merged:
863 return status;
864}
865EXPORT_SYMBOL(bch_btree_insert_key);
866
867/* Lookup */
868
cafe5635
KO
869struct bset_search_iter {
870 struct bkey *l, *r;
871};
872
a85e968e 873static struct bset_search_iter bset_search_write_set(struct bset_tree *t,
cafe5635
KO
874 const struct bkey *search)
875{
876 unsigned li = 0, ri = t->size;
877
cafe5635
KO
878 while (li + 1 != ri) {
879 unsigned m = (li + ri) >> 1;
880
881 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
882 ri = m;
883 else
884 li = m;
885 }
886
887 return (struct bset_search_iter) {
888 table_to_bkey(t, li),
fafff81c 889 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
cafe5635
KO
890 };
891}
892
a85e968e 893static struct bset_search_iter bset_search_tree(struct bset_tree *t,
cafe5635
KO
894 const struct bkey *search)
895{
896 struct bkey *l, *r;
897 struct bkey_float *f;
898 unsigned inorder, j, n = 1;
899
900 do {
901 unsigned p = n << 4;
902 p &= ((int) (p - t->size)) >> 31;
903
904 prefetch(&t->tree[p]);
905
906 j = n;
907 f = &t->tree[j];
908
909 /*
910 * n = (f->mantissa > bfloat_mantissa())
911 * ? j * 2
912 * : j * 2 + 1;
913 *
914 * We need to subtract 1 from f->mantissa for the sign bit trick
915 * to work - that's done in make_bfloat()
916 */
917 if (likely(f->exponent != 127))
918 n = j * 2 + (((unsigned)
919 (f->mantissa -
920 bfloat_mantissa(search, f))) >> 31);
921 else
922 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
923 ? j * 2
924 : j * 2 + 1;
925 } while (n < t->size);
926
927 inorder = to_inorder(j, t);
928
929 /*
930 * n would have been the node we recursed to - the low bit tells us if
931 * we recursed left or recursed right.
932 */
933 if (n & 1) {
934 l = cacheline_to_bkey(t, inorder, f->m);
935
936 if (++inorder != t->size) {
937 f = &t->tree[inorder_next(j, t->size)];
938 r = cacheline_to_bkey(t, inorder, f->m);
939 } else
fafff81c 940 r = bset_bkey_last(t->data);
cafe5635
KO
941 } else {
942 r = cacheline_to_bkey(t, inorder, f->m);
943
944 if (--inorder) {
945 f = &t->tree[inorder_prev(j, t->size)];
946 l = cacheline_to_bkey(t, inorder, f->m);
947 } else
948 l = t->data->start;
949 }
950
951 return (struct bset_search_iter) {l, r};
952}
953
c052dd9a 954struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
cafe5635
KO
955 const struct bkey *search)
956{
957 struct bset_search_iter i;
958
959 /*
960 * First, we search for a cacheline, then lastly we do a linear search
961 * within that cacheline.
962 *
963 * To search for the cacheline, there's three different possibilities:
964 * * The set is too small to have a search tree, so we just do a linear
965 * search over the whole set.
966 * * The set is the one we're currently inserting into; keeping a full
967 * auxiliary search tree up to date would be too expensive, so we
968 * use a much simpler lookup table to do a binary search -
969 * bset_search_write_set().
970 * * Or we use the auxiliary search tree we constructed earlier -
971 * bset_search_tree()
972 */
973
974 if (unlikely(!t->size)) {
975 i.l = t->data->start;
fafff81c 976 i.r = bset_bkey_last(t->data);
c052dd9a 977 } else if (bset_written(b, t)) {
cafe5635
KO
978 /*
979 * Each node in the auxiliary search tree covers a certain range
980 * of bits, and keys above and below the set it covers might
981 * differ outside those bits - so we have to special case the
982 * start and end - handle that here:
983 */
984
985 if (unlikely(bkey_cmp(search, &t->end) >= 0))
fafff81c 986 return bset_bkey_last(t->data);
cafe5635
KO
987
988 if (unlikely(bkey_cmp(search, t->data->start) < 0))
989 return t->data->start;
990
a85e968e
KO
991 i = bset_search_tree(t, search);
992 } else {
c052dd9a 993 BUG_ON(!b->nsets &&
a85e968e
KO
994 t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
995
996 i = bset_search_write_set(t, search);
997 }
cafe5635 998
c052dd9a
KO
999 if (btree_keys_expensive_checks(b)) {
1000 BUG_ON(bset_written(b, t) &&
280481d0
KO
1001 i.l != t->data->start &&
1002 bkey_cmp(tree_to_prev_bkey(t,
1003 inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
1004 search) > 0);
cafe5635 1005
fafff81c 1006 BUG_ON(i.r != bset_bkey_last(t->data) &&
280481d0
KO
1007 bkey_cmp(i.r, search) <= 0);
1008 }
cafe5635
KO
1009
1010 while (likely(i.l != i.r) &&
1011 bkey_cmp(i.l, search) <= 0)
1012 i.l = bkey_next(i.l);
1013
1014 return i.l;
1015}
a85e968e 1016EXPORT_SYMBOL(__bch_bset_search);
cafe5635
KO
1017
1018/* Btree iterator */
1019
911c9610
KO
1020typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
1021 struct btree_iter_set);
1022
cafe5635
KO
1023static inline bool btree_iter_cmp(struct btree_iter_set l,
1024 struct btree_iter_set r)
1025{
911c9610 1026 return bkey_cmp(l.k, r.k) > 0;
cafe5635
KO
1027}
1028
1029static inline bool btree_iter_end(struct btree_iter *iter)
1030{
1031 return !iter->used;
1032}
1033
1034void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
1035 struct bkey *end)
1036{
1037 if (k != end)
1038 BUG_ON(!heap_add(iter,
1039 ((struct btree_iter_set) { k, end }),
1040 btree_iter_cmp));
1041}
1042
c052dd9a 1043static struct bkey *__bch_btree_iter_init(struct btree_keys *b,
911c9610
KO
1044 struct btree_iter *iter,
1045 struct bkey *search,
1046 struct bset_tree *start)
cafe5635
KO
1047{
1048 struct bkey *ret = NULL;
1049 iter->size = ARRAY_SIZE(iter->data);
1050 iter->used = 0;
1051
280481d0
KO
1052#ifdef CONFIG_BCACHE_DEBUG
1053 iter->b = b;
1054#endif
1055
c052dd9a 1056 for (; start <= bset_tree_last(b); start++) {
cafe5635 1057 ret = bch_bset_search(b, start, search);
fafff81c 1058 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
cafe5635
KO
1059 }
1060
1061 return ret;
1062}
1063
c052dd9a 1064struct bkey *bch_btree_iter_init(struct btree_keys *b,
911c9610
KO
1065 struct btree_iter *iter,
1066 struct bkey *search)
1067{
c052dd9a 1068 return __bch_btree_iter_init(b, iter, search, b->set);
911c9610 1069}
a85e968e 1070EXPORT_SYMBOL(bch_btree_iter_init);
911c9610
KO
1071
1072static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
1073 btree_iter_cmp_fn *cmp)
cafe5635
KO
1074{
1075 struct btree_iter_set unused;
1076 struct bkey *ret = NULL;
1077
1078 if (!btree_iter_end(iter)) {
280481d0
KO
1079 bch_btree_iter_next_check(iter);
1080
cafe5635
KO
1081 ret = iter->data->k;
1082 iter->data->k = bkey_next(iter->data->k);
1083
1084 if (iter->data->k > iter->data->end) {
cc0f4eaa 1085 WARN_ONCE(1, "bset was corrupt!\n");
cafe5635
KO
1086 iter->data->k = iter->data->end;
1087 }
1088
1089 if (iter->data->k == iter->data->end)
911c9610 1090 heap_pop(iter, unused, cmp);
cafe5635 1091 else
911c9610 1092 heap_sift(iter, 0, cmp);
cafe5635
KO
1093 }
1094
1095 return ret;
1096}
1097
911c9610
KO
1098struct bkey *bch_btree_iter_next(struct btree_iter *iter)
1099{
1100 return __bch_btree_iter_next(iter, btree_iter_cmp);
1101
1102}
a85e968e 1103EXPORT_SYMBOL(bch_btree_iter_next);
911c9610 1104
cafe5635 1105struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
a85e968e 1106 struct btree_keys *b, ptr_filter_fn fn)
cafe5635
KO
1107{
1108 struct bkey *ret;
1109
1110 do {
1111 ret = bch_btree_iter_next(iter);
1112 } while (ret && fn(b, ret));
1113
1114 return ret;
1115}
1116
cafe5635
KO
1117/* Mergesort */
1118
67539e85
KO
1119void bch_bset_sort_state_free(struct bset_sort_state *state)
1120{
1121 if (state->pool)
1122 mempool_destroy(state->pool);
1123}
1124
1125int bch_bset_sort_state_init(struct bset_sort_state *state, unsigned page_order)
1126{
1127 spin_lock_init(&state->time.lock);
1128
1129 state->page_order = page_order;
1130 state->crit_factor = int_sqrt(1 << page_order);
1131
1132 state->pool = mempool_create_page_pool(1, page_order);
1133 if (!state->pool)
1134 return -ENOMEM;
1135
1136 return 0;
1137}
a85e968e 1138EXPORT_SYMBOL(bch_bset_sort_state_init);
67539e85 1139
a85e968e 1140static void btree_mergesort(struct btree_keys *b, struct bset *out,
cafe5635
KO
1141 struct btree_iter *iter,
1142 bool fixup, bool remove_stale)
1143{
911c9610 1144 int i;
cafe5635 1145 struct bkey *k, *last = NULL;
ef71ec00 1146 BKEY_PADDED(k) tmp;
a85e968e 1147 bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale
cafe5635
KO
1148 ? bch_ptr_bad
1149 : bch_ptr_invalid;
1150
911c9610
KO
1151 /* Heapify the iterator, using our comparison function */
1152 for (i = iter->used / 2 - 1; i >= 0; --i)
65d45231 1153 heap_sift(iter, i, b->ops->sort_cmp);
911c9610 1154
cafe5635 1155 while (!btree_iter_end(iter)) {
65d45231
KO
1156 if (b->ops->sort_fixup && fixup)
1157 k = b->ops->sort_fixup(iter, &tmp.k);
ef71ec00
KO
1158 else
1159 k = NULL;
1160
1161 if (!k)
65d45231 1162 k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
cafe5635 1163
cafe5635
KO
1164 if (bad(b, k))
1165 continue;
1166
1167 if (!last) {
1168 last = out->start;
1169 bkey_copy(last, k);
65d45231 1170 } else if (!bch_bkey_try_merge(b, last, k)) {
cafe5635
KO
1171 last = bkey_next(last);
1172 bkey_copy(last, k);
1173 }
1174 }
1175
1176 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1177
1178 pr_debug("sorted %i keys", out->keys);
cafe5635
KO
1179}
1180
a85e968e 1181static void __btree_sort(struct btree_keys *b, struct btree_iter *iter,
67539e85
KO
1182 unsigned start, unsigned order, bool fixup,
1183 struct bset_sort_state *state)
cafe5635
KO
1184{
1185 uint64_t start_time;
0a451145 1186 bool used_mempool = false;
501d52a9 1187 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT,
cafe5635
KO
1188 order);
1189 if (!out) {
3572324a
KO
1190 struct page *outp;
1191
67539e85
KO
1192 BUG_ON(order > state->page_order);
1193
3572324a
KO
1194 outp = mempool_alloc(state->pool, GFP_NOIO);
1195 out = page_address(outp);
0a451145 1196 used_mempool = true;
a85e968e 1197 order = state->page_order;
cafe5635
KO
1198 }
1199
1200 start_time = local_clock();
1201
67539e85 1202 btree_mergesort(b, out, iter, fixup, false);
cafe5635
KO
1203 b->nsets = start;
1204
cafe5635
KO
1205 if (!start && order == b->page_order) {
1206 /*
1207 * Our temporary buffer is the same size as the btree node's
1208 * buffer, we can just swap buffers instead of doing a big
1209 * memcpy()
1210 */
1211
a85e968e
KO
1212 out->magic = b->set->data->magic;
1213 out->seq = b->set->data->seq;
1214 out->version = b->set->data->version;
1215 swap(out, b->set->data);
cafe5635 1216 } else {
a85e968e
KO
1217 b->set[start].data->keys = out->keys;
1218 memcpy(b->set[start].data->start, out->start,
fafff81c 1219 (void *) bset_bkey_last(out) - (void *) out->start);
cafe5635
KO
1220 }
1221
0a451145 1222 if (used_mempool)
67539e85 1223 mempool_free(virt_to_page(out), state->pool);
cafe5635
KO
1224 else
1225 free_pages((unsigned long) out, order);
1226
a85e968e 1227 bch_bset_build_written_tree(b);
cafe5635 1228
65d22e91 1229 if (!start)
67539e85 1230 bch_time_stats_update(&state->time, start_time);
cafe5635
KO
1231}
1232
89ebb4a2 1233void bch_btree_sort_partial(struct btree_keys *b, unsigned start,
67539e85 1234 struct bset_sort_state *state)
cafe5635 1235{
89ebb4a2 1236 size_t order = b->page_order, keys = 0;
cafe5635 1237 struct btree_iter iter;
89ebb4a2 1238 int oldsize = bch_count_data(b);
280481d0 1239
89ebb4a2 1240 __bch_btree_iter_init(b, &iter, NULL, &b->set[start]);
cafe5635
KO
1241
1242 if (start) {
1243 unsigned i;
1244
89ebb4a2
KO
1245 for (i = start; i <= b->nsets; i++)
1246 keys += b->set[i].data->keys;
cafe5635 1247
89ebb4a2 1248 order = get_order(__set_bytes(b->set->data, keys));
cafe5635
KO
1249 }
1250
89ebb4a2 1251 __btree_sort(b, &iter, start, order, false, state);
cafe5635 1252
89ebb4a2 1253 EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize);
cafe5635 1254}
65d45231 1255EXPORT_SYMBOL(bch_btree_sort_partial);
cafe5635 1256
a85e968e
KO
1257void bch_btree_sort_and_fix_extents(struct btree_keys *b,
1258 struct btree_iter *iter,
67539e85 1259 struct bset_sort_state *state)
cafe5635 1260{
67539e85 1261 __btree_sort(b, iter, 0, b->page_order, true, state);
cafe5635
KO
1262}
1263
89ebb4a2 1264void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
67539e85 1265 struct bset_sort_state *state)
cafe5635
KO
1266{
1267 uint64_t start_time = local_clock();
1268
1269 struct btree_iter iter;
89ebb4a2 1270 bch_btree_iter_init(b, &iter, NULL);
cafe5635 1271
89ebb4a2 1272 btree_mergesort(b, new->set->data, &iter, false, true);
cafe5635 1273
67539e85 1274 bch_time_stats_update(&state->time, start_time);
cafe5635 1275
89ebb4a2 1276 new->set->size = 0; // XXX: why?
cafe5635
KO
1277}
1278
6ded34d1
KO
1279#define SORT_CRIT (4096 / sizeof(uint64_t))
1280
89ebb4a2 1281void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state)
cafe5635 1282{
6ded34d1
KO
1283 unsigned crit = SORT_CRIT;
1284 int i;
cafe5635 1285
6ded34d1 1286 /* Don't sort if nothing to do */
89ebb4a2 1287 if (!b->nsets)
6ded34d1 1288 goto out;
cafe5635 1289
89ebb4a2 1290 for (i = b->nsets - 1; i >= 0; --i) {
67539e85 1291 crit *= state->crit_factor;
cafe5635 1292
89ebb4a2 1293 if (b->set[i].data->keys < crit) {
67539e85 1294 bch_btree_sort_partial(b, i, state);
cafe5635
KO
1295 return;
1296 }
1297 }
1298
6ded34d1 1299 /* Sort if we'd overflow */
89ebb4a2 1300 if (b->nsets + 1 == MAX_BSETS) {
67539e85 1301 bch_btree_sort(b, state);
6ded34d1
KO
1302 return;
1303 }
1304
1305out:
89ebb4a2 1306 bch_bset_build_written_tree(b);
cafe5635 1307}
a85e968e 1308EXPORT_SYMBOL(bch_btree_sort_lazy);
cafe5635 1309
f67342dd 1310void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats)
cafe5635 1311{
cafe5635
KO
1312 unsigned i;
1313
f67342dd
KO
1314 for (i = 0; i <= b->nsets; i++) {
1315 struct bset_tree *t = &b->set[i];
cafe5635
KO
1316 size_t bytes = t->data->keys * sizeof(uint64_t);
1317 size_t j;
1318
f67342dd 1319 if (bset_written(b, t)) {
cafe5635
KO
1320 stats->sets_written++;
1321 stats->bytes_written += bytes;
1322
1323 stats->floats += t->size - 1;
1324
1325 for (j = 1; j < t->size; j++)
1326 if (t->tree[j].exponent == 127)
1327 stats->failed++;
1328 } else {
1329 stats->sets_unwritten++;
1330 stats->bytes_unwritten += bytes;
1331 }
1332 }
cafe5635 1333}