1 // SPDX-License-Identifier: GPL-2.0
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
9 #include "bcachefs.h"
10 #include "btree_cache.h"
11 #include "bset.h"
12 #include "eytzinger.h"
13 #include "trace.h"
14 #include "util.h"
15
16 #include <asm/unaligned.h>
17 #include <linux/console.h>
18 #include <linux/random.h>
19 #include <linux/prefetch.h>
20
21 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *,
22 struct btree *);
23
24 static inline unsigned __btree_node_iter_used(struct btree_node_iter *iter)
25 {
26 unsigned n = ARRAY_SIZE(iter->data);
27
28 while (n && __btree_node_iter_set_end(iter, n - 1))
29 --n;
30
31 return n;
32 }
33
34 struct bset_tree *bch2_bkey_to_bset(struct btree *b, struct bkey_packed *k)
35 {
36 return bch2_bkey_to_bset_inlined(b, k);
37 }
38
39 /*
40 * There are never duplicate live keys in the btree - but including keys that
41 * have been flagged as deleted (and will be cleaned up later) we _will_ see
42 * duplicates.
43 *
44 * Thus the sort order is: usual key comparison first, but for keys that compare
45 * equal the deleted key(s) come first, and the (at most one) live version comes
46 * last.
47 *
48 * The main reason for this is insertion: to handle overwrites, we first iterate
49 * over keys that compare equal to our insert key, and then insert immediately
50 * prior to the first key greater than the key we're inserting - our insert
51 * position will be after all keys that compare equal to our insert key, which
52 * by the time we actually do the insert will all be deleted.
53 */
54
55 void bch2_dump_bset(struct bch_fs *c, struct btree *b,
56 struct bset *i, unsigned set)
57 {
58 struct bkey_packed *_k, *_n;
59 struct bkey uk, n;
60 struct bkey_s_c k;
61 struct printbuf buf = PRINTBUF;
62
63 if (!i->u64s)
64 return;
65
66 for (_k = i->start;
67 _k < vstruct_last(i);
68 _k = _n) {
69 _n = bkey_p_next(_k);
70
71 if (!_k->u64s) {
72 printk(KERN_ERR "block %u key %5zu - u64s 0? aieee!\n", set,
73 _k->_data - i->_data);
74 break;
75 }
76
77 k = bkey_disassemble(b, _k, &uk);
78
79 printbuf_reset(&buf);
80 if (c)
81 bch2_bkey_val_to_text(&buf, c, k);
82 else
83 bch2_bkey_to_text(&buf, k.k);
84 printk(KERN_ERR "block %u key %5zu: %s\n", set,
85 _k->_data - i->_data, buf.buf);
86
87 if (_n == vstruct_last(i))
88 continue;
89
90 n = bkey_unpack_key(b, _n);
91
92 if (bpos_lt(n.p, k.k->p)) {
93 printk(KERN_ERR "Key skipped backwards\n");
94 continue;
95 }
96
97 if (!bkey_deleted(k.k) && bpos_eq(n.p, k.k->p))
98 printk(KERN_ERR "Duplicate keys\n");
99 }
100
101 printbuf_exit(&buf);
102 }
103
104 void bch2_dump_btree_node(struct bch_fs *c, struct btree *b)
105 {
106 console_lock();
107 for_each_bset(b, t)
108 bch2_dump_bset(c, b, bset(b, t), t - b->set);
109 console_unlock();
110 }
111
112 void bch2_dump_btree_node_iter(struct btree *b,
113 struct btree_node_iter *iter)
114 {
115 struct btree_node_iter_set *set;
116 struct printbuf buf = PRINTBUF;
117
118 printk(KERN_ERR "btree node iter with %u/%u sets:\n",
119 __btree_node_iter_used(iter), b->nsets);
120
121 btree_node_iter_for_each(iter, set) {
122 struct bkey_packed *k = __btree_node_offset_to_key(b, set->k);
123 struct bset_tree *t = bch2_bkey_to_bset(b, k);
124 struct bkey uk = bkey_unpack_key(b, k);
125
126 printbuf_reset(&buf);
127 bch2_bkey_to_text(&buf, &uk);
128 printk(KERN_ERR "set %zu key %u: %s\n",
129 t - b->set, set->k, buf.buf);
130 }
131
132 printbuf_exit(&buf);
133 }
134
135 struct btree_nr_keys bch2_btree_node_count_keys(struct btree *b)
136 {
137 struct bkey_packed *k;
138 struct btree_nr_keys nr = {};
139
140 for_each_bset(b, t)
141 bset_tree_for_each_key(b, t, k)
142 if (!bkey_deleted(k))
143 btree_keys_account_key_add(&nr, t - b->set, k);
144 return nr;
145 }
146
147 #ifdef CONFIG_BCACHEFS_DEBUG
148
149 void __bch2_verify_btree_nr_keys(struct btree *b)
150 {
151 struct btree_nr_keys nr = bch2_btree_node_count_keys(b);
152
153 BUG_ON(memcmp(&nr, &b->nr, sizeof(nr)));
154 }
155
156 static void bch2_btree_node_iter_next_check(struct btree_node_iter *_iter,
157 struct btree *b)
158 {
159 struct btree_node_iter iter = *_iter;
160 const struct bkey_packed *k, *n;
161
162 k = bch2_btree_node_iter_peek_all(&iter, b);
163 __bch2_btree_node_iter_advance(&iter, b);
164 n = bch2_btree_node_iter_peek_all(&iter, b);
165
166 bkey_unpack_key(b, k);
167
168 if (n &&
169 bkey_iter_cmp(b, k, n) > 0) {
170 struct btree_node_iter_set *set;
171 struct bkey ku = bkey_unpack_key(b, k);
172 struct bkey nu = bkey_unpack_key(b, n);
173 struct printbuf buf1 = PRINTBUF;
174 struct printbuf buf2 = PRINTBUF;
175
176 bch2_dump_btree_node(NULL, b);
177 bch2_bkey_to_text(&buf1, &ku);
178 bch2_bkey_to_text(&buf2, &nu);
179 printk(KERN_ERR "out of order/overlapping:\n%s\n%s\n",
180 buf1.buf, buf2.buf);
181 printk(KERN_ERR "iter was:");
182
183 btree_node_iter_for_each(_iter, set) {
184 struct bkey_packed *k2 = __btree_node_offset_to_key(b, set->k);
185 struct bset_tree *t = bch2_bkey_to_bset(b, k2);
186 printk(" [%zi %zi]", t - b->set,
187 k2->_data - bset(b, t)->_data);
188 }
189 panic("\n");
190 }
191 }
192
193 void bch2_btree_node_iter_verify(struct btree_node_iter *iter,
194 struct btree *b)
195 {
196 struct btree_node_iter_set *set, *s2;
197 struct bkey_packed *k, *p;
198
199 if (bch2_btree_node_iter_end(iter))
200 return;
201
202 /* Verify no duplicates: */
203 btree_node_iter_for_each(iter, set) {
204 BUG_ON(set->k > set->end);
205 btree_node_iter_for_each(iter, s2)
206 BUG_ON(set != s2 && set->end == s2->end);
207 }
208
209 /* Verify that set->end is correct: */
210 btree_node_iter_for_each(iter, set) {
211 for_each_bset(b, t)
212 if (set->end == t->end_offset) {
213 BUG_ON(set->k < btree_bkey_first_offset(t) ||
214 set->k >= t->end_offset);
215 goto found;
216 }
217 BUG();
218 found:
219 do {} while (0);
220 }
221
222 /* Verify iterator is sorted: */
223 btree_node_iter_for_each(iter, set)
224 BUG_ON(set != iter->data &&
225 btree_node_iter_cmp(b, set[-1], set[0]) > 0);
226
227 k = bch2_btree_node_iter_peek_all(iter, b);
228
229 for_each_bset(b, t) {
230 if (iter->data[0].end == t->end_offset)
231 continue;
232
233 p = bch2_bkey_prev_all(b, t,
234 bch2_btree_node_iter_bset_pos(iter, b, t));
235
236 BUG_ON(p && bkey_iter_cmp(b, k, p) < 0);
237 }
238 }
239
240 void bch2_verify_insert_pos(struct btree *b, struct bkey_packed *where,
241 struct bkey_packed *insert, unsigned clobber_u64s)
242 {
243 struct bset_tree *t = bch2_bkey_to_bset(b, where);
244 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, where);
245 struct bkey_packed *next = (void *) ((u64 *) where->_data + clobber_u64s);
246 struct printbuf buf1 = PRINTBUF;
247 struct printbuf buf2 = PRINTBUF;
248 #if 0
249 BUG_ON(prev &&
250 bkey_iter_cmp(b, prev, insert) > 0);
251 #else
252 if (prev &&
253 bkey_iter_cmp(b, prev, insert) > 0) {
254 struct bkey k1 = bkey_unpack_key(b, prev);
255 struct bkey k2 = bkey_unpack_key(b, insert);
256
257 bch2_dump_btree_node(NULL, b);
258 bch2_bkey_to_text(&buf1, &k1);
259 bch2_bkey_to_text(&buf2, &k2);
260
261 panic("prev > insert:\n"
262 "prev key %s\n"
263 "insert key %s\n",
264 buf1.buf, buf2.buf);
265 }
266 #endif
267 #if 0
268 BUG_ON(next != btree_bkey_last(b, t) &&
269 bkey_iter_cmp(b, insert, next) > 0);
270 #else
271 if (next != btree_bkey_last(b, t) &&
272 bkey_iter_cmp(b, insert, next) > 0) {
273 struct bkey k1 = bkey_unpack_key(b, insert);
274 struct bkey k2 = bkey_unpack_key(b, next);
275
276 bch2_dump_btree_node(NULL, b);
277 bch2_bkey_to_text(&buf1, &k1);
278 bch2_bkey_to_text(&buf2, &k2);
279
280 panic("insert > next:\n"
281 "insert key %s\n"
282 "next key %s\n",
283 buf1.buf, buf2.buf);
284 }
285 #endif
286 }
287
288 #else
289
290 static inline void bch2_btree_node_iter_next_check(struct btree_node_iter *iter,
291 struct btree *b) {}
292
293 #endif
294
295 /* Auxiliary search trees */
296
297 #define BFLOAT_FAILED_UNPACKED U8_MAX
298 #define BFLOAT_FAILED U8_MAX
299
300 struct bkey_float {
301 u8 exponent;
302 u8 key_offset;
303 u16 mantissa;
304 };
305 #define BKEY_MANTISSA_BITS 16
306
307 static unsigned bkey_float_byte_offset(unsigned idx)
308 {
309 return idx * sizeof(struct bkey_float);
310 }
311
312 struct ro_aux_tree {
313 u8 nothing[0];
314 struct bkey_float f[];
315 };
316
317 struct rw_aux_tree {
318 u16 offset;
319 struct bpos k;
320 };
321
322 static unsigned bset_aux_tree_buf_end(const struct bset_tree *t)
323 {
324 BUG_ON(t->aux_data_offset == U16_MAX);
325
326 switch (bset_aux_tree_type(t)) {
327 case BSET_NO_AUX_TREE:
328 return t->aux_data_offset;
329 case BSET_RO_AUX_TREE:
330 return t->aux_data_offset +
331 DIV_ROUND_UP(t->size * sizeof(struct bkey_float) +
332 t->size * sizeof(u8), 8);
333 case BSET_RW_AUX_TREE:
334 return t->aux_data_offset +
335 DIV_ROUND_UP(sizeof(struct rw_aux_tree) * t->size, 8);
336 default:
337 BUG();
338 }
339 }
340
341 static unsigned bset_aux_tree_buf_start(const struct btree *b,
342 const struct bset_tree *t)
343 {
344 return t == b->set
345 ? DIV_ROUND_UP(b->unpack_fn_len, 8)
346 : bset_aux_tree_buf_end(t - 1);
347 }
348
349 static void *__aux_tree_base(const struct btree *b,
350 const struct bset_tree *t)
351 {
352 return b->aux_data + t->aux_data_offset * 8;
353 }
354
355 static struct ro_aux_tree *ro_aux_tree_base(const struct btree *b,
356 const struct bset_tree *t)
357 {
358 EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
359
360 return __aux_tree_base(b, t);
361 }
362
363 static u8 *ro_aux_tree_prev(const struct btree *b,
364 const struct bset_tree *t)
365 {
366 EBUG_ON(bset_aux_tree_type(t) != BSET_RO_AUX_TREE);
367
368 return __aux_tree_base(b, t) + bkey_float_byte_offset(t->size);
369 }
370
371 static struct bkey_float *bkey_float(const struct btree *b,
372 const struct bset_tree *t,
373 unsigned idx)
374 {
375 return ro_aux_tree_base(b, t)->f + idx;
376 }
377
378 static void bset_aux_tree_verify(struct btree *b)
379 {
380 #ifdef CONFIG_BCACHEFS_DEBUG
381 for_each_bset(b, t) {
382 if (t->aux_data_offset == U16_MAX)
383 continue;
384
385 BUG_ON(t != b->set &&
386 t[-1].aux_data_offset == U16_MAX);
387
388 BUG_ON(t->aux_data_offset < bset_aux_tree_buf_start(b, t));
389 BUG_ON(t->aux_data_offset > btree_aux_data_u64s(b));
390 BUG_ON(bset_aux_tree_buf_end(t) > btree_aux_data_u64s(b));
391 }
392 #endif
393 }
394
395 void bch2_btree_keys_init(struct btree *b)
396 {
397 unsigned i;
398
399 b->nsets = 0;
400 memset(&b->nr, 0, sizeof(b->nr));
401
402 for (i = 0; i < MAX_BSETS; i++)
403 b->set[i].data_offset = U16_MAX;
404
405 bch2_bset_set_no_aux_tree(b, b->set);
406 }
407
408 /* Binary tree stuff for auxiliary search trees */
409
410 /*
411 * Cacheline/offset <-> bkey pointer arithmetic:
412 *
413 * t->tree is a binary search tree in an array; each node corresponds to a key
414 * in one cacheline in t->set (BSET_CACHELINE bytes).
415 *
416 * This means we don't have to store the full index of the key that a node in
417 * the binary tree points to; eytzinger1_to_inorder() gives us the cacheline, and
418 * then bkey_float->m gives us the offset within that cacheline, in units of 8
419 * bytes.
420 *
421 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
422 * make this work.
423 *
424 * To construct the bfloat for an arbitrary key we need to know what the key
425 * immediately preceding it is: we have to check if the two keys differ in the
426 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
427 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
428 */
429
430 static inline void *bset_cacheline(const struct btree *b,
431 const struct bset_tree *t,
432 unsigned cacheline)
433 {
434 return (void *) round_down((unsigned long) btree_bkey_first(b, t),
435 L1_CACHE_BYTES) +
436 cacheline * BSET_CACHELINE;
437 }
438
439 static struct bkey_packed *cacheline_to_bkey(const struct btree *b,
440 const struct bset_tree *t,
441 unsigned cacheline,
442 unsigned offset)
443 {
444 return bset_cacheline(b, t, cacheline) + offset * 8;
445 }
446
447 static unsigned bkey_to_cacheline(const struct btree *b,
448 const struct bset_tree *t,
449 const struct bkey_packed *k)
450 {
451 return ((void *) k - bset_cacheline(b, t, 0)) / BSET_CACHELINE;
452 }
453
454 static ssize_t __bkey_to_cacheline_offset(const struct btree *b,
455 const struct bset_tree *t,
456 unsigned cacheline,
457 const struct bkey_packed *k)
458 {
459 return (u64 *) k - (u64 *) bset_cacheline(b, t, cacheline);
460 }
461
462 static unsigned bkey_to_cacheline_offset(const struct btree *b,
463 const struct bset_tree *t,
464 unsigned cacheline,
465 const struct bkey_packed *k)
466 {
467 size_t m = __bkey_to_cacheline_offset(b, t, cacheline, k);
468
469 EBUG_ON(m > U8_MAX);
470 return m;
471 }
472
473 static inline struct bkey_packed *tree_to_bkey(const struct btree *b,
474 const struct bset_tree *t,
475 unsigned j)
476 {
477 return cacheline_to_bkey(b, t,
478 __eytzinger1_to_inorder(j, t->size - 1, t->extra),
479 bkey_float(b, t, j)->key_offset);
480 }
481
482 static struct bkey_packed *tree_to_prev_bkey(const struct btree *b,
483 const struct bset_tree *t,
484 unsigned j)
485 {
486 unsigned prev_u64s = ro_aux_tree_prev(b, t)[j];
487
488 return (void *) ((u64 *) tree_to_bkey(b, t, j)->_data - prev_u64s);
489 }
490
491 static struct rw_aux_tree *rw_aux_tree(const struct btree *b,
492 const struct bset_tree *t)
493 {
494 EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
495
496 return __aux_tree_base(b, t);
497 }
498
499 /*
500 * For the write set - the one we're currently inserting keys into - we don't
501 * maintain a full search tree, we just keep a simple lookup table in t->prev.
502 */
503 static struct bkey_packed *rw_aux_to_bkey(const struct btree *b,
504 struct bset_tree *t,
505 unsigned j)
506 {
507 return __btree_node_offset_to_key(b, rw_aux_tree(b, t)[j].offset);
508 }
509
510 static void rw_aux_tree_set(const struct btree *b, struct bset_tree *t,
511 unsigned j, struct bkey_packed *k)
512 {
513 EBUG_ON(k >= btree_bkey_last(b, t));
514
515 rw_aux_tree(b, t)[j] = (struct rw_aux_tree) {
516 .offset = __btree_node_key_to_offset(b, k),
517 .k = bkey_unpack_pos(b, k),
518 };
519 }
520
521 static void bch2_bset_verify_rw_aux_tree(struct btree *b,
522 struct bset_tree *t)
523 {
524 struct bkey_packed *k = btree_bkey_first(b, t);
525 unsigned j = 0;
526
527 if (!bch2_expensive_debug_checks)
528 return;
529
530 BUG_ON(bset_has_ro_aux_tree(t));
531
532 if (!bset_has_rw_aux_tree(t))
533 return;
534
535 BUG_ON(t->size < 1);
536 BUG_ON(rw_aux_to_bkey(b, t, j) != k);
537
538 goto start;
539 while (1) {
540 if (rw_aux_to_bkey(b, t, j) == k) {
541 BUG_ON(!bpos_eq(rw_aux_tree(b, t)[j].k,
542 bkey_unpack_pos(b, k)));
543 start:
544 if (++j == t->size)
545 break;
546
547 BUG_ON(rw_aux_tree(b, t)[j].offset <=
548 rw_aux_tree(b, t)[j - 1].offset);
549 }
550
551 k = bkey_p_next(k);
552 BUG_ON(k >= btree_bkey_last(b, t));
553 }
554 }
555
556 /* returns idx of first entry >= offset: */
557 static unsigned rw_aux_tree_bsearch(struct btree *b,
558 struct bset_tree *t,
559 unsigned offset)
560 {
561 unsigned bset_offs = offset - btree_bkey_first_offset(t);
562 unsigned bset_u64s = t->end_offset - btree_bkey_first_offset(t);
563 unsigned idx = bset_u64s ? bset_offs * t->size / bset_u64s : 0;
564
565 EBUG_ON(bset_aux_tree_type(t) != BSET_RW_AUX_TREE);
566 EBUG_ON(!t->size);
567 EBUG_ON(idx > t->size);
568
569 while (idx < t->size &&
570 rw_aux_tree(b, t)[idx].offset < offset)
571 idx++;
572
573 while (idx &&
574 rw_aux_tree(b, t)[idx - 1].offset >= offset)
575 idx--;
576
577 EBUG_ON(idx < t->size &&
578 rw_aux_tree(b, t)[idx].offset < offset);
579 EBUG_ON(idx && rw_aux_tree(b, t)[idx - 1].offset >= offset);
580 EBUG_ON(idx + 1 < t->size &&
581 rw_aux_tree(b, t)[idx].offset ==
582 rw_aux_tree(b, t)[idx + 1].offset);
583
584 return idx;
585 }
586
587 static inline unsigned bkey_mantissa(const struct bkey_packed *k,
588 const struct bkey_float *f,
589 unsigned idx)
590 {
591 u64 v;
592
593 EBUG_ON(!bkey_packed(k));
594
595 v = get_unaligned((u64 *) (((u8 *) k->_data) + (f->exponent >> 3)));
596
597 /*
598 * In little endian, we're shifting off low bits (and then the bits we
599 * want are at the low end), in big endian we're shifting off high bits
600 * (and then the bits we want are at the high end, so we shift them
601 * back down):
602 */
603 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
604 v >>= f->exponent & 7;
605 #else
606 v >>= 64 - (f->exponent & 7) - BKEY_MANTISSA_BITS;
607 #endif
608 return (u16) v;
609 }
610
611 static __always_inline void make_bfloat(struct btree *b, struct bset_tree *t,
612 unsigned j,
613 struct bkey_packed *min_key,
614 struct bkey_packed *max_key)
615 {
616 struct bkey_float *f = bkey_float(b, t, j);
617 struct bkey_packed *m = tree_to_bkey(b, t, j);
618 struct bkey_packed *l = is_power_of_2(j)
619 ? min_key
620 : tree_to_prev_bkey(b, t, j >> ffs(j));
621 struct bkey_packed *r = is_power_of_2(j + 1)
622 ? max_key
623 : tree_to_bkey(b, t, j >> (ffz(j) + 1));
624 unsigned mantissa;
625 int shift, exponent, high_bit;
626
627 /*
628 * for failed bfloats, the lookup code falls back to comparing against
629 * the original key.
630 */
631
632 if (!bkey_packed(l) || !bkey_packed(r) || !bkey_packed(m) ||
633 !b->nr_key_bits) {
634 f->exponent = BFLOAT_FAILED_UNPACKED;
635 return;
636 }
637
638 /*
639 * The greatest differing bit of l and r is the first bit we must
640 * include in the bfloat mantissa we're creating in order to do
641 * comparisons - that bit always becomes the high bit of
642 * bfloat->mantissa, and thus the exponent we're calculating here is
643 * the position of what will become the low bit in bfloat->mantissa:
644 *
645 * Note that this may be negative - we may be running off the low end
646 * of the key: we handle this later:
647 */
648 high_bit = max(bch2_bkey_greatest_differing_bit(b, l, r),
649 min_t(unsigned, BKEY_MANTISSA_BITS, b->nr_key_bits) - 1);
650 exponent = high_bit - (BKEY_MANTISSA_BITS - 1);
651
652 /*
653 * Then we calculate the actual shift value, from the start of the key
654 * (k->_data), to get the key bits starting at exponent:
655 */
656 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
657 shift = (int) (b->format.key_u64s * 64 - b->nr_key_bits) + exponent;
658
659 EBUG_ON(shift + BKEY_MANTISSA_BITS > b->format.key_u64s * 64);
660 #else
661 shift = high_bit_offset +
662 b->nr_key_bits -
663 exponent -
664 BKEY_MANTISSA_BITS;
665
666 EBUG_ON(shift < KEY_PACKED_BITS_START);
667 #endif
668 EBUG_ON(shift < 0 || shift >= BFLOAT_FAILED);
669
670 f->exponent = shift;
671 mantissa = bkey_mantissa(m, f, j);
672
673 /*
674 * If we've got garbage bits, set them to all 1s - it's legal for the
675 * bfloat to compare larger than the original key, but not smaller:
676 */
677 if (exponent < 0)
678 mantissa |= ~(~0U << -exponent);
679
680 f->mantissa = mantissa;
681 }
682
683 /* bytes remaining - only valid for last bset: */
684 static unsigned __bset_tree_capacity(struct btree *b, const struct bset_tree *t)
685 {
686 bset_aux_tree_verify(b);
687
688 return btree_aux_data_bytes(b) - t->aux_data_offset * sizeof(u64);
689 }
690
691 static unsigned bset_ro_tree_capacity(struct btree *b, const struct bset_tree *t)
692 {
693 return __bset_tree_capacity(b, t) /
694 (sizeof(struct bkey_float) + sizeof(u8));
695 }
696
697 static unsigned bset_rw_tree_capacity(struct btree *b, const struct bset_tree *t)
698 {
699 return __bset_tree_capacity(b, t) / sizeof(struct rw_aux_tree);
700 }
701
702 static noinline void __build_rw_aux_tree(struct btree *b, struct bset_tree *t)
703 {
704 struct bkey_packed *k;
705
706 t->size = 1;
707 t->extra = BSET_RW_AUX_TREE_VAL;
708 rw_aux_tree(b, t)[0].offset =
709 __btree_node_key_to_offset(b, btree_bkey_first(b, t));
710
711 bset_tree_for_each_key(b, t, k) {
712 if (t->size == bset_rw_tree_capacity(b, t))
713 break;
714
715 if ((void *) k - (void *) rw_aux_to_bkey(b, t, t->size - 1) >
716 L1_CACHE_BYTES)
717 rw_aux_tree_set(b, t, t->size++, k);
718 }
719 }
720
721 static noinline void __build_ro_aux_tree(struct btree *b, struct bset_tree *t)
722 {
723 struct bkey_packed *prev = NULL, *k = btree_bkey_first(b, t);
724 struct bkey_i min_key, max_key;
725 unsigned cacheline = 1;
726
727 t->size = min(bkey_to_cacheline(b, t, btree_bkey_last(b, t)),
728 bset_ro_tree_capacity(b, t));
729 retry:
730 if (t->size < 2) {
731 t->size = 0;
732 t->extra = BSET_NO_AUX_TREE_VAL;
733 return;
734 }
735
736 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
737
738 /* First we figure out where the first key in each cacheline is */
739 eytzinger1_for_each(j, t->size - 1) {
740 while (bkey_to_cacheline(b, t, k) < cacheline)
741 prev = k, k = bkey_p_next(k);
742
743 if (k >= btree_bkey_last(b, t)) {
744 /* XXX: this path sucks */
745 t->size--;
746 goto retry;
747 }
748
749 ro_aux_tree_prev(b, t)[j] = prev->u64s;
750 bkey_float(b, t, j)->key_offset =
751 bkey_to_cacheline_offset(b, t, cacheline++, k);
752
753 EBUG_ON(tree_to_prev_bkey(b, t, j) != prev);
754 EBUG_ON(tree_to_bkey(b, t, j) != k);
755 }
756
757 while (k != btree_bkey_last(b, t))
758 prev = k, k = bkey_p_next(k);
759
760 if (!bkey_pack_pos(bkey_to_packed(&min_key), b->data->min_key, b)) {
761 bkey_init(&min_key.k);
762 min_key.k.p = b->data->min_key;
763 }
764
765 if (!bkey_pack_pos(bkey_to_packed(&max_key), b->data->max_key, b)) {
766 bkey_init(&max_key.k);
767 max_key.k.p = b->data->max_key;
768 }
769
770 /* Then we build the tree */
771 eytzinger1_for_each(j, t->size - 1)
772 make_bfloat(b, t, j,
773 bkey_to_packed(&min_key),
774 bkey_to_packed(&max_key));
775 }
776
777 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
778 {
779 struct bset_tree *i;
780
781 for (i = b->set; i != t; i++)
782 BUG_ON(bset_has_rw_aux_tree(i));
783
784 bch2_bset_set_no_aux_tree(b, t);
785
786 /* round up to next cacheline: */
787 t->aux_data_offset = round_up(bset_aux_tree_buf_start(b, t),
788 SMP_CACHE_BYTES / sizeof(u64));
789
790 bset_aux_tree_verify(b);
791 }
792
793 void bch2_bset_build_aux_tree(struct btree *b, struct bset_tree *t,
794 bool writeable)
795 {
796 if (writeable
797 ? bset_has_rw_aux_tree(t)
798 : bset_has_ro_aux_tree(t))
799 return;
800
801 bset_alloc_tree(b, t);
802
803 if (!__bset_tree_capacity(b, t))
804 return;
805
806 if (writeable)
807 __build_rw_aux_tree(b, t);
808 else
809 __build_ro_aux_tree(b, t);
810
811 bset_aux_tree_verify(b);
812 }
813
814 void bch2_bset_init_first(struct btree *b, struct bset *i)
815 {
816 struct bset_tree *t;
817
818 BUG_ON(b->nsets);
819
820 memset(i, 0, sizeof(*i));
821 get_random_bytes(&i->seq, sizeof(i->seq));
822 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
823
824 t = &b->set[b->nsets++];
825 set_btree_bset(b, t, i);
826 }
827
828 void bch2_bset_init_next(struct btree *b, struct btree_node_entry *bne)
829 {
830 struct bset *i = &bne->keys;
831 struct bset_tree *t;
832
833 BUG_ON(bset_byte_offset(b, bne) >= btree_buf_bytes(b));
834 BUG_ON((void *) bne < (void *) btree_bkey_last(b, bset_tree_last(b)));
835 BUG_ON(b->nsets >= MAX_BSETS);
836
837 memset(i, 0, sizeof(*i));
838 i->seq = btree_bset_first(b)->seq;
839 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
840
841 t = &b->set[b->nsets++];
842 set_btree_bset(b, t, i);
843 }
844
845 /*
846 * find _some_ key in the same bset as @k that precedes @k - not necessarily the
847 * immediate predecessor:
848 */
849 static struct bkey_packed *__bkey_prev(struct btree *b, struct bset_tree *t,
850 struct bkey_packed *k)
851 {
852 struct bkey_packed *p;
853 unsigned offset;
854 int j;
855
856 EBUG_ON(k < btree_bkey_first(b, t) ||
857 k > btree_bkey_last(b, t));
858
859 if (k == btree_bkey_first(b, t))
860 return NULL;
861
862 switch (bset_aux_tree_type(t)) {
863 case BSET_NO_AUX_TREE:
864 p = btree_bkey_first(b, t);
865 break;
866 case BSET_RO_AUX_TREE:
867 j = min_t(unsigned, t->size - 1, bkey_to_cacheline(b, t, k));
868
869 do {
870 p = j ? tree_to_bkey(b, t,
871 __inorder_to_eytzinger1(j--,
872 t->size - 1, t->extra))
873 : btree_bkey_first(b, t);
874 } while (p >= k);
875 break;
876 case BSET_RW_AUX_TREE:
877 offset = __btree_node_key_to_offset(b, k);
878 j = rw_aux_tree_bsearch(b, t, offset);
879 p = j ? rw_aux_to_bkey(b, t, j - 1)
880 : btree_bkey_first(b, t);
881 break;
882 }
883
884 return p;
885 }
886
887 struct bkey_packed *bch2_bkey_prev_filter(struct btree *b,
888 struct bset_tree *t,
889 struct bkey_packed *k,
890 unsigned min_key_type)
891 {
892 struct bkey_packed *p, *i, *ret = NULL, *orig_k = k;
893
894 while ((p = __bkey_prev(b, t, k)) && !ret) {
895 for (i = p; i != k; i = bkey_p_next(i))
896 if (i->type >= min_key_type)
897 ret = i;
898
899 k = p;
900 }
901
902 if (bch2_expensive_debug_checks) {
903 BUG_ON(ret >= orig_k);
904
905 for (i = ret
906 ? bkey_p_next(ret)
907 : btree_bkey_first(b, t);
908 i != orig_k;
909 i = bkey_p_next(i))
910 BUG_ON(i->type >= min_key_type);
911 }
912
913 return ret;
914 }
915
916 /* Insert */
917
918 static void bch2_bset_fix_lookup_table(struct btree *b,
919 struct bset_tree *t,
920 struct bkey_packed *_where,
921 unsigned clobber_u64s,
922 unsigned new_u64s)
923 {
924 int shift = new_u64s - clobber_u64s;
925 unsigned l, j, where = __btree_node_key_to_offset(b, _where);
926
927 EBUG_ON(bset_has_ro_aux_tree(t));
928
929 if (!bset_has_rw_aux_tree(t))
930 return;
931
932 /* returns first entry >= where */
933 l = rw_aux_tree_bsearch(b, t, where);
934
935 if (!l) /* never delete first entry */
936 l++;
937 else if (l < t->size &&
938 where < t->end_offset &&
939 rw_aux_tree(b, t)[l].offset == where)
940 rw_aux_tree_set(b, t, l++, _where);
941
942 /* l now > where */
943
944 for (j = l;
945 j < t->size &&
946 rw_aux_tree(b, t)[j].offset < where + clobber_u64s;
947 j++)
948 ;
949
950 if (j < t->size &&
951 rw_aux_tree(b, t)[j].offset + shift ==
952 rw_aux_tree(b, t)[l - 1].offset)
953 j++;
954
955 memmove(&rw_aux_tree(b, t)[l],
956 &rw_aux_tree(b, t)[j],
957 (void *) &rw_aux_tree(b, t)[t->size] -
958 (void *) &rw_aux_tree(b, t)[j]);
959 t->size -= j - l;
960
961 for (j = l; j < t->size; j++)
962 rw_aux_tree(b, t)[j].offset += shift;
963
964 EBUG_ON(l < t->size &&
965 rw_aux_tree(b, t)[l].offset ==
966 rw_aux_tree(b, t)[l - 1].offset);
967
968 if (t->size < bset_rw_tree_capacity(b, t) &&
969 (l < t->size
970 ? rw_aux_tree(b, t)[l].offset
971 : t->end_offset) -
972 rw_aux_tree(b, t)[l - 1].offset >
973 L1_CACHE_BYTES / sizeof(u64)) {
974 struct bkey_packed *start = rw_aux_to_bkey(b, t, l - 1);
975 struct bkey_packed *end = l < t->size
976 ? rw_aux_to_bkey(b, t, l)
977 : btree_bkey_last(b, t);
978 struct bkey_packed *k = start;
979
980 while (1) {
981 k = bkey_p_next(k);
982 if (k == end)
983 break;
984
985 if ((void *) k - (void *) start >= L1_CACHE_BYTES) {
986 memmove(&rw_aux_tree(b, t)[l + 1],
987 &rw_aux_tree(b, t)[l],
988 (void *) &rw_aux_tree(b, t)[t->size] -
989 (void *) &rw_aux_tree(b, t)[l]);
990 t->size++;
991 rw_aux_tree_set(b, t, l, k);
992 break;
993 }
994 }
995 }
996
997 bch2_bset_verify_rw_aux_tree(b, t);
998 bset_aux_tree_verify(b);
999 }
1000
1001 void bch2_bset_insert(struct btree *b,
1002 struct btree_node_iter *iter,
1003 struct bkey_packed *where,
1004 struct bkey_i *insert,
1005 unsigned clobber_u64s)
1006 {
1007 struct bkey_format *f = &b->format;
1008 struct bset_tree *t = bset_tree_last(b);
1009 struct bkey_packed packed, *src = bkey_to_packed(insert);
1010
1011 bch2_bset_verify_rw_aux_tree(b, t);
1012 bch2_verify_insert_pos(b, where, bkey_to_packed(insert), clobber_u64s);
1013
1014 if (bch2_bkey_pack_key(&packed, &insert->k, f))
1015 src = &packed;
1016
1017 if (!bkey_deleted(&insert->k))
1018 btree_keys_account_key_add(&b->nr, t - b->set, src);
1019
1020 if (src->u64s != clobber_u64s) {
1021 u64 *src_p = (u64 *) where->_data + clobber_u64s;
1022 u64 *dst_p = (u64 *) where->_data + src->u64s;
1023
1024 EBUG_ON((int) le16_to_cpu(bset(b, t)->u64s) <
1025 (int) clobber_u64s - src->u64s);
1026
1027 memmove_u64s(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1028 le16_add_cpu(&bset(b, t)->u64s, src->u64s - clobber_u64s);
1029 set_btree_bset_end(b, t);
1030 }
1031
1032 memcpy_u64s_small(where, src,
1033 bkeyp_key_u64s(f, src));
1034 memcpy_u64s(bkeyp_val(f, where), &insert->v,
1035 bkeyp_val_u64s(f, src));
1036
1037 if (src->u64s != clobber_u64s)
1038 bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, src->u64s);
1039
1040 bch2_verify_btree_nr_keys(b);
1041 }
1042
1043 void bch2_bset_delete(struct btree *b,
1044 struct bkey_packed *where,
1045 unsigned clobber_u64s)
1046 {
1047 struct bset_tree *t = bset_tree_last(b);
1048 u64 *src_p = (u64 *) where->_data + clobber_u64s;
1049 u64 *dst_p = where->_data;
1050
1051 bch2_bset_verify_rw_aux_tree(b, t);
1052
1053 EBUG_ON(le16_to_cpu(bset(b, t)->u64s) < clobber_u64s);
1054
1055 memmove_u64s_down(dst_p, src_p, btree_bkey_last(b, t)->_data - src_p);
1056 le16_add_cpu(&bset(b, t)->u64s, -clobber_u64s);
1057 set_btree_bset_end(b, t);
1058
1059 bch2_bset_fix_lookup_table(b, t, where, clobber_u64s, 0);
1060 }
1061
1062 /* Lookup */
1063
1064 __flatten
1065 static struct bkey_packed *bset_search_write_set(const struct btree *b,
1066 struct bset_tree *t,
1067 struct bpos *search)
1068 {
1069 unsigned l = 0, r = t->size;
1070
1071 while (l + 1 != r) {
1072 unsigned m = (l + r) >> 1;
1073
1074 if (bpos_lt(rw_aux_tree(b, t)[m].k, *search))
1075 l = m;
1076 else
1077 r = m;
1078 }
1079
1080 return rw_aux_to_bkey(b, t, l);
1081 }
1082
1083 static inline void prefetch_four_cachelines(void *p)
1084 {
1085 #ifdef CONFIG_X86_64
1086 asm("prefetcht0 (-127 + 64 * 0)(%0);"
1087 "prefetcht0 (-127 + 64 * 1)(%0);"
1088 "prefetcht0 (-127 + 64 * 2)(%0);"
1089 "prefetcht0 (-127 + 64 * 3)(%0);"
1090 :
1091 : "r" (p + 127));
1092 #else
1093 prefetch(p + L1_CACHE_BYTES * 0);
1094 prefetch(p + L1_CACHE_BYTES * 1);
1095 prefetch(p + L1_CACHE_BYTES * 2);
1096 prefetch(p + L1_CACHE_BYTES * 3);
1097 #endif
1098 }
1099
1100 static inline bool bkey_mantissa_bits_dropped(const struct btree *b,
1101 const struct bkey_float *f,
1102 unsigned idx)
1103 {
1104 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1105 unsigned key_bits_start = b->format.key_u64s * 64 - b->nr_key_bits;
1106
1107 return f->exponent > key_bits_start;
1108 #else
1109 unsigned key_bits_end = high_bit_offset + b->nr_key_bits;
1110
1111 return f->exponent + BKEY_MANTISSA_BITS < key_bits_end;
1112 #endif
1113 }
1114
1115 __flatten
1116 static struct bkey_packed *bset_search_tree(const struct btree *b,
1117 const struct bset_tree *t,
1118 const struct bpos *search,
1119 const struct bkey_packed *packed_search)
1120 {
1121 struct ro_aux_tree *base = ro_aux_tree_base(b, t);
1122 struct bkey_float *f;
1123 struct bkey_packed *k;
1124 unsigned inorder, n = 1, l, r;
1125 int cmp;
1126
1127 do {
1128 if (likely(n << 4 < t->size))
1129 prefetch(&base->f[n << 4]);
1130
1131 f = &base->f[n];
1132 if (unlikely(f->exponent >= BFLOAT_FAILED))
1133 goto slowpath;
1134
1135 l = f->mantissa;
1136 r = bkey_mantissa(packed_search, f, n);
1137
1138 if (unlikely(l == r) && bkey_mantissa_bits_dropped(b, f, n))
1139 goto slowpath;
1140
1141 n = n * 2 + (l < r);
1142 continue;
1143 slowpath:
1144 k = tree_to_bkey(b, t, n);
1145 cmp = bkey_cmp_p_or_unp(b, k, packed_search, search);
1146 if (!cmp)
1147 return k;
1148
1149 n = n * 2 + (cmp < 0);
1150 } while (n < t->size);
1151
1152 inorder = __eytzinger1_to_inorder(n >> 1, t->size - 1, t->extra);
1153
1154 /*
1155 * n would have been the node we recursed to - the low bit tells us if
1156 * we recursed left or recursed right.
1157 */
1158 if (likely(!(n & 1))) {
1159 --inorder;
1160 if (unlikely(!inorder))
1161 return btree_bkey_first(b, t);
1162
1163 f = &base->f[eytzinger1_prev(n >> 1, t->size - 1)];
1164 }
1165
1166 return cacheline_to_bkey(b, t, inorder, f->key_offset);
1167 }
1168
1169 static __always_inline __flatten
1170 struct bkey_packed *__bch2_bset_search(struct btree *b,
1171 struct bset_tree *t,
1172 struct bpos *search,
1173 const struct bkey_packed *lossy_packed_search)
1174 {
1175
1176 /*
1177 * First, we search for a cacheline, then lastly we do a linear search
1178 * within that cacheline.
1179 *
1180 * To search for the cacheline, there's three different possibilities:
1181 * * The set is too small to have a search tree, so we just do a linear
1182 * search over the whole set.
1183 * * The set is the one we're currently inserting into; keeping a full
1184 * auxiliary search tree up to date would be too expensive, so we
1185 * use a much simpler lookup table to do a binary search -
1186 * bset_search_write_set().
1187 * * Or we use the auxiliary search tree we constructed earlier -
1188 * bset_search_tree()
1189 */
1190
1191 switch (bset_aux_tree_type(t)) {
1192 case BSET_NO_AUX_TREE:
1193 return btree_bkey_first(b, t);
1194 case BSET_RW_AUX_TREE:
1195 return bset_search_write_set(b, t, search);
1196 case BSET_RO_AUX_TREE:
1197 return bset_search_tree(b, t, search, lossy_packed_search);
1198 default:
1199 BUG();
1200 }
1201 }
1202
1203 static __always_inline __flatten
1204 struct bkey_packed *bch2_bset_search_linear(struct btree *b,
1205 struct bset_tree *t,
1206 struct bpos *search,
1207 struct bkey_packed *packed_search,
1208 const struct bkey_packed *lossy_packed_search,
1209 struct bkey_packed *m)
1210 {
1211 if (lossy_packed_search)
1212 while (m != btree_bkey_last(b, t) &&
1213 bkey_iter_cmp_p_or_unp(b, m,
1214 lossy_packed_search, search) < 0)
1215 m = bkey_p_next(m);
1216
1217 if (!packed_search)
1218 while (m != btree_bkey_last(b, t) &&
1219 bkey_iter_pos_cmp(b, m, search) < 0)
1220 m = bkey_p_next(m);
1221
1222 if (bch2_expensive_debug_checks) {
1223 struct bkey_packed *prev = bch2_bkey_prev_all(b, t, m);
1224
1225 BUG_ON(prev &&
1226 bkey_iter_cmp_p_or_unp(b, prev,
1227 packed_search, search) >= 0);
1228 }
1229
1230 return m;
1231 }
1232
1233 /* Btree node iterator */
1234
1235 static inline void __bch2_btree_node_iter_push(struct btree_node_iter *iter,
1236 struct btree *b,
1237 const struct bkey_packed *k,
1238 const struct bkey_packed *end)
1239 {
1240 if (k != end) {
1241 struct btree_node_iter_set *pos;
1242
1243 btree_node_iter_for_each(iter, pos)
1244 ;
1245
1246 BUG_ON(pos >= iter->data + ARRAY_SIZE(iter->data));
1247 *pos = (struct btree_node_iter_set) {
1248 __btree_node_key_to_offset(b, k),
1249 __btree_node_key_to_offset(b, end)
1250 };
1251 }
1252 }
1253
1254 void bch2_btree_node_iter_push(struct btree_node_iter *iter,
1255 struct btree *b,
1256 const struct bkey_packed *k,
1257 const struct bkey_packed *end)
1258 {
1259 __bch2_btree_node_iter_push(iter, b, k, end);
1260 bch2_btree_node_iter_sort(iter, b);
1261 }
1262
1263 noinline __flatten __cold
1264 static void btree_node_iter_init_pack_failed(struct btree_node_iter *iter,
1265 struct btree *b, struct bpos *search)
1266 {
1267 struct bkey_packed *k;
1268
1269 trace_bkey_pack_pos_fail(search);
1270
1271 bch2_btree_node_iter_init_from_start(iter, b);
1272
1273 while ((k = bch2_btree_node_iter_peek(iter, b)) &&
1274 bkey_iter_pos_cmp(b, k, search) < 0)
1275 bch2_btree_node_iter_advance(iter, b);
1276 }
1277
1278 /**
1279 * bch2_btree_node_iter_init - initialize a btree node iterator, starting from a
1280 * given position
1281 *
1282 * @iter: iterator to initialize
1283 * @b: btree node to search
1284 * @search: search key
1285 *
1286 * Main entry point to the lookup code for individual btree nodes:
1287 *
1288 * NOTE:
1289 *
1290 * When you don't filter out deleted keys, btree nodes _do_ contain duplicate
1291 * keys. This doesn't matter for most code, but it does matter for lookups.
1292 *
1293 * Some adjacent keys with a string of equal keys:
1294 * i j k k k k l m
1295 *
1296 * If you search for k, the lookup code isn't guaranteed to return you any
1297 * specific k. The lookup code is conceptually doing a binary search and
1298 * iterating backwards is very expensive so if the pivot happens to land at the
1299 * last k that's what you'll get.
1300 *
1301 * This works out ok, but it's something to be aware of:
1302 *
1303 * - For non extents, we guarantee that the live key comes last - see
1304 * btree_node_iter_cmp(), keys_out_of_order(). So the duplicates you don't
1305 * see will only be deleted keys you don't care about.
1306 *
1307 * - For extents, deleted keys sort last (see the comment at the top of this
1308 * file). But when you're searching for extents, you actually want the first
1309 * key strictly greater than your search key - an extent that compares equal
1310 * to the search key is going to have 0 sectors after the search key.
1311 *
1312 * But this does mean that we can't just search for
1313 * bpos_successor(start_of_range) to get the first extent that overlaps with
1314 * the range we want - if we're unlucky and there's an extent that ends
1315 * exactly where we searched, then there could be a deleted key at the same
1316 * position and we'd get that when we search instead of the preceding extent
1317 * we needed.
1318 *
1319 * So we've got to search for start_of_range, then after the lookup iterate
1320 * past any extents that compare equal to the position we searched for.
1321 */
1322 __flatten
1323 void bch2_btree_node_iter_init(struct btree_node_iter *iter,
1324 struct btree *b, struct bpos *search)
1325 {
1326 struct bkey_packed p, *packed_search = NULL;
1327 struct btree_node_iter_set *pos = iter->data;
1328 struct bkey_packed *k[MAX_BSETS];
1329 unsigned i;
1330
1331 EBUG_ON(bpos_lt(*search, b->data->min_key));
1332 EBUG_ON(bpos_gt(*search, b->data->max_key));
1333 bset_aux_tree_verify(b);
1334
1335 memset(iter, 0, sizeof(*iter));
1336
1337 switch (bch2_bkey_pack_pos_lossy(&p, *search, b)) {
1338 case BKEY_PACK_POS_EXACT:
1339 packed_search = &p;
1340 break;
1341 case BKEY_PACK_POS_SMALLER:
1342 packed_search = NULL;
1343 break;
1344 case BKEY_PACK_POS_FAIL:
1345 btree_node_iter_init_pack_failed(iter, b, search);
1346 return;
1347 }
1348
1349 for (i = 0; i < b->nsets; i++) {
1350 k[i] = __bch2_bset_search(b, b->set + i, search, &p);
1351 prefetch_four_cachelines(k[i]);
1352 }
1353
1354 for (i = 0; i < b->nsets; i++) {
1355 struct bset_tree *t = b->set + i;
1356 struct bkey_packed *end = btree_bkey_last(b, t);
1357
1358 k[i] = bch2_bset_search_linear(b, t, search,
1359 packed_search, &p, k[i]);
1360 if (k[i] != end)
1361 *pos++ = (struct btree_node_iter_set) {
1362 __btree_node_key_to_offset(b, k[i]),
1363 __btree_node_key_to_offset(b, end)
1364 };
1365 }
1366
1367 bch2_btree_node_iter_sort(iter, b);
1368 }
1369
1370 void bch2_btree_node_iter_init_from_start(struct btree_node_iter *iter,
1371 struct btree *b)
1372 {
1373 memset(iter, 0, sizeof(*iter));
1374
1375 for_each_bset(b, t)
1376 __bch2_btree_node_iter_push(iter, b,
1377 btree_bkey_first(b, t),
1378 btree_bkey_last(b, t));
1379 bch2_btree_node_iter_sort(iter, b);
1380 }
1381
1382 struct bkey_packed *bch2_btree_node_iter_bset_pos(struct btree_node_iter *iter,
1383 struct btree *b,
1384 struct bset_tree *t)
1385 {
1386 struct btree_node_iter_set *set;
1387
1388 btree_node_iter_for_each(iter, set)
1389 if (set->end == t->end_offset)
1390 return __btree_node_offset_to_key(b, set->k);
1391
1392 return btree_bkey_last(b, t);
1393 }
1394
1395 static inline bool btree_node_iter_sort_two(struct btree_node_iter *iter,
1396 struct btree *b,
1397 unsigned first)
1398 {
1399 bool ret;
1400
1401 if ((ret = (btree_node_iter_cmp(b,
1402 iter->data[first],
1403 iter->data[first + 1]) > 0)))
1404 swap(iter->data[first], iter->data[first + 1]);
1405 return ret;
1406 }
1407
1408 void bch2_btree_node_iter_sort(struct btree_node_iter *iter,
1409 struct btree *b)
1410 {
1411 /* unrolled bubble sort: */
1412
1413 if (!__btree_node_iter_set_end(iter, 2)) {
1414 btree_node_iter_sort_two(iter, b, 0);
1415 btree_node_iter_sort_two(iter, b, 1);
1416 }
1417
1418 if (!__btree_node_iter_set_end(iter, 1))
1419 btree_node_iter_sort_two(iter, b, 0);
1420 }
1421
1422 void bch2_btree_node_iter_set_drop(struct btree_node_iter *iter,
1423 struct btree_node_iter_set *set)
1424 {
1425 struct btree_node_iter_set *last =
1426 iter->data + ARRAY_SIZE(iter->data) - 1;
1427
1428 memmove(&set[0], &set[1], (void *) last - (void *) set);
1429 *last = (struct btree_node_iter_set) { 0, 0 };
1430 }
1431
1432 static inline void __bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1433 struct btree *b)
1434 {
1435 iter->data->k += __bch2_btree_node_iter_peek_all(iter, b)->u64s;
1436
1437 EBUG_ON(iter->data->k > iter->data->end);
1438
1439 if (unlikely(__btree_node_iter_set_end(iter, 0))) {
1440 /* avoid an expensive memmove call: */
1441 iter->data[0] = iter->data[1];
1442 iter->data[1] = iter->data[2];
1443 iter->data[2] = (struct btree_node_iter_set) { 0, 0 };
1444 return;
1445 }
1446
1447 if (__btree_node_iter_set_end(iter, 1))
1448 return;
1449
1450 if (!btree_node_iter_sort_two(iter, b, 0))
1451 return;
1452
1453 if (__btree_node_iter_set_end(iter, 2))
1454 return;
1455
1456 btree_node_iter_sort_two(iter, b, 1);
1457 }
1458
1459 void bch2_btree_node_iter_advance(struct btree_node_iter *iter,
1460 struct btree *b)
1461 {
1462 if (bch2_expensive_debug_checks) {
1463 bch2_btree_node_iter_verify(iter, b);
1464 bch2_btree_node_iter_next_check(iter, b);
1465 }
1466
1467 __bch2_btree_node_iter_advance(iter, b);
1468 }
1469
1470 /*
1471 * Expensive:
1472 */
1473 struct bkey_packed *bch2_btree_node_iter_prev_all(struct btree_node_iter *iter,
1474 struct btree *b)
1475 {
1476 struct bkey_packed *k, *prev = NULL;
1477 struct btree_node_iter_set *set;
1478 unsigned end = 0;
1479
1480 if (bch2_expensive_debug_checks)
1481 bch2_btree_node_iter_verify(iter, b);
1482
1483 for_each_bset(b, t) {
1484 k = bch2_bkey_prev_all(b, t,
1485 bch2_btree_node_iter_bset_pos(iter, b, t));
1486 if (k &&
1487 (!prev || bkey_iter_cmp(b, k, prev) > 0)) {
1488 prev = k;
1489 end = t->end_offset;
1490 }
1491 }
1492
1493 if (!prev)
1494 return NULL;
1495
1496 /*
1497 * We're manually memmoving instead of just calling sort() to ensure the
1498 * prev we picked ends up in slot 0 - sort won't necessarily put it
1499 * there because of duplicate deleted keys:
1500 */
1501 btree_node_iter_for_each(iter, set)
1502 if (set->end == end)
1503 goto found;
1504
1505 BUG_ON(set != &iter->data[__btree_node_iter_used(iter)]);
1506 found:
1507 BUG_ON(set >= iter->data + ARRAY_SIZE(iter->data));
1508
1509 memmove(&iter->data[1],
1510 &iter->data[0],
1511 (void *) set - (void *) &iter->data[0]);
1512
1513 iter->data[0].k = __btree_node_key_to_offset(b, prev);
1514 iter->data[0].end = end;
1515
1516 if (bch2_expensive_debug_checks)
1517 bch2_btree_node_iter_verify(iter, b);
1518 return prev;
1519 }
1520
1521 struct bkey_packed *bch2_btree_node_iter_prev(struct btree_node_iter *iter,
1522 struct btree *b)
1523 {
1524 struct bkey_packed *prev;
1525
1526 do {
1527 prev = bch2_btree_node_iter_prev_all(iter, b);
1528 } while (prev && bkey_deleted(prev));
1529
1530 return prev;
1531 }
1532
1533 struct bkey_s_c bch2_btree_node_iter_peek_unpack(struct btree_node_iter *iter,
1534 struct btree *b,
1535 struct bkey *u)
1536 {
1537 struct bkey_packed *k = bch2_btree_node_iter_peek(iter, b);
1538
1539 return k ? bkey_disassemble(b, k, u) : bkey_s_c_null;
1540 }
1541
1542 /* Mergesort */
1543
1544 void bch2_btree_keys_stats(const struct btree *b, struct bset_stats *stats)
1545 {
1546 for_each_bset_c(b, t) {
1547 enum bset_aux_tree_type type = bset_aux_tree_type(t);
1548 size_t j;
1549
1550 stats->sets[type].nr++;
1551 stats->sets[type].bytes += le16_to_cpu(bset(b, t)->u64s) *
1552 sizeof(u64);
1553
1554 if (bset_has_ro_aux_tree(t)) {
1555 stats->floats += t->size - 1;
1556
1557 for (j = 1; j < t->size; j++)
1558 stats->failed +=
1559 bkey_float(b, t, j)->exponent ==
1560 BFLOAT_FAILED;
1561 }
1562 }
1563 }
1564
1565 void bch2_bfloat_to_text(struct printbuf *out, struct btree *b,
1566 struct bkey_packed *k)
1567 {
1568 struct bset_tree *t = bch2_bkey_to_bset(b, k);
1569 struct bkey uk;
1570 unsigned j, inorder;
1571
1572 if (!bset_has_ro_aux_tree(t))
1573 return;
1574
1575 inorder = bkey_to_cacheline(b, t, k);
1576 if (!inorder || inorder >= t->size)
1577 return;
1578
1579 j = __inorder_to_eytzinger1(inorder, t->size - 1, t->extra);
1580 if (k != tree_to_bkey(b, t, j))
1581 return;
1582
1583 switch (bkey_float(b, t, j)->exponent) {
1584 case BFLOAT_FAILED:
1585 uk = bkey_unpack_key(b, k);
1586 prt_printf(out,
1587 " failed unpacked at depth %u\n"
1588 "\t",
1589 ilog2(j));
1590 bch2_bpos_to_text(out, uk.p);
1591 prt_printf(out, "\n");
1592 break;
1593 }
1594 }
1595
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.