xref: /linux/drivers/md/bcache/bset.c (revision 8e07e0e3964ca4e23ce7b68e2096fe660a888942)
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 #define pr_fmt(fmt) "bcache: %s() " fmt, __func__
10 
11 #include "util.h"
12 #include "bset.h"
13 
14 #include <linux/console.h>
15 #include <linux/sched/clock.h>
16 #include <linux/random.h>
17 #include <linux/prefetch.h>
18 
19 #ifdef CONFIG_BCACHE_DEBUG
20 
21 void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned int 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 
28 		pr_err("block %u key %u/%u: ", set,
29 		       (unsigned int) ((u64 *) k - i->d), i->keys);
30 
31 		if (b->ops->key_dump)
32 			b->ops->key_dump(b, k);
33 		else
34 			pr_cont("%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 			pr_err("Key skipped backwards\n");
40 	}
41 }
42 
43 void bch_dump_bucket(struct btree_keys *b)
44 {
45 	unsigned int 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 
54 int __bch_count_data(struct btree_keys *b)
55 {
56 	unsigned int 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 
66 void __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;
101 bug:
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 
111 static 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 
125 static inline void bch_btree_iter_next_check(struct btree_iter *iter) {}
126 
127 #endif
128 
129 /* Keylists */
130 
131 int __bch_keylist_realloc(struct keylist *l, unsigned int u64s)
132 {
133 	size_t oldsize = bch_keylist_nkeys(l);
134 	size_t newsize = oldsize + u64s;
135 	uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
136 	uint64_t *new_keys;
137 
138 	newsize = roundup_pow_of_two(newsize);
139 
140 	if (newsize <= KEYLIST_INLINE ||
141 	    roundup_pow_of_two(oldsize) == newsize)
142 		return 0;
143 
144 	new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
145 
146 	if (!new_keys)
147 		return -ENOMEM;
148 
149 	if (!old_keys)
150 		memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
151 
152 	l->keys_p = new_keys;
153 	l->top_p = new_keys + oldsize;
154 
155 	return 0;
156 }
157 
158 /* Pop the top key of keylist by pointing l->top to its previous key */
159 struct bkey *bch_keylist_pop(struct keylist *l)
160 {
161 	struct bkey *k = l->keys;
162 
163 	if (k == l->top)
164 		return NULL;
165 
166 	while (bkey_next(k) != l->top)
167 		k = bkey_next(k);
168 
169 	return l->top = k;
170 }
171 
172 /* Pop the bottom key of keylist and update l->top_p */
173 void bch_keylist_pop_front(struct keylist *l)
174 {
175 	l->top_p -= bkey_u64s(l->keys);
176 
177 	memmove(l->keys,
178 		bkey_next(l->keys),
179 		bch_keylist_bytes(l));
180 }
181 
182 /* Key/pointer manipulation */
183 
184 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
185 			      unsigned int i)
186 {
187 	BUG_ON(i > KEY_PTRS(src));
188 
189 	/* Only copy the header, key, and one pointer. */
190 	memcpy(dest, src, 2 * sizeof(uint64_t));
191 	dest->ptr[0] = src->ptr[i];
192 	SET_KEY_PTRS(dest, 1);
193 	/* We didn't copy the checksum so clear that bit. */
194 	SET_KEY_CSUM(dest, 0);
195 }
196 
197 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
198 {
199 	unsigned int i, len = 0;
200 
201 	if (bkey_cmp(where, &START_KEY(k)) <= 0)
202 		return false;
203 
204 	if (bkey_cmp(where, k) < 0)
205 		len = KEY_OFFSET(k) - KEY_OFFSET(where);
206 	else
207 		bkey_copy_key(k, where);
208 
209 	for (i = 0; i < KEY_PTRS(k); i++)
210 		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
211 
212 	BUG_ON(len > KEY_SIZE(k));
213 	SET_KEY_SIZE(k, len);
214 	return true;
215 }
216 
217 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
218 {
219 	unsigned int len = 0;
220 
221 	if (bkey_cmp(where, k) >= 0)
222 		return false;
223 
224 	BUG_ON(KEY_INODE(where) != KEY_INODE(k));
225 
226 	if (bkey_cmp(where, &START_KEY(k)) > 0)
227 		len = KEY_OFFSET(where) - KEY_START(k);
228 
229 	bkey_copy_key(k, where);
230 
231 	BUG_ON(len > KEY_SIZE(k));
232 	SET_KEY_SIZE(k, len);
233 	return true;
234 }
235 
236 /* Auxiliary search trees */
237 
238 /* 32 bits total: */
239 #define BKEY_MID_BITS		3
240 #define BKEY_EXPONENT_BITS	7
241 #define BKEY_MANTISSA_BITS	(32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS)
242 #define BKEY_MANTISSA_MASK	((1 << BKEY_MANTISSA_BITS) - 1)
243 
244 struct bkey_float {
245 	unsigned int	exponent:BKEY_EXPONENT_BITS;
246 	unsigned int	m:BKEY_MID_BITS;
247 	unsigned int	mantissa:BKEY_MANTISSA_BITS;
248 } __packed;
249 
250 /*
251  * BSET_CACHELINE was originally intended to match the hardware cacheline size -
252  * it used to be 64, but I realized the lookup code would touch slightly less
253  * memory if it was 128.
254  *
255  * It definites the number of bytes (in struct bset) per struct bkey_float in
256  * the auxiliar search tree - when we're done searching the bset_float tree we
257  * have this many bytes left that we do a linear search over.
258  *
259  * Since (after level 5) every level of the bset_tree is on a new cacheline,
260  * we're touching one fewer cacheline in the bset tree in exchange for one more
261  * cacheline in the linear search - but the linear search might stop before it
262  * gets to the second cacheline.
263  */
264 
265 #define BSET_CACHELINE		128
266 
267 /* Space required for the btree node keys */
268 static inline size_t btree_keys_bytes(struct btree_keys *b)
269 {
270 	return PAGE_SIZE << b->page_order;
271 }
272 
273 static inline size_t btree_keys_cachelines(struct btree_keys *b)
274 {
275 	return btree_keys_bytes(b) / BSET_CACHELINE;
276 }
277 
278 /* Space required for the auxiliary search trees */
279 static inline size_t bset_tree_bytes(struct btree_keys *b)
280 {
281 	return btree_keys_cachelines(b) * sizeof(struct bkey_float);
282 }
283 
284 /* Space required for the prev pointers */
285 static inline size_t bset_prev_bytes(struct btree_keys *b)
286 {
287 	return btree_keys_cachelines(b) * sizeof(uint8_t);
288 }
289 
290 /* Memory allocation */
291 
292 void bch_btree_keys_free(struct btree_keys *b)
293 {
294 	struct bset_tree *t = b->set;
295 
296 	if (bset_prev_bytes(b) < PAGE_SIZE)
297 		kfree(t->prev);
298 	else
299 		free_pages((unsigned long) t->prev,
300 			   get_order(bset_prev_bytes(b)));
301 
302 	if (bset_tree_bytes(b) < PAGE_SIZE)
303 		kfree(t->tree);
304 	else
305 		free_pages((unsigned long) t->tree,
306 			   get_order(bset_tree_bytes(b)));
307 
308 	free_pages((unsigned long) t->data, b->page_order);
309 
310 	t->prev = NULL;
311 	t->tree = NULL;
312 	t->data = NULL;
313 }
314 
315 int bch_btree_keys_alloc(struct btree_keys *b,
316 			 unsigned int page_order,
317 			 gfp_t gfp)
318 {
319 	struct bset_tree *t = b->set;
320 
321 	BUG_ON(t->data);
322 
323 	b->page_order = page_order;
324 
325 	t->data = (void *) __get_free_pages(__GFP_COMP|gfp, b->page_order);
326 	if (!t->data)
327 		goto err;
328 
329 	t->tree = bset_tree_bytes(b) < PAGE_SIZE
330 		? kmalloc(bset_tree_bytes(b), gfp)
331 		: (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
332 	if (!t->tree)
333 		goto err;
334 
335 	t->prev = bset_prev_bytes(b) < PAGE_SIZE
336 		? kmalloc(bset_prev_bytes(b), gfp)
337 		: (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
338 	if (!t->prev)
339 		goto err;
340 
341 	return 0;
342 err:
343 	bch_btree_keys_free(b);
344 	return -ENOMEM;
345 }
346 
347 void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
348 			 bool *expensive_debug_checks)
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 	/*
356 	 * struct btree_keys in embedded in struct btree, and struct
357 	 * bset_tree is embedded into struct btree_keys. They are all
358 	 * initialized as 0 by kzalloc() in mca_bucket_alloc(), and
359 	 * b->set[0].data is allocated in bch_btree_keys_alloc(), so we
360 	 * don't have to initiate b->set[].size and b->set[].data here
361 	 * any more.
362 	 */
363 }
364 
365 /* Binary tree stuff for auxiliary search trees */
366 
367 /*
368  * return array index next to j when does in-order traverse
369  * of a binary tree which is stored in a linear array
370  */
371 static unsigned int inorder_next(unsigned int j, unsigned int size)
372 {
373 	if (j * 2 + 1 < size) {
374 		j = j * 2 + 1;
375 
376 		while (j * 2 < size)
377 			j *= 2;
378 	} else
379 		j >>= ffz(j) + 1;
380 
381 	return j;
382 }
383 
384 /*
385  * return array index previous to j when does in-order traverse
386  * of a binary tree which is stored in a linear array
387  */
388 static unsigned int inorder_prev(unsigned int j, unsigned int size)
389 {
390 	if (j * 2 < size) {
391 		j = j * 2;
392 
393 		while (j * 2 + 1 < size)
394 			j = j * 2 + 1;
395 	} else
396 		j >>= ffs(j);
397 
398 	return j;
399 }
400 
401 /*
402  * I have no idea why this code works... and I'm the one who wrote it
403  *
404  * However, I do know what it does:
405  * Given a binary tree constructed in an array (i.e. how you normally implement
406  * a heap), it converts a node in the tree - referenced by array index - to the
407  * index it would have if you did an inorder traversal.
408  *
409  * Also tested for every j, size up to size somewhere around 6 million.
410  *
411  * The binary tree starts at array index 1, not 0
412  * extra is a function of size:
413  *   extra = (size - rounddown_pow_of_two(size - 1)) << 1;
414  */
415 static unsigned int __to_inorder(unsigned int j,
416 				  unsigned int size,
417 				  unsigned int extra)
418 {
419 	unsigned int b = fls(j);
420 	unsigned int shift = fls(size - 1) - b;
421 
422 	j  ^= 1U << (b - 1);
423 	j <<= 1;
424 	j  |= 1;
425 	j <<= shift;
426 
427 	if (j > extra)
428 		j -= (j - extra) >> 1;
429 
430 	return j;
431 }
432 
433 /*
434  * Return the cacheline index in bset_tree->data, where j is index
435  * from a linear array which stores the auxiliar binary tree
436  */
437 static unsigned int to_inorder(unsigned int j, struct bset_tree *t)
438 {
439 	return __to_inorder(j, t->size, t->extra);
440 }
441 
442 static unsigned int __inorder_to_tree(unsigned int j,
443 				      unsigned int size,
444 				      unsigned int extra)
445 {
446 	unsigned int shift;
447 
448 	if (j > extra)
449 		j += j - extra;
450 
451 	shift = ffs(j);
452 
453 	j >>= shift;
454 	j  |= roundup_pow_of_two(size) >> shift;
455 
456 	return j;
457 }
458 
459 /*
460  * Return an index from a linear array which stores the auxiliar binary
461  * tree, j is the cacheline index of t->data.
462  */
463 static unsigned int inorder_to_tree(unsigned int j, struct bset_tree *t)
464 {
465 	return __inorder_to_tree(j, t->size, t->extra);
466 }
467 
468 #if 0
469 void inorder_test(void)
470 {
471 	unsigned long done = 0;
472 	ktime_t start = ktime_get();
473 
474 	for (unsigned int size = 2;
475 	     size < 65536000;
476 	     size++) {
477 		unsigned int extra =
478 			(size - rounddown_pow_of_two(size - 1)) << 1;
479 		unsigned int i = 1, j = rounddown_pow_of_two(size - 1);
480 
481 		if (!(size % 4096))
482 			pr_notice("loop %u, %llu per us\n", size,
483 			       done / ktime_us_delta(ktime_get(), start));
484 
485 		while (1) {
486 			if (__inorder_to_tree(i, size, extra) != j)
487 				panic("size %10u j %10u i %10u", size, j, i);
488 
489 			if (__to_inorder(j, size, extra) != i)
490 				panic("size %10u j %10u i %10u", size, j, i);
491 
492 			if (j == rounddown_pow_of_two(size) - 1)
493 				break;
494 
495 			BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
496 
497 			j = inorder_next(j, size);
498 			i++;
499 		}
500 
501 		done += size - 1;
502 	}
503 }
504 #endif
505 
506 /*
507  * Cacheline/offset <-> bkey pointer arithmetic:
508  *
509  * t->tree is a binary search tree in an array; each node corresponds to a key
510  * in one cacheline in t->set (BSET_CACHELINE bytes).
511  *
512  * This means we don't have to store the full index of the key that a node in
513  * the binary tree points to; to_inorder() gives us the cacheline, and then
514  * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
515  *
516  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
517  * make this work.
518  *
519  * To construct the bfloat for an arbitrary key we need to know what the key
520  * immediately preceding it is: we have to check if the two keys differ in the
521  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
522  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
523  */
524 
525 static struct bkey *cacheline_to_bkey(struct bset_tree *t,
526 				      unsigned int cacheline,
527 				      unsigned int offset)
528 {
529 	return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
530 }
531 
532 static unsigned int bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
533 {
534 	return ((void *) k - (void *) t->data) / BSET_CACHELINE;
535 }
536 
537 static unsigned int bkey_to_cacheline_offset(struct bset_tree *t,
538 					 unsigned int cacheline,
539 					 struct bkey *k)
540 {
541 	return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0);
542 }
543 
544 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned int j)
545 {
546 	return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
547 }
548 
549 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned int j)
550 {
551 	return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
552 }
553 
554 /*
555  * For the write set - the one we're currently inserting keys into - we don't
556  * maintain a full search tree, we just keep a simple lookup table in t->prev.
557  */
558 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned int cacheline)
559 {
560 	return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
561 }
562 
563 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
564 {
565 	low >>= shift;
566 	low  |= (high << 1) << (63U - shift);
567 	return low;
568 }
569 
570 /*
571  * Calculate mantissa value for struct bkey_float.
572  * If most significant bit of f->exponent is not set, then
573  *  - f->exponent >> 6 is 0
574  *  - p[0] points to bkey->low
575  *  - p[-1] borrows bits from KEY_INODE() of bkey->high
576  * if most isgnificant bits of f->exponent is set, then
577  *  - f->exponent >> 6 is 1
578  *  - p[0] points to bits from KEY_INODE() of bkey->high
579  *  - p[-1] points to other bits from KEY_INODE() of
580  *    bkey->high too.
581  * See make_bfloat() to check when most significant bit of f->exponent
582  * is set or not.
583  */
584 static inline unsigned int bfloat_mantissa(const struct bkey *k,
585 				       struct bkey_float *f)
586 {
587 	const uint64_t *p = &k->low - (f->exponent >> 6);
588 
589 	return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
590 }
591 
592 static void make_bfloat(struct bset_tree *t, unsigned int j)
593 {
594 	struct bkey_float *f = &t->tree[j];
595 	struct bkey *m = tree_to_bkey(t, j);
596 	struct bkey *p = tree_to_prev_bkey(t, j);
597 
598 	struct bkey *l = is_power_of_2(j)
599 		? t->data->start
600 		: tree_to_prev_bkey(t, j >> ffs(j));
601 
602 	struct bkey *r = is_power_of_2(j + 1)
603 		? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
604 		: tree_to_bkey(t, j >> (ffz(j) + 1));
605 
606 	BUG_ON(m < l || m > r);
607 	BUG_ON(bkey_next(p) != m);
608 
609 	/*
610 	 * If l and r have different KEY_INODE values (different backing
611 	 * device), f->exponent records how many least significant bits
612 	 * are different in KEY_INODE values and sets most significant
613 	 * bits to 1 (by +64).
614 	 * If l and r have same KEY_INODE value, f->exponent records
615 	 * how many different bits in least significant bits of bkey->low.
616 	 * See bfloat_mantiss() how the most significant bit of
617 	 * f->exponent is used to calculate bfloat mantissa value.
618 	 */
619 	if (KEY_INODE(l) != KEY_INODE(r))
620 		f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
621 	else
622 		f->exponent = fls64(r->low ^ l->low);
623 
624 	f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
625 
626 	/*
627 	 * Setting f->exponent = 127 flags this node as failed, and causes the
628 	 * lookup code to fall back to comparing against the original key.
629 	 */
630 
631 	if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
632 		f->mantissa = bfloat_mantissa(m, f) - 1;
633 	else
634 		f->exponent = 127;
635 }
636 
637 static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t)
638 {
639 	if (t != b->set) {
640 		unsigned int j = roundup(t[-1].size,
641 				     64 / sizeof(struct bkey_float));
642 
643 		t->tree = t[-1].tree + j;
644 		t->prev = t[-1].prev + j;
645 	}
646 
647 	while (t < b->set + MAX_BSETS)
648 		t++->size = 0;
649 }
650 
651 static void bch_bset_build_unwritten_tree(struct btree_keys *b)
652 {
653 	struct bset_tree *t = bset_tree_last(b);
654 
655 	BUG_ON(b->last_set_unwritten);
656 	b->last_set_unwritten = 1;
657 
658 	bset_alloc_tree(b, t);
659 
660 	if (t->tree != b->set->tree + btree_keys_cachelines(b)) {
661 		t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start);
662 		t->size = 1;
663 	}
664 }
665 
666 void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic)
667 {
668 	if (i != b->set->data) {
669 		b->set[++b->nsets].data = i;
670 		i->seq = b->set->data->seq;
671 	} else
672 		get_random_bytes(&i->seq, sizeof(uint64_t));
673 
674 	i->magic	= magic;
675 	i->version	= 0;
676 	i->keys		= 0;
677 
678 	bch_bset_build_unwritten_tree(b);
679 }
680 
681 /*
682  * Build auxiliary binary tree 'struct bset_tree *t', this tree is used to
683  * accelerate bkey search in a btree node (pointed by bset_tree->data in
684  * memory). After search in the auxiliar tree by calling bset_search_tree(),
685  * a struct bset_search_iter is returned which indicates range [l, r] from
686  * bset_tree->data where the searching bkey might be inside. Then a followed
687  * linear comparison does the exact search, see __bch_bset_search() for how
688  * the auxiliary tree is used.
689  */
690 void bch_bset_build_written_tree(struct btree_keys *b)
691 {
692 	struct bset_tree *t = bset_tree_last(b);
693 	struct bkey *prev = NULL, *k = t->data->start;
694 	unsigned int j, cacheline = 1;
695 
696 	b->last_set_unwritten = 0;
697 
698 	bset_alloc_tree(b, t);
699 
700 	t->size = min_t(unsigned int,
701 			bkey_to_cacheline(t, bset_bkey_last(t->data)),
702 			b->set->tree + btree_keys_cachelines(b) - t->tree);
703 
704 	if (t->size < 2) {
705 		t->size = 0;
706 		return;
707 	}
708 
709 	t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
710 
711 	/* First we figure out where the first key in each cacheline is */
712 	for (j = inorder_next(0, t->size);
713 	     j;
714 	     j = inorder_next(j, t->size)) {
715 		while (bkey_to_cacheline(t, k) < cacheline) {
716 			prev = k;
717 			k = bkey_next(k);
718 		}
719 
720 		t->prev[j] = bkey_u64s(prev);
721 		t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k);
722 	}
723 
724 	while (bkey_next(k) != bset_bkey_last(t->data))
725 		k = bkey_next(k);
726 
727 	t->end = *k;
728 
729 	/* Then we build the tree */
730 	for (j = inorder_next(0, t->size);
731 	     j;
732 	     j = inorder_next(j, t->size))
733 		make_bfloat(t, j);
734 }
735 
736 /* Insert */
737 
738 void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k)
739 {
740 	struct bset_tree *t;
741 	unsigned int inorder, j = 1;
742 
743 	for (t = b->set; t <= bset_tree_last(b); t++)
744 		if (k < bset_bkey_last(t->data))
745 			goto found_set;
746 
747 	BUG();
748 found_set:
749 	if (!t->size || !bset_written(b, t))
750 		return;
751 
752 	inorder = bkey_to_cacheline(t, k);
753 
754 	if (k == t->data->start)
755 		goto fix_left;
756 
757 	if (bkey_next(k) == bset_bkey_last(t->data)) {
758 		t->end = *k;
759 		goto fix_right;
760 	}
761 
762 	j = inorder_to_tree(inorder, t);
763 
764 	if (j &&
765 	    j < t->size &&
766 	    k == tree_to_bkey(t, j))
767 fix_left:	do {
768 			make_bfloat(t, j);
769 			j = j * 2;
770 		} while (j < t->size);
771 
772 	j = inorder_to_tree(inorder + 1, t);
773 
774 	if (j &&
775 	    j < t->size &&
776 	    k == tree_to_prev_bkey(t, j))
777 fix_right:	do {
778 			make_bfloat(t, j);
779 			j = j * 2 + 1;
780 		} while (j < t->size);
781 }
782 
783 static void bch_bset_fix_lookup_table(struct btree_keys *b,
784 				      struct bset_tree *t,
785 				      struct bkey *k)
786 {
787 	unsigned int shift = bkey_u64s(k);
788 	unsigned int j = bkey_to_cacheline(t, k);
789 
790 	/* We're getting called from btree_split() or btree_gc, just bail out */
791 	if (!t->size)
792 		return;
793 
794 	/*
795 	 * k is the key we just inserted; we need to find the entry in the
796 	 * lookup table for the first key that is strictly greater than k:
797 	 * it's either k's cacheline or the next one
798 	 */
799 	while (j < t->size &&
800 	       table_to_bkey(t, j) <= k)
801 		j++;
802 
803 	/*
804 	 * Adjust all the lookup table entries, and find a new key for any that
805 	 * have gotten too big
806 	 */
807 	for (; j < t->size; j++) {
808 		t->prev[j] += shift;
809 
810 		if (t->prev[j] > 7) {
811 			k = table_to_bkey(t, j - 1);
812 
813 			while (k < cacheline_to_bkey(t, j, 0))
814 				k = bkey_next(k);
815 
816 			t->prev[j] = bkey_to_cacheline_offset(t, j, k);
817 		}
818 	}
819 
820 	if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree)
821 		return;
822 
823 	/* Possibly add a new entry to the end of the lookup table */
824 
825 	for (k = table_to_bkey(t, t->size - 1);
826 	     k != bset_bkey_last(t->data);
827 	     k = bkey_next(k))
828 		if (t->size == bkey_to_cacheline(t, k)) {
829 			t->prev[t->size] =
830 				bkey_to_cacheline_offset(t, t->size, k);
831 			t->size++;
832 		}
833 }
834 
835 /*
836  * Tries to merge l and r: l should be lower than r
837  * Returns true if we were able to merge. If we did merge, l will be the merged
838  * key, r will be untouched.
839  */
840 bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r)
841 {
842 	if (!b->ops->key_merge)
843 		return false;
844 
845 	/*
846 	 * Generic header checks
847 	 * Assumes left and right are in order
848 	 * Left and right must be exactly aligned
849 	 */
850 	if (!bch_bkey_equal_header(l, r) ||
851 	     bkey_cmp(l, &START_KEY(r)))
852 		return false;
853 
854 	return b->ops->key_merge(b, l, r);
855 }
856 
857 void bch_bset_insert(struct btree_keys *b, struct bkey *where,
858 		     struct bkey *insert)
859 {
860 	struct bset_tree *t = bset_tree_last(b);
861 
862 	BUG_ON(!b->last_set_unwritten);
863 	BUG_ON(bset_byte_offset(b, t->data) +
864 	       __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
865 	       PAGE_SIZE << b->page_order);
866 
867 	memmove((uint64_t *) where + bkey_u64s(insert),
868 		where,
869 		(void *) bset_bkey_last(t->data) - (void *) where);
870 
871 	t->data->keys += bkey_u64s(insert);
872 	bkey_copy(where, insert);
873 	bch_bset_fix_lookup_table(b, t, where);
874 }
875 
876 unsigned int bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
877 			      struct bkey *replace_key)
878 {
879 	unsigned int status = BTREE_INSERT_STATUS_NO_INSERT;
880 	struct bset *i = bset_tree_last(b)->data;
881 	struct bkey *m, *prev = NULL;
882 	struct btree_iter iter;
883 	struct bkey preceding_key_on_stack = ZERO_KEY;
884 	struct bkey *preceding_key_p = &preceding_key_on_stack;
885 
886 	BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
887 
888 	/*
889 	 * If k has preceding key, preceding_key_p will be set to address
890 	 *  of k's preceding key; otherwise preceding_key_p will be set
891 	 * to NULL inside preceding_key().
892 	 */
893 	if (b->ops->is_extents)
894 		preceding_key(&START_KEY(k), &preceding_key_p);
895 	else
896 		preceding_key(k, &preceding_key_p);
897 
898 	m = bch_btree_iter_init(b, &iter, preceding_key_p);
899 
900 	if (b->ops->insert_fixup(b, k, &iter, replace_key))
901 		return status;
902 
903 	status = BTREE_INSERT_STATUS_INSERT;
904 
905 	while (m != bset_bkey_last(i) &&
906 	       bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0) {
907 		prev = m;
908 		m = bkey_next(m);
909 	}
910 
911 	/* prev is in the tree, if we merge we're done */
912 	status = BTREE_INSERT_STATUS_BACK_MERGE;
913 	if (prev &&
914 	    bch_bkey_try_merge(b, prev, k))
915 		goto merged;
916 #if 0
917 	status = BTREE_INSERT_STATUS_OVERWROTE;
918 	if (m != bset_bkey_last(i) &&
919 	    KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
920 		goto copy;
921 #endif
922 	status = BTREE_INSERT_STATUS_FRONT_MERGE;
923 	if (m != bset_bkey_last(i) &&
924 	    bch_bkey_try_merge(b, k, m))
925 		goto copy;
926 
927 	bch_bset_insert(b, m, k);
928 copy:	bkey_copy(m, k);
929 merged:
930 	return status;
931 }
932 
933 /* Lookup */
934 
935 struct bset_search_iter {
936 	struct bkey *l, *r;
937 };
938 
939 static struct bset_search_iter bset_search_write_set(struct bset_tree *t,
940 						     const struct bkey *search)
941 {
942 	unsigned int li = 0, ri = t->size;
943 
944 	while (li + 1 != ri) {
945 		unsigned int m = (li + ri) >> 1;
946 
947 		if (bkey_cmp(table_to_bkey(t, m), search) > 0)
948 			ri = m;
949 		else
950 			li = m;
951 	}
952 
953 	return (struct bset_search_iter) {
954 		table_to_bkey(t, li),
955 		ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
956 	};
957 }
958 
959 static struct bset_search_iter bset_search_tree(struct bset_tree *t,
960 						const struct bkey *search)
961 {
962 	struct bkey *l, *r;
963 	struct bkey_float *f;
964 	unsigned int inorder, j, n = 1;
965 
966 	do {
967 		unsigned int p = n << 4;
968 
969 		if (p < t->size)
970 			prefetch(&t->tree[p]);
971 
972 		j = n;
973 		f = &t->tree[j];
974 
975 		if (likely(f->exponent != 127)) {
976 			if (f->mantissa >= bfloat_mantissa(search, f))
977 				n = j * 2;
978 			else
979 				n = j * 2 + 1;
980 		} else {
981 			if (bkey_cmp(tree_to_bkey(t, j), search) > 0)
982 				n = j * 2;
983 			else
984 				n = j * 2 + 1;
985 		}
986 	} while (n < t->size);
987 
988 	inorder = to_inorder(j, t);
989 
990 	/*
991 	 * n would have been the node we recursed to - the low bit tells us if
992 	 * we recursed left or recursed right.
993 	 */
994 	if (n & 1) {
995 		l = cacheline_to_bkey(t, inorder, f->m);
996 
997 		if (++inorder != t->size) {
998 			f = &t->tree[inorder_next(j, t->size)];
999 			r = cacheline_to_bkey(t, inorder, f->m);
1000 		} else
1001 			r = bset_bkey_last(t->data);
1002 	} else {
1003 		r = cacheline_to_bkey(t, inorder, f->m);
1004 
1005 		if (--inorder) {
1006 			f = &t->tree[inorder_prev(j, t->size)];
1007 			l = cacheline_to_bkey(t, inorder, f->m);
1008 		} else
1009 			l = t->data->start;
1010 	}
1011 
1012 	return (struct bset_search_iter) {l, r};
1013 }
1014 
1015 struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
1016 			       const struct bkey *search)
1017 {
1018 	struct bset_search_iter i;
1019 
1020 	/*
1021 	 * First, we search for a cacheline, then lastly we do a linear search
1022 	 * within that cacheline.
1023 	 *
1024 	 * To search for the cacheline, there's three different possibilities:
1025 	 *  * The set is too small to have a search tree, so we just do a linear
1026 	 *    search over the whole set.
1027 	 *  * The set is the one we're currently inserting into; keeping a full
1028 	 *    auxiliary search tree up to date would be too expensive, so we
1029 	 *    use a much simpler lookup table to do a binary search -
1030 	 *    bset_search_write_set().
1031 	 *  * Or we use the auxiliary search tree we constructed earlier -
1032 	 *    bset_search_tree()
1033 	 */
1034 
1035 	if (unlikely(!t->size)) {
1036 		i.l = t->data->start;
1037 		i.r = bset_bkey_last(t->data);
1038 	} else if (bset_written(b, t)) {
1039 		/*
1040 		 * Each node in the auxiliary search tree covers a certain range
1041 		 * of bits, and keys above and below the set it covers might
1042 		 * differ outside those bits - so we have to special case the
1043 		 * start and end - handle that here:
1044 		 */
1045 
1046 		if (unlikely(bkey_cmp(search, &t->end) >= 0))
1047 			return bset_bkey_last(t->data);
1048 
1049 		if (unlikely(bkey_cmp(search, t->data->start) < 0))
1050 			return t->data->start;
1051 
1052 		i = bset_search_tree(t, search);
1053 	} else {
1054 		BUG_ON(!b->nsets &&
1055 		       t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
1056 
1057 		i = bset_search_write_set(t, search);
1058 	}
1059 
1060 	if (btree_keys_expensive_checks(b)) {
1061 		BUG_ON(bset_written(b, t) &&
1062 		       i.l != t->data->start &&
1063 		       bkey_cmp(tree_to_prev_bkey(t,
1064 			  inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
1065 				search) > 0);
1066 
1067 		BUG_ON(i.r != bset_bkey_last(t->data) &&
1068 		       bkey_cmp(i.r, search) <= 0);
1069 	}
1070 
1071 	while (likely(i.l != i.r) &&
1072 	       bkey_cmp(i.l, search) <= 0)
1073 		i.l = bkey_next(i.l);
1074 
1075 	return i.l;
1076 }
1077 
1078 /* Btree iterator */
1079 
1080 typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
1081 				 struct btree_iter_set);
1082 
1083 static inline bool btree_iter_cmp(struct btree_iter_set l,
1084 				  struct btree_iter_set r)
1085 {
1086 	return bkey_cmp(l.k, r.k) > 0;
1087 }
1088 
1089 static inline bool btree_iter_end(struct btree_iter *iter)
1090 {
1091 	return !iter->used;
1092 }
1093 
1094 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
1095 			 struct bkey *end)
1096 {
1097 	if (k != end)
1098 		BUG_ON(!heap_add(iter,
1099 				 ((struct btree_iter_set) { k, end }),
1100 				 btree_iter_cmp));
1101 }
1102 
1103 static struct bkey *__bch_btree_iter_init(struct btree_keys *b,
1104 					  struct btree_iter *iter,
1105 					  struct bkey *search,
1106 					  struct bset_tree *start)
1107 {
1108 	struct bkey *ret = NULL;
1109 
1110 	iter->size = ARRAY_SIZE(iter->data);
1111 	iter->used = 0;
1112 
1113 #ifdef CONFIG_BCACHE_DEBUG
1114 	iter->b = b;
1115 #endif
1116 
1117 	for (; start <= bset_tree_last(b); start++) {
1118 		ret = bch_bset_search(b, start, search);
1119 		bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
1120 	}
1121 
1122 	return ret;
1123 }
1124 
1125 struct bkey *bch_btree_iter_init(struct btree_keys *b,
1126 				 struct btree_iter *iter,
1127 				 struct bkey *search)
1128 {
1129 	return __bch_btree_iter_init(b, iter, search, b->set);
1130 }
1131 
1132 static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
1133 						 btree_iter_cmp_fn *cmp)
1134 {
1135 	struct btree_iter_set b __maybe_unused;
1136 	struct bkey *ret = NULL;
1137 
1138 	if (!btree_iter_end(iter)) {
1139 		bch_btree_iter_next_check(iter);
1140 
1141 		ret = iter->data->k;
1142 		iter->data->k = bkey_next(iter->data->k);
1143 
1144 		if (iter->data->k > iter->data->end) {
1145 			WARN_ONCE(1, "bset was corrupt!\n");
1146 			iter->data->k = iter->data->end;
1147 		}
1148 
1149 		if (iter->data->k == iter->data->end)
1150 			heap_pop(iter, b, cmp);
1151 		else
1152 			heap_sift(iter, 0, cmp);
1153 	}
1154 
1155 	return ret;
1156 }
1157 
1158 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
1159 {
1160 	return __bch_btree_iter_next(iter, btree_iter_cmp);
1161 
1162 }
1163 
1164 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
1165 					struct btree_keys *b, ptr_filter_fn fn)
1166 {
1167 	struct bkey *ret;
1168 
1169 	do {
1170 		ret = bch_btree_iter_next(iter);
1171 	} while (ret && fn(b, ret));
1172 
1173 	return ret;
1174 }
1175 
1176 /* Mergesort */
1177 
1178 void bch_bset_sort_state_free(struct bset_sort_state *state)
1179 {
1180 	mempool_exit(&state->pool);
1181 }
1182 
1183 int bch_bset_sort_state_init(struct bset_sort_state *state,
1184 			     unsigned int page_order)
1185 {
1186 	spin_lock_init(&state->time.lock);
1187 
1188 	state->page_order = page_order;
1189 	state->crit_factor = int_sqrt(1 << page_order);
1190 
1191 	return mempool_init_page_pool(&state->pool, 1, page_order);
1192 }
1193 
1194 static void btree_mergesort(struct btree_keys *b, struct bset *out,
1195 			    struct btree_iter *iter,
1196 			    bool fixup, bool remove_stale)
1197 {
1198 	int i;
1199 	struct bkey *k, *last = NULL;
1200 	BKEY_PADDED(k) tmp;
1201 	bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale
1202 		? bch_ptr_bad
1203 		: bch_ptr_invalid;
1204 
1205 	/* Heapify the iterator, using our comparison function */
1206 	for (i = iter->used / 2 - 1; i >= 0; --i)
1207 		heap_sift(iter, i, b->ops->sort_cmp);
1208 
1209 	while (!btree_iter_end(iter)) {
1210 		if (b->ops->sort_fixup && fixup)
1211 			k = b->ops->sort_fixup(iter, &tmp.k);
1212 		else
1213 			k = NULL;
1214 
1215 		if (!k)
1216 			k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
1217 
1218 		if (bad(b, k))
1219 			continue;
1220 
1221 		if (!last) {
1222 			last = out->start;
1223 			bkey_copy(last, k);
1224 		} else if (!bch_bkey_try_merge(b, last, k)) {
1225 			last = bkey_next(last);
1226 			bkey_copy(last, k);
1227 		}
1228 	}
1229 
1230 	out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1231 
1232 	pr_debug("sorted %i keys\n", out->keys);
1233 }
1234 
1235 static void __btree_sort(struct btree_keys *b, struct btree_iter *iter,
1236 			 unsigned int start, unsigned int order, bool fixup,
1237 			 struct bset_sort_state *state)
1238 {
1239 	uint64_t start_time;
1240 	bool used_mempool = false;
1241 	struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT,
1242 						     order);
1243 	if (!out) {
1244 		struct page *outp;
1245 
1246 		BUG_ON(order > state->page_order);
1247 
1248 		outp = mempool_alloc(&state->pool, GFP_NOIO);
1249 		out = page_address(outp);
1250 		used_mempool = true;
1251 		order = state->page_order;
1252 	}
1253 
1254 	start_time = local_clock();
1255 
1256 	btree_mergesort(b, out, iter, fixup, false);
1257 	b->nsets = start;
1258 
1259 	if (!start && order == b->page_order) {
1260 		/*
1261 		 * Our temporary buffer is the same size as the btree node's
1262 		 * buffer, we can just swap buffers instead of doing a big
1263 		 * memcpy()
1264 		 *
1265 		 * Don't worry event 'out' is allocated from mempool, it can
1266 		 * still be swapped here. Because state->pool is a page mempool
1267 		 * created by mempool_init_page_pool(), which allocates
1268 		 * pages by alloc_pages() indeed.
1269 		 */
1270 
1271 		out->magic	= b->set->data->magic;
1272 		out->seq	= b->set->data->seq;
1273 		out->version	= b->set->data->version;
1274 		swap(out, b->set->data);
1275 	} else {
1276 		b->set[start].data->keys = out->keys;
1277 		memcpy(b->set[start].data->start, out->start,
1278 		       (void *) bset_bkey_last(out) - (void *) out->start);
1279 	}
1280 
1281 	if (used_mempool)
1282 		mempool_free(virt_to_page(out), &state->pool);
1283 	else
1284 		free_pages((unsigned long) out, order);
1285 
1286 	bch_bset_build_written_tree(b);
1287 
1288 	if (!start)
1289 		bch_time_stats_update(&state->time, start_time);
1290 }
1291 
1292 void bch_btree_sort_partial(struct btree_keys *b, unsigned int start,
1293 			    struct bset_sort_state *state)
1294 {
1295 	size_t order = b->page_order, keys = 0;
1296 	struct btree_iter iter;
1297 	int oldsize = bch_count_data(b);
1298 
1299 	__bch_btree_iter_init(b, &iter, NULL, &b->set[start]);
1300 
1301 	if (start) {
1302 		unsigned int i;
1303 
1304 		for (i = start; i <= b->nsets; i++)
1305 			keys += b->set[i].data->keys;
1306 
1307 		order = get_order(__set_bytes(b->set->data, keys));
1308 	}
1309 
1310 	__btree_sort(b, &iter, start, order, false, state);
1311 
1312 	EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize);
1313 }
1314 
1315 void bch_btree_sort_and_fix_extents(struct btree_keys *b,
1316 				    struct btree_iter *iter,
1317 				    struct bset_sort_state *state)
1318 {
1319 	__btree_sort(b, iter, 0, b->page_order, true, state);
1320 }
1321 
1322 void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
1323 			 struct bset_sort_state *state)
1324 {
1325 	uint64_t start_time = local_clock();
1326 	struct btree_iter iter;
1327 
1328 	bch_btree_iter_init(b, &iter, NULL);
1329 
1330 	btree_mergesort(b, new->set->data, &iter, false, true);
1331 
1332 	bch_time_stats_update(&state->time, start_time);
1333 
1334 	new->set->size = 0; // XXX: why?
1335 }
1336 
1337 #define SORT_CRIT	(4096 / sizeof(uint64_t))
1338 
1339 void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state)
1340 {
1341 	unsigned int crit = SORT_CRIT;
1342 	int i;
1343 
1344 	/* Don't sort if nothing to do */
1345 	if (!b->nsets)
1346 		goto out;
1347 
1348 	for (i = b->nsets - 1; i >= 0; --i) {
1349 		crit *= state->crit_factor;
1350 
1351 		if (b->set[i].data->keys < crit) {
1352 			bch_btree_sort_partial(b, i, state);
1353 			return;
1354 		}
1355 	}
1356 
1357 	/* Sort if we'd overflow */
1358 	if (b->nsets + 1 == MAX_BSETS) {
1359 		bch_btree_sort(b, state);
1360 		return;
1361 	}
1362 
1363 out:
1364 	bch_bset_build_written_tree(b);
1365 }
1366 
1367 void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats)
1368 {
1369 	unsigned int i;
1370 
1371 	for (i = 0; i <= b->nsets; i++) {
1372 		struct bset_tree *t = &b->set[i];
1373 		size_t bytes = t->data->keys * sizeof(uint64_t);
1374 		size_t j;
1375 
1376 		if (bset_written(b, t)) {
1377 			stats->sets_written++;
1378 			stats->bytes_written += bytes;
1379 
1380 			stats->floats += t->size - 1;
1381 
1382 			for (j = 1; j < t->size; j++)
1383 				if (t->tree[j].exponent == 127)
1384 					stats->failed++;
1385 		} else {
1386 			stats->sets_unwritten++;
1387 			stats->bytes_unwritten += bytes;
1388 		}
1389 	}
1390 }
1391