xref: /openbmc/linux/drivers/md/bcache/bset.c (revision 37be287c)
1 /*
2  * Code for working with individual keys, and sorted sets of keys with in a
3  * btree node
4  *
5  * Copyright 2012 Google, Inc.
6  */
7 
8 #define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__
9 
10 #include "util.h"
11 #include "bset.h"
12 
13 #include <linux/console.h>
14 #include <linux/random.h>
15 #include <linux/prefetch.h>
16 
17 #ifdef CONFIG_BCACHE_DEBUG
18 
19 void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned set)
20 {
21 	struct bkey *k, *next;
22 
23 	for (k = i->start; k < bset_bkey_last(i); k = next) {
24 		next = bkey_next(k);
25 
26 		printk(KERN_ERR "block %u key %zi/%u: ", set,
27 		       (uint64_t *) k - i->d, i->keys);
28 
29 		if (b->ops->key_dump)
30 			b->ops->key_dump(b, k);
31 		else
32 			printk("%llu:%llu\n", KEY_INODE(k), KEY_OFFSET(k));
33 
34 		if (next < bset_bkey_last(i) &&
35 		    bkey_cmp(k, b->ops->is_extents ?
36 			     &START_KEY(next) : next) > 0)
37 			printk(KERN_ERR "Key skipped backwards\n");
38 	}
39 }
40 
41 void bch_dump_bucket(struct btree_keys *b)
42 {
43 	unsigned i;
44 
45 	console_lock();
46 	for (i = 0; i <= b->nsets; i++)
47 		bch_dump_bset(b, b->set[i].data,
48 			      bset_sector_offset(b, b->set[i].data));
49 	console_unlock();
50 }
51 
52 int __bch_count_data(struct btree_keys *b)
53 {
54 	unsigned ret = 0;
55 	struct btree_iter iter;
56 	struct bkey *k;
57 
58 	if (b->ops->is_extents)
59 		for_each_key(b, k, &iter)
60 			ret += KEY_SIZE(k);
61 	return ret;
62 }
63 
64 void __bch_check_keys(struct btree_keys *b, const char *fmt, ...)
65 {
66 	va_list args;
67 	struct bkey *k, *p = NULL;
68 	struct btree_iter iter;
69 	const char *err;
70 
71 	for_each_key(b, k, &iter) {
72 		if (b->ops->is_extents) {
73 			err = "Keys out of order";
74 			if (p && bkey_cmp(&START_KEY(p), &START_KEY(k)) > 0)
75 				goto bug;
76 
77 			if (bch_ptr_invalid(b, k))
78 				continue;
79 
80 			err =  "Overlapping keys";
81 			if (p && bkey_cmp(p, &START_KEY(k)) > 0)
82 				goto bug;
83 		} else {
84 			if (bch_ptr_bad(b, k))
85 				continue;
86 
87 			err = "Duplicate keys";
88 			if (p && !bkey_cmp(p, k))
89 				goto bug;
90 		}
91 		p = k;
92 	}
93 #if 0
94 	err = "Key larger than btree node key";
95 	if (p && bkey_cmp(p, &b->key) > 0)
96 		goto bug;
97 #endif
98 	return;
99 bug:
100 	bch_dump_bucket(b);
101 
102 	va_start(args, fmt);
103 	vprintk(fmt, args);
104 	va_end(args);
105 
106 	panic("bch_check_keys error:  %s:\n", err);
107 }
108 
109 static void bch_btree_iter_next_check(struct btree_iter *iter)
110 {
111 	struct bkey *k = iter->data->k, *next = bkey_next(k);
112 
113 	if (next < iter->data->end &&
114 	    bkey_cmp(k, iter->b->ops->is_extents ?
115 		     &START_KEY(next) : next) > 0) {
116 		bch_dump_bucket(iter->b);
117 		panic("Key skipped backwards\n");
118 	}
119 }
120 
121 #else
122 
123 static inline void bch_btree_iter_next_check(struct btree_iter *iter) {}
124 
125 #endif
126 
127 /* Keylists */
128 
129 int __bch_keylist_realloc(struct keylist *l, unsigned u64s)
130 {
131 	size_t oldsize = bch_keylist_nkeys(l);
132 	size_t newsize = oldsize + u64s;
133 	uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
134 	uint64_t *new_keys;
135 
136 	newsize = roundup_pow_of_two(newsize);
137 
138 	if (newsize <= KEYLIST_INLINE ||
139 	    roundup_pow_of_two(oldsize) == newsize)
140 		return 0;
141 
142 	new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
143 
144 	if (!new_keys)
145 		return -ENOMEM;
146 
147 	if (!old_keys)
148 		memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
149 
150 	l->keys_p = new_keys;
151 	l->top_p = new_keys + oldsize;
152 
153 	return 0;
154 }
155 
156 struct bkey *bch_keylist_pop(struct keylist *l)
157 {
158 	struct bkey *k = l->keys;
159 
160 	if (k == l->top)
161 		return NULL;
162 
163 	while (bkey_next(k) != l->top)
164 		k = bkey_next(k);
165 
166 	return l->top = k;
167 }
168 
169 void bch_keylist_pop_front(struct keylist *l)
170 {
171 	l->top_p -= bkey_u64s(l->keys);
172 
173 	memmove(l->keys,
174 		bkey_next(l->keys),
175 		bch_keylist_bytes(l));
176 }
177 
178 /* Key/pointer manipulation */
179 
180 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
181 			      unsigned i)
182 {
183 	BUG_ON(i > KEY_PTRS(src));
184 
185 	/* Only copy the header, key, and one pointer. */
186 	memcpy(dest, src, 2 * sizeof(uint64_t));
187 	dest->ptr[0] = src->ptr[i];
188 	SET_KEY_PTRS(dest, 1);
189 	/* We didn't copy the checksum so clear that bit. */
190 	SET_KEY_CSUM(dest, 0);
191 }
192 
193 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
194 {
195 	unsigned i, len = 0;
196 
197 	if (bkey_cmp(where, &START_KEY(k)) <= 0)
198 		return false;
199 
200 	if (bkey_cmp(where, k) < 0)
201 		len = KEY_OFFSET(k) - KEY_OFFSET(where);
202 	else
203 		bkey_copy_key(k, where);
204 
205 	for (i = 0; i < KEY_PTRS(k); i++)
206 		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
207 
208 	BUG_ON(len > KEY_SIZE(k));
209 	SET_KEY_SIZE(k, len);
210 	return true;
211 }
212 
213 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
214 {
215 	unsigned len = 0;
216 
217 	if (bkey_cmp(where, k) >= 0)
218 		return false;
219 
220 	BUG_ON(KEY_INODE(where) != KEY_INODE(k));
221 
222 	if (bkey_cmp(where, &START_KEY(k)) > 0)
223 		len = KEY_OFFSET(where) - KEY_START(k);
224 
225 	bkey_copy_key(k, where);
226 
227 	BUG_ON(len > KEY_SIZE(k));
228 	SET_KEY_SIZE(k, len);
229 	return true;
230 }
231 
232 /* Auxiliary search trees */
233 
234 /* 32 bits total: */
235 #define BKEY_MID_BITS		3
236 #define BKEY_EXPONENT_BITS	7
237 #define BKEY_MANTISSA_BITS	(32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS)
238 #define BKEY_MANTISSA_MASK	((1 << BKEY_MANTISSA_BITS) - 1)
239 
240 struct bkey_float {
241 	unsigned	exponent:BKEY_EXPONENT_BITS;
242 	unsigned	m:BKEY_MID_BITS;
243 	unsigned	mantissa:BKEY_MANTISSA_BITS;
244 } __packed;
245 
246 /*
247  * BSET_CACHELINE was originally intended to match the hardware cacheline size -
248  * it used to be 64, but I realized the lookup code would touch slightly less
249  * memory if it was 128.
250  *
251  * It definites the number of bytes (in struct bset) per struct bkey_float in
252  * the auxiliar search tree - when we're done searching the bset_float tree we
253  * have this many bytes left that we do a linear search over.
254  *
255  * Since (after level 5) every level of the bset_tree is on a new cacheline,
256  * we're touching one fewer cacheline in the bset tree in exchange for one more
257  * cacheline in the linear search - but the linear search might stop before it
258  * gets to the second cacheline.
259  */
260 
261 #define BSET_CACHELINE		128
262 
263 /* Space required for the btree node keys */
264 static inline size_t btree_keys_bytes(struct btree_keys *b)
265 {
266 	return PAGE_SIZE << b->page_order;
267 }
268 
269 static inline size_t btree_keys_cachelines(struct btree_keys *b)
270 {
271 	return btree_keys_bytes(b) / BSET_CACHELINE;
272 }
273 
274 /* Space required for the auxiliary search trees */
275 static inline size_t bset_tree_bytes(struct btree_keys *b)
276 {
277 	return btree_keys_cachelines(b) * sizeof(struct bkey_float);
278 }
279 
280 /* Space required for the prev pointers */
281 static inline size_t bset_prev_bytes(struct btree_keys *b)
282 {
283 	return btree_keys_cachelines(b) * sizeof(uint8_t);
284 }
285 
286 /* Memory allocation */
287 
288 void bch_btree_keys_free(struct btree_keys *b)
289 {
290 	struct bset_tree *t = b->set;
291 
292 	if (bset_prev_bytes(b) < PAGE_SIZE)
293 		kfree(t->prev);
294 	else
295 		free_pages((unsigned long) t->prev,
296 			   get_order(bset_prev_bytes(b)));
297 
298 	if (bset_tree_bytes(b) < PAGE_SIZE)
299 		kfree(t->tree);
300 	else
301 		free_pages((unsigned long) t->tree,
302 			   get_order(bset_tree_bytes(b)));
303 
304 	free_pages((unsigned long) t->data, b->page_order);
305 
306 	t->prev = NULL;
307 	t->tree = NULL;
308 	t->data = NULL;
309 }
310 EXPORT_SYMBOL(bch_btree_keys_free);
311 
312 int bch_btree_keys_alloc(struct btree_keys *b, unsigned page_order, gfp_t gfp)
313 {
314 	struct bset_tree *t = b->set;
315 
316 	BUG_ON(t->data);
317 
318 	b->page_order = page_order;
319 
320 	t->data = (void *) __get_free_pages(gfp, b->page_order);
321 	if (!t->data)
322 		goto err;
323 
324 	t->tree = bset_tree_bytes(b) < PAGE_SIZE
325 		? kmalloc(bset_tree_bytes(b), gfp)
326 		: (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
327 	if (!t->tree)
328 		goto err;
329 
330 	t->prev = bset_prev_bytes(b) < PAGE_SIZE
331 		? kmalloc(bset_prev_bytes(b), gfp)
332 		: (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
333 	if (!t->prev)
334 		goto err;
335 
336 	return 0;
337 err:
338 	bch_btree_keys_free(b);
339 	return -ENOMEM;
340 }
341 EXPORT_SYMBOL(bch_btree_keys_alloc);
342 
343 void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
344 			 bool *expensive_debug_checks)
345 {
346 	unsigned i;
347 
348 	b->ops = ops;
349 	b->expensive_debug_checks = expensive_debug_checks;
350 	b->nsets = 0;
351 	b->last_set_unwritten = 0;
352 
353 	/* XXX: shouldn't be needed */
354 	for (i = 0; i < MAX_BSETS; i++)
355 		b->set[i].size = 0;
356 	/*
357 	 * Second loop starts at 1 because b->keys[0]->data is the memory we
358 	 * allocated
359 	 */
360 	for (i = 1; i < MAX_BSETS; i++)
361 		b->set[i].data = NULL;
362 }
363 EXPORT_SYMBOL(bch_btree_keys_init);
364 
365 /* Binary tree stuff for auxiliary search trees */
366 
367 static unsigned inorder_next(unsigned j, unsigned size)
368 {
369 	if (j * 2 + 1 < size) {
370 		j = j * 2 + 1;
371 
372 		while (j * 2 < size)
373 			j *= 2;
374 	} else
375 		j >>= ffz(j) + 1;
376 
377 	return j;
378 }
379 
380 static unsigned inorder_prev(unsigned j, unsigned size)
381 {
382 	if (j * 2 < size) {
383 		j = j * 2;
384 
385 		while (j * 2 + 1 < size)
386 			j = j * 2 + 1;
387 	} else
388 		j >>= ffs(j);
389 
390 	return j;
391 }
392 
393 /* I have no idea why this code works... and I'm the one who wrote it
394  *
395  * However, I do know what it does:
396  * Given a binary tree constructed in an array (i.e. how you normally implement
397  * a heap), it converts a node in the tree - referenced by array index - to the
398  * index it would have if you did an inorder traversal.
399  *
400  * Also tested for every j, size up to size somewhere around 6 million.
401  *
402  * The binary tree starts at array index 1, not 0
403  * extra is a function of size:
404  *   extra = (size - rounddown_pow_of_two(size - 1)) << 1;
405  */
406 static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
407 {
408 	unsigned b = fls(j);
409 	unsigned shift = fls(size - 1) - b;
410 
411 	j  ^= 1U << (b - 1);
412 	j <<= 1;
413 	j  |= 1;
414 	j <<= shift;
415 
416 	if (j > extra)
417 		j -= (j - extra) >> 1;
418 
419 	return j;
420 }
421 
422 static unsigned to_inorder(unsigned j, struct bset_tree *t)
423 {
424 	return __to_inorder(j, t->size, t->extra);
425 }
426 
427 static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
428 {
429 	unsigned shift;
430 
431 	if (j > extra)
432 		j += j - extra;
433 
434 	shift = ffs(j);
435 
436 	j >>= shift;
437 	j  |= roundup_pow_of_two(size) >> shift;
438 
439 	return j;
440 }
441 
442 static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
443 {
444 	return __inorder_to_tree(j, t->size, t->extra);
445 }
446 
447 #if 0
448 void inorder_test(void)
449 {
450 	unsigned long done = 0;
451 	ktime_t start = ktime_get();
452 
453 	for (unsigned size = 2;
454 	     size < 65536000;
455 	     size++) {
456 		unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
457 		unsigned i = 1, j = rounddown_pow_of_two(size - 1);
458 
459 		if (!(size % 4096))
460 			printk(KERN_NOTICE "loop %u, %llu per us\n", size,
461 			       done / ktime_us_delta(ktime_get(), start));
462 
463 		while (1) {
464 			if (__inorder_to_tree(i, size, extra) != j)
465 				panic("size %10u j %10u i %10u", size, j, i);
466 
467 			if (__to_inorder(j, size, extra) != i)
468 				panic("size %10u j %10u i %10u", size, j, i);
469 
470 			if (j == rounddown_pow_of_two(size) - 1)
471 				break;
472 
473 			BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
474 
475 			j = inorder_next(j, size);
476 			i++;
477 		}
478 
479 		done += size - 1;
480 	}
481 }
482 #endif
483 
484 /*
485  * Cacheline/offset <-> bkey pointer arithmetic:
486  *
487  * t->tree is a binary search tree in an array; each node corresponds to a key
488  * in one cacheline in t->set (BSET_CACHELINE bytes).
489  *
490  * This means we don't have to store the full index of the key that a node in
491  * the binary tree points to; to_inorder() gives us the cacheline, and then
492  * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
493  *
494  * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
495  * make this work.
496  *
497  * To construct the bfloat for an arbitrary key we need to know what the key
498  * immediately preceding it is: we have to check if the two keys differ in the
499  * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
500  * of the previous key so we can walk backwards to it from t->tree[j]'s key.
501  */
502 
503 static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
504 				      unsigned offset)
505 {
506 	return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
507 }
508 
509 static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
510 {
511 	return ((void *) k - (void *) t->data) / BSET_CACHELINE;
512 }
513 
514 static unsigned bkey_to_cacheline_offset(struct bset_tree *t,
515 					 unsigned cacheline,
516 					 struct bkey *k)
517 {
518 	return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0);
519 }
520 
521 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
522 {
523 	return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
524 }
525 
526 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
527 {
528 	return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
529 }
530 
531 /*
532  * For the write set - the one we're currently inserting keys into - we don't
533  * maintain a full search tree, we just keep a simple lookup table in t->prev.
534  */
535 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
536 {
537 	return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
538 }
539 
540 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
541 {
542 	low >>= shift;
543 	low  |= (high << 1) << (63U - shift);
544 	return low;
545 }
546 
547 static inline unsigned bfloat_mantissa(const struct bkey *k,
548 				       struct bkey_float *f)
549 {
550 	const uint64_t *p = &k->low - (f->exponent >> 6);
551 	return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
552 }
553 
554 static void make_bfloat(struct bset_tree *t, unsigned j)
555 {
556 	struct bkey_float *f = &t->tree[j];
557 	struct bkey *m = tree_to_bkey(t, j);
558 	struct bkey *p = tree_to_prev_bkey(t, j);
559 
560 	struct bkey *l = is_power_of_2(j)
561 		? t->data->start
562 		: tree_to_prev_bkey(t, j >> ffs(j));
563 
564 	struct bkey *r = is_power_of_2(j + 1)
565 		? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
566 		: tree_to_bkey(t, j >> (ffz(j) + 1));
567 
568 	BUG_ON(m < l || m > r);
569 	BUG_ON(bkey_next(p) != m);
570 
571 	if (KEY_INODE(l) != KEY_INODE(r))
572 		f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
573 	else
574 		f->exponent = fls64(r->low ^ l->low);
575 
576 	f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
577 
578 	/*
579 	 * Setting f->exponent = 127 flags this node as failed, and causes the
580 	 * lookup code to fall back to comparing against the original key.
581 	 */
582 
583 	if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
584 		f->mantissa = bfloat_mantissa(m, f) - 1;
585 	else
586 		f->exponent = 127;
587 }
588 
589 static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t)
590 {
591 	if (t != b->set) {
592 		unsigned j = roundup(t[-1].size,
593 				     64 / sizeof(struct bkey_float));
594 
595 		t->tree = t[-1].tree + j;
596 		t->prev = t[-1].prev + j;
597 	}
598 
599 	while (t < b->set + MAX_BSETS)
600 		t++->size = 0;
601 }
602 
603 static void bch_bset_build_unwritten_tree(struct btree_keys *b)
604 {
605 	struct bset_tree *t = bset_tree_last(b);
606 
607 	BUG_ON(b->last_set_unwritten);
608 	b->last_set_unwritten = 1;
609 
610 	bset_alloc_tree(b, t);
611 
612 	if (t->tree != b->set->tree + btree_keys_cachelines(b)) {
613 		t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start);
614 		t->size = 1;
615 	}
616 }
617 
618 void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic)
619 {
620 	if (i != b->set->data) {
621 		b->set[++b->nsets].data = i;
622 		i->seq = b->set->data->seq;
623 	} else
624 		get_random_bytes(&i->seq, sizeof(uint64_t));
625 
626 	i->magic	= magic;
627 	i->version	= 0;
628 	i->keys		= 0;
629 
630 	bch_bset_build_unwritten_tree(b);
631 }
632 EXPORT_SYMBOL(bch_bset_init_next);
633 
634 void bch_bset_build_written_tree(struct btree_keys *b)
635 {
636 	struct bset_tree *t = bset_tree_last(b);
637 	struct bkey *prev = NULL, *k = t->data->start;
638 	unsigned j, cacheline = 1;
639 
640 	b->last_set_unwritten = 0;
641 
642 	bset_alloc_tree(b, t);
643 
644 	t->size = min_t(unsigned,
645 			bkey_to_cacheline(t, bset_bkey_last(t->data)),
646 			b->set->tree + btree_keys_cachelines(b) - t->tree);
647 
648 	if (t->size < 2) {
649 		t->size = 0;
650 		return;
651 	}
652 
653 	t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
654 
655 	/* First we figure out where the first key in each cacheline is */
656 	for (j = inorder_next(0, t->size);
657 	     j;
658 	     j = inorder_next(j, t->size)) {
659 		while (bkey_to_cacheline(t, k) < cacheline)
660 			prev = k, k = bkey_next(k);
661 
662 		t->prev[j] = bkey_u64s(prev);
663 		t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k);
664 	}
665 
666 	while (bkey_next(k) != bset_bkey_last(t->data))
667 		k = bkey_next(k);
668 
669 	t->end = *k;
670 
671 	/* Then we build the tree */
672 	for (j = inorder_next(0, t->size);
673 	     j;
674 	     j = inorder_next(j, t->size))
675 		make_bfloat(t, j);
676 }
677 EXPORT_SYMBOL(bch_bset_build_written_tree);
678 
679 /* Insert */
680 
681 void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k)
682 {
683 	struct bset_tree *t;
684 	unsigned inorder, j = 1;
685 
686 	for (t = b->set; t <= bset_tree_last(b); t++)
687 		if (k < bset_bkey_last(t->data))
688 			goto found_set;
689 
690 	BUG();
691 found_set:
692 	if (!t->size || !bset_written(b, t))
693 		return;
694 
695 	inorder = bkey_to_cacheline(t, k);
696 
697 	if (k == t->data->start)
698 		goto fix_left;
699 
700 	if (bkey_next(k) == bset_bkey_last(t->data)) {
701 		t->end = *k;
702 		goto fix_right;
703 	}
704 
705 	j = inorder_to_tree(inorder, t);
706 
707 	if (j &&
708 	    j < t->size &&
709 	    k == tree_to_bkey(t, j))
710 fix_left:	do {
711 			make_bfloat(t, j);
712 			j = j * 2;
713 		} while (j < t->size);
714 
715 	j = inorder_to_tree(inorder + 1, t);
716 
717 	if (j &&
718 	    j < t->size &&
719 	    k == tree_to_prev_bkey(t, j))
720 fix_right:	do {
721 			make_bfloat(t, j);
722 			j = j * 2 + 1;
723 		} while (j < t->size);
724 }
725 EXPORT_SYMBOL(bch_bset_fix_invalidated_key);
726 
727 static void bch_bset_fix_lookup_table(struct btree_keys *b,
728 				      struct bset_tree *t,
729 				      struct bkey *k)
730 {
731 	unsigned shift = bkey_u64s(k);
732 	unsigned j = bkey_to_cacheline(t, k);
733 
734 	/* We're getting called from btree_split() or btree_gc, just bail out */
735 	if (!t->size)
736 		return;
737 
738 	/* k is the key we just inserted; we need to find the entry in the
739 	 * lookup table for the first key that is strictly greater than k:
740 	 * it's either k's cacheline or the next one
741 	 */
742 	while (j < t->size &&
743 	       table_to_bkey(t, j) <= k)
744 		j++;
745 
746 	/* Adjust all the lookup table entries, and find a new key for any that
747 	 * have gotten too big
748 	 */
749 	for (; j < t->size; j++) {
750 		t->prev[j] += shift;
751 
752 		if (t->prev[j] > 7) {
753 			k = table_to_bkey(t, j - 1);
754 
755 			while (k < cacheline_to_bkey(t, j, 0))
756 				k = bkey_next(k);
757 
758 			t->prev[j] = bkey_to_cacheline_offset(t, j, k);
759 		}
760 	}
761 
762 	if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree)
763 		return;
764 
765 	/* Possibly add a new entry to the end of the lookup table */
766 
767 	for (k = table_to_bkey(t, t->size - 1);
768 	     k != bset_bkey_last(t->data);
769 	     k = bkey_next(k))
770 		if (t->size == bkey_to_cacheline(t, k)) {
771 			t->prev[t->size] = bkey_to_cacheline_offset(t, t->size, k);
772 			t->size++;
773 		}
774 }
775 
776 /*
777  * Tries to merge l and r: l should be lower than r
778  * Returns true if we were able to merge. If we did merge, l will be the merged
779  * key, r will be untouched.
780  */
781 bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r)
782 {
783 	if (!b->ops->key_merge)
784 		return false;
785 
786 	/*
787 	 * Generic header checks
788 	 * Assumes left and right are in order
789 	 * Left and right must be exactly aligned
790 	 */
791 	if (!bch_bkey_equal_header(l, r) ||
792 	     bkey_cmp(l, &START_KEY(r)))
793 		return false;
794 
795 	return b->ops->key_merge(b, l, r);
796 }
797 EXPORT_SYMBOL(bch_bkey_try_merge);
798 
799 void bch_bset_insert(struct btree_keys *b, struct bkey *where,
800 		     struct bkey *insert)
801 {
802 	struct bset_tree *t = bset_tree_last(b);
803 
804 	BUG_ON(!b->last_set_unwritten);
805 	BUG_ON(bset_byte_offset(b, t->data) +
806 	       __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
807 	       PAGE_SIZE << b->page_order);
808 
809 	memmove((uint64_t *) where + bkey_u64s(insert),
810 		where,
811 		(void *) bset_bkey_last(t->data) - (void *) where);
812 
813 	t->data->keys += bkey_u64s(insert);
814 	bkey_copy(where, insert);
815 	bch_bset_fix_lookup_table(b, t, where);
816 }
817 EXPORT_SYMBOL(bch_bset_insert);
818 
819 unsigned bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
820 			      struct bkey *replace_key)
821 {
822 	unsigned status = BTREE_INSERT_STATUS_NO_INSERT;
823 	struct bset *i = bset_tree_last(b)->data;
824 	struct bkey *m, *prev = NULL;
825 	struct btree_iter iter;
826 
827 	BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
828 
829 	m = bch_btree_iter_init(b, &iter, b->ops->is_extents
830 				? PRECEDING_KEY(&START_KEY(k))
831 				: PRECEDING_KEY(k));
832 
833 	if (b->ops->insert_fixup(b, k, &iter, replace_key))
834 		return status;
835 
836 	status = BTREE_INSERT_STATUS_INSERT;
837 
838 	while (m != bset_bkey_last(i) &&
839 	       bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0)
840 		prev = m, m = bkey_next(m);
841 
842 	/* prev is in the tree, if we merge we're done */
843 	status = BTREE_INSERT_STATUS_BACK_MERGE;
844 	if (prev &&
845 	    bch_bkey_try_merge(b, prev, k))
846 		goto merged;
847 #if 0
848 	status = BTREE_INSERT_STATUS_OVERWROTE;
849 	if (m != bset_bkey_last(i) &&
850 	    KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
851 		goto copy;
852 #endif
853 	status = BTREE_INSERT_STATUS_FRONT_MERGE;
854 	if (m != bset_bkey_last(i) &&
855 	    bch_bkey_try_merge(b, k, m))
856 		goto copy;
857 
858 	bch_bset_insert(b, m, k);
859 copy:	bkey_copy(m, k);
860 merged:
861 	return status;
862 }
863 EXPORT_SYMBOL(bch_btree_insert_key);
864 
865 /* Lookup */
866 
867 struct bset_search_iter {
868 	struct bkey *l, *r;
869 };
870 
871 static struct bset_search_iter bset_search_write_set(struct bset_tree *t,
872 						     const struct bkey *search)
873 {
874 	unsigned li = 0, ri = t->size;
875 
876 	while (li + 1 != ri) {
877 		unsigned m = (li + ri) >> 1;
878 
879 		if (bkey_cmp(table_to_bkey(t, m), search) > 0)
880 			ri = m;
881 		else
882 			li = m;
883 	}
884 
885 	return (struct bset_search_iter) {
886 		table_to_bkey(t, li),
887 		ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
888 	};
889 }
890 
891 static struct bset_search_iter bset_search_tree(struct bset_tree *t,
892 						const struct bkey *search)
893 {
894 	struct bkey *l, *r;
895 	struct bkey_float *f;
896 	unsigned inorder, j, n = 1;
897 
898 	do {
899 		unsigned p = n << 4;
900 		p &= ((int) (p - t->size)) >> 31;
901 
902 		prefetch(&t->tree[p]);
903 
904 		j = n;
905 		f = &t->tree[j];
906 
907 		/*
908 		 * n = (f->mantissa > bfloat_mantissa())
909 		 *	? j * 2
910 		 *	: j * 2 + 1;
911 		 *
912 		 * We need to subtract 1 from f->mantissa for the sign bit trick
913 		 * to work  - that's done in make_bfloat()
914 		 */
915 		if (likely(f->exponent != 127))
916 			n = j * 2 + (((unsigned)
917 				      (f->mantissa -
918 				       bfloat_mantissa(search, f))) >> 31);
919 		else
920 			n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
921 				? j * 2
922 				: j * 2 + 1;
923 	} while (n < t->size);
924 
925 	inorder = to_inorder(j, t);
926 
927 	/*
928 	 * n would have been the node we recursed to - the low bit tells us if
929 	 * we recursed left or recursed right.
930 	 */
931 	if (n & 1) {
932 		l = cacheline_to_bkey(t, inorder, f->m);
933 
934 		if (++inorder != t->size) {
935 			f = &t->tree[inorder_next(j, t->size)];
936 			r = cacheline_to_bkey(t, inorder, f->m);
937 		} else
938 			r = bset_bkey_last(t->data);
939 	} else {
940 		r = cacheline_to_bkey(t, inorder, f->m);
941 
942 		if (--inorder) {
943 			f = &t->tree[inorder_prev(j, t->size)];
944 			l = cacheline_to_bkey(t, inorder, f->m);
945 		} else
946 			l = t->data->start;
947 	}
948 
949 	return (struct bset_search_iter) {l, r};
950 }
951 
952 struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
953 			       const struct bkey *search)
954 {
955 	struct bset_search_iter i;
956 
957 	/*
958 	 * First, we search for a cacheline, then lastly we do a linear search
959 	 * within that cacheline.
960 	 *
961 	 * To search for the cacheline, there's three different possibilities:
962 	 *  * The set is too small to have a search tree, so we just do a linear
963 	 *    search over the whole set.
964 	 *  * The set is the one we're currently inserting into; keeping a full
965 	 *    auxiliary search tree up to date would be too expensive, so we
966 	 *    use a much simpler lookup table to do a binary search -
967 	 *    bset_search_write_set().
968 	 *  * Or we use the auxiliary search tree we constructed earlier -
969 	 *    bset_search_tree()
970 	 */
971 
972 	if (unlikely(!t->size)) {
973 		i.l = t->data->start;
974 		i.r = bset_bkey_last(t->data);
975 	} else if (bset_written(b, t)) {
976 		/*
977 		 * Each node in the auxiliary search tree covers a certain range
978 		 * of bits, and keys above and below the set it covers might
979 		 * differ outside those bits - so we have to special case the
980 		 * start and end - handle that here:
981 		 */
982 
983 		if (unlikely(bkey_cmp(search, &t->end) >= 0))
984 			return bset_bkey_last(t->data);
985 
986 		if (unlikely(bkey_cmp(search, t->data->start) < 0))
987 			return t->data->start;
988 
989 		i = bset_search_tree(t, search);
990 	} else {
991 		BUG_ON(!b->nsets &&
992 		       t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
993 
994 		i = bset_search_write_set(t, search);
995 	}
996 
997 	if (btree_keys_expensive_checks(b)) {
998 		BUG_ON(bset_written(b, t) &&
999 		       i.l != t->data->start &&
1000 		       bkey_cmp(tree_to_prev_bkey(t,
1001 			  inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
1002 				search) > 0);
1003 
1004 		BUG_ON(i.r != bset_bkey_last(t->data) &&
1005 		       bkey_cmp(i.r, search) <= 0);
1006 	}
1007 
1008 	while (likely(i.l != i.r) &&
1009 	       bkey_cmp(i.l, search) <= 0)
1010 		i.l = bkey_next(i.l);
1011 
1012 	return i.l;
1013 }
1014 EXPORT_SYMBOL(__bch_bset_search);
1015 
1016 /* Btree iterator */
1017 
1018 typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
1019 				 struct btree_iter_set);
1020 
1021 static inline bool btree_iter_cmp(struct btree_iter_set l,
1022 				  struct btree_iter_set r)
1023 {
1024 	return bkey_cmp(l.k, r.k) > 0;
1025 }
1026 
1027 static inline bool btree_iter_end(struct btree_iter *iter)
1028 {
1029 	return !iter->used;
1030 }
1031 
1032 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
1033 			 struct bkey *end)
1034 {
1035 	if (k != end)
1036 		BUG_ON(!heap_add(iter,
1037 				 ((struct btree_iter_set) { k, end }),
1038 				 btree_iter_cmp));
1039 }
1040 
1041 static struct bkey *__bch_btree_iter_init(struct btree_keys *b,
1042 					  struct btree_iter *iter,
1043 					  struct bkey *search,
1044 					  struct bset_tree *start)
1045 {
1046 	struct bkey *ret = NULL;
1047 	iter->size = ARRAY_SIZE(iter->data);
1048 	iter->used = 0;
1049 
1050 #ifdef CONFIG_BCACHE_DEBUG
1051 	iter->b = b;
1052 #endif
1053 
1054 	for (; start <= bset_tree_last(b); start++) {
1055 		ret = bch_bset_search(b, start, search);
1056 		bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
1057 	}
1058 
1059 	return ret;
1060 }
1061 
1062 struct bkey *bch_btree_iter_init(struct btree_keys *b,
1063 				 struct btree_iter *iter,
1064 				 struct bkey *search)
1065 {
1066 	return __bch_btree_iter_init(b, iter, search, b->set);
1067 }
1068 EXPORT_SYMBOL(bch_btree_iter_init);
1069 
1070 static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
1071 						 btree_iter_cmp_fn *cmp)
1072 {
1073 	struct btree_iter_set unused;
1074 	struct bkey *ret = NULL;
1075 
1076 	if (!btree_iter_end(iter)) {
1077 		bch_btree_iter_next_check(iter);
1078 
1079 		ret = iter->data->k;
1080 		iter->data->k = bkey_next(iter->data->k);
1081 
1082 		if (iter->data->k > iter->data->end) {
1083 			WARN_ONCE(1, "bset was corrupt!\n");
1084 			iter->data->k = iter->data->end;
1085 		}
1086 
1087 		if (iter->data->k == iter->data->end)
1088 			heap_pop(iter, unused, cmp);
1089 		else
1090 			heap_sift(iter, 0, cmp);
1091 	}
1092 
1093 	return ret;
1094 }
1095 
1096 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
1097 {
1098 	return __bch_btree_iter_next(iter, btree_iter_cmp);
1099 
1100 }
1101 EXPORT_SYMBOL(bch_btree_iter_next);
1102 
1103 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
1104 					struct btree_keys *b, ptr_filter_fn fn)
1105 {
1106 	struct bkey *ret;
1107 
1108 	do {
1109 		ret = bch_btree_iter_next(iter);
1110 	} while (ret && fn(b, ret));
1111 
1112 	return ret;
1113 }
1114 
1115 /* Mergesort */
1116 
1117 void bch_bset_sort_state_free(struct bset_sort_state *state)
1118 {
1119 	if (state->pool)
1120 		mempool_destroy(state->pool);
1121 }
1122 
1123 int bch_bset_sort_state_init(struct bset_sort_state *state, unsigned page_order)
1124 {
1125 	spin_lock_init(&state->time.lock);
1126 
1127 	state->page_order = page_order;
1128 	state->crit_factor = int_sqrt(1 << page_order);
1129 
1130 	state->pool = mempool_create_page_pool(1, page_order);
1131 	if (!state->pool)
1132 		return -ENOMEM;
1133 
1134 	return 0;
1135 }
1136 EXPORT_SYMBOL(bch_bset_sort_state_init);
1137 
1138 static void btree_mergesort(struct btree_keys *b, struct bset *out,
1139 			    struct btree_iter *iter,
1140 			    bool fixup, bool remove_stale)
1141 {
1142 	int i;
1143 	struct bkey *k, *last = NULL;
1144 	BKEY_PADDED(k) tmp;
1145 	bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale
1146 		? bch_ptr_bad
1147 		: bch_ptr_invalid;
1148 
1149 	/* Heapify the iterator, using our comparison function */
1150 	for (i = iter->used / 2 - 1; i >= 0; --i)
1151 		heap_sift(iter, i, b->ops->sort_cmp);
1152 
1153 	while (!btree_iter_end(iter)) {
1154 		if (b->ops->sort_fixup && fixup)
1155 			k = b->ops->sort_fixup(iter, &tmp.k);
1156 		else
1157 			k = NULL;
1158 
1159 		if (!k)
1160 			k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
1161 
1162 		if (bad(b, k))
1163 			continue;
1164 
1165 		if (!last) {
1166 			last = out->start;
1167 			bkey_copy(last, k);
1168 		} else if (!bch_bkey_try_merge(b, last, k)) {
1169 			last = bkey_next(last);
1170 			bkey_copy(last, k);
1171 		}
1172 	}
1173 
1174 	out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1175 
1176 	pr_debug("sorted %i keys", out->keys);
1177 }
1178 
1179 static void __btree_sort(struct btree_keys *b, struct btree_iter *iter,
1180 			 unsigned start, unsigned order, bool fixup,
1181 			 struct bset_sort_state *state)
1182 {
1183 	uint64_t start_time;
1184 	bool used_mempool = false;
1185 	struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO,
1186 						     order);
1187 	if (!out) {
1188 		BUG_ON(order > state->page_order);
1189 
1190 		out = page_address(mempool_alloc(state->pool, GFP_NOIO));
1191 		used_mempool = true;
1192 		order = state->page_order;
1193 	}
1194 
1195 	start_time = local_clock();
1196 
1197 	btree_mergesort(b, out, iter, fixup, false);
1198 	b->nsets = start;
1199 
1200 	if (!start && order == b->page_order) {
1201 		/*
1202 		 * Our temporary buffer is the same size as the btree node's
1203 		 * buffer, we can just swap buffers instead of doing a big
1204 		 * memcpy()
1205 		 */
1206 
1207 		out->magic	= b->set->data->magic;
1208 		out->seq	= b->set->data->seq;
1209 		out->version	= b->set->data->version;
1210 		swap(out, b->set->data);
1211 	} else {
1212 		b->set[start].data->keys = out->keys;
1213 		memcpy(b->set[start].data->start, out->start,
1214 		       (void *) bset_bkey_last(out) - (void *) out->start);
1215 	}
1216 
1217 	if (used_mempool)
1218 		mempool_free(virt_to_page(out), state->pool);
1219 	else
1220 		free_pages((unsigned long) out, order);
1221 
1222 	bch_bset_build_written_tree(b);
1223 
1224 	if (!start)
1225 		bch_time_stats_update(&state->time, start_time);
1226 }
1227 
1228 void bch_btree_sort_partial(struct btree_keys *b, unsigned start,
1229 			    struct bset_sort_state *state)
1230 {
1231 	size_t order = b->page_order, keys = 0;
1232 	struct btree_iter iter;
1233 	int oldsize = bch_count_data(b);
1234 
1235 	__bch_btree_iter_init(b, &iter, NULL, &b->set[start]);
1236 
1237 	if (start) {
1238 		unsigned i;
1239 
1240 		for (i = start; i <= b->nsets; i++)
1241 			keys += b->set[i].data->keys;
1242 
1243 		order = get_order(__set_bytes(b->set->data, keys));
1244 	}
1245 
1246 	__btree_sort(b, &iter, start, order, false, state);
1247 
1248 	EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize);
1249 }
1250 EXPORT_SYMBOL(bch_btree_sort_partial);
1251 
1252 void bch_btree_sort_and_fix_extents(struct btree_keys *b,
1253 				    struct btree_iter *iter,
1254 				    struct bset_sort_state *state)
1255 {
1256 	__btree_sort(b, iter, 0, b->page_order, true, state);
1257 }
1258 
1259 void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
1260 			 struct bset_sort_state *state)
1261 {
1262 	uint64_t start_time = local_clock();
1263 
1264 	struct btree_iter iter;
1265 	bch_btree_iter_init(b, &iter, NULL);
1266 
1267 	btree_mergesort(b, new->set->data, &iter, false, true);
1268 
1269 	bch_time_stats_update(&state->time, start_time);
1270 
1271 	new->set->size = 0; // XXX: why?
1272 }
1273 
1274 #define SORT_CRIT	(4096 / sizeof(uint64_t))
1275 
1276 void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state)
1277 {
1278 	unsigned crit = SORT_CRIT;
1279 	int i;
1280 
1281 	/* Don't sort if nothing to do */
1282 	if (!b->nsets)
1283 		goto out;
1284 
1285 	for (i = b->nsets - 1; i >= 0; --i) {
1286 		crit *= state->crit_factor;
1287 
1288 		if (b->set[i].data->keys < crit) {
1289 			bch_btree_sort_partial(b, i, state);
1290 			return;
1291 		}
1292 	}
1293 
1294 	/* Sort if we'd overflow */
1295 	if (b->nsets + 1 == MAX_BSETS) {
1296 		bch_btree_sort(b, state);
1297 		return;
1298 	}
1299 
1300 out:
1301 	bch_bset_build_written_tree(b);
1302 }
1303 EXPORT_SYMBOL(bch_btree_sort_lazy);
1304 
1305 void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats)
1306 {
1307 	unsigned i;
1308 
1309 	for (i = 0; i <= b->nsets; i++) {
1310 		struct bset_tree *t = &b->set[i];
1311 		size_t bytes = t->data->keys * sizeof(uint64_t);
1312 		size_t j;
1313 
1314 		if (bset_written(b, t)) {
1315 			stats->sets_written++;
1316 			stats->bytes_written += bytes;
1317 
1318 			stats->floats += t->size - 1;
1319 
1320 			for (j = 1; j < t->size; j++)
1321 				if (t->tree[j].exponent == 127)
1322 					stats->failed++;
1323 		} else {
1324 			stats->sets_unwritten++;
1325 			stats->bytes_unwritten += bytes;
1326 		}
1327 	}
1328 }
1329