1 // SPDX-License-Identifier: GPL-2.0
3 * Code for working with individual keys, and sorted sets of keys with in a
6 * Copyright 2012 Google, Inc.
9 #define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__
14 #include <linux/console.h>
15 #include <linux/sched/clock.h>
16 #include <linux/random.h>
17 #include <linux/prefetch.h>
19 #ifdef CONFIG_BCACHE_DEBUG
21 void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned int set)
23 struct bkey *k, *next;
25 for (k = i->start; k < bset_bkey_last(i); k = next) {
28 pr_err("block %u key %u/%u: ", set,
29 (unsigned int) ((u64 *) k - i->d), i->keys);
32 b->ops->key_dump(b, k);
34 pr_err("%llu:%llu\n", KEY_INODE(k), KEY_OFFSET(k));
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");
43 void bch_dump_bucket(struct btree_keys *b)
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));
54 int __bch_count_data(struct btree_keys *b)
57 struct btree_iter iter;
60 if (b->ops->is_extents)
61 for_each_key(b, k, &iter)
66 void __bch_check_keys(struct btree_keys *b, const char *fmt, ...)
69 struct bkey *k, *p = NULL;
70 struct btree_iter iter;
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)
79 if (bch_ptr_invalid(b, k))
82 err = "Overlapping keys";
83 if (p && bkey_cmp(p, &START_KEY(k)) > 0)
86 if (bch_ptr_bad(b, k))
89 err = "Duplicate keys";
90 if (p && !bkey_cmp(p, k))
96 err = "Key larger than btree node key";
97 if (p && bkey_cmp(p, &b->key) > 0)
108 panic("bch_check_keys error: %s:\n", err);
111 static void bch_btree_iter_next_check(struct btree_iter *iter)
113 struct bkey *k = iter->data->k, *next = bkey_next(k);
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");
125 static inline void bch_btree_iter_next_check(struct btree_iter *iter) {}
131 int __bch_keylist_realloc(struct keylist *l, unsigned int u64s)
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;
138 newsize = roundup_pow_of_two(newsize);
140 if (newsize <= KEYLIST_INLINE ||
141 roundup_pow_of_two(oldsize) == newsize)
144 new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
150 memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
152 l->keys_p = new_keys;
153 l->top_p = new_keys + oldsize;
158 /* Pop the top key of keylist by pointing l->top to its previous key */
159 struct bkey *bch_keylist_pop(struct keylist *l)
161 struct bkey *k = l->keys;
166 while (bkey_next(k) != l->top)
172 /* Pop the bottom key of keylist and update l->top_p */
173 void bch_keylist_pop_front(struct keylist *l)
175 l->top_p -= bkey_u64s(l->keys);
179 bch_keylist_bytes(l));
182 /* Key/pointer manipulation */
184 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
187 BUG_ON(i > KEY_PTRS(src));
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);
197 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
199 unsigned int i, len = 0;
201 if (bkey_cmp(where, &START_KEY(k)) <= 0)
204 if (bkey_cmp(where, k) < 0)
205 len = KEY_OFFSET(k) - KEY_OFFSET(where);
207 bkey_copy_key(k, where);
209 for (i = 0; i < KEY_PTRS(k); i++)
210 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
212 BUG_ON(len > KEY_SIZE(k));
213 SET_KEY_SIZE(k, len);
217 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
219 unsigned int len = 0;
221 if (bkey_cmp(where, k) >= 0)
224 BUG_ON(KEY_INODE(where) != KEY_INODE(k));
226 if (bkey_cmp(where, &START_KEY(k)) > 0)
227 len = KEY_OFFSET(where) - KEY_START(k);
229 bkey_copy_key(k, where);
231 BUG_ON(len > KEY_SIZE(k));
232 SET_KEY_SIZE(k, len);
236 /* Auxiliary search trees */
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)
245 unsigned int exponent:BKEY_EXPONENT_BITS;
246 unsigned int m:BKEY_MID_BITS;
247 unsigned int mantissa:BKEY_MANTISSA_BITS;
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.
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.
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.
265 #define BSET_CACHELINE 128
267 /* Space required for the btree node keys */
268 static inline size_t btree_keys_bytes(struct btree_keys *b)
270 return PAGE_SIZE << b->page_order;
273 static inline size_t btree_keys_cachelines(struct btree_keys *b)
275 return btree_keys_bytes(b) / BSET_CACHELINE;
278 /* Space required for the auxiliary search trees */
279 static inline size_t bset_tree_bytes(struct btree_keys *b)
281 return btree_keys_cachelines(b) * sizeof(struct bkey_float);
284 /* Space required for the prev pointers */
285 static inline size_t bset_prev_bytes(struct btree_keys *b)
287 return btree_keys_cachelines(b) * sizeof(uint8_t);
290 /* Memory allocation */
292 void bch_btree_keys_free(struct btree_keys *b)
294 struct bset_tree *t = b->set;
296 if (bset_prev_bytes(b) < PAGE_SIZE)
299 free_pages((unsigned long) t->prev,
300 get_order(bset_prev_bytes(b)));
302 if (bset_tree_bytes(b) < PAGE_SIZE)
305 free_pages((unsigned long) t->tree,
306 get_order(bset_tree_bytes(b)));
308 free_pages((unsigned long) t->data, b->page_order);
314 EXPORT_SYMBOL(bch_btree_keys_free);
316 int bch_btree_keys_alloc(struct btree_keys *b,
317 unsigned int page_order,
320 struct bset_tree *t = b->set;
324 b->page_order = page_order;
326 t->data = (void *) __get_free_pages(gfp, b->page_order);
330 t->tree = bset_tree_bytes(b) < PAGE_SIZE
331 ? kmalloc(bset_tree_bytes(b), gfp)
332 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
336 t->prev = bset_prev_bytes(b) < PAGE_SIZE
337 ? kmalloc(bset_prev_bytes(b), gfp)
338 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
344 bch_btree_keys_free(b);
347 EXPORT_SYMBOL(bch_btree_keys_alloc);
349 void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
350 bool *expensive_debug_checks)
353 b->expensive_debug_checks = expensive_debug_checks;
355 b->last_set_unwritten = 0;
358 * struct btree_keys in embedded in struct btree, and struct
359 * bset_tree is embedded into struct btree_keys. They are all
360 * initialized as 0 by kzalloc() in mca_bucket_alloc(), and
361 * b->set[0].data is allocated in bch_btree_keys_alloc(), so we
362 * don't have to initiate b->set[].size and b->set[].data here
366 EXPORT_SYMBOL(bch_btree_keys_init);
368 /* Binary tree stuff for auxiliary search trees */
371 * return array index next to j when does in-order traverse
372 * of a binary tree which is stored in a linear array
374 static unsigned int inorder_next(unsigned int j, unsigned int size)
376 if (j * 2 + 1 < size) {
388 * return array index previous to j when does in-order traverse
389 * of a binary tree which is stored in a linear array
391 static unsigned int inorder_prev(unsigned int j, unsigned int size)
396 while (j * 2 + 1 < size)
405 * I have no idea why this code works... and I'm the one who wrote it
407 * However, I do know what it does:
408 * Given a binary tree constructed in an array (i.e. how you normally implement
409 * a heap), it converts a node in the tree - referenced by array index - to the
410 * index it would have if you did an inorder traversal.
412 * Also tested for every j, size up to size somewhere around 6 million.
414 * The binary tree starts at array index 1, not 0
415 * extra is a function of size:
416 * extra = (size - rounddown_pow_of_two(size - 1)) << 1;
418 static unsigned int __to_inorder(unsigned int j,
422 unsigned int b = fls(j);
423 unsigned int shift = fls(size - 1) - b;
431 j -= (j - extra) >> 1;
437 * Return the cacheline index in bset_tree->data, where j is index
438 * from a linear array which stores the auxiliar binary tree
440 static unsigned int to_inorder(unsigned int j, struct bset_tree *t)
442 return __to_inorder(j, t->size, t->extra);
445 static unsigned int __inorder_to_tree(unsigned int j,
457 j |= roundup_pow_of_two(size) >> shift;
463 * Return an index from a linear array which stores the auxiliar binary
464 * tree, j is the cacheline index of t->data.
466 static unsigned int inorder_to_tree(unsigned int j, struct bset_tree *t)
468 return __inorder_to_tree(j, t->size, t->extra);
472 void inorder_test(void)
474 unsigned long done = 0;
475 ktime_t start = ktime_get();
477 for (unsigned int size = 2;
481 (size - rounddown_pow_of_two(size - 1)) << 1;
482 unsigned int i = 1, j = rounddown_pow_of_two(size - 1);
485 pr_notice("loop %u, %llu per us\n", size,
486 done / ktime_us_delta(ktime_get(), start));
489 if (__inorder_to_tree(i, size, extra) != j)
490 panic("size %10u j %10u i %10u", size, j, i);
492 if (__to_inorder(j, size, extra) != i)
493 panic("size %10u j %10u i %10u", size, j, i);
495 if (j == rounddown_pow_of_two(size) - 1)
498 BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
500 j = inorder_next(j, size);
510 * Cacheline/offset <-> bkey pointer arithmetic:
512 * t->tree is a binary search tree in an array; each node corresponds to a key
513 * in one cacheline in t->set (BSET_CACHELINE bytes).
515 * This means we don't have to store the full index of the key that a node in
516 * the binary tree points to; to_inorder() gives us the cacheline, and then
517 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
519 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
522 * To construct the bfloat for an arbitrary key we need to know what the key
523 * immediately preceding it is: we have to check if the two keys differ in the
524 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
525 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
528 static struct bkey *cacheline_to_bkey(struct bset_tree *t,
529 unsigned int cacheline,
532 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
535 static unsigned int bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
537 return ((void *) k - (void *) t->data) / BSET_CACHELINE;
540 static unsigned int bkey_to_cacheline_offset(struct bset_tree *t,
541 unsigned int cacheline,
544 return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0);
547 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned int j)
549 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
552 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned int j)
554 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
558 * For the write set - the one we're currently inserting keys into - we don't
559 * maintain a full search tree, we just keep a simple lookup table in t->prev.
561 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned int cacheline)
563 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
566 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
569 low |= (high << 1) << (63U - shift);
574 * Calculate mantissa value for struct bkey_float.
575 * If most significant bit of f->exponent is not set, then
576 * - f->exponent >> 6 is 0
577 * - p[0] points to bkey->low
578 * - p[-1] borrows bits from KEY_INODE() of bkey->high
579 * if most isgnificant bits of f->exponent is set, then
580 * - f->exponent >> 6 is 1
581 * - p[0] points to bits from KEY_INODE() of bkey->high
582 * - p[-1] points to other bits from KEY_INODE() of
584 * See make_bfloat() to check when most significant bit of f->exponent
587 static inline unsigned int bfloat_mantissa(const struct bkey *k,
588 struct bkey_float *f)
590 const uint64_t *p = &k->low - (f->exponent >> 6);
592 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
595 static void make_bfloat(struct bset_tree *t, unsigned int j)
597 struct bkey_float *f = &t->tree[j];
598 struct bkey *m = tree_to_bkey(t, j);
599 struct bkey *p = tree_to_prev_bkey(t, j);
601 struct bkey *l = is_power_of_2(j)
603 : tree_to_prev_bkey(t, j >> ffs(j));
605 struct bkey *r = is_power_of_2(j + 1)
606 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
607 : tree_to_bkey(t, j >> (ffz(j) + 1));
609 BUG_ON(m < l || m > r);
610 BUG_ON(bkey_next(p) != m);
613 * If l and r have different KEY_INODE values (different backing
614 * device), f->exponent records how many least significant bits
615 * are different in KEY_INODE values and sets most significant
616 * bits to 1 (by +64).
617 * If l and r have same KEY_INODE value, f->exponent records
618 * how many different bits in least significant bits of bkey->low.
619 * See bfloat_mantiss() how the most significant bit of
620 * f->exponent is used to calculate bfloat mantissa value.
622 if (KEY_INODE(l) != KEY_INODE(r))
623 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
625 f->exponent = fls64(r->low ^ l->low);
627 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
630 * Setting f->exponent = 127 flags this node as failed, and causes the
631 * lookup code to fall back to comparing against the original key.
634 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
635 f->mantissa = bfloat_mantissa(m, f) - 1;
640 static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t)
643 unsigned int j = roundup(t[-1].size,
644 64 / sizeof(struct bkey_float));
646 t->tree = t[-1].tree + j;
647 t->prev = t[-1].prev + j;
650 while (t < b->set + MAX_BSETS)
654 static void bch_bset_build_unwritten_tree(struct btree_keys *b)
656 struct bset_tree *t = bset_tree_last(b);
658 BUG_ON(b->last_set_unwritten);
659 b->last_set_unwritten = 1;
661 bset_alloc_tree(b, t);
663 if (t->tree != b->set->tree + btree_keys_cachelines(b)) {
664 t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start);
669 void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic)
671 if (i != b->set->data) {
672 b->set[++b->nsets].data = i;
673 i->seq = b->set->data->seq;
675 get_random_bytes(&i->seq, sizeof(uint64_t));
681 bch_bset_build_unwritten_tree(b);
683 EXPORT_SYMBOL(bch_bset_init_next);
686 * Build auxiliary binary tree 'struct bset_tree *t', this tree is used to
687 * accelerate bkey search in a btree node (pointed by bset_tree->data in
688 * memory). After search in the auxiliar tree by calling bset_search_tree(),
689 * a struct bset_search_iter is returned which indicates range [l, r] from
690 * bset_tree->data where the searching bkey might be inside. Then a followed
691 * linear comparison does the exact search, see __bch_bset_search() for how
692 * the auxiliary tree is used.
694 void bch_bset_build_written_tree(struct btree_keys *b)
696 struct bset_tree *t = bset_tree_last(b);
697 struct bkey *prev = NULL, *k = t->data->start;
698 unsigned int j, cacheline = 1;
700 b->last_set_unwritten = 0;
702 bset_alloc_tree(b, t);
704 t->size = min_t(unsigned int,
705 bkey_to_cacheline(t, bset_bkey_last(t->data)),
706 b->set->tree + btree_keys_cachelines(b) - t->tree);
713 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
715 /* First we figure out where the first key in each cacheline is */
716 for (j = inorder_next(0, t->size);
718 j = inorder_next(j, t->size)) {
719 while (bkey_to_cacheline(t, k) < cacheline)
720 prev = k, k = bkey_next(k);
722 t->prev[j] = bkey_u64s(prev);
723 t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k);
726 while (bkey_next(k) != bset_bkey_last(t->data))
731 /* Then we build the tree */
732 for (j = inorder_next(0, t->size);
734 j = inorder_next(j, t->size))
737 EXPORT_SYMBOL(bch_bset_build_written_tree);
741 void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k)
744 unsigned int inorder, j = 1;
746 for (t = b->set; t <= bset_tree_last(b); t++)
747 if (k < bset_bkey_last(t->data))
752 if (!t->size || !bset_written(b, t))
755 inorder = bkey_to_cacheline(t, k);
757 if (k == t->data->start)
760 if (bkey_next(k) == bset_bkey_last(t->data)) {
765 j = inorder_to_tree(inorder, t);
769 k == tree_to_bkey(t, j))
773 } while (j < t->size);
775 j = inorder_to_tree(inorder + 1, t);
779 k == tree_to_prev_bkey(t, j))
783 } while (j < t->size);
785 EXPORT_SYMBOL(bch_bset_fix_invalidated_key);
787 static void bch_bset_fix_lookup_table(struct btree_keys *b,
791 unsigned int shift = bkey_u64s(k);
792 unsigned int j = bkey_to_cacheline(t, k);
794 /* We're getting called from btree_split() or btree_gc, just bail out */
799 * k is the key we just inserted; we need to find the entry in the
800 * lookup table for the first key that is strictly greater than k:
801 * it's either k's cacheline or the next one
803 while (j < t->size &&
804 table_to_bkey(t, j) <= k)
808 * Adjust all the lookup table entries, and find a new key for any that
809 * have gotten too big
811 for (; j < t->size; j++) {
814 if (t->prev[j] > 7) {
815 k = table_to_bkey(t, j - 1);
817 while (k < cacheline_to_bkey(t, j, 0))
820 t->prev[j] = bkey_to_cacheline_offset(t, j, k);
824 if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree)
827 /* Possibly add a new entry to the end of the lookup table */
829 for (k = table_to_bkey(t, t->size - 1);
830 k != bset_bkey_last(t->data);
832 if (t->size == bkey_to_cacheline(t, k)) {
834 bkey_to_cacheline_offset(t, t->size, k);
840 * Tries to merge l and r: l should be lower than r
841 * Returns true if we were able to merge. If we did merge, l will be the merged
842 * key, r will be untouched.
844 bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r)
846 if (!b->ops->key_merge)
850 * Generic header checks
851 * Assumes left and right are in order
852 * Left and right must be exactly aligned
854 if (!bch_bkey_equal_header(l, r) ||
855 bkey_cmp(l, &START_KEY(r)))
858 return b->ops->key_merge(b, l, r);
860 EXPORT_SYMBOL(bch_bkey_try_merge);
862 void bch_bset_insert(struct btree_keys *b, struct bkey *where,
865 struct bset_tree *t = bset_tree_last(b);
867 BUG_ON(!b->last_set_unwritten);
868 BUG_ON(bset_byte_offset(b, t->data) +
869 __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
870 PAGE_SIZE << b->page_order);
872 memmove((uint64_t *) where + bkey_u64s(insert),
874 (void *) bset_bkey_last(t->data) - (void *) where);
876 t->data->keys += bkey_u64s(insert);
877 bkey_copy(where, insert);
878 bch_bset_fix_lookup_table(b, t, where);
880 EXPORT_SYMBOL(bch_bset_insert);
882 unsigned int bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
883 struct bkey *replace_key)
885 unsigned int status = BTREE_INSERT_STATUS_NO_INSERT;
886 struct bset *i = bset_tree_last(b)->data;
887 struct bkey *m, *prev = NULL;
888 struct btree_iter iter;
889 struct bkey preceding_key_on_stack = ZERO_KEY;
890 struct bkey *preceding_key_p = &preceding_key_on_stack;
892 BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
895 * If k has preceding key, preceding_key_p will be set to address
896 * of k's preceding key; otherwise preceding_key_p will be set
897 * to NULL inside preceding_key().
899 if (b->ops->is_extents)
900 preceding_key(&START_KEY(k), &preceding_key_p);
902 preceding_key(k, &preceding_key_p);
904 m = bch_btree_iter_init(b, &iter, preceding_key_p);
906 if (b->ops->insert_fixup(b, k, &iter, replace_key))
909 status = BTREE_INSERT_STATUS_INSERT;
911 while (m != bset_bkey_last(i) &&
912 bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0)
913 prev = m, m = bkey_next(m);
915 /* prev is in the tree, if we merge we're done */
916 status = BTREE_INSERT_STATUS_BACK_MERGE;
918 bch_bkey_try_merge(b, prev, k))
921 status = BTREE_INSERT_STATUS_OVERWROTE;
922 if (m != bset_bkey_last(i) &&
923 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
926 status = BTREE_INSERT_STATUS_FRONT_MERGE;
927 if (m != bset_bkey_last(i) &&
928 bch_bkey_try_merge(b, k, m))
931 bch_bset_insert(b, m, k);
932 copy: bkey_copy(m, k);
936 EXPORT_SYMBOL(bch_btree_insert_key);
940 struct bset_search_iter {
944 static struct bset_search_iter bset_search_write_set(struct bset_tree *t,
945 const struct bkey *search)
947 unsigned int li = 0, ri = t->size;
949 while (li + 1 != ri) {
950 unsigned int m = (li + ri) >> 1;
952 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
958 return (struct bset_search_iter) {
959 table_to_bkey(t, li),
960 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
964 static struct bset_search_iter bset_search_tree(struct bset_tree *t,
965 const struct bkey *search)
968 struct bkey_float *f;
969 unsigned int inorder, j, n = 1;
972 unsigned int p = n << 4;
975 prefetch(&t->tree[p]);
980 if (likely(f->exponent != 127)) {
981 if (f->mantissa >= bfloat_mantissa(search, f))
986 if (bkey_cmp(tree_to_bkey(t, j), search) > 0)
991 } while (n < t->size);
993 inorder = to_inorder(j, t);
996 * n would have been the node we recursed to - the low bit tells us if
997 * we recursed left or recursed right.
1000 l = cacheline_to_bkey(t, inorder, f->m);
1002 if (++inorder != t->size) {
1003 f = &t->tree[inorder_next(j, t->size)];
1004 r = cacheline_to_bkey(t, inorder, f->m);
1006 r = bset_bkey_last(t->data);
1008 r = cacheline_to_bkey(t, inorder, f->m);
1011 f = &t->tree[inorder_prev(j, t->size)];
1012 l = cacheline_to_bkey(t, inorder, f->m);
1017 return (struct bset_search_iter) {l, r};
1020 struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
1021 const struct bkey *search)
1023 struct bset_search_iter i;
1026 * First, we search for a cacheline, then lastly we do a linear search
1027 * within that cacheline.
1029 * To search for the cacheline, there's three different possibilities:
1030 * * The set is too small to have a search tree, so we just do a linear
1031 * search over the whole set.
1032 * * The set is the one we're currently inserting into; keeping a full
1033 * auxiliary search tree up to date would be too expensive, so we
1034 * use a much simpler lookup table to do a binary search -
1035 * bset_search_write_set().
1036 * * Or we use the auxiliary search tree we constructed earlier -
1037 * bset_search_tree()
1040 if (unlikely(!t->size)) {
1041 i.l = t->data->start;
1042 i.r = bset_bkey_last(t->data);
1043 } else if (bset_written(b, t)) {
1045 * Each node in the auxiliary search tree covers a certain range
1046 * of bits, and keys above and below the set it covers might
1047 * differ outside those bits - so we have to special case the
1048 * start and end - handle that here:
1051 if (unlikely(bkey_cmp(search, &t->end) >= 0))
1052 return bset_bkey_last(t->data);
1054 if (unlikely(bkey_cmp(search, t->data->start) < 0))
1055 return t->data->start;
1057 i = bset_search_tree(t, search);
1060 t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
1062 i = bset_search_write_set(t, search);
1065 if (btree_keys_expensive_checks(b)) {
1066 BUG_ON(bset_written(b, t) &&
1067 i.l != t->data->start &&
1068 bkey_cmp(tree_to_prev_bkey(t,
1069 inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
1072 BUG_ON(i.r != bset_bkey_last(t->data) &&
1073 bkey_cmp(i.r, search) <= 0);
1076 while (likely(i.l != i.r) &&
1077 bkey_cmp(i.l, search) <= 0)
1078 i.l = bkey_next(i.l);
1082 EXPORT_SYMBOL(__bch_bset_search);
1084 /* Btree iterator */
1086 typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
1087 struct btree_iter_set);
1089 static inline bool btree_iter_cmp(struct btree_iter_set l,
1090 struct btree_iter_set r)
1092 return bkey_cmp(l.k, r.k) > 0;
1095 static inline bool btree_iter_end(struct btree_iter *iter)
1100 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
1104 BUG_ON(!heap_add(iter,
1105 ((struct btree_iter_set) { k, end }),
1109 static struct bkey *__bch_btree_iter_init(struct btree_keys *b,
1110 struct btree_iter *iter,
1111 struct bkey *search,
1112 struct bset_tree *start)
1114 struct bkey *ret = NULL;
1116 iter->size = ARRAY_SIZE(iter->data);
1119 #ifdef CONFIG_BCACHE_DEBUG
1123 for (; start <= bset_tree_last(b); start++) {
1124 ret = bch_bset_search(b, start, search);
1125 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
1131 struct bkey *bch_btree_iter_init(struct btree_keys *b,
1132 struct btree_iter *iter,
1133 struct bkey *search)
1135 return __bch_btree_iter_init(b, iter, search, b->set);
1137 EXPORT_SYMBOL(bch_btree_iter_init);
1139 static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
1140 btree_iter_cmp_fn *cmp)
1142 struct btree_iter_set b __maybe_unused;
1143 struct bkey *ret = NULL;
1145 if (!btree_iter_end(iter)) {
1146 bch_btree_iter_next_check(iter);
1148 ret = iter->data->k;
1149 iter->data->k = bkey_next(iter->data->k);
1151 if (iter->data->k > iter->data->end) {
1152 WARN_ONCE(1, "bset was corrupt!\n");
1153 iter->data->k = iter->data->end;
1156 if (iter->data->k == iter->data->end)
1157 heap_pop(iter, b, cmp);
1159 heap_sift(iter, 0, cmp);
1165 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
1167 return __bch_btree_iter_next(iter, btree_iter_cmp);
1170 EXPORT_SYMBOL(bch_btree_iter_next);
1172 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
1173 struct btree_keys *b, ptr_filter_fn fn)
1178 ret = bch_btree_iter_next(iter);
1179 } while (ret && fn(b, ret));
1186 void bch_bset_sort_state_free(struct bset_sort_state *state)
1188 mempool_exit(&state->pool);
1191 int bch_bset_sort_state_init(struct bset_sort_state *state,
1192 unsigned int page_order)
1194 spin_lock_init(&state->time.lock);
1196 state->page_order = page_order;
1197 state->crit_factor = int_sqrt(1 << page_order);
1199 return mempool_init_page_pool(&state->pool, 1, page_order);
1201 EXPORT_SYMBOL(bch_bset_sort_state_init);
1203 static void btree_mergesort(struct btree_keys *b, struct bset *out,
1204 struct btree_iter *iter,
1205 bool fixup, bool remove_stale)
1208 struct bkey *k, *last = NULL;
1210 bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale
1214 /* Heapify the iterator, using our comparison function */
1215 for (i = iter->used / 2 - 1; i >= 0; --i)
1216 heap_sift(iter, i, b->ops->sort_cmp);
1218 while (!btree_iter_end(iter)) {
1219 if (b->ops->sort_fixup && fixup)
1220 k = b->ops->sort_fixup(iter, &tmp.k);
1225 k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
1233 } else if (!bch_bkey_try_merge(b, last, k)) {
1234 last = bkey_next(last);
1239 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1241 pr_debug("sorted %i keys", out->keys);
1244 static void __btree_sort(struct btree_keys *b, struct btree_iter *iter,
1245 unsigned int start, unsigned int order, bool fixup,
1246 struct bset_sort_state *state)
1248 uint64_t start_time;
1249 bool used_mempool = false;
1250 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT,
1255 BUG_ON(order > state->page_order);
1257 outp = mempool_alloc(&state->pool, GFP_NOIO);
1258 out = page_address(outp);
1259 used_mempool = true;
1260 order = state->page_order;
1263 start_time = local_clock();
1265 btree_mergesort(b, out, iter, fixup, false);
1268 if (!start && order == b->page_order) {
1270 * Our temporary buffer is the same size as the btree node's
1271 * buffer, we can just swap buffers instead of doing a big
1275 out->magic = b->set->data->magic;
1276 out->seq = b->set->data->seq;
1277 out->version = b->set->data->version;
1278 swap(out, b->set->data);
1280 b->set[start].data->keys = out->keys;
1281 memcpy(b->set[start].data->start, out->start,
1282 (void *) bset_bkey_last(out) - (void *) out->start);
1286 mempool_free(virt_to_page(out), &state->pool);
1288 free_pages((unsigned long) out, order);
1290 bch_bset_build_written_tree(b);
1293 bch_time_stats_update(&state->time, start_time);
1296 void bch_btree_sort_partial(struct btree_keys *b, unsigned int start,
1297 struct bset_sort_state *state)
1299 size_t order = b->page_order, keys = 0;
1300 struct btree_iter iter;
1301 int oldsize = bch_count_data(b);
1303 __bch_btree_iter_init(b, &iter, NULL, &b->set[start]);
1308 for (i = start; i <= b->nsets; i++)
1309 keys += b->set[i].data->keys;
1311 order = get_order(__set_bytes(b->set->data, keys));
1314 __btree_sort(b, &iter, start, order, false, state);
1316 EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize);
1318 EXPORT_SYMBOL(bch_btree_sort_partial);
1320 void bch_btree_sort_and_fix_extents(struct btree_keys *b,
1321 struct btree_iter *iter,
1322 struct bset_sort_state *state)
1324 __btree_sort(b, iter, 0, b->page_order, true, state);
1327 void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
1328 struct bset_sort_state *state)
1330 uint64_t start_time = local_clock();
1331 struct btree_iter iter;
1333 bch_btree_iter_init(b, &iter, NULL);
1335 btree_mergesort(b, new->set->data, &iter, false, true);
1337 bch_time_stats_update(&state->time, start_time);
1339 new->set->size = 0; // XXX: why?
1342 #define SORT_CRIT (4096 / sizeof(uint64_t))
1344 void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state)
1346 unsigned int crit = SORT_CRIT;
1349 /* Don't sort if nothing to do */
1353 for (i = b->nsets - 1; i >= 0; --i) {
1354 crit *= state->crit_factor;
1356 if (b->set[i].data->keys < crit) {
1357 bch_btree_sort_partial(b, i, state);
1362 /* Sort if we'd overflow */
1363 if (b->nsets + 1 == MAX_BSETS) {
1364 bch_btree_sort(b, state);
1369 bch_bset_build_written_tree(b);
1371 EXPORT_SYMBOL(bch_btree_sort_lazy);
1373 void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats)
1377 for (i = 0; i <= b->nsets; i++) {
1378 struct bset_tree *t = &b->set[i];
1379 size_t bytes = t->data->keys * sizeof(uint64_t);
1382 if (bset_written(b, t)) {
1383 stats->sets_written++;
1384 stats->bytes_written += bytes;
1386 stats->floats += t->size - 1;
1388 for (j = 1; j < t->size; j++)
1389 if (t->tree[j].exponent == 127)
1392 stats->sets_unwritten++;
1393 stats->bytes_unwritten += bytes;