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 set)
23 struct bkey *k, *next;
25 for (k = i->start; k < bset_bkey_last(i); k = next) {
28 printk(KERN_ERR "block %u key %u/%u: ", set,
29 (unsigned) ((u64 *) k - i->d), i->keys);
32 b->ops->key_dump(b, k);
34 printk("%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 printk(KERN_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 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 struct bkey *bch_keylist_pop(struct keylist *l)
160 struct bkey *k = l->keys;
165 while (bkey_next(k) != l->top)
171 void bch_keylist_pop_front(struct keylist *l)
173 l->top_p -= bkey_u64s(l->keys);
177 bch_keylist_bytes(l));
180 /* Key/pointer manipulation */
182 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
185 BUG_ON(i > KEY_PTRS(src));
187 /* Only copy the header, key, and one pointer. */
188 memcpy(dest, src, 2 * sizeof(uint64_t));
189 dest->ptr[0] = src->ptr[i];
190 SET_KEY_PTRS(dest, 1);
191 /* We didn't copy the checksum so clear that bit. */
192 SET_KEY_CSUM(dest, 0);
195 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
199 if (bkey_cmp(where, &START_KEY(k)) <= 0)
202 if (bkey_cmp(where, k) < 0)
203 len = KEY_OFFSET(k) - KEY_OFFSET(where);
205 bkey_copy_key(k, where);
207 for (i = 0; i < KEY_PTRS(k); i++)
208 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
210 BUG_ON(len > KEY_SIZE(k));
211 SET_KEY_SIZE(k, len);
215 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
219 if (bkey_cmp(where, k) >= 0)
222 BUG_ON(KEY_INODE(where) != KEY_INODE(k));
224 if (bkey_cmp(where, &START_KEY(k)) > 0)
225 len = KEY_OFFSET(where) - KEY_START(k);
227 bkey_copy_key(k, where);
229 BUG_ON(len > KEY_SIZE(k));
230 SET_KEY_SIZE(k, len);
234 /* Auxiliary search trees */
237 #define BKEY_MID_BITS 3
238 #define BKEY_EXPONENT_BITS 7
239 #define BKEY_MANTISSA_BITS (32 - BKEY_MID_BITS - BKEY_EXPONENT_BITS)
240 #define BKEY_MANTISSA_MASK ((1 << BKEY_MANTISSA_BITS) - 1)
243 unsigned exponent:BKEY_EXPONENT_BITS;
244 unsigned m:BKEY_MID_BITS;
245 unsigned mantissa:BKEY_MANTISSA_BITS;
249 * BSET_CACHELINE was originally intended to match the hardware cacheline size -
250 * it used to be 64, but I realized the lookup code would touch slightly less
251 * memory if it was 128.
253 * It definites the number of bytes (in struct bset) per struct bkey_float in
254 * the auxiliar search tree - when we're done searching the bset_float tree we
255 * have this many bytes left that we do a linear search over.
257 * Since (after level 5) every level of the bset_tree is on a new cacheline,
258 * we're touching one fewer cacheline in the bset tree in exchange for one more
259 * cacheline in the linear search - but the linear search might stop before it
260 * gets to the second cacheline.
263 #define BSET_CACHELINE 128
265 /* Space required for the btree node keys */
266 static inline size_t btree_keys_bytes(struct btree_keys *b)
268 return PAGE_SIZE << b->page_order;
271 static inline size_t btree_keys_cachelines(struct btree_keys *b)
273 return btree_keys_bytes(b) / BSET_CACHELINE;
276 /* Space required for the auxiliary search trees */
277 static inline size_t bset_tree_bytes(struct btree_keys *b)
279 return btree_keys_cachelines(b) * sizeof(struct bkey_float);
282 /* Space required for the prev pointers */
283 static inline size_t bset_prev_bytes(struct btree_keys *b)
285 return btree_keys_cachelines(b) * sizeof(uint8_t);
288 /* Memory allocation */
290 void bch_btree_keys_free(struct btree_keys *b)
292 struct bset_tree *t = b->set;
294 if (bset_prev_bytes(b) < PAGE_SIZE)
297 free_pages((unsigned long) t->prev,
298 get_order(bset_prev_bytes(b)));
300 if (bset_tree_bytes(b) < PAGE_SIZE)
303 free_pages((unsigned long) t->tree,
304 get_order(bset_tree_bytes(b)));
306 free_pages((unsigned long) t->data, b->page_order);
312 EXPORT_SYMBOL(bch_btree_keys_free);
314 int bch_btree_keys_alloc(struct btree_keys *b, unsigned page_order, gfp_t gfp)
316 struct bset_tree *t = b->set;
320 b->page_order = page_order;
322 t->data = (void *) __get_free_pages(__GFP_COMP|gfp, b->page_order);
326 t->tree = bset_tree_bytes(b) < PAGE_SIZE
327 ? kmalloc(bset_tree_bytes(b), gfp)
328 : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b)));
332 t->prev = bset_prev_bytes(b) < PAGE_SIZE
333 ? kmalloc(bset_prev_bytes(b), gfp)
334 : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b)));
340 bch_btree_keys_free(b);
343 EXPORT_SYMBOL(bch_btree_keys_alloc);
345 void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
346 bool *expensive_debug_checks)
351 b->expensive_debug_checks = expensive_debug_checks;
353 b->last_set_unwritten = 0;
355 /* XXX: shouldn't be needed */
356 for (i = 0; i < MAX_BSETS; i++)
359 * Second loop starts at 1 because b->keys[0]->data is the memory we
362 for (i = 1; i < MAX_BSETS; i++)
363 b->set[i].data = NULL;
365 EXPORT_SYMBOL(bch_btree_keys_init);
367 /* Binary tree stuff for auxiliary search trees */
369 static unsigned inorder_next(unsigned j, unsigned size)
371 if (j * 2 + 1 < size) {
382 static unsigned inorder_prev(unsigned j, unsigned size)
387 while (j * 2 + 1 < size)
395 /* I have no idea why this code works... and I'm the one who wrote it
397 * However, I do know what it does:
398 * Given a binary tree constructed in an array (i.e. how you normally implement
399 * a heap), it converts a node in the tree - referenced by array index - to the
400 * index it would have if you did an inorder traversal.
402 * Also tested for every j, size up to size somewhere around 6 million.
404 * The binary tree starts at array index 1, not 0
405 * extra is a function of size:
406 * extra = (size - rounddown_pow_of_two(size - 1)) << 1;
408 static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
411 unsigned shift = fls(size - 1) - b;
419 j -= (j - extra) >> 1;
424 static unsigned to_inorder(unsigned j, struct bset_tree *t)
426 return __to_inorder(j, t->size, t->extra);
429 static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
439 j |= roundup_pow_of_two(size) >> shift;
444 static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
446 return __inorder_to_tree(j, t->size, t->extra);
450 void inorder_test(void)
452 unsigned long done = 0;
453 ktime_t start = ktime_get();
455 for (unsigned size = 2;
458 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
459 unsigned i = 1, j = rounddown_pow_of_two(size - 1);
462 printk(KERN_NOTICE "loop %u, %llu per us\n", size,
463 done / ktime_us_delta(ktime_get(), start));
466 if (__inorder_to_tree(i, size, extra) != j)
467 panic("size %10u j %10u i %10u", size, j, i);
469 if (__to_inorder(j, size, extra) != i)
470 panic("size %10u j %10u i %10u", size, j, i);
472 if (j == rounddown_pow_of_two(size) - 1)
475 BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
477 j = inorder_next(j, size);
487 * Cacheline/offset <-> bkey pointer arithmetic:
489 * t->tree is a binary search tree in an array; each node corresponds to a key
490 * in one cacheline in t->set (BSET_CACHELINE bytes).
492 * This means we don't have to store the full index of the key that a node in
493 * the binary tree points to; to_inorder() gives us the cacheline, and then
494 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
496 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
499 * To construct the bfloat for an arbitrary key we need to know what the key
500 * immediately preceding it is: we have to check if the two keys differ in the
501 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
502 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
505 static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
508 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
511 static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
513 return ((void *) k - (void *) t->data) / BSET_CACHELINE;
516 static unsigned bkey_to_cacheline_offset(struct bset_tree *t,
520 return (u64 *) k - (u64 *) cacheline_to_bkey(t, cacheline, 0);
523 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
525 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
528 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
530 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
534 * For the write set - the one we're currently inserting keys into - we don't
535 * maintain a full search tree, we just keep a simple lookup table in t->prev.
537 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
539 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
542 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
545 low |= (high << 1) << (63U - shift);
549 static inline unsigned bfloat_mantissa(const struct bkey *k,
550 struct bkey_float *f)
552 const uint64_t *p = &k->low - (f->exponent >> 6);
553 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
556 static void make_bfloat(struct bset_tree *t, unsigned j)
558 struct bkey_float *f = &t->tree[j];
559 struct bkey *m = tree_to_bkey(t, j);
560 struct bkey *p = tree_to_prev_bkey(t, j);
562 struct bkey *l = is_power_of_2(j)
564 : tree_to_prev_bkey(t, j >> ffs(j));
566 struct bkey *r = is_power_of_2(j + 1)
567 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
568 : tree_to_bkey(t, j >> (ffz(j) + 1));
570 BUG_ON(m < l || m > r);
571 BUG_ON(bkey_next(p) != m);
573 if (KEY_INODE(l) != KEY_INODE(r))
574 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
576 f->exponent = fls64(r->low ^ l->low);
578 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
581 * Setting f->exponent = 127 flags this node as failed, and causes the
582 * lookup code to fall back to comparing against the original key.
585 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
586 f->mantissa = bfloat_mantissa(m, f) - 1;
591 static void bset_alloc_tree(struct btree_keys *b, struct bset_tree *t)
594 unsigned j = roundup(t[-1].size,
595 64 / sizeof(struct bkey_float));
597 t->tree = t[-1].tree + j;
598 t->prev = t[-1].prev + j;
601 while (t < b->set + MAX_BSETS)
605 static void bch_bset_build_unwritten_tree(struct btree_keys *b)
607 struct bset_tree *t = bset_tree_last(b);
609 BUG_ON(b->last_set_unwritten);
610 b->last_set_unwritten = 1;
612 bset_alloc_tree(b, t);
614 if (t->tree != b->set->tree + btree_keys_cachelines(b)) {
615 t->prev[0] = bkey_to_cacheline_offset(t, 0, t->data->start);
620 void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic)
622 if (i != b->set->data) {
623 b->set[++b->nsets].data = i;
624 i->seq = b->set->data->seq;
626 get_random_bytes(&i->seq, sizeof(uint64_t));
632 bch_bset_build_unwritten_tree(b);
634 EXPORT_SYMBOL(bch_bset_init_next);
636 void bch_bset_build_written_tree(struct btree_keys *b)
638 struct bset_tree *t = bset_tree_last(b);
639 struct bkey *prev = NULL, *k = t->data->start;
640 unsigned j, cacheline = 1;
642 b->last_set_unwritten = 0;
644 bset_alloc_tree(b, t);
646 t->size = min_t(unsigned,
647 bkey_to_cacheline(t, bset_bkey_last(t->data)),
648 b->set->tree + btree_keys_cachelines(b) - t->tree);
655 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
657 /* First we figure out where the first key in each cacheline is */
658 for (j = inorder_next(0, t->size);
660 j = inorder_next(j, t->size)) {
661 while (bkey_to_cacheline(t, k) < cacheline)
662 prev = k, k = bkey_next(k);
664 t->prev[j] = bkey_u64s(prev);
665 t->tree[j].m = bkey_to_cacheline_offset(t, cacheline++, k);
668 while (bkey_next(k) != bset_bkey_last(t->data))
673 /* Then we build the tree */
674 for (j = inorder_next(0, t->size);
676 j = inorder_next(j, t->size))
679 EXPORT_SYMBOL(bch_bset_build_written_tree);
683 void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k)
686 unsigned inorder, j = 1;
688 for (t = b->set; t <= bset_tree_last(b); t++)
689 if (k < bset_bkey_last(t->data))
694 if (!t->size || !bset_written(b, t))
697 inorder = bkey_to_cacheline(t, k);
699 if (k == t->data->start)
702 if (bkey_next(k) == bset_bkey_last(t->data)) {
707 j = inorder_to_tree(inorder, t);
711 k == tree_to_bkey(t, j))
715 } while (j < t->size);
717 j = inorder_to_tree(inorder + 1, t);
721 k == tree_to_prev_bkey(t, j))
725 } while (j < t->size);
727 EXPORT_SYMBOL(bch_bset_fix_invalidated_key);
729 static void bch_bset_fix_lookup_table(struct btree_keys *b,
733 unsigned shift = bkey_u64s(k);
734 unsigned j = bkey_to_cacheline(t, k);
736 /* We're getting called from btree_split() or btree_gc, just bail out */
740 /* k is the key we just inserted; we need to find the entry in the
741 * lookup table for the first key that is strictly greater than k:
742 * it's either k's cacheline or the next one
744 while (j < t->size &&
745 table_to_bkey(t, j) <= k)
748 /* Adjust all the lookup table entries, and find a new key for any that
749 * have gotten too big
751 for (; j < t->size; j++) {
754 if (t->prev[j] > 7) {
755 k = table_to_bkey(t, j - 1);
757 while (k < cacheline_to_bkey(t, j, 0))
760 t->prev[j] = bkey_to_cacheline_offset(t, j, k);
764 if (t->size == b->set->tree + btree_keys_cachelines(b) - t->tree)
767 /* Possibly add a new entry to the end of the lookup table */
769 for (k = table_to_bkey(t, t->size - 1);
770 k != bset_bkey_last(t->data);
772 if (t->size == bkey_to_cacheline(t, k)) {
773 t->prev[t->size] = bkey_to_cacheline_offset(t, t->size, k);
779 * Tries to merge l and r: l should be lower than r
780 * Returns true if we were able to merge. If we did merge, l will be the merged
781 * key, r will be untouched.
783 bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r)
785 if (!b->ops->key_merge)
789 * Generic header checks
790 * Assumes left and right are in order
791 * Left and right must be exactly aligned
793 if (!bch_bkey_equal_header(l, r) ||
794 bkey_cmp(l, &START_KEY(r)))
797 return b->ops->key_merge(b, l, r);
799 EXPORT_SYMBOL(bch_bkey_try_merge);
801 void bch_bset_insert(struct btree_keys *b, struct bkey *where,
804 struct bset_tree *t = bset_tree_last(b);
806 BUG_ON(!b->last_set_unwritten);
807 BUG_ON(bset_byte_offset(b, t->data) +
808 __set_bytes(t->data, t->data->keys + bkey_u64s(insert)) >
809 PAGE_SIZE << b->page_order);
811 memmove((uint64_t *) where + bkey_u64s(insert),
813 (void *) bset_bkey_last(t->data) - (void *) where);
815 t->data->keys += bkey_u64s(insert);
816 bkey_copy(where, insert);
817 bch_bset_fix_lookup_table(b, t, where);
819 EXPORT_SYMBOL(bch_bset_insert);
821 unsigned bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
822 struct bkey *replace_key)
824 unsigned status = BTREE_INSERT_STATUS_NO_INSERT;
825 struct bset *i = bset_tree_last(b)->data;
826 struct bkey *m, *prev = NULL;
827 struct btree_iter iter;
828 struct bkey preceding_key_on_stack = ZERO_KEY;
829 struct bkey *preceding_key_p = &preceding_key_on_stack;
831 BUG_ON(b->ops->is_extents && !KEY_SIZE(k));
834 * If k has preceding key, preceding_key_p will be set to address
835 * of k's preceding key; otherwise preceding_key_p will be set
836 * to NULL inside preceding_key().
838 if (b->ops->is_extents)
839 preceding_key(&START_KEY(k), &preceding_key_p);
841 preceding_key(k, &preceding_key_p);
843 m = bch_btree_iter_init(b, &iter, preceding_key_p);
845 if (b->ops->insert_fixup(b, k, &iter, replace_key))
848 status = BTREE_INSERT_STATUS_INSERT;
850 while (m != bset_bkey_last(i) &&
851 bkey_cmp(k, b->ops->is_extents ? &START_KEY(m) : m) > 0)
852 prev = m, m = bkey_next(m);
854 /* prev is in the tree, if we merge we're done */
855 status = BTREE_INSERT_STATUS_BACK_MERGE;
857 bch_bkey_try_merge(b, prev, k))
860 status = BTREE_INSERT_STATUS_OVERWROTE;
861 if (m != bset_bkey_last(i) &&
862 KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m))
865 status = BTREE_INSERT_STATUS_FRONT_MERGE;
866 if (m != bset_bkey_last(i) &&
867 bch_bkey_try_merge(b, k, m))
870 bch_bset_insert(b, m, k);
871 copy: bkey_copy(m, k);
875 EXPORT_SYMBOL(bch_btree_insert_key);
879 struct bset_search_iter {
883 static struct bset_search_iter bset_search_write_set(struct bset_tree *t,
884 const struct bkey *search)
886 unsigned li = 0, ri = t->size;
888 while (li + 1 != ri) {
889 unsigned m = (li + ri) >> 1;
891 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
897 return (struct bset_search_iter) {
898 table_to_bkey(t, li),
899 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
903 static struct bset_search_iter bset_search_tree(struct bset_tree *t,
904 const struct bkey *search)
907 struct bkey_float *f;
908 unsigned inorder, j, n = 1;
912 p &= ((int) (p - t->size)) >> 31;
914 prefetch(&t->tree[p]);
920 * n = (f->mantissa > bfloat_mantissa())
924 * We need to subtract 1 from f->mantissa for the sign bit trick
925 * to work - that's done in make_bfloat()
927 if (likely(f->exponent != 127))
928 n = j * 2 + (((unsigned)
930 bfloat_mantissa(search, f))) >> 31);
932 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
935 } while (n < t->size);
937 inorder = to_inorder(j, t);
940 * n would have been the node we recursed to - the low bit tells us if
941 * we recursed left or recursed right.
944 l = cacheline_to_bkey(t, inorder, f->m);
946 if (++inorder != t->size) {
947 f = &t->tree[inorder_next(j, t->size)];
948 r = cacheline_to_bkey(t, inorder, f->m);
950 r = bset_bkey_last(t->data);
952 r = cacheline_to_bkey(t, inorder, f->m);
955 f = &t->tree[inorder_prev(j, t->size)];
956 l = cacheline_to_bkey(t, inorder, f->m);
961 return (struct bset_search_iter) {l, r};
964 struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
965 const struct bkey *search)
967 struct bset_search_iter i;
970 * First, we search for a cacheline, then lastly we do a linear search
971 * within that cacheline.
973 * To search for the cacheline, there's three different possibilities:
974 * * The set is too small to have a search tree, so we just do a linear
975 * search over the whole set.
976 * * The set is the one we're currently inserting into; keeping a full
977 * auxiliary search tree up to date would be too expensive, so we
978 * use a much simpler lookup table to do a binary search -
979 * bset_search_write_set().
980 * * Or we use the auxiliary search tree we constructed earlier -
984 if (unlikely(!t->size)) {
985 i.l = t->data->start;
986 i.r = bset_bkey_last(t->data);
987 } else if (bset_written(b, t)) {
989 * Each node in the auxiliary search tree covers a certain range
990 * of bits, and keys above and below the set it covers might
991 * differ outside those bits - so we have to special case the
992 * start and end - handle that here:
995 if (unlikely(bkey_cmp(search, &t->end) >= 0))
996 return bset_bkey_last(t->data);
998 if (unlikely(bkey_cmp(search, t->data->start) < 0))
999 return t->data->start;
1001 i = bset_search_tree(t, search);
1004 t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
1006 i = bset_search_write_set(t, search);
1009 if (btree_keys_expensive_checks(b)) {
1010 BUG_ON(bset_written(b, t) &&
1011 i.l != t->data->start &&
1012 bkey_cmp(tree_to_prev_bkey(t,
1013 inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
1016 BUG_ON(i.r != bset_bkey_last(t->data) &&
1017 bkey_cmp(i.r, search) <= 0);
1020 while (likely(i.l != i.r) &&
1021 bkey_cmp(i.l, search) <= 0)
1022 i.l = bkey_next(i.l);
1026 EXPORT_SYMBOL(__bch_bset_search);
1028 /* Btree iterator */
1030 typedef bool (btree_iter_cmp_fn)(struct btree_iter_set,
1031 struct btree_iter_set);
1033 static inline bool btree_iter_cmp(struct btree_iter_set l,
1034 struct btree_iter_set r)
1036 return bkey_cmp(l.k, r.k) > 0;
1039 static inline bool btree_iter_end(struct btree_iter *iter)
1044 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
1048 BUG_ON(!heap_add(iter,
1049 ((struct btree_iter_set) { k, end }),
1053 static struct bkey *__bch_btree_iter_init(struct btree_keys *b,
1054 struct btree_iter *iter,
1055 struct bkey *search,
1056 struct bset_tree *start)
1058 struct bkey *ret = NULL;
1059 iter->size = ARRAY_SIZE(iter->data);
1062 #ifdef CONFIG_BCACHE_DEBUG
1066 for (; start <= bset_tree_last(b); start++) {
1067 ret = bch_bset_search(b, start, search);
1068 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
1074 struct bkey *bch_btree_iter_init(struct btree_keys *b,
1075 struct btree_iter *iter,
1076 struct bkey *search)
1078 return __bch_btree_iter_init(b, iter, search, b->set);
1080 EXPORT_SYMBOL(bch_btree_iter_init);
1082 static inline struct bkey *__bch_btree_iter_next(struct btree_iter *iter,
1083 btree_iter_cmp_fn *cmp)
1085 struct btree_iter_set unused;
1086 struct bkey *ret = NULL;
1088 if (!btree_iter_end(iter)) {
1089 bch_btree_iter_next_check(iter);
1091 ret = iter->data->k;
1092 iter->data->k = bkey_next(iter->data->k);
1094 if (iter->data->k > iter->data->end) {
1095 WARN_ONCE(1, "bset was corrupt!\n");
1096 iter->data->k = iter->data->end;
1099 if (iter->data->k == iter->data->end)
1100 heap_pop(iter, unused, cmp);
1102 heap_sift(iter, 0, cmp);
1108 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
1110 return __bch_btree_iter_next(iter, btree_iter_cmp);
1113 EXPORT_SYMBOL(bch_btree_iter_next);
1115 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
1116 struct btree_keys *b, ptr_filter_fn fn)
1121 ret = bch_btree_iter_next(iter);
1122 } while (ret && fn(b, ret));
1129 void bch_bset_sort_state_free(struct bset_sort_state *state)
1132 mempool_destroy(state->pool);
1135 int bch_bset_sort_state_init(struct bset_sort_state *state, unsigned page_order)
1137 spin_lock_init(&state->time.lock);
1139 state->page_order = page_order;
1140 state->crit_factor = int_sqrt(1 << page_order);
1142 state->pool = mempool_create_page_pool(1, page_order);
1148 EXPORT_SYMBOL(bch_bset_sort_state_init);
1150 static void btree_mergesort(struct btree_keys *b, struct bset *out,
1151 struct btree_iter *iter,
1152 bool fixup, bool remove_stale)
1155 struct bkey *k, *last = NULL;
1157 bool (*bad)(struct btree_keys *, const struct bkey *) = remove_stale
1161 /* Heapify the iterator, using our comparison function */
1162 for (i = iter->used / 2 - 1; i >= 0; --i)
1163 heap_sift(iter, i, b->ops->sort_cmp);
1165 while (!btree_iter_end(iter)) {
1166 if (b->ops->sort_fixup && fixup)
1167 k = b->ops->sort_fixup(iter, &tmp.k);
1172 k = __bch_btree_iter_next(iter, b->ops->sort_cmp);
1180 } else if (!bch_bkey_try_merge(b, last, k)) {
1181 last = bkey_next(last);
1186 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1188 pr_debug("sorted %i keys", out->keys);
1191 static void __btree_sort(struct btree_keys *b, struct btree_iter *iter,
1192 unsigned start, unsigned order, bool fixup,
1193 struct bset_sort_state *state)
1195 uint64_t start_time;
1196 bool used_mempool = false;
1197 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOWAIT,
1202 BUG_ON(order > state->page_order);
1204 outp = mempool_alloc(state->pool, GFP_NOIO);
1205 out = page_address(outp);
1206 used_mempool = true;
1207 order = state->page_order;
1210 start_time = local_clock();
1212 btree_mergesort(b, out, iter, fixup, false);
1215 if (!start && order == b->page_order) {
1217 * Our temporary buffer is the same size as the btree node's
1218 * buffer, we can just swap buffers instead of doing a big
1222 out->magic = b->set->data->magic;
1223 out->seq = b->set->data->seq;
1224 out->version = b->set->data->version;
1225 swap(out, b->set->data);
1227 b->set[start].data->keys = out->keys;
1228 memcpy(b->set[start].data->start, out->start,
1229 (void *) bset_bkey_last(out) - (void *) out->start);
1233 mempool_free(virt_to_page(out), state->pool);
1235 free_pages((unsigned long) out, order);
1237 bch_bset_build_written_tree(b);
1240 bch_time_stats_update(&state->time, start_time);
1243 void bch_btree_sort_partial(struct btree_keys *b, unsigned start,
1244 struct bset_sort_state *state)
1246 size_t order = b->page_order, keys = 0;
1247 struct btree_iter iter;
1248 int oldsize = bch_count_data(b);
1250 __bch_btree_iter_init(b, &iter, NULL, &b->set[start]);
1255 for (i = start; i <= b->nsets; i++)
1256 keys += b->set[i].data->keys;
1258 order = get_order(__set_bytes(b->set->data, keys));
1261 __btree_sort(b, &iter, start, order, false, state);
1263 EBUG_ON(oldsize >= 0 && bch_count_data(b) != oldsize);
1265 EXPORT_SYMBOL(bch_btree_sort_partial);
1267 void bch_btree_sort_and_fix_extents(struct btree_keys *b,
1268 struct btree_iter *iter,
1269 struct bset_sort_state *state)
1271 __btree_sort(b, iter, 0, b->page_order, true, state);
1274 void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
1275 struct bset_sort_state *state)
1277 uint64_t start_time = local_clock();
1279 struct btree_iter iter;
1280 bch_btree_iter_init(b, &iter, NULL);
1282 btree_mergesort(b, new->set->data, &iter, false, true);
1284 bch_time_stats_update(&state->time, start_time);
1286 new->set->size = 0; // XXX: why?
1289 #define SORT_CRIT (4096 / sizeof(uint64_t))
1291 void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state)
1293 unsigned crit = SORT_CRIT;
1296 /* Don't sort if nothing to do */
1300 for (i = b->nsets - 1; i >= 0; --i) {
1301 crit *= state->crit_factor;
1303 if (b->set[i].data->keys < crit) {
1304 bch_btree_sort_partial(b, i, state);
1309 /* Sort if we'd overflow */
1310 if (b->nsets + 1 == MAX_BSETS) {
1311 bch_btree_sort(b, state);
1316 bch_bset_build_written_tree(b);
1318 EXPORT_SYMBOL(bch_btree_sort_lazy);
1320 void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *stats)
1324 for (i = 0; i <= b->nsets; i++) {
1325 struct bset_tree *t = &b->set[i];
1326 size_t bytes = t->data->keys * sizeof(uint64_t);
1329 if (bset_written(b, t)) {
1330 stats->sets_written++;
1331 stats->bytes_written += bytes;
1333 stats->floats += t->size - 1;
1335 for (j = 1; j < t->size; j++)
1336 if (t->tree[j].exponent == 127)
1339 stats->sets_unwritten++;
1340 stats->bytes_unwritten += bytes;