2 * Code for working with individual keys, and sorted sets of keys with in a
5 * Copyright 2012 Google, Inc.
12 #include <linux/random.h>
13 #include <linux/prefetch.h>
17 void bch_keylist_copy(struct keylist *dest, struct keylist *src)
21 if (src->list == src->d) {
22 size_t n = (uint64_t *) src->top - src->d;
23 dest->top = (struct bkey *) &dest->d[n];
28 int bch_keylist_realloc(struct keylist *l, int nptrs, struct cache_set *c)
30 unsigned oldsize = (uint64_t *) l->top - l->list;
31 unsigned newsize = oldsize + 2 + nptrs;
34 /* The journalling code doesn't handle the case where the keys to insert
35 * is bigger than an empty write: If we just return -ENOMEM here,
36 * bio_insert() and bio_invalidate() will insert the keys created so far
37 * and finish the rest when the keylist is empty.
39 if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
42 newsize = roundup_pow_of_two(newsize);
44 if (newsize <= KEYLIST_INLINE ||
45 roundup_pow_of_two(oldsize) == newsize)
48 new = krealloc(l->list == l->d ? NULL : l->list,
49 sizeof(uint64_t) * newsize, GFP_NOIO);
55 memcpy(new, l->list, sizeof(uint64_t) * KEYLIST_INLINE);
58 l->top = (struct bkey *) (&l->list[oldsize]);
63 struct bkey *bch_keylist_pop(struct keylist *l)
65 struct bkey *k = l->bottom;
70 while (bkey_next(k) != l->top)
76 /* Pointer validation */
78 bool __bch_ptr_invalid(struct cache_set *c, int level, const struct bkey *k)
83 if (level && (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k)))
86 if (!level && KEY_SIZE(k) > KEY_OFFSET(k))
92 for (i = 0; i < KEY_PTRS(k); i++)
93 if (ptr_available(c, k, i)) {
94 struct cache *ca = PTR_CACHE(c, k, i);
95 size_t bucket = PTR_BUCKET_NR(c, k, i);
96 size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
98 if (KEY_SIZE(k) + r > c->sb.bucket_size ||
99 bucket < ca->sb.first_bucket ||
100 bucket >= ca->sb.nbuckets)
106 bch_bkey_to_text(buf, sizeof(buf), k);
107 cache_bug(c, "spotted bad key %s: %s", buf, bch_ptr_status(c, k));
111 bool bch_ptr_bad(struct btree *b, const struct bkey *k)
116 if (!bkey_cmp(k, &ZERO_KEY) ||
118 bch_ptr_invalid(b, k))
121 if (KEY_PTRS(k) && PTR_DEV(k, 0) == PTR_CHECK_DEV)
124 for (i = 0; i < KEY_PTRS(k); i++)
125 if (ptr_available(b->c, k, i)) {
126 g = PTR_BUCKET(b->c, k, i);
127 stale = ptr_stale(b->c, k, i);
129 btree_bug_on(stale > 96, b,
130 "key too stale: %i, need_gc %u",
131 stale, b->c->need_gc);
133 btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k),
134 b, "stale dirty pointer");
139 #ifdef CONFIG_BCACHE_EDEBUG
140 if (!mutex_trylock(&b->c->bucket_lock))
145 g->prio != BTREE_PRIO ||
146 (b->c->gc_mark_valid &&
147 GC_MARK(g) != GC_MARK_METADATA))
151 if (g->prio == BTREE_PRIO)
155 b->c->gc_mark_valid &&
156 GC_MARK(g) != GC_MARK_DIRTY)
159 mutex_unlock(&b->c->bucket_lock);
164 #ifdef CONFIG_BCACHE_EDEBUG
166 mutex_unlock(&b->c->bucket_lock);
171 bch_bkey_to_text(buf, sizeof(buf), k);
173 "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i",
174 buf, PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin),
175 g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen);
181 /* Key/pointer manipulation */
183 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
186 BUG_ON(i > KEY_PTRS(src));
188 /* Only copy the header, key, and one pointer. */
189 memcpy(dest, src, 2 * sizeof(uint64_t));
190 dest->ptr[0] = src->ptr[i];
191 SET_KEY_PTRS(dest, 1);
192 /* We didn't copy the checksum so clear that bit. */
193 SET_KEY_CSUM(dest, 0);
196 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
200 if (bkey_cmp(where, &START_KEY(k)) <= 0)
203 if (bkey_cmp(where, k) < 0)
204 len = KEY_OFFSET(k) - KEY_OFFSET(where);
206 bkey_copy_key(k, where);
208 for (i = 0; i < KEY_PTRS(k); i++)
209 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
211 BUG_ON(len > KEY_SIZE(k));
212 SET_KEY_SIZE(k, len);
216 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
220 if (bkey_cmp(where, k) >= 0)
223 BUG_ON(KEY_INODE(where) != KEY_INODE(k));
225 if (bkey_cmp(where, &START_KEY(k)) > 0)
226 len = KEY_OFFSET(where) - KEY_START(k);
228 bkey_copy_key(k, where);
230 BUG_ON(len > KEY_SIZE(k));
231 SET_KEY_SIZE(k, len);
235 static uint64_t merge_chksums(struct bkey *l, struct bkey *r)
237 return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) &
238 ~((uint64_t)1 << 63);
241 /* Tries to merge l and r: l should be lower than r
242 * Returns true if we were able to merge. If we did merge, l will be the merged
243 * key, r will be untouched.
245 bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r)
249 if (key_merging_disabled(b->c))
252 if (KEY_PTRS(l) != KEY_PTRS(r) ||
253 KEY_DIRTY(l) != KEY_DIRTY(r) ||
254 bkey_cmp(l, &START_KEY(r)))
257 for (i = 0; i < KEY_PTRS(l); i++)
258 if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] ||
259 PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i))
262 /* Keys with no pointers aren't restricted to one bucket and could
265 if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) {
266 SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l));
267 SET_KEY_SIZE(l, USHRT_MAX);
275 l->ptr[KEY_PTRS(l)] = merge_chksums(l, r);
280 SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r));
281 SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r));
286 /* Binary tree stuff for auxiliary search trees */
288 static unsigned inorder_next(unsigned j, unsigned size)
290 if (j * 2 + 1 < size) {
301 static unsigned inorder_prev(unsigned j, unsigned size)
306 while (j * 2 + 1 < size)
314 /* I have no idea why this code works... and I'm the one who wrote it
316 * However, I do know what it does:
317 * Given a binary tree constructed in an array (i.e. how you normally implement
318 * a heap), it converts a node in the tree - referenced by array index - to the
319 * index it would have if you did an inorder traversal.
321 * Also tested for every j, size up to size somewhere around 6 million.
323 * The binary tree starts at array index 1, not 0
324 * extra is a function of size:
325 * extra = (size - rounddown_pow_of_two(size - 1)) << 1;
327 static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
330 unsigned shift = fls(size - 1) - b;
338 j -= (j - extra) >> 1;
343 static unsigned to_inorder(unsigned j, struct bset_tree *t)
345 return __to_inorder(j, t->size, t->extra);
348 static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
358 j |= roundup_pow_of_two(size) >> shift;
363 static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
365 return __inorder_to_tree(j, t->size, t->extra);
369 void inorder_test(void)
371 unsigned long done = 0;
372 ktime_t start = ktime_get();
374 for (unsigned size = 2;
377 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
378 unsigned i = 1, j = rounddown_pow_of_two(size - 1);
381 printk(KERN_NOTICE "loop %u, %llu per us\n", size,
382 done / ktime_us_delta(ktime_get(), start));
385 if (__inorder_to_tree(i, size, extra) != j)
386 panic("size %10u j %10u i %10u", size, j, i);
388 if (__to_inorder(j, size, extra) != i)
389 panic("size %10u j %10u i %10u", size, j, i);
391 if (j == rounddown_pow_of_two(size) - 1)
394 BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
396 j = inorder_next(j, size);
406 * Cacheline/offset <-> bkey pointer arithmetic:
408 * t->tree is a binary search tree in an array; each node corresponds to a key
409 * in one cacheline in t->set (BSET_CACHELINE bytes).
411 * This means we don't have to store the full index of the key that a node in
412 * the binary tree points to; to_inorder() gives us the cacheline, and then
413 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
415 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
418 * To construct the bfloat for an arbitrary key we need to know what the key
419 * immediately preceding it is: we have to check if the two keys differ in the
420 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
421 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
424 static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
427 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
430 static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
432 return ((void *) k - (void *) t->data) / BSET_CACHELINE;
435 static unsigned bkey_to_cacheline_offset(struct bkey *k)
437 return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t);
440 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
442 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
445 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
447 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
451 * For the write set - the one we're currently inserting keys into - we don't
452 * maintain a full search tree, we just keep a simple lookup table in t->prev.
454 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
456 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
459 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
462 asm("shrd %[shift],%[high],%[low]"
469 low |= (high << 1) << (63U - shift);
474 static inline unsigned bfloat_mantissa(const struct bkey *k,
475 struct bkey_float *f)
477 const uint64_t *p = &k->low - (f->exponent >> 6);
478 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
481 static void make_bfloat(struct bset_tree *t, unsigned j)
483 struct bkey_float *f = &t->tree[j];
484 struct bkey *m = tree_to_bkey(t, j);
485 struct bkey *p = tree_to_prev_bkey(t, j);
487 struct bkey *l = is_power_of_2(j)
489 : tree_to_prev_bkey(t, j >> ffs(j));
491 struct bkey *r = is_power_of_2(j + 1)
492 ? node(t->data, t->data->keys - bkey_u64s(&t->end))
493 : tree_to_bkey(t, j >> (ffz(j) + 1));
495 BUG_ON(m < l || m > r);
496 BUG_ON(bkey_next(p) != m);
498 if (KEY_INODE(l) != KEY_INODE(r))
499 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
501 f->exponent = fls64(r->low ^ l->low);
503 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
506 * Setting f->exponent = 127 flags this node as failed, and causes the
507 * lookup code to fall back to comparing against the original key.
510 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
511 f->mantissa = bfloat_mantissa(m, f) - 1;
516 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
519 unsigned j = roundup(t[-1].size,
520 64 / sizeof(struct bkey_float));
522 t->tree = t[-1].tree + j;
523 t->prev = t[-1].prev + j;
526 while (t < b->sets + MAX_BSETS)
530 static void bset_build_unwritten_tree(struct btree *b)
532 struct bset_tree *t = b->sets + b->nsets;
534 bset_alloc_tree(b, t);
536 if (t->tree != b->sets->tree + bset_tree_space(b)) {
537 t->prev[0] = bkey_to_cacheline_offset(t->data->start);
542 static void bset_build_written_tree(struct btree *b)
544 struct bset_tree *t = b->sets + b->nsets;
545 struct bkey *k = t->data->start;
546 unsigned j, cacheline = 1;
548 bset_alloc_tree(b, t);
550 t->size = min_t(unsigned,
551 bkey_to_cacheline(t, end(t->data)),
552 b->sets->tree + bset_tree_space(b) - t->tree);
559 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
561 /* First we figure out where the first key in each cacheline is */
562 for (j = inorder_next(0, t->size);
564 j = inorder_next(j, t->size)) {
565 while (bkey_to_cacheline(t, k) != cacheline)
568 t->prev[j] = bkey_u64s(k);
571 t->tree[j].m = bkey_to_cacheline_offset(k);
574 while (bkey_next(k) != end(t->data))
579 /* Then we build the tree */
580 for (j = inorder_next(0, t->size);
582 j = inorder_next(j, t->size))
586 void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k)
589 unsigned inorder, j = 1;
591 for (t = b->sets; t <= &b->sets[b->nsets]; t++)
592 if (k < end(t->data))
597 if (!t->size || !bset_written(b, t))
600 inorder = bkey_to_cacheline(t, k);
602 if (k == t->data->start)
605 if (bkey_next(k) == end(t->data)) {
610 j = inorder_to_tree(inorder, t);
614 k == tree_to_bkey(t, j))
618 } while (j < t->size);
620 j = inorder_to_tree(inorder + 1, t);
624 k == tree_to_prev_bkey(t, j))
628 } while (j < t->size);
631 void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k)
633 struct bset_tree *t = &b->sets[b->nsets];
634 unsigned shift = bkey_u64s(k);
635 unsigned j = bkey_to_cacheline(t, k);
637 /* We're getting called from btree_split() or btree_gc, just bail out */
641 /* k is the key we just inserted; we need to find the entry in the
642 * lookup table for the first key that is strictly greater than k:
643 * it's either k's cacheline or the next one
646 table_to_bkey(t, j) <= k)
649 /* Adjust all the lookup table entries, and find a new key for any that
650 * have gotten too big
652 for (; j < t->size; j++) {
655 if (t->prev[j] > 7) {
656 k = table_to_bkey(t, j - 1);
658 while (k < cacheline_to_bkey(t, j, 0))
661 t->prev[j] = bkey_to_cacheline_offset(k);
665 if (t->size == b->sets->tree + bset_tree_space(b) - t->tree)
668 /* Possibly add a new entry to the end of the lookup table */
670 for (k = table_to_bkey(t, t->size - 1);
673 if (t->size == bkey_to_cacheline(t, k)) {
674 t->prev[t->size] = bkey_to_cacheline_offset(k);
679 void bch_bset_init_next(struct btree *b)
681 struct bset *i = write_block(b);
683 if (i != b->sets[0].data) {
684 b->sets[++b->nsets].data = i;
685 i->seq = b->sets[0].data->seq;
687 get_random_bytes(&i->seq, sizeof(uint64_t));
689 i->magic = bset_magic(b->c);
693 bset_build_unwritten_tree(b);
696 struct bset_search_iter {
700 static struct bset_search_iter bset_search_write_set(struct btree *b,
702 const struct bkey *search)
704 unsigned li = 0, ri = t->size;
707 t->size < bkey_to_cacheline(t, end(t->data)));
709 while (li + 1 != ri) {
710 unsigned m = (li + ri) >> 1;
712 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
718 return (struct bset_search_iter) {
719 table_to_bkey(t, li),
720 ri < t->size ? table_to_bkey(t, ri) : end(t->data)
724 static struct bset_search_iter bset_search_tree(struct btree *b,
726 const struct bkey *search)
729 struct bkey_float *f;
730 unsigned inorder, j, n = 1;
734 p &= ((int) (p - t->size)) >> 31;
736 prefetch(&t->tree[p]);
742 * n = (f->mantissa > bfloat_mantissa())
746 * We need to subtract 1 from f->mantissa for the sign bit trick
747 * to work - that's done in make_bfloat()
749 if (likely(f->exponent != 127))
750 n = j * 2 + (((unsigned)
752 bfloat_mantissa(search, f))) >> 31);
754 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
757 } while (n < t->size);
759 inorder = to_inorder(j, t);
762 * n would have been the node we recursed to - the low bit tells us if
763 * we recursed left or recursed right.
766 l = cacheline_to_bkey(t, inorder, f->m);
768 if (++inorder != t->size) {
769 f = &t->tree[inorder_next(j, t->size)];
770 r = cacheline_to_bkey(t, inorder, f->m);
774 r = cacheline_to_bkey(t, inorder, f->m);
777 f = &t->tree[inorder_prev(j, t->size)];
778 l = cacheline_to_bkey(t, inorder, f->m);
783 return (struct bset_search_iter) {l, r};
786 struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t,
787 const struct bkey *search)
789 struct bset_search_iter i;
792 * First, we search for a cacheline, then lastly we do a linear search
793 * within that cacheline.
795 * To search for the cacheline, there's three different possibilities:
796 * * The set is too small to have a search tree, so we just do a linear
797 * search over the whole set.
798 * * The set is the one we're currently inserting into; keeping a full
799 * auxiliary search tree up to date would be too expensive, so we
800 * use a much simpler lookup table to do a binary search -
801 * bset_search_write_set().
802 * * Or we use the auxiliary search tree we constructed earlier -
806 if (unlikely(!t->size)) {
807 i.l = t->data->start;
809 } else if (bset_written(b, t)) {
811 * Each node in the auxiliary search tree covers a certain range
812 * of bits, and keys above and below the set it covers might
813 * differ outside those bits - so we have to special case the
814 * start and end - handle that here:
817 if (unlikely(bkey_cmp(search, &t->end) >= 0))
820 if (unlikely(bkey_cmp(search, t->data->start) < 0))
821 return t->data->start;
823 i = bset_search_tree(b, t, search);
825 i = bset_search_write_set(b, t, search);
827 #ifdef CONFIG_BCACHE_EDEBUG
828 BUG_ON(bset_written(b, t) &&
829 i.l != t->data->start &&
830 bkey_cmp(tree_to_prev_bkey(t,
831 inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
834 BUG_ON(i.r != end(t->data) &&
835 bkey_cmp(i.r, search) <= 0);
838 while (likely(i.l != i.r) &&
839 bkey_cmp(i.l, search) <= 0)
840 i.l = bkey_next(i.l);
847 static inline bool btree_iter_cmp(struct btree_iter_set l,
848 struct btree_iter_set r)
850 int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k));
852 return c ? c > 0 : l.k < r.k;
855 static inline bool btree_iter_end(struct btree_iter *iter)
860 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
864 BUG_ON(!heap_add(iter,
865 ((struct btree_iter_set) { k, end }),
869 struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter,
870 struct bkey *search, struct bset_tree *start)
872 struct bkey *ret = NULL;
873 iter->size = ARRAY_SIZE(iter->data);
876 for (; start <= &b->sets[b->nsets]; start++) {
877 ret = bch_bset_search(b, start, search);
878 bch_btree_iter_push(iter, ret, end(start->data));
884 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
886 struct btree_iter_set unused;
887 struct bkey *ret = NULL;
889 if (!btree_iter_end(iter)) {
891 iter->data->k = bkey_next(iter->data->k);
893 if (iter->data->k > iter->data->end) {
894 WARN_ONCE(1, "bset was corrupt!\n");
895 iter->data->k = iter->data->end;
898 if (iter->data->k == iter->data->end)
899 heap_pop(iter, unused, btree_iter_cmp);
901 heap_sift(iter, 0, btree_iter_cmp);
907 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
908 struct btree *b, ptr_filter_fn fn)
913 ret = bch_btree_iter_next(iter);
914 } while (ret && fn(b, ret));
919 struct bkey *bch_next_recurse_key(struct btree *b, struct bkey *search)
921 struct btree_iter iter;
923 bch_btree_iter_init(b, &iter, search);
924 return bch_btree_iter_next_filter(&iter, b, bch_ptr_bad);
929 static void btree_sort_fixup(struct btree_iter *iter)
931 while (iter->used > 1) {
932 struct btree_iter_set *top = iter->data, *i = top + 1;
935 if (iter->used > 2 &&
936 btree_iter_cmp(i[0], i[1]))
940 k != i->end && bkey_cmp(top->k, &START_KEY(k)) > 0;
943 __bch_cut_front(top->k, k);
944 else if (KEY_SIZE(k))
945 bch_cut_back(&START_KEY(k), top->k);
947 if (top->k < i->k || k == i->k)
950 heap_sift(iter, i - top, btree_iter_cmp);
954 static void btree_mergesort(struct btree *b, struct bset *out,
955 struct btree_iter *iter,
956 bool fixup, bool remove_stale)
958 struct bkey *k, *last = NULL;
959 bool (*bad)(struct btree *, const struct bkey *) = remove_stale
963 while (!btree_iter_end(iter)) {
964 if (fixup && !b->level)
965 btree_sort_fixup(iter);
967 k = bch_btree_iter_next(iter);
974 } else if (b->level ||
975 !bch_bkey_try_merge(b, last, k)) {
976 last = bkey_next(last);
981 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
983 pr_debug("sorted %i keys", out->keys);
984 bch_check_key_order(b, out);
987 static void __btree_sort(struct btree *b, struct btree_iter *iter,
988 unsigned start, unsigned order, bool fixup)
991 bool remove_stale = !b->written;
992 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO,
995 mutex_lock(&b->c->sort_lock);
997 order = ilog2(bucket_pages(b->c));
1000 start_time = local_clock();
1002 btree_mergesort(b, out, iter, fixup, remove_stale);
1005 if (!fixup && !start && b->written)
1006 bch_btree_verify(b, out);
1008 if (!start && order == b->page_order) {
1010 * Our temporary buffer is the same size as the btree node's
1011 * buffer, we can just swap buffers instead of doing a big
1015 out->magic = bset_magic(b->c);
1016 out->seq = b->sets[0].data->seq;
1017 out->version = b->sets[0].data->version;
1018 swap(out, b->sets[0].data);
1020 if (b->c->sort == b->sets[0].data)
1023 b->sets[start].data->keys = out->keys;
1024 memcpy(b->sets[start].data->start, out->start,
1025 (void *) end(out) - (void *) out->start);
1028 if (out == b->c->sort)
1029 mutex_unlock(&b->c->sort_lock);
1031 free_pages((unsigned long) out, order);
1034 bset_build_written_tree(b);
1037 spin_lock(&b->c->sort_time_lock);
1038 bch_time_stats_update(&b->c->sort_time, start_time);
1039 spin_unlock(&b->c->sort_time_lock);
1043 void bch_btree_sort_partial(struct btree *b, unsigned start)
1045 size_t oldsize = 0, order = b->page_order, keys = 0;
1046 struct btree_iter iter;
1047 __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]);
1049 BUG_ON(b->sets[b->nsets].data == write_block(b) &&
1050 (b->sets[b->nsets].size || b->nsets));
1053 oldsize = bch_count_data(b);
1058 for (i = start; i <= b->nsets; i++)
1059 keys += b->sets[i].data->keys;
1061 order = roundup_pow_of_two(__set_bytes(b->sets->data,
1064 order = ilog2(order);
1067 __btree_sort(b, &iter, start, order, false);
1069 EBUG_ON(b->written && bch_count_data(b) != oldsize);
1072 void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter)
1074 BUG_ON(!b->written);
1075 __btree_sort(b, iter, 0, b->page_order, true);
1078 void bch_btree_sort_into(struct btree *b, struct btree *new)
1080 uint64_t start_time = local_clock();
1082 struct btree_iter iter;
1083 bch_btree_iter_init(b, &iter, NULL);
1085 btree_mergesort(b, new->sets->data, &iter, false, true);
1087 spin_lock(&b->c->sort_time_lock);
1088 bch_time_stats_update(&b->c->sort_time, start_time);
1089 spin_unlock(&b->c->sort_time_lock);
1091 bkey_copy_key(&new->key, &b->key);
1092 new->sets->size = 0;
1095 #define SORT_CRIT (4096 / sizeof(uint64_t))
1097 void bch_btree_sort_lazy(struct btree *b)
1099 unsigned crit = SORT_CRIT;
1102 /* Don't sort if nothing to do */
1106 /* If not a leaf node, always sort */
1112 for (i = b->nsets - 1; i >= 0; --i) {
1113 crit *= b->c->sort_crit_factor;
1115 if (b->sets[i].data->keys < crit) {
1116 bch_btree_sort_partial(b, i);
1121 /* Sort if we'd overflow */
1122 if (b->nsets + 1 == MAX_BSETS) {
1128 bset_build_written_tree(b);
1135 size_t sets_written, sets_unwritten;
1136 size_t bytes_written, bytes_unwritten;
1137 size_t floats, failed;
1140 static int bch_btree_bset_stats(struct btree *b, struct btree_op *op,
1141 struct bset_stats *stats)
1148 for (i = 0; i <= b->nsets; i++) {
1149 struct bset_tree *t = &b->sets[i];
1150 size_t bytes = t->data->keys * sizeof(uint64_t);
1153 if (bset_written(b, t)) {
1154 stats->sets_written++;
1155 stats->bytes_written += bytes;
1157 stats->floats += t->size - 1;
1159 for (j = 1; j < t->size; j++)
1160 if (t->tree[j].exponent == 127)
1163 stats->sets_unwritten++;
1164 stats->bytes_unwritten += bytes;
1169 struct btree_iter iter;
1171 for_each_key_filter(b, k, &iter, bch_ptr_bad) {
1172 int ret = btree(bset_stats, k, b, op, stats);
1181 int bch_bset_print_stats(struct cache_set *c, char *buf)
1184 struct bset_stats t;
1187 bch_btree_op_init_stack(&op);
1188 memset(&t, 0, sizeof(struct bset_stats));
1190 ret = btree_root(bset_stats, c, &op, &t);
1194 return snprintf(buf, PAGE_SIZE,
1195 "btree nodes: %zu\n"
1196 "written sets: %zu\n"
1197 "unwritten sets: %zu\n"
1198 "written key bytes: %zu\n"
1199 "unwritten key bytes: %zu\n"
1203 t.sets_written, t.sets_unwritten,
1204 t.bytes_written, t.bytes_unwritten,
1205 t.floats, t.failed);