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 int __bch_keylist_realloc(struct keylist *l, unsigned u64s)
19 size_t oldsize = bch_keylist_nkeys(l);
20 size_t newsize = oldsize + u64s;
21 uint64_t *old_keys = l->keys_p == l->inline_keys ? NULL : l->keys_p;
24 newsize = roundup_pow_of_two(newsize);
26 if (newsize <= KEYLIST_INLINE ||
27 roundup_pow_of_two(oldsize) == newsize)
30 new_keys = krealloc(old_keys, sizeof(uint64_t) * newsize, GFP_NOIO);
36 memcpy(new_keys, l->inline_keys, sizeof(uint64_t) * oldsize);
39 l->top_p = new_keys + oldsize;
44 struct bkey *bch_keylist_pop(struct keylist *l)
46 struct bkey *k = l->keys;
51 while (bkey_next(k) != l->top)
57 void bch_keylist_pop_front(struct keylist *l)
59 l->top_p -= bkey_u64s(l->keys);
63 bch_keylist_bytes(l));
66 /* Pointer validation */
68 static bool __ptr_invalid(struct cache_set *c, const struct bkey *k)
72 for (i = 0; i < KEY_PTRS(k); i++)
73 if (ptr_available(c, k, i)) {
74 struct cache *ca = PTR_CACHE(c, k, i);
75 size_t bucket = PTR_BUCKET_NR(c, k, i);
76 size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
78 if (KEY_SIZE(k) + r > c->sb.bucket_size ||
79 bucket < ca->sb.first_bucket ||
80 bucket >= ca->sb.nbuckets)
87 bool bch_btree_ptr_invalid(struct cache_set *c, const struct bkey *k)
91 if (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))
94 if (__ptr_invalid(c, k))
99 bch_bkey_to_text(buf, sizeof(buf), k);
100 cache_bug(c, "spotted btree ptr %s: %s", buf, bch_ptr_status(c, k));
104 bool bch_extent_ptr_invalid(struct cache_set *c, const struct bkey *k)
111 if (KEY_SIZE(k) > KEY_OFFSET(k))
114 if (__ptr_invalid(c, k))
119 bch_bkey_to_text(buf, sizeof(buf), k);
120 cache_bug(c, "spotted extent %s: %s", buf, bch_ptr_status(c, k));
124 static bool ptr_bad_expensive_checks(struct btree *b, const struct bkey *k,
127 struct bucket *g = PTR_BUCKET(b->c, k, ptr);
130 if (mutex_trylock(&b->c->bucket_lock)) {
133 g->prio != BTREE_PRIO ||
134 (b->c->gc_mark_valid &&
135 GC_MARK(g) != GC_MARK_METADATA))
139 if (g->prio == BTREE_PRIO)
143 b->c->gc_mark_valid &&
144 GC_MARK(g) != GC_MARK_DIRTY)
147 mutex_unlock(&b->c->bucket_lock);
152 mutex_unlock(&b->c->bucket_lock);
153 bch_bkey_to_text(buf, sizeof(buf), k);
155 "inconsistent pointer %s: bucket %zu pin %i prio %i gen %i last_gc %i mark %llu gc_gen %i",
156 buf, PTR_BUCKET_NR(b->c, k, ptr), atomic_read(&g->pin),
157 g->prio, g->gen, g->last_gc, GC_MARK(g), g->gc_gen);
161 bool bch_ptr_bad(struct btree *b, const struct bkey *k)
166 if (!bkey_cmp(k, &ZERO_KEY) ||
168 bch_ptr_invalid(b, k))
171 for (i = 0; i < KEY_PTRS(k); i++)
172 if (!ptr_available(b->c, k, i))
175 if (!expensive_debug_checks(b->c) && KEY_DIRTY(k))
178 for (i = 0; i < KEY_PTRS(k); i++) {
179 g = PTR_BUCKET(b->c, k, i);
180 stale = ptr_stale(b->c, k, i);
182 btree_bug_on(stale > 96, b,
183 "key too stale: %i, need_gc %u",
184 stale, b->c->need_gc);
186 btree_bug_on(stale && KEY_DIRTY(k) && KEY_SIZE(k),
187 b, "stale dirty pointer");
192 if (expensive_debug_checks(b->c) &&
193 ptr_bad_expensive_checks(b, k, i))
200 /* Key/pointer manipulation */
202 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
205 BUG_ON(i > KEY_PTRS(src));
207 /* Only copy the header, key, and one pointer. */
208 memcpy(dest, src, 2 * sizeof(uint64_t));
209 dest->ptr[0] = src->ptr[i];
210 SET_KEY_PTRS(dest, 1);
211 /* We didn't copy the checksum so clear that bit. */
212 SET_KEY_CSUM(dest, 0);
215 bool __bch_cut_front(const struct bkey *where, struct bkey *k)
219 if (bkey_cmp(where, &START_KEY(k)) <= 0)
222 if (bkey_cmp(where, k) < 0)
223 len = KEY_OFFSET(k) - KEY_OFFSET(where);
225 bkey_copy_key(k, where);
227 for (i = 0; i < KEY_PTRS(k); i++)
228 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + KEY_SIZE(k) - len);
230 BUG_ON(len > KEY_SIZE(k));
231 SET_KEY_SIZE(k, len);
235 bool __bch_cut_back(const struct bkey *where, struct bkey *k)
239 if (bkey_cmp(where, k) >= 0)
242 BUG_ON(KEY_INODE(where) != KEY_INODE(k));
244 if (bkey_cmp(where, &START_KEY(k)) > 0)
245 len = KEY_OFFSET(where) - KEY_START(k);
247 bkey_copy_key(k, where);
249 BUG_ON(len > KEY_SIZE(k));
250 SET_KEY_SIZE(k, len);
254 static uint64_t merge_chksums(struct bkey *l, struct bkey *r)
256 return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) &
257 ~((uint64_t)1 << 63);
260 /* Tries to merge l and r: l should be lower than r
261 * Returns true if we were able to merge. If we did merge, l will be the merged
262 * key, r will be untouched.
264 bool bch_bkey_try_merge(struct btree *b, struct bkey *l, struct bkey *r)
268 if (key_merging_disabled(b->c))
271 if (KEY_PTRS(l) != KEY_PTRS(r) ||
272 KEY_DIRTY(l) != KEY_DIRTY(r) ||
273 bkey_cmp(l, &START_KEY(r)))
276 for (i = 0; i < KEY_PTRS(l); i++)
277 if (l->ptr[i] + PTR(0, KEY_SIZE(l), 0) != r->ptr[i] ||
278 PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i))
281 /* Keys with no pointers aren't restricted to one bucket and could
284 if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) {
285 SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l));
286 SET_KEY_SIZE(l, USHRT_MAX);
294 l->ptr[KEY_PTRS(l)] = merge_chksums(l, r);
299 SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r));
300 SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r));
305 /* Binary tree stuff for auxiliary search trees */
307 static unsigned inorder_next(unsigned j, unsigned size)
309 if (j * 2 + 1 < size) {
320 static unsigned inorder_prev(unsigned j, unsigned size)
325 while (j * 2 + 1 < size)
333 /* I have no idea why this code works... and I'm the one who wrote it
335 * However, I do know what it does:
336 * Given a binary tree constructed in an array (i.e. how you normally implement
337 * a heap), it converts a node in the tree - referenced by array index - to the
338 * index it would have if you did an inorder traversal.
340 * Also tested for every j, size up to size somewhere around 6 million.
342 * The binary tree starts at array index 1, not 0
343 * extra is a function of size:
344 * extra = (size - rounddown_pow_of_two(size - 1)) << 1;
346 static unsigned __to_inorder(unsigned j, unsigned size, unsigned extra)
349 unsigned shift = fls(size - 1) - b;
357 j -= (j - extra) >> 1;
362 static unsigned to_inorder(unsigned j, struct bset_tree *t)
364 return __to_inorder(j, t->size, t->extra);
367 static unsigned __inorder_to_tree(unsigned j, unsigned size, unsigned extra)
377 j |= roundup_pow_of_two(size) >> shift;
382 static unsigned inorder_to_tree(unsigned j, struct bset_tree *t)
384 return __inorder_to_tree(j, t->size, t->extra);
388 void inorder_test(void)
390 unsigned long done = 0;
391 ktime_t start = ktime_get();
393 for (unsigned size = 2;
396 unsigned extra = (size - rounddown_pow_of_two(size - 1)) << 1;
397 unsigned i = 1, j = rounddown_pow_of_two(size - 1);
400 printk(KERN_NOTICE "loop %u, %llu per us\n", size,
401 done / ktime_us_delta(ktime_get(), start));
404 if (__inorder_to_tree(i, size, extra) != j)
405 panic("size %10u j %10u i %10u", size, j, i);
407 if (__to_inorder(j, size, extra) != i)
408 panic("size %10u j %10u i %10u", size, j, i);
410 if (j == rounddown_pow_of_two(size) - 1)
413 BUG_ON(inorder_prev(inorder_next(j, size), size) != j);
415 j = inorder_next(j, size);
425 * Cacheline/offset <-> bkey pointer arithmetic:
427 * t->tree is a binary search tree in an array; each node corresponds to a key
428 * in one cacheline in t->set (BSET_CACHELINE bytes).
430 * This means we don't have to store the full index of the key that a node in
431 * the binary tree points to; to_inorder() gives us the cacheline, and then
432 * bkey_float->m gives us the offset within that cacheline, in units of 8 bytes.
434 * cacheline_to_bkey() and friends abstract out all the pointer arithmetic to
437 * To construct the bfloat for an arbitrary key we need to know what the key
438 * immediately preceding it is: we have to check if the two keys differ in the
439 * bits we're going to store in bkey_float->mantissa. t->prev[j] stores the size
440 * of the previous key so we can walk backwards to it from t->tree[j]'s key.
443 static struct bkey *cacheline_to_bkey(struct bset_tree *t, unsigned cacheline,
446 return ((void *) t->data) + cacheline * BSET_CACHELINE + offset * 8;
449 static unsigned bkey_to_cacheline(struct bset_tree *t, struct bkey *k)
451 return ((void *) k - (void *) t->data) / BSET_CACHELINE;
454 static unsigned bkey_to_cacheline_offset(struct bkey *k)
456 return ((size_t) k & (BSET_CACHELINE - 1)) / sizeof(uint64_t);
459 static struct bkey *tree_to_bkey(struct bset_tree *t, unsigned j)
461 return cacheline_to_bkey(t, to_inorder(j, t), t->tree[j].m);
464 static struct bkey *tree_to_prev_bkey(struct bset_tree *t, unsigned j)
466 return (void *) (((uint64_t *) tree_to_bkey(t, j)) - t->prev[j]);
470 * For the write set - the one we're currently inserting keys into - we don't
471 * maintain a full search tree, we just keep a simple lookup table in t->prev.
473 static struct bkey *table_to_bkey(struct bset_tree *t, unsigned cacheline)
475 return cacheline_to_bkey(t, cacheline, t->prev[cacheline]);
478 static inline uint64_t shrd128(uint64_t high, uint64_t low, uint8_t shift)
481 low |= (high << 1) << (63U - shift);
485 static inline unsigned bfloat_mantissa(const struct bkey *k,
486 struct bkey_float *f)
488 const uint64_t *p = &k->low - (f->exponent >> 6);
489 return shrd128(p[-1], p[0], f->exponent & 63) & BKEY_MANTISSA_MASK;
492 static void make_bfloat(struct bset_tree *t, unsigned j)
494 struct bkey_float *f = &t->tree[j];
495 struct bkey *m = tree_to_bkey(t, j);
496 struct bkey *p = tree_to_prev_bkey(t, j);
498 struct bkey *l = is_power_of_2(j)
500 : tree_to_prev_bkey(t, j >> ffs(j));
502 struct bkey *r = is_power_of_2(j + 1)
503 ? bset_bkey_idx(t->data, t->data->keys - bkey_u64s(&t->end))
504 : tree_to_bkey(t, j >> (ffz(j) + 1));
506 BUG_ON(m < l || m > r);
507 BUG_ON(bkey_next(p) != m);
509 if (KEY_INODE(l) != KEY_INODE(r))
510 f->exponent = fls64(KEY_INODE(r) ^ KEY_INODE(l)) + 64;
512 f->exponent = fls64(r->low ^ l->low);
514 f->exponent = max_t(int, f->exponent - BKEY_MANTISSA_BITS, 0);
517 * Setting f->exponent = 127 flags this node as failed, and causes the
518 * lookup code to fall back to comparing against the original key.
521 if (bfloat_mantissa(m, f) != bfloat_mantissa(p, f))
522 f->mantissa = bfloat_mantissa(m, f) - 1;
527 static void bset_alloc_tree(struct btree *b, struct bset_tree *t)
530 unsigned j = roundup(t[-1].size,
531 64 / sizeof(struct bkey_float));
533 t->tree = t[-1].tree + j;
534 t->prev = t[-1].prev + j;
537 while (t < b->sets + MAX_BSETS)
541 static void bset_build_unwritten_tree(struct btree *b)
543 struct bset_tree *t = b->sets + b->nsets;
545 bset_alloc_tree(b, t);
547 if (t->tree != b->sets->tree + bset_tree_space(b)) {
548 t->prev[0] = bkey_to_cacheline_offset(t->data->start);
553 static void bset_build_written_tree(struct btree *b)
555 struct bset_tree *t = b->sets + b->nsets;
556 struct bkey *k = t->data->start;
557 unsigned j, cacheline = 1;
559 bset_alloc_tree(b, t);
561 t->size = min_t(unsigned,
562 bkey_to_cacheline(t, bset_bkey_last(t->data)),
563 b->sets->tree + bset_tree_space(b) - t->tree);
570 t->extra = (t->size - rounddown_pow_of_two(t->size - 1)) << 1;
572 /* First we figure out where the first key in each cacheline is */
573 for (j = inorder_next(0, t->size);
575 j = inorder_next(j, t->size)) {
576 while (bkey_to_cacheline(t, k) != cacheline)
579 t->prev[j] = bkey_u64s(k);
582 t->tree[j].m = bkey_to_cacheline_offset(k);
585 while (bkey_next(k) != bset_bkey_last(t->data))
590 /* Then we build the tree */
591 for (j = inorder_next(0, t->size);
593 j = inorder_next(j, t->size))
597 void bch_bset_fix_invalidated_key(struct btree *b, struct bkey *k)
600 unsigned inorder, j = 1;
602 for (t = b->sets; t <= &b->sets[b->nsets]; t++)
603 if (k < bset_bkey_last(t->data))
608 if (!t->size || !bset_written(b, t))
611 inorder = bkey_to_cacheline(t, k);
613 if (k == t->data->start)
616 if (bkey_next(k) == bset_bkey_last(t->data)) {
621 j = inorder_to_tree(inorder, t);
625 k == tree_to_bkey(t, j))
629 } while (j < t->size);
631 j = inorder_to_tree(inorder + 1, t);
635 k == tree_to_prev_bkey(t, j))
639 } while (j < t->size);
642 void bch_bset_fix_lookup_table(struct btree *b, struct bkey *k)
644 struct bset_tree *t = &b->sets[b->nsets];
645 unsigned shift = bkey_u64s(k);
646 unsigned j = bkey_to_cacheline(t, k);
648 /* We're getting called from btree_split() or btree_gc, just bail out */
652 /* k is the key we just inserted; we need to find the entry in the
653 * lookup table for the first key that is strictly greater than k:
654 * it's either k's cacheline or the next one
657 table_to_bkey(t, j) <= k)
660 /* Adjust all the lookup table entries, and find a new key for any that
661 * have gotten too big
663 for (; j < t->size; j++) {
666 if (t->prev[j] > 7) {
667 k = table_to_bkey(t, j - 1);
669 while (k < cacheline_to_bkey(t, j, 0))
672 t->prev[j] = bkey_to_cacheline_offset(k);
676 if (t->size == b->sets->tree + bset_tree_space(b) - t->tree)
679 /* Possibly add a new entry to the end of the lookup table */
681 for (k = table_to_bkey(t, t->size - 1);
682 k != bset_bkey_last(t->data);
684 if (t->size == bkey_to_cacheline(t, k)) {
685 t->prev[t->size] = bkey_to_cacheline_offset(k);
690 void bch_bset_init_next(struct btree *b)
692 struct bset *i = write_block(b);
694 if (i != b->sets[0].data) {
695 b->sets[++b->nsets].data = i;
696 i->seq = b->sets[0].data->seq;
698 get_random_bytes(&i->seq, sizeof(uint64_t));
700 i->magic = bset_magic(&b->c->sb);
704 bset_build_unwritten_tree(b);
707 struct bset_search_iter {
711 static struct bset_search_iter bset_search_write_set(struct btree *b,
713 const struct bkey *search)
715 unsigned li = 0, ri = t->size;
718 t->size < bkey_to_cacheline(t, bset_bkey_last(t->data)));
720 while (li + 1 != ri) {
721 unsigned m = (li + ri) >> 1;
723 if (bkey_cmp(table_to_bkey(t, m), search) > 0)
729 return (struct bset_search_iter) {
730 table_to_bkey(t, li),
731 ri < t->size ? table_to_bkey(t, ri) : bset_bkey_last(t->data)
735 static struct bset_search_iter bset_search_tree(struct btree *b,
737 const struct bkey *search)
740 struct bkey_float *f;
741 unsigned inorder, j, n = 1;
745 p &= ((int) (p - t->size)) >> 31;
747 prefetch(&t->tree[p]);
753 * n = (f->mantissa > bfloat_mantissa())
757 * We need to subtract 1 from f->mantissa for the sign bit trick
758 * to work - that's done in make_bfloat()
760 if (likely(f->exponent != 127))
761 n = j * 2 + (((unsigned)
763 bfloat_mantissa(search, f))) >> 31);
765 n = (bkey_cmp(tree_to_bkey(t, j), search) > 0)
768 } while (n < t->size);
770 inorder = to_inorder(j, t);
773 * n would have been the node we recursed to - the low bit tells us if
774 * we recursed left or recursed right.
777 l = cacheline_to_bkey(t, inorder, f->m);
779 if (++inorder != t->size) {
780 f = &t->tree[inorder_next(j, t->size)];
781 r = cacheline_to_bkey(t, inorder, f->m);
783 r = bset_bkey_last(t->data);
785 r = cacheline_to_bkey(t, inorder, f->m);
788 f = &t->tree[inorder_prev(j, t->size)];
789 l = cacheline_to_bkey(t, inorder, f->m);
794 return (struct bset_search_iter) {l, r};
797 struct bkey *__bch_bset_search(struct btree *b, struct bset_tree *t,
798 const struct bkey *search)
800 struct bset_search_iter i;
803 * First, we search for a cacheline, then lastly we do a linear search
804 * within that cacheline.
806 * To search for the cacheline, there's three different possibilities:
807 * * The set is too small to have a search tree, so we just do a linear
808 * search over the whole set.
809 * * The set is the one we're currently inserting into; keeping a full
810 * auxiliary search tree up to date would be too expensive, so we
811 * use a much simpler lookup table to do a binary search -
812 * bset_search_write_set().
813 * * Or we use the auxiliary search tree we constructed earlier -
817 if (unlikely(!t->size)) {
818 i.l = t->data->start;
819 i.r = bset_bkey_last(t->data);
820 } else if (bset_written(b, t)) {
822 * Each node in the auxiliary search tree covers a certain range
823 * of bits, and keys above and below the set it covers might
824 * differ outside those bits - so we have to special case the
825 * start and end - handle that here:
828 if (unlikely(bkey_cmp(search, &t->end) >= 0))
829 return bset_bkey_last(t->data);
831 if (unlikely(bkey_cmp(search, t->data->start) < 0))
832 return t->data->start;
834 i = bset_search_tree(b, t, search);
836 i = bset_search_write_set(b, t, search);
838 if (expensive_debug_checks(b->c)) {
839 BUG_ON(bset_written(b, t) &&
840 i.l != t->data->start &&
841 bkey_cmp(tree_to_prev_bkey(t,
842 inorder_to_tree(bkey_to_cacheline(t, i.l), t)),
845 BUG_ON(i.r != bset_bkey_last(t->data) &&
846 bkey_cmp(i.r, search) <= 0);
849 while (likely(i.l != i.r) &&
850 bkey_cmp(i.l, search) <= 0)
851 i.l = bkey_next(i.l);
859 * Returns true if l > r - unless l == r, in which case returns true if l is
862 * Necessary for btree_sort_fixup() - if there are multiple keys that compare
863 * equal in different sets, we have to process them newest to oldest.
865 static inline bool btree_iter_cmp(struct btree_iter_set l,
866 struct btree_iter_set r)
868 int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k));
870 return c ? c > 0 : l.k < r.k;
873 static inline bool btree_iter_end(struct btree_iter *iter)
878 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
882 BUG_ON(!heap_add(iter,
883 ((struct btree_iter_set) { k, end }),
887 struct bkey *__bch_btree_iter_init(struct btree *b, struct btree_iter *iter,
888 struct bkey *search, struct bset_tree *start)
890 struct bkey *ret = NULL;
891 iter->size = ARRAY_SIZE(iter->data);
894 #ifdef CONFIG_BCACHE_DEBUG
898 for (; start <= &b->sets[b->nsets]; start++) {
899 ret = bch_bset_search(b, start, search);
900 bch_btree_iter_push(iter, ret, bset_bkey_last(start->data));
906 struct bkey *bch_btree_iter_next(struct btree_iter *iter)
908 struct btree_iter_set unused;
909 struct bkey *ret = NULL;
911 if (!btree_iter_end(iter)) {
912 bch_btree_iter_next_check(iter);
915 iter->data->k = bkey_next(iter->data->k);
917 if (iter->data->k > iter->data->end) {
918 WARN_ONCE(1, "bset was corrupt!\n");
919 iter->data->k = iter->data->end;
922 if (iter->data->k == iter->data->end)
923 heap_pop(iter, unused, btree_iter_cmp);
925 heap_sift(iter, 0, btree_iter_cmp);
931 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
932 struct btree *b, ptr_filter_fn fn)
937 ret = bch_btree_iter_next(iter);
938 } while (ret && fn(b, ret));
945 static void sort_key_next(struct btree_iter *iter,
946 struct btree_iter_set *i)
948 i->k = bkey_next(i->k);
951 *i = iter->data[--iter->used];
954 static struct bkey *btree_sort_fixup(struct btree_iter *iter, struct bkey *tmp)
956 while (iter->used > 1) {
957 struct btree_iter_set *top = iter->data, *i = top + 1;
959 if (iter->used > 2 &&
960 btree_iter_cmp(i[0], i[1]))
963 if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0)
966 if (!KEY_SIZE(i->k)) {
967 sort_key_next(iter, i);
968 heap_sift(iter, i - top, btree_iter_cmp);
973 if (bkey_cmp(top->k, i->k) >= 0)
974 sort_key_next(iter, i);
976 bch_cut_front(top->k, i->k);
978 heap_sift(iter, i - top, btree_iter_cmp);
980 /* can't happen because of comparison func */
981 BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k)));
983 if (bkey_cmp(i->k, top->k) < 0) {
984 bkey_copy(tmp, top->k);
986 bch_cut_back(&START_KEY(i->k), tmp);
987 bch_cut_front(i->k, top->k);
988 heap_sift(iter, 0, btree_iter_cmp);
992 bch_cut_back(&START_KEY(i->k), top->k);
1000 static void btree_mergesort(struct btree *b, struct bset *out,
1001 struct btree_iter *iter,
1002 bool fixup, bool remove_stale)
1004 struct bkey *k, *last = NULL;
1006 bool (*bad)(struct btree *, const struct bkey *) = remove_stale
1010 while (!btree_iter_end(iter)) {
1011 if (fixup && !b->level)
1012 k = btree_sort_fixup(iter, &tmp.k);
1017 k = bch_btree_iter_next(iter);
1025 } else if (b->level ||
1026 !bch_bkey_try_merge(b, last, k)) {
1027 last = bkey_next(last);
1032 out->keys = last ? (uint64_t *) bkey_next(last) - out->d : 0;
1034 pr_debug("sorted %i keys", out->keys);
1037 static void __btree_sort(struct btree *b, struct btree_iter *iter,
1038 unsigned start, unsigned order, bool fixup)
1040 uint64_t start_time;
1041 bool remove_stale = !b->written;
1042 bool used_mempool = false;
1043 struct bset *out = (void *) __get_free_pages(__GFP_NOWARN|GFP_NOIO,
1046 out = page_address(mempool_alloc(b->c->sort_pool, GFP_NOIO));
1047 used_mempool = true;
1048 order = ilog2(bucket_pages(b->c));
1051 start_time = local_clock();
1053 btree_mergesort(b, out, iter, fixup, remove_stale);
1056 if (!start && order == b->page_order) {
1058 * Our temporary buffer is the same size as the btree node's
1059 * buffer, we can just swap buffers instead of doing a big
1063 out->magic = bset_magic(&b->c->sb);
1064 out->seq = b->sets[0].data->seq;
1065 out->version = b->sets[0].data->version;
1066 swap(out, b->sets[0].data);
1068 b->sets[start].data->keys = out->keys;
1069 memcpy(b->sets[start].data->start, out->start,
1070 (void *) bset_bkey_last(out) - (void *) out->start);
1074 mempool_free(virt_to_page(out), b->c->sort_pool);
1076 free_pages((unsigned long) out, order);
1079 bset_build_written_tree(b);
1082 bch_time_stats_update(&b->c->sort_time, start_time);
1085 void bch_btree_sort_partial(struct btree *b, unsigned start)
1087 size_t order = b->page_order, keys = 0;
1088 struct btree_iter iter;
1089 int oldsize = bch_count_data(b);
1091 __bch_btree_iter_init(b, &iter, NULL, &b->sets[start]);
1093 BUG_ON(b->sets[b->nsets].data == write_block(b) &&
1094 (b->sets[b->nsets].size || b->nsets));
1100 for (i = start; i <= b->nsets; i++)
1101 keys += b->sets[i].data->keys;
1103 order = roundup_pow_of_two(__set_bytes(b->sets->data,
1106 order = ilog2(order);
1109 __btree_sort(b, &iter, start, order, false);
1111 EBUG_ON(b->written && oldsize >= 0 && bch_count_data(b) != oldsize);
1114 void bch_btree_sort_and_fix_extents(struct btree *b, struct btree_iter *iter)
1116 BUG_ON(!b->written);
1117 __btree_sort(b, iter, 0, b->page_order, true);
1120 void bch_btree_sort_into(struct btree *b, struct btree *new)
1122 uint64_t start_time = local_clock();
1124 struct btree_iter iter;
1125 bch_btree_iter_init(b, &iter, NULL);
1127 btree_mergesort(b, new->sets->data, &iter, false, true);
1129 bch_time_stats_update(&b->c->sort_time, start_time);
1131 bkey_copy_key(&new->key, &b->key);
1132 new->sets->size = 0;
1135 #define SORT_CRIT (4096 / sizeof(uint64_t))
1137 void bch_btree_sort_lazy(struct btree *b)
1139 unsigned crit = SORT_CRIT;
1142 /* Don't sort if nothing to do */
1146 /* If not a leaf node, always sort */
1152 for (i = b->nsets - 1; i >= 0; --i) {
1153 crit *= b->c->sort_crit_factor;
1155 if (b->sets[i].data->keys < crit) {
1156 bch_btree_sort_partial(b, i);
1161 /* Sort if we'd overflow */
1162 if (b->nsets + 1 == MAX_BSETS) {
1168 bset_build_written_tree(b);
1176 size_t sets_written, sets_unwritten;
1177 size_t bytes_written, bytes_unwritten;
1178 size_t floats, failed;
1181 static int btree_bset_stats(struct btree_op *op, struct btree *b)
1183 struct bset_stats *stats = container_of(op, struct bset_stats, op);
1188 for (i = 0; i <= b->nsets; i++) {
1189 struct bset_tree *t = &b->sets[i];
1190 size_t bytes = t->data->keys * sizeof(uint64_t);
1193 if (bset_written(b, t)) {
1194 stats->sets_written++;
1195 stats->bytes_written += bytes;
1197 stats->floats += t->size - 1;
1199 for (j = 1; j < t->size; j++)
1200 if (t->tree[j].exponent == 127)
1203 stats->sets_unwritten++;
1204 stats->bytes_unwritten += bytes;
1208 return MAP_CONTINUE;
1211 int bch_bset_print_stats(struct cache_set *c, char *buf)
1213 struct bset_stats t;
1216 memset(&t, 0, sizeof(struct bset_stats));
1217 bch_btree_op_init(&t.op, -1);
1219 ret = bch_btree_map_nodes(&t.op, c, &ZERO_KEY, btree_bset_stats);
1223 return snprintf(buf, PAGE_SIZE,
1224 "btree nodes: %zu\n"
1225 "written sets: %zu\n"
1226 "unwritten sets: %zu\n"
1227 "written key bytes: %zu\n"
1228 "unwritten key bytes: %zu\n"
1232 t.sets_written, t.sets_unwritten,
1233 t.bytes_written, t.bytes_unwritten,
1234 t.floats, t.failed);