// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
-#include "bkey_on_stack.h"
+#include "bkey_buf.h"
#include "bkey_sort.h"
#include "bset.h"
#include "extents.h"
-/* too many iterators, need to clean this up */
-
-/* btree_node_iter_large: */
-
-#define btree_node_iter_cmp_heap(h, _l, _r) btree_node_iter_cmp(b, _l, _r)
+typedef int (*sort_cmp_fn)(struct btree *,
+ struct bkey_packed *,
+ struct bkey_packed *);
-static inline bool
-bch2_btree_node_iter_large_end(struct btree_node_iter_large *iter)
+static inline bool sort_iter_end(struct sort_iter *iter)
{
return !iter->used;
}
-static inline struct bkey_packed *
-bch2_btree_node_iter_large_peek_all(struct btree_node_iter_large *iter,
- struct btree *b)
-{
- return bch2_btree_node_iter_large_end(iter)
- ? NULL
- : __btree_node_offset_to_key(b, iter->data->k);
-}
-
-static void
-bch2_btree_node_iter_large_advance(struct btree_node_iter_large *iter,
- struct btree *b)
-{
- iter->data->k += __btree_node_offset_to_key(b, iter->data->k)->u64s;
-
- EBUG_ON(!iter->used);
- EBUG_ON(iter->data->k > iter->data->end);
-
- if (iter->data->k == iter->data->end)
- heap_del(iter, 0, btree_node_iter_cmp_heap, NULL);
- else
- heap_sift_down(iter, 0, btree_node_iter_cmp_heap, NULL);
-}
-
-static inline struct bkey_packed *
-bch2_btree_node_iter_large_next_all(struct btree_node_iter_large *iter,
- struct btree *b)
-{
- struct bkey_packed *ret = bch2_btree_node_iter_large_peek_all(iter, b);
-
- if (ret)
- bch2_btree_node_iter_large_advance(iter, b);
-
- return ret;
-}
-
-void bch2_btree_node_iter_large_push(struct btree_node_iter_large *iter,
- struct btree *b,
- const struct bkey_packed *k,
- const struct bkey_packed *end)
-{
- if (k != end) {
- struct btree_node_iter_set n =
- ((struct btree_node_iter_set) {
- __btree_node_key_to_offset(b, k),
- __btree_node_key_to_offset(b, end)
- });
-
- __heap_add(iter, n, btree_node_iter_cmp_heap, NULL);
- }
-}
-
-static void sort_key_next(struct btree_node_iter_large *iter,
- struct btree *b,
- struct btree_node_iter_set *i)
-{
- i->k += __btree_node_offset_to_key(b, i->k)->u64s;
-
- while (i->k != i->end &&
- !__btree_node_offset_to_key(b, i->k)->u64s)
- i->k++;
-
- if (i->k == i->end)
- *i = iter->data[--iter->used];
-}
-
-/* regular sort_iters */
-
-typedef int (*sort_cmp_fn)(struct btree *,
- struct bkey_packed *,
- struct bkey_packed *);
-
static inline void __sort_iter_sift(struct sort_iter *iter,
unsigned from,
sort_cmp_fn cmp)
static inline struct bkey_packed *sort_iter_peek(struct sort_iter *iter)
{
- return iter->used ? iter->data->k : NULL;
+ return !sort_iter_end(iter) ? iter->data->k : NULL;
}
-static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
+static inline void __sort_iter_advance(struct sort_iter *iter,
+ unsigned idx, sort_cmp_fn cmp)
{
- iter->data->k = bkey_next_skip_noops(iter->data->k, iter->data->end);
+ struct sort_iter_set *i = iter->data + idx;
+
+ BUG_ON(idx >= iter->used);
- BUG_ON(iter->data->k > iter->data->end);
+ i->k = bkey_next_skip_noops(i->k, i->end);
- if (iter->data->k == iter->data->end)
- array_remove_item(iter->data, iter->used, 0);
+ BUG_ON(i->k > i->end);
+
+ if (i->k == i->end)
+ array_remove_item(iter->data, iter->used, idx);
else
- sort_iter_sift(iter, cmp);
+ __sort_iter_sift(iter, idx, cmp);
+}
+
+static inline void sort_iter_advance(struct sort_iter *iter, sort_cmp_fn cmp)
+{
+ __sort_iter_advance(iter, 0, cmp);
}
static inline struct bkey_packed *sort_iter_next(struct sort_iter *iter,
}
/*
- * Returns true if l > r - unless l == r, in which case returns true if l is
- * older than r.
- *
- * Necessary for btree_sort_fixup() - if there are multiple keys that compare
- * equal in different sets, we have to process them newest to oldest.
+ * If keys compare equal, compare by pointer order:
*/
-#define key_sort_cmp(h, l, r) \
-({ \
- bkey_cmp_packed(b, \
- __btree_node_offset_to_key(b, (l).k), \
- __btree_node_offset_to_key(b, (r).k)) \
- \
- ?: (l).k - (r).k; \
-})
-
-static inline bool should_drop_next_key(struct btree_node_iter_large *iter,
- struct btree *b)
+static inline int key_sort_fix_overlapping_cmp(struct btree *b,
+ struct bkey_packed *l,
+ struct bkey_packed *r)
{
- struct btree_node_iter_set *l = iter->data, *r = iter->data + 1;
- struct bkey_packed *k = __btree_node_offset_to_key(b, l->k);
-
- if (bkey_whiteout(k))
- return true;
-
- if (iter->used < 2)
- return false;
-
- if (iter->used > 2 &&
- key_sort_cmp(iter, r[0], r[1]) >= 0)
- r++;
+ return bch2_bkey_cmp_packed(b, l, r) ?:
+ cmp_int((unsigned long) l, (unsigned long) r);
+}
+static inline bool should_drop_next_key(struct sort_iter *iter)
+{
/*
* key_sort_cmp() ensures that when keys compare equal the older key
- * comes first; so if l->k compares equal to r->k then l->k is older and
- * should be dropped.
+ * comes first; so if l->k compares equal to r->k then l->k is older
+ * and should be dropped.
*/
- return !bkey_cmp_packed(b,
- __btree_node_offset_to_key(b, l->k),
- __btree_node_offset_to_key(b, r->k));
+ return iter->used >= 2 &&
+ !bch2_bkey_cmp_packed(iter->b,
+ iter->data[0].k,
+ iter->data[1].k);
}
-struct btree_nr_keys bch2_key_sort_fix_overlapping(struct bset *dst,
- struct btree *b,
- struct btree_node_iter_large *iter)
+struct btree_nr_keys
+bch2_key_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
+ struct sort_iter *iter)
{
struct bkey_packed *out = dst->start;
+ struct bkey_packed *k;
struct btree_nr_keys nr;
memset(&nr, 0, sizeof(nr));
- heap_resort(iter, key_sort_cmp, NULL);
-
- while (!bch2_btree_node_iter_large_end(iter)) {
- if (!should_drop_next_key(iter, b)) {
- struct bkey_packed *k =
- __btree_node_offset_to_key(b, iter->data->k);
+ sort_iter_sort(iter, key_sort_fix_overlapping_cmp);
+ while ((k = sort_iter_peek(iter))) {
+ if (!bkey_whiteout(k) &&
+ !should_drop_next_key(iter)) {
bkey_copy(out, k);
btree_keys_account_key_add(&nr, 0, out);
out = bkey_next(out);
}
- sort_key_next(iter, b, iter->data);
- heap_sift_down(iter, 0, key_sort_cmp, NULL);
+ sort_iter_advance(iter, key_sort_fix_overlapping_cmp);
}
dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
return nr;
}
-/*
- * If keys compare equal, compare by pointer order:
- *
- * Necessary for sort_fix_overlapping() - if there are multiple keys that
- * compare equal in different sets, we have to process them newest to oldest.
- */
-#define extent_sort_cmp(h, l, r) \
-({ \
- struct bkey _ul = bkey_unpack_key(b, \
- __btree_node_offset_to_key(b, (l).k)); \
- struct bkey _ur = bkey_unpack_key(b, \
- __btree_node_offset_to_key(b, (r).k)); \
- \
- bkey_cmp(bkey_start_pos(&_ul), \
- bkey_start_pos(&_ur)) ?: (r).k - (l).k; \
-})
-
-static inline void extent_sort_sift(struct btree_node_iter_large *iter,
- struct btree *b, size_t i)
-{
- heap_sift_down(iter, i, extent_sort_cmp, NULL);
-}
-
-static inline void extent_sort_next(struct btree_node_iter_large *iter,
- struct btree *b,
- struct btree_node_iter_set *i)
-{
- sort_key_next(iter, b, i);
- heap_sift_down(iter, i - iter->data, extent_sort_cmp, NULL);
-}
-
-static void extent_sort_advance_prev(struct bkey_format *f,
- struct btree_nr_keys *nr,
- struct bkey_packed *start,
- struct bkey_packed **prev)
-{
- if (*prev) {
- bch2_bkey_pack(*prev, (void *) *prev, f);
-
- btree_keys_account_key_add(nr, 0, *prev);
- *prev = bkey_next(*prev);
- } else {
- *prev = start;
- }
-}
-
static void extent_sort_append(struct bch_fs *c,
struct bkey_format *f,
struct btree_nr_keys *nr,
- struct bkey_packed *start,
- struct bkey_packed **prev,
+ struct bkey_packed **out,
struct bkey_s k)
{
- if (bkey_whiteout(k.k))
- return;
-
- /*
- * prev is always unpacked, for key merging - until right before we
- * advance it:
- */
-
- if (*prev &&
- bch2_bkey_merge(c, bkey_i_to_s((void *) *prev), k) ==
- BCH_MERGE_MERGE)
- return;
-
- extent_sort_advance_prev(f, nr, start, prev);
-
- bkey_reassemble((void *) *prev, k.s_c);
-}
-
-struct btree_nr_keys bch2_extent_sort_fix_overlapping(struct bch_fs *c,
- struct bset *dst,
- struct btree *b,
- struct btree_node_iter_large *iter)
-{
- struct bkey_format *f = &b->format;
- struct btree_node_iter_set *_l = iter->data, *_r;
- struct bkey_packed *prev = NULL, *lk, *rk;
- struct bkey l_unpacked, r_unpacked;
- struct bkey_s l, r;
- struct btree_nr_keys nr;
- struct bkey_on_stack split;
-
- memset(&nr, 0, sizeof(nr));
- bkey_on_stack_init(&split);
-
- heap_resort(iter, extent_sort_cmp, NULL);
-
- while (!bch2_btree_node_iter_large_end(iter)) {
- lk = __btree_node_offset_to_key(b, _l->k);
- l = __bkey_disassemble(b, lk, &l_unpacked);
-
- if (iter->used == 1) {
- extent_sort_append(c, f, &nr, dst->start, &prev, l);
- extent_sort_next(iter, b, _l);
- continue;
- }
-
- _r = iter->data + 1;
- if (iter->used > 2 &&
- extent_sort_cmp(iter, _r[0], _r[1]) >= 0)
- _r++;
-
- rk = __btree_node_offset_to_key(b, _r->k);
- r = __bkey_disassemble(b, rk, &r_unpacked);
-
- /* If current key and next key don't overlap, just append */
- if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
- extent_sort_append(c, f, &nr, dst->start, &prev, l);
- extent_sort_next(iter, b, _l);
- continue;
- }
-
- /* Skip 0 size keys */
- if (!r.k->size) {
- extent_sort_next(iter, b, _r);
- continue;
- }
+ if (!bkey_whiteout(k.k)) {
+ if (!bch2_bkey_pack_key(*out, k.k, f))
+ memcpy_u64s_small(*out, k.k, BKEY_U64s);
- /*
- * overlap: keep the newer key and trim the older key so they
- * don't overlap. comparing pointers tells us which one is
- * newer, since the bsets are appended one after the other.
- */
+ memcpy_u64s_small(bkeyp_val(f, *out), k.v, bkey_val_u64s(k.k));
- /* can't happen because of comparison func */
- BUG_ON(_l->k < _r->k &&
- !bkey_cmp(bkey_start_pos(l.k), bkey_start_pos(r.k)));
-
- if (_l->k > _r->k) {
- /* l wins, trim r */
- if (bkey_cmp(l.k->p, r.k->p) >= 0) {
- sort_key_next(iter, b, _r);
- } else {
- bch2_cut_front_s(l.k->p, r);
- extent_save(b, rk, r.k);
- }
-
- extent_sort_sift(iter, b, _r - iter->data);
- } else if (bkey_cmp(l.k->p, r.k->p) > 0) {
- bkey_on_stack_realloc(&split, c, l.k->u64s);
-
- /*
- * r wins, but it overlaps in the middle of l - split l:
- */
- bkey_reassemble(split.k, l.s_c);
- bch2_cut_back(bkey_start_pos(r.k), split.k);
-
- bch2_cut_front_s(r.k->p, l);
- extent_save(b, lk, l.k);
-
- extent_sort_sift(iter, b, 0);
-
- extent_sort_append(c, f, &nr, dst->start,
- &prev, bkey_i_to_s(split.k));
- } else {
- bch2_cut_back_s(bkey_start_pos(r.k), l);
- extent_save(b, lk, l.k);
- }
+ btree_keys_account_key_add(nr, 0, *out);
+ *out = bkey_next(*out);
}
-
- extent_sort_advance_prev(f, &nr, dst->start, &prev);
-
- dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
-
- bkey_on_stack_exit(&split, c);
- return nr;
}
/* Sort + repack in a new format: */
return nr;
}
-/* Sort, repack, and merge: */
+/* Sort, repack, and call bch2_bkey_normalize() to drop stale pointers: */
struct btree_nr_keys
bch2_sort_repack_merge(struct bch_fs *c,
struct bset *dst, struct btree *src,
struct bkey_format *out_f,
bool filter_whiteouts)
{
- struct bkey_packed *prev = NULL, *k_packed;
- struct bkey_s k;
+ struct bkey_packed *out = vstruct_last(dst), *k_packed;
+ struct bkey_buf k;
struct btree_nr_keys nr;
- struct bkey unpacked;
memset(&nr, 0, sizeof(nr));
+ bch2_bkey_buf_init(&k);
while ((k_packed = bch2_btree_node_iter_next_all(iter, src))) {
if (filter_whiteouts && bkey_whiteout(k_packed))
continue;
- k = __bkey_disassemble(src, k_packed, &unpacked);
+ /*
+ * NOTE:
+ * bch2_bkey_normalize may modify the key we pass it (dropping
+ * stale pointers) and we don't have a write lock on the src
+ * node; we have to make a copy of the entire key before calling
+ * normalize
+ */
+ bch2_bkey_buf_realloc(&k, c, k_packed->u64s + BKEY_U64s);
+ bch2_bkey_unpack(src, k.k, k_packed);
if (filter_whiteouts &&
- bch2_bkey_normalize(c, k))
+ bch2_bkey_normalize(c, bkey_i_to_s(k.k)))
continue;
- extent_sort_append(c, out_f, &nr, vstruct_last(dst), &prev, k);
+ extent_sort_append(c, out_f, &nr, &out, bkey_i_to_s(k.k));
}
- extent_sort_advance_prev(out_f, &nr, vstruct_last(dst), &prev);
-
- dst->u64s = cpu_to_le16((u64 *) prev - dst->_data);
+ dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
+ bch2_bkey_buf_exit(&k, c);
return nr;
}
struct bkey_packed *l,
struct bkey_packed *r)
{
- return bkey_cmp_packed(b, l, r) ?:
- (int) bkey_whiteout(r) - (int) bkey_whiteout(l) ?:
+ return bch2_bkey_cmp_packed(b, l, r) ?:
+ (int) bkey_deleted(r) - (int) bkey_deleted(l) ?:
(int) l->needs_whiteout - (int) r->needs_whiteout;
}
sort_iter_sort(iter, sort_keys_cmp);
while ((in = sort_iter_next(iter, sort_keys_cmp))) {
+ bool needs_whiteout = false;
+
if (bkey_whiteout(in) &&
(filter_whiteouts || !in->needs_whiteout))
continue;
- if (bkey_whiteout(in) &&
- (next = sort_iter_peek(iter)) &&
- !bkey_cmp_packed(iter->b, in, next)) {
+ while ((next = sort_iter_peek(iter)) &&
+ !bch2_bkey_cmp_packed(iter->b, in, next)) {
BUG_ON(in->needs_whiteout &&
next->needs_whiteout);
- /*
- * XXX racy, called with read lock from write path
- *
- * leads to spurious BUG_ON() in bkey_unpack_key() in
- * debug mode
- */
- next->needs_whiteout |= in->needs_whiteout;
- continue;
+ needs_whiteout |= in->needs_whiteout;
+ in = sort_iter_next(iter, sort_keys_cmp);
}
if (bkey_whiteout(in)) {
} else {
bkey_copy(out, in);
}
+ out->needs_whiteout |= needs_whiteout;
out = bkey_next(out);
}
return (u64 *) out - (u64 *) dst;
}
+/* Compat code for btree_node_old_extent_overwrite: */
+
+/*
+ * If keys compare equal, compare by pointer order:
+ *
+ * Necessary for sort_fix_overlapping() - if there are multiple keys that
+ * compare equal in different sets, we have to process them newest to oldest.
+ */
+static inline int extent_sort_fix_overlapping_cmp(struct btree *b,
+ struct bkey_packed *l,
+ struct bkey_packed *r)
+{
+ struct bkey ul = bkey_unpack_key(b, l);
+ struct bkey ur = bkey_unpack_key(b, r);
+
+ return bkey_cmp(bkey_start_pos(&ul),
+ bkey_start_pos(&ur)) ?:
+ cmp_int((unsigned long) r, (unsigned long) l);
+}
+
+/*
+ * The algorithm in extent_sort_fix_overlapping() relies on keys in the same
+ * bset being ordered by start offset - but 0 size whiteouts (which are always
+ * KEY_TYPE_deleted) break this ordering, so we need to skip over them:
+ */
+static void extent_iter_advance(struct sort_iter *iter, unsigned idx)
+{
+ struct sort_iter_set *i = iter->data + idx;
+
+ do {
+ i->k = bkey_next_skip_noops(i->k, i->end);
+ } while (i->k != i->end && bkey_deleted(i->k));
+
+ if (i->k == i->end)
+ array_remove_item(iter->data, iter->used, idx);
+ else
+ __sort_iter_sift(iter, idx, extent_sort_fix_overlapping_cmp);
+}
+
+struct btree_nr_keys
+bch2_extent_sort_fix_overlapping(struct bch_fs *c, struct bset *dst,
+ struct sort_iter *iter)
+{
+ struct btree *b = iter->b;
+ struct bkey_format *f = &b->format;
+ struct sort_iter_set *_l = iter->data, *_r = iter->data + 1;
+ struct bkey_packed *out = dst->start;
+ struct bkey l_unpacked, r_unpacked;
+ struct bkey_s l, r;
+ struct btree_nr_keys nr;
+ struct bkey_buf split;
+ unsigned i;
+
+ memset(&nr, 0, sizeof(nr));
+ bch2_bkey_buf_init(&split);
+
+ sort_iter_sort(iter, extent_sort_fix_overlapping_cmp);
+ for (i = 0; i < iter->used;) {
+ if (bkey_deleted(iter->data[i].k))
+ __sort_iter_advance(iter, i,
+ extent_sort_fix_overlapping_cmp);
+ else
+ i++;
+ }
+
+ while (!sort_iter_end(iter)) {
+ l = __bkey_disassemble(b, _l->k, &l_unpacked);
+
+ if (iter->used == 1) {
+ extent_sort_append(c, f, &nr, &out, l);
+ extent_iter_advance(iter, 0);
+ continue;
+ }
+
+ r = __bkey_disassemble(b, _r->k, &r_unpacked);
+
+ /* If current key and next key don't overlap, just append */
+ if (bkey_cmp(l.k->p, bkey_start_pos(r.k)) <= 0) {
+ extent_sort_append(c, f, &nr, &out, l);
+ extent_iter_advance(iter, 0);
+ continue;
+ }
+
+ /* Skip 0 size keys */
+ if (!r.k->size) {
+ extent_iter_advance(iter, 1);
+ continue;
+ }
+
+ /*
+ * overlap: keep the newer key and trim the older key so they
+ * don't overlap. comparing pointers tells us which one is
+ * newer, since the bsets are appended one after the other.
+ */
+
+ /* can't happen because of comparison func */
+ BUG_ON(_l->k < _r->k &&
+ !bkey_cmp(bkey_start_pos(l.k), bkey_start_pos(r.k)));
+
+ if (_l->k > _r->k) {
+ /* l wins, trim r */
+ if (bkey_cmp(l.k->p, r.k->p) >= 0) {
+ extent_iter_advance(iter, 1);
+ } else {
+ bch2_cut_front_s(l.k->p, r);
+ extent_save(b, _r->k, r.k);
+ __sort_iter_sift(iter, 1,
+ extent_sort_fix_overlapping_cmp);
+ }
+ } else if (bkey_cmp(l.k->p, r.k->p) > 0) {
+
+ /*
+ * r wins, but it overlaps in the middle of l - split l:
+ */
+ bch2_bkey_buf_reassemble(&split, c, l.s_c);
+ bch2_cut_back(bkey_start_pos(r.k), split.k);
+
+ bch2_cut_front_s(r.k->p, l);
+ extent_save(b, _l->k, l.k);
+
+ __sort_iter_sift(iter, 0,
+ extent_sort_fix_overlapping_cmp);
+
+ extent_sort_append(c, f, &nr, &out,
+ bkey_i_to_s(split.k));
+ } else {
+ bch2_cut_back_s(bkey_start_pos(r.k), l);
+ extent_save(b, _l->k, l.k);
+ }
+ }
+
+ dst->u64s = cpu_to_le16((u64 *) out - dst->_data);
+
+ bch2_bkey_buf_exit(&split, c);
+ return nr;
+}
+
static inline int sort_extents_cmp(struct btree *b,
struct bkey_packed *l,
struct bkey_packed *r)
{
- return bkey_cmp_packed(b, l, r) ?:
+ return bch2_bkey_cmp_packed(b, l, r) ?:
(int) bkey_deleted(l) - (int) bkey_deleted(r);
}
return (u64 *) out - (u64 *) dst;
}
-static inline int sort_key_whiteouts_cmp(struct btree *b,
- struct bkey_packed *l,
- struct bkey_packed *r)
-{
- return bkey_cmp_packed(b, l, r);
-}
-
-unsigned bch2_sort_key_whiteouts(struct bkey_packed *dst,
- struct sort_iter *iter)
-{
- struct bkey_packed *in, *out = dst;
-
- sort_iter_sort(iter, sort_key_whiteouts_cmp);
-
- while ((in = sort_iter_next(iter, sort_key_whiteouts_cmp))) {
- bkey_copy(out, in);
- out = bkey_next(out);
- }
-
- return (u64 *) out - (u64 *) dst;
-}
-
static inline int sort_extent_whiteouts_cmp(struct btree *b,
struct bkey_packed *l,
struct bkey_packed *r)
projects / bcachefs-tools-debian / blobdiff