iter->l[l].b != BTREE_ITER_NOT_END;
}
+/* Returns < 0 if @k is before iter pos, > 0 if @k is after */
+static inline int __btree_iter_pos_cmp(struct btree_iter *iter,
+ const struct btree *b,
+ const struct bkey_packed *k,
+ bool interior_node)
+{
+ int cmp = bkey_cmp_left_packed(b, k, &iter->pos);
+
+ if (cmp)
+ return cmp;
+ if (bkey_deleted(k))
+ return -1;
+
+ /*
+ * Normally, for extents we want the first key strictly greater than
+ * the iterator position - with the exception that for interior nodes,
+ * we don't want to advance past the last key if the iterator position
+ * is POS_MAX:
+ */
+ if (iter->flags & BTREE_ITER_IS_EXTENTS &&
+ (!interior_node ||
+ bkey_cmp_left_packed_byval(b, k, POS_MAX)))
+ return -1;
+ return 1;
+}
+
+static inline int btree_iter_pos_cmp(struct btree_iter *iter,
+ const struct btree *b,
+ const struct bkey_packed *k)
+{
+ return __btree_iter_pos_cmp(iter, b, k, b->level != 0);
+}
+
/* Btree node locking: */
/*
struct btree_iter *linked;
EBUG_ON(iter->l[b->level].b != b);
- EBUG_ON(iter->lock_seq[b->level] + 1 != b->lock.state.seq);
+ EBUG_ON(iter->l[b->level].lock_seq + 1 != b->lock.state.seq);
for_each_btree_iter_with_node(iter, b, linked)
- linked->lock_seq[b->level] += 2;
+ linked->l[b->level].lock_seq += 2;
six_unlock_write(&b->lock);
}
&b->lock.state.counter);
}
-/*
- * Lock a btree node if we already have it locked on one of our linked
- * iterators:
- */
-static inline bool btree_node_lock_increment(struct btree_iter *iter,
- struct btree *b, unsigned level,
- enum btree_node_locked_type want)
-{
- struct btree_iter *linked;
-
- for_each_linked_btree_iter(iter, linked)
- if (linked->l[level].b == b &&
- btree_node_locked_type(linked, level) >= want) {
- six_lock_increment(&b->lock, want);
- return true;
- }
-
- return false;
-}
-
bool __bch2_btree_node_relock(struct btree_iter *iter, unsigned level)
{
struct btree *b = btree_iter_node(iter, level);
if (race_fault())
return false;
- if (!six_relock_type(&b->lock, want, iter->lock_seq[level]) &&
- !(iter->lock_seq[level] >> 1 == b->lock.state.seq >> 1 &&
+ if (!six_relock_type(&b->lock, want, iter->l[level].lock_seq) &&
+ !(iter->l[level].lock_seq >> 1 == b->lock.state.seq >> 1 &&
btree_node_lock_increment(iter, b, level, want)))
return false;
if (btree_node_locked(iter, level)
? six_lock_tryupgrade(&b->lock)
- : six_relock_type(&b->lock, SIX_LOCK_intent, iter->lock_seq[level]))
+ : six_relock_type(&b->lock, SIX_LOCK_intent, iter->l[level].lock_seq))
goto success;
- if (iter->lock_seq[level] >> 1 == b->lock.state.seq >> 1 &&
+ if (iter->l[level].lock_seq >> 1 == b->lock.state.seq >> 1 &&
btree_node_lock_increment(iter, b, level, BTREE_NODE_INTENT_LOCKED)) {
btree_node_unlock(iter, level);
goto success;
struct btree_iter *linked;
bool ret = true;
- /* Can't have children locked before ancestors: */
- EBUG_ON(iter->nodes_locked && level > __ffs(iter->nodes_locked));
-
- /*
- * Can't hold any read locks while we block taking an intent lock - see
- * below for reasoning, and we should have already dropped any read
- * locks in the current iterator
- */
- EBUG_ON(type == SIX_LOCK_intent &&
- iter->nodes_locked != iter->nodes_intent_locked);
-
- if (btree_node_lock_increment(iter, b, level, type))
- return true;
-
- /*
- * Must lock btree nodes in key order - this case happens when locking
- * the prev sibling in btree node merging:
- */
- if (iter->nodes_locked &&
- __ffs(iter->nodes_locked) <= level &&
- __btree_iter_cmp(iter->btree_id, pos, iter))
- return false;
-
- for_each_linked_btree_iter(iter, linked) {
+ /* Check if it's safe to block: */
+ for_each_btree_iter(iter, linked) {
if (!linked->nodes_locked)
continue;
- /* We have to lock btree nodes in key order: */
+ /* * Must lock btree nodes in key order: */
if (__btree_iter_cmp(iter->btree_id, pos, linked) < 0)
ret = false;
if (linked->btree_id == iter->btree_id &&
level > __fls(linked->nodes_locked)) {
if (may_drop_locks) {
- linked->locks_want = max_t(unsigned,
- linked->locks_want,
- iter->locks_want);
+ linked->locks_want =
+ max(level + 1, max_t(unsigned,
+ linked->locks_want,
+ iter->locks_want));
btree_iter_get_locks(linked, true);
}
ret = false;
if (ret)
__btree_node_lock_type(c, b, type);
+ else
+ trans_restart();
+
return ret;
}
/* Btree iterator locking: */
#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_iter_verify_locks(struct btree_iter *iter)
+void __bch2_btree_iter_verify_locks(struct btree_iter *iter)
{
unsigned l;
+ BUG_ON((iter->flags & BTREE_ITER_NOUNLOCK) &&
+ !btree_node_locked(iter, 0));
+
for (l = 0; btree_iter_node(iter, l); l++) {
if (iter->uptodate >= BTREE_ITER_NEED_RELOCK &&
!btree_node_locked(iter, l))
btree_node_locked_type(iter, l));
}
}
+
+void bch2_btree_iter_verify_locks(struct btree_iter *iter)
+{
+ struct btree_iter *linked;
+
+ for_each_btree_iter(iter, linked)
+ __bch2_btree_iter_verify_locks(linked);
+
+}
#endif
__flatten
break;
}
}
-
- bch2_btree_iter_verify_locks(linked);
}
+
+ bch2_btree_iter_verify_locks(iter);
}
int bch2_btree_iter_unlock(struct btree_iter *iter)
struct btree_node_iter tmp = l->iter;
struct bkey_packed *k;
+ if (iter->uptodate > BTREE_ITER_NEED_PEEK)
+ return;
+
bch2_btree_node_iter_verify(&l->iter, b);
/*
* For interior nodes, the iterator will have skipped past
* deleted keys:
+ *
+ * For extents, the iterator may have skipped past deleted keys (but not
+ * whiteouts)
*/
- k = b->level
- ? bch2_btree_node_iter_prev(&tmp, b)
+ k = b->level || iter->flags & BTREE_ITER_IS_EXTENTS
+ ? bch2_btree_node_iter_prev_filter(&tmp, b, KEY_TYPE_discard)
: bch2_btree_node_iter_prev_all(&tmp, b);
- if (k && btree_iter_pos_cmp_packed(b, &iter->pos, k,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ if (k && btree_iter_pos_cmp(iter, b, k) > 0) {
char buf[100];
struct bkey uk = bkey_unpack_key(b, k);
- bch2_bkey_to_text(buf, sizeof(buf), &uk);
- panic("prev key should be before after pos:\n%s\n%llu:%llu\n",
+ bch2_bkey_to_text(&PBUF(buf), &uk);
+ panic("prev key should be before iter pos:\n%s\n%llu:%llu\n",
buf, iter->pos.inode, iter->pos.offset);
}
k = bch2_btree_node_iter_peek_all(&l->iter, b);
- if (k && !btree_iter_pos_cmp_packed(b, &iter->pos, k,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ if (k && btree_iter_pos_cmp(iter, b, k) < 0) {
char buf[100];
struct bkey uk = bkey_unpack_key(b, k);
- bch2_bkey_to_text(buf, sizeof(buf), &uk);
- panic("next key should be before iter pos:\n%llu:%llu\n%s\n",
+ bch2_bkey_to_text(&PBUF(buf), &uk);
+ panic("iter should be after current key:\n"
+ "iter pos %llu:%llu\n"
+ "cur key %s\n",
iter->pos.inode, iter->pos.offset, buf);
}
- if (iter->uptodate == BTREE_ITER_UPTODATE &&
- (iter->flags & BTREE_ITER_TYPE) != BTREE_ITER_NODES) {
- BUG_ON(!bkey_whiteout(&iter->k) &&
- bch2_btree_node_iter_end(&l->iter));
- }
+ BUG_ON(iter->uptodate == BTREE_ITER_UPTODATE &&
+ (iter->flags & BTREE_ITER_TYPE) == BTREE_ITER_KEYS &&
+ !bkey_whiteout(&iter->k) &&
+ bch2_btree_node_iter_end(&l->iter));
}
void bch2_btree_iter_verify(struct btree_iter *iter, struct btree *b)
__bch2_btree_iter_verify(linked, b);
}
+#else
+
+static inline void __bch2_btree_iter_verify(struct btree_iter *iter,
+ struct btree *b) {}
+
#endif
static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
struct btree_node_iter_set *set;
unsigned offset = __btree_node_key_to_offset(b, where);
int shift = new_u64s - clobber_u64s;
- unsigned old_end = (int) __btree_node_key_to_offset(b, end) - shift;
+ unsigned old_end = t->end_offset - shift;
btree_node_iter_for_each(node_iter, set)
if (set->end == old_end)
/* didn't find the bset in the iterator - might have to readd it: */
if (new_u64s &&
- btree_iter_pos_cmp_packed(b, &iter->pos, where,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ btree_iter_pos_cmp(iter, b, where) > 0) {
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
bch2_btree_node_iter_push(node_iter, b, where, end);
}
return;
found:
- set->end = (int) set->end + shift;
+ set->end = t->end_offset;
/* Iterator hasn't gotten to the key that changed yet: */
if (set->k < offset)
return;
if (new_u64s &&
- btree_iter_pos_cmp_packed(b, &iter->pos, where,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ btree_iter_pos_cmp(iter, b, where) > 0) {
set->k = offset;
} else if (set->k < offset + clobber_u64s) {
set->k = offset + new_u64s;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
bch2_btree_node_iter_sort(node_iter, b);
- if (!b->level && node_iter == &iter->l[0].iter)
+ if (!b->level && node_iter == &iter->l[0].iter) {
+ /*
+ * not legal to call bkey_debugcheck() here, because we're
+ * called midway through the update path after update has been
+ * marked but before deletes have actually happened:
+ */
+#if 0
__btree_iter_peek_all(iter, &iter->l[0], &iter->k);
+#endif
+ struct btree_iter_level *l = &iter->l[0];
+ struct bkey_packed *k =
+ bch2_btree_node_iter_peek_all(&l->iter, l->b);
+
+ if (unlikely(!k))
+ iter->k.type = KEY_TYPE_deleted;
+ else
+ bkey_disassemble(l->b, k, &iter->k);
+ }
iter_current_key_not_modified:
/*
* always point to the key for the child node the btree iterator points
* to.
*/
- if (b->level && new_u64s && !bkey_deleted(where) &&
- btree_iter_pos_cmp_packed(b, &iter->pos, where,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ if (b->level && new_u64s &&
+ btree_iter_pos_cmp(iter, b, where) > 0) {
struct bset_tree *t;
struct bkey_packed *k;
k = bch2_bkey_prev_all(b, t,
bch2_btree_node_iter_bset_pos(node_iter, b, t));
if (k &&
- __btree_node_iter_cmp(node_iter, b,
- k, where) > 0) {
+ bkey_iter_cmp(b, k, where) > 0) {
struct btree_node_iter_set *set;
unsigned offset =
__btree_node_key_to_offset(b, bkey_next(k));
}
void bch2_btree_node_iter_fix(struct btree_iter *iter,
- struct btree *b,
- struct btree_node_iter *node_iter,
- struct bset_tree *t,
- struct bkey_packed *where,
- unsigned clobber_u64s,
- unsigned new_u64s)
+ struct btree *b,
+ struct btree_node_iter *node_iter,
+ struct bkey_packed *where,
+ unsigned clobber_u64s,
+ unsigned new_u64s)
{
+ struct bset_tree *t = bch2_bkey_to_bset(b, where);
struct btree_iter *linked;
if (node_iter != &iter->l[b->level].iter)
__bch2_btree_node_iter_fix(linked, b,
&linked->l[b->level].iter, t,
where, clobber_u64s, new_u64s);
-
- /* interior node iterators are... special... */
- if (!b->level)
- bch2_btree_iter_verify(iter, b);
}
static inline struct bkey_s_c __btree_iter_unpack(struct btree_iter *iter,
* signal to bch2_btree_iter_peek_slot() that we're currently at
* a hole
*/
- u->type = KEY_TYPE_DELETED;
+ u->type = KEY_TYPE_deleted;
return bkey_s_c_null;
}
bch2_btree_node_iter_peek(&l->iter, l->b));
}
-static inline void __btree_iter_advance(struct btree_iter_level *l)
+static inline bool btree_iter_advance_to_pos(struct btree_iter *iter,
+ struct btree_iter_level *l,
+ int max_advance)
{
- bch2_btree_node_iter_advance(&l->iter, l->b);
+ struct bkey_packed *k;
+ int nr_advanced = 0;
+
+ while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
+ btree_iter_pos_cmp(iter, l->b, k) < 0) {
+ if (max_advance > 0 && nr_advanced >= max_advance)
+ return false;
+
+ bch2_btree_node_iter_advance(&l->iter, l->b);
+ nr_advanced++;
+ }
+
+ return true;
}
/*
char buf[100];
struct bkey uk = bkey_unpack_key(b, k);
- bch2_bkey_to_text(buf, sizeof(buf), &uk);
+ bch2_bkey_to_text(&PBUF(buf), &uk);
panic("parent iter doesn't point to new node:\n%s\n%llu:%llu\n",
buf, b->key.k.p.inode, b->key.k.p.offset);
}
btree_node_unlock(iter, b->level + 1);
}
-/* Returns true if @k is after iterator position @pos */
-static inline bool btree_iter_pos_cmp(struct btree_iter *iter,
- const struct bkey *k)
-{
- int cmp = bkey_cmp(k->p, iter->pos);
-
- return cmp > 0 ||
- (cmp == 0 &&
- !(iter->flags & BTREE_ITER_IS_EXTENTS) && !bkey_deleted(k));
-}
-
static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
struct btree *b)
{
- return !btree_iter_pos_cmp(iter, &b->key.k) &&
- bkey_cmp(b->key.k.p, POS_MAX);
+ return __btree_iter_pos_cmp(iter, NULL,
+ bkey_to_packed(&b->key), true) < 0;
}
static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
}
static inline void __btree_iter_init(struct btree_iter *iter,
- struct btree *b)
+ unsigned level)
{
- struct btree_iter_level *l = &iter->l[b->level];
+ struct btree_iter_level *l = &iter->l[level];
+
+ bch2_btree_node_iter_init(&l->iter, l->b, &iter->pos);
- bch2_btree_node_iter_init(&l->iter, b, iter->pos,
- iter->flags & BTREE_ITER_IS_EXTENTS,
- btree_node_is_extents(b));
+ if (iter->flags & BTREE_ITER_IS_EXTENTS)
+ btree_iter_advance_to_pos(iter, l, -1);
/* Skip to first non whiteout: */
- if (b->level)
- bch2_btree_node_iter_peek(&l->iter, b);
+ if (level)
+ bch2_btree_node_iter_peek(&l->iter, l->b);
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
}
EBUG_ON(!btree_iter_pos_in_node(iter, b));
EBUG_ON(b->lock.state.seq & 1);
- iter->lock_seq[b->level] = b->lock.state.seq;
+ iter->l[b->level].lock_seq = b->lock.state.seq;
iter->l[b->level].b = b;
- __btree_iter_init(iter, b);
+ __btree_iter_init(iter, b->level);
}
/*
struct btree_iter *linked;
unsigned level = b->level;
+ /* caller now responsible for unlocking @b */
+
+ BUG_ON(iter->l[level].b != b);
+ BUG_ON(!btree_node_intent_locked(iter, level));
+
+ iter->l[level].b = BTREE_ITER_NOT_END;
+ mark_btree_node_unlocked(iter, level);
+
for_each_btree_iter(iter, linked)
if (linked->l[level].b == b) {
- btree_node_unlock(linked, level);
+ __btree_node_unlock(linked, level);
linked->l[level].b = BTREE_ITER_NOT_END;
}
}
struct btree_iter *linked;
for_each_btree_iter_with_node(iter, b, linked)
- __btree_iter_init(linked, b);
+ __btree_iter_init(linked, b->level);
}
static inline int btree_iter_lock_root(struct btree_iter *iter,
*/
iter->level = depth_want;
iter->l[iter->level].b = NULL;
- return 0;
+ return 1;
}
lock_type = __btree_lock_want(iter, iter->level);
if (__bch2_btree_iter_relock(iter))
return 0;
- iter->flags &= ~BTREE_ITER_AT_END_OF_LEAF;
-
/*
* XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
* here unnecessary
*
* XXX correctly using BTREE_ITER_UPTODATE should make this unnecessary
*/
- if (btree_iter_node(iter, iter->level)) {
- struct btree_iter_level *l = &iter->l[iter->level];
- struct bkey_s_c k;
- struct bkey u;
-
- while ((k = __btree_iter_peek_all(iter, l, &u)).k &&
- !btree_iter_pos_cmp(iter, k.k))
- __btree_iter_advance(l);
- }
+ if (btree_iter_node(iter, iter->level))
+ btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
/*
* Note: iter->nodes[iter->level] may be temporarily NULL here - that
? btree_iter_down(iter)
: btree_iter_lock_root(iter, depth_want);
if (unlikely(ret)) {
+ if (ret == 1)
+ return 0;
+
iter->level = depth_want;
iter->l[iter->level].b = BTREE_ITER_NOT_END;
return ret;
}
iter->uptodate = BTREE_ITER_NEED_PEEK;
+
bch2_btree_iter_verify_locks(iter);
+ __bch2_btree_iter_verify(iter, iter->l[iter->level].b);
return 0;
}
enum btree_iter_type type)
{
EBUG_ON(iter->btree_id >= BTREE_ID_NR);
- EBUG_ON((iter->flags & BTREE_ITER_TYPE) != type);
EBUG_ON(!!(iter->flags & BTREE_ITER_IS_EXTENTS) !=
(iter->btree_id == BTREE_ID_EXTENTS &&
type != BTREE_ITER_NODES));
void bch2_btree_iter_set_pos_same_leaf(struct btree_iter *iter, struct bpos new_pos)
{
struct btree_iter_level *l = &iter->l[0];
- struct bkey_packed *k;
EBUG_ON(iter->level != 0);
EBUG_ON(bkey_cmp(new_pos, iter->pos) < 0);
iter->pos = new_pos;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
- while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
- !btree_iter_pos_cmp_packed(l->b, &iter->pos, k,
- iter->flags & BTREE_ITER_IS_EXTENTS))
- __btree_iter_advance(l);
+ btree_iter_advance_to_pos(iter, l, -1);
- if (!k && btree_iter_pos_after_node(iter, l->b)) {
+ if (bch2_btree_node_iter_end(&l->iter) &&
+ btree_iter_pos_after_node(iter, l->b))
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
- iter->flags |= BTREE_ITER_AT_END_OF_LEAF;
- }
}
void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
level = btree_iter_up_until_locked(iter, true);
if (btree_iter_node(iter, level)) {
- unsigned nr_advanced = 0;
- struct btree_iter_level *l = &iter->l[level];
- struct bkey_s_c k;
- struct bkey u;
-
/*
* We might have to skip over many keys, or just a few: try
* advancing the node iterator, and if we have to skip over too
* many keys just reinit it (or if we're rewinding, since that
* is expensive).
*/
- if (cmp > 0) {
- while ((k = __btree_iter_peek_all(iter, l, &u)).k &&
- !btree_iter_pos_cmp(iter, k.k)) {
- if (nr_advanced > 8)
- goto reinit_node;
-
- __btree_iter_advance(l);
- nr_advanced++;
- }
- } else {
-reinit_node:
- __btree_iter_init(iter, iter->l[level].b);
- }
+ if (cmp < 0 ||
+ !btree_iter_advance_to_pos(iter, &iter->l[level], 8))
+ __btree_iter_init(iter, level);
/* Don't leave it locked if we're not supposed to: */
if (btree_lock_want(iter, level) == BTREE_NODE_UNLOCKED)
}
do {
- __btree_iter_advance(l);
+ bch2_btree_node_iter_advance(&l->iter, l->b);
p = bch2_btree_node_iter_peek_all(&l->iter, l->b);
if (unlikely(!p))
return bch2_btree_iter_peek_next_leaf(iter);
}
static inline struct bkey_s_c
-__bch2_btree_iter_peek_slot(struct btree_iter *iter)
+__bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
{
struct btree_iter_level *l = &iter->l[0];
+ struct btree_node_iter node_iter;
struct bkey_s_c k;
struct bkey n;
int ret;
while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
bkey_deleted(k.k) &&
bkey_cmp(bkey_start_pos(k.k), iter->pos) == 0)
- __btree_iter_advance(l);
+ bch2_btree_node_iter_advance(&l->iter, l->b);
+
+ /*
+ * iterator is now at the correct position for inserting at iter->pos,
+ * but we need to keep iterating until we find the first non whiteout so
+ * we know how big a hole we have, if any:
+ */
+
+ node_iter = l->iter;
+ if (k.k && bkey_whiteout(k.k))
+ k = __btree_iter_unpack(iter, l, &iter->k,
+ bch2_btree_node_iter_peek(&node_iter, l->b));
/*
* If we got to the end of the node, check if we need to traverse to the
if (k.k &&
!bkey_whiteout(k.k) &&
bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
+ /*
+ * if we skipped forward to find the first non whiteout and
+ * there _wasn't_ actually a hole, we want the iterator to be
+ * pointed at the key we found:
+ */
+ l->iter = node_iter;
+
EBUG_ON(bkey_cmp(k.k->p, iter->pos) < 0);
EBUG_ON(bkey_deleted(k.k));
iter->uptodate = BTREE_ITER_UPTODATE;
}
/* hole */
+
+ /* holes can't span inode numbers: */
+ if (iter->pos.offset == KEY_OFFSET_MAX) {
+ if (iter->pos.inode == KEY_INODE_MAX)
+ return bkey_s_c_null;
+
+ iter->pos = bkey_successor(iter->pos);
+ goto recheck;
+ }
+
+ if (!k.k)
+ k.k = &l->b->key.k;
+
bkey_init(&n);
n.p = iter->pos;
+ bch2_key_resize(&n,
+ min_t(u64, KEY_SIZE_MAX,
+ (k.k->p.inode == n.p.inode
+ ? bkey_start_offset(k.k)
+ : KEY_OFFSET_MAX) -
+ n.p.offset));
- if (iter->flags & BTREE_ITER_IS_EXTENTS) {
- if (n.p.offset == KEY_OFFSET_MAX) {
- if (n.p.inode == KEY_INODE_MAX)
- return bkey_s_c_null;
+ EBUG_ON(!n.size);
- iter->pos = bkey_successor(iter->pos);
- goto recheck;
- }
+ iter->k = n;
+ iter->uptodate = BTREE_ITER_UPTODATE;
+ return (struct bkey_s_c) { &iter->k, NULL };
+}
- if (k.k && bkey_whiteout(k.k)) {
- struct btree_node_iter node_iter = l->iter;
+static inline struct bkey_s_c
+__bch2_btree_iter_peek_slot(struct btree_iter *iter)
+{
+ struct btree_iter_level *l = &iter->l[0];
+ struct bkey_s_c k;
+ int ret;
- k = __btree_iter_unpack(iter, l, &iter->k,
- bch2_btree_node_iter_peek(&node_iter, l->b));
- }
+ if (iter->flags & BTREE_ITER_IS_EXTENTS)
+ return __bch2_btree_iter_peek_slot_extents(iter);
- if (!k.k)
- k.k = &l->b->key.k;
+recheck:
+ while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
+ bkey_deleted(k.k) &&
+ bkey_cmp(k.k->p, iter->pos) == 0)
+ bch2_btree_node_iter_advance(&l->iter, l->b);
- bch2_key_resize(&n,
- min_t(u64, KEY_SIZE_MAX,
- (k.k->p.inode == n.p.inode
- ? bkey_start_offset(k.k)
- : KEY_OFFSET_MAX) -
- n.p.offset));
+ /*
+ * If we got to the end of the node, check if we need to traverse to the
+ * next node:
+ */
+ if (unlikely(!k.k && btree_iter_pos_after_node(iter, l->b))) {
+ btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
+ ret = bch2_btree_iter_traverse(iter);
+ if (unlikely(ret))
+ return bkey_s_c_err(ret);
- EBUG_ON(!n.size);
+ goto recheck;
}
- iter->k = n;
- iter->uptodate = BTREE_ITER_UPTODATE;
- return (struct bkey_s_c) { &iter->k, NULL };
+ if (k.k &&
+ !bkey_deleted(k.k) &&
+ !bkey_cmp(iter->pos, k.k->p)) {
+ iter->uptodate = BTREE_ITER_UPTODATE;
+ return k;
+ } else {
+ /* hole */
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos;
+
+ iter->uptodate = BTREE_ITER_UPTODATE;
+ return (struct bkey_s_c) { &iter->k, NULL };
+ }
}
struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
}
if (!bkey_deleted(&iter->k))
- __btree_iter_advance(&iter->l[0]);
+ bch2_btree_node_iter_advance(&iter->l[0].iter, iter->l[0].b);
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
for_each_linked_btree_iter(iter, linked)
if (linked->next == iter) {
linked->next = iter->next;
+ iter->next = iter;
return;
}
if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
unsigned nr_iters = 0;
- for_each_btree_iter(iter, new)
- nr_iters++;
+ for_each_btree_iter(new, iter)
+ if (iter->btree_id == new->btree_id)
+ nr_iters++;
BUG_ON(nr_iters > SIX_LOCK_MAX_RECURSE);
}
void bch2_btree_iter_copy(struct btree_iter *dst, struct btree_iter *src)
{
+ unsigned i;
+
__bch2_btree_iter_unlock(dst);
memcpy(dst, src, offsetof(struct btree_iter, next));
- dst->nodes_locked = dst->nodes_intent_locked = 0;
- dst->uptodate = BTREE_ITER_NEED_RELOCK;
+
+ for (i = 0; i < BTREE_MAX_DEPTH; i++)
+ if (btree_node_locked(dst, i))
+ six_lock_increment(&dst->l[i].b->lock,
+ __btree_lock_want(dst, i));
+}
+
+/* new transactional stuff: */
+
+static void btree_trans_verify(struct btree_trans *trans)
+{
+ unsigned i;
+
+ for (i = 0; i < trans->nr_iters; i++) {
+ struct btree_iter *iter = &trans->iters[i];
+
+ BUG_ON(btree_iter_linked(iter) !=
+ ((trans->iters_linked & (1 << i)) &&
+ !is_power_of_2(trans->iters_linked)));
+ }
+}
+
+static inline unsigned btree_trans_iter_idx(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ ssize_t idx = iter - trans->iters;
+
+ BUG_ON(idx < 0 || idx >= trans->nr_iters);
+ BUG_ON(!(trans->iters_live & (1U << idx)));
+
+ return idx;
+}
+
+void bch2_trans_iter_put(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ ssize_t idx = btree_trans_iter_idx(trans, iter);
+
+ trans->iters_live &= ~(1U << idx);
+}
+
+void bch2_trans_iter_free(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ ssize_t idx = btree_trans_iter_idx(trans, iter);
+
+ trans->iters_live &= ~(1U << idx);
+ trans->iters_linked &= ~(1U << idx);
+ bch2_btree_iter_unlink(iter);
+}
+
+static int btree_trans_realloc_iters(struct btree_trans *trans)
+{
+ struct btree_iter *new_iters;
+ unsigned i;
+
+ bch2_trans_unlock(trans);
+
+ new_iters = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
+
+ memcpy(new_iters, trans->iters,
+ sizeof(struct btree_iter) * trans->nr_iters);
+ trans->iters = new_iters;
+
+ for (i = 0; i < trans->nr_iters; i++)
+ trans->iters[i].next = &trans->iters[i];
+
+ if (trans->iters_linked) {
+ unsigned first_linked = __ffs(trans->iters_linked);
+
+ for (i = first_linked + 1; i < trans->nr_iters; i++)
+ if (trans->iters_linked & (1 << i))
+ bch2_btree_iter_link(&trans->iters[first_linked],
+ &trans->iters[i]);
+ }
+
+ btree_trans_verify(trans);
+
+ if (trans->iters_live) {
+ trans_restart();
+ return -EINTR;
+ }
+
+ return 0;
+}
+
+void bch2_trans_preload_iters(struct btree_trans *trans)
+{
+ if (trans->iters == trans->iters_onstack)
+ btree_trans_realloc_iters(trans);
+}
+
+static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
+ unsigned btree_id,
+ unsigned flags, u64 iter_id)
+{
+ struct btree_iter *iter;
+ int idx;
+
+ BUG_ON(trans->nr_iters > BTREE_ITER_MAX);
+
+ for (idx = 0; idx < trans->nr_iters; idx++)
+ if (trans->iter_ids[idx] == iter_id)
+ goto found;
+ idx = -1;
+found:
+ if (idx < 0) {
+ idx = ffz(trans->iters_linked);
+ if (idx < trans->nr_iters)
+ goto got_slot;
+
+ BUG_ON(trans->nr_iters == BTREE_ITER_MAX);
+
+ if (trans->iters == trans->iters_onstack &&
+ trans->nr_iters == ARRAY_SIZE(trans->iters_onstack)) {
+ int ret = btree_trans_realloc_iters(trans);
+ if (ret)
+ return ERR_PTR(ret);
+ }
+
+ idx = trans->nr_iters++;
+got_slot:
+ trans->iter_ids[idx] = iter_id;
+ iter = &trans->iters[idx];
+
+ bch2_btree_iter_init(iter, trans->c, btree_id, POS_MIN, flags);
+ } else {
+ iter = &trans->iters[idx];
+
+ iter->flags &= ~(BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
+ iter->flags |= flags & (BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
+ }
+
+ BUG_ON(trans->iters_live & (1 << idx));
+ trans->iters_live |= 1 << idx;
+
+ if (trans->iters_linked &&
+ !(trans->iters_linked & (1 << idx)))
+ bch2_btree_iter_link(&trans->iters[__ffs(trans->iters_linked)],
+ iter);
+
+ trans->iters_linked |= 1 << idx;
+
+ btree_trans_verify(trans);
+
+ BUG_ON(iter->btree_id != btree_id);
+ BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
+
+ return iter;
+}
+
+struct btree_iter *__bch2_trans_get_iter(struct btree_trans *trans,
+ enum btree_id btree_id,
+ struct bpos pos, unsigned flags,
+ u64 iter_id)
+{
+ struct btree_iter *iter =
+ __btree_trans_get_iter(trans, btree_id, flags, iter_id);
+
+ if (!IS_ERR(iter))
+ bch2_btree_iter_set_pos(iter, pos);
+ return iter;
+}
+
+struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
+ struct btree_iter *src,
+ u64 iter_id)
+{
+ struct btree_iter *iter =
+ __btree_trans_get_iter(trans, src->btree_id,
+ src->flags, iter_id);
+
+ if (!IS_ERR(iter))
+ bch2_btree_iter_copy(iter, src);
+ return iter;
+}
+
+void *bch2_trans_kmalloc(struct btree_trans *trans,
+ size_t size)
+{
+ void *ret;
+
+ if (trans->mem_top + size > trans->mem_bytes) {
+ size_t old_bytes = trans->mem_bytes;
+ size_t new_bytes = roundup_pow_of_two(trans->mem_top + size);
+ void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
+
+ if (!new_mem)
+ return ERR_PTR(-ENOMEM);
+
+ trans->mem = new_mem;
+ trans->mem_bytes = new_bytes;
+
+ if (old_bytes) {
+ trans_restart();
+ return ERR_PTR(-EINTR);
+ }
+ }
+
+ ret = trans->mem + trans->mem_top;
+ trans->mem_top += size;
+ return ret;
+}
+
+int bch2_trans_unlock(struct btree_trans *trans)
+{
+ unsigned iters = trans->iters_linked;
+ int ret = 0;
+
+ while (iters) {
+ unsigned idx = __ffs(iters);
+ struct btree_iter *iter = &trans->iters[idx];
+
+ if (iter->flags & BTREE_ITER_ERROR)
+ ret = -EIO;
+
+ __bch2_btree_iter_unlock(iter);
+ iters ^= 1 << idx;
+ }
+
+ return ret;
+}
+
+void __bch2_trans_begin(struct btree_trans *trans)
+{
+ unsigned idx;
+
+ btree_trans_verify(trans);
+
+ /*
+ * On transaction restart, the transaction isn't required to allocate
+ * all the same iterators it on the last iteration:
+ *
+ * Unlink any iterators it didn't use this iteration, assuming it got
+ * further (allocated an iter with a higher idx) than where the iter
+ * was originally allocated:
+ */
+ while (trans->iters_linked &&
+ trans->iters_live &&
+ (idx = __fls(trans->iters_linked)) >
+ __fls(trans->iters_live)) {
+ trans->iters_linked ^= 1 << idx;
+ bch2_btree_iter_unlink(&trans->iters[idx]);
+ }
+
+ trans->iters_live = 0;
+ trans->nr_updates = 0;
+ trans->mem_top = 0;
+
+ btree_trans_verify(trans);
+}
+
+void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c)
+{
+ trans->c = c;
+ trans->nr_restarts = 0;
+ trans->nr_iters = 0;
+ trans->iters_live = 0;
+ trans->iters_linked = 0;
+ trans->nr_updates = 0;
+ trans->mem_top = 0;
+ trans->mem_bytes = 0;
+ trans->mem = NULL;
+ trans->iters = trans->iters_onstack;
+}
+
+int bch2_trans_exit(struct btree_trans *trans)
+{
+ int ret = bch2_trans_unlock(trans);
+
+ kfree(trans->mem);
+ if (trans->iters != trans->iters_onstack)
+ mempool_free(trans->iters, &trans->c->btree_iters_pool);
+ trans->mem = (void *) 0x1;
+ trans->iters = (void *) 0x1;
+ return ret;
}
projects / bcachefs-tools-debian / blobdiff