+// SPDX-License-Identifier: GPL-2.0
#include "bcachefs.h"
#include "bkey_methods.h"
#include "btree_cache.h"
#include "btree_iter.h"
#include "btree_locking.h"
+#include "btree_update.h"
#include "debug.h"
#include "extents.h"
#include <linux/prefetch.h>
#include <trace/events/bcachefs.h>
-static inline struct bkey_s_c __btree_iter_peek_all(struct btree_iter *,
- struct btree_iter_level *,
- struct bkey *);
-
-#define BTREE_ITER_NOT_END ((struct btree *) 1)
+#define BTREE_ITER_NO_NODE_GET_LOCKS ((struct btree *) 1)
+#define BTREE_ITER_NO_NODE_DROP ((struct btree *) 2)
+#define BTREE_ITER_NO_NODE_LOCK_ROOT ((struct btree *) 3)
+#define BTREE_ITER_NO_NODE_UP ((struct btree *) 4)
+#define BTREE_ITER_NO_NODE_DOWN ((struct btree *) 5)
+#define BTREE_ITER_NO_NODE_INIT ((struct btree *) 6)
+#define BTREE_ITER_NO_NODE_ERROR ((struct btree *) 7)
static inline bool is_btree_node(struct btree_iter *iter, unsigned l)
{
return l < BTREE_MAX_DEPTH &&
- iter->l[l].b &&
- iter->l[l].b != BTREE_ITER_NOT_END;
+ (unsigned long) iter->l[l].b >= 128;
}
-/* Btree node locking: */
+static inline struct bpos btree_iter_search_key(struct btree_iter *iter)
+{
+ struct bpos pos = iter->pos;
-/*
- * Updates the saved lock sequence number, so that bch2_btree_node_relock() will
- * succeed:
- */
-void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
+ if ((iter->flags & BTREE_ITER_IS_EXTENTS) &&
+ bkey_cmp(pos, POS_MAX))
+ pos = bkey_successor(pos);
+ return pos;
+}
+
+static inline bool btree_iter_pos_before_node(struct btree_iter *iter,
+ struct btree *b)
{
- struct btree_iter *linked;
+ return bkey_cmp(iter->pos, b->data->min_key) < 0;
+}
+
+static inline bool btree_iter_pos_after_node(struct btree_iter *iter,
+ struct btree *b)
+{
+ return bkey_cmp(b->key.k.p, btree_iter_search_key(iter)) < 0;
+}
- EBUG_ON(iter->l[b->level].b != b);
- EBUG_ON(iter->l[b->level].lock_seq + 1 != b->lock.state.seq);
+static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
+ struct btree *b)
+{
+ return iter->btree_id == b->btree_id &&
+ !btree_iter_pos_before_node(iter, b) &&
+ !btree_iter_pos_after_node(iter, b);
+}
- for_each_btree_iter_with_node(iter, b, linked)
- linked->l[b->level].lock_seq += 2;
+/* Btree node locking: */
- six_unlock_write(&b->lock);
+void bch2_btree_node_unlock_write(struct btree *b, struct btree_iter *iter)
+{
+ bch2_btree_node_unlock_write_inlined(b, iter);
}
void __bch2_btree_node_lock_write(struct btree *b, struct btree_iter *iter)
{
- struct bch_fs *c = iter->c;
struct btree_iter *linked;
unsigned readers = 0;
- EBUG_ON(btree_node_read_locked(iter, b->level));
+ EBUG_ON(!btree_node_intent_locked(iter, b->level));
- for_each_linked_btree_iter(iter, linked)
+ trans_for_each_iter(iter->trans, linked)
if (linked->l[b->level].b == b &&
btree_node_read_locked(linked, b->level))
readers++;
*/
atomic64_sub(__SIX_VAL(read_lock, readers),
&b->lock.state.counter);
- btree_node_lock_type(c, b, SIX_LOCK_write);
+ btree_node_lock_type(iter->trans->c, b, SIX_LOCK_write);
atomic64_add(__SIX_VAL(read_lock, readers),
&b->lock.state.counter);
}
struct btree *b = btree_iter_node(iter, level);
int want = __btree_lock_want(iter, level);
- if (!b || b == BTREE_ITER_NOT_END)
+ if (!is_btree_node(iter, level))
return false;
if (race_fault())
return false;
- 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)))
+ if (six_relock_type(&b->lock, want, iter->l[level].lock_seq) ||
+ (btree_node_lock_seq_matches(iter, b, level) &&
+ btree_node_lock_increment(iter, b, level, want))) {
+ mark_btree_node_locked(iter, level, want);
+ return true;
+ } else {
return false;
-
- mark_btree_node_locked(iter, level, want);
- return true;
+ }
}
static bool bch2_btree_node_upgrade(struct btree_iter *iter, unsigned level)
: six_relock_type(&b->lock, SIX_LOCK_intent, iter->l[level].lock_seq))
goto success;
- if (iter->l[level].lock_seq >> 1 == b->lock.state.seq >> 1 &&
+ if (btree_node_lock_seq_matches(iter, b, level) &&
btree_node_lock_increment(iter, b, level, BTREE_NODE_INTENT_LOCKED)) {
btree_node_unlock(iter, level);
goto success;
}
static inline bool btree_iter_get_locks(struct btree_iter *iter,
- bool upgrade)
+ bool upgrade, bool trace)
{
unsigned l = iter->level;
int fail_idx = -1;
if (!(upgrade
? bch2_btree_node_upgrade(iter, l)
: bch2_btree_node_relock(iter, l))) {
+ if (trace)
+ (upgrade
+ ? trace_node_upgrade_fail
+ : trace_node_relock_fail)(l, iter->l[l].lock_seq,
+ is_btree_node(iter, l)
+ ? 0
+ : (unsigned long) iter->l[l].b,
+ is_btree_node(iter, l)
+ ? iter->l[l].b->lock.state.seq
+ : 0);
+
fail_idx = l;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
}
*/
while (fail_idx >= 0) {
btree_node_unlock(iter, fail_idx);
- iter->l[fail_idx].b = BTREE_ITER_NOT_END;
+ iter->l[fail_idx].b = BTREE_ITER_NO_NODE_GET_LOCKS;
--fail_idx;
}
if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
iter->uptodate = BTREE_ITER_NEED_PEEK;
- bch2_btree_iter_verify_locks(iter);
+ bch2_btree_trans_verify_locks(iter->trans);
+
return iter->uptodate < BTREE_ITER_NEED_RELOCK;
}
bool __bch2_btree_node_lock(struct btree *b, struct bpos pos,
unsigned level,
struct btree_iter *iter,
- enum six_lock_type type,
- bool may_drop_locks)
+ enum six_lock_type type)
{
- struct bch_fs *c = iter->c;
struct btree_iter *linked;
bool ret = true;
/* Check if it's safe to block: */
- for_each_btree_iter(iter, linked) {
+ trans_for_each_iter(iter->trans, linked) {
if (!linked->nodes_locked)
continue;
*/
if (type == SIX_LOCK_intent &&
linked->nodes_locked != linked->nodes_intent_locked) {
- if (may_drop_locks) {
+ if (!(iter->trans->nounlock)) {
linked->locks_want = max_t(unsigned,
linked->locks_want,
__fls(linked->nodes_locked) + 1);
- btree_iter_get_locks(linked, true);
+ btree_iter_get_locks(linked, true, false);
}
ret = false;
}
*/
if (linked->btree_id == iter->btree_id &&
level > __fls(linked->nodes_locked)) {
- if (may_drop_locks) {
+ if (!(iter->trans->nounlock)) {
linked->locks_want =
max(level + 1, max_t(unsigned,
linked->locks_want,
iter->locks_want));
- btree_iter_get_locks(linked, true);
+ btree_iter_get_locks(linked, true, false);
}
ret = false;
}
}
- if (ret)
- __btree_node_lock_type(c, b, type);
- else
- trans_restart();
+ if (unlikely(!ret)) {
+ trace_trans_restart_would_deadlock(iter->trans->ip);
+ return false;
+ }
- return ret;
+ __btree_node_lock_type(iter->trans->c, b, type);
+ return true;
}
/* Btree iterator locking: */
#ifdef CONFIG_BCACHEFS_DEBUG
-void bch2_btree_iter_verify_locks(struct btree_iter *iter)
+static void bch2_btree_iter_verify_locks(struct btree_iter *iter)
{
unsigned l;
btree_node_locked_type(iter, l));
}
}
-#endif
-__flatten
-static bool __bch2_btree_iter_relock(struct btree_iter *iter)
+void bch2_btree_trans_verify_locks(struct btree_trans *trans)
{
- return iter->uptodate >= BTREE_ITER_NEED_RELOCK
- ? btree_iter_get_locks(iter, false)
- : true;
+ struct btree_iter *iter;
+
+ trans_for_each_iter(trans, iter)
+ bch2_btree_iter_verify_locks(iter);
}
+#else
+static inline void bch2_btree_iter_verify_locks(struct btree_iter *iter) {}
+#endif
-bool bch2_btree_iter_relock(struct btree_iter *iter)
+__flatten
+static bool bch2_btree_iter_relock(struct btree_iter *iter, bool trace)
{
- struct btree_iter *linked;
- bool ret = true;
-
- for_each_btree_iter(iter, linked)
- ret &= __bch2_btree_iter_relock(linked);
-
- return ret;
+ return btree_iter_get_locks(iter, false, trace);
}
bool __bch2_btree_iter_upgrade(struct btree_iter *iter,
iter->locks_want = new_locks_want;
- if (btree_iter_get_locks(iter, true))
+ if (btree_iter_get_locks(iter, true, true))
return true;
/*
* on iterators that might lock ancestors before us to avoid getting
* -EINTR later:
*/
- for_each_linked_btree_iter(iter, linked)
- if (linked->btree_id == iter->btree_id &&
- btree_iter_cmp(linked, iter) <= 0 &&
+ trans_for_each_iter(iter->trans, linked)
+ if (linked != iter &&
+ linked->btree_id == iter->btree_id &&
linked->locks_want < new_locks_want) {
linked->locks_want = new_locks_want;
- btree_iter_get_locks(linked, true);
+ btree_iter_get_locks(linked, true, false);
}
return false;
* might have had to modify locks_want on linked iterators due to lock
* ordering:
*/
- for_each_btree_iter(iter, linked) {
+ trans_for_each_iter(iter->trans, linked) {
unsigned new_locks_want = downgrade_to ?:
(linked->flags & BTREE_ITER_INTENT ? 1 : 0);
break;
}
}
-
- bch2_btree_iter_verify_locks(linked);
}
+
+ bch2_btree_trans_verify_locks(iter->trans);
}
-int bch2_btree_iter_unlock(struct btree_iter *iter)
+/* Btree transaction locking: */
+
+bool bch2_trans_relock(struct btree_trans *trans)
{
- struct btree_iter *linked;
+ struct btree_iter *iter;
+ bool ret = true;
+
+ trans_for_each_iter(trans, iter)
+ if (iter->uptodate == BTREE_ITER_NEED_RELOCK)
+ ret &= bch2_btree_iter_relock(iter, true);
+
+ return ret;
+}
- for_each_btree_iter(iter, linked)
- __bch2_btree_iter_unlock(linked);
+void bch2_trans_unlock(struct btree_trans *trans)
+{
+ struct btree_iter *iter;
- return iter->flags & BTREE_ITER_ERROR ? -EIO : 0;
+ trans_for_each_iter(trans, iter)
+ __bch2_btree_iter_unlock(iter);
}
/* Btree iterator: */
#ifdef CONFIG_BCACHEFS_DEBUG
-static void __bch2_btree_iter_verify(struct btree_iter *iter,
- struct btree *b)
+static void bch2_btree_iter_verify_level(struct btree_iter *iter,
+ unsigned level)
{
- struct btree_iter_level *l = &iter->l[b->level];
+ struct bpos pos = btree_iter_search_key(iter);
+ struct btree_iter_level *l = &iter->l[level];
struct btree_node_iter tmp = l->iter;
- struct bkey_packed *k;
+ bool locked = btree_node_locked(iter, level);
+ struct bkey_packed *p, *k;
+ char buf1[100], buf2[100];
+ const char *msg;
+
+ if (!debug_check_iterators(iter->trans->c))
+ return;
+
+ BUG_ON(iter->level < iter->min_depth);
+
+ if (!btree_iter_node(iter, level))
+ return;
- if (iter->uptodate > BTREE_ITER_NEED_PEEK)
+ if (!bch2_btree_node_relock(iter, level))
return;
- bch2_btree_node_iter_verify(&l->iter, b);
+ /*
+ * Ideally this invariant would always be true, and hopefully in the
+ * future it will be, but for now set_pos_same_leaf() breaks it:
+ */
+ BUG_ON(iter->uptodate < BTREE_ITER_NEED_TRAVERSE &&
+ !btree_iter_pos_in_node(iter, l->b));
+
+ /*
+ * node iterators don't use leaf node iterator:
+ */
+ if (btree_iter_type(iter) == BTREE_ITER_NODES &&
+ level <= iter->min_depth)
+ goto unlock;
+
+ bch2_btree_node_iter_verify(&l->iter, l->b);
/*
* For interior nodes, the iterator will have skipped past
* For extents, the iterator may have skipped past deleted keys (but not
* whiteouts)
*/
- 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)) {
- char buf[100];
- struct bkey uk = bkey_unpack_key(b, k);
+ p = level || btree_node_type_is_extents(iter->btree_id)
+ ? bch2_btree_node_iter_prev_filter(&tmp, l->b, KEY_TYPE_discard)
+ : bch2_btree_node_iter_prev_all(&tmp, l->b);
+ k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
- bch2_bkey_to_text(buf, sizeof(buf), &uk);
- panic("prev key should be before iter pos:\n%s\n%llu:%llu\n",
- buf, iter->pos.inode, iter->pos.offset);
+ if (p && bkey_iter_pos_cmp(l->b, p, &pos) >= 0) {
+ msg = "before";
+ goto err;
}
- 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)) {
- char buf[100];
- struct bkey uk = bkey_unpack_key(b, k);
+ if (k && bkey_iter_pos_cmp(l->b, k, &pos) < 0) {
+ msg = "after";
+ goto err;
+ }
+unlock:
+ if (!locked)
+ btree_node_unlock(iter, level);
+ return;
+err:
+ strcpy(buf1, "(none)");
+ strcpy(buf2, "(none)");
+
+ if (p) {
+ struct bkey uk = bkey_unpack_key(l->b, p);
+ bch2_bkey_to_text(&PBUF(buf1), &uk);
+ }
- bch2_bkey_to_text(buf, sizeof(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 (k) {
+ struct bkey uk = bkey_unpack_key(l->b, k);
+ bch2_bkey_to_text(&PBUF(buf2), &uk);
}
- 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));
+ panic("iterator should be %s key at level %u:\n"
+ "iter pos %s %llu:%llu\n"
+ "prev key %s\n"
+ "cur key %s\n",
+ msg, level,
+ iter->flags & BTREE_ITER_IS_EXTENTS ? ">" : "=>",
+ iter->pos.inode, iter->pos.offset,
+ buf1, buf2);
}
-void bch2_btree_iter_verify(struct btree_iter *iter, struct btree *b)
+static void bch2_btree_iter_verify(struct btree_iter *iter)
{
- struct btree_iter *linked;
+ unsigned i;
+
+ bch2_btree_trans_verify_locks(iter->trans);
+
+ for (i = 0; i < BTREE_MAX_DEPTH; i++)
+ bch2_btree_iter_verify_level(iter, i);
+}
+
+void bch2_btree_trans_verify_iters(struct btree_trans *trans, struct btree *b)
+{
+ struct btree_iter *iter;
+
+ if (!debug_check_iterators(trans->c))
+ return;
- for_each_btree_iter_with_node(iter, b, linked)
- __bch2_btree_iter_verify(linked, b);
+ trans_for_each_iter_with_node(trans, b, iter)
+ bch2_btree_iter_verify_level(iter, b->level);
}
#else
-static inline void __bch2_btree_iter_verify(struct btree_iter *iter,
- struct btree *b) {}
+static inline void bch2_btree_iter_verify_level(struct btree_iter *iter, unsigned l) {}
+static inline void bch2_btree_iter_verify(struct btree_iter *iter) {}
#endif
+static void btree_node_iter_set_set_pos(struct btree_node_iter *iter,
+ struct btree *b,
+ struct bset_tree *t,
+ struct bkey_packed *k)
+{
+ struct btree_node_iter_set *set;
+
+ btree_node_iter_for_each(iter, set)
+ if (set->end == t->end_offset) {
+ set->k = __btree_node_key_to_offset(b, k);
+ bch2_btree_node_iter_sort(iter, b);
+ return;
+ }
+
+ bch2_btree_node_iter_push(iter, b, k, btree_bkey_last(b, t));
+}
+
+static void __bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
+ struct btree *b,
+ struct bkey_packed *where)
+{
+ struct btree_iter_level *l = &iter->l[b->level];
+ struct bpos pos = btree_iter_search_key(iter);
+
+ if (where != bch2_btree_node_iter_peek_all(&l->iter, l->b))
+ return;
+
+ if (bkey_iter_pos_cmp(l->b, where, &pos) < 0)
+ bch2_btree_node_iter_advance(&l->iter, l->b);
+
+ btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+}
+
+void bch2_btree_iter_fix_key_modified(struct btree_iter *iter,
+ struct btree *b,
+ struct bkey_packed *where)
+{
+ struct btree_iter *linked;
+
+ trans_for_each_iter_with_node(iter->trans, b, linked) {
+ __bch2_btree_iter_fix_key_modified(linked, b, where);
+ bch2_btree_iter_verify_level(linked, b->level);
+ }
+}
+
static void __bch2_btree_node_iter_fix(struct btree_iter *iter,
struct btree *b,
struct btree_node_iter *node_iter,
unsigned offset = __btree_node_key_to_offset(b, where);
int shift = new_u64s - clobber_u64s;
unsigned old_end = t->end_offset - shift;
+ unsigned orig_iter_pos = node_iter->data[0].k;
+ bool iter_current_key_modified =
+ orig_iter_pos >= offset &&
+ orig_iter_pos <= offset + clobber_u64s;
+ struct bpos iter_pos = btree_iter_search_key(iter);
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_set_dirty(iter, BTREE_ITER_NEED_PEEK);
-
+ bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
bch2_btree_node_iter_push(node_iter, b, where, end);
-
- if (!b->level &&
- node_iter == &iter->l[0].iter)
- bkey_disassemble(b,
- bch2_btree_node_iter_peek_all(node_iter, b),
- &iter->k);
+ goto fixup_done;
+ } else {
+ /* Iterator is after key that changed */
+ return;
}
- return;
found:
set->end = t->end_offset;
return;
if (new_u64s &&
- btree_iter_pos_cmp_packed(b, &iter->pos, where,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ bkey_iter_pos_cmp(b, where, &iter_pos) >= 0) {
set->k = offset;
} else if (set->k < offset + clobber_u64s) {
set->k = offset + new_u64s;
if (set->k == set->end)
bch2_btree_node_iter_set_drop(node_iter, set);
} else {
+ /* Iterator is after key that changed */
set->k = (int) set->k + shift;
- goto iter_current_key_not_modified;
+ return;
}
- 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)
- __btree_iter_peek_all(iter, &iter->l[0], &iter->k);
-iter_current_key_not_modified:
+fixup_done:
+ if (node_iter->data[0].k != orig_iter_pos)
+ iter_current_key_modified = true;
/*
- * Interior nodes are special because iterators for interior nodes don't
- * obey the usual invariants regarding the iterator position:
- *
- * We may have whiteouts that compare greater than the iterator
- * position, and logically should be in the iterator, but that we
- * skipped past to find the first live key greater than the iterator
- * position. This becomes an issue when we insert a new key that is
- * greater than the current iterator position, but smaller than the
- * whiteouts we've already skipped past - this happens in the course of
- * a btree split.
- *
- * We have to rewind the iterator past to before those whiteouts here,
- * else bkey_node_iter_prev() is not going to work and who knows what
- * else would happen. And we have to do it manually, because here we've
- * already done the insert and the iterator is currently inconsistent:
- *
- * We've got multiple competing invariants, here - we have to be careful
- * about rewinding iterators for interior nodes, because they should
- * always point to the key for the child node the btree iterator points
- * to.
+ * When a new key is added, and the node iterator now points to that
+ * key, the iterator might have skipped past deleted keys that should
+ * come after the key the iterator now points to. We have to rewind to
+ * before those deleted keys - otherwise
+ * bch2_btree_node_iter_prev_all() breaks:
*/
- if (b->level && new_u64s && !bkey_deleted(where) &&
- btree_iter_pos_cmp_packed(b, &iter->pos, where,
- iter->flags & BTREE_ITER_IS_EXTENTS)) {
+ if (!bch2_btree_node_iter_end(node_iter) &&
+ iter_current_key_modified &&
+ (b->level ||
+ btree_node_type_is_extents(iter->btree_id))) {
struct bset_tree *t;
- struct bkey_packed *k;
+ struct bkey_packed *k, *k2, *p;
+
+ k = bch2_btree_node_iter_peek_all(node_iter, b);
for_each_bset(b, t) {
- if (bch2_bkey_to_bset(b, where) == t)
+ bool set_pos = false;
+
+ if (node_iter->data[0].end == t->end_offset)
continue;
- k = bch2_bkey_prev_all(b, t,
- bch2_btree_node_iter_bset_pos(node_iter, b, t));
- if (k &&
- __btree_node_iter_cmp(b, k, where) > 0) {
- struct btree_node_iter_set *set;
- unsigned offset =
- __btree_node_key_to_offset(b, bkey_next(k));
-
- btree_node_iter_for_each(node_iter, set)
- if (set->k == offset) {
- set->k = __btree_node_key_to_offset(b, k);
- bch2_btree_node_iter_sort(node_iter, b);
- goto next_bset;
- }
-
- bch2_btree_node_iter_push(node_iter, b, k,
- btree_bkey_last(b, t));
+ k2 = bch2_btree_node_iter_bset_pos(node_iter, b, t);
+
+ while ((p = bch2_bkey_prev_all(b, t, k2)) &&
+ bkey_iter_cmp(b, k, p) < 0) {
+ k2 = p;
+ set_pos = true;
}
-next_bset:
- t = t;
+
+ if (set_pos)
+ btree_node_iter_set_set_pos(node_iter,
+ b, t, k2);
}
}
+
+ if (!b->level &&
+ node_iter == &iter->l[0].iter &&
+ iter_current_key_modified)
+ btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
}
void bch2_btree_node_iter_fix(struct btree_iter *iter,
struct bset_tree *t = bch2_bkey_to_bset(b, where);
struct btree_iter *linked;
- if (node_iter != &iter->l[b->level].iter)
+ if (node_iter != &iter->l[b->level].iter) {
__bch2_btree_node_iter_fix(iter, b, node_iter, t,
- where, clobber_u64s, new_u64s);
+ where, clobber_u64s, new_u64s);
+
+ if (debug_check_iterators(iter->trans->c))
+ bch2_btree_node_iter_verify(node_iter, b);
+ }
- for_each_btree_iter_with_node(iter, b, linked)
+ trans_for_each_iter_with_node(iter->trans, b, linked) {
__bch2_btree_node_iter_fix(linked, b,
- &linked->l[b->level].iter, t,
- where, clobber_u64s, new_u64s);
+ &linked->l[b->level].iter, t,
+ where, clobber_u64s, new_u64s);
+ bch2_btree_iter_verify_level(linked, b->level);
+ }
}
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;
}
ret = bkey_disassemble(l->b, k, u);
- if (debug_check_bkeys(iter->c))
- bch2_bkey_debugcheck(iter->c, l->b, ret);
+ if (debug_check_bkeys(iter->trans->c))
+ bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
return ret;
}
bch2_btree_node_iter_peek(&l->iter, l->b));
}
-static inline void __btree_iter_advance(struct btree_iter_level *l)
+static inline struct bkey_s_c __btree_iter_prev(struct btree_iter *iter,
+ struct btree_iter_level *l)
+{
+ return __btree_iter_unpack(iter, l, &iter->k,
+ bch2_btree_node_iter_prev(&l->iter, l->b));
+}
+
+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 bpos pos = btree_iter_search_key(iter);
+ struct bkey_packed *k;
+ int nr_advanced = 0;
+
+ while ((k = bch2_btree_node_iter_peek_all(&l->iter, l->b)) &&
+ bkey_iter_pos_cmp(l->b, k, &pos) < 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);
}
-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);
-}
-
-static inline bool btree_iter_pos_in_node(struct btree_iter *iter,
- struct btree *b)
-{
- return iter->btree_id == b->btree_id &&
- bkey_cmp(iter->pos, b->data->min_key) >= 0 &&
- !btree_iter_pos_after_node(iter, b);
-}
-
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 bpos pos = btree_iter_search_key(iter);
+ struct btree_iter_level *l = &iter->l[level];
- bch2_btree_node_iter_init(&l->iter, b, iter->pos,
- iter->flags & BTREE_ITER_IS_EXTENTS);
-
- /* Skip to first non whiteout: */
- if (b->level)
- bch2_btree_node_iter_peek(&l->iter, b);
+ bch2_btree_node_iter_init(&l->iter, l->b, &pos);
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
}
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);
}
/*
enum btree_node_locked_type t;
struct btree_iter *linked;
- for_each_btree_iter(iter, linked)
+ trans_for_each_iter(iter->trans, linked)
if (btree_iter_pos_in_node(linked, b)) {
/*
* bch2_btree_iter_node_drop() has already been called -
btree_iter_node_set(linked, b);
}
-
- six_unlock_intent(&b->lock);
}
void bch2_btree_iter_node_drop(struct btree_iter *iter, struct btree *b)
struct btree_iter *linked;
unsigned level = b->level;
- for_each_btree_iter(iter, linked)
+ trans_for_each_iter(iter->trans, linked)
if (linked->l[level].b == b) {
- btree_node_unlock(linked, level);
- linked->l[level].b = BTREE_ITER_NOT_END;
+ __btree_node_unlock(linked, level);
+ linked->l[level].b = BTREE_ITER_NO_NODE_DROP;
}
}
{
struct btree_iter *linked;
- for_each_btree_iter_with_node(iter, b, linked)
- __btree_iter_init(linked, b);
+ trans_for_each_iter_with_node(iter->trans, b, linked)
+ __btree_iter_init(linked, b->level);
}
static inline int btree_iter_lock_root(struct btree_iter *iter,
unsigned depth_want)
{
- struct bch_fs *c = iter->c;
+ struct bch_fs *c = iter->trans->c;
struct btree *b;
enum six_lock_type lock_type;
unsigned i;
* that depth
*/
iter->level = depth_want;
- iter->l[iter->level].b = NULL;
- return 0;
+ for (i = iter->level; i < BTREE_MAX_DEPTH; i++)
+ iter->l[i].b = NULL;
+ return 1;
}
lock_type = __btree_lock_want(iter, iter->level);
if (unlikely(!btree_node_lock(b, POS_MAX, iter->level,
- iter, lock_type, true)))
+ iter, lock_type)))
return -EINTR;
if (likely(b == c->btree_roots[iter->btree_id].b &&
b->level == iter->level &&
!race_fault())) {
for (i = 0; i < iter->level; i++)
- iter->l[i].b = BTREE_ITER_NOT_END;
+ iter->l[i].b = BTREE_ITER_NO_NODE_LOCK_ROOT;
iter->l[iter->level].b = b;
+ for (i = iter->level + 1; i < BTREE_MAX_DEPTH; i++)
+ iter->l[i].b = NULL;
mark_btree_node_locked(iter, iter->level, lock_type);
btree_iter_node_set(iter, b);
return 0;
-
}
six_unlock_type(&b->lock, lock_type);
noinline
static void btree_iter_prefetch(struct btree_iter *iter)
{
+ struct bch_fs *c = iter->trans->c;
struct btree_iter_level *l = &iter->l[iter->level];
struct btree_node_iter node_iter = l->iter;
struct bkey_packed *k;
BKEY_PADDED(k) tmp;
- unsigned nr = test_bit(BCH_FS_STARTED, &iter->c->flags)
+ unsigned nr = test_bit(BCH_FS_STARTED, &c->flags)
? (iter->level > 1 ? 0 : 2)
: (iter->level > 1 ? 1 : 16);
bool was_locked = btree_node_locked(iter, iter->level);
break;
bch2_bkey_unpack(l->b, &tmp.k, k);
- bch2_btree_node_prefetch(iter->c, iter, &tmp.k,
- iter->level - 1);
+ bch2_btree_node_prefetch(c, iter, &tmp.k, iter->level - 1);
}
if (!was_locked)
btree_node_unlock(iter, iter->level);
}
-static inline int btree_iter_down(struct btree_iter *iter)
+static noinline void btree_node_mem_ptr_set(struct btree_iter *iter,
+ unsigned plevel, struct btree *b)
+{
+ struct btree_iter_level *l = &iter->l[plevel];
+ bool locked = btree_node_locked(iter, plevel);
+ struct bkey_packed *k;
+ struct bch_btree_ptr_v2 *bp;
+
+ if (!bch2_btree_node_relock(iter, plevel))
+ return;
+
+ k = bch2_btree_node_iter_peek_all(&l->iter, l->b);
+ BUG_ON(k->type != KEY_TYPE_btree_ptr_v2);
+
+ bp = (void *) bkeyp_val(&l->b->format, k);
+ bp->mem_ptr = (unsigned long)b;
+
+ if (!locked)
+ btree_node_unlock(iter, plevel);
+}
+
+static __always_inline int btree_iter_down(struct btree_iter *iter)
{
+ struct bch_fs *c = iter->trans->c;
struct btree_iter_level *l = &iter->l[iter->level];
struct btree *b;
unsigned level = iter->level - 1;
enum six_lock_type lock_type = __btree_lock_want(iter, level);
BKEY_PADDED(k) tmp;
- BUG_ON(!btree_node_locked(iter, iter->level));
+ EBUG_ON(!btree_node_locked(iter, iter->level));
bch2_bkey_unpack(l->b, &tmp.k,
bch2_btree_node_iter_peek(&l->iter, l->b));
- b = bch2_btree_node_get(iter->c, iter, &tmp.k, level, lock_type, true);
+ b = bch2_btree_node_get(c, iter, &tmp.k, level, lock_type);
if (unlikely(IS_ERR(b)))
return PTR_ERR(b);
mark_btree_node_locked(iter, level, lock_type);
btree_iter_node_set(iter, b);
+ if (tmp.k.k.type == KEY_TYPE_btree_ptr_v2 &&
+ unlikely(b != btree_node_mem_ptr(&tmp.k)))
+ btree_node_mem_ptr_set(iter, level + 1, b);
+
if (iter->flags & BTREE_ITER_PREFETCH)
btree_iter_prefetch(iter);
btree_node_unlock(iter, iter->level++);
}
-int __must_check __bch2_btree_iter_traverse(struct btree_iter *);
+static int btree_iter_traverse_one(struct btree_iter *);
-static int btree_iter_traverse_error(struct btree_iter *iter, int ret)
+static int __btree_iter_traverse_all(struct btree_trans *trans,
+ struct btree_iter *orig_iter, int ret)
{
- struct bch_fs *c = iter->c;
- struct btree_iter *linked, *sorted_iters, **i;
-retry_all:
- bch2_btree_iter_unlock(iter);
+ struct bch_fs *c = trans->c;
+ struct btree_iter *iter;
+ u8 sorted[BTREE_ITER_MAX];
+ unsigned i, nr_sorted = 0;
- if (ret != -ENOMEM && ret != -EINTR)
- goto io_error;
+ trans_for_each_iter(trans, iter)
+ sorted[nr_sorted++] = iter - trans->iters;
+
+#define btree_iter_cmp_by_idx(_l, _r) \
+ btree_iter_cmp(&trans->iters[_l], &trans->iters[_r])
+
+ bubble_sort(sorted, nr_sorted, btree_iter_cmp_by_idx);
+#undef btree_iter_cmp_by_idx
+
+retry_all:
+ bch2_trans_unlock(trans);
- if (ret == -ENOMEM) {
+ if (unlikely(ret == -ENOMEM)) {
struct closure cl;
closure_init_stack(&cl);
} while (ret);
}
- /*
- * Linked iters are normally a circular singly linked list - break cycle
- * while we sort them:
- */
- linked = iter->next;
- iter->next = NULL;
- sorted_iters = NULL;
-
- while (linked) {
- iter = linked;
- linked = linked->next;
-
- i = &sorted_iters;
- while (*i && btree_iter_cmp(iter, *i) > 0)
- i = &(*i)->next;
-
- iter->next = *i;
- *i = iter;
+ if (unlikely(ret == -EIO)) {
+ trans->error = true;
+ if (orig_iter) {
+ orig_iter->flags |= BTREE_ITER_ERROR;
+ orig_iter->l[orig_iter->level].b =
+ BTREE_ITER_NO_NODE_ERROR;
+ }
+ goto out;
}
- /* Make list circular again: */
- iter = sorted_iters;
- while (iter->next)
- iter = iter->next;
- iter->next = sorted_iters;
+ BUG_ON(ret && ret != -EINTR);
/* Now, redo traversals in correct order: */
+ for (i = 0; i < nr_sorted; i++) {
+ iter = &trans->iters[sorted[i]];
- iter = sorted_iters;
- do {
-retry:
- ret = __bch2_btree_iter_traverse(iter);
- if (unlikely(ret)) {
- if (ret == -EINTR)
- goto retry;
+ ret = btree_iter_traverse_one(iter);
+ if (ret)
goto retry_all;
- }
-
- iter = iter->next;
- } while (iter != sorted_iters);
+ }
- ret = btree_iter_linked(iter) ? -EINTR : 0;
+ if (hweight64(trans->iters_live) > 1)
+ ret = -EINTR;
+ else
+ trans_for_each_iter(trans, iter)
+ if (iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) {
+ ret = -EINTR;
+ break;
+ }
out:
bch2_btree_cache_cannibalize_unlock(c);
return ret;
-io_error:
- BUG_ON(ret != -EIO);
+}
+
+int bch2_btree_iter_traverse_all(struct btree_trans *trans)
+{
+ return __btree_iter_traverse_all(trans, NULL, 0);
+}
- iter->flags |= BTREE_ITER_ERROR;
- iter->l[iter->level].b = BTREE_ITER_NOT_END;
- goto out;
+static inline bool btree_iter_good_node(struct btree_iter *iter,
+ unsigned l, int check_pos)
+{
+ if (!is_btree_node(iter, l) ||
+ !bch2_btree_node_relock(iter, l))
+ return false;
+
+ if (check_pos <= 0 && btree_iter_pos_before_node(iter, iter->l[l].b))
+ return false;
+ if (check_pos >= 0 && btree_iter_pos_after_node(iter, iter->l[l].b))
+ return false;
+ return true;
}
-static unsigned btree_iter_up_until_locked(struct btree_iter *iter,
- bool check_pos)
+static inline unsigned btree_iter_up_until_good_node(struct btree_iter *iter,
+ int check_pos)
{
unsigned l = iter->level;
while (btree_iter_node(iter, l) &&
- !(is_btree_node(iter, l) &&
- bch2_btree_node_relock(iter, l) &&
- (!check_pos ||
- btree_iter_pos_in_node(iter, iter->l[l].b)))) {
+ !btree_iter_good_node(iter, l, check_pos)) {
btree_node_unlock(iter, l);
- iter->l[l].b = BTREE_ITER_NOT_END;
+ iter->l[l].b = BTREE_ITER_NO_NODE_UP;
l++;
}
* Returns 0 on success, -EIO on error (error reading in a btree node).
*
* On error, caller (peek_node()/peek_key()) must return NULL; the error is
- * stashed in the iterator and returned from bch2_btree_iter_unlock().
+ * stashed in the iterator and returned from bch2_trans_exit().
*/
-int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
+static int btree_iter_traverse_one(struct btree_iter *iter)
{
unsigned depth_want = iter->level;
if (unlikely(iter->level >= BTREE_MAX_DEPTH))
return 0;
- if (__bch2_btree_iter_relock(iter))
+ /*
+ * if we need interior nodes locked, call btree_iter_relock() to make
+ * sure we walk back up enough that we lock them:
+ */
+ if (iter->uptodate == BTREE_ITER_NEED_RELOCK ||
+ iter->locks_want > 1)
+ bch2_btree_iter_relock(iter, false);
+
+ if (iter->uptodate < BTREE_ITER_NEED_RELOCK)
return 0;
/*
* XXX: correctly using BTREE_ITER_UPTODATE should make using check_pos
* here unnecessary
*/
- iter->level = btree_iter_up_until_locked(iter, true);
+ iter->level = btree_iter_up_until_good_node(iter, 0);
/*
* If we've got a btree node locked (i.e. we aren't about to relock the
* 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;
+ BUG_ON(!btree_iter_pos_in_node(iter, iter->l[iter->level].b));
- while ((k = __btree_iter_peek_all(iter, l, &u)).k &&
- !btree_iter_pos_cmp(iter, k.k))
- __btree_iter_advance(l);
+ btree_iter_advance_to_pos(iter, &iter->l[iter->level], -1);
}
/*
? 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;
+ iter->l[iter->level].b = BTREE_ITER_NO_NODE_DOWN;
return ret;
}
}
iter->uptodate = BTREE_ITER_NEED_PEEK;
- bch2_btree_iter_verify_locks(iter);
- __bch2_btree_iter_verify(iter, iter->l[iter->level].b);
+ bch2_btree_iter_verify(iter);
return 0;
}
-int __must_check bch2_btree_iter_traverse(struct btree_iter *iter)
+int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter)
{
int ret;
- ret = __bch2_btree_iter_traverse(iter);
+ ret = bch2_trans_cond_resched(iter->trans) ?:
+ btree_iter_traverse_one(iter);
if (unlikely(ret))
- ret = btree_iter_traverse_error(iter, ret);
-
- BUG_ON(ret == -EINTR && !btree_iter_linked(iter));
+ ret = __btree_iter_traverse_all(iter->trans, iter, ret);
return ret;
}
enum btree_iter_type type)
{
EBUG_ON(iter->btree_id >= BTREE_ID_NR);
- EBUG_ON(!!(iter->flags & BTREE_ITER_IS_EXTENTS) !=
- (iter->btree_id == BTREE_ID_EXTENTS &&
- type != BTREE_ITER_NODES));
+ EBUG_ON(btree_iter_type(iter) != type);
+
+ BUG_ON(type == BTREE_ITER_KEYS &&
+ (bkey_cmp(iter->pos, bkey_start_pos(&iter->k)) < 0 ||
+ bkey_cmp(iter->pos, iter->k.p) > 0));
bch2_btree_iter_verify_locks(iter);
+ bch2_btree_iter_verify_level(iter, iter->level);
}
/* Iterate across nodes (leaf and interior nodes) */
iter->pos = b->key.k.p;
iter->uptodate = BTREE_ITER_UPTODATE;
+ bch2_btree_iter_verify(iter);
+
return b;
}
-struct btree *bch2_btree_iter_next_node(struct btree_iter *iter, unsigned depth)
+struct btree *bch2_btree_iter_next_node(struct btree_iter *iter)
{
struct btree *b;
int ret;
if (!btree_iter_node(iter, iter->level))
return NULL;
+ bch2_trans_cond_resched(iter->trans);
+
btree_iter_up(iter);
if (!bch2_btree_node_relock(iter, iter->level))
iter->pos = iter->btree_id == BTREE_ID_INODES
? btree_type_successor(iter->btree_id, iter->pos)
: bkey_successor(iter->pos);
- iter->level = depth;
+ iter->level = iter->min_depth;
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
ret = bch2_btree_iter_traverse(iter);
iter->pos = b->key.k.p;
iter->uptodate = BTREE_ITER_UPTODATE;
+ bch2_btree_iter_verify(iter);
+
return b;
}
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);
EBUG_ON(!btree_node_locked(iter, 0));
EBUG_ON(bkey_cmp(new_pos, l->b->key.k.p) > 0);
- iter->pos = new_pos;
+ bkey_init(&iter->k);
+ iter->k.p = 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);
}
-void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
+static void btree_iter_pos_changed(struct btree_iter *iter, int cmp)
{
- int cmp = bkey_cmp(new_pos, iter->pos);
- unsigned level;
+ unsigned l = iter->level;
if (!cmp)
- return;
-
- iter->pos = new_pos;
+ goto out;
- 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;
+ l = btree_iter_up_until_good_node(iter, cmp);
+ if (btree_iter_node(iter, l)) {
/*
* 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[l], 8))
+ __btree_iter_init(iter, l);
/* Don't leave it locked if we're not supposed to: */
- if (btree_lock_want(iter, level) == BTREE_NODE_UNLOCKED)
- btree_node_unlock(iter, level);
+ if (btree_lock_want(iter, l) == BTREE_NODE_UNLOCKED)
+ btree_node_unlock(iter, l);
}
-
- if (level != iter->level)
+out:
+ if (l != iter->level)
btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
else
btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
}
-static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
+void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos,
+ bool strictly_greater)
{
- struct btree_iter_level *l = &iter->l[0];
- struct bkey_s_c ret = { .k = &iter->k };
+ struct bpos old = btree_iter_search_key(iter);
+ int cmp;
- if (!bkey_deleted(&iter->k)) {
- EBUG_ON(bch2_btree_node_iter_end(&l->iter));
- ret.v = bkeyp_val(&l->b->format,
- __bch2_btree_node_iter_peek_all(&l->iter, l->b));
- }
+ iter->flags &= ~BTREE_ITER_IS_EXTENTS;
+ iter->flags |= strictly_greater ? BTREE_ITER_IS_EXTENTS : 0;
- if (debug_check_bkeys(iter->c) &&
- !bkey_deleted(ret.k))
- bch2_bkey_debugcheck(iter->c, l->b, ret);
- return ret;
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos = new_pos;
+
+ cmp = bkey_cmp(btree_iter_search_key(iter), old);
+
+ btree_iter_pos_changed(iter, cmp);
}
-struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
+void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
{
- struct btree_iter_level *l = &iter->l[0];
- struct bkey_s_c k;
+ int cmp = bkey_cmp(new_pos, iter->pos);
+
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos = new_pos;
+
+ btree_iter_pos_changed(iter, cmp);
+}
+
+static inline bool btree_iter_set_pos_to_next_leaf(struct btree_iter *iter)
+{
+ struct btree_iter_level *l = &iter->l[0];
+ bool ret;
+
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos = l->b->key.k.p;
+
+ ret = bkey_cmp(iter->pos, POS_MAX) != 0;
+ if (ret)
+ iter->k.p = iter->pos = btree_type_successor(iter->btree_id, iter->pos);
+
+ btree_iter_pos_changed(iter, 1);
+ return ret;
+}
+
+static inline bool btree_iter_set_pos_to_prev_leaf(struct btree_iter *iter)
+{
+ struct btree_iter_level *l = &iter->l[0];
+ bool ret;
+
+ bkey_init(&iter->k);
+ iter->k.p = iter->pos = l->b->data->min_key;
+ iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
+
+ ret = bkey_cmp(iter->pos, POS_MIN) != 0;
+ if (ret)
+ iter->k.p = iter->pos = btree_type_predecessor(iter->btree_id, iter->pos);
+
+ btree_iter_pos_changed(iter, -1);
+ return ret;
+}
+
+/**
+ * btree_iter_peek_uptodate - given an iterator that is uptodate, return the key
+ * it currently points to
+ */
+static inline struct bkey_s_c btree_iter_peek_uptodate(struct btree_iter *iter)
+{
+ struct btree_iter_level *l = &iter->l[0];
+ struct bkey_s_c ret = { .k = &iter->k };
+
+ if (!bkey_deleted(&iter->k)) {
+ struct bkey_packed *_k =
+ __bch2_btree_node_iter_peek_all(&l->iter, l->b);
+
+ ret.v = bkeyp_val(&l->b->format, _k);
+
+ if (debug_check_iterators(iter->trans->c)) {
+ struct bkey k = bkey_unpack_key(l->b, _k);
+
+ BUG_ON(memcmp(&k, &iter->k, sizeof(k)));
+ }
+
+ if (debug_check_bkeys(iter->trans->c))
+ bch2_bkey_debugcheck(iter->trans->c, l->b, ret);
+ }
+
+ return ret;
+}
+
+/**
+ * bch2_btree_iter_peek: returns first key greater than or equal to iterator's
+ * current position
+ */
+struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
+{
+ struct btree_iter_level *l = &iter->l[0];
+ struct bkey_s_c k;
int ret;
bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
- if (iter->uptodate == BTREE_ITER_UPTODATE)
+ if (iter->uptodate == BTREE_ITER_UPTODATE &&
+ !bkey_deleted(&iter->k))
return btree_iter_peek_uptodate(iter);
while (1) {
if (likely(k.k))
break;
- /* got to the end of the leaf, iterator needs to be traversed: */
- iter->pos = l->b->key.k.p;
- iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
-
- if (!bkey_cmp(iter->pos, POS_MAX))
+ if (!btree_iter_set_pos_to_next_leaf(iter))
return bkey_s_c_null;
-
- iter->pos = btree_type_successor(iter->btree_id, iter->pos);
}
/*
iter->pos = bkey_start_pos(k.k);
iter->uptodate = BTREE_ITER_UPTODATE;
+
+ bch2_btree_iter_verify_level(iter, 0);
return k;
}
-static noinline
-struct bkey_s_c bch2_btree_iter_peek_next_leaf(struct btree_iter *iter)
+/**
+ * bch2_btree_iter_next: returns first key greater than iterator's current
+ * position
+ */
+struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
{
- struct btree_iter_level *l = &iter->l[0];
-
- iter->pos = l->b->key.k.p;
- iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
-
- if (!bkey_cmp(iter->pos, POS_MAX))
+ if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
return bkey_s_c_null;
- iter->pos = btree_type_successor(iter->btree_id, iter->pos);
+ bch2_btree_iter_set_pos(iter,
+ btree_type_successor(iter->btree_id, iter->k.p));
return bch2_btree_iter_peek(iter);
}
-struct bkey_s_c bch2_btree_iter_next(struct btree_iter *iter)
+static struct bkey_s_c __btree_trans_updates_peek(struct btree_iter *iter)
+{
+ struct bpos pos = btree_iter_search_key(iter);
+ struct btree_trans *trans = iter->trans;
+ struct btree_insert_entry *i;
+
+ trans_for_each_update2(trans, i)
+ if ((cmp_int(iter->btree_id, i->iter->btree_id) ?:
+ bkey_cmp(pos, i->k->k.p)) <= 0)
+ break;
+
+ return i < trans->updates2 + trans->nr_updates2 &&
+ iter->btree_id == i->iter->btree_id
+ ? bkey_i_to_s_c(i->k)
+ : bkey_s_c_null;
+}
+
+static struct bkey_s_c __bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
{
struct btree_iter_level *l = &iter->l[0];
- struct bkey_packed *p;
+ struct bkey_s_c k = __btree_iter_peek(iter, l);
+ struct bkey_s_c u = __btree_trans_updates_peek(iter);
+
+ if (k.k && (!u.k || bkey_cmp(k.k->p, u.k->p) < 0))
+ return k;
+ if (u.k && bkey_cmp(u.k->p, l->b->key.k.p) <= 0) {
+ iter->k = *u.k;
+ return u;
+ }
+ return bkey_s_c_null;
+}
+
+struct bkey_s_c bch2_btree_iter_peek_with_updates(struct btree_iter *iter)
+{
struct bkey_s_c k;
+ int ret;
bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
- if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
- k = bch2_btree_iter_peek(iter);
- if (IS_ERR_OR_NULL(k.k))
- return k;
- }
+ while (1) {
+ ret = bch2_btree_iter_traverse(iter);
+ if (unlikely(ret))
+ return bkey_s_c_err(ret);
- do {
- __btree_iter_advance(l);
- p = bch2_btree_node_iter_peek_all(&l->iter, l->b);
- if (unlikely(!p))
- return bch2_btree_iter_peek_next_leaf(iter);
- } while (bkey_whiteout(p));
+ k = __bch2_btree_iter_peek_with_updates(iter);
+
+ if (k.k && bkey_deleted(k.k)) {
+ bch2_btree_iter_set_pos(iter,
+ btree_type_successor(iter->btree_id, iter->k.p));
+ continue;
+ }
+
+ if (likely(k.k))
+ break;
+
+ if (!btree_iter_set_pos_to_next_leaf(iter))
+ return bkey_s_c_null;
+ }
- k = __btree_iter_unpack(iter, l, &iter->k, p);
+ /*
+ * iter->pos should always be equal to the key we just
+ * returned - except extents can straddle iter->pos:
+ */
+ if (!(iter->flags & BTREE_ITER_IS_EXTENTS) ||
+ bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0)
+ iter->pos = bkey_start_pos(k.k);
- EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) < 0);
- iter->pos = bkey_start_pos(k.k);
+ iter->uptodate = BTREE_ITER_UPTODATE;
return k;
}
-struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
+struct bkey_s_c bch2_btree_iter_next_with_updates(struct btree_iter *iter)
{
+ if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+ return bkey_s_c_null;
+
+ bch2_btree_iter_set_pos(iter,
+ btree_type_successor(iter->btree_id, iter->k.p));
+
+ return bch2_btree_iter_peek_with_updates(iter);
+}
+
+/**
+ * bch2_btree_iter_peek_prev: returns first key less than or equal to
+ * iterator's current position
+ */
+struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *iter)
+{
+ struct bpos pos = iter->pos;
struct btree_iter_level *l = &iter->l[0];
- struct bkey_packed *p;
struct bkey_s_c k;
int ret;
bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
- if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
- k = bch2_btree_iter_peek(iter);
- if (IS_ERR(k.k))
- return k;
- }
+ if (iter->uptodate == BTREE_ITER_UPTODATE &&
+ !bkey_deleted(&iter->k))
+ return btree_iter_peek_uptodate(iter);
while (1) {
- p = bch2_btree_node_iter_prev(&l->iter, l->b);
- if (likely(p))
- break;
-
- iter->pos = l->b->data->min_key;
- if (!bkey_cmp(iter->pos, POS_MIN))
- return bkey_s_c_null;
-
- bch2_btree_iter_set_pos(iter,
- btree_type_predecessor(iter->btree_id, iter->pos));
-
ret = bch2_btree_iter_traverse(iter);
if (unlikely(ret))
return bkey_s_c_err(ret);
- p = bch2_btree_node_iter_peek(&l->iter, l->b);
- if (p)
- break;
- }
+ k = __btree_iter_peek(iter, l);
+ if (!k.k || bkey_cmp(bkey_start_pos(k.k), pos) > 0)
+ k = __btree_iter_prev(iter, l);
- k = __btree_iter_unpack(iter, l, &iter->k, p);
+ if (likely(k.k))
+ break;
- EBUG_ON(bkey_cmp(bkey_start_pos(k.k), iter->pos) > 0);
+ if (!btree_iter_set_pos_to_prev_leaf(iter))
+ return bkey_s_c_null;
+ }
+ EBUG_ON(bkey_cmp(bkey_start_pos(k.k), pos) > 0);
iter->pos = bkey_start_pos(k.k);
iter->uptodate = BTREE_ITER_UPTODATE;
return k;
}
+/**
+ * bch2_btree_iter_prev: returns first key less than iterator's current
+ * position
+ */
+struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *iter)
+{
+ struct bpos pos = bkey_start_pos(&iter->k);
+
+ bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
+
+ if (unlikely(!bkey_cmp(pos, POS_MIN)))
+ return bkey_s_c_null;
+
+ bch2_btree_iter_set_pos(iter, bkey_predecessor(pos));
+
+ return bch2_btree_iter_peek_prev(iter);
+}
+
static inline struct bkey_s_c
__bch2_btree_iter_peek_slot_extents(struct btree_iter *iter)
{
struct bkey n;
int ret;
-recheck:
- 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);
+ /* keys & 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;
+
+ bch2_btree_iter_set_pos(iter, bkey_successor(iter->pos));
+
+ ret = bch2_btree_iter_traverse(iter);
+ if (unlikely(ret))
+ return bkey_s_c_err(ret);
+ }
/*
* iterator is now at the correct position for inserting at iter->pos,
*/
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
- * 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);
+ k = __btree_iter_unpack(iter, l, &iter->k,
+ bch2_btree_node_iter_peek(&node_iter, l->b));
- goto recheck;
- }
-
- if (k.k &&
- !bkey_whiteout(k.k) &&
- bkey_cmp(bkey_start_pos(k.k), iter->pos) <= 0) {
+ if (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:
+ * We're not setting iter->uptodate because the node iterator
+ * doesn't necessarily point at the key we're returning:
*/
- 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;
+ EBUG_ON(bkey_cmp(k.k->p, iter->pos) <= 0);
+ bch2_btree_iter_verify_level(iter, 0);
return k;
}
/* 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;
: KEY_OFFSET_MAX) -
n.p.offset));
- //EBUG_ON(!n.size);
- if (!n.size) {
- char buf[100];
- bch2_dump_btree_node(iter->l[0].b);
-
- bch2_bkey_to_text(buf, sizeof(buf), k.k);
- panic("iter at %llu:%llu\n"
- "next key %s\n",
- iter->pos.inode,
- iter->pos.offset,
- buf);
- }
+ EBUG_ON(!n.size);
iter->k = n;
iter->uptodate = BTREE_ITER_UPTODATE;
+
+ bch2_btree_iter_verify_level(iter, 0);
return (struct bkey_s_c) { &iter->k, NULL };
}
-static inline struct bkey_s_c
-__bch2_btree_iter_peek_slot(struct btree_iter *iter)
+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;
+ bch2_btree_iter_checks(iter, BTREE_ITER_KEYS);
+
+ if (iter->uptodate == BTREE_ITER_UPTODATE)
+ return btree_iter_peek_uptodate(iter);
+
+ ret = bch2_btree_iter_traverse(iter);
+ if (unlikely(ret))
+ return bkey_s_c_err(ret);
+
if (iter->flags & BTREE_ITER_IS_EXTENTS)
return __bch2_btree_iter_peek_slot_extents(iter);
-recheck:
- while ((k = __btree_iter_peek_all(iter, l, &iter->k)).k &&
- bkey_deleted(k.k) &&
- bkey_cmp(k.k->p, iter->pos) == 0)
- __btree_iter_advance(l);
-
- /*
- * 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);
+ k = __btree_iter_peek_all(iter, l, &iter->k);
- goto recheck;
- }
+ EBUG_ON(k.k && bkey_deleted(k.k) && bkey_cmp(k.k->p, iter->pos) == 0);
- if (k.k &&
- !bkey_deleted(k.k) &&
- !bkey_cmp(iter->pos, k.k->p)) {
- iter->uptodate = BTREE_ITER_UPTODATE;
- return k;
- } else {
+ if (!k.k || bkey_cmp(iter->pos, k.k->p)) {
/* hole */
bkey_init(&iter->k);
iter->k.p = iter->pos;
-
- iter->uptodate = BTREE_ITER_UPTODATE;
- return (struct bkey_s_c) { &iter->k, NULL };
+ k = (struct bkey_s_c) { &iter->k, NULL };
}
-}
-
-struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *iter)
-{
- int ret;
- bch2_btree_iter_checks(iter, BTREE_ITER_SLOTS);
-
- if (iter->uptodate == BTREE_ITER_UPTODATE)
- return btree_iter_peek_uptodate(iter);
-
- ret = bch2_btree_iter_traverse(iter);
- if (unlikely(ret))
- return bkey_s_c_err(ret);
-
- return __bch2_btree_iter_peek_slot(iter);
+ iter->uptodate = BTREE_ITER_UPTODATE;
+ bch2_btree_iter_verify_level(iter, 0);
+ return k;
}
struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *iter)
{
- bch2_btree_iter_checks(iter, BTREE_ITER_SLOTS);
-
- iter->pos = btree_type_successor(iter->btree_id, iter->k.p);
-
- if (unlikely(iter->uptodate != BTREE_ITER_UPTODATE)) {
- /*
- * XXX: when we just need to relock we should be able to avoid
- * calling traverse, but we need to kill BTREE_ITER_NEED_PEEK
- * for that to work
- */
- btree_iter_set_dirty(iter, BTREE_ITER_NEED_TRAVERSE);
-
- return bch2_btree_iter_peek_slot(iter);
- }
-
- if (!bkey_deleted(&iter->k))
- __btree_iter_advance(&iter->l[0]);
+ if (unlikely(!bkey_cmp(iter->k.p, POS_MAX)))
+ return bkey_s_c_null;
- btree_iter_set_dirty(iter, BTREE_ITER_NEED_PEEK);
+ bch2_btree_iter_set_pos(iter,
+ btree_type_successor(iter->btree_id, iter->k.p));
- return __bch2_btree_iter_peek_slot(iter);
+ return bch2_btree_iter_peek_slot(iter);
}
-void __bch2_btree_iter_init(struct btree_iter *iter, struct bch_fs *c,
- enum btree_id btree_id, struct bpos pos,
- unsigned locks_want, unsigned depth,
- unsigned flags)
+static inline void bch2_btree_iter_init(struct btree_trans *trans,
+ struct btree_iter *iter, enum btree_id btree_id,
+ struct bpos pos, unsigned flags)
{
+ struct bch_fs *c = trans->c;
unsigned i;
- EBUG_ON(depth >= BTREE_MAX_DEPTH);
- EBUG_ON(locks_want > BTREE_MAX_DEPTH);
+ if (btree_node_type_is_extents(btree_id) &&
+ !(flags & BTREE_ITER_NODES))
+ flags |= BTREE_ITER_IS_EXTENTS;
- iter->c = c;
+ iter->trans = trans;
iter->pos = pos;
bkey_init(&iter->k);
iter->k.p = pos;
iter->flags = flags;
iter->uptodate = BTREE_ITER_NEED_TRAVERSE;
iter->btree_id = btree_id;
- iter->level = depth;
- iter->locks_want = locks_want;
+ iter->level = 0;
+ iter->min_depth = 0;
+ iter->locks_want = flags & BTREE_ITER_INTENT ? 1 : 0;
iter->nodes_locked = 0;
iter->nodes_intent_locked = 0;
for (i = 0; i < ARRAY_SIZE(iter->l); i++)
- iter->l[i].b = NULL;
- iter->l[iter->level].b = BTREE_ITER_NOT_END;
- iter->next = iter;
+ iter->l[i].b = BTREE_ITER_NO_NODE_INIT;
prefetch(c->btree_roots[btree_id].b);
}
-void bch2_btree_iter_unlink(struct btree_iter *iter)
+/* new transactional stuff: */
+
+static inline void __bch2_trans_iter_free(struct btree_trans *trans,
+ unsigned idx)
{
- struct btree_iter *linked;
+ __bch2_btree_iter_unlock(&trans->iters[idx]);
+ trans->iters_linked &= ~(1ULL << idx);
+ trans->iters_live &= ~(1ULL << idx);
+ trans->iters_touched &= ~(1ULL << idx);
+}
- __bch2_btree_iter_unlock(iter);
+int bch2_trans_iter_put(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ int ret;
- if (!btree_iter_linked(iter))
- return;
+ if (IS_ERR_OR_NULL(iter))
+ return 0;
- for_each_linked_btree_iter(iter, linked)
- if (linked->next == iter) {
- linked->next = iter->next;
- iter->next = iter;
- return;
- }
+ BUG_ON(trans->iters + iter->idx != iter);
- BUG();
-}
+ ret = btree_iter_err(iter);
-void bch2_btree_iter_link(struct btree_iter *iter, struct btree_iter *new)
-{
- BUG_ON(btree_iter_linked(new));
+ if (!(trans->iters_touched & (1ULL << iter->idx)) &&
+ !(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT))
+ __bch2_trans_iter_free(trans, iter->idx);
- new->next = iter->next;
- iter->next = new;
+ trans->iters_live &= ~(1ULL << iter->idx);
+ return ret;
+}
- if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
- unsigned nr_iters = 0;
+int bch2_trans_iter_free(struct btree_trans *trans,
+ struct btree_iter *iter)
+{
+ if (IS_ERR_OR_NULL(iter))
+ return 0;
- for_each_btree_iter(new, iter)
- if (iter->btree_id == new->btree_id)
- nr_iters++;
+ trans->iters_touched &= ~(1ULL << iter->idx);
- BUG_ON(nr_iters > SIX_LOCK_MAX_RECURSE);
- }
+ return bch2_trans_iter_put(trans, iter);
}
-void bch2_btree_iter_copy(struct btree_iter *dst, struct btree_iter *src)
+static int bch2_trans_realloc_iters(struct btree_trans *trans,
+ unsigned new_size)
{
- unsigned i;
+ void *p, *new_iters, *new_updates, *new_updates2;
+ size_t iters_bytes;
+ size_t updates_bytes;
- __bch2_btree_iter_unlock(dst);
- memcpy(dst, src, offsetof(struct btree_iter, next));
+ new_size = roundup_pow_of_two(new_size);
- 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));
-}
+ BUG_ON(new_size > BTREE_ITER_MAX);
-/* new transactional stuff: */
+ if (new_size <= trans->size)
+ return 0;
-static void btree_trans_verify(struct btree_trans *trans)
-{
- unsigned i;
+ BUG_ON(trans->used_mempool);
- for (i = 0; i < trans->nr_iters; i++) {
- struct btree_iter *iter = &trans->iters[i];
+ bch2_trans_unlock(trans);
- BUG_ON(btree_iter_linked(iter) !=
- ((trans->iters_linked & (1 << i)) &&
- !is_power_of_2(trans->iters_linked)));
- }
-}
+ iters_bytes = sizeof(struct btree_iter) * new_size;
+ updates_bytes = sizeof(struct btree_insert_entry) * new_size;
-static inline unsigned btree_trans_iter_idx(struct btree_trans *trans,
- struct btree_iter *iter)
-{
- ssize_t idx = iter - trans->iters;
+ p = kmalloc(iters_bytes +
+ updates_bytes +
+ updates_bytes, GFP_NOFS);
+ if (p)
+ goto success;
- BUG_ON(idx < 0 || idx >= trans->nr_iters);
- BUG_ON(!(trans->iters_live & (1U << idx)));
+ p = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
+ new_size = BTREE_ITER_MAX;
- return idx;
-}
+ trans->used_mempool = true;
+success:
+ new_iters = p; p += iters_bytes;
+ new_updates = p; p += updates_bytes;
+ new_updates2 = p; p += updates_bytes;
-void bch2_trans_iter_put(struct btree_trans *trans,
- struct btree_iter *iter)
-{
- ssize_t idx = btree_trans_iter_idx(trans, iter);
+ memcpy(new_iters, trans->iters,
+ sizeof(struct btree_iter) * trans->nr_iters);
+ memcpy(new_updates, trans->updates,
+ sizeof(struct btree_insert_entry) * trans->nr_updates);
+ memcpy(new_updates2, trans->updates2,
+ sizeof(struct btree_insert_entry) * trans->nr_updates2);
- trans->iters_live &= ~(1U << idx);
-}
+ if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
+ memset(trans->iters, POISON_FREE,
+ sizeof(struct btree_iter) * trans->nr_iters +
+ sizeof(struct btree_insert_entry) * trans->nr_iters);
-void bch2_trans_iter_free(struct btree_trans *trans,
- struct btree_iter *iter)
-{
- ssize_t idx = btree_trans_iter_idx(trans, iter);
+ if (trans->iters != trans->iters_onstack)
+ kfree(trans->iters);
- trans->iters_live &= ~(1U << idx);
- trans->iters_linked &= ~(1U << idx);
- bch2_btree_iter_unlink(iter);
+ trans->iters = new_iters;
+ trans->updates = new_updates;
+ trans->updates2 = new_updates2;
+ trans->size = new_size;
+
+ if (trans->iters_live) {
+ trace_trans_restart_iters_realloced(trans->ip, trans->size);
+ return -EINTR;
+ }
+
+ return 0;
}
-static int btree_trans_realloc_iters(struct btree_trans *trans)
+static struct btree_iter *btree_trans_iter_alloc(struct btree_trans *trans)
{
- struct btree_iter *new_iters;
- unsigned i;
+ unsigned idx = __ffs64(~trans->iters_linked);
- bch2_trans_unlock(trans);
+ if (idx < trans->nr_iters)
+ goto got_slot;
- new_iters = mempool_alloc(&trans->c->btree_iters_pool, GFP_NOFS);
+ if (trans->nr_iters == trans->size) {
+ int ret;
- memcpy(new_iters, trans->iters,
- sizeof(struct btree_iter) * trans->nr_iters);
- trans->iters = new_iters;
+ if (trans->nr_iters >= BTREE_ITER_MAX) {
+ struct btree_iter *iter;
- for (i = 0; i < trans->nr_iters; i++)
- trans->iters[i].next = &trans->iters[i];
+ trans_for_each_iter(trans, iter) {
+ pr_err("iter: btree %s pos %llu:%llu%s%s%s",
+ bch2_btree_ids[iter->btree_id],
+ iter->pos.inode,
+ iter->pos.offset,
+ (trans->iters_live & (1ULL << iter->idx)) ? " live" : "",
+ (trans->iters_touched & (1ULL << iter->idx)) ? " touched" : "",
+ iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT ? " keep" : "");
+ }
- if (trans->iters_linked) {
- unsigned first_linked = __ffs(trans->iters_linked);
+ panic("trans iter oveflow\n");
+ }
- 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]);
+ ret = bch2_trans_realloc_iters(trans, trans->size * 2);
+ if (ret)
+ return ERR_PTR(ret);
}
- btree_trans_verify(trans);
+ idx = trans->nr_iters++;
+ BUG_ON(trans->nr_iters > trans->size);
- if (trans->iters_live) {
- trans_restart();
- return -EINTR;
- }
+ trans->iters[idx].idx = idx;
+got_slot:
+ BUG_ON(trans->iters_linked & (1ULL << idx));
+ trans->iters_linked |= 1ULL << idx;
+ trans->iters[idx].flags = 0;
+ return &trans->iters[idx];
+}
- return 0;
+static inline void btree_iter_copy(struct btree_iter *dst,
+ struct btree_iter *src)
+{
+ unsigned i, idx = dst->idx;
+
+ *dst = *src;
+ dst->idx = idx;
+
+ 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));
+
+ dst->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
+ dst->flags &= ~BTREE_ITER_SET_POS_AFTER_COMMIT;
}
-void bch2_trans_preload_iters(struct btree_trans *trans)
+static inline struct bpos bpos_diff(struct bpos l, struct bpos r)
{
- if (trans->iters == trans->iters_onstack)
- btree_trans_realloc_iters(trans);
+ if (bkey_cmp(l, r) > 0)
+ swap(l, r);
+
+ return POS(r.inode - l.inode, r.offset - l.offset);
}
static struct btree_iter *__btree_trans_get_iter(struct btree_trans *trans,
- unsigned btree_id,
- unsigned flags, u64 iter_id)
+ unsigned btree_id, struct bpos pos,
+ unsigned flags)
{
- struct btree_iter *iter;
- int idx;
+ struct btree_iter *iter, *best = NULL;
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);
- }
+ trans_for_each_iter(trans, iter) {
+ if (btree_iter_type(iter) != (flags & BTREE_ITER_TYPE))
+ continue;
- idx = trans->nr_iters++;
-got_slot:
- trans->iter_ids[idx] = iter_id;
- iter = &trans->iters[idx];
+ if (iter->btree_id != btree_id)
+ continue;
- bch2_btree_iter_init(iter, trans->c, btree_id, POS_MIN, flags);
- } else {
- iter = &trans->iters[idx];
+ if (best &&
+ bkey_cmp(bpos_diff(best->pos, pos),
+ bpos_diff(iter->pos, pos)) < 0)
+ continue;
- iter->flags &= ~(BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
- iter->flags |= flags & (BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
+ best = iter;
}
- BUG_ON(trans->iters_live & (1 << idx));
- trans->iters_live |= 1 << idx;
+ if (!best) {
+ iter = btree_trans_iter_alloc(trans);
+ if (IS_ERR(iter))
+ return iter;
- if (trans->iters_linked &&
- !(trans->iters_linked & (1 << idx)))
- bch2_btree_iter_link(&trans->iters[__ffs(trans->iters_linked)],
- iter);
+ bch2_btree_iter_init(trans, iter, btree_id, pos, flags);
+ } else if ((trans->iters_live & (1ULL << best->idx)) ||
+ (best->flags & BTREE_ITER_KEEP_UNTIL_COMMIT)) {
+ iter = btree_trans_iter_alloc(trans);
+ if (IS_ERR(iter))
+ return iter;
- trans->iters_linked |= 1 << idx;
+ btree_iter_copy(iter, best);
+ } else {
+ iter = best;
+ }
- btree_trans_verify(trans);
+ iter->flags &= ~(BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
+ iter->flags |= flags & (BTREE_ITER_SLOTS|BTREE_ITER_INTENT|BTREE_ITER_PREFETCH);
+
+ if (iter->flags & BTREE_ITER_INTENT)
+ bch2_btree_iter_upgrade(iter, 1);
+ else
+ bch2_btree_iter_downgrade(iter);
BUG_ON(iter->btree_id != btree_id);
BUG_ON((iter->flags ^ flags) & BTREE_ITER_TYPE);
+ BUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
+ BUG_ON(iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT);
+ BUG_ON(trans->iters_live & (1ULL << iter->idx));
+
+ trans->iters_live |= 1ULL << iter->idx;
+ trans->iters_touched |= 1ULL << iter->idx;
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 *bch2_trans_get_iter(struct btree_trans *trans,
+ enum btree_id btree_id,
+ struct bpos pos, unsigned flags)
{
struct btree_iter *iter =
- __btree_trans_get_iter(trans, btree_id, flags, iter_id);
+ __btree_trans_get_iter(trans, btree_id, pos, flags);
if (!IS_ERR(iter))
- bch2_btree_iter_set_pos(iter, pos);
+ __bch2_btree_iter_set_pos(iter, pos,
+ btree_node_type_is_extents(btree_id));
return iter;
}
-struct btree_iter *__bch2_trans_copy_iter(struct btree_trans *trans,
- struct btree_iter *src,
- u64 iter_id)
+struct btree_iter *bch2_trans_get_node_iter(struct btree_trans *trans,
+ enum btree_id btree_id,
+ struct bpos pos,
+ unsigned locks_want,
+ unsigned depth,
+ unsigned flags)
{
struct btree_iter *iter =
- __btree_trans_get_iter(trans, src->btree_id,
- src->flags, iter_id);
+ __btree_trans_get_iter(trans, btree_id, pos,
+ flags|BTREE_ITER_NODES);
+ unsigned i;
+
+ BUG_ON(IS_ERR(iter));
+ BUG_ON(bkey_cmp(iter->pos, pos));
+
+ iter->locks_want = locks_want;
+ iter->level = depth;
+ iter->min_depth = depth;
+
+ for (i = 0; i < ARRAY_SIZE(iter->l); i++)
+ iter->l[i].b = NULL;
+ iter->l[iter->level].b = BTREE_ITER_NO_NODE_INIT;
- if (!IS_ERR(iter))
- bch2_btree_iter_copy(iter, src);
return iter;
}
-void *bch2_trans_kmalloc(struct btree_trans *trans,
- size_t size)
+struct btree_iter *bch2_trans_copy_iter(struct btree_trans *trans,
+ struct btree_iter *src)
{
- void *ret;
+ struct btree_iter *iter;
+
+ iter = btree_trans_iter_alloc(trans);
+ if (IS_ERR(iter))
+ return iter;
+
+ btree_iter_copy(iter, src);
+
+ trans->iters_live |= 1ULL << iter->idx;
+ /*
+ * We don't need to preserve this iter since it's cheap to copy it
+ * again - this will cause trans_iter_put() to free it right away:
+ */
+ trans->iters_touched &= ~(1ULL << iter->idx);
- if (trans->mem_top + size > trans->mem_bytes) {
+ return iter;
+}
+
+static int bch2_trans_preload_mem(struct btree_trans *trans, size_t size)
+{
+ if (size > trans->mem_bytes) {
size_t old_bytes = trans->mem_bytes;
- size_t new_bytes = roundup_pow_of_two(trans->mem_top + size);
+ size_t new_bytes = roundup_pow_of_two(size);
void *new_mem = krealloc(trans->mem, new_bytes, GFP_NOFS);
if (!new_mem)
- return ERR_PTR(-ENOMEM);
+ return -ENOMEM;
trans->mem = new_mem;
trans->mem_bytes = new_bytes;
if (old_bytes) {
- trans_restart();
- return ERR_PTR(-EINTR);
+ trace_trans_restart_mem_realloced(trans->ip, new_bytes);
+ return -EINTR;
}
}
- ret = trans->mem + trans->mem_top;
+ return 0;
+}
+
+void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
+{
+ void *p;
+ int ret;
+
+ ret = bch2_trans_preload_mem(trans, trans->mem_top + size);
+ if (ret)
+ return ERR_PTR(ret);
+
+ p = trans->mem + trans->mem_top;
trans->mem_top += size;
- return ret;
+ return p;
}
-int bch2_trans_unlock(struct btree_trans *trans)
+inline void bch2_trans_unlink_iters(struct btree_trans *trans)
{
- unsigned iters = trans->iters_linked;
- int ret = 0;
+ u64 iters = trans->iters_linked &
+ ~trans->iters_touched &
+ ~trans->iters_live;
while (iters) {
- unsigned idx = __ffs(iters);
- struct btree_iter *iter = &trans->iters[idx];
+ unsigned idx = __ffs64(iters);
- if (iter->flags & BTREE_ITER_ERROR)
- ret = -EIO;
-
- __bch2_btree_iter_unlock(iter);
- iters ^= 1 << idx;
+ iters &= ~(1ULL << idx);
+ __bch2_trans_iter_free(trans, idx);
}
-
- return ret;
}
-void __bch2_trans_begin(struct btree_trans *trans)
+void bch2_trans_reset(struct btree_trans *trans, unsigned flags)
{
- unsigned idx;
+ struct btree_iter *iter;
- btree_trans_verify(trans);
+ trans_for_each_iter(trans, iter)
+ iter->flags &= ~(BTREE_ITER_KEEP_UNTIL_COMMIT|
+ BTREE_ITER_SET_POS_AFTER_COMMIT);
- /*
- * 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]);
- }
+ bch2_trans_unlink_iters(trans);
+
+ trans->iters_touched &= trans->iters_live;
+
+ trans->need_reset = 0;
+ trans->nr_updates = 0;
+ trans->nr_updates2 = 0;
+ trans->mem_top = 0;
- trans->iters_live = 0;
- trans->nr_updates = 0;
- trans->mem_top = 0;
+ if (trans->fs_usage_deltas) {
+ trans->fs_usage_deltas->used = 0;
+ memset(&trans->fs_usage_deltas->memset_start, 0,
+ (void *) &trans->fs_usage_deltas->memset_end -
+ (void *) &trans->fs_usage_deltas->memset_start);
+ }
- btree_trans_verify(trans);
+ if (!(flags & TRANS_RESET_NOTRAVERSE))
+ bch2_btree_iter_traverse_all(trans);
}
-void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c)
+void bch2_trans_init(struct btree_trans *trans, struct bch_fs *c,
+ unsigned expected_nr_iters,
+ size_t expected_mem_bytes)
{
+ memset(trans, 0, offsetof(struct btree_trans, iters_onstack));
+
+ /*
+ * reallocating iterators currently completely breaks
+ * bch2_trans_iter_put():
+ */
+ expected_nr_iters = BTREE_ITER_MAX;
+
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->ip = _RET_IP_;
+ trans->size = ARRAY_SIZE(trans->iters_onstack);
trans->iters = trans->iters_onstack;
+ trans->updates = trans->updates_onstack;
+ trans->updates2 = trans->updates2_onstack;
+ trans->fs_usage_deltas = NULL;
+
+ if (expected_nr_iters > trans->size)
+ bch2_trans_realloc_iters(trans, expected_nr_iters);
+
+ if (expected_mem_bytes)
+ bch2_trans_preload_mem(trans, expected_mem_bytes);
}
int bch2_trans_exit(struct btree_trans *trans)
{
- int ret = bch2_trans_unlock(trans);
+ bch2_trans_unlock(trans);
+ kfree(trans->fs_usage_deltas);
kfree(trans->mem);
- if (trans->iters != trans->iters_onstack)
+ if (trans->used_mempool)
mempool_free(trans->iters, &trans->c->btree_iters_pool);
+ else if (trans->iters != trans->iters_onstack)
+ kfree(trans->iters);
trans->mem = (void *) 0x1;
trans->iters = (void *) 0x1;
- return ret;
+
+ return trans->error ? -EIO : 0;
+}
+
+void bch2_fs_btree_iter_exit(struct bch_fs *c)
+{
+ mempool_exit(&c->btree_iters_pool);
+}
+
+int bch2_fs_btree_iter_init(struct bch_fs *c)
+{
+ unsigned nr = BTREE_ITER_MAX;
+
+ return mempool_init_kmalloc_pool(&c->btree_iters_pool, 1,
+ sizeof(struct btree_iter) * nr +
+ sizeof(struct btree_insert_entry) * nr +
+ sizeof(struct btree_insert_entry) * nr);
}
projects / bcachefs-tools-debian / blobdiff