@@ -543,6 +543,10 @@ update_existing_head_ref(struct btrfs_delayed_ref_node *existing,
* update the reference mod on the head to reflect this new operation
*/
existing->ref_mod += update->ref_mod;
+
+ WARN_ON_ONCE(update->ref_mod > 1);
+ if (update->ref_mod == 1)
+ existing_ref->add_cnt++;
}
/*
@@ -604,6 +608,12 @@ static noinline void add_delayed_ref_head(struct btrfs_fs_info *fs_info,
head_ref->must_insert_reserved = must_insert_reserved;
head_ref->is_data = is_data;
+ /* track added ref, more comments in select_delayed_ref() */
+ if (count_mod == 1)
+ head_ref->add_cnt = 1;
+ else
+ head_ref->add_cnt = 0;
+
INIT_LIST_HEAD(&head_ref->cluster);
mutex_init(&head_ref->mutex);
@@ -84,6 +84,9 @@ struct btrfs_delayed_ref_head {
struct list_head cluster;
struct btrfs_delayed_extent_op *extent_op;
+
+ int add_cnt;
+
/*
* when a new extent is allocated, it is just reserved in memory
* The actual extent isn't inserted into the extent allocation tree
@@ -2287,6 +2287,16 @@ select_delayed_ref(struct btrfs_delayed_ref_head *head)
struct rb_node *node;
struct btrfs_delayed_ref_node *ref;
int action = BTRFS_ADD_DELAYED_REF;
+
+ /*
+ * track the count of BTRFS_ADD_DELAYED_REF,
+ * in the case that there's no BTRFS_ADD_DELAYED_REF while there're a
+ * a great number of BTRFS_DROP_DELAYED_REF,
+ * it'll waste time on searching BTRFS_ADD_DELAYED_REF, usually this
+ * happens with dedup enabled.
+ */
+ if (head->add_cnt == 0)
+ action = BTRFS_DROP_DELAYED_REF;
again:
/*
* select delayed ref of type BTRFS_ADD_DELAYED_REF first.
@@ -2301,8 +2311,11 @@ again:
rb_node);
if (ref->bytenr != head->node.bytenr)
break;
- if (ref->action == action)
+ if (ref->action == action) {
+ if (action == BTRFS_ADD_DELAYED_REF)
+ head->add_cnt--;
return ref;
+ }
node = rb_prev(node);
}
if (action == BTRFS_ADD_DELAYED_REF) {
@@ -2378,6 +2391,8 @@ static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
* there are still refs with lower seq numbers in the
* process of being added. Don't run this ref yet.
*/
+ if (ref->action == BTRFS_ADD_DELAYED_REF)
+ locked_ref->add_cnt++;
list_del_init(&locked_ref->cluster);
btrfs_delayed_ref_unlock(locked_ref);
locked_ref = NULL;
While removing a file with dedup extents, we could have a great number of delayed refs pending to process, and these refs refer to droping a ref of the extent, which is of BTRFS_DROP_DELAYED_REF type. But in order to prevent an extent's ref count from going down to zero when there still are pending delayed refs, we first select those "adding a ref" ones, which is of BTRFS_ADD_DELAYED_REF type. So in removing case, all of our delayed refs are of BTRFS_DROP_DELAYED_REF type, but we have to walk all the refs issued to the extent to find any BTRFS_ADD_DELAYED_REF types and end up there is no such thing, and then start over again to find BTRFS_DROP_DELAYED_REF. This is really unnecessary, we can improve this by tracking how many BTRFS_ADD_DELAYED_REF refs we have and search by the right type. Signed-off-by: Liu Bo <bo.li.liu@oracle.com> --- fs/btrfs/delayed-ref.c | 10 ++++++++++ fs/btrfs/delayed-ref.h | 3 +++ fs/btrfs/extent-tree.c | 17 ++++++++++++++++- 3 files changed, 29 insertions(+), 1 deletion(-)