Btrfs: run chunk allocations while we do delayed refs
Btrfs tries to batch extent allocation tree changes to improve performance and reduce metadata trashing. But it doesn't allocate new metadata chunks while it is doing allocations for the extent allocation tree. This commit changes the delayed refence code to do chunk allocations if we're getting low on room. It prevents crashes and improves performance. Signed-off-by: Chris Mason <chris.mason@oracle.com>
This commit is contained in:
@@ -467,19 +467,12 @@ static int __btrfs_end_transaction(struct btrfs_trans_handle *trans,
|
||||
|
||||
btrfs_trans_release_metadata(trans, root);
|
||||
trans->block_rsv = NULL;
|
||||
while (count < 4) {
|
||||
while (count < 2) {
|
||||
unsigned long cur = trans->delayed_ref_updates;
|
||||
trans->delayed_ref_updates = 0;
|
||||
if (cur &&
|
||||
trans->transaction->delayed_refs.num_heads_ready > 64) {
|
||||
trans->delayed_ref_updates = 0;
|
||||
|
||||
/*
|
||||
* do a full flush if the transaction is trying
|
||||
* to close
|
||||
*/
|
||||
if (trans->transaction->delayed_refs.flushing)
|
||||
cur = 0;
|
||||
btrfs_run_delayed_refs(trans, root, cur);
|
||||
} else {
|
||||
break;
|
||||
|
Reference in New Issue
Block a user