Ceph » rgw

Currently we use shadow objects for every write that overwrites an object. We can avoid that by assuming that objects that are smaller or equal to the read chunk size are inherently atomic.
Things to keep in mind:
- read of object data might be atomic, but we currently first read object xattrs, and only then do we read the data which is not atomic. We should either still use barriers and handle this case, or we should read everything atomically in the first place as a single compound op.
- maybe doing the latter, but adding a method to cls_rgw that verifies ACLs, so that we only read the object if ACLs are ok?