To do so, we allow callers of _flush(Inode) to pass in a Context
as well. This Context is then given to the ObjectCacher as
the completion callback instead of the C_Client_PutInode that
it uses by default. _fsync() passes in a SafeCond which is
pointing to its return code, and presto! The return code gets
a failure if the Objecter returns any.
To preserve existing semantics, _fsync() (and any similar
caller) is now required to take a reference to the inode.
However, this is a behavioral change: previously, a call to fsync
in the userspace code would simply spin as it waited for all the
data to flush out. That obviously wasn't desirable, but now a user
can call fsync, ignore the return value, and think that everything is
hunky-dory when in fact the client is repeatedly trying to flush out
the dirty data. Hrm...