currently fail_abort() in t-f ensures that all the clients are back, using lfs df: the first df ensures that the clients recognize its eviction, initiates reconnect and the second one verifies the clients are really back.
this doesn't work with MDT and the consequence is that such an evicted state is deteted by another MDT(s) at random point, results in EIO and breaks testing in constructions like mkdir $DIR/$tdir || error ...
IMO, fail_abort() should do something similar for MDTs.