OSP objects are not removed immediately, so ll_d_iput() in unlink path finds extent locks and do not reset nlink to 0 leaving the inode in the cache.

we need to figure out another way to purge inode...

on MDS corresponding inode does get nlink=0 only if the inode is actually being destroyed (orphans get nlink=1 being on PENDING/). this zero nlink propagated to the client can be a signal for ELC on the client ?