On Thu, Oct 27, 2016 at 01:22:49PM +0300, Sagi Grimberg wrote:

Christoph, did you manage to leap to the future and solve the
RDMA persistency hole? :)

e.g. what happens with O_DSYNC in this model? Or you did
a message exchange for commits?

Yes, pNFS calls this the layoutcommit.  That being said once we get a RDMA
commit or flush operation we could easily make the layoutcommit optional
for some operations.  There already is a precedence for the in the
flexfiles layout specification.