Previously, we passed the toplevel PlannerInfo, but we actually want
to pass the relevant subroot.  One problem with passing the toplevel
PlannerInfo is that the FDW which wants to push down an UPDATE or
DELETE against a join won't find the relevant joinrel there.
As of commit 1bc0100d, postgres_fdw
tries to do exactly this and can be made to fail an assertion as a
result.

It's possible that this should be regarded as a bug fix and
back-patched to earlier releases, but for lack of a test case that
fails in earlier releases, no back-patch for now.

Etsuro Fujita, reviewed by Amit Langote.

Discussion: http://postgr.es/m/5AF43E02.30000@lab.ntt.co.jp