Fix pg_upgrade to cope with non-default unix_socket_directory scenarios.

When starting either an old or new postmaster, force it to place its Unix
socket in the current directory.  This makes it even harder for accidental
connections to occur during pg_upgrade, and also works around some
scenarios where the default socket location isn't usable.  (For example,
if the default location is something other than "/tmp", it might not exist
during "make check".)

When checking an already-running old postmaster, find out its actual socket
directory location from postmaster.pid, if possible.  This dodges problems
with an old postmaster having a configured location different from the
default built into pg_upgrade's libpq.  We can't find that out if the old
postmaster is pre-9.1, so also document how to cope with such scenarios
manually.

In support of this, centralize handling of the connection-related command
line options passed to pg_upgrade's subsidiary programs, such as pg_dump.
This should make future changes easier.

Bruce Momjian and Tom Lane

1 files changed, 8 insertions, 0 deletions

diff --git a/doc/src/sgml/pgupgrade.sgml b/doc/src/sgml/pgupgrade.sgml
index b62aba2641..9e43f3ce13 100644
--- a/doc/src/sgml/pgupgrade.sgml
+++ b/doc/src/sgml/pgupgrade.sgml
@@ -520,6 +520,14 @@ psql --username postgres --file script.sql postgres
   </para>
 
   <para>
+   If doing <option>--check</> with a running old server of a pre-9.1 version,
+   and the old server is using a Unix-domain socket directory that is
+   different from the default built into the new <productname>PostgreSQL</>
+   installation, set <envar>PGHOST</> to point to the socket location of the
+   old server.  (This is not relevant on Windows.)
+  </para>
+
+  <para>
    A Log-Shipping Standby Server (<xref linkend="warm-standby">) cannot
    be upgraded because the server must allow writes.  The simplest way
    is to upgrade the primary and use rsync to rebuild the standbys.