Log the creation of an init fork unconditionally.
authorRobert Haas <rhaas@postgresql.org>
Thu, 8 Dec 2016 19:09:09 +0000 (14:09 -0500)
committerRobert Haas <rhaas@postgresql.org>
Thu, 8 Dec 2016 19:19:25 +0000 (14:19 -0500)
commita00ac62991fc9cce309e40979ca74ad54bf97b15
tree4f4f0f7d83f810e75861ae87f5da3b98dff99517
parent311bc147ff0924b651f36910925a163c4f3b4a6c
Log the creation of an init fork unconditionally.

Previously, it was thought that this only needed to be done for the
benefit of possible standbys, so wal_level = minimal skipped it.
But that's not safe, because during crash recovery we might replay
XLOG_DBASE_CREATE or XLOG_TBLSPC_CREATE record which recursively
removes the directory that contains the new init fork.  So log it
always.

The user-visible effect of this bug is that if you create a database
or tablespace, then create an unlogged table, then crash without
checkpointing, then restart, accessing the table will fail, because
the it won't have been properly reset.  This commit fixes that.

Michael Paquier, per a report from Konstantin Knizhnik.  Wording of
the comments per a suggestion from me.
src/backend/access/nbtree/nbtree.c
src/backend/access/spgist/spginsert.c
src/backend/catalog/heap.c