Horton

Re: pageinspect: Hash index support

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
Cc: Jesper Pedersen <jesper.pedersen@redhat.com>, Michael Paquier <michael.paquier@gmail.com>, Ashutosh Sharma <ashu.coek88@gmail.com>, Jeff Janes <jeff.janes@gmail.com>, Alvaro Herrera <alvherre@2ndquadrant.com>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2016-11-02T17:54:47Z
Lists: pgsql-hackers 
Peter Eisentraut <peter.eisentraut@2ndquadrant.com> writes:
> So, we're down to crashes in gin_metapage_info() on ia64 and sparc64.
> My guess is that the raw page data that is passed into the function
> needs to be 8-byte aligned before being accessed as GinMetaPageData.

That's what it looks like to me, too.  The "bytea" page image is
guaranteed to be improperly aligned for 64-bit access, since it will have
an int32 length word before the actual page data, breaking the alignment
that would exist for a page sitting in a page buffer.  This is likely to
be a problem for more things than just gin_metapage_info(); sooner or
later it could affect just about everything in pageinspect.

> (Maybe GinPageGetMeta() should do it?)

I think the right thing is likely to be to copy the presented bytea
into a palloc'd (and therefore properly aligned) buffer.  And not
just in this one function.

			regards, tom lane

Commits

  1. Fix incorrect typecast.

  2. In pageinspect/hashfuncs.c, avoid crashes on alignment-picky machines.

  3. pageinspect: Support hash indexes.

  4. Add uuid to the set of types supported by contrib/btree_gist.

  5. Allow CREATE EXTENSION to follow extension update paths.