Re: Setting pd_lower in GIN metapage

Amit Langote <langote_amit_f8@lab.ntt.co.jp>

From: Amit Langote <Langote_Amit_f8@lab.ntt.co.jp>
To: Tom Lane <tgl@sss.pgh.pa.us>
Cc: Amit Kapila <amit.kapila16@gmail.com>, Michael Paquier <michael.paquier@gmail.com>, Masahiko Sawada <sawada.mshk@gmail.com>, Pg Hackers <pgsql-hackers@postgresql.org>
Date: 2017-09-13T05:48:12Z
Lists: pgsql-hackers

Attachments

On 2017/09/13 13:05, Tom Lane wrote:
> Amit Langote <Langote_Amit_f8@lab.ntt.co.jp> writes:
>> On 2017/09/12 23:27, Amit Kapila wrote:
>>> I think one point which might be missed is that the patch needs to
>>> modify pd_lower for all usages of metapage, not only when it is first
>>> time initialized.
> 
>> Maybe I'm missing something, but isn't the metadata size fixed and hence
>> pd_lower won't change once it's initialized?  Maybe, it's not true for all
>> index types?
> 
> No, the point is that you might be dealing with an index recently
> pg_upgraded from v10 or before, which does not have the correct
> value for pd_lower on that page.  This has to be coped with.

Ah, got it.  Thanks for the explanation.

I updated the patches so that the metapage's pd_lower is set to the
correct value just before *every* point where we are about to insert a
full page image of the metapage into WAL.  That's in addition to doing the
same in various metapage init routines, which the original patch did
already anyway.  I guess this now ensures that wal_consistency_checking
masking of these metapages as standard layout pages always works, even for
pre-v11 indexes that were upgraded.

Also, we now pass the metapage buffer as containing a page of standard
layout to XLogRegisterBuffer(), so that any hole in it is compressed when
actually writing to WAL.

Thanks,
Amit

Commits

  1. Flag index metapages as standard-format in xlog.c calls.

  2. Set the metapage's pd_lower correctly in brin, gin, and spgist indexes.