Re: [HACKERS] Setting pd_lower in GIN metapage

On 2017/06/19 23:31, Tom Lane wrote:
> Amit Kapila <amit.kapila16@gmail.com> writes:
>> On Mon, Jun 19, 2017 at 11:37 AM, Amit Langote
>> <Langote_Amit_f8@lab.ntt.co.jp> wrote:
>>> What are some arguments against setting pd_lower in the GIN metapage as
>>> follows?
> 
>> Actually, hash index also has similar code (See _hash_init_metabuffer)
>> and I see no harm in doing this at similar other places.
> 
> Seems reasonable.

Here is a patch that does it for the GIN metapage.  (I am not sure if the
changes to gin_mask() that are included in the patch are really necessary.)

>>> How about porting such a change to the back-branches if we do this at all?
>>> The reason I'm asking is that a certain backup tool relies on pd_lower
>>> values of data pages (disk blocks in relation files that are known to have
>>> a valid PageHeaderData) to be correct to discard the portion of every page
>>> that supposedly does not contain any useful information.  The assumption
>>> doesn't hold in the case of GIN metapage, so any GIN indexes contain
>>> corrupted metapage after recovery (metadata overwritten with zeros).
> 
> I'm not in favor of back-porting such a change.  Even if we did, it would
> only affect subsequently-created indexes not existing ones.  That means
> your tool has to cope with an unset pd_lower in any case --- and will for
> the foreseeable future, because of pg_upgrade.
> 
> I'd suggest a rule like "if pd_lower is smaller than SizeOfPageHeaderData
> then don't trust it, but assume all of the page is valid data".

Actually, such a check is already in place in the tool, whose condition
looks like:

    if (PageGetPageSize(header) == BLCKSZ &&
        PageGetPageLayoutVersion(header) == PG_PAGE_LAYOUT_VERSION &&
        (header->pd_flags & ~PD_VALID_FLAG_BITS) == 0 &&
        header->pd_lower >= SizeOfPageHeaderData &&
        header->pd_lower <= header->pd_upper &&
        header->pd_upper <= header->pd_special &&
        header->pd_special <= BLCKSZ &&
        header->pd_special == MAXALIGN(header->pd_special) && ...

which even GIN metapage passes, making it an eligible data page and hence
for omitting the hole between pd_lower and pd_upper.

That's because a GIN metapage will always have undergone PageInit() that
sets pd_lower to SizeOfPageHeaderData.  Which means the tool has to look
beyond the standard PageHeaderData to determine whether the area between
pd_lower and pd_upper is really a hole.  Amit K also suggested the same,
but that seems to require either duplicating GIN's private struct
definition (of GinMetaPageData) in the tool or including backend's
gin_private.h, either of which doesn't seem to be a good thing to do in
what is FRONTEND code, but maybe there is no other way.  Am I missing
something?

Thanks,
Amit

-- 
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers