Re: Handling GIN incomplete splits

On Wed, Nov 13, 2013 at 8:49 AM, Heikki Linnakangas <hlinnakangas(at)vmware(dot)com
> wrote:

> Here's another part of my crusade against xlog cleanup routines. This
> series of patches gets rid of the gin_cleanup() function, which is
> currently used to finish splits of GIN b-tree pages, if the system crashes
> (or an error occurs) between splitting a page and inserting its downlink to
> the parent.
>
> The first three patches just move code around. IMHO they make the code
> more readable, so they should be committed in any case. The meat is in the
> fourth patch.
>
> Thoughts, objections?
>
> Alexander, I'm sorry if this conflicts with your GIN patches. Feel free to
> post the latest versions of your patches against the current master,
> ignoring patches. I can fix the bitrot. That said, I think these
> refactorings will make your code look a little bit nicer too, so you might
> want to rebase because of that anyway.

Hi Heikki,

The commit 04eee1fa9ee80dabf7 of this series causes a self-deadlock in the
LWLock code during the operation below, with it trying to take
an LW_EXCLUSIVE on a high, even-numbered lockid when it already holds the
same lockid.

CREATE INDEX planet_osm_ways_nodes ON planet_osm_ways USING gin (nodes)
WITH (FASTUPDATE=OFF);

It happens pretty reliably using osm2pgsql.

I will try to come up with a simple reproducible demonstration, and stack
trace, over the weekend.

Cheers,

Jeff