Re: Latent cache flush hazard in RelationInitIndexAccessInfo

Michael Paquier <michael(dot)paquier(at)gmail(dot)com> writes:
> On Fri, Sep 11, 2015 at 11:29 AM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> On Wed, Sep 9, 2015 at 3:00 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>>> The specific reason there's a problem is that there's a disconnect between
>>> RelationClearRelation's test for whether a relcache entry has valid index
>>> info (it checks whether rd_indexcxt != NULL) and the order of operations
>>> in RelationInitIndexAccessInfo (creating the index context is not the
>>> first thing it does). Given that if RelationClearRelation thinks the
>>> info is valid, what it does is call RelationReloadIndexInfo which needs
>>> rd_index to be valid, I'm thinking the best fix is to leave the order of
>>> operations in RelationInitIndexAccessInfo alone and instead make the
>>> "relcache entry has index info" check be "rd_index != NULL". There might
>>> need to be some additional work to keep RelationReloadIndexInfo from
>>> setting rd_isvalid = true too soon.

>> Doesn't seem like a bad thing to clean up.

> Doesn't sound bad to me either to reinforce things on HEAD... It's
> been months since this has been posted, are you working on a patch?

This dropped off my radar screen somehow. (I have a feeling I fixed it in
Salesforce's code and forgot to transfer the fix to community.)

Given that this is just latent for the community's purposes, it seems
probably inappropriate to fix it post-beta1. I'm inclined to let it
slide till 9.7^H^H^H10 development opens. Thoughts?

regards, tom lane