On 04/06/2018 07:46 PM, Andres Freund wrote:
> On 2018-04-06 19:40:59 +0200, Tomas Vondra wrote:
>> In any case, I wouldn't call LockBufHdr/UnlockBufHdr a "side channel"
>> interlock. It's a pretty direct and intentional interlock, I think.
>
> I mean it's a side-channel as far as DataChecksumsNeedWrite() is
> concerned. You're banking on all callers using a barrier implying
> operation around it.
Ah, OK.
>
>> Sure. But what would that be? I can't think of anything. A process that
>> modifies a buffer (or any other piece of shared state) without holding
>> some sort of lock seems broken by default.
>
> You can quite possibly already *hold* a lock if it's not an exclusive
> one.
>
Sure, but if you're holding the buffer lock when the checksum version is
changed, then the checksumhelper is obviously not running yet. In which
case it will update the checksum on the buffer later.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services