Tom Lane wrote:
> Heikki Linnakangas <heikki@enterprisedb.com> writes:
>> ... MarkBufferDirty needs to be called
>> before XLogInsert to avoid a race condition in checkpoint, see comments
>> in SyncOneBuffer in bufmgr.c for an explanation.
>
> Right, see also the "Write-Ahead Log coding" section in
> src/backend/access/transam/README (which is maybe not a very good place
> for it, but it doesn't seem like bufmgr's turf either).
Yeah it could be documented more visibly, I didn't know about (or didn't
remember) that rule until I saw that comment today. I found that issue
in GIN by just quickly grepping for callers of MarkBufferDirty.
How about adding an Assert to XLogInsert to check that all buffers given
to it are already marked as dirty? It wouldn't be completely
water-tight, sometimes we don't pass the buffer to XLogInsert even
though we stamp the LSN, but it would catch most cases.
-- Heikki Linnakangas EnterpriseDB http://www.enterprisedb.com