Re: MarkBufferDirtyHint() and LSN update

Kyotaro Horiguchi <horikyota(dot)ntt(at)gmail(dot)com> wrote:

> At Mon, 11 Nov 2019 10:03:14 +0100, Antonin Houska <ah(at)cybertec(dot)at> wrote in
> > Michael Paquier <michael(at)paquier(dot)xyz> wrote:
> > > Does something like the attached patch make sense? Reviews are
> > > welcome.
> >
> > This looks good to me.
>
> I have a qustion.
>
> The current code assumes that !BM_DIRTY means that the function is
> dirtying the page. But if !BM_JUST_DIRTIED, the function actually is
> going to re-dirty the page even if BM_DIRTY.

It makes sense to me. I can imagine the following:

1. FlushBuffer() cleared BM_JUST_DIRTIED, wrote the page to disk but hasn't
yet cleared BM_DIRTY.

2. Another backend changed a hint bit in shared memory and called
MarkBufferDirtyHint().

Thus FlushBuffer() missed the current hint bit change, so we need to dirty the
page again.

> If this is correct, the trigger for stats update is not !BM_DIRTY but
> !BM_JUST_DIRTIED, or the fact that we passed the line of
> XLogSaveBufferForHint() could be the trigger, regardless whether the
> LSN is valid or not.

I agree.

--
Antonin Houska
Web: https://www.cybertec-postgresql.com