Horton

Re: straightening out backend process startup

Andres Freund <andres@anarazel.de>

View thread View on postgr.es
From: Andres Freund <andres@anarazel.de>
To: Kyotaro Horiguchi <horikyota.ntt@gmail.com>
Cc: pgsql-hackers@postgresql.org
Date: 2021-09-14T03:11:29Z
Lists: pgsql-hackers

Attachments

Hi,

On 2021-08-05 12:50:15 -0700, Andres Freund wrote:
> On 2021-08-03 16:50:24 +0900, Kyotaro Horiguchi wrote:
> > > - My first attempt at PostgresMainSingle() separated the single/multi user
> > >   cases a bit more than the code right now, by having a PostgresMainCommon()
> > >   which was called by PostgresMainSingle(), PostgresMain(). *Common only
> > >   started with the MessageContext allocation, which did have the advantage of
> > >   splitting out a few of the remaining conditional actions in PostgresMain()
> > >   (PostmasterContext, welcome banner, Log_disconnections). But lead to a bit
> > >   more duplication. I don't really have an opinion on what's better.
> >
> > I'm not sure how it looked like, but isn't it reasonable that quickdie
> > and log_disconnections(). handle IsUnderPostmaster instead?  Or for
> > log_disconnections, Log_disconnections should be turned off at
> > standalone-initialization?
> 
> I was wondering about log_disconnections too. The conditional addition of the
> callback is all that forces log_disconnections to be PGC_SU_BACKEND rather
> than PGC_SUSET too. So I agree that moving a check for Log_disconnections and
> IsUnderPostmaster into log_disconnections is a good idea.

I did that, and it didn't seem a clear improvement..


> > > - I had to move the PgStartTime computation to a bit earlier for single user
> > >   mode. That seems to make sense to me anyway, given that postmaster does so
> > >   fairly early too.
> > >
> > >   Any reason that'd be a bad idea?
> > >
> > >   Arguably it should even be a tad earlier to be symmetric.
> >
> > Why don't you move the code for multiuser as earlier as standalone does?
> 
> I think it's the other way round - right now the standalone case is much later
> than the multiuser case. Postmaster determines PgStartTime after creating
> shared memory, just before starting checkpointer / startup process - whereas
> single user mode in HEAD does it just before accepting input for the first
> time.

Did that in the attached.


I've attached the three remaining patches, after some more polish. Unless
somebody argues against I plan to commit these soon-ish.

Greetings,

Andres Freund

Commits

  1. process startup: Split single user code out of PostgresMain().

  2. process startup: Do InitProcess() at the same time regardless of EXEC_BACKEND.

  3. process startup: Initialize PgStartTime earlier in single user mode.

  4. pgstat: Schedule per-backend pgstat shutdown via before_shmem_exit().

  5. pgstat: Bring up pgstat in BaseInit() to fix uninitialized use of pgstat by AV.

  6. process startup: Always call Init[Auxiliary]Process() before BaseInit().

  7. process startup: Centralize pgwin32_signal_initialize() calls.

  8. process startup: Remove bootstrap / checker modes from AuxProcType.

  9. process startup: Move AuxiliaryProcessMain into its own file.

  10. process startup: auxprocess: reindent block

  11. process startup: Separate out BootstrapModeMain from AuxiliaryProcessMain.

  12. process startup: Rename postmaster's --forkboot to --forkaux.