Re: A micro-optimisation for ProcSendSignal()

On Tue, Aug 3, 2021 at 2:56 PM Andres Freund <andres@anarazel.de> wrote:
> On 2021-08-03 13:44:58 +1200, Thomas Munro wrote:
> > In the case of buffer pin waits, we switch to storing the pgprocno of
> > the waiter.  In the case of SERIALIZABLE READ ONLY DEFERRABLE waits, we
> > derive the pgprocno from the vxid (though that's not yet as efficient as
> > it could be).
>
> That's kind of an understatement :)

I abandoned the vxid part for now and went back to v3.  If/when
BackendId is replaced with or becomes synonymous with pgprocno, we can
make this change and drop the pgprocno member from SERIALIZABLEXACT.

> > +     SetLatch(&ProcGlobal->allProcs[pgprocno].procLatch);

> I think some validation of the pgprocno here would be a good idea. I'm worried
> that something could cause e.g. INVALID_PGPROCNO to be passed in, and suddenly
> we're modifying random memory. That could end up being a pretty hard bug to
> catch, because we might not even notice that the right latch isn't set...

Added.

> >  /* Accessor for PGPROC given a pgprocno. */
> >  #define GetPGProcByNumber(n) (&ProcGlobal->allProcs[(n)])
> > +/* Accessor for pgprocno given a pointer to PGPROC. */
> > +#define GetPGProcNumber(proc) ((proc) - ProcGlobal->allProcs)
>
> I'm not sure this is a good idea. There's no really cheap way for the compiler
> to compute this, because sizeof(PGPROC) isn't a power of two. Given that
> PGPROC is ~880 bytes, I don't see all that much gain in getting rid of
> ->pgprocno.

Yeah, that would need some examination; 0002 patch abandoned for now.

Pushed.