Re: 2026 TODO list!

Justin Hibbits <[email protected]> Tue, 13 Jan 2026 17:44:32 -0500
Newsgroups gmane.os.freebsd.devel.ppc
Message-ID <[email protected]>
On Tue, 13 Jan 2026 15:27:05 -0600 (CST)
Timothy Pearson <[email protected]> wrote:

> ----- Original Message -----
> > From: "Lexi Winter" <[email protected]>
> > To: "freebsd-ppc" <[email protected]>
> > Sent: Tuesday, January 13, 2026 3:42:28 AM
> > Subject: Re: 2026 TODO list!  
> 
> > Adrian Chadd wrote in
> > <CAJ-VmonYHdVhdPiaLHUqzPv33_iAGiTY=_hGrhg4rXF5+YZPQQ@mail.gmail.com>:
> >  
> >> I'd like to assemble a 2026 TODO list for PPC64 work. Justin, Ivy,
> >> Warner, I and and others have been poking at the sharp edges and
> >> trying to use/abuse various PPC64 hardware and VMs.
> >> 
> >> I'd appreciate some brainstorming from y'all about what you'd like
> >> to see fixed, implemented and working.  
> > 
> > i've been meaning to look at making lldb work on powerpc64le for
> > both kernel and userland debugging.  this looks like it should be
> > pretty simple, since it already supports FreeBSD/ppc64 and
> > Linux/ppc64le, so it's just a matter of combining the two.
> > 
> > i've been busy with non-FreeBSD stuff recently but i'm hoping i can
> > look at this soon.
> > 
> > Raptor has kindly given me access to a Talos II so i can do dev work
> > and also test patches there if needed.
> > 
> > a couple of other things i think are important:
> > 
> > - non-OpenFirmware ppc64le loader - i see a couple of people have
> >  already mentioned this.
> > 
> > - SMP performance is quite bad; this 44-core SMT4 Talos II runs
> >  buildworld at the same speed as my 8-core 2.6GHz Xeon system.
> >  i mentioned this to Justin and he suspects this could be from
> >   locking issues in vm/pmap.  
> 
> After digging around in the PCI driver for the past month, I think
> we're also not getting DMA enabled on potentially *any* PCI devices.
> It looks like the DMA tag was removed for PHB4 (POWER9) support, but
> it might have worked on PHB3 (POWER8).  Once I finish fixing the rest
> of the PE 1:1 mapping mess I'll start looking at that and see if we
> can't get DMA reenabled -- no DMA would definitely explain poor
> performance on disk intensive activity...
> 

DMA should be active already.  Without a bus tag it defaults to nexus,
which allows all as 1:1, so no translation, restrictions, or anything
like that.  Also no IOMMU is attached.

- Justin