[OpenPower-Firmware] HDAT documentation

[Stewart Smith]

On Fri, May 21, 2021, at 10:58 AM, Krystian Hebel wrote:
> On 20.05.2021 17:29, Stewart Smith wrote:
> > On Thu, May 20, 2021, at 8:09 AM, Krystian Hebel wrote:
> >> Hi,
> >>
> >> We are working on coreboot support for POWER9.
> >> Recently we have achieved the milestone where we could initialize the
> >> memory on the Talos II platform and start the coreboot ramstage phase [1].
> > Oh, that's really cool!
> >
> > I'd love to know more and any thoughts on getting it going on a Blackbird.
> 
> First we have to finish work on Talos, then we'll see. Given that Blackbird
> is by design using one CPU and what we did so far is "only" RAM 
> initialization
> it may even be better supported by the current state of code than Talos,
> unless there are more subtle differences in hardware. Unfortunately, we
> don't have access to that platform so we have no way of testing.

I'd be tempted to give it a try on mine tbh. Keeping the SBE code/entry-point?


> >> What we would like to achieve next is to start skiboot from coreboot.
> >>
> >> We can see that there is this HDAT interface between Hostboot and skiboot
> >> [2] which we probably would need to reimplement.
> >>
> >> Sadly, the documentation says that "The HDAT specification is currently
> >> not public.".
> >> Any idea why is that? Can we get some more information about
> >> the HDAT somewhere (apart from reading the Hostboot code)?
> > I have even better news: skiboot will *happily* take Device Tree instead.
> >
> > I'd solidly recommend just producing a dtb Device Tree blob for skiboot. It'll save you significant pain.
> That is great news! There is support for Device Tree for Arm in coreboot.
> I'm not sure how much of it is actually implemented, coreboot tends to
> keep code base as small as possible, but this is still better than nothing.

libdtb and libdevicetree are fairly small and self contained, and using these to serialize out whatever way you keep track of what hardware exists shouldn't be too difficult if it isn't there already.

> Is there any list of the minimal set of entries required in such Device 
> Tree?
> So far I've seen only documentation about the DT produced by skiboot, not
> the one consumed by it.

Yeah, that we never really documented :)

So, you're basically going to want to end up with a device tree equivalent to what exists after parse_hdat() in core/init.c - which you can get from a live machine either by hacking skiboot to save off a dtb somewhere after parse_hdat() and looking at the result, or dumping out the HDAT and using the skiboot hdat_to_dt unit test in hdat/test/ (there's a test there for a Witherspoon, but not any other P9 platform... patches are probably welcome).

You'll at least need CPU nodes, xscom and lpc to boot... but doing the above would likely give you a goal to get full functionality at runtime.

Skiboot should be pretty good at complaining on anything essential that's missing (we'll assert on it), and also really forgiving of missing bits (it'll just skip that bit of code).

> > Background:
> > During initial bringup on POWER7 machines, the FSP would generate HDAT, and thus we had to read it in skiboot (there was no way we were going to convince everyone else to change for something that was just a bit of research at the time), and this was also true for POWER8 when it was a bit of a mixed responsibility between Hostboot and the FSP to generate HDAT. On OpenPOWER POWER8 systems, we got a device tree from hostboot, but this mean Hostboot was producing two data structures with roughly the same data, but sometimes the DT was missing things... and it wasn't as fun as we would have liked. For POWER9, we said "just give us HDAT, it's fine" as a compromise, as at least it's only one bit of code for Hostboot to maintain.
> > There are places where P9 skiboot gets everything in DT such as some simulation environments, and pdbg also will use device tree exclusively.
> 
> Thanks for this background. It is exactly this transition from DT on POWER8
> to no DT on POWER9 that made me believe HDAT is the way to go.

I have no idea where POWER10 may be going, but I'd expect skiboot to always be able to deal with device tree as an input as it's just soooo much nicer for humans to construct for use in simulation or alternate ways of IPLing the box.