Re: [RFC] devicetree: new FDT format version

[David Gibson <david-xT8FGy+AXnRB3Ne2BGzF6laj5H9X9Tb+@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 7658 bytes --]

On Mon, Jan 29, 2018 at 06:32:35PM +0000, Grant Likely wrote:
> On Thu, Jan 25, 2018 at 8:01 PM, Frank Rowand <frowand.list-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> > On 01/25/18 04:29, David Gibson wrote:
> >> On Wed, Jan 24, 2018 at 04:22:15PM -0800, Frank Rowand wrote:
> >>> On 01/24/18 13:16, Frank Rowand wrote:
> >>>> On 01/24/18 07:47, Rob Herring wrote:
> >>>>> On Tue, Jan 23, 2018 at 3:17 PM, Frank Rowand <frowand.list@gmail.com> wrote:
> >>>>>> On 01/23/18 04:42, David Gibson wrote:
> >>>>>>> On Mon, Jan 22, 2018 at 03:01:52PM -0600, Rob Herring wrote:
> >>>>>>>> On Mon, Jan 22, 2018 at 2:09 AM, Frank Rowand <frowand.list@gmail.com> wrote:
> >>>>>>>>> Hi All,
> >>>>>>>>>
> >>>>>>>>> I've tried to create a decent distribution list, but I'm sure I've missed
> >>>>>>>>> someone or some important list.  Please share this with anyone you think
> >>>>>>>>> will be affected.
> >>>>>>>>>
> >>>>>>>>> I have been playing around with some thoughts for some additions to
> >>>>>>>>> the devicetree FDT aka blob format.
> >>>>>>>>>
> >>>>>>>>> I would like to get the affected parties thinking about how additions to
> >>>>>>>>> the format could improve whichever pieces of FDT related technology you
> >>>>>>>>> work on or care about.  In my opinion, the FDT format should change
> >>>>>>>>> very infrequently because of the impact on so many projects that have
> >>>>>>>>> to work together to create a final solution, plus the many many users
> >>>>>>>>> of those projects.
> >>>>>>>>
> >>>>>>>> A few things discussed before:
> >>>>>>>> - Adding type information Even just tagging phandles would be good.
> >>>>>>>
> >>>>>>> I'm a bit dubious about this.  It would have to be "hints" only -
> >>>>>>> there's not really anyway we can put in authoritative type
> >>>>>>> information, since dtc itself doesn't really know that.  It's also
> >>>>>>> hard to see how that could be done in a way which wouldn't either a)
> >>>>>>> require very awkward parallel lookup of the data and type information
> >>>>>>> or b) not be backwards compatible (even read only).
> >>>>>
> >>>>> I never said it was possible. :) I'm just trying to enumerate possible
> >>>>> FDT format changes because I don't think we want to continuously
> >>>>> trickle out FDT changes even if they are backwards compatible.
> >>>>
> >>>> Yes, I'm trying to capture any pending changes in a single version change.
> >>>>
> >>>>
> >>>>>>>> - Allow applying overlays by just appending to the blob. The need for
> >>>>>>>> this is somewhat gone now that libfdt can just fully apply overlays.
> >>>>>>>
> >>>>>>> I'm not really sure what you want here.  I mean you could easily allow
> >>>>>>> the format to allow multiple appended overlays, and define that to
> >>>>>>> mean the overlaid result.  At some point *something* is going to have
> >>>>>>> to really do the application, so I'm not sure that it really buys you
> >>>>>>> that much.  It also makes nearly every operation on the tree in libfdt
> >>>>>>> horrible to implement, at least within the other constraints the
> >>>>>>> interface was designed around; you'll continually have to scan the
> >>>>>>> entire tree just in case some other overlay fragment has some bearing
> >>>>>>> on the thing you're looking at.  It confuses the interface too: what
> >>>>>>> does "node offset" mean if the same node could be built up from
> >>>>>>> overlay fragments at multiple offsets.
> >>>>>
> >>>>> The idea was to avoid applying overlays to flattened trees at all.
> >>>>> You're just passing the problem to the next stage (typically the
> >>>>> kernel). But we have applying overlays to flattened trees now, so
> >>>>> maybe there's no need anymore.
> >>>>>
> >>>>>> Somewhat echoing David's comment, I'm also not sure what you mean.
> >>>>>> And trying to not overly influence this conversation with preconceptions
> >>>>>> from what I'm going to propose.
> >>>>>>
> >>>>>> My first shot at the new format added a field to the FDT to indicate
> >>>>>> that an overlay FDT was concatenated to the FDT (and the overlay FDT
> >>>>>> in turn could set it's field to concatenate another overlay FDT).
> >>>>>
> >>>>> Yes, something like this is what I meant. This was something Grant had
> >>>>> talked about.
> >>>>>
> >>>>>> But in the end I decided that information belonged outside the FDT,
> >>>>>> and it was sufficient to require that all FDTs be padded out so that
> >>>>>> if an overlay FDT was concatenated to the FDT, the overlay FDT would
> >>>>>> be properly aligned.
> >>>>>
> >>>>> I can't think of why this wouldn't work either.
> >>>>>
> >>>>>> For ease of typing, I'll call this "chaining" or "chained".  For
> >>>>>> Linux, I am envisioning no kernel use of data from a chained FDT
> >>>>>> until after the tree is unflattened.  I haven't done an exhaustive
> >>>>>> search to determine all of the uses of data directly from the
> >>>>>> flattened FDT, but I am optimistic that there will not be any such
> >>>>>> access that would require data from a chained FDT (and we could
> >>>>>> make this a rule).
> >>>>>
> >>>>> This would be a major downside to leaving it up to the kernel because
> >>>>> what can't be touched is hard to enumerate and could change. For
> >>>>> example, we added earlycon and now the uart node can't be modified.
> >>>>
> >>>> What you say makes sense.  So I'll reverse myself and say that for
> >>>> Linux, we should update the FDT read code to scan all chained
> >>>> overlay FDT(s) as well as the base FDT.
> >>>
> >>> < snip >
> >>>
> >>> What I wrote was somewhat ambiguous.  What I meant by "FDT read
> >>> code" was functions that check for existence of nodes in the
> >>> FDT or read property values from the FDT.
> >>
> >> Oh.. not just FDT unflattening code.
> >>
> >> The trouble with this is that scanning for a specific node or property
> >> in a set of chained overlays is highly nontrivial.  Even if you set
> >> aside the arguably self-imposed design constraints in libfdt, you
> >> can't just do the same lookup in each fragment: along the way you need
> >> to resolve the path at which each fragment applies.  That in itself is
> >> non-trivial.  If you have overlays applying on top of other overlays,
> >> that could involve recursive lookups of things from previous overlays.
> >> It's spectacularly complicated and we have to do it on *every single*
> >> read operation.
> >
> > I totally overlooked having to resolve each fragment.  You are right,
> > that makes the problem very complex instead of trivial.
> 
> It would be possible to do if each fragment was pre-resolved at append
> time to the node that it modifies. ie. each fragment node points back
> with an offset to the node it modifies Then searching the tree could
> be done by walking backwards through the fragments instead of
> searching forwards. Would need to research the best way to do that,
> and it would also mean that merely appending a DTB fragment isn't an
> option.

I guess.  But if you're going to pre-resolve, why not just apply the
overlay.  Yes, it's a bit simpler at the time, but you increase
overall complexity in that: 1) you split one logical operation into
two places, 2) you have to define an intermediate
resolved-but-not-applied overlay format.

We already have a full overlay application implementation that works
on flattened trees.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]