Re: [RFC][PATCH 0/3] Overlay manager for predefined DT overlay fragments

[Rob Herring]

On Fri, Oct 13, 2017 at 04:51:03PM -0700, John Stultz wrote:
> On Fri, Oct 13, 2017 at 2:41 PM, Rob Herring <robh@xxxxxxxxxx> wrote:
> > On Mon, Oct 09, 2017 at 10:34:17PM -0700, John Stultz wrote:
> >> In working with the HiKey and HiKey960 as targets for AOSP,
> >> Dmitry developed the following overlay manager driver, which
> >> allows a number of pre-determined DT overlay configurations to
> >> be defined, and the configurations to be enabled at boot time
> >> via a kernel boot argument.
> >>
> >> This has been submitted before, but while the earlier discussion
> >> didn't really resolve to any sort of actionable direction, this
> >> issue cropped up again and was a major discussion topic at the
> >> Linux Plumbers Conference Android Microconference, so I suspect
> >> it is worth revisiting this solution again.
> >
> > The Treble way to handle this is it is the bootloader's problem, right?
> 
> I believe so, though with the treble approach, the device
> configuration is still somewhat fixed. Just that configuration is
> spread out between multiple partitions. Its a bit of a separate issue.
> 
> >> The overall use case is being able to configure devboards that
> >> support a number of different mezzanine peripherals which
> >> unfortunately cannot be probed. Some example mezzanines are
> >> LCD panels or sensor hubs, as well as other options.
> >
> > There's probably some usecases for putting the overlay into the base
> > DT, but I don't think mezzanine boards is one of them especially for dev
> > boards where we may want to share overlays.
> >
> > I could see it in cases where you have 2nd source components for a
> > product and want to select one.
> >
> >> The new functionality this driver provides is a mechanism to
> >> specify multiple pre-determined dt overlay fragments in a
> >> dtb file, and providing a way to select which dt fragments
> >> should be applied via the kernel boot argument.
> >>
> >> The desire to use a kernel boot-argument as the selection
> >> mechanism, comes from the Android Boot Image format not handling
> >> dtbs independently. Usually with Android, the dtb is appended
> >> to the kernel image, and modifying that is much more difficult
> >> then changing the boot argugments. There is also a usability
> >> argument that using a kernel command option to select
> >> pre-defined entries is simpler for users to navigate.
> >
> > Doesn't Treble address the handling of overlays? Of course, none of what
> > has been outlined for Treble has been reviewed upstream either... The
> > concept seems fine, it's the vendor implementations that worry me.
> 
> Again, I'm not sure how that connects to the proposal here. Having
> treble's multiple overlays being picked up in a standard fashion is
> nice, but how does that translate into a user booting a board with one
> mezzanine and then change the mezzanine and be able to get the new
> mezzanine to work with a minimal amount of fuss.

Perhaps I need to study how exactly treble systems will select all their 
overlays. Supporting different panels or other 2nd source components is 
a pretty common problem in production devices. Those aren't probe-able 
either and folks don't want to flash different images for each 
combination of h/w.

> > As far as usability, a RPi user would probably say listing
> > overlays in a text file is easiest as that's how the RPi firmware works.
> 
> Right, and I agree that usability wise, its very similar to the
> cmdline arguments. But that requires firmware that keeps track of
> various overlays and has some small filesystem where that config file
> can be stored and easily updated.
> 
> >> Also, since the mezzanines are unable to be probed, we cannot
> >> use other solutions, like having the bootloader specify
> >> additional dtb overlays to the kernel.
> >
> > Not sure I follow.
> >
> > Someone has to decide what to put on the kernel command line. If you are
> > setting the overlay in the command line at bootimage build time or in
> > the bootloader, then you could just apply the overlay at that point in
> > time. Maybe it's slightly easier to change the kernel command line than
> > change the dtb.
> 
> Not, its not slightly easier, its much *much* easier to tweak the boot
> arguments. Since the dtb is appended to the kernel image w/ the
> abootimg format, its not simple at all to cut one out and replace it.

Appended dtb was to support legacy non-DT aware bootloaders. It never 
should have been carried over to arm64.

> While the cmdline is trivial to do with the abootimg command:
> abootimg -x boot.img
> vi bootimg.cfg
> abootimg -u boot.img -f ./bootimg.cfg

It's easy because there is a tool to do so, not because a the kernel 
command line is inheritly easier to modify than a dtb.

It would be trivial to write a tool or add to this one the code to find 
the dtb within the image and modify it. That's what every bootloader can 
do.

Of course, abootimg is a horrible creation IMO. Why Android can't use 
something called a filesystem for the boot partition is beyond me. Then 
modifying it would not require special tools that could only modify 
whatever someone had the itch to modify. And combining the kernel, dtb, 
and overlays into a single file would not be necessary.

> and re-flash.  This doesn't require access to the kernel source or
> rebuilding anything.
> 
> It seems you're suggesting that there be some sort of special overlay
> partition which users have to flash with pre-built images containing
> the appropriate overlay dtbs, so that something like the treble
> overlay-per-partition approach could be used.

I'm not really suggesting anything. I'm not going to take something that 
*only* solves your usecase of apply an overlay embedded in a base dtb 
based on the kernel command line. I have no issue really with either one 
of those. What I don't want is another "overlay manager" for each 
usecase. And sorry, you don't win for being first (you're not really).
All this I said before.

> And in the case with hikey/hikey960, this may be doable, as we can
> modify the UEFI source, but it would also require users to have all
> the various overlay images to flash as well, which adds complexity.
> Additionally on other boards which don't have open firmware, such a
> solution may not be feasible. The benefit to this approach is it
> doesn't require such standardized firmware behavior.

Unless there's some great, overwhelming demand to support such devices 
with closed firmware, there's no benefit to me to care. What board do we 
have with closed firmware that needs this?

Rob