[PATCH 2/3] binman: Remove header from compressed data

[Simon Glass]

Hi Stefan,

On Thu, 4 Aug 2022 at 01:50, Stefan Herbrechtsmeier
<stefan.herbrechtsmeier-oss at weidmueller.com> wrote:
>
> Hi Simon,
>
> Am 03.08.2022 um 20:14 schrieb Simon Glass:
> > Hi Stefan,
> >
> > On Tue, 2 Aug 2022 at 07:45, Stefan Herbrechtsmeier
> > <stefan.herbrechtsmeier-oss at weidmueller.com> wrote:
> >>
> >> Hi Simon,
> >>
> >> Am 02.08.2022 um 14:41 schrieb Simon Glass:
> >>> Hi Stefan,
> >>>
> >>> On Tue, 2 Aug 2022 at 06:29, Stefan Herbrechtsmeier
> >>> <stefan.herbrechtsmeier-oss at weidmueller.com> wrote:
> >>>>
> >>>> From: Stefan Herbrechtsmeier <stefan.herbrechtsmeier at weidmueller.com>
> >>>>
> >>>> Remove header from compressed data because this is uncommon, not
> >>>> supported by U-Boot and incompatible with external compressed artifacts.
> >>>>
> >>>> The header was introduced as part of commit eb0f4a4cb402 ("binman:
> >>>> Support replacing data in a cbfs") to allow device tree entries to be
> >>>> larger that the compressed contents. Regarding the commit "this is
> >>>> necessary to cope with a compressed device tree being updated in such a
> >>>> way that it shrinks after the entry size is already set (an obscure
> >>>> case)". This case need to be fixed without influence the compressed data
> >>>> by itself.
> >>>
> >>> I was not able to find a way around this due to the chicken-and egg
> >>> problem. Compressed data has an unpredictable size and adding an extra
> >>> uncompressed byte may increase or decrease the compressed size.
> >>
> >> Is it possible to use the `pad-after` attribute to record the unused
> >> space. In this case it is possible to calculate the size of the
> >> compressed data.
> >
> > Well if you update that attribute it can change the size of the DTB
> > which is the chicken-and-egg problem again. As far as I know, if
> > people set the size of the region to something a bit larger than
> > needed, the problem is avoided. But the image generation does need to
> > be deterministic.
>
> Does this means the size is only needed for the creation of the fitImage
> and not for decompression in u-boot?

Possibly, but of course we cannot do that. As you say, U-Boot mainline
does not expect or support the header, at present.

>
> >
> >>
> >> Do you have a test for this use case?
> >
> > There are various ones, e.g. testCompressDtb()
>
> Thanks. Now I understand the problem and why you call it a
> chicken-and-egg problem. It wasn't clear to me that the attributes are
> inside the DTB.

OK good.

>
> But I wonder how the decompression of the DTB works if the fitImage
> implementation doesn't support the header.

It doesn't. Something needs to change here for compression to work.

>
> >>> So my solution was to add the header.
> >>
> >> Is the header used outside of binman? I don't spot it in the decompress
> >> fitImage implementation.
> >
> > It is used in the Chromium OS verified boot implementation, but not elsewhere.
> >
> >>
> >>
> >>> It is optional though, so can we perhaps have a property in the
> >>> description to enable it?
> >>
> >> Is this header needed and supported outside of binman?
> >>
> >> At the moment the header is incompatible and not well documented. It
> >> took me some time to find out why my gzip compression via binman doesn't
> >> work as expected because I assume a compatibility between binman
> >> compress and fitImage decompress.
> >
> > Yes I understand that, however you can pass a parameter to not include
> > the size value.
>
> Do we need the header outside of the DTB? Otherwise we could handle this
> special or we could add a special compression type.
>
> > It would also be possible to add a property to the image (top-level
> > section) that indicates whether this field is present, such property
> > to apply to the whole image. We could have it default to off, if you
> > like.
>
> Do we really need the header outside of the DTB entry?

That's an interesting question. It is possible that we only need it if
DTB is present and is compressed. It should be possible to check this
by adjusting the tests and checking for failures.

But I am not sure it is a good idea, since it is wildly inconsistent.
I do prefer things to be deterministic - i.e.  you specify what you
want and you get it. If binman starts adopting obscure conventions it
could be confusing.

Regards,
Simon