[PATCH 0/2] handle DPKG_ARCH=all case for transitive deps

7 views
Skip to first unread message

Felix Moessbauer

unread,
Aug 20, 2025, 8:42:02 AMAug 20
to isar-...@googlegroups.com, jan.k...@siemens.com, cedric.h...@siemens.com, adriaan...@siemens.com, christop...@siemens.com, Felix Moessbauer
This series fixes an issue with DPKG_ARCH=all packages that have
dependencies to other DPKG_ARCH=all packages. Hereby, the transitive
packages were built for the wrong architecture. The issue became
apparent when testing the SBOM series by Christoph, which builds
packages to be used during the ISAR build (not for the final target).

While fixing this, another issue surfaced where files are missing in
dpkg-raw packages.

Best regards,
Felix

Felix Moessbauer (2):
dpkg-raw: add files to source package
handle DPKG_ARCH=all case for transitive deps

meta/classes/dpkg-raw.bbclass | 4 ++--
meta/classes/multiarch.bbclass | 18 +++++++++++++++++-
2 files changed, 19 insertions(+), 3 deletions(-)

--
2.50.1

Felix Moessbauer

unread,
Aug 20, 2025, 8:42:06 AMAug 20
to isar-...@googlegroups.com, jan.k...@siemens.com, cedric.h...@siemens.com, adriaan...@siemens.com, christop...@siemens.com, Felix Moessbauer
In dpkg-raw, the user can place files in ${D} which are then installed
into the resulting binary package. Hereby, ${D} is a directory outside
of ${S} (since b19cd25) to not interfere with other data added to ${S}.
However, by that the files are not added to the source package. This
remained unnoticed, as the directory dh_install installs from is set
to absolute path, hence the installed files actually came from the
absolute path and not from the extracted source package. In case of
${PN} == ${BPN}, this path was always there as it has been created by
previous tasks. However, with the switch in 2ca3a7e5 to only build the
source package once, the path is not always there.

We fix this by adding the files to the source package (under image) and
install from a relative base. We further use a sub-path (image) in ${S}
as a temporary location to not run into the issue solved in b19cd25.

Fixes: 2ca3a7e5 ("dpkg-source: Build source package only once")
Signed-off-by: Felix Moessbauer <felix.mo...@siemens.com>
---
meta/classes/dpkg-raw.bbclass | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/meta/classes/dpkg-raw.bbclass b/meta/classes/dpkg-raw.bbclass
index a7bf204a..d4cb7d8a 100644
--- a/meta/classes/dpkg-raw.bbclass
+++ b/meta/classes/dpkg-raw.bbclass
@@ -5,7 +5,7 @@

inherit dpkg

-D = "${WORKDIR}/image"
+D = "${S}/image"

# Default to creating a binary-indep package
DPKG_ARCH ??= "all"
@@ -30,6 +30,6 @@ do_prepare_build() {
cat <<EOF >> ${S}/debian/rules

override_dh_install:
- dh_install --sourcedir=${PP}/image
+ dh_install --sourcedir=image
EOF
}
--
2.50.1

Felix Moessbauer

unread,
Aug 20, 2025, 8:42:07 AMAug 20
to isar-...@googlegroups.com, jan.k...@siemens.com, cedric.h...@siemens.com, adriaan...@siemens.com, christop...@siemens.com, Felix Moessbauer
Arch=all packages might build depend on other arch=all packages, hence we
need to correctly model the dependency chain. Otherwise the transitive
dependencies are built for the distro arch instead of the native arch.

We implement this by dispatching the non-native variant of DPKG_ARCH=all
packages to the -native variant by adding a dependency. We further
replace the non-native do_deploy_dep task with a noop to preserve the
dependency chain.

Co-developed-by: Adriaan Schmidt <adriaan...@siemens.com>
Signed-off-by: Felix Moessbauer <felix.mo...@siemens.com>
---
meta/classes/multiarch.bbclass | 18 +++++++++++++++++-
1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/meta/classes/multiarch.bbclass b/meta/classes/multiarch.bbclass
index babdfbd4..c2bba21f 100644
--- a/meta/classes/multiarch.bbclass
+++ b/meta/classes/multiarch.bbclass
@@ -29,7 +29,11 @@ python() {
d.appendVar('BBCLASSEXTEND', ' compat')

# build native separately only when it differs from the target variant
- if not archIsAll and archDiffers:
+ # We must not short-circuit for DPKG_ARCH=all packages, as they might
+ # have transitive dependencies which need to be built for -native.
+ # This special handling for DPKG_ARCH=all packages is left to the
+ # multiarch_virtclass_handler
+ if archDiffers:
d.appendVar('BBCLASSEXTEND', ' native')
else:
extend_provides(pn, 'native', d)
@@ -86,6 +90,8 @@ python multiarch_virtclass_handler() {
d.setVar(var, ' '.join(multiarch_var))

pn = e.data.getVar('PN')
+ archDiffers = d.getVar('HOST_ARCH') != d.getVar('DISTRO_ARCH')
+ archIsAll = d.getVar('DPKG_ARCH') == 'all'
if pn.endswith('-compat'):
e.data.setVar('BPN', pn[:-len('-compat')])
e.data.appendVar('OVERRIDES', ':class-compat')
@@ -96,6 +102,16 @@ python multiarch_virtclass_handler() {
e.data.appendVar('OVERRIDES', ':class-native')
fixup_pn_in_vars(e.data)
fixup_depends('-native', e.data)
+ elif archIsAll and archDiffers:
+ # Arch=all packages might build depend on other arch=all packages,
+ # hence we need to correctly model the dependency chain.
+ # We implement this by dispatching the non-native variant to the -native
+ # variant by adding a dependency. We further replace the non-native
+ # do_deploy_dep task with a noop to preserve the dependency chain.
+ e.data.setVar('do_deploy_deb', '')
+ bb.build.deltask('deploy_deb', e.data)
+ bb.build.addtask('deploy_deb', 'do_build', '', e.data)
+ e.data.setVarFlag('do_deploy_deb', 'depends', f'{pn}-native:do_deploy_deb')
}
addhandler multiarch_virtclass_handler
multiarch_virtclass_handler[eventmask] = "bb.event.RecipePreFinalise"
--
2.50.1

Andreas Naumann

unread,
Sep 10, 2025, 7:00:40 AM (4 days ago) Sep 10
to isar-...@googlegroups.com, felix.mo...@siemens.com
Hi Felix, all,

Am 20.08.25 um 14:41 schrieb 'Felix Moessbauer' via isar-users:
> Arch=all packages might build depend on other arch=all packages, hence we
> need to correctly model the dependency chain. Otherwise the transitive
> dependencies are built for the distro arch instead of the native arch.
>
> We implement this by dispatching the non-native variant of DPKG_ARCH=all
> packages to the -native variant by adding a dependency. We further
> replace the non-native do_deploy_dep task with a noop to preserve the
> dependency chain.
>
> Co-developed-by: Adriaan Schmidt <adriaan...@siemens.com>
> Signed-off-by: Felix Moessbauer <felix.mo...@siemens.com>
> ---
> meta/classes/multiarch.bbclass | 18 +++++++++++++++++-
> 1 file changed, 17 insertions(+), 1 deletion(-)
>
> diff --git a/meta/classes/multiarch.bbclass b/meta/classes/multiarch.bbclass
> index babdfbd4..c2bba21f 100644
> --- a/meta/classes/multiarch.bbclass
> +++ b/meta/classes/multiarch.bbclass
> @@ -29,7 +29,11 @@ python() {
> d.appendVar('BBCLASSEXTEND', ' compat')
>
> # build native separately only when it differs from the target variant
> - if not archIsAll and archDiffers:
> + # We must not short-circuit for DPKG_ARCH=all packages, as they might
> + # have transitive dependencies which need to be built for -native.

Funny enough i suspected to have the same problem yesterday, but it
turned out to
be different: When a package is single architecture only, e.g. arm64, it
should not be
extended as native if HOST_ARCH is amd64.

So I added a check to BBCLASSEXTEND only if DPKG_ARCH != HOST_ARCH.

> + # This special handling for DPKG_ARCH=all packages is left to the
> + # multiarch_virtclass_handler
> + if archDiffers:
> d.appendVar('BBCLASSEXTEND', ' native')
Now, seeing your patch to remove the not-"all" case also, I wonder if
that could be used
as the sole check before BBCLASSEXTEND. Like e.g.
archIsNative = d.getVar('DPKG_ARCH') == d.getVar('HOST_ARCH')
if isNative:
   d.appendVar('BBCLASSEXTEND', ' native')
> else:
> extend_provides(pn, 'native', d)

but now I wonder what the extend_provides is needed for. If HOST_ARCH ==
DISTRO_ARCH, what's
the use of providing a native variant?

But back to your Patch: Is it possible that -compat "all" packages also have
dependencies on other "all" packages, which in turn have dependencies on
architecture specific packages?

regards,
Andreas
Andreas Naumann

emlix GmbH
Headquarters: Berliner Str. 12, 37073 Goettingen, Germany
Phone +49 (0)551 30664-0, e-mail in...@emlix.com
District Court of Goettingen, Registry Number HR B 3160
Managing Directors: Heike Jordan, Dr. Uwe Kracke
VAT ID No. DE 205 198 055
Office Berlin: Panoramastr. 1, 10178 Berlin, Germany
Office Bonn: Bachstr. 6, 53115 Bonn, Germany
http://www.emlix.com

MOESSBAUER, Felix

unread,
Sep 11, 2025, 6:20:27 AM (3 days ago) Sep 11
to Schmidt, Adriaan, anau...@emlix.com, isar-...@googlegroups.com
Good catch. This case is currently not handled. But I would like to
have it in a separate patch, as it tackles another issue.

>
> > + # This special handling for DPKG_ARCH=all packages is left to the
> > + # multiarch_virtclass_handler
> > + if archDiffers:
> > d.appendVar('BBCLASSEXTEND', ' native')
> Now, seeing your patch to remove the not-"all" case also, I wonder if
> that could be used
> as the sole check before BBCLASSEXTEND. Like e.g.
> archIsNative = d.getVar('DPKG_ARCH') == d.getVar('HOST_ARCH')
> if isNative:
>    d.appendVar('BBCLASSEXTEND', ' native')
> > else:
> > extend_provides(pn, 'native', d)
>
> but now I wonder what the extend_provides is needed for. If HOST_ARCH ==
> DISTRO_ARCH, what's
> the use of providing a native variant?

If we are building for the native arch, we can just add the -native
variants to the PROVIDES, so dependencies of packages depending
explicitly on <foo>-native can be resolved (without the need for extra
recipe variants).

>
> But back to your Patch: Is it possible that -compat "all" packages also have
> dependencies on other "all" packages, which in turn have dependencies on
> architecture specific packages?

Good question. In general, package dependencies are resolved for the
native architecture (or the requested one). AFAIK this selection is
propagated down the dependency tree. By that, A:i386 -> B:all -> C:x
should always resolve C as either :all or i386 (except if there is a
explicit cross-arch dependency specifier like :any, which we don't
model in bitbake). Anyways, The A:i386 -> B:all -> C:i386 chain should
already be resolved correctly with this patch.

Tested by's are welcome :)

Felix
--
Siemens AG
Linux Expert Center
Friedrich-Ludwig-Bauer-Str. 3
85748 Garching, Germany

Andreas Naumann

unread,
Sep 12, 2025, 11:50:59 AM (yesterday) Sep 12
to MOESSBAUER, Felix, Schmidt, Adriaan, isar-...@googlegroups.com

Am 11.09.25 um 12:20 schrieb MOESSBAUER, Felix:
I thought it's possible to make the nesting of if-clauses easier that
way, but after some more thought and fiddling... unfortunately it's the
opposite. So yes, separate patch.

>
>>> + # This special handling for DPKG_ARCH=all packages is left to the
>>> + # multiarch_virtclass_handler
>>> + if archDiffers:
>>> d.appendVar('BBCLASSEXTEND', ' native')
>> Now, seeing your patch to remove the not-"all" case also, I wonder if
>> that could be used
>> as the sole check before BBCLASSEXTEND. Like e.g.
>> archIsNative = d.getVar('DPKG_ARCH') == d.getVar('HOST_ARCH')
>> if isNative:
>>    d.appendVar('BBCLASSEXTEND', ' native')
>>> else:
>>> extend_provides(pn, 'native', d)
>> but now I wonder what the extend_provides is needed for. If HOST_ARCH ==
>> DISTRO_ARCH, what's
>> the use of providing a native variant?
> If we are building for the native arch, we can just add the -native
> variants to the PROVIDES, so dependencies of packages depending
> explicitly on <foo>-native can be resolved (without the need for extra
> recipe variants).

I understand that its a simpler solution in that case. But, maybe I'm not
seeing the obvious, what would require a -native package then? If
target-arch
is host-arch, why would anything else than the base package be in any
dependency chain?

>
>> But back to your Patch: Is it possible that -compat "all" packages also have
>> dependencies on other "all" packages, which in turn have dependencies on
>> architecture specific packages?
> Good question. In general, package dependencies are resolved for the
> native architecture (or the requested one). AFAIK this selection is
> propagated down the dependency tree. By that, A:i386 -> B:all -> C:x
> should always resolve C as either :all or i386 (except if there is a
> explicit cross-arch dependency specifier like :any, which we don't
> model in bitbake). Anyways, The A:i386 -> B:all -> C:i386 chain should
> already be resolved correctly with this patch.
The patch doesnt change that -compat is "short-circuited" when
archIsAll, doesnt it?
Or maybe I dont fully understand how the code works yet.
>
> Tested by's are welcome :)

I'm working with it and keep an eye on it.

regards,
Andreas
Reply all
Reply to author
Forward
0 new messages