Re: [PATCH v3] base: Fix HOST_ARCH for native builds

[Jan Kiszka <jan.kiszka@siemens.com>]

On 05.10.23 10:36, 'Jan Kiszka' via isar-users wrote:
> On 05.10.23 08:56, 'Jan Kiszka' via isar-users wrote:
>> On 05.10.23 07:48, Uladzimir Bely wrote:
>>> On Sun, 2023-10-01 at 11:09 +0200, Jan Kiszka wrote:
>>>> From: Jan Kiszka <jan.kiszka@siemens.com>
>>>>
>>>> HOST_ARCH must be DISTRO_ARCH when we are not cross-building.
>>>> Otherwise,
>>>> recipes that set PACKAGE_ARCH to it will fail in native builds.
>>>>
>>>> To avoid recursions, we have to rework the ISAR_CROSS_COMPILE setting
>>>> in
>>>> imagetypes.bbclass to an anonymous python function.
>>>>
>>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
>>>> ---
>>>>
>>>> This looks better now.
>>>>
>>>> Maybe we can even kill BUILD_HOST_ARCH, now that HOST_ARCH is fixed.
>>>>
>>>>  meta/classes/base.bbclass       | 6 ++++--
>>>>  meta/classes/imagetypes.bbclass | 6 ++++--
>>>>  2 files changed, 8 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/meta/classes/base.bbclass b/meta/classes/base.bbclass
>>>> index 88004120..f315a9d5 100644
>>>> --- a/meta/classes/base.bbclass
>>>> +++ b/meta/classes/base.bbclass
>>>> @@ -49,13 +49,15 @@ def oe_import(d):
>>>>  # We need the oe module name space early (before INHERITs get added)
>>>>  OE_IMPORTED := "${@oe_import(d)}"
>>>>  
>>>> -def get_deb_host_arch():
>>>> +def get_deb_host_arch(d):
>>>>      import subprocess
>>>> +    if d.getVar("ISAR_CROSS_COMPILE") != "1":
>>>> +        return d.getVar("DISTRO_ARCH")
>>>>      host_arch = subprocess.check_output(
>>>>          ["dpkg", "--print-architecture"]
>>>>      ).decode('utf-8').strip()
>>>>      return host_arch
>>>> -HOST_ARCH ??= "${@get_deb_host_arch()}"
>>>> +HOST_ARCH ??= "${@get_deb_host_arch(d)}"
>>>>  HOST_DISTRO ??= "${DISTRO}"
>>>>  
>>>>  die() {
>>>> diff --git a/meta/classes/imagetypes.bbclass
>>>> b/meta/classes/imagetypes.bbclass
>>>> index a3be0a1d..205377b1 100644
>>>> --- a/meta/classes/imagetypes.bbclass
>>>> +++ b/meta/classes/imagetypes.bbclass
>>>> @@ -65,8 +65,10 @@ UBIFS_IMG ?=
>>>> "${PP_DEPLOY}/${IMAGE_FULLNAME}.ubifs"
>>>>  
>>>>  # glibc bug 23960
>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=23960
>>>>  # should not use QEMU on armhf target with mkfs.ubifs < v2.1.3
>>>> -THIS_ISAR_CROSS_COMPILE := "${ISAR_CROSS_COMPILE}"
>>>> -ISAR_CROSS_COMPILE:armhf = "${@bb.utils.contains('IMAGE_BASETYPES',
>>>> 'ubifs', '1', '${THIS_ISAR_CROSS_COMPILE}', d)}"
>>>> +python() {
>>>> +    if d.getVar('DISTRO_ARCH') == 'armhf' and
>>>> bb.utils.contains('IMAGE_BASETYPES', 'ubifs', True, False, d):
>>>> +        d.setVar('ISAR_CROSS_COMPILE', '1')
>>>> +}
>>>>  
>>>>  IMAGE_CMD:ubifs() {
>>>>      ${SUDO_CHROOT} /usr/sbin/mkfs.ubifs ${MKUBIFS_ARGS} \
>>>
>>> The patch is merged now, but it seems to bring a regression.
>>>
>>> Now, native compilation if imx6-sabrelite fails, because we set
>>> ISAR_CROSS_COMPILE to "1", while it seems should be "0" (like
>>> previously used temporary value of THIS_ISAR_CROSS_COMPILE.
>>>
>>
>> No, this is correct: The image recipe must have ISAR_CROSS_COMPILE = "1"
>> because of
>>
>> # glibc bug 23960 https://sourceware.org/bugzilla/show_bug.cgi?id=23960
>> # should not use QEMU on armhf target with mkfs.ubifs < v2.1.3
>>
>> Package recipes have it disabled, as desired.
>>
> 
> ISAR_CROSS_COMPILE is indeed correct - but too late. The problem is that
> the crossvars does not see the python block results of imagetypes, thus
> switches to native build for the imager.
> 
> I'm playing with yet another variant that looks more like the original
> code but avoids the recursion-triggering override.
> 

Nope, does not fly, we need to revert: It is glued into isar that
HOST_ARCH is static and does not depend on ISAR_CROSS_COMPILE. Users of
HOST_ARCH that expect such a dependency should have taken
BUILD_HOST_ARCH. Now I see what we have two.

So the bugs are elsewhere in the code, e.g. in those recipes that do

PACKAGE_ARCH = "${HOST_ARCH}"

but probably also in sdk.bbclass, native.bbclass, multiarch.bbclass,
essential.bbclass - if not more. Checking.

Sigh.

Jan

-- 
Siemens AG, Technology
Linux Expert Center