Re: [OE-core] [PATCH v4 3/7] classes-global/package_rpm.bbclass: Remove %pre and %postin for libc6

[Robert Yang <liezhi.yang@windriver.com>]

Hi RP,

On 2/10/26 16:52, Richard Purdie wrote:
> On Mon, 2026-02-09 at 23:10 -0800, Robert Yang via lists.openembedded.org wrote:
>> From: Robert Yang <liezhi.yang@windriver.com>
>>
>> Fixed:
>> $ bitbake core-image-full-cmdline:do_testimage
>> %post(busybox-1.37.0-r0.x86_64_x32): execv(/bin/sh) pid 624
>> error: failed to exec scriptlet interpreter /bin/sh: No such file or directory
>>
>> It is because busybox and libc6 depends on each other, busybox' elf files
>> depends on libc6, and libc6's postin depends on busybox' /bin/sh, the do_rootfs
>> works well is because dnf-native has set RPM_NO_CHROOT_FOR_SCRIPTS=1, but it
>> would be failed for a fresh rootfs.
>>
>> In rpm 4.20.1, it let the installed files' Requires win, so it installed
>> busybox firstly, but in rpm 6.0.1, it let the postin's Requires win since
>> postin would run immediately after the files are installed, this does make
>> sense, so it installed busybox (which provides /bin/sh required by libc6'
>> postin) firstly, then we got the errors. I couldn't find which commit made this
>> change because a lot of files and functions are refactored during 4.20.1 and
>> 6.0.0 (not .1), I also tried bisect, but failed because a lot of do_patch or
>> do_configure/do_compile failures for each build.
>>
>> For libc6's the postin is:
>>
>>    #!/bin/sh
>>    if [ x"$D" = "x" ]; then
>>        if [ -x /sbin/ldconfig ]; then /sbin/ldconfig ; fi
>>    fi
>>
>> This doesn't make sense for lib6 since there is no /bin/sh or ldconfig when
>> libc6 is not ready yet, so we can just remove libc6's postin to fix the
>> problem.
>>
>> And also remove the workarounds in oeqa/runtime/cases/dnf.py, they are not
>> needed any more since the circular dependency is fixed.
>>
>> Signed-off-by: Robert Yang <liezhi.yang@windriver.com>
>> ---
>>   meta/classes-global/package_rpm.bbclass | 6 ++++--
>>   meta/lib/oeqa/runtime/cases/dnf.py      | 8 --------
>>   2 files changed, 4 insertions(+), 10 deletions(-)
>>
>> diff --git a/meta/classes-global/package_rpm.bbclass b/meta/classes-global/package_rpm.bbclass
>> index f4dd779a52..526ac57982 100644
>> --- a/meta/classes-global/package_rpm.bbclass
>> +++ b/meta/classes-global/package_rpm.bbclass
>> @@ -421,12 +421,14 @@ python write_specfile () {
>>           spec_preamble_bottom.append('')
>>   
>>           # Now process scriptlets
>> -        if splitrpreinst:
>> +        # The libc6 shouldn't have %pre or %post to avoid circular dependency
>> +        libc6 = '%slibc6' % (d.getVar('MLPREFIX') or '')
>> +        if splitrpreinst and splitname != libc6:
>>               spec_scriptlets_bottom.append('%%pre -n %s' % splitname)
>>               spec_scriptlets_bottom.append('# %s - preinst' % splitname)
>>               spec_scriptlets_bottom.append(splitrpreinst)
>>               spec_scriptlets_bottom.append('')
>> -        if splitrpostinst:
>> +        if splitrpostinst and splitname != libc6:
>>               spec_scriptlets_bottom.append('%%post -n %s' % splitname)
>>               spec_scriptlets_bottom.append('# %s - postinst' % splitname)
>>               spec_scriptlets_bottom.append(splitrpostinst)
> 
> I'm not happy about coding a "libc6" reference into the generic package
> class. It does make me wonder if there is a similar issue with musl for
> example and we'd end up with a longer hardcoded list.
> 
> Could/shouldn't we just not add that postinst for libc6? You could for
> example force the value of pkg_postinst:libc6 during the packaging
> process and empty the postinst instead.

I think you meant add the following lines in glibc-package.inc:

pkg_postinst:glibc () {
     echo
}

We need an "echo" here, otherwise, meta/lib/oe/package.py would add the
"#!/bin/sh" to the postinst:

         if needs_ldconfig:
             bb.debug(1, 'adding ldconfig call to postinst for %s' % pkg)
             postinst = d.getVar('pkg_postinst:%s' % pkg)
             if not postinst:
                 postinst = '#!/bin/sh\n'
             postinst += d.getVar('ldconfig_postinst_fragment')
             d.setVar('pkg_postinst:%s' % pkg, postinst)

But this still doesn't work, because rpm will add the /bin/sh as the default
interpreter when not specified. And currently, meta/lib/oe/package.py
add the ldconfig_postinst_fragment automatically when find so files,
that the root cause, so how about we add a value like:

NEEDS_LDCONFIG:<pkg> = "0" or "1"

Then we can add the following line to glibc-package.inc to skip that.
NEEDS_LDCONFIG:libc6 = "0"

And also we need update meta/lib/oe/package.py to check the
NEEDS_LDCONFIG:pkg.

> 
> Also the commit message explains the problem with the postinst but not
> really why the preinst is a problem? What does that contain which is
> problematic?
> 

It is similar to posint, the prinst has a harder situation than postinst.

// Robert


> Cheers,
> 
> Richard