Skip to contentSkip to conversation

bash: Drop /bin/sh from RPROVIDES

RFC PATCH In review Marek Vasut 4 participants 11 messages Jan 23, 2025 – Jan 24, 2025

lore
Share & tools
lore
#0 lore raw

RFC PATCH 1/1 bash: Drop /bin/sh from RPROVIDES

Remove /bin/sh from bash RPROVIDES as this has a side-effect which confuses rpm package manager when also busybox provides /bin/sh and base-files depend on /bin/sh . The problem is broken down below.

First, bash depends on base-files and bash pkg_postinst must run after base-files was installed, because it requires /etc/shells provided by base-files to be in place.

Second, base-files depends on /bin/sh, which is provided by either bash or busybox in this case. This is the actual problem here, if bash is selected as /bin/sh provider, then there is cyclic dependency between bash and base-files, and that confuses dnf which may install the packages in the wrong order, bash first and base-files second .

To make this worse, if busybox is also /bin/sh provider, it can and does happen that some systems pick busybox as the /bin/sh provider, while others pick bash as the /bin/sh provider, and that cyclic dependency does not always appear.

Attempt to break this dependency, drop RPROVIDES /bin/sh from bash recipe to always force base-files to pick /bin/sh from busybox . (I am really unsure about this approach)

---
Cc: Alexander Kanavin <[email protected]>
Cc: Alexandre Belloni <[email protected]>
Cc: Richard Purdie <[email protected]>

---
 meta/recipes-extended/bash/bash.inc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/recipes-extended/bash/bash.inc +1 -1
index 634209c9115..6a061b8a612 100644
@@ -134,4 +134,4 @@ FILES:${PN}-loadable += "${libdir}/bash/*"
 
 # Limit the RPROVIDES here to class target so that if usrmerge is enabled for nativesdk, it does not
 # include host system paths in /bin/
-RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
+RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"
Signature · 1 line
2.45.2
#1 lore raw
Quoted 38 lines · Marek Vasut via lists.openembedded.org

Remove /bin/sh from bash RPROVIDES as this has a side-effect which confuses rpm package manager when also busybox provides /bin/sh and base-files depend on /bin/sh . The problem is broken down below.

First, bash depends on base-files and bash pkg_postinst must run after base-files was installed, because it requires /etc/shells provided by base-files to be in place.

Second, base-files depends on /bin/sh, which is provided by either bash or busybox in this case. This is the actual problem here, if bash is selected as /bin/sh provider, then there is cyclic dependency between bash and base-files, and that confuses dnf which may install the packages in the wrong order, bash first and base-files second .

To make this worse, if busybox is also /bin/sh provider, it can and does happen that some systems pick busybox as the /bin/sh provider, while others pick bash as the /bin/sh provider, and that cyclic dependency does not always appear.

Attempt to break this dependency, drop RPROVIDES /bin/sh from bash recipe to always force base-files to pick /bin/sh from busybox . (I am really unsure about this approach)

---
Cc: Alexander Kanavin <[email protected]>
Cc: Alexandre Belloni <[email protected]>
Cc: Richard Purdie <[email protected]>

---
 meta/recipes-extended/bash/bash.inc | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
meta/recipes-extended/bash/bash.inc +0 -0
index 634209c9115..6a061b8a612 100644

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 5 lines
meta/recipes-extended/bash/bash.inc +1 -1
@@ -134,4 +134,4 @@ FILES:${PN}-loadable += "${libdir}/bash/*"
 
 # Limit the RPROVIDES here to class target so that if usrmerge is enabled for nativesdk, it does not
 # include host system paths in /bin/
-RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
+RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"

Does this now mean that system will always need busybox ? some distros may not want that and simply use
other variants to fill in the shoes. I wonder if shell choice could be a distro option or image feature which can then
help sort this out ?

Quoted 12 lines

--
2.45.2

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#210212): https://lists.openembedded.org/g/openembedded-core/message/210212
Mute This Topic: https://lists.openembedded.org/mt/110778651/1997914
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#2 lore raw
Quoted 41 lines · Khem Raj
Quoted 39 lines · Marek Vasut via lists.openembedded.org

Remove /bin/sh from bash RPROVIDES as this has a side-effect which confuses rpm package manager when also busybox provides /bin/sh and base-files depend on /bin/sh . The problem is broken down below.

First, bash depends on base-files and bash pkg_postinst must run after base-files was installed, because it requires /etc/shells provided by base-files to be in place.

Second, base-files depends on /bin/sh, which is provided by either bash or busybox in this case. This is the actual problem here, if bash is selected as /bin/sh provider, then there is cyclic dependency between bash and base-files, and that confuses dnf which may install the packages in the wrong order, bash first and base-files second .

To make this worse, if busybox is also /bin/sh provider, it can and does happen that some systems pick busybox as the /bin/sh provider, while others pick bash as the /bin/sh provider, and that cyclic dependency does not always appear.

Attempt to break this dependency, drop RPROVIDES /bin/sh from bash recipe to always force base-files to pick /bin/sh from busybox . (I am really unsure about this approach)

---
Cc: Alexander Kanavin <[email protected]>
Cc: Alexandre Belloni <[email protected]>
Cc: Richard Purdie <[email protected]>

---
  meta/recipes-extended/bash/bash.inc | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
meta/recipes-extended/bash/bash.inc +0 -0
index 634209c9115..6a061b8a612 100644
@@ -134,4 +134,4 @@ FILES:${PN}-loadable += "${libdir}/bash/*"
 
Quoted 6 lines

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 4 lines
meta/recipes-extended/bash/bash.inc +1 -1
  # Limit the RPROVIDES here to class target so that if usrmerge is enabled for nativesdk, it does not
  # include host system paths in /bin/
-RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
+RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"

Does this now mean that system will always need busybox ?

The system will require some other /bin/sh provider than bash .

I suspect there will always be one , but maybe I am wrong and there are systems without busybox and with bash only (probably yes) ?

Quoted 5 lines

some distros
may not want that and simply use
other variants to fill in the shoes. I wonder if shell choice could be a distro option or image feature which can then
help sort this out ?

I don't think this specific fix is the correct solution, I think what I need to figure out is how to make sure base-files is installed before bash, even if there is this cyclic dependency which confuses rpm/dnf into installing the two packages in arbitrary order:

base-files -depends_on-> /bin/sh -depends_on-> bash (the /bin/sh provider)

bash -depends_on-> base-files

I still have to check the input from Martin Jansa , there was some hint which might be the better fix for this.

Any ideas ?

#3 lore raw

I wonder if we can just drop /etc/shells altogether from base-files?

Alex

Quoted 89 lines · Marek Vasut via lists.openembedded.org
Quoted 35 lines · Khem Raj
Quoted 33 lines · Marek Vasut via lists.openembedded.org

Remove /bin/sh from bash RPROVIDES as this has a side-effect which confuses rpm package manager when also busybox provides /bin/sh and base-files depend on /bin/sh . The problem is broken down below.

First, bash depends on base-files and bash pkg_postinst must run after base-files was installed, because it requires /etc/shells provided by base-files to be in place.

Second, base-files depends on /bin/sh, which is provided by either bash or busybox in this case. This is the actual problem here, if bash is selected as /bin/sh provider, then there is cyclic dependency between bash and base-files, and that confuses dnf which may install the packages in the wrong order, bash first and base-files second .

To make this worse, if busybox is also /bin/sh provider, it can and does happen that some systems pick busybox as the /bin/sh provider, while others pick bash as the /bin/sh provider, and that cyclic dependency does not always appear.

Attempt to break this dependency, drop RPROVIDES /bin/sh from bash recipe to always force base-files to pick /bin/sh from busybox . (I am really unsure about this approach)

---
Cc: Alexander Kanavin <[email protected]>
Cc: Alexandre Belloni <[email protected]>
Cc: Richard Purdie <[email protected]>

---
  meta/recipes-extended/bash/bash.inc | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 1 line

Review

Quoted 1 line
(quoted diff) +0 -0

b/meta/recipes-extended/bash/bash.inc

Quoted 6 lines

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 135

Quoted 6 lines

index 634209c9115..6a061b8a612 100644

meta/recipes-extended/bash/bash.inc +0 -0
@@ -134,4 +134,4 @@ FILES:${PN}-loadable += "${libdir}/bash/*"
 
  # Limit the RPROVIDES here to class target so that if usrmerge is

enabled for nativesdk, it does not

Quoted 2 lines

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 2 lines
meta/recipes-extended/bash/bash.inc +0 -1
  # include host system paths in /bin/
-RPROVIDES:${PN}:append:class-target = "

${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"

Quoted 1 line

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 1 line
meta/recipes-extended/bash/bash.inc +1 -0
+RPROVIDES:${PN}:append:class-target = "

${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"

Quoted 2 lines

Does this now mean that system will always need busybox ?

The system will require some other /bin/sh provider than bash .

I suspect there will always be one , but maybe I am wrong and there are systems without busybox and with bash only (probably yes) ?

Quoted 5 lines

some distros
may not want that and simply use
other variants to fill in the shoes. I wonder if shell choice could be a distro option or image feature which can then
help sort this out ?

I don't think this specific fix is the correct solution, I think what I need to figure out is how to make sure base-files is installed before bash, even if there is this cyclic dependency which confuses rpm/dnf into installing the two packages in arbitrary order:

base-files -depends_on-> /bin/sh -depends_on-> bash (the /bin/sh provider)

bash -depends_on-> base-files

I still have to check the input from Martin Jansa , there was some hint which might be the better fix for this.

Any ideas ?

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group. View/Reply Online (#210223):
https://lists.openembedded.org/g/openembedded-core/message/210223
Mute This Topic: https://lists.openembedded.org/mt/110778651/1686489
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [ [email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#4 lore raw
Quoted 1 line · Alexander Kanavin via lists.openembedded.org

I wonder if we can just drop /etc/shells altogether from base-files?

/etc/shells is used by pam as one of it's multiple validation steps. If the shell is not listed there then remote login can be denied. So the existence of /etc/shells is definitely required. The file doesn't need to come from base-files (but usually should).

With that said, "something" needs to provide /bin/sh. Just dropping /bin/sh from a shell, by itself, it's the right answer as there are ABSOLUTELY systems out there where '/bin/bash' _is_ the only shell installed on the system.

Just like we have busybox only systems providing /bin/sh...

or dash based systems where it is the only thing providing /bin/sh...

I'm more inclined to say that base-files depending on /bin/sh is the error. AFAIK nothing in base-files should require /bin/sh (or are there post-install scripts?). Once that is resolved, then various POSIX /bin/sh providers can each provide it -- using update-alternatives to have the 'best' matching providing the right /bin/sh link. (This is why update-alternatives [in the past at least] run directly and didn't require a shell to execute. I don't know that we have the ability to do this with our multi-package support approach though.)

So essential install order may be something like:

base-files
libc (libraries only)
update-alternatives
<a shell>
- shell calls update-alternatives to register it as /bin/sh, as well as add
this as a valid shell to /etc/shells.

End result (for bash)

/bin/sh -> /bin/bash

/etc/shell contains:
/bin/sh
/bin/bash

(Note, circular dependencies for some base components have been considered normal in the workstation world. They work around this by use the 'install' to temporarily provide the /bin/sh and some other resources until they can be installed into the chroot. This really isn't anything different from what we're doing, we expect some host-system (native) items to be available in order to populate the image.)

--Mark

Quoted 98 lines

Alex

On Fri 24. Jan 2025 at 16.26, Marek Vasut via lists.openembedded.org <http://lists.openembedded.org> <[email protected] <mailto:[email protected]>> wrote:

On 1/23/25 10:18 PM, Khem Raj wrote:
> On Thu, Jan 23, 2025 at 12:12 PM Marek Vasut via
> lists.openembedded.org <http://lists.openembedded.org>
<[email protected] <mailto:[email protected]>> wrote:
>>
>> Remove /bin/sh from bash RPROVIDES as this has a side-effect which >> confuses rpm package manager when also busybox provides /bin/sh and >> base-files depend on /bin/sh . The problem is broken down below. >>
>> First, bash depends on base-files and bash pkg_postinst must run >> after base-files was installed, because it requires /etc/shells >> provided by base-files to be in place.
>>
>> Second, base-files depends on /bin/sh, which is provided by either >> bash or busybox in this case. This is the actual problem here, if >> bash is selected as /bin/sh provider, then there is cyclic dependency >> between bash and base-files, and that confuses dnf which may install >> the packages in the wrong order, bash first and base-files second . >>
>> To make this worse, if busybox is also /bin/sh provider, it can and >> does happen that some systems pick busybox as the /bin/sh provider, >> while others pick bash as the /bin/sh provider, and that cyclic >> dependency does not always appear.
>>
>> Attempt to break this dependency, drop RPROVIDES /bin/sh from bash >> recipe to always force base-files to pick /bin/sh from busybox . >> (I am really unsure about this approach)
>>
>> Signed-off-by: Marek Vasut <[email protected] <mailto:[email protected]>> >> ---
>> Cc: Alexander Kanavin <[email protected] <mailto:[email protected]>> >> Cc: Alexandre Belloni <[email protected]
<mailto:[email protected]>>
>> Cc: Richard Purdie <[email protected]
<mailto:[email protected]>>
>> ---
>>   meta/recipes-extended/bash/bash.inc | 2 +-
>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/meta/recipes-extended/bash/bash.inc
b/meta/recipes-extended/bash/bash.inc
>> index 634209c9115..6a061b8a612 100644
>> --- a/meta/recipes-extended/bash/bash.inc
>> +++ b/meta/recipes-extended/bash/bash.inc
>> @@ -134,4 +134,4 @@ FILES:${PN}-loadable += "${libdir}/bash/*" >>
>>   # Limit the RPROVIDES here to class target so that if usrmerge is
enabled for nativesdk, it does not
>>   # include host system paths in /bin/
>> -RPROVIDES:${PN}:append:class-target = "
${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
>> +RPROVIDES:${PN}:append:class-target = "
${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"
>
> Does this now mean that system will always need busybox ?

The system will require some other /bin/sh provider than bash .

I suspect there will always be one , but maybe I am wrong and there are systems without busybox and with bash only (probably yes) ?

>  some distros
> may not want that and simply use
> other variants to fill in the shoes. I wonder if shell choice could be > a distro option or image feature which can then
> help sort this out ?
I don't think this specific fix is the correct solution, I think what I need to figure out is how to make sure base-files is installed before bash, even if there is this cyclic dependency which confuses rpm/dnf into installing the two packages in arbitrary order:

base-files -depends_on-> /bin/sh -depends_on-> bash (the /bin/sh provider)

bash -depends_on-> base-files

I still have to check the input from Martin Jansa , there was some hint which might be the better fix for this.

Any ideas ?

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#210242): https://lists.openembedded.org/g/openembedded-core/message/210242
Mute This Topic: https://lists.openembedded.org/mt/110778651/3616948
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#6 lore raw
Quoted 10 lines · Alexander Kanavin via lists.openembedded.org
Quoted 3 lines · Mark Hatle

I'm more inclined to say that base-files depending on /bin/sh is the error. AFAIK nothing in base-files should require /bin/sh (or are there post-install scripts?).

There is a pre-install script, but it's for a long obsolete issue:

https://git.yoctoproject.org/poky/commit/?id=6f03a8d7ae383d63bfed8b85c06a697b86719004

Let's just drop it from the recipe, and then base-files won't depend on /bin/sh.

Agreed, base-files should be doing this kind of thing anyway.

--Mark

Quoted 13 lines

Alex

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.
View/Reply Online (#210248): https://lists.openembedded.org/g/openembedded-core/message/210248
Mute This Topic: https://lists.openembedded.org/mt/110778651/3616948
Group Owner: [email protected]
Unsubscribe: https://lists.openembedded.org/g/openembedded-core/unsub [[email protected]]
-=-=-=-=-=-=-=-=-=-=-=-

#7 lore raw
Quoted 7 lines · Mark Hatle
Quoted 5 lines · Alexander Kanavin via lists.openembedded.org
Quoted 3 lines · Mark Hatle

I'm more inclined to say that base-files depending on /bin/sh is the error. AFAIK nothing in base-files should require /bin/sh (or are there post-install scripts?).

Quoted 7 lines
Quoted 5 lines

There is a pre-install script, but it's for a long obsolete issue:

https://git.yoctoproject.org/poky/commit/? id=6f03a8d7ae383d63bfed8b85c06a697b86719004

Let's just drop it from the recipe, and then base-files won't depend on /bin/sh.

Agreed, base-files should be doing this kind of thing anyway.

This works too, v2 sent, thanks !

#8 lore raw
In reply to Mark Hatle
Quoted 7 lines · Mark Hatle
Quoted 1 line · Alexander Kanavin via lists.openembedded.org

I wonder if we can just drop /etc/shells altogether from base-files?

/etc/shells is used by pam as one of it's multiple validation steps.  If the shell is not listed there then remote login can be denied.  So the existence of /etc/shells is definitely required.  The file doesn't need to come from base-files (but usually should).

Quoted 8 lines

With that said, "something" needs to provide /bin/sh.  Just dropping / bin/sh from a shell, by itself, it's the right answer as there are ABSOLUTELY systems out there where '/bin/bash' _is_ the only shell installed on the system.

Just like we have busybox only systems providing /bin/sh...

or dash based systems where it is the only thing providing /bin/sh...

I'm more inclined to say that base-files depending on /bin/sh is the error. AFAIK nothing in base-files should require /bin/sh (or are there post-install scripts?).

There is a pre-install script:

$ git grep bin.sh meta/recipes-core/base-files/base-files_3.0.14.bb
meta/recipes-core/base-files/base-files_3.0.14.bb:    #!/bin/sh -e
..
    #!/bin/sh -e
    if [ x"$D" = "x" ]; then
        if [ -h "/var/lock" ]; then
            # Remove the symbolic link
            rm -f /var/lock
        fi

if [ -h "/run" ]; then
# Remove the symbolic link
rm -f /run
fi
fi
..

Quoted 20 lines

Once that is resolved, then various POSIX / bin/sh providers can each provide it -- using update-alternatives to have the 'best' matching providing the right /bin/sh link.  (This is why update-alternatives [in the past at least] run directly and didn't require a shell to execute.  I don't know that we have the ability to do this with our multi-package support approach though.)

So essential install order may be something like:

base-files
libc (libraries only)
update-alternatives
<a shell>
 - shell calls update-alternatives to register it as /bin/sh, as well
as add this as a valid shell to /etc/shells.

End result (for bash)

/bin/sh -> /bin/bash

/etc/shell contains:
  /bin/sh
  /bin/bash

(Note, circular dependencies for some base components have been considered normal in the workstation world.  They work around this by use the 'install' to temporarily provide the /bin/sh and some other resources until they can be installed into the chroot.   This really isn't anything different from what we're doing, we expect some host- system (native) items to be available in order to populate the image.)

[...]

#9 lore raw
In reply to Marek Vasut

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 4 lines · Marek Vasut via lists.openembedded.org

# Limit the RPROVIDES here to class target so that if usrmerge is enabled for nativesdk, it does not

 # include host system paths in /bin/
meta/recipes-extended/bash/bash.inc +1 -1
-RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
+RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"

This was added for a reason (that you can find in git history), if we partially remove it, then the reason will come back and bite us.

What I'd like to understand is how bash's postinst is actually failing when /etc/shells is not there yet:

grep -q "^${base_bindir}/bash$" $D${sysconfdir}/shells || echo ${base_bindir}/bash >> $D${sysconfdir}/shells

Basically you either need to show me how to reproduce the issue, or show an error message you're getting, or an incorrect outcome on the rootfs.

Alex

#10 lore raw
Quoted 10 lines · Alexander Kanavin

meta/recipes-extended/bash/bash.inc FILES:${PN}-loadable += "${libdir}/bash/*" · line 137

Quoted 4 lines · Marek Vasut via lists.openembedded.org

# Limit the RPROVIDES here to class target so that if usrmerge is enabled for nativesdk, it does not

  # include host system paths in /bin/
meta/recipes-extended/bash/bash.inc +1 -1
-RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/sh /bin/bash', '', d)}"
+RPROVIDES:${PN}:append:class-target = " ${@bb.utils.contains('DISTRO_FEATURES', 'usrmerge', '/bin/bash', '', d)}"

This was added for a reason (that you can find in git history), if we partially remove it, then the reason will come back and bite us.

That's why there is [RFC] tag in subject.

Quoted 5 lines

What I'd like to understand is how bash's postinst is actually failing when /etc/shells is not there yet:

grep -q "^${base_bindir}/bash$" $D${sysconfdir}/shells || echo ${base_bindir}/bash >> $D${sysconfdir}/shells

Grep will complain about the missing file /etc/shells .

I can totally imagine we can take the easy way out here for all the postinst scripts and do something like :

"
# Make sure $D${sysconfdir}/shells exists
touch $D${sysconfdir}/shells
# Remove previous bash entry (if any)
sed -i '/^${base_bindir}/bash$/ d' $D${sysconfdir}/shells # Add new bash entry , now surely a singular one echo ${base_bindir}/bash >> $D${sysconfdir}/shells "

But I have the bad feeling that this would be hiding some deeper issue.

Quoted 1 line

Basically you either need to show me how to reproduce the issue

I cannot, because so far this only happens in internal CI. I am not able to reproduce it on multiple local build machines.

Quoted 2 lines

, or
show an error message you're getting

The grep -q fails on missing /etc/shells , because base-files which was supposed to add it was not installed yet.

Quoted 2 lines

, or an incorrect outcome on the
rootfs.

That incorrect rootfs will only contain /etc/shells from base-files , the base-files which were installed AFTER bash was installed already , which will be missing the bash entry because that was never added because the bash postinst ran before base-files was installed .

(and I still plan to reply to Khem and also got some hint from Martin Jansa, which I have yet to explore)