Re: [PATCH v2 09/15] bootlinux: fix os detection for 9p build dependency installation

On 08/11/2025 17.11, Chuck Lever wrote:
> On 11/7/25 6:20 PM, Daniel Gomez wrote:
>>>> +- name: Red Hat-specific setup (controller node)
>>>>    ansible.builtin.import_tasks: redhat/main.yml
>>>>    when:
>>>>      - distro_redhat_based|bool
>>>> +    - bootlinux_9p|bool
>>>> +
>>>> +- name: Red Hat-specific setup (builder node)
>>>> +  ansible.builtin.import_tasks: redhat/main.yml
>>>> +  when:
>>>> +    - bootlinux_builder|bool
>>>> +    - ansible_os_family == "RedHat"
>>>>
>>>> IIUC, this should return the correct ansible_os_family value when builder target
>>>> is selected because the task is executed on the guest with --limit baseline:dev.
>>> Hi Daniel -
>>>
>>> None of these tasks will execute if bootlinux_targets == true, AFAICT.
>> Sorry, I missed that target. I meant to add that condition too when we use
>> ansible_os_family case.
>>
>>> Why can't this work based on which host the install-deps tasks are
>>> running on? That was the original design I had in mind, and seems
>>> most Ansible-like.
>> I can't replicate the issue. I think we can revert the patch. Here my tests:
>>
>> fedora: https://github.com/linux-kdevops/kdevops/actions/runs/19183690762
>> debian: https://github.com/linux-kdevops/kdevops/actions/runs/19183543710
> 
> OK, let me try to be less terse.  :-)
> 
> I'd like to see this instance of install-deps look like most of the
> others: basically, key off only the ansible_os_family settings. I think
> that will be more maintainable.

100%.

> 
> Those settings, however, need to properly reflect the environment where
> install-deps is running. For the 9p case, it will run on both the
> controller (where the kernel is built) and the targets (where the kernel
> is installed); and I think bootlinux even runs the install-deps steps
> twice for 9p to deal exactly with that duplication.

I think you are referring to "Install essential build tools for 9P builds" task
in main.yml + "Install dependencies to build the Linux kernel" in build/9p.yml.
The first one executes on the guests while the second one is delegated to the
localhost.

> 
> The same is true for the "separate builder" case: the kernel builder
> runs on one target node, and the built artifacts are installed on other
> target nodes; and bootlinux has to ensure the correct dependencies are
> installed everywhere they are needed.

This is referring to the task "Install dependencies to build and install
the Linux kernel" in main.yml. Which is skipped for 9p (but included through
build/9p.yml + delegated_to).

> 
> The way I tried to deal with that was to use "delegate_to:". But I think
> you might have advanced a different mechanism, and I missed the
> subtleties of that. Or, maybe providing execution locale information on
> the Ansible command line just isn't appropriate in these cases. I
> haven't thought it through.

This is where it gets tricky. I've just found out that delegated tasks to a
host that is not included in the scope (such as localhost), do not have gathered
facts.

https://docs.ansible.com/ansible/latest/playbook_guide/playbooks_delegation.html#delegating-facts

That explains why ansible_os_family returns the guest fact (e.g. RedHat
guest), not the localhost fact (e.g. Debian host) when we import_tasks:
install-deps/main.yml in build/9p.yml.

The use of distro_yada_based in patch "bootlinux: fix os detection for 9p build
dependency installation" allows to access the fact (os family) via Kconfig for
the localhost. But breaks builder and targets sub-roles.

> 
> I'm very happy to help drill into the 9p case that was not working for
> you, instead of standing nearby lobbing uninformed suggestions.
> Meanwhile, I'll keep a patch that reverts 0e9c8ac8bc68 in my private
> tree.
> 

Thanks Chuck. I think the correct approach is to revert patch 0e9c8ac8
"bootlinux: fix os detection for 9p build dependency installation" upstream,
while we try to sort this out in 9p.

For the revert part, please push:
Acked-by: Daniel Gomez <[email protected]>

FYI, I'm already working on a solution using delegate_facts + hostvars as
described in the Ansible documentation link above. That seems to be the right
approach. I'll send something soon for testing.