bug#30948: [PATCH core-updates] guix: Reap finished child processes in build containers.

[Maxim Cournoyer <maxim.cournoyer@gmail.com>]

Hi,

Ludovic Courtès <ludo@gnu.org> writes:

> Maxim Cournoyer <maxim.cournoyer@gmail.com> skribis:
>
>> Ludovic Courtès <ludo@gnu.org> writes:
>
> [...]
>
>>> If I write:
>>>
>>>   (gexp->derivation "foo" #~(mkdir #$output))
>>>
>>> I can be sure that my derivation depends on nothing but (default-guile).
>>> This is important for tests, but also to make sure we can use this
>>> primitive everywhere—if it pulled in the Shepherd, I wouldn’t be able to
>>> use to build glibc, because there’d be a cycle.
>>
>> I was not suggesting to pull in extra dependencies such as Shepherd, but
>> to weave the to-be-added signal handling logic at a much lower level.
>> One idea could be to arrange so that the correct signal handlers always
>> get installed for any Guile code running in the build side (I'm not sure
>> how, but perhaps by adjusting the gexp "compiler"?).
>>
>> The handlers could be defined in (guix build signal-handling) or
>> similar.  Users wouldn't need to explicitly import the module and
>> install its signal handlers, that'd be taken care of automatically, all
>> the time.
>>
>> Does that sound feasible?
>
> Not like this: the imported-modules derivation for (guix build
> signal-handling) would be a dependency in themselves.

I see a couple of options for the lowest place to inject the minimal
signal handling of a PID.

1. In Guile itself.  We could make it detect when it's running as PID 1
and then set up the required signal handling.  This is apparently what
Bash does, a peculiarity exploited by NixOS (they launch their builder
scripts via Bash, which is PID 1 and takes care of reaping the dead
processes)

2. In a Guile wrapper.  Instead of running Guile directly in the
container, guix-daemon would run it through a wrapper that acts as PID 1.
This would make it a tool comparable to dumb-init [0] or tini [1],
except written in Scheme.

[0] https://github.com/Yelp/dumb-init/
[1] https://github.com/krallin/tini

If we implement 1, it'd make Guile potentially useful as a wrapper
itself to launch scripts in containerized environment (the same as
tini), and it alleviates any integration overhead for us, so I find it
attractive.

What do you think?

For further reading, see [2], which I found interesting.

[2]  https://medium.com/hackernoon/my-process-became-pid-1-and-now-signals-behave-strangely-b05c52cc551c

-- 
Thanks,
Maxim