Passing app arguments on the kernel command line quietly mangles some of them

[pocopepe]

For Linux guests (including the default urunit path) we hand the application its arguments through the kernel command line — init=/urunit -- <args>, joined with spaces. When kernel splits on white spaces and treats -- specially. Anything that does't fit those rules exact gets changed silently, when the guest boots fine,

A few shapes I ran into,

• An empty argument disappears, since everything is glued together with spaces — ["a", "", "b"] reaches the app as a, b.
• An argument that's literally -- cuts off everything after it — ["a", "--", "b", "c"] arrives as just a.
• A long argument list gets chopped at 2048 bytes, the last argument is cut mid value and anything past it is gone.
• Go past 32 arguments and the guest doesn't start at all, it panics, Kernel panic - not syncing: Too many boot init vars.

I don't think these are four separate parsing bugs so much as one thing: the command line was never really able to carry an arbitrary argument list.

[pocopepe]

I feel mirroring what freebsd guests do (passing arguments through urunit config) can fix this. This issue's been opened as a superset of #897

[indrajithek]

I hit the -- case from this issue while running Argo Workflows on urunc v0.8.0 (Linux guest, QEMU, virtiofs rootfs).

Arguments

Argo rewrites the main container's command to argoexec emissary -- <user command>. urunc adds its own -- after init=, so the guest gets:

init=/var/run/argo/argoexec -- emissary --loglevel info --log-format text --gloglevel 0 -- python /argo/staging/script

Everything after the second -- is dropped. emissary starts with no command and panics:

panic: runtime error: slice bounds out of range [1:0] cmd/argoexec/commands/emissary.go:99

Environment variables
In v0.8.0, env goes into the urunit config only when the container's argv[0] contains urunit:

// pkg/unikontainers/unikernels/linux.go:237
l.InitrdConf = strings.Contains(l.App, "urunit")

Otherwise every variable is added to the kernel command line (linux.go:109). Under Argo, argv[0] is /var/run/argo/argoexec, so the config file is not used. In my run, the kernel command line had no URUNIT_CONFIG= and 22 environment variables, so we can easily exceed the limit of 32 in more complicated scenarios. This is also after keeping headroom by the flag enableServiceLinks: false

So two things: for the argument matter, we can do what @pocopepe suggested. For the second, I saw (actually, my Claude agent saw) the feat/unchanged_containers branch always boots urunit, so env already goes through the config there, while args still go on the command line (rdinit=/init -- <cmd>). Would it make sense to base the fix on that branch? @cmainas

Plus, I'm happy to work on this, as this has to be sorted to proceed with Argo integration work, which I'm currently doing. Plus, I don't feel we need to open a separate issue on this