Avoiding the stage-2 interactive shell in runInLinuxImage

charles-dyfis-net · October 30, 2018, 3:45am

I’m trying to understand how an invocation of runInLinuxImage such as vmTools.buildRPM actually kicks off the included build phases, when invoked with a diskImage previously constructed with runInLinuxVM.

runInLinuxVM hardcodes use of qemuCommandLinux, which references stage2Init.
Only one stage2Init definition appears to exist, and that definition is built to run only once (dropping to an interactive shell if the VM is re-invoked).

However, runInLinuxImage – as used by buildRPM – takes an arbitrary derivation, and runs it inside the guest… but how? That guest, when it runs, starts up with an init= parameter telling it to run a command that drops to an interactive shell (which I don’t see an obvious way to interact with from nix-build, though that’s a rather secondary concern), rather than invoking the builder in question.

Is there documentation available on how this is expected to work?

Mic92 · October 30, 2018, 3:20pm

buildRPM overrides builder of the the runInLinuxVM functions, which is then executed within the VM on this line:

github.com

NixOS/nixpkgs/blob/master/pkgs/build-support/vm/default.nix#L180


      
          out="$1"
          cd "$NIX_BUILD_TOP"
          
          
if ! test -e /bin/sh; then
            ${coreutils}/bin/mkdir -p /bin
            ${coreutils}/bin/ln -s ${bash}/bin/sh /bin/sh
          fi
          
          
# Set up automatic kernel module loading.
          export MODULE_DIR=${kernel}/lib/modules/
          ${coreutils}/bin/cat <<EOF > /run/modprobe
          #! ${bash}/bin/sh
          export MODULE_DIR=$MODULE_DIR
          exec ${kmod}/bin/modprobe "\$@"
          EOF
          ${coreutils}/bin/chmod 755 /run/modprobe
          echo /run/modprobe > /proc/sys/kernel/modprobe
          
          
# For debugging: if this is the second time this image is run,
          # then don't start the build again, but instead drop the user into
          # an interactive shell.

I don’t think there is documentation on those rather low-level implementation details.

charles-dyfis-net · October 30, 2018, 3:49pm

Ahh! What I missed was this critical line in fillDiskWithRPMs:

rm /mnt/.debug

Thank you for your time and patience.