fix(ci): move Build workflow off the deprecated ubuntu-20.04 runner

.github/workflows/build.yml

@@ -10,12 +10,17 @@ on:
 
 jobs:
   build:
-    runs-on: ubuntu-20.04
+    runs-on: ubuntu-22.04
+    env:
+      # BuildKit mounts each RUN step's rootfs with nosuid, which breaks `sudo` inside
+      # ev3dev-service-control's Dockerfile (still needed on the newer Docker Engine that ships
+      # with this runner image, even though that Dockerfile itself is unchanged). The legacy
+      # builder doesn't impose this restriction.
+      DOCKER_BUILDKIT: "0"
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
         with:
           submodules: recursive
-      - uses: satackey/action-docker-layer-caching@v0.0.8
       - name: install brickstrap dependencies
         run: |
           sudo apt-get update
@@ -47,7 +52,7 @@ jobs:
         run: sudo chmod 755 /boot/vmlinuz*
       - name: build image
         run: ./build_image.sh
-      - uses: actions/upload-artifact@v2
+      - uses: actions/upload-artifact@v4
         if: ${{ github.event_name == 'workflow_dispatch' }}
         with:
           name: image

brickstrap.sh

@@ -201,9 +201,25 @@ function brickstrap_create_image()
     # orphan_file, compat bit 12), which systemd-fsck-root.service treats as fatal and drops the
     # whole boot into emergency mode. Explicitly excluding the newer features keeps the
     # filesystem on a feature set the image's own e2fsck actually understands. Harmless on an
-    # older host (e.g. this project's current ubuntu-20.04 CI runner) whose mke2fs never enabled
-    # these features anyway - excluding an already-absent feature is a no-op.
-    ROOTFS_MKE2FS_FEATURES="^orphan_file,^metadata_csum_seed,^fast_commit,^sparse_super2,^stable_inodes"
+    # older host whose mke2fs never enabled these features anyway - excluding an already-absent
+    # feature is normally a no-op.
+    #
+    # That assumption broke moving to a newer CI runner image: its mke2fs (1.46.5) is old enough
+    # to not recognize one of these five names at all (added in a later e2fsprogs release), and
+    # unlike "silently no-op on an absent feature", an unrecognized name makes mke2fs reject the
+    # entire -O list as invalid, aborting the whole image build. Rather than hardcode which name(s)
+    # are safe for whichever e2fsprogs happens to be installed, detect it directly: keep only the
+    # names this host's own mke2fs actually accepts negating, on a real scratch filesystem via the
+    # same -O parser guestfish's appliance will use (built from this same host's packages).
+    ROOTFS_MKE2FS_FEATURES=""
+    for feat in orphan_file metadata_csum_seed fast_commit sparse_super2 stable_inodes; do
+        feat_test_img="$(mktemp)"
+        truncate -s 16M "$feat_test_img"
+        if mke2fs -F -q -O "^$feat" "$feat_test_img" >/dev/null 2>&1; then
+            ROOTFS_MKE2FS_FEATURES="${ROOTFS_MKE2FS_FEATURES:+$ROOTFS_MKE2FS_FEATURES,}^$feat"
+        fi
+        rm -f "$feat_test_img"
+    done
 
     guestfish -N "$BRICKSTRAP_IMAGE_FILE_NAME"=disk:$BRICKSTRAP_IMAGE_FILE_SIZE -- \
         part-init /dev/sda mbr : \

build_control_panel.sh

@@ -7,6 +7,24 @@ REPO_DIR="ev3dev-service-control"
 IMAGE_NAME="sourceacademy/ev3-vala-compiler"
 
 cd "$SCRIPT_DIR/$REPO_DIR"
+# The ev3dev/debian-stretch-armel-cross base image's own /usr/bin/sudo is missing its setuid bit
+# ("effective uid is not 0") - a defect in that specific image, not a runner/host issue: the
+# sibling debian-stretch-cross image (Dockerfile.sling) uses the exact same sudo pattern and
+# builds fine on the same runner. This image's default user genuinely isn't root either (a plain
+# `sed -i` on /etc/apt/sources.list without sudo fails with a real permission error), so strip
+# sudo AND force USER root directly - that doesn't depend on sudo's (broken) setuid mechanism at
+# all. Patched here rather than in the upstream submodule repo, which we don't own.
+
+# Debian Stretch has been end-of-life for years and its archive's Release-file signatures have
+# since expired, so a plain `apt install -y` refuses to proceed ("unauthenticated packages"). This
+# is the standard, narrowly-scoped workaround for an archived/dead release's own expired signing
+# keys - it does not affect any live, currently-maintained repository.
+sed -i \
+  -e 's/^RUN sudo /RUN /' \
+  -e 's/"sudo", //' \
+  -e '/^FROM /a USER root' \
+  -e 's/apt install --yes/apt -o APT::Get::AllowUnauthenticated=true install --yes/' \
+  Dockerfile
 docker build -t "$IMAGE_NAME" .
 
 cd "$SCRIPT_DIR"