Skip to content

A rootless Android app that boots Alpine Linux: run containers (Podman/Docker/LXC) and GUI desktop apps.

fix(rootfs,init): rescue mount-move, await block devices, verify QEMU IP, daemon liveness, validate resize, exclude bridges from NIC pick, random password salt

ExTvc6392ad2115d
diff --git a/build-rootfs/build-rootfs.sh b/build-rootfs/build-rootfs.sh
index ef9837d..1ee8eaa 100755
--- a/build-rootfs/build-rootfs.sh
+++ b/build-rootfs/build-rootfs.sh
@@ -74,7 +74,9 @@ chmod 0440 "$ROOTFS/etc/sudoers.d/wheel"
# Set root password to "podroid" (pre-hashed with openssl).
# We can't run chpasswd inside the aarch64 rootfs from an x86_64 host,
# so write the SHA-512 hash directly into /etc/shadow.
-ROOT_HASH=$(openssl passwd -6 -salt podroid podroid)
+# No fixed -salt: openssl generates a random salt so the stored hash differs
+# per build (the password stays the documented default "podroid").
+ROOT_HASH=$(openssl passwd -6 podroid)
sed -i "s|^root:[^:]*:|root:${ROOT_HASH}:|" "$ROOTFS/etc/shadow"
# Strip docs/man/locale to shrink squashfs
diff --git a/build-rootfs/files/etc/init.d/podroid-bootstrap b/build-rootfs/files/etc/init.d/podroid-bootstrap
index 4cd89ee..ffee78f 100644
--- a/build-rootfs/files/etc/init.d/podroid-bootstrap
+++ b/build-rootfs/files/etc/init.d/podroid-bootstrap
@@ -84,17 +84,15 @@ start() {
if [ -b /dev/zram0 ]; then
_mem_kb=$(awk '/^MemTotal:/{print $2}' /proc/meminfo)
echo lz4 > /sys/block/zram0/comp_algorithm 2>/dev/null
+ # *1536 = KiB→bytes (*1024) × 1.5 ratio; disksize wants bytes.
echo $((_mem_kb * 1536)) > /sys/block/zram0/disksize 2>/dev/null
mkswap /dev/zram0 >/dev/null 2>&1 && swapon -p 100 /dev/zram0 2>/dev/null
fi
- # OOM tuning: prefer killing leaf user-apps (Firefox, browsers, npm,
- # webpack) over our infrastructure. Higher score = killed first. -1000
- # = OOM-immune (set in podroid-ready for dropbear/Xvnc/agent so all
- # those services exist by then). Bumping the OOM "score adjustment"
- # for non-protected processes is left to the kernel default. The big
- # win is making sure VNC + SSH + the host-vsock relay survive even
- # when a memory-hog browser tab gets reaped.
+ # OOM behavior: re-assert the kernel default (0 = let the OOM killer pick
+ # the highest-badness victim, rather than killing whatever triggered the
+ # allocation). Per-service immunity (-1000 for dropbear/Xvnc/agent) is set
+ # in podroid-ready once those services exist.
[ -w /proc/sys/vm/oom_kill_allocating_task ] && \
echo 0 > /proc/sys/vm/oom_kill_allocating_task
# overcommit_memory=1 (always overcommit) lets us actually hand out
diff --git a/build-rootfs/files/etc/init.d/podroid-network b/build-rootfs/files/etc/init.d/podroid-network
index db22cf5..bb8beb7 100644
--- a/build-rootfs/files/etc/init.d/podroid-network
+++ b/build-rootfs/files/etc/init.d/podroid-network
@@ -13,8 +13,11 @@ start() {
ip link set lo up 2>/dev/null
NETIF=""
for _i in $(seq 1 10); do
+ # Exclude virtual/bridge devices so the real virtio NIC is picked.
+ # lxcbr0 is created by podroid-bootstrap (which runs *before* us), so
+ # it would otherwise be a candidate; br-/cni cover other bridge names.
NETIF=$(ip -o link show 2>/dev/null | awk -F': ' '{print $2}' | \
- grep -vE '^(lo|dummy[0-9]*|veth|podman|cni-|docker)' | head -1)
+ grep -vE '^(lo|dummy[0-9]*|veth|podman|cni|docker|lxcbr|br-)' | head -1)
[ -n "$NETIF" ] && break
sleep 0.05
done
@@ -49,6 +52,14 @@ start() {
ip route add default via 10.0.2.2 dev "$NETIF" 2>/dev/null
printf 'nameserver 10.0.2.3\nnameserver 8.8.8.8\nnameserver 1.1.1.1\n' \
> /etc/resolv.conf
+ # Mirror the AVF post-check: if the static assignment didn't stick
+ # (wrong NIC, link not up), don't report success. Dependents that
+ # `need` us would assume the link is up and then SSH/apk fail with no
+ # obvious cause. Surface it instead.
+ if ! ip -o -4 addr show dev "$NETIF" 2>/dev/null | grep -q 'inet '; then
+ eend 1 "static IP did not stick on $NETIF"
+ return 1
+ fi
fi
echo "Network found" > /dev/console
eend 0
diff --git a/build-rootfs/files/etc/init.d/podroid-vsock b/build-rootfs/files/etc/init.d/podroid-vsock
index 115520a..d4aeb2d 100644
--- a/build-rootfs/files/etc/init.d/podroid-vsock
+++ b/build-rootfs/files/etc/init.d/podroid-vsock
@@ -25,6 +25,20 @@ start() {
--stderr /var/log/podroid-vsock-agent.log \
--exec /usr/local/bin/podroid-vsock-agent
_agent_rc=$?
+ # --background returns 0 the instant the child forks, even if the agent
+ # exits immediately (e.g. AF_VSOCK bind failure). Poll the pidfile's PID:
+ # a healthy agent breaks out on the first ~0.1s tick; a crash-on-launch
+ # misses every tick and the final kill -0 flips the rc.
+ if [ "$_agent_rc" -eq 0 ]; then
+ _agent_pid=$(cat "$pidfile" 2>/dev/null)
+ _i=0
+ while [ "$_i" -lt 10 ]; do
+ sleep 0.1
+ kill -0 "$_agent_pid" 2>/dev/null && break
+ _i=$((_i + 1))
+ done
+ kill -0 "$_agent_pid" 2>/dev/null || _agent_rc=1
+ fi
[ "$_agent_rc" -eq 0 ] || ewarn "podroid-vsock-agent failed to start (see /var/log/podroid-vsock-agent.log)"
eend "$_agent_rc"
}
diff --git a/build-rootfs/files/etc/init.d/podroid-x11 b/build-rootfs/files/etc/init.d/podroid-x11
index e58bc31..a18d928 100644
--- a/build-rootfs/files/etc/init.d/podroid-x11
+++ b/build-rootfs/files/etc/init.d/podroid-x11
@@ -53,6 +53,21 @@ start() {
-dpi $XVNC_DPI \
-AcceptSetDesktopSize
_xvnc_rc=$?
+ # --background returns 0 once the child is forked, even if it dies an
+ # instant later (bad -geometry, port already bound). Poll the pidfile's
+ # PID instead of a fixed sleep: a healthy daemon is confirmed alive on the
+ # first ~0.1s tick and we stop; a crash-on-launch leaves no live PID, so
+ # all ticks miss and the final kill -0 flips the rc. Caps at ~1s.
+ if [ "$_xvnc_rc" -eq 0 ]; then
+ _xvnc_pid=$(cat /run/xvnc.pid 2>/dev/null)
+ _i=0
+ while [ "$_i" -lt 10 ]; do
+ sleep 0.1
+ kill -0 "$_xvnc_pid" 2>/dev/null && break
+ _i=$((_i + 1))
+ done
+ kill -0 "$_xvnc_pid" 2>/dev/null || _xvnc_rc=1
+ fi
[ "$_xvnc_rc" -eq 0 ] || ewarn "Xvnc failed to start (see $XVNC_LOGFILE)"
# PulseAudio: system-wide-style daemon writing raw S16LE PCM to TCP 4713.
@@ -71,6 +86,20 @@ start() {
--load="module-simple-protocol-tcp source=podroid_sink.monitor record=true rate=44100 format=s16le channels=2 listen=0.0.0.0 port=$PA_PORT" \
--load="module-native-protocol-unix"
_pa_rc=$?
+ # Same liveness poll as Xvnc: --background hides an immediate crash
+ # (missing module, port 4713 already bound). A healthy daemon breaks out
+ # on the first ~0.1s tick; an instant crash misses every tick and the
+ # final kill -0 flips the rc.
+ if [ "$_pa_rc" -eq 0 ]; then
+ _pa_pid=$(cat /run/pulseaudio.pid 2>/dev/null)
+ _i=0
+ while [ "$_i" -lt 10 ]; do
+ sleep 0.1
+ kill -0 "$_pa_pid" 2>/dev/null && break
+ _i=$((_i + 1))
+ done
+ kill -0 "$_pa_pid" 2>/dev/null || _pa_rc=1
+ fi
[ "$_pa_rc" -eq 0 ] || ewarn "pulseaudio failed to start (see $PA_LOGFILE)"
# eend reflects either daemon failing: a single `eend $?` would have
diff --git a/build-rootfs/files/usr/local/bin/podroid-resize b/build-rootfs/files/usr/local/bin/podroid-resize
index b4629da..e4d88c5 100644
--- a/build-rootfs/files/usr/local/bin/podroid-resize
+++ b/build-rootfs/files/usr/local/bin/podroid-resize
@@ -23,6 +23,11 @@ while IFS= read -r _line; do
case "$_line" in
RESIZE\ *)
set -- $_line
+ # Reject empty / non-numeric / zero rows or cols before persisting:
+ # a malformed line must not stamp junk into /run/term_size (which
+ # podroid-login reads back) or stty a 0x0 winsize onto hvc0.
+ case "$2" in ''|*[!0-9]*|0) continue;; esac
+ case "$3" in ''|*[!0-9]*|0) continue;; esac
stty rows "$2" cols "$3" </dev/hvc0 2>/dev/null
printf '%s %s\n' "$2" "$3" > /run/term_size 2>/dev/null
;;
diff --git a/init-podroid b/init-podroid
index 521d804..7c97204 100755
--- a/init-podroid
+++ b/init-podroid
@@ -17,6 +17,15 @@ echo "Mounting storage..." > /dev/console
# /dev/vdb = read-only Alpine squashfs (system files)
mkdir -p /mnt/persist /mnt/lower /mnt/overlay
+# Bounded wait for the virtio block nodes: on most boots they already exist
+# (loop breaks on the first iteration), but a slow virtio_blk probe can lag
+# behind init. Cap the wait so a genuinely-missing node still drops to the
+# FATAL diagnostics below instead of hanging forever.
+for _i in $(seq 1 30); do
+ [ -b /dev/vda ] && [ -b /dev/vdb ] && break
+ sleep 0.1
+done
+
# Persistent: format on first boot if needed.
# noatime: stop atime writes on every read.
# commit=60: journal commits every 60s (default 5s) — 12× less journal IO
@@ -27,34 +36,43 @@ mkdir -p /mnt/persist /mnt/lower /mnt/overlay
# Saves journal-commit syscalls on every metadata change.
if ! mount -t ext4 -o rw,noatime,commit=60,barrier=0 /dev/vda /mnt/persist 2>/dev/null; then
e2fsck -y /dev/vda >/dev/null 2>&1 || mkfs.ext4 -q -F /dev/vda
- mount -t ext4 -o rw,noatime,commit=60,barrier=0 /dev/vda /mnt/persist
+ mount -t ext4 -o rw,noatime,commit=60,barrier=0 /dev/vda /mnt/persist \
+ || { echo "FATAL: persistent ext4 (/dev/vda) mount failed" > /dev/console; ls -l /dev/vd* > /dev/console 2>&1; exec sh; }
fi
echo "Mounting system..." > /dev/console
mount -t squashfs -o ro /dev/vdb /mnt/lower \
- || { echo "FATAL: squashfs (/dev/vdb) mount failed" > /dev/console; exec sh; }
+ || { echo "FATAL: squashfs (/dev/vdb) mount failed" > /dev/console; ls -l /dev/vd* > /dev/console 2>&1; exec sh; }
-mkdir -p /mnt/persist/upper /mnt/persist/work
+mkdir -p /mnt/persist/upper /mnt/persist/work \
+ || { echo "FATAL: mkdir overlay upper/work failed" > /dev/console; exec sh; }
echo "Stacking overlay..." > /dev/console
mount -t overlay overlay \
-o lowerdir=/mnt/lower,upperdir=/mnt/persist/upper,workdir=/mnt/persist/work,metacopy=on,redirect_dir=on,index=on \
/mnt/overlay \
- || { echo "FATAL: overlay mount failed" > /dev/console; exec sh; }
+ || { echo "FATAL: overlay mount failed" > /dev/console; ls -l /dev/vd* > /dev/console 2>&1; exec sh; }
# Move /mnt/persist into the new root so users can reach raw storage,
# and so switch_root can fully clear initramfs (no orphan mounts left
# under /mnt that would block its unlink pass).
-mkdir -p /mnt/overlay/mnt/persist
-mount --move /mnt/persist /mnt/overlay/mnt/persist
+mkdir -p /mnt/overlay/mnt/persist \
+ || { echo "FATAL: mkdir /mnt/overlay/mnt/persist failed" > /dev/console; exec sh; }
+mount --move /mnt/persist /mnt/overlay/mnt/persist \
+ || { echo "FATAL: mount --move /mnt/persist failed" > /dev/console; exec sh; }
# Move /mnt/lower (the squashfs) too — keeps /dev/vdb mount visible in the
# new root, and lets initramfs cleanup unlink /mnt cleanly.
-mkdir -p /mnt/overlay/mnt/lower
-mount --move /mnt/lower /mnt/overlay/mnt/lower
+mkdir -p /mnt/overlay/mnt/lower \
+ || { echo "FATAL: mkdir /mnt/overlay/mnt/lower failed" > /dev/console; exec sh; }
+mount --move /mnt/lower /mnt/overlay/mnt/lower \
+ || { echo "FATAL: mount --move /mnt/lower failed" > /dev/console; exec sh; }
# Move our early mounts into the new root so OpenRC inherits them
-mount --move /proc /mnt/overlay/proc
-mount --move /sys /mnt/overlay/sys
-mount --move /dev /mnt/overlay/dev
+mount --move /proc /mnt/overlay/proc \
+ || { echo "FATAL: mount --move /proc failed" > /dev/console; exec sh; }
+mount --move /sys /mnt/overlay/sys \
+ || { echo "FATAL: mount --move /sys failed" > /dev/console; exec sh; }
+mount --move /dev /mnt/overlay/dev \
+ || { echo "FATAL: mount --move /dev failed" > /dev/console; exec sh; }
exec /sbin/switch_root /mnt/overlay /sbin/init