merge: harden ephvm

lock /home/matej/.config to 0700 (was 0755). post-process qcow2
with parallel zstd on qcow2 v3 via qemu-img convert; smaller
image and faster decompress than the built-in qcow2-compressed.

features/vm-guest.nix

@@ -43,19 +43,15 @@
       config = lib.mkIf cfg.enable (
         lib.mkMerge [
           {
-            services.qemuGuest.enable = true;
             services.spice-vdagentd.enable = lib.mkIf (!cfg.headless) true;
 
             boot.kernelParams = lib.mkIf cfg.headless [ "console=ttyS0,115200" ];
 
+            # 9p autoloads on first mount
             boot.initrd.availableKernelModules = [
               "9p"
               "9pnet_virtio"
             ];
-            boot.kernelModules = [
-              "9p"
-              "9pnet_virtio"
-            ];
 
             networking = {
               useDHCP = true;
@@ -68,7 +64,6 @@
               curl
               wget
               htop
-              sshfs
             ];
           }
 

hosts/ephvm/configuration.nix

@@ -13,19 +13,38 @@
   documentation.enable = false;
   environment.defaultPackages = [ ];
 
-  # compressed qcow2, no channel copy
+  # qcow2, no channel copy; post-processed with parallel zstd on qcow2 v3
+  # (~half the size of zlib v2, faster decompress)
   image.modules.qemu =
     { config, modulesPath, ... }:
     {
       system.build.image = lib.mkForce (
-        import (modulesPath + "/../lib/make-disk-image.nix") {
+        let
+          rawImage = import (modulesPath + "/../lib/make-disk-image.nix") {
             inherit lib config pkgs;
             inherit (config.virtualisation) diskSize;
             inherit (config.image) baseName;
-          format = "qcow2-compressed";
+            format = "qcow2";
             copyChannel = false;
             partitionTableType = "legacy";
-        }
+          };
+          inherit (config.image) baseName;
+        in
+        pkgs.runCommand baseName { nativeBuildInputs = [ pkgs.qemu-utils ]; } ''
+          mkdir -p $out
+          # qemu-img caps -m at 16
+          cores="''${NIX_BUILD_CORES:-4}"
+          [ "$cores" -gt 0 ] || cores=4
+          [ "$cores" -gt 16 ] && cores=16
+          qemu-img convert \
+            -f qcow2 \
+            -O qcow2 \
+            -c \
+            -o compression_type=zstd \
+            -m "$cores" \
+            ${rawImage}/${baseName}.qcow2 \
+            $out/${baseName}.qcow2
+        ''
       );
     };
 
@@ -70,7 +89,7 @@
   features.neovim.dotfiles = inputs.nvim;
 
   # ensure .config exists with correct ownership before automount
-  systemd.tmpfiles.rules = [ "d /home/matej/.config 0755 matej users -" ];
+  systemd.tmpfiles.rules = [ "d /home/matej/.config 0700 matej users -" ];
 
   # TODO:(@janezicmatej) replace ssh with virtio-console (hvc0) when qemu 11.0 lands
   # https://www.mail-archive.com/qemu-devel@nongnu.org/msg1162844.html

scripts/ephvm-run.sh

@@ -27,14 +27,32 @@ info() {
 
 # globals for cleanup trap
 CLEANUP_OVERLAY=""
+CLEANUP_TMPDIR=""
 QEMU_PID=""
+VM_READY=false
 cleanup() {
 	[ -n "$QEMU_PID" ] && kill "$QEMU_PID" 2>/dev/null && wait "$QEMU_PID" 2>/dev/null
 	[ -n "$CLEANUP_OVERLAY" ] && rm -rf "$CLEANUP_OVERLAY"
+	# preserve tmpdir on abnormal exit so the qemu log survives for inspection
+	if [ -n "$CLEANUP_TMPDIR" ]; then
+		if [ "$VM_READY" = true ]; then
+			rm -rf "$CLEANUP_TMPDIR"
+		else
+			echo "qemu log preserved: $CLEANUP_TMPDIR/qemu.log" >&2
+		fi
+	fi
 	return 0
 }
 trap cleanup EXIT
 
+# returns 0 once the guest's sshd is speaking (first bytes are "SSH-")
+awaiting_ssh_banner() {
+	local port="$1"
+	local banner
+	banner=$(timeout 2 bash -c "exec 3<>/dev/tcp/localhost/$port; head -c 4 <&3" 2>/dev/null) || return 1
+	[ "$banner" = "SSH-" ]
+}
+
 usage() {
 	cat <<EOF
 Usage: ephvm-run.sh [options]
@@ -55,6 +73,8 @@ EOF
 main() {
 	setup_colors
 
+	[ "$EUID" -eq 0 ] && die "ephvm-run.sh must not run as root"
+
 	local ssh_port="" memory=4G cpus=2 claude=true disk_size="" serial=false
 	local -a mounts=()
 
@@ -110,15 +130,13 @@ main() {
 		CLEANUP_OVERLAY=$(mktemp -d)
 		local overlay="$CLEANUP_OVERLAY/overlay.qcow2"
 		qemu-img create -f qcow2 -b "$(realpath "$image")" -F qcow2 "$overlay" "$disk_size"
-		drive_arg="file=$overlay,format=qcow2"
+		drive_arg="if=none,id=hd0,file=$overlay,format=qcow2,cache=writeback,aio=threads,discard=unmap,detect-zeroes=unmap"
 	else
-		drive_arg="file=$image,format=qcow2,snapshot=on"
+		drive_arg="if=none,id=hd0,file=$image,format=qcow2,snapshot=on,cache=writeback,aio=threads,discard=unmap,detect-zeroes=unmap"
 	fi
 
 	command -v qemu-system-x86_64 &>/dev/null || die "qemu-system-x86_64 not found"
-
+	[ -r /dev/kvm ] || die "/dev/kvm not readable; kvm is required"
-	local accel="tcg"
-	[ -r /dev/kvm ] && accel="kvm"
 
 	# auto-allocate ssh port unless serial mode
 	if [ "$serial" = false ] && [ -z "$ssh_port" ]; then
@@ -128,28 +146,33 @@ main() {
 		done
 	fi
 
-	local nic_arg="user"
+	local nic_arg="user,model=virtio-net-pci"
 	if [ -n "$ssh_port" ]; then
-		nic_arg="user,hostfwd=tcp::${ssh_port}-:22"
+		nic_arg="user,model=virtio-net-pci,hostfwd=tcp:127.0.0.1:${ssh_port}-:22"
 	fi
 
 	local -a qemu_args=(
 		qemu-system-x86_64
-		-accel "$accel"
+		-accel kvm
+		-cpu host
 		-m "$memory"
 		-smp "$cpus"
 		-drive "$drive_arg"
+		-device "virtio-blk-pci,drive=hd0"
+		-device virtio-rng-pci
 		-nic "$nic_arg"
 		-nographic
+		-sandbox "on,obsolete=deny,elevateprivileges=deny,spawn=deny,resourcecontrol=deny"
 	)
 
-	if [ "$accel" != "tcg" ]; then
-		qemu_args+=(-cpu host)
-	fi
-
 	local fs_id=0 mount_path name tag
 	for mount_path in "${mounts[@]}"; do
+		[ -e "$mount_path" ] || die "--mount path does not exist: $mount_path"
 		mount_path=$(realpath "$mount_path")
+		# qemu parses -virtfs as csv, a comma in the path would inject options
+		case "$mount_path" in
+		*,*) die "--mount path may not contain commas: $mount_path" ;;
+		esac
 		name=$(basename "$mount_path")
 		tag="m_${name:0:29}"
 		qemu_args+=(
@@ -163,6 +186,9 @@ main() {
 		mkdir -p "$CLAUDE_CONFIG_DIR"
 		local claude_dir
 		claude_dir=$(realpath "$CLAUDE_CONFIG_DIR")
+		case "$claude_dir" in
+		*,*) die "claude config dir may not contain commas: $claude_dir" ;;
+		esac
 
 		qemu_args+=(
 			-virtfs "local,path=$claude_dir,mount_tag=claude,security_model=none,id=fs${fs_id}"
@@ -171,27 +197,38 @@ main() {
 	fi
 
 	info "---"
-	info "Accel: $accel"
+	[ -n "$ssh_port" ] && info "SSH: ssh -p $ssh_port matej@localhost"
 	info "---"
 
 	if [ "$serial" = true ]; then
 		exec "${qemu_args[@]}"
 	fi
 
+	CLEANUP_TMPDIR=$(mktemp -d)
+	local qemu_log="$CLEANUP_TMPDIR/qemu.log"
+
 	# start qemu in background and auto-ssh
-	"${qemu_args[@]}" &>/dev/null &
+	"${qemu_args[@]}" &>"$qemu_log" &
 	QEMU_PID=$!
 
+	# throwaway ssh key (vm accepts any key via AuthorizedKeysCommand)
+	local ssh_key="$CLEANUP_TMPDIR/id_ed25519"
+	ssh-keygen -t ed25519 -f "$ssh_key" -N "" -q
+
 	info "waiting for vm (port $ssh_port)..."
 	local attempts=0
-	while ! (echo > /dev/tcp/localhost/"$ssh_port") 2>/dev/null; do
+	# poll for the real SSH banner, not TCP accept: qemu's user-mode nic
+	# accepts host-side the moment qemu starts, well before guest sshd is up
+	while ! awaiting_ssh_banner "$ssh_port"; do
 		attempts=$((attempts + 1))
-		[ $attempts -gt 60 ] && die "vm did not become ready in 60s"
+		[ $attempts -gt 120 ] && die "vm did not become ready in 60s"
 		kill -0 "$QEMU_PID" 2>/dev/null || die "qemu exited unexpectedly"
-		sleep 1
+		sleep 0.5
 	done
+	VM_READY=true
 
 	ssh -p "$ssh_port" -t \
+		-i "$ssh_key" \
 		-o StrictHostKeyChecking=no \
 		-o UserKnownHostsFile=/dev/null \
 		-o LogLevel=ERROR \