Create the Windows data folder with its ACL instead of setting it after

On Windows the console sends Ctrl+C to both processes. The firewall
exited at once with 0xC000013A, handing the prompt back while a .cmd
shim's cmd.exe still asked "Terminate batch job (Y/N)?". It now ignores
Ctrl+C and Ctrl+Break while the manager runs and exits with its code.

A killed firewall also left the manager and its children running. The
firewall now joins a kill-on-close job before spawning, so whatever is
left dies with it.

An argument with a line break, which Rust refuses to pass to a batch
file, now says so instead of "batch file arguments are invalid".

Co-authored-by: Ian Graham Martinez <[email protected]>

core/Cargo.toml

@@ -38,14 +38,17 @@ path = "src/main.rs"
 [[bin]]
 name = "firewall-service"
 path = "src/bin/firewall-service.rs"
+required-features = ["machine-wide"]
 
 [[bin]]
 name = "firewall-tray"
 path = "src/bin/firewall-tray.rs"
+required-features = ["machine-wide"]
 
 [[bin]]
 name = "firewall-uninstall"
 path = "src/bin/firewall-uninstall.rs"
+required-features = ["machine-wide"]
 
 [[example]]
 name = "proxy-test"
@@ -87,6 +90,10 @@ path = "tests/tpm.rs"
 name = "managers"
 path = "tests/managers.rs"
 
+[features]
+# Builds the machine-wide binaries: the Windows service and tray, and the macOS uninstaller.
+machine-wide = []
+
 [dependencies]
 clap = "4.6.6"
 directories = "6.0.0"
@@ -118,7 +125,7 @@ schannel = "0.1.29"
 tokio = { version = "1.53.1", features = ["net"] }
 tracing-appender = "0.2.5"
 tray-icon = { version = "0.26.1", default-features = false }
-windows = { version = "0.62", features = ["Win32_Foundation", "Win32_Security", "Win32_Security_Authorization", "Win32_Security_Cryptography", "Win32_System_Com", "Win32_System_Environment", "Win32_System_Threading", "Win32_UI_Shell", "Win32_UI_WindowsAndMessaging"] }
+windows = { version = "0.62", features = ["Win32_Foundation", "Win32_Security", "Win32_Security_Authorization", "Win32_Security_Cryptography", "Win32_Storage_FileSystem", "Win32_System_Com", "Win32_System_Console", "Win32_System_Environment", "Win32_System_JobObjects", "Win32_System_Threading", "Win32_UI_Shell", "Win32_UI_WindowsAndMessaging"] }
 windows-registry = "0.6.1"
 windows-service = "0.8.0"
 

core/src/aliasing/mod.rs

@@ -1,5 +1,7 @@
 pub mod forward;
 pub mod managers;
+#[cfg(windows)]
+mod windows;
 
 pub use forward::spawn;
 pub use managers::*;
@@ -34,7 +36,8 @@ impl Invocation {
 
 pub fn run(port: u16, invocation: &Invocation, dir: Option<&Path>) -> std::io::Result<i32> {
     let proxy = crate::local_url(port, "");
-    let mut command = Command::new(resolve(&invocation.program)?);
+    let program = resolve(&invocation.program)?;
+    let mut command = Command::new(&program);
     command.args(invocation.arguments(&proxy));
     if let Some(dir) = dir {
         command.current_dir(dir);
@@ -42,7 +45,12 @@ pub fn run(port: u16, invocation: &Invocation, dir: Option<&Path>) -> std::io::R
     for (key, value) in managers::variables(&proxy) {
         command.env(key, value);
     }
-    Ok(command.status()?.code().unwrap_or(1))
+    #[cfg(windows)]
+    windows::prepare();
+    let status = command.status();
+    #[cfg(windows)]
+    let status = status.map_err(|error| windows::explain(error, &program));
+    Ok(status?.code().unwrap_or(1))
 }
 
 // Command::new only finds .exe on Windows; npm, pnpm and gem are .cmd shims found via PATHEXT.

core/src/aliasing/windows.rs

@@ -0,0 +1,64 @@
+use std::io::{Error, ErrorKind};
+use std::path::Path;
+use std::sync::Once;
+use tracing::debug;
+use windows::core::{BOOL, PCWSTR};
+use windows::Win32::System::Console::{SetConsoleCtrlHandler, CTRL_BREAK_EVENT, CTRL_C_EVENT};
+use windows::Win32::System::JobObjects::{
+    AssignProcessToJobObject, CreateJobObjectW, JobObjectExtendedLimitInformation,
+    SetInformationJobObject, JOBOBJECT_EXTENDED_LIMIT_INFORMATION,
+    JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE,
+};
+use windows::Win32::System::Threading::GetCurrentProcess;
+
+/// Called before spawning, so the manager behaves as it would run from the shell directly.
+/// Once per process: a caller may run several managers.
+pub fn prepare() {
+    static PREPARED: Once = Once::new();
+    PREPARED.call_once(|| {
+        // The console sends Ctrl+C to the manager too. Exiting first would hand the prompt back
+        // while a .cmd shim still asks "Terminate batch job (Y/N)?", so wait for its exit instead.
+        // A handler, unlike ignoring the signal outright, is not inherited by the manager.
+        let _ = unsafe { SetConsoleCtrlHandler(Some(ignore_interrupt), true) };
+        if let Err(error) = kill_children_on_exit() {
+            debug!(%error, "the manager may outlive the firewall");
+        }
+    });
+}
+
+unsafe extern "system" fn ignore_interrupt(event: u32) -> BOOL {
+    (event == CTRL_C_EVENT || event == CTRL_BREAK_EVENT).into()
+}
+
+/// Joins a job that kills what is left in it once the firewall exits, even if killed itself.
+/// The manager inherits the job when spawned, as do its own children.
+fn kill_children_on_exit() -> windows::core::Result<()> {
+    let mut limits = JOBOBJECT_EXTENDED_LIMIT_INFORMATION::default();
+    limits.BasicLimitInformation.LimitFlags = JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE;
+    // Never closed: the handle closing as the firewall exits is what kills the rest.
+    let job = unsafe { CreateJobObjectW(None, PCWSTR::null())? };
+    unsafe {
+        SetInformationJobObject(
+            job,
+            JobObjectExtendedLimitInformation,
+            (&raw const limits).cast(),
+            size_of_val(&limits) as u32,
+        )?
+    };
+    unsafe { AssignProcessToJobObject(job, GetCurrentProcess()) }
+}
+
+/// Rust refuses to pass a line break to a batch file, where it would cut the command short.
+pub fn explain(error: Error, program: &Path) -> Error {
+    let batch = program
+        .extension()
+        .is_some_and(|ext| ext.eq_ignore_ascii_case("cmd") || ext.eq_ignore_ascii_case("bat"));
+    if !batch || error.kind() != ErrorKind::InvalidInput {
+        return error;
+    }
+    let name = program.file_name().unwrap_or(program.as_os_str()).to_string_lossy();
+    Error::new(
+        ErrorKind::InvalidInput,
+        format!("{name} is a batch file, which cannot be passed an argument with a line break"),
+    )
+}

core/src/os/windows/ca.rs

@@ -11,6 +11,7 @@ use std::time::{SystemTime, UNIX_EPOCH};
 use tracing::warn;
 use windows::Win32::Security::Authorization::{GetNamedSecurityInfoW, SE_FILE_OBJECT};
 use windows::Win32::Security::*;
+use windows::Win32::Storage::FileSystem::CreateDirectoryW;
 use windows_registry::Type;
 
 /// SYSTEM and admins write, users only read, so no one else can plant a CA for node.
@@ -65,25 +66,39 @@ pub fn load_or_create_ca(
     Ok(ca)
 }
 
-/// Created on first use, its owner and ACL reset every time. One a user made first, or a link,
+/// Created on first use with its ACL, which is reset every time. One a user made first, or a link,
 /// is replaced: its owner could grant itself access again, and a link would redirect our writes.
 pub(super) fn secure_data_dir() -> io::Result<PathBuf> {
     let dir = data_dir();
-    match std::fs::symlink_metadata(&dir) {
-        Ok(meta) if meta.file_type().is_symlink() => std::fs::remove_dir(&dir)?,
+    let exists = match std::fs::symlink_metadata(&dir) {
+        Ok(meta) if meta.file_type().is_symlink() => {
+            std::fs::remove_dir(&dir)?;
+            false
+        }
         Ok(_) if !owned_by_admins(&dir)? => {
             warn!(dir = %dir.display(), "replacing a data folder not created by an admin");
-            std::fs::remove_dir_all(&dir)?
+            std::fs::remove_dir_all(&dir)?;
+            false
         }
-        _ => {}
-    }
-    std::fs::create_dir_all(&dir)?;
+        Ok(_) => true,
+        Err(_) => false,
+    };
     let info = OWNER_SECURITY_INFORMATION
         | DACL_SECURITY_INFORMATION
         | PROTECTED_DACL_SECURITY_INFORMATION;
     let path = HSTRING::from(dir.as_path());
     let acl = security_descriptor(DATA_DIR_ACL)?;
-    unsafe { SetFileSecurityW(&path, info, PSECURITY_DESCRIPTOR(acl.0)).ok()? };
+    if exists {
+        unsafe { SetFileSecurityW(&path, info, PSECURITY_DESCRIPTOR(acl.0)).ok()? };
+    } else {
+        let attributes = SECURITY_ATTRIBUTES {
+            nLength: size_of::<SECURITY_ATTRIBUTES>() as u32,
+            lpSecurityDescriptor: acl.0,
+            bInheritHandle: false.into(),
+        };
+        // Never inherits ProgramData's ACL, which lets users add files; fails if it was recreated.
+        unsafe { CreateDirectoryW(&path, Some(&attributes))? };
+    }
     Ok(dir)
 }
 

core/src/os/windows/service.rs

@@ -26,6 +26,7 @@ use std::net::{IpAddr, Ipv4Addr};
 use std::sync::Arc;
 use std::time::Duration;
 use tracing_appender::rolling::{RollingFileAppender, Rotation};
+use windows::Win32::Foundation::ERROR_GEN_FAILURE;
 use windows_service::service::*;
 use windows_service::service_control_handler::{
     self, ServiceControlHandlerResult, ServiceStatusHandle,
@@ -73,8 +74,12 @@ fn run_until_stopped() -> Result<(), BoxError> {
         | ServiceControl::Shutdown
         | ServiceControl::Pause
         | ServiceControl::Continue => {
-            let _ = control_sender.send(control);
-            ServiceControlHandlerResult::NoError
+            // Answered once handled, so a failed pause fails the caller's ControlService.
+            let (reply, handled) = tokio::sync::oneshot::channel();
+            let _ = control_sender.send((control, reply));
+            handled
+                .blocking_recv()
+                .unwrap_or(ServiceControlHandlerResult::NoError)
         }
         ServiceControl::Interrogate => ServiceControlHandlerResult::NoError,
         _ => ServiceControlHandlerResult::NotImplemented,
@@ -89,16 +94,26 @@ fn run_until_stopped() -> Result<(), BoxError> {
             }
         };
         // Only pausing removes our hosts entries; a stop or crash keeps them, failing closed.
-        while let Some(control) = control_receiver.recv().await {
+        while let Some((control, reply)) = control_receiver.recv().await {
             let (hosts, state) = match control {
                 ServiceControl::Pause => (&[][..], ServiceState::Paused),
                 ServiceControl::Continue => (&hosts[..], ServiceState::Running),
                 _ => break,
             };
-            match write_hosts_entries(hosts) {
-                Ok(()) => set_status(&status, state, false)?,
-                Err(error) => tracing::error!(%error, "could not rewrite the hosts file"),
-            }
+            let result = match write_hosts_entries(hosts) {
+                Ok(()) => {
+                    set_status(&status, state, false)?;
+                    ServiceControlHandlerResult::NoError
+                }
+                Err(error) => {
+                    tracing::error!(%error, "could not rewrite the hosts file");
+                    let code = error
+                        .raw_os_error()
+                        .map_or(ERROR_GEN_FAILURE.0, |code| code as u32);
+                    ServiceControlHandlerResult::Other(code)
+                }
+            };
+            let _ = reply.send(result);
         }
         Ok::<_, BoxError>(())
     });
@@ -147,7 +162,7 @@ async fn start_screening(status: &ServiceStatusHandle) -> Result<Vec<&'static st
                 resolver: resolver.clone(),
                 service: service.clone(),
             };
-            let upstream = ConsumeErrLayer::trace_as_debug().into_layer(upstream);
+            let upstream = ConsumeErrLayer::trace_as_warning().into_layer(upstream);
             tokio::spawn(listen_on(address, port, &exec)?.serve(upstream));
         }
     }

core/src/os/windows/tray.rs

@@ -144,7 +144,11 @@ impl Tray {
 }
 
 fn show_error(error: BoxError) {
-    let text = HSTRING::from(error.to_string());
+    let text = match error.source() {
+        Some(source) => format!("{error}: {source}"),
+        None => error.to_string(),
+    };
+    let text = HSTRING::from(text);
     unsafe { MessageBoxW(None, &text, w!("Strix Firewall"), MB_ICONERROR) };
 }
 

dist-workspace.toml

@@ -7,6 +7,7 @@ ci = "github"
 installers = ["shell", "powershell", "homebrew", "msi"]
 # core/wix/main.wxs is edited to install the service.
 allow-dirty = ["msi"]
+features = ["machine-wide"]
 tap = "usestrix/homebrew-tap"
 publish-jobs = ["homebrew", "./publish-crates"]
 targets = ["aarch64-apple-darwin", "x86_64-apple-darwin", "aarch64-unknown-linux-gnu", "x86_64-unknown-linux-gnu", "x86_64-pc-windows-msvc"]

macos/project.yml

@@ -133,7 +133,7 @@ targets:
               arm64|arm64e) triple=aarch64-apple-darwin ;;
               x86_64)       triple=x86_64-apple-darwin ;;
             esac
-            cargo build --bin firewall-uninstall --target $triple $([ "$CONFIGURATION" = Debug ] || echo --release)
+            cargo build --bin firewall-uninstall --features machine-wide --target $triple $([ "$CONFIGURATION" = Debug ] || echo --release)
             bins="$bins target/$triple/$CARGO_PROFILE/firewall-uninstall"
           done
           lipo -create $bins -output "$TARGET_BUILD_DIR/$EXECUTABLE_FOLDER_PATH/firewall-uninstall"