The argument builder was written based on the incorrect assumption that bwrap arguments are unordered. The argument builder is replaced in this commit to correct that mistake.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
Bubblewrap as init is a bit awkward and don't support a few setup actions fortify will need, such as starting/supervising nscd.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
Both machinectl and sudo launch methods launch shim as shim is now responsible for setting up the sandbox. Various app structures are adapted to accommodate bwrap configuration and mediated wayland access.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
ACL operations are now tagged with the enablement causing them. At the end of child process's life, enablements of all remaining launchers are resolved and inverted. This allows Wait to only revert operations targeting resources no longer required by other launchers.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
Launch methods serve the primary purpose of setting UID in the init namespace, which bubblewrap does not do. Furthermore, all applications will start within a bubblewrap sandbox once it has been implemented.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
There was an earlier attempt of cleaning up the app package however it ended up creating even more of a mess and the code structure largely still looked like Ego with state setup scattered everywhere and a bunch of ugly hacks had to be implemented to keep track of all of them. In this commit the entire app package is rewritten to track everything that has to do with an app in one thread safe value.
In anticipation of the client/server split also made changes:
- Console messages are cleaned up to be consistent
- State tracking is fully rewritten to be cleaner and usable for multiple process and client/server
- Encapsulate errors to easier identify type of action causing the error as well as additional info
- System-level setup operations is grouped in a way that can be collectively committed/reverted
and gracefully handles errors returned by each operation
- Resource sharing is made more fine-grained with PID-scoped resources whenever possible,
a few remnants (X11, Wayland, PulseAudio) will be addressed when a generic proxy is available
- Application setup takes a JSON-friendly config struct and deterministically generates system setup operations
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
Exit cleanup state information is now stored in a dedicated struct and built up using methods of that struct.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
State query command has been moved to main where it belongs, "system" information are now fetched in app.New and stored in *App with accessors for relevant values. Exit (cleanup-related) functions are separated into its dedicated "final" package.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
This commit adds code that starts and registers the D-Bus proxy, as well as cleanup code that tracks and closes the daemon once our child exits. A few more flags were added to pass D-Bus config to xdg-dbus-proxy.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
This is yet another remnant of Ego, as Ego unconditionally shares these resources and the absence of them are ignored and warned about in verbose logging. In our case they are individually opt-in so silently dropping them while the enablement is still set makes very little sense.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
In the past Wayland, X and PulseAudio are shared unconditionally. This can unnecessarily increase attack surface as some of these resources might not be needed at all. This commit moves all environment preparation code to the internal app package and selectively call them based on flags.
An "enablements" bitfield is introduced tracking all enabled shares. This value is registered after successful child process launch and stored in launcher states.
Code responsible for running the child process is isolated to its own app/run file and cleaned up. Launch method selection is also extensively cleaned up.
The internal state/track readLaunchers function now takes uid as an argument. Launcher state is now printed using text/tabwriter and argv is only emitted when verbose.
Signed-off-by: Ophestra Umiker <cat@ophivana.moe>
