hst: add direct hardware option
All checks were successful
Test / Create distribution (push) Successful in 34s
Test / Sandbox (push) Successful in 2m19s
Test / Hakurei (push) Successful in 47s
Test / Sandbox (race detector) (push) Successful in 2m17s
Test / Hakurei (race detector) (push) Successful in 3m12s
Test / Hpkg (push) Successful in 3m26s
Test / Flake checks (push) Successful in 1m37s

This is unfortunately the only possible setup to securely expose PipeWire to the container. Further explanation explained in the doc comment and #29.

This will be implemented in a future commit.

Signed-off-by: Ophestra <cat@gensokyo.uk>
This commit is contained in:
2025-12-15 12:29:32 +09:00
parent fae910a1ad
commit 0c0e3d6fc2

View File

@@ -30,12 +30,46 @@ type Config struct {
// This option is unsupported and most likely enables full control over the Wayland
// session. Do not set this to true unless you are sure you know what you are doing.
DirectWayland bool `json:"direct_wayland,omitempty"`
// Direct access to the PipeWire socket established via SecurityContext::Create, no
// attempt is made to start the pipewire-pulse server.
//
// The SecurityContext machinery is fatally flawed, it blindly sets read and execute
// bits on all objects for clients with the lowest achievable privilege level (by
// setting PW_KEY_ACCESS to "restricted"). This enables them to call any method
// targeting any object, and since Registry::Destroy checks for the read and execute bit,
// allows the destruction of any object other than PW_ID_CORE as well. This behaviour
// is implemented separately in media-session and wireplumber, with the wireplumber
// implementation in Lua via an embedded Lua vm. In all known setups, wireplumber is
// in use, and there is no known way to change its behaviour and set permissions
// differently without replacing the Lua script. Also, since PipeWire relies on these
// permissions to work, reducing them is not possible.
//
// Currently, the only other sandboxed use case is flatpak, which is not aware of
// PipeWire and blindly exposes the bare PulseAudio socket to the container (behaves
// like DirectPulse). This socket is backed by the pipewire-pulse compatibility daemon,
// which obtains client pid via the SO_PEERCRED option. The PipeWire daemon, pipewire-pulse
// daemon and the session manager daemon then separately performs the /.flatpak-info hack
// described in https://git.gensokyo.uk/security/hakurei/issues/21. Under such use case,
// since the client has no direct access to PipeWire, insecure parts of the protocol are
// obscured by pipewire-pulse simply not implementing them, and thus hiding the flaws
// described above.
//
// Hakurei does not rely on the /.flatpak-info hack. Instead, a socket is sets up via
// SecurityContext. A pipewire-pulse server connected through it achieves the same
// permissions as flatpak does via the /.flatpak-info hack and is maintained for the
// life of the container.
//
// This option is unsupported and enables a denial-of-service attack as the sandboxed
// client is able to destroy any client object and thus disconnecting them from PipeWire,
// or destroy the SecurityContext object preventing any further container creation.
// Do not set this to true, it is insecure under any configuration.
DirectPipeWire bool `json:"direct_pipewire,omitempty"`
// Direct access to PulseAudio socket, no attempt is made to establish pipewire-pulse
// server via a PipeWire socket with a SecurityContext attached and the bare socket
// is made available to the container.
//
// This option is unsupported and enables arbitrary code execution as the PulseAudio
// server. Do not set this to true, this is insecure under any configuration.
// server. Do not set this to true, it is insecure under any configuration.
DirectPulse bool `json:"direct_pulse,omitempty"`
// Extra acl updates to perform before setuid.