Skip to content

controller-tests.md

Latest commit

 

History

History
111 lines (96 loc) · 6.65 KB

controller-tests.md

File metadata and controls

111 lines (96 loc) · 6.65 KB
Raw

Virtual controller acceptance

test-input OUTPUT starts Monado's official remote driver in a temporary configuration, with its own IPC directory and an ephemeral TCP port. Run it inside the development VM or an isolated Linux container. The client connects only to loopback; no physical driver is selected. The automated VM has no network interface. All processes have finite deadlines and are cleaned up.

The protocol matches installed Monado 25.1.0^20260820git01c1f6b-1.fc44, source 01c1f6b23ab73c1459c8f4cfd2eb16d5142214c2. The wire layout comes from src/xrt/drivers/remote/r_interface.h: the native 376-byte mndrmt3 packet contains a head and two 120-byte controller records. The harness receives both server templates and checks their protocol header before changing fields. The configuration needs view_count: 2 as well as version and port; omitting view count makes this release silently use port 4242. See u_config_json_get_remote_settings and target_builder_remote.c in that same revision. Revalidate the layout and settings when Monado changes.

The real OpenXR probe binds each Index controller's grip pose, trigger click, trigger value and haptic action. Over 600 samples, it requires both controllers to press and release, reach trigger value 0.75, become inactive, and become active again. It also checks stereo views, valid poses, head movement, monotonic timestamps and orderly session exit. The initial run on Quest-derived Linux 5.10.246 passed with 600 valid stereo samples and 541 tracked samples per controller; the deliberate one-second disconnect accounts for the rest.

The clean runtime image repeats this acceptance on both Fedora's 7.1 kernel and the Quest-derived 5.10.246 QEMU kernel. Five successive runtime start/test/stop cycles also passed. After each cycle's process cleanup, the test cgroup retained 0.88–0.97 MiB; this is residual accounting, not the peak during rendering. The bounded soak used a 700 MiB cap and a 130-second deadline. Its log is output/steamvr-session/diagnostics-kernel-v4/openxr-input-soak.txt.

Both haptic start/stop requests succeed at the OpenXR API. The remote driver's output callback is a default stub, so this result establishes API handling only. It does not establish delivery to a haptic device. SteamVR controller actions and game audio remain separate tests.

test-input OUTPUT --windows runs the same C++ action probe as a Windows x86-64 executable through the pinned Khronos loader, ARM64 Proton/FEX and native Monado. Windows uses XR_KHR_win32_convert_performance_counter_time; native Linux uses the timespec conversion extension. The harness waits for the application's explicit ready marker before changing controls. Windows writes its results to an explicit file because its Wine console does not inherit the Linux output pipe reliably. The test does not substitute a mock OpenXR implementation.

Fresh image output/runtime-vm-v5/ passes both native and Windows action tests on Fedora 7.1 and Quest-derived 5.10.246. Each application checks 600 stereo samples and every required transition for both hands. These tests use Monado's headless compositor; the separate Vulkan 1 and Vulkan 2 tests establish rendered stereo output.

The combined test now exercises both paths together:

xvfb-run -a --server-args='-screen 0 1280x800x24 -noreset' \
  test-input OUTPUT --render
xvfb-run -a --server-args='-screen 0 1280x800x24 -noreset' \
  test-input OUTPUT --windows --render

The probe creates real Vulkan 2/OpenXR swapchains and submits a stereo projection layer on each rendered frame. Left/right trigger actions change the corresponding eye's color. Four compositor captures must show the initial colors, each separate press, and release. The checker rejects stale frames, unexpected colors and reversed eye ordering. It also requires all pose/action/disconnection checks and normal session exit. The ready marker is emitted only after the session gains input focus.

Both maintained binaries pass on Quest-derived QEMU Linux: 600 samples, 599 submitted projection frames, and approximately 90 pressed frames per hand. Validation runs confirm the actual native Vulkan instance/device validation layer loaded in both paths, with no Vulkan validation errors. Images, logs and hashes are under output/rendered-input-validation-native/ and output/rendered-input-validation-windows/. The compositor captures were visually inspected as well as checked numerically. These are functional test patterns, not game or headset-performance results.

The logs retain Monado's unimplemented remote-haptic callback messages and a space-destruction diagnostic during teardown. Applications and the service exit successfully; those runtime diagnostics have not been resolved. The test uses a disabled D-Bus endpoint to avoid an unrelated desktop portal mounting into its temporary runtime directory; an absent filesystem path breaks Runtime 4's bind validation, so it uses the existing /dev/null path.

An additional native/Windows diagnostic moves the virtual head 35 cm along X, then invokes the installed monado-ctl -c against only that test's private IPC socket. Both applications receive one XrEventDataReferenceSpaceChangePending for LOCAL space and observe their view-space center change from 0.35 m to 0 m while stereo/controller tests continue to pass. The runtime sets poseValid false, so the test does not interpret poseInPreviousSpace; it verifies the coordinates through subsequent xrLocateViews calls. This is allowed by the OpenXR event specification. The first diagnostic incorrectly required that optional pose and failed; its correction preserves the event and actual coordinate requirements. Evidence: output/recenter-native-v2/ and output/recenter-windows-v2/, with temporary test sources in .agents/. This supplemental check is not yet part of the fresh-image boot suite.

Five consecutive Windows rendered-input runs also pass on the Quest-derived kernel after reboot. Each starts its own Xvfb/Monado session and relaunches Proton with the existing test prefix. The 2 GiB cgroup reports zero memory limit/OOM events; post-cycle accounting stays around 526 MiB. memory.peak is unavailable on this kernel, so no peak-memory claim is made. All five application logs and capture measurements are preserved in output/windows-rendered-soak-v1/. This bounded test does not establish SteamVR, game or long-session stability.

Monado documents this driver for hardware-free development and conformance testing.