TheCreateGM / mvgal-docs / blob / main / STATUS.md

MVGAL Project Status

Source version

The source repository reports 0.7.31 in its CMake and Cargo workspace metadata. The source changelog has entries through 0.7.31. This documentation snapshot was reconciled against that source on 2026-10-30. Package publication and physical hardware validation are separate questions; a successful build or installed artifact does not establish runtime support.

What the code provides

The source includes a Linux DRM-facing kernel module, GPU discovery and capability reporting, a userspace library/API, the mvgald runtime daemon, several API integration targets, management/diagnostic tools, Rust safety crates, language bindings, UI sources, tests and distribution packaging definitions. Exact components depend on build options, dependency discovery and package selection.

v0.7.31 is a field-report-driven correctness release addressing the 0.7.30 test report on a single-GPU Secure Boot host (Fedora 43, kernel 7.2.13-200.fc44.x86_64, Intel UHD Graphics 770, plus a blacklisted MTT S30). COPR build #11229933 succeeded on fedora-43-x86_64. Key fixes: C1 — the kernel module now includes PCI_CLASS_DISPLAY_OTHER (0x0380) so Intel integrated GPUs are found alongside discrete cards, and skips display-class PCI devices whose native driver is not bound (blacklisted MTT S30), so kernel and userspace enumeration agree on GPU indices; C2 — mvgal_mtt.c now checks MVGAL_FULL_STACK (the real compile-time guard) instead of the non-existent CONFIG_MVGal_MTT typo, and the find_module() calls removed from kernel 7.x are replaced by a /proc/modules read; C3 — udev rules and the pkexec helper now set GROUP="mvgal" and MODE="0660" (were GROUP="video" and MODE="0666"); H1 — mvgal-enum and mvgal list-gpus now scan PCI for display-class GPUs with no DRM node and warn about present-but-unusable devices; H2 — mvgal start now checks for /sys/module/mvgal after launching the daemon and prints a degraded-mode warning if the kernel module is absent.

v0.7.30 is a field-report-driven correctness and packaging release from the 0.7.30 test campaign on a single-GPU Secure Boot host (Fedora 44, kernel 7.2.12-200.fc44.x86_64, Intel UHD Graphics 770). Multi-GPU aggregation could not be exercised (single GPU), and the kernel module path remains unavailable under Secure Boot until the MOK is enrolled (see SECURE_BOOT.md). The headline fixes: mvgal-vulkan-env.sh no longer forces VK_ICD_FILENAMES/VK_DRIVER_FILES to only the MVGAL aggregating ICD (which reports zero devices when the kernel module is absent), defaulting to MVGAL_ICD_ONLY=0 so vendor ICDs are preserved; the OpenCL ICD reports zero platforms when the kernel module is absent instead of exposing a half-working platform whose context creation fails; the GUI dashboard status cards auto-refresh on startup via a single-shot QTimer instead of staying "Checking…"; mvgal-bench compute and other single-suite paths now print the "What is actually measured" caveat like all; usage/error paths are standardized to exit 2 across tools (documented in man pages); the two configuration log lines are DEBUG-only and gated behind --verbose; %post removes leftover /etc/OpenCL/vendors/mvgal.icd.disabled; man pages were added for mvgal-gui, mvgal-dashboard, mvgal-load, mvgal-unload, mvgal-workload, mvgal-probe; and mvgal-daemon.preset (disabled) documents that %post enables the unit on first install. COPR build #11217288 succeeded on all 18 enabled chroots.

v0.7.29 is a test-report-driven correctness release addressing 15 issues from the v0.7.28 field validation on a single-GPU Secure Boot host. The headline fixes: the Vulkan layer activation variable was documented as VK_LAYER_MVGAL=1/ENABLE_MVGAL=1 but the manifest only enables on MVGAL_VULKAN_ENABLE=1; /etc/mvgal is now explicitly 0755 root:root in the package (was 0775 group:mvgal, a privilege-escalation path); mvgal status and mvgal-status now use a three-tier exit code (0 healthy, 1 degraded, 2 daemon down) instead of always returning 1; benchmark subcommands forward mvgal-bench output instead of suppressing it; mvgal set-strategy now refuses multi-GPU strategies on single-GPU machines; mvgal-config show-config handles ENUM-typed entries and shows only the effective view by default; -c/--config now rejects paths outside /etc/mvgal/; start/stop are idempotent (exit 0 when already in the target state); the OpenCL ICD park location moved to /usr/lib/mvgal/; mvgal-workload skips the MVGAL aggregating ICD platform to avoid double-counting; mvgal-probe demotes its /dev/mvgal0 open failure to a notice; and listen_addr defaults to 127.0.0.1.

v0.7.28 is a correctness and packaging release. The CUDA shim reported the MVGAL GPU count from cuDeviceGetCount/cudaGetDeviceCount instead of the native driver's, and mapped CUDA ordinals to MVGAL GPUs by modulo, so CUDA device 0 (RTX 3050) was attributed to MVGAL GPU 0 (Intel Arc). The shim now probes each ordinal's PCI BDF once and matches it against the NVIDIA MVGAL GPUs, passes native counts through, rejects the -1 sentinel before context creation, allocation and launch (the old mvgal_gpu < gpu_count guard was true for -1), restricts kernel launches to eligible GPUs, and lets MVGAL_CUDA_DEBUG=1 raise the log level. A new test_cuda_multidevice harness brings ctest to 27/27 (4.73 s). The release also adds a Windows cross-compilation packaging pipeline (NSIS and WiX installers from Jenkinsfile.windows), and the RPM ships a systemd-sysusers fragment so the mvgal group is declared before dependency resolution — dnf install no longer fails with nothing provides group(mvgal). See Changelog § v0.7.28.

v0.7.27 fixes three hardware-detection and reporting defects. The Vulkan ICD filename check did not match Fedora's radeon_icd.x86_64.json, so AMD GPUs were reported as having no Vulkan driver; the AMD shader-backend probe pointed at mesa's software rasterizer instead of RADV; the WoW64 layer check only looked in /usr/lib and reported "not built" on a correctly installed multilib Fedora package; and mvgal-info left NVIDIA temperature and utilization at their -1 sentinel — the proprietary driver exposes no hwmon source for them — and now fills only those empty fields via NVML, never fabricating a value. See Changelog § v0.7.27.

v0.7.26 is a packaging release with no runtime behaviour changes. The RPM files list now always includes mvgal-workload, which the build produces whenever CMake finds OpenCL — the previous conditional listing made a Fedora build that found OpenCL fail in %install. The AppImage dependency self-check now resolves not found entries and verifies against the staged usr/lib, and the application icons are normalized to 256×256 for the Flatpak, RPM and AppImage exports. See Changelog § v0.7.26.

v0.7.25 is a packaging and reporting release. Its headline defect made the modules unloadable and the documented Secure Boot recovery unable to repair them: the modules were compressed with xz -9e, producing a CRC64 xz stream that the kernel's decoder rejects (it accepts CRC32 and below), so the module failed at decompression — before the signature was examined. A user who enrolled the MOK correctly still had a system that would not load MVGAL, and the journal pointed at a signature problem that was not the cause. CRC64 is xz's default check at every compression level, so xz -9 was equally broken. v0.7.25 uses the kernel's own settings from scripts/Makefile.modinst (xz --check=crc32 --lzma2=dict=1MiB), which is also byte-identical to what DKMS builds.

Two related gaps were closed with it: the daemon had chowned systemd's RuntimeDirectory, which made every systemctl reload mvgald fail with status=233 while restart kept working; and SIGHUP was installed but the handler had no case for it, so a reload did nothing and reported success. The remainder of the release is reporting: the Vulkan probe called a staged manifest "ACTIVE", the D-Bus methods were registered without SD_BUS_VTABLE_UNPRIVILEGED (so sd-bus refused every non-root call and left the GUI and dashboard dead for members of the mvgal group), the OpenCL ICD registration was frozen at install time, and a multi-GPU strategy was accepted on a single-GPU host. See Changelog § v0.7.25.

The 0.7.23–0.7.24 releases mainly fixed correctness and observability issues found by packaging and exercising the installed system. The latest audit fixed daemon device access and D-Bus name ownership checks; PCIe generation, NUMA and class-code reporting; sysfs setup and kernel work/reference lifecycle defects; GPU memory totals; and inert power-control paths. It added mvgal-workload, an end-to-end OpenCL workload gate requiring at least two distinct vendors. See Changelog for the full source-derived details.

Runtime boundaries

The kernel module is above-driver discovery/control scaffolding. It does not take ownership of GPU PCI functions; native vendor drivers remain in charge of GPU submission and their memory managers.
Capabilities are intended to come from runtime probes. Unsupported kernel submission and VRAM allocation are not made functional by ioctl declarations or adapter files; callers must honor explicit unsupported results and fallbacks.
The Vulkan layer and Vulkan ICD are distinct components. The ICD does not provide a fabricated physical device representing multiple vendors.
API interception, strategy identifiers and memory APIs do not prove safe work distribution. Correctness depends on workload/API semantics, available backend operations and probed capabilities.
DMA-BUF/P2P paths depend on the device pair and driver support. Host staging is the fallback where supported by the particular userspace path.
Moore Threads and Adreno adapters and Windows/Steam compatibility code are experimental integration surfaces. Treat each as unvalidated until the exact hardware, driver and workload path has been exercised.

Build and validation

CMake is the primary broad build; Meson and Cargo cover separate portions. Optional dependencies and build switches determine which targets exist. Unit/integration tests, synthetic topology, successful package builds and physical multi-vendor validation provide different evidence and should be reported separately. The synthetic topology backend is a test fixture and does not create devices.

The source test table records 27/27 CTest tests (4.73 s, measured 2026-10-25) and 71/71 across the full suite, including the new test_cuda_multidevice and uapi_probe on the three-vendor bench host (AMD Navi 24, NVIDIA RTX 3050, Intel Arc A380). This is bench evidence for that host, not a universal hardware claim.

The source changelog's 0.7.24 known-issues entry records a PCI enumeration reference-lifetime fix that still needs a clean-boot hardware check on the host where it was found; the 0.7.28 entry does not restate it. The 2026-10-25 bench run (ctest 27/27, including uapi_probe, on the three-vendor bench host) does not substitute for that check on the affected machine. Do not turn this source-level fix into a claim of completed physical validation.

Package state

This docs repository's package/ directory contains 0.7.28 RPM, DEB, AppImage, tarball and Flatpak artifacts as of this update. Version 0.7.31 has been built and pushed to COPR (axogm/mvgal); install with pkexec dnf install mvgal from the repository configured for v0.7.31. COPR build #11229933 succeeded on fedora-43-x86_64. Its package README and individual manifests should be consulted for exact contents. Presence here does not establish that a remote repository has published the same version or that the artifacts have passed hardware validation.