Add Windows platform support (x86_64-pc-windows-msvc)

[arndawg]

Summary

Native Windows port of siomon with full feature parity for user-mode hardware
monitoring. The Linux version is completely unaffected — all Windows code is
behind #[cfg(windows)] and [target.'cfg(windows)'.dependencies].

What it provides on Windows

154 real-time sensors without any kernel driver:

• Per-core CPU frequency (via CallNtPowerInformation) and utilization
• NVIDIA GPU temperature, fan, power, clocks, utilization, VRAM (via NVML/nvml.dll)
• Per-disk I/O and per-adapter network throughput
• WHEA hardware error monitoring (Windows Event Log)
• ACPI thermal zones

Full system inventory:

• CPU: CPUID, topology via sysinfo, cache, features, microcode from registry
• Memory: sysinfo + SMBIOS via GetSystemFirmwareTable (DIMM details)
• Motherboard: WMI + registry (board, BIOS, Secure Boot, chipset, serial, UUID)
• GPU: NVML + EnumDisplayDevices (display outputs)
• Storage: sysinfo + NVMe SMART (IOCTL_STORAGE_QUERY_PROPERTY) + SATA SMART (SMART_RCV_DRIVE_DATA)
• Network: GetAdaptersAddresses (MAC, IPs, speed, MTU, driver name)
• PCI: WMI + pci_ids name resolution (87+ devices)
• USB: WMI with speed classification (Full/High/Super)
• Audio: WMI with HDAUDIO codec detection
• Battery: WMI Win32_Battery

WinRing0 optional (~210+ sensors with driver):

• SuperIO temps/fans/voltages (NCT67xx, IT87xx — register code already portable)
• RAPL power via MSR
• AMD HSMP via SMN mailbox
• PCIe link speed/width
• I2C/SMBus VRM + DDR5 DIMM temps
• AMD ADL GPU sensors

Architecture

All Windows code uses #[cfg(windows)] / #[cfg(not(unix))]. New files:

src/platform/
  winring0.rs, port_io_win.rs, msr_win.rs    — WinRing0 abstraction
  sinfo_io_win.rs                              — HwmAccess for SuperIO
  nvme_win.rs, sata_win.rs                     — SMART ioctls
  smbus_win.rs                                 — SMBus host controller
  adl.rs                                       — AMD ADL GPU wrapper

src/sensors/
  rapl_win.rs, hsmp_win.rs                     — Power/telemetry via MSR/SMN
  smbus_win.rs                                 — PMBus VRM + SPD5118 DIMM
  gpu_sensors_adl.rs                           — AMD GPU sensor source
  whea.rs                                      — WHEA error monitoring
  ipmi_win.rs                                  — IPMI via ipmitool
  acpi_thermal_win.rs                          — ACPI thermal zones

Dependencies (Windows-only)

[target.'cfg(windows)'.dependencies]
sysinfo = "0.33"
winapi = { version = "0.3", features = [...] }
libloading = "0.8"

No impact on Linux Cargo.lock or build.

CI

Added windows-check, windows-build, windows-test jobs to ci.yml.
Added build-windows job to release.yml (zip + SHA256 artifact).

Testing

36/36 automated CLI smoke tests passing on Windows (tests/cli_smoke_test.sh).
Zero warnings with cargo clippy and cargo fmt --check.

Build

cargo build --release --target x86_64-pc-windows-msvc

Stats

~70 files changed, +8,595 / -453 lines across the port.

[Copilot]

Pull request overview

Adds native Windows (x86_64-pc-windows-msvc) support to siomon, including Windows-specific collectors/sensors, WinRing0-gated low-level access, updated packaging (winget), and CI/release workflows to build/test on Windows while keeping Linux behavior behind #[cfg] gates.

Changes:

• Implement Windows collectors + sensor sources (CPU freq/util, disk/network throughput, WHEA, ACPI thermal, IPMI, RAPL/WinRing0 paths).
• Add Windows platform modules for WinRing0, SMBus, NVMe/SATA SMART IOCTLs, and related accessors.
• Add Windows CI/release jobs and winget manifests/docs for distribution.

Reviewed changes

Copilot reviewed 66 out of 67 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
winget/manifests/a/arndawg/sio/0.1.3-win.3/arndawg.sio.yaml	Winget version manifest for Windows release.
winget/manifests/a/arndawg/sio/0.1.3-win.3/arndawg.sio.locale.en-US.yaml	Winget locale/metadata (description, tags, release notes).
winget/manifests/a/arndawg/sio/0.1.3-win.3/arndawg.sio.installer.yaml	Winget installer manifest (zip portable, URL/SHA, release date).
tests/cli_smoke_test.sh	Windows CLI smoke test script for sio.exe.
src/sensors/whea.rs	New WHEA sensor source via wevtutil.
src/sensors/superio/nct67xx.rs	Platform-conditional HWM access import for Windows.
src/sensors/superio/ite87xx.rs	Platform-conditional Port I/O import for Windows.
src/sensors/superio/chip_detect.rs	Windows Port I/O support + kernel-driver-loaded check gating.
src/sensors/rapl_win.rs	New Windows RAPL power source via MSR reads (WinRing0).
src/sensors/poller.rs	Split unix vs non-unix source discovery; add Windows sources.
src/sensors/network_stats.rs	Add Windows implementation using sysinfo::Networks.
src/sensors/mod.rs	Add Windows-only sensor modules (ACPI thermal, WHEA, IPMI, RAPL, etc.).
src/sensors/ipmi_win.rs	New Windows IPMI source via ipmitool background poller.
src/sensors/gpu_sensors_adl.rs	New AMD GPU sensors via ADL (Windows).
src/sensors/gpu_sensors.rs	Gate Linux AMD GPU sysfs polling behind #[cfg(unix)].
src/sensors/disk_activity.rs	Add Windows disk throughput implementation via sysinfo::Disks.
src/sensors/cpu_util.rs	Add Windows CPU utilization implementation via sysinfo.
src/sensors/cpu_freq.rs	Split Linux vs Windows CPU frequency implementations.
src/sensors/acpi_thermal_win.rs	New ACPI thermal zone temperatures via PowerShell/WMI JSON.
src/platform/winring0.rs	WinRing0 DLL wrapper singleton (port I/O, MSR, PCI config).
src/platform/smbus_win.rs	SMBus controller access via WinRing0 (AMD FCH).
src/platform/sinfo_io_win.rs	Windows HwmAccess implementation for Super I/O register reads.
src/platform/sata_win.rs	SATA SMART IOCTL implementation + SMART parsing helpers.
src/platform/port_io_win.rs	Windows Port I/O adapter backed by WinRing0.
src/platform/nvml.rs	Load NVML from nvml.dll on Windows.
src/platform/nvme_win.rs	NVMe SMART IOCTL implementation + SMART parsing helpers.
src/platform/msr_win.rs	Windows MSR read helper via WinRing0.
src/platform/mod.rs	Add Windows platform modules + is_elevated() implementation.
src/parsers/smbios.rs	Add Windows SMBIOS parsing via GetSystemFirmwareTable('RSMB') + Type 16 support.
src/output/tui/mod.rs	Use cross-platform is_elevated() and adjust privilege hint for Windows.
src/output/text.rs	Use is_elevated(), Windows storage labels, and improved NIC detail formatting.
src/output/html.rs	Make device label portable (/dev/* only on unix).
src/model/network.rs	Add Display impl for NetworkInterfaceType.
src/main.rs	Windows hostname/kernel version paths + chipset fill heuristic + Windows OS/version helpers.
src/collectors/network.rs	Windows adapter enumeration via GetAdaptersAddresses + WMI driver names.
src/collectors/motherboard.rs	Windows motherboard/BIOS fields via wmic/registry + cfg gating for unix helpers.
src/collectors/memory.rs	Windows memory info via sysinfo + DIMM details via SMBIOS parse.
src/collectors/cpu.rs	Windows topology/frequency stubs via sysinfo + microcode from registry.
src/collectors/battery.rs	Windows battery via PowerShell/WMI + non-unix/non-windows fallback.
src/cli.rs	Update CLI description and direct-io help text for Windows.
phase_status.json	Tracks port phases/status and planned ZETA/ETA work items.
cli_testing_results.md	Windows test results summary (smoke + manual).
archive/PHASE_GAMMA_EXIT_REPORT.md	Archived phase report (WinRing0-gated sensors).
archive/PHASE_EPSILON_EXIT_REPORT.md	Archived phase report (CI/release/winget).
archive/PHASE_EPSILON.md	Archived plan doc for CI/release/winget phase.
archive/PHASE_DELTA_EXIT_REPORT.md	Archived phase report (SMBIOS/DIMM/microcode polish).
archive/PHASE_DELTA.md	Archived plan doc for polish/upstream prep.
archive/PHASE_BETA_EXIT_REPORT.md	Archived phase report (user-mode parity items).
README.md	Document Windows build/install and Windows-port feature summary.
PHASE_ZETA.md	Plan doc for distribution/upstream merge and winget automation.
Cargo.toml	Move nix to unix-only deps; add Windows-only deps (sysinfo, winapi, libloading).
Cargo.lock	Lockfile updates for Windows dependency graph.
.github/workflows/release.yml	Add Windows release build + zip/SHA artifacts; adjust GitHub Release job dependencies.
.github/workflows/ci.yml	Add Windows check/build/test jobs.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[arndawg]

@copilot apply changes based on this feedback

[arndawg]

Fixed CI lint failures — formatting drift in 4 files (storage.rs, text.rs, mod.rs, nvme_win.rs) that were modified after the initial lint cleanup. The fix has been merged to main on the fork. CI should pass on the next run.

Commit: arndawg/siomon-win@e83ba46

[arndawg]

Additional clippy fix pushed — unused_mut on motherboard variable (only mutated on Windows for chipset-from-PCI). Added #[allow(unused_mut)] since the mutation is platform-conditional.

CI should be green on the next run.

[eous]

Hey, really appreciate the effort here. This is a serious amount of work and the Windows platform code is genuinely well done. The FFI is clean, handles get closed properly, the cfg gating strategy makes sense, and it doesn't break Linux. I want to get this merged.

That said there's a bunch of stuff that needs to be cleaned up before we can pull this in. Nothing that changes the core work, mostly housekeeping and a few bugs.

Files to remove

About a third of this PR is development planning docs and fork specific infrastructure that shouldn't land in the upstream repo.

Please remove all of these

PHASE_ZETA.md
PHASE_ZETA_EXIT_REPORT.md
archive/PHASE_ALPHA.md
archive/PHASE_ALPHA_EXIT_REPORT.md
archive/PHASE_BETA.md
archive/PHASE_BETA_EXIT_REPORT.md
archive/PHASE_DELTA.md
archive/PHASE_DELTA_EXIT_REPORT.md
archive/PHASE_EPSILON.md
archive/PHASE_EPSILON_EXIT_REPORT.md
archive/PHASE_GAMMA.md
archive/PHASE_GAMMA_EXIT_REPORT.md
cli_testing_plan.md
cli_testing_results.md
phase_status.json
tests/cli_smoke_test.sh

Also remove the entire winget/ directory and .github/workflows/winget-update.yml. The winget manifests point to your fork releases and the workflow references arndawg.sio plus a WINGET_TOKEN secret that doesn't exist in this repo. If we want winget support later we'll set it up under level1techs.

README

The README has references to arndawg/siomon-win and winget install arndawg.sio that need to come out. Also the "Windows Port" section is pretty long at 75 lines. Would be great to trim it down to a concise "Platform Support" section. The "Recent fixes" subsection reads like a changelog and the references to cli_testing_results.md and cli_testing_plan.md point to files that won't exist after cleanup.

Bugs in src/platform/adl.rs

Found three issues in the AMD ADL wrapper that should be fixed before merge.

1. The adl_malloc callback uses std::alloc::alloc with alignment 8, but ADL internally calls C free() on memory returned by this callback. That's an allocator mismatch. Also no guard against negative or zero size from the c_int parameter. Fix is to use libc::malloc instead and return null for size <= 0.

2. The AdlAdapterInfo struct has no padding for forward compatibility. AMD has added fields to this struct across SDK versions. If someone has a newer driver, ADL writes past the end of your allocation. Add a _reserved: [u8; 128] at the end or use a flat buffer approach.

3. read_c_string uses CStr::from_ptr which scans forward with no length bound. If a char array isn't null terminated (especially possible in combination with the struct size issue above) this reads past the buffer. Safer to scan within the slice length manually.

cfg(not(unix)) should be cfg(windows)

There are about 37 places where #[cfg(not(unix))] is used on code that is clearly Windows specific (PowerShell commands, WMI queries, winapi calls, registry access). cfg(not(unix)) would also match WASM and bare metal targets. Please swap these to #[cfg(windows)] for correctness.

Files affected include collectors/cpu.rs, collectors/memory.rs, collectors/motherboard.rs, collectors/network.rs, collectors/storage.rs, sensors/cpu_util.rs, sensors/disk_activity.rs, sensors/network_stats.rs, sensors/poller.rs, sensors/superio/chip_detect.rs, and main.rs.

wmic deprecation (non blocking, but worth noting)

src/collectors/motherboard.rs uses wmic baseboard get, wmic bios get etc. Microsoft deprecated wmic back in Windows 10 21H1. All your other collectors already use PowerShell Get-CimInstance which is the replacement. Would be nice to migrate this one too for consistency. If you want to do it as a follow up PR that's fine.

Minor stuff (non blocking)

These don't need to block merge but would be nice to clean up

to_wide() is defined in three places (nvme_win.rs, sata_win.rs, storage.rs). Could be one shared utility.

The PowerShell JSON normalization pattern (wrapping bare objects in an array) is repeated in 5+ collectors. Good candidate for a helper function.

PortIo on Windows calls WinRing0::try_load() on every single I/O operation. The OnceLock means it only loads once but it's still an Option check per register read during SuperIO polling. Holding a &'static WinRing0 reference would be cleaner.

SMBus operations in smbus_win.rs do multi step register sequences that aren't atomic. Currently safe because the poller is sequential but worth a comment or a Mutex for future proofing.

Storage interface classification maps all SSDs to NVMe initially. SATA SSDs get corrected when SMART attaches but drives without SMART stay mislabeled.

What's good

The WinRing0 singleton pattern with OnceLock is clean. Handle lifetime management is correct everywhere I checked. The NVML wrapper sharing between platforms via cfg on the DLL name is elegant. The SensorSource trait impls all conform properly. The cpu_freq.rs pattern with internal mod platform and pub use is the best approach in the PR and could be a model for the others.

No security concerns with the actual Rust code. All 14 shell command invocations use hardcoded arguments with no user input. Dependencies are all legitimate high download count crates.

Looking forward to the cleaned up version. This will be a great addition to the project.

[eous]

std::alloc::alloc with alignment 8 here, but ADL internally calls C free() on memory returned by this callback. That is an allocator mismatch that can corrupt the heap.

Also no guard against negative or zero size. If ADL passes a negative value (error condition, driver bug), the size as usize cast wraps to ~4GB and Layout::from_size_align_unchecked hits undefined behavior.

Fix would be to use libc::malloc for guaranteed CRT compatibility and return null for size <= 0.

extern "system" fn adl_malloc(size: c_int) -> *mut c_void {
    if size <= 0 {
        return std::ptr::null_mut();
    }
    unsafe { libc::malloc(size as usize) as *mut c_void }
}

[eous]

This struct has no padding for forward compatibility. AMD has added fields to AdlAdapterInfo across SDK versions (iInfoMask, iInfoValue in newer releases). If someone has a newer driver than the SDK version this struct was based on, ADL2_Adapter_AdapterInfo_Get writes past the end of each array element. With count adapters that is silent memory corruption across the entire info array.

Either add safety padding at the end

pub os_display_index: c_int,
_reserved: [u8; 128],

or allocate a generously sized flat buffer and read individual entries at offsets based on the size field that ADL populates in each entry.

[eous]

CStr::from_ptr scans forward until it finds a null byte with no length bound. If the char array is not null terminated (possible with the struct size mismatch above, or corrupted data from the driver) this reads past the buffer into whatever is next in memory.

Safer to bound the scan to the slice length

fn read_c_string(buf: &[c_char]) -> String {
    let bytes = unsafe {
        std::slice::from_raw_parts(buf.as_ptr() as *const u8, buf.len())
    };
    let end = bytes.iter().position(|&b| b == 0).unwrap_or(bytes.len());
    String::from_utf8_lossy(&bytes[..end]).into_owned()
}

[arndawg]

gotcha, let me clean up a bit, of course i want my local repo to have the history because it helps make the agentic work smooth. i'll work on cleaning stuff up. --- Arni

…

On Tue, Mar 17, 2026 at 12:18 PM Patrick Buckley ***@***.***> wrote: ***@***.**** requested changes on this pull request. Hey, really appreciate the effort here. This is a serious amount of work and the Windows platform code is genuinely well done. The FFI is clean, handles get closed properly, the cfg gating strategy makes sense, and it doesn't break Linux. I want to get this merged. That said there's a bunch of stuff that needs to be cleaned up before we can pull this in. Nothing that changes the core work, mostly housekeeping and a few bugs. Files to remove About a third of this PR is development planning docs and fork specific infrastructure that shouldn't land in the upstream repo. Please remove all of these PHASE_ZETA.md PHASE_ZETA_EXIT_REPORT.md archive/PHASE_ALPHA.md archive/PHASE_ALPHA_EXIT_REPORT.md archive/PHASE_BETA.md archive/PHASE_BETA_EXIT_REPORT.md archive/PHASE_DELTA.md archive/PHASE_DELTA_EXIT_REPORT.md archive/PHASE_EPSILON.md archive/PHASE_EPSILON_EXIT_REPORT.md archive/PHASE_GAMMA.md archive/PHASE_GAMMA_EXIT_REPORT.md cli_testing_plan.md cli_testing_results.md phase_status.json tests/cli_smoke_test.sh Also remove the entire winget/ directory and .github/workflows/winget-update.yml. The winget manifests point to your fork releases and the workflow references arndawg.sio plus a WINGET_TOKEN secret that doesn't exist in this repo. If we want winget support later we'll set it up under level1techs. README The README has references to arndawg/siomon-win and winget install arndawg.sio that need to come out. Also the "Windows Port" section is pretty long at 75 lines. Would be great to trim it down to a concise "Platform Support" section. The "Recent fixes" subsection reads like a changelog and the references to cli_testing_results.md and cli_testing_plan.md point to files that won't exist after cleanup. Bugs in src/platform/adl.rs Found three issues in the AMD ADL wrapper that should be fixed before merge. 1. The adl_malloc callback uses std::alloc::alloc with alignment 8, but ADL internally calls C free() on memory returned by this callback. That's an allocator mismatch. Also no guard against negative or zero size from the c_int parameter. Fix is to use libc::malloc instead and return null for size <= 0. 2. The AdlAdapterInfo struct has no padding for forward compatibility. AMD has added fields to this struct across SDK versions. If someone has a newer driver, ADL writes past the end of your allocation. Add a _reserved: [u8; 128] at the end or use a flat buffer approach. 3. read_c_string uses CStr::from_ptr which scans forward with no length bound. If a char array isn't null terminated (especially possible in combination with the struct size issue above) this reads past the buffer. Safer to scan within the slice length manually. cfg(not(unix)) should be cfg(windows) There are about 37 places where #[cfg(not(unix))] is used on code that is clearly Windows specific (PowerShell commands, WMI queries, winapi calls, registry access). cfg(not(unix)) would also match WASM and bare metal targets. Please swap these to #[cfg(windows)] for correctness. Files affected include collectors/cpu.rs, collectors/memory.rs, collectors/motherboard.rs, collectors/network.rs, collectors/storage.rs, sensors/cpu_util.rs, sensors/disk_activity.rs, sensors/network_stats.rs, sensors/poller.rs, sensors/superio/chip_detect.rs, and main.rs. wmic deprecation (non blocking, but worth noting) src/collectors/motherboard.rs uses wmic baseboard get, wmic bios get etc. Microsoft deprecated wmic back in Windows 10 21H1. All your other collectors already use PowerShell Get-CimInstance which is the replacement. Would be nice to migrate this one too for consistency. If you want to do it as a follow up PR that's fine. Minor stuff (non blocking) These don't need to block merge but would be nice to clean up to_wide() is defined in three places (nvme_win.rs, sata_win.rs, storage.rs). Could be one shared utility. The PowerShell JSON normalization pattern (wrapping bare objects in an array) is repeated in 5+ collectors. Good candidate for a helper function. PortIo on Windows calls WinRing0::try_load() on every single I/O operation. The OnceLock means it only loads once but it's still an Option check per register read during SuperIO polling. Holding a &'static WinRing0 reference would be cleaner. SMBus operations in smbus_win.rs do multi step register sequences that aren't atomic. Currently safe because the poller is sequential but worth a comment or a Mutex for future proofing. Storage interface classification maps all SSDs to NVMe initially. SATA SSDs get corrected when SMART attaches but drives without SMART stay mislabeled. What's good The WinRing0 singleton pattern with OnceLock is clean. Handle lifetime management is correct everywhere I checked. The NVML wrapper sharing between platforms via cfg on the DLL name is elegant. The SensorSource trait impls all conform properly. The cpu_freq.rs pattern with internal mod platform and pub use is the best approach in the PR and could be a model for the others. No security concerns with the actual Rust code. All 14 shell command invocations use hardcoded arguments with no user input. Dependencies are all legitimate high download count crates. Looking forward to the cleaned up version. This will be a great addition to the project. — Reply to this email directly, view it on GitHub <#14 (review)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABAQEMFSFLLQGL3DMPS7VZD4RGQJPAVCNFSM6AAAAACWT63QMKVHI2DSMVQWIX3LMV43YUDVNRWFEZLROVSXG5CSMV3GSZLXHMZTSNRTGE4DSNZZHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[emansom]

Is the unsecure WinRing0 still the only way to get to the hardware? Has Microsoft not implemented a better mechanism since the whole 2024 CrowdStrike incident?

[arndawg]

They haven't implemented anything yet as far as I know. On linux you've already got the infrastructure setup.

…

--- Arni

On Wed, Mar 18, 2026, 11:22 AM Ewout van Mansom ***@***.***> wrote: *emansom* left a comment (level1techs/siomon#14) <#14 (comment)> Is the unsecure WinRing0 still the only way to get to the hardware? Has Microsoft not implemented a better mechanism since the whole 2024 CrowdStrike incident <https://en.wikipedia.org/wiki/2024_CrowdStrike-related_IT_outages>? — Reply to this email directly, view it on GitHub <#14 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABAQEMHPW2XV34PRQE4VW2T4RLSOXAVCNFSM6AAAAACWT63QMKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DAOBUGY4TONZRHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>