feat(aarch64): add WHP backend for ARM64 Windows by cshung · Pull Request #1638 · hyperlight-dev/hyperlight

cshung · 2026-07-12T15:02:55Z

Summary

Implements the WHP (Windows Hypervisor Platform) hypervisor backend for aarch64, enabling hyperlight to run micro-VMs on Windows ARM64 systems.

Resolves #1544

Changes

Structural

Restructure whp.rs into whp/ directory (mod.rs + x86_64.rs) to support per-architecture implementations, matching the existing kvm/ and mshv/ pattern
Fix super::x86_64::hw_interrupts module path after directory restructure

New: `whp/aarch64.rs`

Manual FFI bindings for ARM64 WHP register names, exit reasons, and exit context layout (extracted from Windows SDK WinHvPlatformDefs.h)
Full VirtualMachine trait implementation with:
- MMIO-based I/O handling (ARM64 has no IO ports)
- Register get/set via WHvGet/SetVirtualProcessorRegisters
- Surrogate process support (same pattern as x86_64)

Integration

Wire WhpVm into hyperlight_vm/aarch64.rs for Windows platform
Add WindowsInterruptHandle for aarch64 Windows
Add CpuVendor::current() for aarch64 Windows target

Cross-compilation fix

Move vmm-sys-util to unix-only dependencies (it doesn't compile on Windows)

Verification

Verified compilation on three targets:

x86_64-pc-windows-msvc (native) — just clippy debug/release pass
aarch64-pc-windows-msvc (cross-compile) — cargo check passes
aarch64-unknown-linux-gnu (cross-compile) — cargo check --features kvm passes

Note: No runtime testing was possible — requires actual ARM64 Windows hardware with Hyper-V enabled. CI has no ARM64 Windows runners.

Why manual FFI bindings?

The windows crate (v0.62) does not expose ARM64 WHP types (register names, exit reasons, exit context structs). All definitions were extracted from the Windows SDK header WinHvPlatformDefs.h (SDK 10.0.26100.0) which has full ARM64 support behind #ifdef _ARM64_.

Copilot

Pull request overview

Adds a Windows Hypervisor Platform (WHP) backend for Windows/aarch64 so Hyperlight can run micro-VMs on ARM64 Windows systems, aligning WHP’s structure with existing per-arch hypervisor layouts.

Changes:

Introduces whp/aarch64.rs implementing the VirtualMachine trait for ARM64 WHP, including MMIO-based exit handling and register get/set via WHvGet/SetVirtualProcessorRegisters.
Restructures the WHP backend into whp/ with per-architecture modules and fixes x86_64 interrupt helper module paths.
Wires WHP into the aarch64 Hyperlight VM path on Windows and moves vmm-sys-util into unix-only dependencies to fix Windows builds.

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
src/hyperlight_host/src/sandbox/snapshot/file/config.rs	Adds Windows/aarch64 CPU vendor token for snapshot config.
src/hyperlight_host/src/hypervisor/virtual_machine/whp/x86_64.rs	Updates `hw_interrupts` module paths after WHP directory restructure.
src/hyperlight_host/src/hypervisor/virtual_machine/whp/mod.rs	New per-arch WHP module dispatcher (`x86_64` vs `aarch64`).
src/hyperlight_host/src/hypervisor/virtual_machine/whp/aarch64.rs	New ARM64 WHP backend implementation and manual ARM64 WHP FFI bindings.
src/hyperlight_host/src/hypervisor/hyperlight_vm/aarch64.rs	Selects WHP on Windows/aarch64 and adds Windows interrupt handle wiring.
src/hyperlight_host/Cargo.toml	Makes `vmm-sys-util` unix-only to avoid Windows compilation failures.

cshung · 2026-08-05T17:48:19Z

End-to-end cargo-hyperlight validation completed on the ARM64 WHP hardware.

Using this PR's 0.16 crates, a freshly scaffolded sample successfully built aarch64-hyperlight-none guests and ran through WHP in debug and release. Host callbacks, typed calls, persistent state, snapshot, and restore passed. I also ran the release workflow 10 times with surrogate mappings and 10 times with surrogates disabled. All runs produced the expected output.

The backend works. The published scaffold is the remaining user-experience gap: cargo-hyperlight pins Hyperlight 0.15 and hardcodes the generated host's x64 guest path. Filed hyperlight-dev/cargo-hyperlight#70 with the reproduction and results.

syntactically

This looks great! I have a bunch of little nits, and also I (more majorly) think that the addition of reset_vcpu to InteruptHandle seems wrong.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Signed-off-by: cshung <3410332+cshung@users.noreply.github.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 6f20a05d-6bee-4e2e-b320-12f8d9759bbc Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Use WHvResetPartition so snapshots preserve partition mappings while resetting virtual processor state. Keep reset ownership in the virtual machine abstraction. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Wait for process-wide partition capacity before allocating GIC-backed ARM64 partitions. Release capacity only after WHP teardown completes. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Match WHV_RUN_VP_EXIT_CONTEXT's 16-byte ABI alignment. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 6f20a05d-6bee-4e2e-b320-12f8d9759bbc Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Fall back to generic MMIO for undecodable ARM64 writes and share file-mapping cleanup across WHP backends. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 6f20a05d-6bee-4e2e-b320-12f8d9759bbc Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

hyperlight-gh-bot · 2026-09-29T17:21:26Z

Benchmark Results

Measured commit: 4ad16d1bd69c
Baseline commit: 5131e8332f88

kvm / amd (Linux) (➖ stable)

No benchmark improved or regressed.

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`752.93 ns` (➖ 1.12x faster)
`vec_bytes`	`582.74 ns` (➖ 1.02x slower)
		`374.36 µs` (➖ 1.00x slower)

payload_allocation

	`slot_pool_segmented`
`262144`	`537.05 ns` (➖ 1.01x slower)
`65536`	`144.34 ns` (➖ 1.01x slower)

sandboxes

	`create_initialized_and_drop`
`medium`	`80.25 ms` (➖ 1.06x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`7.76 ns` (➖ 1.00x slower)	`7.72 ns` (➖ 1.00x faster)	`7.95 ns` (➖ 1.05x faster)

snapshot_files

	`load_snapshot_unverified`
`small`	`89.66 µs` (➖ 1.04x faster)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`7.53 µs` (➖ 1.07x faster)	`6.77 µs` (➖ 1.07x faster)
`65536`	`1.93 µs` (➖ 1.00x slower)	`2.06 µs` (➖ 1.06x slower)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`6.43 µs` (➖ 1.07x faster)	`6.30 µs` (➖ 1.04x faster)
`8192`	`1.06 µs` (➖ 1.00x slower)	`1.07 µs` (➖ 1.00x faster)
`262144`		`26.92 µs` (➖ 1.05x faster)

kvm / intel (Linux) (➖ stable)

No benchmark improved or regressed.

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`677.16 ns` (➖ 1.39x faster)
`vec_bytes`	`510.21 ns` (➖ 1.17x faster)
		`675.09 µs` (➖ 1.20x faster)

payload_allocation

	`slot_pool_segmented`
`262144`	`499.45 ns` (➖ 1.12x faster)
`65536`	`136.76 ns` (➖ 1.11x faster)

sandboxes

	`create_initialized_and_drop`
`medium`	`81.08 ms` (➖ 1.28x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`7.11 ns` (➖ 1.08x faster)	`6.93 ns` (➖ 1.20x faster)	`7.04 ns` (➖ 1.10x faster)

snapshot_files

	`load_snapshot_unverified`
`small`	`45.29 µs` (➖ 1.08x faster)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`7.60 µs` (➖ 1.11x faster)	`7.56 µs` (➖ 1.10x faster)
`65536`	`2.13 µs` (➖ 1.09x faster)	`2.13 µs` (➖ 1.10x faster)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`7.14 µs` (➖ 1.22x faster)	`7.13 µs` (➖ 1.24x faster)
`8192`	`749.74 ns` (➖ 1.19x faster)	`728.88 ns` (➖ 1.18x faster)
`262144`		`29.41 µs` (➖ 1.20x faster)

mshv3 / amd (Linux) (❌ 1.83x)

Top regressions

sandboxes/create_initialized_and_drop/medium — ❌ 1.83x slower

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`957.07 ns` (➖ 1.04x faster)
`vec_bytes`	`728.02 ns` (➖ 1.08x slower)
		`269.47 µs` (➖ 1.29x faster)

payload_allocation

	`slot_pool_segmented`
`262144`	`720.19 ns` (➖ 1.01x slower)
`65536`	`192.87 ns` (➖ 1.01x slower)

sandboxes

	`create_initialized_and_drop`
`medium`	`55.68 ms` (❌ 1.83x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`9.72 ns` (➖ 1.00x faster)	`10.02 ns` (➖ 1.04x slower)	`9.73 ns` (➖ 1.00x faster)

snapshot_files

	`load_snapshot_unverified`
`small`	`84.90 µs` (➖ 1.09x faster)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`9.16 µs` (➖ 1.04x faster)	`9.03 µs` (➖ 1.00x faster)
`65536`	`2.24 µs` (➖ 1.06x faster)	`2.24 µs` (➖ 1.05x faster)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`8.19 µs` (➖ 1.10x faster)	`8.14 µs` (➖ 1.05x faster)
`8192`	`1.29 µs` (➖ 1.02x slower)	`1.36 µs` (➖ 1.09x slower)
`262144`		`36.98 µs` (➖ 1.05x faster)

mshv3 / intel (Linux) (❌ 1.57x)

Top regressions

sandboxes/create_initialized_and_drop/medium — ❌ 1.57x slower

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`823.49 ns` (➖ 1.18x faster)
`vec_bytes`	`595.75 ns` (➖ 1.03x faster)
		`770.42 µs` (➖ 1.07x slower)

payload_allocation

	`slot_pool_segmented`
`262144`	`603.91 ns` (➖ 1.01x slower)
`65536`	`162.65 ns` (➖ 1.04x slower)

sandboxes

	`create_initialized_and_drop`
`medium`	`61.28 ms` (❌ 1.57x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`8.25 ns` (➖ 1.03x slower)	`8.23 ns` (➖ 1.02x slower)	`8.25 ns` (➖ 1.03x slower)

snapshot_files

	`load_snapshot_unverified`
`small`	`45.74 µs` (➖ 1.01x slower)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`8.92 µs` (➖ 1.03x slower)	`8.97 µs` (➖ 1.03x slower)
`65536`	`2.52 µs` (➖ 1.04x slower)	`2.51 µs` (➖ 1.02x slower)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`8.50 µs` (➖ 1.06x faster)	`8.50 µs` (➖ 1.05x faster)
`8192`	`882.12 ns` (➖ 1.01x slower)	`880.96 ns` (➖ 1.05x faster)
`262144`		`34.63 µs` (➖ 1.05x faster)

hyperv-ws2025 / amd (Windows) (❌ 1.76x)

Top regressions

sandboxes/create_initialized_and_drop/medium — ❌ 1.76x slower

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`1.12 µs` (➖ 1.03x faster)
`vec_bytes`	`799.56 ns` (➖ 1.00x faster)
		`2.15 ms` (➖ 1.18x slower)

payload_allocation

	`slot_pool_segmented`
`262144`	`780.81 ns` (➖ 1.03x faster)
`65536`	`242.46 ns` (➖ 1.10x slower)

sandboxes

	`create_initialized_and_drop`
`medium`	`113.51 ms` (❌ 1.76x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`10.56 ns` (➖ 1.05x slower)	`10.56 ns` (➖ 1.04x slower)	`10.57 ns` (➖ 1.04x slower)

snapshot_files

	`load_snapshot_unverified`
`small`	`725.07 µs` (➖ 1.11x slower)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`8.83 µs` (➖ 1.09x faster)	`10.14 µs` (➖ 1.05x slower)
`65536`	`2.57 µs` (➖ 1.07x slower)	`2.57 µs` (➖ 1.08x slower)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`8.76 µs` (➖ 1.02x faster)	`9.10 µs` (➖ 1.02x slower)
`8192`	`1.45 µs` (➖ 1.12x slower)	`1.44 µs` (➖ 1.12x slower)
`262144`		`39.45 µs` (➖ 1.01x faster)

hyperv-ws2025 / intel (Windows) (❌ 1.51x)

Top regressions

virtq_readwrite/slot_pool_segmented/262144 — ❌ 1.51x slower

Benchmark Results

function_call_codec

	`encode_control`	`decode_vec_bytes_copy`
`byte_chunks`	`1.21 µs` (➖ 1.04x slower)
`vec_bytes`	`801.52 ns` (➖ 1.04x slower)
		`2.96 ms` (➖ 1.01x slower)

payload_allocation

	`slot_pool_segmented`
`262144`	`755.80 ns` (➖ 1.05x faster)
`65536`	`212.12 ns` (➖ 1.11x faster)

sandboxes

	`create_initialized_and_drop`
`medium`	`106.77 ms` (➖ 1.18x slower)

slot_pool

	`alloc_dealloc_1500`	`alloc_dealloc_4096`	`alloc_dealloc_128`
	`10.48 ns` (➖ 1.04x slower)	`9.97 ns` (➖ 1.07x faster)	`9.97 ns` (➖ 1.01x faster)

snapshot_files

	`load_snapshot_unverified`
`small`	`608.08 µs` (➖ 1.26x slower)

virtq_readonly

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`262144`	`7.81 µs` (➖ 1.00x faster)	`7.64 µs` (➖ 1.02x faster)
`65536`	`2.38 µs` (➖ 1.02x slower)	`2.31 µs` (➖ 1.00x faster)

virtq_readwrite

	`slot_pool_segmented_fragmented`	`slot_pool_segmented`
`65536`	`8.49 µs` (➖ 1.01x faster)	`8.30 µs` (➖ 1.00x slower)
`8192`	`1.15 µs` (➖ 1.05x slower)	`1.11 µs` (➖ 1.04x faster)
`262144`		`58.49 µs` (❌ 1.51x slower)

_{Reported by cargo ci bench-report --candidate run:36597233968 --baseline run:36361996946 --config-file bench_report.toml.}

Copilot AI review requested due to automatic review settings July 12, 2026 15:02

cshung added kind/enhancement For PRs adding features, improving functionality, docs, tests, etc. ready-for-review PR is ready for (re-)review labels Jul 12, 2026

cshung requested review from andreiltd, danbugs, dblnz, devigned, jprendes, jsturtevant, ludfjig, simongdavies, squillace and syntactically as code owners July 12, 2026 15:02

Copilot started reviewing on behalf of cshung July 12, 2026 15:06 View session

Copilot AI reviewed Jul 12, 2026

View reviewed changes

cshung marked this pull request as draft July 13, 2026 02:13

github-actions Bot removed the ready-for-review PR is ready for (re-)review label Jul 13, 2026

syntactically reviewed Jul 15, 2026

View reviewed changes

Comment thread src/hyperlight_host/src/sandbox/snapshot/file/config.rs Outdated

cshung force-pushed the cshung/whp-aarch64 branch from 5f992dc to 1d3a299 Compare July 30, 2026 22:14

jsturtevant mentioned this pull request Aug 5, 2026

cargo hyperlight aarch64 support #1546

Closed

2 tasks

syntactically requested changes Aug 5, 2026

View reviewed changes

cshung force-pushed the cshung/whp-aarch64 branch 5 times, most recently from 7e93813 to 4c2f4e4 Compare August 18, 2026 23:59

cshung force-pushed the cshung/whp-aarch64 branch 3 times, most recently from 404a655 to 189317a Compare August 19, 2026 15:45

cshung and others added 11 commits September 28, 2026 21:32

Limit ARM64 smoke to Rust guests

49762ae

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Serialize Windows ARM64 smoke tests

f6869e0

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Use native libclang in ARM64 smoke

4ad5ded

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Reduce WHP interrupt path churn

0811106

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Harden ARM64 WHP FFI layouts

70a44ae

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Remove unsupported ARM64 WHP tracing

864b87e

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Move ARM64 WHP register definitions

845aaff

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Batch ARM64 WHP register access

36bddb7

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

cshung force-pushed the cshung/whp-aarch64 branch from 9f9058d to b6ecb9d Compare September 29, 2026 05:30

cshung added 2 commits September 29, 2026 09:11

Fix ARM64 WHP exit context alignment

1a24730

Match WHV_RUN_VP_EXIT_CONTEXT's 16-byte ABI alignment. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 6f20a05d-6bee-4e2e-b320-12f8d9759bbc Signed-off-by: cshung <3410332+cshung@users.noreply.github.com>

Conversation

cshung commented Jul 12, 2026

Summary

Changes

Structural

New: whp/aarch64.rs

Integration

Cross-compilation fix

Verification

Why manual FFI bindings?

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cshung commented Aug 5, 2026

Uh oh!

syntactically left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

hyperlight-gh-bot Bot commented Sep 29, 2026

Benchmark Results

Benchmark Results

function_call_codec

payload_allocation

sandboxes

slot_pool

snapshot_files

virtq_readonly

virtq_readwrite

Benchmark Results

function_call_codec

payload_allocation

sandboxes

slot_pool

snapshot_files

virtq_readonly

virtq_readwrite

Top regressions

Benchmark Results

function_call_codec

payload_allocation

sandboxes

slot_pool

snapshot_files

virtq_readonly

virtq_readwrite

Top regressions

Benchmark Results

New: `whp/aarch64.rs`