Porker

A practical, minimal, high-performance VMM with first-class device passthrough. Forked from Firecracker.

Porker is what Firecracker is, plus VFIO PCI passthrough that works for real GPUs.

What is Porker?

Porker is a Rust-based virtual machine monitor that uses Linux KVM to run microVMs. It is a hard fork of Amazon Web Services' Firecracker, which targets multi-tenant serverless workloads with a deliberately minimal device model. Porker keeps Firecracker's small surface area and fast startup, but adds the device-passthrough machinery needed to run GPU and other PCIe-attached workloads end-to-end.

If you want maximum density, snapshotting, balloon-style memory overcommit, and the upstream support story — use Firecracker. If you want a small VMM that can pass a GPU through to a guest, run real CUDA workloads on it, and tear down cleanly under load — that's the Porker bet.

Porker inherits Firecracker's Apache 2.0 license. The vast majority of this codebase is upstream Firecracker; the VFIO passthrough additions are the contribution. See CREDITS.md and CHANGELOG.md.

What works today

Validated on real hardware (see docs/HARDWARE.md):

• PCIe device passthrough via VFIO type1 — Realtek RTL8111 NIC and NVIDIA RTX 4080 Max-Q (AD104M)
• NVIDIA GPU compute end-to-end — driver bind, nvidia-smi, CUDA kernels, 4-stream concurrency, ~111K kernel launches/sec sustained, 8 GiB IOMMU-mapped DMA round-trip
• Clean teardown under load — SIGTERM during sustained kernel launches triggers vfio-pci FLR and a fresh VM boot re-acquires the device cleanly (15/15 cycles validated)
• All of Firecracker's existing virtio support — block, net, vsock, entropy, pmem, balloon — when no VFIO device is attached

What does NOT work

Honest about the holes. These are explicit non-goals for v0.1.0:

• Snapshot + VFIO — incompatible. Porker rejects at the API level. (GPU device state isn't snapshottable through any VFIO API.)
• Balloon + VFIO — incompatible. Porker rejects at the API level. (Balloon-reclaimed pages remain IOMMU-mapped → DMA use-after-free risk.)
• PCIe hot-plug — cold-plug only. Devices are assigned at VM creation.
• Multi-GPU per VM — single VFIO device per VM.
• ResizableBAR negotiation — disabled (pci=realloc=off on the guest cmdline). The device's default BAR sizes are used.
• MSI-X on RTX 4080 Max-Q — this specific GPU has its MSI-X capability orphaned outside the standard config-space cap chain. The nvidia driver falls back to polling and works correctly. Behavior on other GPUs is untested.
• aarch64 VFIO — untested. The upstream Firecracker aarch64 paths still exist in tree but the passthrough additions are x86_64-only for now.

See docs/THREAT_MODEL.md for the security guarantees that change versus upstream Firecracker — most importantly, that DMA isolation now depends on a correctly configured host IOMMU.

Quickstart

End-to-end walkthrough: host setup → bind device to vfio-pci → boot a passthrough VM. See docs/QUICKSTART.md.

Building from source

Porker uses Firecracker's existing build system unchanged. You need Docker running on the host (the build runs in a development container).

git clone <fork-url>.git porker
cd porker
tools/devtool build
toolchain="$(uname -m)-unknown-linux-musl"

The firecracker and jailer binaries are placed at build/cargo_target/${toolchain}/debug/. The binaries are still named firecracker and jailer — only the project and release artifacts are rebranded. See docs/UPSTREAM.md for why.

Releases

Latest release: see GitHub releases. Changes are recorded in CHANGELOG.md; Porker entries are at the top, upstream Firecracker entries are preserved below.

Project status

Alpha. The platform works on the validated hardware path; expect rough edges anywhere else. v0.1.0 is the first release intended for consumption by anyone-not-the-author.

We track upstream Firecracker for security fixes; we do not pursue upstream merge for the VFIO/passthrough work — those goals diverge enough from Firecracker's design center that upstreaming isn't realistic. See docs/UPSTREAM.md for the sync cadence and the file-level drift surface where upstream changes will require manual merging.

Contributing

Open a GitHub issue. Keep PRs small and focused. Most of the developer workflow in CONTRIBUTING.md (inherited from Firecracker) still applies.

Security disclosure

Security-sensitive issues: do not open a public issue. The current SECURITY.md is inherited from upstream Firecracker and points to AWS contacts; that policy will be updated in a subsequent release to point at the Porker maintainer. Until then, if you find something sensitive, contact the maintainer directly via GitHub.

Design

Porker's overall architecture is described in the design document, inherited from Firecracker and still accurate for everything except the VFIO subsystem. The VFIO additions live under src/vmm/src/vfio/ and src/vmm/src/devices/pci/vfio_pci.rs.