pvm

PVM (pagetable-based virtualization, from virt-pvm/linux) provides a KVM-compatible /dev/kvm on hosts without hardware virtualization: the kvm_pvm module runs guests as host-user-mode processes behind shadow paging. Stock Cloud Hypervisor binaries work unchanged; the guest kernel does not — it must be a PVM-enlightened, position-independent build.

This repository builds and releases both halves from one pinned tree:

• Host kernel — Ubuntu-installable debs with kvm_pvm
• Guest kernel — module-free direct-boot vmlinux-pvm-amd64
• Upstream tracking — how the pin is maintained
• Why no Windows guests — the boundary, and why it stays there

Validated baseline

Smoke-tested on a GCP e2 instance (no nested virtualization): stock Cloud Hypervisor v53.0 static binary, direct kernel boot of a PVM guest vmlinux, EROFS/ext4 rootfs, kernel to userspace in about 1.6 seconds. The serial log carries the PVM marker:

Kernel/User page tables isolation: force enabled on kvm pvm guest

Hard limits

These follow from PVM’s design, not from this packaging:

• x86_64 only. No arm64.
• Direct kernel boot only. The UEFI firmware path (CLOUDHV.fd, rust-hypervisor-firmware) cannot run under PVM.
• No Windows guests. The guest kernel runs at hardware CPL3 and must be recompiled against PVM’s paravirtual ABI; the instructions Windows relies on most are the ones PVM’s design removes. Windows workloads stay on nodes with hardware virtualization — see why.
• Feature matrix untested. Cloud Hypervisor’s advanced features (snapshot/restore, virtio-fs, memory hotplug, ballooning) need per-item validation under PVM before production use.

This site is open source. Improve this page.