Port to ARM64?

[zoeleu]

Many cloud platforms now run on ARM architectures. Is it possible to port this to ARM64?

[bysui]

Many cloud platforms now run on ARM architectures. Is it possible to port this to ARM64?

Yes, it's possible to port the PVM idea to ARM64. In fact, it is on our long‑term roadmap. However, compared with x86, implementing PVM on ARM64 requires more work because some key mechanisms used by PVM are not currently available on ARM. Perhaps with AI‑assisted coding, implementing PVM on ARM64 will become much easier. :)

These are some of the key mechanisms I can recall from memory, as I have been away from the ARM architecture for quite a while, so some details may not be entirely accurate.

Function Module	Purpose	On x86	On ARM64	How on ARM64
Privilege isolation	Trap privileged instructions	Guest kernel runs in CPL3	Guest kernel runs in EL0	Similar, but requires more work, e.g., use ttbr1 for guest kernel?
Instruction emulator	Emulate privileged instructions	Reuse the KVM x86 emulator	None	Port the QEMU instruction emulator or replace privileged instructions with PVOPS
PVOPS	Replace frequently executed privileged instructions to improve performance	Reuse Xen PVOPS	None	Introduce a PVOPS framework in the guest kernel via instruction patching
Switcher	Guest/host world switching	Switcher integrated into the host entry path	Switcher integrated into the host vector table	Similar design
MMU virtualization	Provide memory isolation and virtual memory translation for the guest	Reuse shadow page tables	None	ARM64 KVM mainly relies on Stage‑2 translation; shadow paging is only used for nested virtualization, so reusing it would require some redesign. Alternatively, introduce a new PV ABI for MMU virtualization
Address space isolation	Separate guest and host kernels for the integrated switcher	PIE kernel (new)	Reuse	The ARM64 kernel is already PIE
Interrupt controller	Manage hardware interrupts for the guest	Reuse in‑kernel APIC emulation	None	Port QEMU GIC emulation
Timer	Provide virtualized timing facilities to the guest	Reuse in‑kernel APIC timer emulation	None	Trap and emulate using hrtimer
I/O virtualization	Allow guest access to devices through controlled interfaces	MMIO + port I/O emulation	MMIO emulation	Similar
KVM compatibility	Maintain compatibility with the existing KVM ecosystem	Implemented as a KVM vendor (kvm‑pvm)	Implemented as a KVM mode	Requires ARM64 KVM mode hooks

In short, KVM on x86 has a long history, so there are various software emulation components available. In contrast, KVM on ARM64 was designed primarily around hardware-assisted virtualization, so some of these software mechanisms would need to be implemented if we want to support PVM on ARM64.