Hacker News Reader

VMs also don't always require hardware virtualization - Alibaba's PVM https://lkml.org/lkml/2024/2/26/1263 didn't get upstreamed, but, theoretically the MMU is all you need for complete isolation. This kind of idea is also how VM software worked before VT-x was introduced. And of course QEMU has the TCG which works with no kernel support at all.

View on HN

I think you could also add Xen to that list. IIRC, the old Xen PV mode was purely paravirtualized without using any hardware extensions.

Yes, Xen was big on paravirtualisation but started supporting the other kinds pretty soon, too. (At least they were supported around 2009-2012, when I was working on XenServer.)

I think things are swinging back the other way if I have understood the more recent PVHv2 stuff correctly.

In my experience TCG (or any method that doesn't require root / admin power) is pretty slow. But I'd be happy to be wrong about that, for an odd project I have

It depends a bit on your workload. If you have a pure computation workload, without much IO, TCG etc doesn't need to be slow.

It also depends on the architectures. x86 on ARM is tough to do efficiently because of the memory model differences. One of the keys to Rosetta 2 being so good was being able to make the underlying ARM processor obey the x86 memory model (even though it was still executing ARM instructions).

Reply on news.ycombinator.com