RFC - TSC virtualization for KVM
A message that is not in the archive (sent off-list or not yet received): [email protected]
Hi Zachary,
Quoted 6 lines · Zachary Amsden
The overall goal is to have a stable platform TSC and a mixed set of virtualization approaches that can be tailored for the need of the particular operating system running, which runs on systems regardless of hardware TSC frequency changes, CPU hotplug, or destabilization events, and which allows virtualization of multiple different TSC speeds. I realize this is ambitious, but it is achievable, and I believe these patches provide a solid foundation.
I have some high-level comments on your patchset.
First, I like the idea of introducing a new ioctl to set the guest tsc rate. But I think it could be a little bit simpler. How about a single __u32 parameter instead of a struct. This parameter is the tsc rate for the guest in khz. If this value is configured to be 0 this means the cpu tsc frequency should be used.
Second, the patchset implements very sophisticated code to keep the tsc of different cpus synchronized. This code looks a bit fragile to me, it does not protect against SMI or NMI (and there is no way to do so). Have you thought about using the kernel system time as a reference and calculate a tsc value for the vcpu? The kernel timekeeping solves most of the problems you have here already.
Regards,
Joerg