On 12/06/2013 01:38 AM, Paolo Bonzini wrote:
Il 05/12/2013 17:17, Marcelo Tosatti ha scritto:
I agree it is a bit ugly, but in my testing QEMU seemed to loop over all
the VCPUS fast enough for the kernel side kvm_write_tsc() to do a
reasonable job of matching the offsets (the Linux guest did not mark
the TSC unstable due to the TSCs being unsynchronized). Am I missing
something?
Right, modern kernels (see kvm_write_tsc) perform synchronization, so in
theory the "KVM is yet unable to synchronize ..." code is not necessary
anymore.
I vote for dropping the thing entirely.
When I was writing the original patch I was tempted to do that,
but I feared that it could break older kernels that do not have
TSC synchronization code. Should we care about such uses
(recent QEMU user space + old kernel)?
I also wanted to make sure that the initialization that we do
in kvm_arch_vcpu_postcreate on power up and the subsequent
TSC writeback work well together, but I didn't have time to
test it (reading the code, I would say that the TSC generation
counter may end up being increased a few times but the TSCs
would eventually converge).
If it can be dropped entirely, I certainly have no problem with starting
with a simple patch first.
Could we start with the patch that I already sent? It's been
tested, it is conservative in the sense that it does the minimum
necessary to fix an existing bug, and should be easy to
backport. I will be replying to this email with an updated
version that has a more appropriate and less scary patch
description.
I will also be sending a patch that makes the TSC writeback
unconditional, but this one should probably be kept on hold
until it is properly tested.
As a follow-up effort we can work on Paolo's suggestions.
Is this an acceptable way forward?
Thanks,
Fernando
