Re: [PATCH] target-i386: clear guest TSC on reset

[Paolo Bonzini <pbonzini@redhat.com>]

Il 05/12/2013 17:12, Marcelo Tosatti ha scritto:
>> > - call kvm_set_ticks() from cpu_set_ticks() and cpu_enable_ticks()
> env->tsc is just a placeholder for the vcpu TSC.
> 
> A vcpus TSC from QEMU's point of view is a register initialized to zero,
> which requires read/write from KVM, and migration.

QEMU already tracks the TSC in cpu_get_ticks().  So far this is used
only for TCG, but for example the code is there that preserves the TSC
when you stop/resume the VM and when you migrate the VM.  Reset is not
yet there, which is a bug similar to the one Fernando is trying to solve
for KVM.

So, from QEMU's point of view the TSC should be a global value across
the whole system (timer_state.cpu_ticks_offset) + a per-VCPU TSC offset
(env->tsc_adjust).  When talking to KVM, the per-VCPU TSC offset in turn
has two parts, both set with KVM_SET_MSRS: one is computed from
MSR_IA32_TSC, the other comes from MSR_IA32_TSC_ADJUST.

The point here would be to treat it as such.

With this change, env->tsc need not be migrated.  The global value
timer_state.cpu_ticks_offset is migrated already.  The host-side TSC
adjust can be computed from rdtsc()-timer_state.cpu_ticks_offset on the
destination machine and/or at reset time.  The guest-side TSC adjust is
env->tsc_adjust as it is now.

Paolo