On 04/25/2018 09:49 AM, Jan Beulich wrote:
>>>> On 24.04.18 at 20:51, <andrew.coop...@citrix.com> wrote:
>> --- a/xen/arch/x86/hvm/svm/svm.c
>> +++ b/xen/arch/x86/hvm/svm/svm.c
>> @@ -1883,6 +1883,22 @@ static int svm_msr_read_intercept(unsigned int msr,
>> uint64_t *msr_content)
>> switch ( msr )
>> {
>> case MSR_IA32_SYSENTER_CS:
>> + case MSR_IA32_SYSENTER_ESP:
>> + case MSR_IA32_SYSENTER_EIP:
>
> These three do not require sync-ing, as their values aren't read from the
> VMCB.
> (They do require sync-ing on the write path).
>
> I also don't think this is going to fully resolve Razvan's issue (not the
> least
> because the code paths you adjust aren't involved in his scenario): As
> pointed out in a private mail, I think vmcb_in_sync needs to start out as
> true for a vCPU, and may need setting to true upon context set and/or
> reset/init emulation.
Doing arch_svm->vmcb_in_sync = 1; in construct_vmcb() does solve the
issue. I can't unfortunately test if it also needs setting in other
places as our internal patches still need some work so introspection is
not yet fully functional on SVM (mem_access events specifically are a
bit of a problem).
Thanks,
Razvan
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xenproject.org
https://lists.xenproject.org/mailman/listinfo/xen-devel
