Re: [PATCH RFC] pvh: clearly specify used parameters in vcpu_guest_context

[George Dunlap <george.dunlap@eu.citrix.com>]

On 15/11/13 23:56, Mukesh Rathor wrote:
> On Fri, 15 Nov 2013 13:56:54 -0800
> Mukesh Rathor <mukesh.rathor@oracle.com> wrote:
>
>> On Fri, 15 Nov 2013 16:45:47 +0000
>> "Jan Beulich" <JBeulich@suse.com> wrote:
>>
>>>>>> On 15.11.13 at 17:40, George Dunlap
>>>>>> <george.dunlap@eu.citrix.com> wrote:
>>>> On 15/11/13 16:32, Jan Beulich wrote:
>>>>>>>> On 15.11.13 at 16:50, Roger Pau Monne <roger.pau@citrix.com>
>>>>>>>> wrote:
>>>>>> --- a/xen/arch/x86/domain.c
>>>>>> +++ b/xen/arch/x86/domain.c
>>>>>> @@ -704,9 +704,11 @@ int arch_set_info_guest(
>>>>>>            /* PVH 32bitfixme */
>>>>>>            ASSERT(!compat);
>>>>>>    
>>>>>> -        if ( c(ctrlreg[1]) || c(ldt_base) || c(ldt_ents) ||
>>>>>> +        if ( c(ctrlreg[0]) || c(ctrlreg[1]) || c(ctrlreg[2]) ||
>>>>>> +             c(ctrlreg[4]) || c(ldt_base) || c(ldt_ents) ||
>>>>> I think it should actually be a bug for the guest to request an
>>>>> all blank CR0 or CR4. Minimally CR0.PE, CR0.PG, and CR4.PAE
>>>>> would seem to be a valid requirement to be set.
>>>>>
>>>>> Apart from that ctrlreg[] is an 8-element array... And I don't
>>>>> see debugreg[] being verified at all.
>>>>>
>>>>>>                 c(user_regs.cs) || c(user_regs.ss) ||
>>>>>> c(user_regs.es) || c(user_regs.ds) || c(user_regs.fs) ||
>>>>>> c(user_regs.gs) ||
>>>>>> +             c(kernel_ss) || c(kernel_sp) ||
>>>>>> c.nat->gs_base_kernel ||
>>>>> So George and/or Mukesh found it necessary to set
>>>>> gs_base_kernel, and you rip it out? I'm curious as to what
>>>>> they're going to say...
>>>> I didn't find it necessary; I was mostly focused on merging the
>>>> PVH and HVM codepaths without causing any regressions.  It's not
>>>> obvious to me what's special about gs_base_kernel, and I haven't
>>>> yet gone back to try to find out why Mukesh did it that way.
>> Hi,
>>
>> We had talked about this while ago, but upon boot, the first thing
>> a vcpu needs is access to kernel data structure. (A secondary vcpu
>> is bootstrapped way up into the kernel). It would be possible to
>> get rid of gs_base_kernel, but will take some work on the linux side.
>> I can try and test it out, and let you guys know.
> Ok, looking at this more, I can hack cpu_bringup_and_idle() in
> linux to include a static variable for cpuid, which is the least a
> vcpu needs to know first thing. But, I think that would not work when
> vcpu hotplug support is added. Another option would be to pass cpuid
> in one of the registers, say rdi. Thus, rdi == cpuid will be passed
> to VCPUOP_initialise. In bringup function, the booting vcpu can then
> load it's own gs based on the cpuid. If linux folks, konrad (CCd), is OK
> with this, we can remove gs_base_kernel.  Otherwise, it's such a small
> thing, hopefually it can stay.

So Linux has the exact same problem on native, and (AFAICT) they solve 
it by simply writing gs to a global variable called initial_gs.  Is 
there any reason why we can't just do what native Linux does here?  We 
should, in fact, be able to use the exact same variable.

  -George