Re: [PATCH 7/7] x86/pv: Rewrite %dr6 handling

[Andrew Cooper <andrew.cooper3@xxxxxxxxxx>]

On 16/09/2023 2:10 pm, Jinoh Kang wrote:
> On 9/16/23 21:56, Andrew Cooper wrote:
>>>> -    /* Save debug status register where guest OS can peek at it */
>>>> -    v->arch.dr6 |= (dr6 & ~X86_DR6_DEFAULT);
>>>> -    v->arch.dr6 &= (dr6 | ~X86_DR6_DEFAULT);
>>>> -
>>>>      if ( guest_kernel_mode(v, regs) && v->domain->debugger_attached )
>>>>      {
>>>> +        /* Save debug status register where gdbsx can peek at it */
>>>> +        v->arch.dr6 = x86_merge_dr6(v->domain->arch.cpu_policy,
>>>> +                                    v->arch.dr6, pending_dbg);
>>>>          domain_pause_for_debugger();
>>>>          return;
>>>>      }
>>>>  
>>>> -    pv_inject_hw_exception(X86_EXC_DB, X86_EVENT_NO_EC);
>>>> +    pv_inject_DB(pending_dbg);
>> ... this, which will merge all new pending bits into the guest's DR6.
>>
>> If the guest chooses not to clean out DR6 each time, then it will see
>> content accumulate.
>>
>> To your earlier point of shadowing, we already have to do that.  DR6 is
>> just one of many registers we need to context switch for the vCPU.
>>
>> PV guests, being ring-deprivieged, trap into Xen for every DR access,
>> and Xen handles every one of their #DB exceptions.  This is one reason
>> why I split the work into several parts.  PV guests are easier to get
>> sorted properly, and patch 1,2,5,6 are all common improvements relevant
>> to HVM too.
> That confirms my knowledge.  Honestly if I got such a foolish remark
> I would question the reviewer's understanding of PV mode (not that you did
> anything wrong).  Sorry for wasting your time.

Not at all.

This is exceptionally tricky, and as you can see from the v1 series,
don't assume that I know it all perfectly.

The specifics of PV guests are unique to Xen, and there are 0 people who
understand it fully.  For example, doing the sanity testing for this v2
series, I rediscovered the completely undocumented properly that a PV
guests IOPL is separate from an architectural ideal of IOPL, defaults to
0 which has the meaning "guest kernel takes a #GP for access to IO ports".

Please do continue to ask questions like this if you're not sure.  It is
not a waste of time cross-checking that the logic is correct.

~Andrew