Re: [Qemu-devel] [PATCH 4/7 v2] KVM regsync: Add register bitmap parameter to do_kvm_cpu_synchronize_state

[Scott Wood]

On 01/24/2013 12:22:56 PM, Peter Maydell wrote:
> On 24 January 2013 18:17, Scott Wood <address@hidden> wrote:
> > On 01/24/2013 06:41:05 AM, Alexander Graf wrote:
> >> But maybe the better solution would be a special "write to clear"  
> ONE_REG
> >> register to clear specific bits and a big hammer "set" ONE_REG  
> (which we
> >> have already) for reset only.
> >>
> >> That would make things easier, right? Scott, any ideas on this?
>
> > Yes, we should have a ONE_REG write-to-clear TSR.  Maybe a  
> write-to-set
> > version as well, in case userspace wants to inject something.   
> Likewise for
> > MCSR.
>
> This would go against the principle we adopted for ARM ONE_REG,
> which is that the "registers" accessed this way never have side
> effects, so userspace could do a complete migration via
> "get complete list of regs; get each reg in list; transfer
> all values to remote end; remote end sets each reg".
>
> Obviously PPC isn't bound to abide by the same ground rules
> for the ioctl, but I figured I'd mention the principle so
> you can decide if you like it :-)

I don't like it. :-)

Such a rule would only force us to make a new mechanism for cases where  
we need side effects.  As Alex points out, the set of registers to be  
synced can be a subset of all "registers".  We do provide a  
non-side-effect version (it's all we have at the moment), and that's  
not going away.

-Scott