Re: [PATCH 1/3] Implement emulator_write_phys()

[Anthony Liguori <aliguori@us.ibm.com>]

On Mon, 2007-08-27 at 20:26 +0300, Avi Kivity wrote:
> Anthony Liguori wrote:
> > On Mon, 2007-08-27 at 18:45 +0300, Avi Kivity wrote:
> >   
> >> Anthony Liguori wrote:
> >>     
> >>> Since a hypercall may span two pages and is a gva, we need a function to write
> >>> to a gva that may span multiple pages.  emulator_write_phys() seems like the
> >>> logical choice for this.
> >>>
> >>> @@ -962,8 +962,35 @@ static int emulator_write_std(unsigned long addr,
> >>>  			      unsigned int bytes,
> >>>  			      struct kvm_vcpu *vcpu
> >>>       
> >> I think that emulator_write_emulated(), except for being awkwardly 
> >> named, should do the job.  We have enough APIs.
> >>
> >> But!  We may not overwrite the hypercall instruction while a vcpu may be 
> >> executing, since there's no atomicity guarantee for code fetch.  We have 
> >> to to be out of guest mode while writing that insn.
> >>     
> >
> >
> > Hrm, good catch.
> >
> > How can we get out of guest mode given SMP guest support?
> >
> >   
> 
> kvm_flush_remote_tlbs() is something that can be generalized.  
> Basically, you set a bit in each vcpu and send an IPI to take them out.
> 
> But that's deadlock prone and complex.  Maybe you can just take 
> kvm->lock, zap the mmu and the flush tlbs, and patch the instruction at 
> your leisure, as no vcpu will be able to map memory until the lock is 
> released.

This works for shadow paging but not necessarily with NPT.  Do code
fetches really not respect atomic writes?  We could switch to a 32-bit
atomic operation and that should result in no worse than the code being
patched twice.

Regards,

Anthony Liguori