Re: [PATCH 2/2] KVM, Fix QEMU-KVM is killed by guest SRAO MCE

[Huang Ying <ying.huang@intel.com>]

On Wed, 2010-04-28 at 17:47 +0800, Avi Kivity wrote:
> On 04/28/2010 05:56 AM, Huang Ying wrote:
> >    
> >>>
> >>> Just want to use the side effect of copy_from_user, SIGBUS will be sent
> >>> to current process because the page touched is marked as poisoned. That
> >>> is, failure is expected, so the return value is not checked.
> >>>
> >>>        
> >> What if the failure doesn't happen?  Say, someone mmap()ed over the page.
> >>      
> > Sorry, not get your idea clearly. hva is re-mmap()ed? We just read the
> > hva, not write, so I think it should be OK here.
> >
> >    
> 
> We don't generate a signal in this case.  Does the code continue to work 
> correctly (not sure what correctly is in this case... should probably 
> just continue).
> 
> There's also the possibility of -EFAULT.

I think signal should be generated for copy_from_user, because the hva
is poisoned now. The signal will not generated only if the hva is
re-mmap()ped to some other physical page, but this should be impossible
unless we have memory hotadd/hotremove in KVM.

If the signal is not generated, lost or overwritten, guest will
continue, and if the hva is still poisoned, the page fault will be
triggered again; if the hva is not poisoned, there will be no further
page fault.

Best Regards,
Huang Ying