Xen project Mailing List

On Fri, Apr 14, 2017 at 1:03 PM, Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx> wrote:

On 04/14/2017 09:08 PM, Tamas K Lengyel wrote:
>
>
> On Thu, Apr 13, 2017 at 4:20 AM, Razvan Cojocaru
> <rcojocaru@xxxxxxxxxxxxxxx <mailto:rcojocaru@bitdefender.com>> wrote:
>
> On 04/12/2017 08:11 PM, Tamas K Lengyel wrote:
> >
> >
> > On Mon, Apr 10, 2017 at 3:44 AM, Razvan Cojocaru
> > + emulate = 1;
> > + memaccess = 1;
> > + }
> > #if defined(__i386__) || defined(__x86_64__)
> > else if ( !strcmp(argv[0], "breakpoint") )
> > {
> > @@ -536,7 +551,7 @@ int main(int argc, char *argv[])
> > }
> >
> > rc = xc_set_mem_access(xch, domain_id, default_access,
> > START_PFN,
> > - (xenaccess->max_gpfn - START_PFN) );
> > + emulate ? 1000 :
> > (xenaccess->max_gpfn - START_PFN));
> >
> >
> > Why only 1000? What if the domain has less then 1000?
>
> Because it will kill the guest to emulate everything, and the emulator
> still can't handle all instructions (this is easy to see by using all
> the guest's pages and looking at the output of xl dmesg with loglvl=all
> guest_loglvl=all on the Xen command line).
>
>
> So what's the guarantee that the emulator will work if you only do it
> only up to the first 1000 pages? Seems totally arbitrary to me. If the
> emulator can't handle all instructions then you would have to check that
> the instruction for which you are returning the emulate flag is in the
> list of instruction that can be handled.. Can such a list be derived
> right now?

If an instruction can't be emulated it will be shown as such in the ring
buffer used by xl dmesg. Speaking of that, I'd like to, at some point,
send a patch that sends a vm_event saying that emulation failed to
userspace when that is the case, to give it a chance to do something
else (for example use altp2m, or lift the page restrictions).

I think that would be a much needed addition to make this system more robust.

We can also probably go through the emulator code and build an exact
list of all the officially supported instructions, but I believe that
that would have to be manual work - I am not aware of a tool to extract
them or a header file that lists them in some structure. I'd love to be
wrong about this.

As for your question, there's no guarantee that the emulator will
work,obut that's not why I chose 1000. I chose that number because the
application will get less EPT events, and the guest will not be bogged
down by handling them. But in my experiments it's also less likely to
hit unhandleable instructions in the first 1000 pages since those are
usually used by the guest kernel, drivers, and so on, and are less
likely to cause problems.

In any case, I don't mind dropping the 1000 pages limit - I can always
build a custom xen-access when I need it.

I don't mind setting it only for a 1000 in the test program, just wanted to understand rationale behind it. I think a comment in the program explaining what has been discussed here would also be helpful.

Tamas

Re: [Xen-devel] [PATCH] tests/xen-access: Added vm_event emulation tests