Re: xl/xm save -c fails - set_vcpucontext EOPNOTSUPP (was Re: xl save -c issues with Windows 7 Ultimate)

["Jan Beulich" <JBeulich@novell.com>]

>>> On 10.05.11 at 17:50, Shriram Rajagopalan <rshriram@cs.ubc.ca> wrote:
> On Tue, May 10, 2011 at 10:02 AM, Jan Beulich <JBeulich@novell.com> wrote:
> 
>> >>> On 10.05.11 at 16:52, Shriram Rajagopalan <rshriram@cs.ubc.ca> wrote:
>> > On Tue, May 10, 2011 at 3:41 AM, Ian Campbell <Ian.Campbell@citrix.com 
>> >wrote:
>> >> The most plausible looking EOPNOTSUPP from that code is in
>> >> xen/arch/x86/domain.c:arch_set_info_guest() but that is on a PV only
>> >> path.
>> >>
>> >> And that rings with the pv guests I am using. It makes perfect sense,
>> > looking
>> > at that function and especially at the code that returns EOPNOTSUPP (the
>> > only
>> > place in the entire file).
>> >    else
>> >     {
>> >         bool_t fail = v->arch.pv_vcpu.ctrlreg[3] != c(ctrlreg[3]);
>> >
>> > #ifdef CONFIG_X86_64
>> >         fail |= v->arch.pv_vcpu.ctrlreg[1] != c(ctrlreg[1]);
>> > #endif
>> >
>> >         for ( i = 0; i < ARRAY_SIZE(v->arch.pv_vcpu.gdt_frames); ++i )
>> >             fail |= v->arch.pv_vcpu.gdt_frames[i] != c(gdt_frames[i]);
>> >         fail |= v->arch.pv_vcpu.gdt_ents != c(gdt_ents);
>> >
>> >         fail |= v->arch.pv_vcpu.ldt_base != c(ldt_base);
>> >         fail |= v->arch.pv_vcpu.ldt_ents != c(ldt_ents);
>> >
>> >         if ( fail )
>> >            return -EOPNOTSUPP;
>> >     }
>> >
>> > This change was introduced by c/s
>> > changeset:   23142:f5e8d152a565
>> > user:        Jan Beulich <jbeulich@novell.com>
>> > date:        Tue Apr 05 13:01:25 2011 +0100
>> > x86: split struct vcpu
>> >
>> > I think I am missing something really obvious in this piece of code. The
>> > xc_domain_resume code tries to modify the return value of shutdown
>> hypercall
>> > (i.e eax register is set to 1) and this code doesnt seem to check those
>> > registers.
>>
>> Correct - the code here checks only for values where the logic
>> needed to support changing the on an already initialized vCPU isn't
>> implemented. Previously, actual vCPU state and what was tracked
>> in struct vcpu could get out of sync in this case, potentially
>> confusing things further down (including possible security issues).
>>
>> You'll want to figure out which part(s) actually differ, and why.
>> Only then we'll be able to tell whether mentioned c/s introduced
>> false positives.
>>
>> Bit confused. If I understand correctly, this piece of code checks new
> values
> of certain registers against old ones, for an already initialized VCPU. And
> AFAIT,
> it is checking the gdts, ldts & control registers. The xc_domain_resume code
> just
> changes one general purpose register eax. basically,
> 
>  get_vcpucontext()
>  set_field(eax, 1) //to indicate SUSPEND_CANCEL
>  set_vcpucontext()
> 
> I dont understand what you mean by "which parts actually differ & why".

Quite obviously there are differences in one or more of the now
checked fields, and we need to find out where they are (and
why). This is regardless of the tools apparently only modifying
eax.

> And just a trivial question:
>  is the hypervisor binary always compiled to a 32-bit elf? somehow, the
> symbols file xen-syms-* is getting compiled to 64 bit ELF binary while
> the xen binary is getting compiled to 32-bit binary.

Yes, that's because it wants to boot from 32-bit GrUB (and the
multiboot protocol also is 32-bit only afaik).

Jan