Re: [PATCH] KVM: X86: Fix the decoding of segment overrides in 64bit mode

From: Andrew Cooper
Date: Fri Mar 23 2018 - 10:43:51 EST


On 23/03/18 14:27, Wanpeng Li wrote:
> 2018-03-22 21:53 GMT+08:00 Andrew Cooper <andrew.cooper3@xxxxxxxxxx>:
>> On 22/03/18 13:39, Wanpeng Li wrote:
>>> 2018-03-22 20:38 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>>>> On 22/03/2018 12:04, Andrew Cooper wrote:
>>>>> We've got a Force Emulation Prefix (ud2a; .ascii "xen") for doing
>>>>> magic. Originally, this was used for PV guests to explicitly request an
>>>>> emulated CPUID, but I extended it to HVM guests for "emulate the next
>>>>> instruction", after we had some guest user => guest kernel privilege
>>>>> escalations because of incorrect emulation.
>>>> Wanpeng, why don't you add it behind a new kvm module parameter? :)
>>> Great point! I will have a try. Thanks Paolo and Andrew. :)
>> Using the force emulation prefix requires intercepting #UD, which is in
>> general a BadThing(tm) for security. Therefore, we have a build time
> Yeah, however kvm intercepts and emulates #UD by default, should we
> add a new kvm module parameter to enable it and disable by default?
> Paolo.
>
>> configuration option to compile in support, and require that test
>> systems explicitly opt into using it via a command line parameter.
>>
>> http://xenbits.xen.org/gitweb/?p=xen.git;a=blob;f=xen/arch/x86/hvm/hvm.c;h=db52312882205e65b32e587106ca795f4bfab2eb;hb=refs/heads/staging#l3741
>> is the general #UD intercept handler if you want a reference. (You can
> Thanks Andrew, it is useful. :) In addition, I didn't see the
> test-memop-seg testcase has "Forced Emulation Prefix", when the prefix
> is added to each instruction in the testcase?

It has ended up substantially more ugly than I first intended, due to
several assembler bugs in older GCC and Clang toolchains.

http://xenbits.xen.org/gitweb/?p=xtf.git;a=blob;f=tests/memop-seg/asm.S;h=698661425bcdc9c181b235e323c2460e06c6e986;hb=HEAD#l35

I previously has FEP passed as a second parameter, but that becomes
prohibitively complicated to extract when testing %ss or %esp. FEP is
now encoded in the bottom bit of the address passed in.

This was the cleanest way I could find of testing every combination, but
I'm open to improvements if anyone can spot any.

~Andrew