Re: [Xen-devel] [PATCH -tip 0/2] x86: Prohibit kprobes on XEN_EMULATE_PREFIX

From: Andrew Cooper
Date: Thu Sep 05 2019 - 09:32:07 EST

On 05/09/2019 14:09, Masami Hiramatsu wrote:
> On Thu, 5 Sep 2019 20:32:24 +0900
> Masami Hiramatsu <mhiramat@xxxxxxxxxx> wrote:
>> On Thu, 5 Sep 2019 08:54:17 +0100
>> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> wrote:
>>> On 05/09/2019 02:49, Masami Hiramatsu wrote:
>>>> On Wed, 4 Sep 2019 12:54:55 +0100
>>>> Andrew Cooper <andrew.cooper3@xxxxxxxxxx> wrote:
>>>>> On 04/09/2019 12:45, Masami Hiramatsu wrote:
>>>>>> Hi,
>>>>>> These patches allow x86 instruction decoder to decode
>>>>>> xen-cpuid which has XEN_EMULATE_PREFIX, and prohibit
>>>>>> kprobes to probe on it.
>>>>>> Josh reported that the objtool can not decode such special
>>>>>> prefixed instructions, and I found that we also have to
>>>>>> prohibit kprobes to probe on such instruction.
>>>>>> This series can be applied on -tip master branch which
>>>>>> has merged Josh's objtool/perf sharing common x86 insn
>>>>>> decoder series.
>>>>> The paravirtualised xen-cpuid is were you'll see it most in a regular
>>>>> kernel, but be aware that it is also used for testing purposes in other
>>>>> circumstances, and there is an equivalent KVM prefix which is used for
>>>>> KVM testing.
>>>> Good catch! I didn't notice that. Is that really same sequance or KVM uses
>>>> another sequence of instructions for KVM prefix?
>>> I don't know if you've spotted, but the prefix is a ud2a instruction
>>> followed by 'xen' in ascii.
>>> The KVM version was added in c/s 6c86eedc206dd1f9d37a2796faa8e6f2278215d2
> Hmm, I think I might misunderstand what the "emulate prefix"... that is not
> a prefix which replace actual prefix, but just works like an escape sequence.
> Thus the next instruction can have any x86 prefix, correct?

There is a bit of history here :)

Originally, 13 years ago, Xen invented the "Force Emulate Prefix", which
was the sequence:

ud2a; .ascii 'xen'; cpuid

which hit the #UD handler and was recognised as a request for
virtualised CPUID information. This was for ring-deprivileged
virtualisation, and is needed because the CPUID instruction itself
doesn't trap to the hypervisor.

Following some security issues in our instruction emulator, I reused
this prefix with VT-x/SVM guests for testing purposes. It behaves in a
similar manner - when enabled, it is recognised in #UD exception
intercept, and causes Xen to add 5 to the instruction pointer, then
emulate the instruction starting there.

Then various folk thought that having the same kind of ability to test
KVM's instruction emulator would be a good idea, so they borrowed the idea.

>From a behaviour point of view, it is an opaque 5 bytes which means
"break into the hypervisor, then emulate the following instruction".

The name "prefix" is unfortunate. It was named thusly because from the
programmers point of view, it was something you put before the CPUID
instruction which wanted to be emulated. It is not related to x86
instruction concept of a prefix.