Re: [PATCH v3 3/5] x86/paravirt: Add paravirt_{read,write}_msr

[Paolo Bonzini <pbonzini@redhat.com>]

On 14/03/2016 18:02, Andy Lutomirski wrote:
> On Mon, Mar 14, 2016 at 9:58 AM, Linus Torvalds
> <torvalds@linux-foundation.org> wrote:
>>
>> On Mar 14, 2016 9:53 AM, "Andy Lutomirski" <luto@amacapital.net> wrote:
>>>
>>> Can you clarify?  KVM uses the native version, and the native version
>>> only oopses with this series applied if panic_on_oops is set.
>>
>> Can we please remove that idiocy?
>>
>> There is no reason to panic whatsoever. Seriously. What's the upside of that
>> logic?
> 
> I imagine that people who set panic_on_oops want their systems to stop
> running user code if something happens that could corrupt the state or
> if there's any sign that user code is trying some non-deterministic
> exploit.  So I'm guessing that they'd want this type of "the kernel
> screwed up -- abort" to actually result in a panic.
> 
> As a concrete, although somewhat silly, example, suppose that a write
> to MSR_SYSENTER_STACK fails.  If that happened, then user code could
> subsequently try to take over the kernel by evil manipulation of TF
> and/or perf.
> 
> I'd be okay with removing this too, though, since arranging for MSR
> access to fail seems unlikely as an exploit vector.
> 
> Borislav: SUSE actually uses panic_on_oops, right?  What's their goal?

RHEL also does, and it's mostly to trap kernel page faults before they
do more damage such as filesystem corruption.  The debug kernel has
panic_on_oops=0, while the production kernel has panic_on_oops=1.

Paolo