Message-ID: <20200227001117.GX9940@linux.intel.com>
Date:   Wed, 26 Feb 2020 16:11:17 -0800
From:   Sean Christopherson <sean.j.christopherson@...el.com>
To:     Paolo Bonzini <pbonzini@...hat.com>
Cc:     Thomas Gleixner <tglx@...utronix.de>,
        Xiaoyao Li <xiaoyao.li@...el.com>,
        Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
        Andy Lutomirski <luto@...capital.net>, x86@...nel.org,
        kvm@...r.kernel.org, linux-kernel@...r.kernel.org,
        David Laight <David.Laight@...lab.com>
Subject: Re: [PATCH v2 3/6] kvm: x86: Emulate split-lock access as a write

On Tue, Feb 11, 2020 at 02:34:18PM +0100, Paolo Bonzini wrote:
> On 11/02/20 14:22, Thomas Gleixner wrote:
> > Paolo Bonzini <pbonzini@...hat.com> writes:
> >> On 03/02/20 16:16, Xiaoyao Li wrote:
> >>> A sane guest should never tigger emulation on a split-lock access, but
> >>> it cannot prevent malicous guest from doing this. So just emulating the
> >>> access as a write if it's a split-lock access to avoid malicous guest
> >>> polluting the kernel log.
> >>
> >> Saying that anything doing a split lock access is malicious makes little
> >> sense.
> > 
> > Correct, but we also have to accept, that split lock access can be used
> > in a malicious way, aka. DoS.
> 
> Indeed, a more accurate emulation such as temporarily disabling
> split-lock detection in the emulator would allow the guest to use split
> lock access as a vehicle for DoS, but that's not what the commit message
> says.  If it were only about polluting the kernel log, there's
> printk_ratelimited for that.  (In fact, if we went for incorrect
> emulation as in this patch, a rate-limited pr_warn would be a good idea).
> 
> It is much more convincing to say that since this is pretty much a
> theoretical case, we can assume that it is only done with the purpose of
> DoS-ing the host or something like that, and therefore we kill the guest.

The problem with "kill the guest", and the reason I'd prefer to emulate the
split-lock as a write, is that killing the guest in this case is annoyingly
difficult.

Returning X86EMUL_UNHANDLEABLE / EMULATION_FAILED gets KVM to
handle_emulation_failure(), but handle_emulation_failure() will only "kill"
the guest if emulation failed in L1 CPL==0.  For all other modes, it will
inject a #UD and resume the guest.  KVM also injects a #UD for L1 CPL==0,
but that's the least annoying thing.

Adding a new emulation type isn't an option because this code can be
triggered through normal emulation.  A new return type could be added for
split-lock, but that's code I'd really not add, both from an Intel
perspective and a KVM maintenance perspective.  And, we'd still have the
conundrum of what to do if/when split-lock #AC is exposed to L1, e.g. in
that case, KVM should inject an #AC into L1, not kill the guest.  Again,
totally doable, but ugly and IMO an unnecessary maintenance burden.

I completely agree that poorly emulating the instruction from the (likely)
malicious guest is a hack, but it's a simple and easy to maintain hack.

> >> Split lock detection is essentially a debugging feature, there's a
> >> reason why the MSR is called "TEST_CTL".  So you don't want to make the
> > 
> > The fact that it ended up in MSR_TEST_CTL does not say anything. That's
> > where they it ended up to be as it was hastily cobbled together for
> > whatever reason.
> 
> Or perhaps it was there all the time in test silicon or something like
> that...  That would be a very plausible reason for all the quirks behind it.
> 
> Paolo
> 

Powered by blists - more mailing lists