Content-Type: text/plain; charset=utf-8
Mime-Version: 1.0 (Mac OS X Mail 10.1 \(3251\))
Subject: Re: [PATCH] KVM: x86: restore IP after all far jump failures
From: Nadav Amit <nadav.amit@gmail.com>
In-Reply-To: <20161122205600.GC12634@potion>
Date: Tue, 22 Nov 2016 15:18:28 -0800
Cc: Paolo Bonzini <pbonzini@redhat.com>,
        LKML <linux-kernel@vger.kernel.org>, KVM list <kvm@vger.kernel.org>,
        stable@vger.kernel.org, Dmitry Vyukov <dvyukov@google.com>,
        Steve Rutherford <srutherford@google.com>,
        Andrew Honig <ahonig@google.com>, Prasad Pandit <ppandit@redhat.com>
Message-Id: <217EBCAE-2976-41C1-9F5C-4B54C1FB2D6E@gmail.com>
References: <20161122192104.19836-1-rkrcmar@redhat.com>
 <b6cfa04f-7188-da7e-8bd5-4c4171093c96@redhat.com>
 <F346D675-FC84-4595-BC1D-C1049F7D6B7E@gmail.com>
 <20161122205600.GC12634@potion>
To: =?utf-8?B?UmFkaW0gS3LEjW3DocWZ?= <rkrcmar@redhat.com>
Sender: linux-kernel-owner@vger.kernel.org
Content-Transfer-Encoding: 8bit
Content-Length: 1565
Lines: 38


> On Nov 22, 2016, at 12:56 PM, Radim Krčmář <rkrcmar@redhat.com> wrote:
> 
> 2016-11-22 11:43-0800, Nadav Amit:
>> I admit my wrongdoings, but I still think the fix should have been to
>> remove the entire recovery logic and just return X86EMUL_UNHANDLEABLE if
>> something goes wrong (exception). This will kill the misbehaving process
>> but keep the VM running.
> 
> X86EMUL_UNHANDLEABLE will kill the whole VM (on QEMU, other userspaces
> might handle the instruction and resume KVM).

I don’t think so. If CPL is not 0, handle_emulation_failure() will be called
and will inject #UD.

> 
> The recovery path is in the spec, which means that nothing goes wrong.
> I think we implement the spec quite well now, so keeping the #GP and CS
> recovery is slightly better, although not safer.
> 
>> Otherwise, a malicious VM process, which can somehow control descriptors
>> (LDT?) may modify the descriptor during the emulation and get the system
>> to inconsistent state and prevent the VM-entry.
> 
> We restore the original CS -- malicious guest would get killed on a
> failed VM entry anyway, so the difference is only in KVM internal error
> code (assuming there are no other bugs).
> 
> Am I misunderstanding something?

Most likely you are right, but after doing one mistake, I don’t want
to be accountable for another.

Note there is another problematic case in em_ret_far(). If an exception
occurs when RIP is assigned, the RSP updates (of emulate_pop() ) are not 
going to be reverted. Can it be used for anything malicious? I don’t know.

Nadav