Message-ID: <539F243D.6090803@gmail.com>
Date: Mon, 16 Jun 2014 20:07:09 +0300
From: Nadav Amit <nadav.amit@gmail.com>
User-Agent: Mozilla/5.0 (Macintosh; Intel Mac OS X 10.9; rv:24.0) Gecko/20100101 Thunderbird/24.6.0
MIME-Version: 1.0
To: Paolo Bonzini <pbonzini@redhat.com>, Nadav Amit <namit@cs.technion.ac.il>
CC: gleb@kernel.org, tglx@linutronix.de, mingo@redhat.com, hpa@zytor.com,
        x86@kernel.org, kvm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH 6/6] KVM: x86: check DR6/7 high-bits are clear only on
 long-mode
References: <1402837982-24959-1-git-send-email-namit@cs.technion.ac.il> <1402837982-24959-7-git-send-email-namit@cs.technion.ac.il> <539EC43F.607@redhat.com> <539EC7F7.2000208@gmail.com> <539ED084.2000906@redhat.com> <539EDABF.5060701@gmail.com> <539F059F.8050501@redhat.com>
In-Reply-To: <539F059F.8050501@redhat.com>
Content-Type: text/plain; charset=ISO-8859-15; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org

On 6/16/14, 5:56 PM, Paolo Bonzini wrote:
> Il 16/06/2014 13:53, Nadav Amit ha scritto:
>> On 6/16/14, 2:09 PM, Paolo Bonzini wrote:
>>> Il 16/06/2014 12:33, Nadav Amit ha scritto:
>>>>>
>>>>> Do you get this if the input register has bit 31 set?
>>>> No. To be frank, the scenario may be considered a bit synthetic: the
>>>> guest assigns a value to a general-purpose register in 64-bit mode,
>>>> setting the high 32-bits to some non-zero value. Then, later, in 32-bit
>>>> mode, the guest performs MOV DR instruction. In between the two
>>>> assignments, the general purpose register is unmodified, so the high
>>>> 32-bits of the general purpose registers are still set.
>>>>
>>>> Note that this scenario does not occur when MOV DR is emulated, but
>>>> when
>>>> handle_dr() is called. In this case, the entire 64-bits of the general
>>>> purpose register used for MOV DR are read, regardless to the execution
>>>> mode of the guest.
>>>
>>> I wonder if the same bug happens elsewhere.  For example,
>>> kvm_emulate_hypercall doesn't look at CS.L/CS.DB, which is really a
>>> corner case but arguably also a bug.  kvm_hv_hypercall instead does it
>>> right.
>>>
>>> Perhaps we need a variant of kvm_register_read that (on 64-bit hosts)
>>> checks EFER/CS.L/CS.DB and masks the returned value accordingly.  You
>>> could call it kvm_register_readl.
>>
>> There are two questions that come in mind:
>> 1. Should we ignore CS.DB? It would make it consistent with
>> kvm_hv_hypercall and handle_dr. I think this is the proper behavior.
>
> It depends on what you're using it for, but as a start yes.
>
>> 2. Reading CS.L once and masking all the registers (i.e., changing the
>> is_long_mode in kvm_emulate_hypercall to is_64_bit_mode) is likely to be
>> more efficient.
>
> Yes, for the case of kvm_emulate_hypercall.  Then you can build
> kvm_register_readl on top of is_64bit_mode and fix this bug with that
> function.  Did you check that handle_cr is unaffected?

handle_cr is affected, and there are some other instances of affected 
register reads. Actually, most of the vmx instruction exit handling 
routines ignore long-mode and cs.l when considering the register operand 
size.

I'll make the necessary changes and resubmit.

Nadav

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/