Re: [Qemu-devel] [PATCH 4/4] Warning messages on net devices hotplug

[Corey Bryant <coreyb@linux.vnet.ibm.com>]

On 10/24/2012 11:45 AM, Paolo Bonzini wrote:
> Il 24/10/2012 17:39, Corey Bryant ha scritto:
>>
>>
>> On 10/24/2012 11:21 AM, Paolo Bonzini wrote:
>>> Il 24/10/2012 16:18, Corey Bryant ha scritto:
>>>>
>>>>
>>>> On 10/18/2012 11:15 AM, Paolo Bonzini wrote:
>>>>> Il 17/10/2012 15:15, Eduardo Otubo ha scritto:
>>>>>> With the inclusion of the new "double whitelist" seccomp filter, Qemu
>>>>>> won't be able to execve() in runtime, thus, no hotplug net devices
>>>>>> allowed.
>>>>>>
>>>>>> Signed-off-by: Eduardo Otubo <otubo@linux.vnet.ibm.com>
>>>>>
>>>>> Please check this in net_init_tap instead.  When using libvirt, hotplug
>>>>> is done with a completely different mechanism that involves
>>>>> file-descriptor passing and does not require executing a helper.
>>>>>
>>>>> Paolo
>>>>>
>>>>
>>>> Are you sure net_init_tap() is the right place for this check?
>>>
>>> Yes, assuming there is a global that says whether the seccomp sandbox is
>>> in effect.  Even something like "if (sandbox_active && !tap->has_fd)
>>> error(...)" can be enough.
>>>
>>> Paolo
>>>
>>
>> What do you think about this? It moves the checks into the functions
>> that actually cause execve() to be called, and it only prevents the
>> commands after QEMU is done with initialization in main().
>
> It doesn't do error reporting correctly because these functions do not
> get an Error **.  If you change that and use error_setg instead of
> error_report, it should be okay.

I just wanted to follow up on a few things..

All of the following functions currently use qerror_report().  I'm 
thinking conversion of these and sub-functions to pass an Error ** 
parameter should be a separate undertaking.

net_init_nic
net_init_slirp
net_init_tap
net_init_socket
net_init_vde
net_init_dump
net_init_bridge
net_init_hubport

>
> However, I really think what your testing is not
> runstate_is_prelaunch(), it is seccomp_effective().  If you structure
> the test like that, it also lets you eliminate the #ifdef (which in
> general we prefer to avoid).

The reason for testing runstate_is_prelaunch() is because seccomp will 
be effective during and after prelaunch.  The only difference is that a 
more restrictive syscall whitelist will be in effect after prelaunch. So 
perhaps the tests can be similar to the following so that we can get rid 
of the preprocessor #ifdef:

if (seccomp_is_effective() && !runstate_is_prelaunch()) {
     error_report("Cannot execute network helper from QEMU monitor "
                  "when -sandbox is in effect");
     return -1;
}

-- 
Regards,
Corey Bryant