Re: [Qemu-devel] [PATCH] Add support for fd: protocol

[Kevin Wolf]

Am 23.05.2011 17:24, schrieb Markus Armbruster:
> Kevin Wolf <address@hidden> writes:
> 
>> Am 20.05.2011 21:53, schrieb Blue Swirl:
>>> On Fri, May 20, 2011 at 10:42 PM, Anthony Liguori <address@hidden> wrote:
>>>> On 05/20/2011 02:25 PM, Blue Swirl wrote:
>>>>>
>>>>> On Fri, May 20, 2011 at 9:48 PM, Corey Bryant<address@hidden>  wrote:
>>>>>>
>>>>>> sVirt provides SELinux MAC isolation for Qemu guest processes and their
>>>>>> corresponding resources (image files). sVirt provides this support
>>>>>> by labeling guests and resources with security labels that are stored
>>>>>> in file system extended attributes. Some file systems, such as NFS, do
>>>>>> not support the extended attribute security namespace, which is needed
>>>>>> for image file isolation when using the sVirt SELinux security driver
>>>>>> in libvirt.
>>>>>>
>>>>>> The proposed solution entails a combination of Qemu, libvirt, and
>>>>>> SELinux patches that work together to isolate multiple guests' images
>>>>>> when they're stored in the same NFS mount. This results in an
>>>>>> environment where sVirt isolation and NFS image file isolation can both
>>>>>> be provided.
>>>>>
>>>>> Very nice. QEMU should use this to support privilege separation. We
>>>>> already have chroot and runas switches, a new switch should convert
>>>>> all file references to fd references internally for that process. If
>>>>> this can be made transparent, this should even be the default way of
>>>>> operation.
>>>>
>>>> You mean, QEMU starts up, opens all disk images, reinvokes itself in a
>>>> confined context, and then passes fds to the child?
>>>
>>> And exit after that, or do the same without forking.
>>>
>>> This wouldn't work now for the native CDROM devices which need to
>>> reopen the device. For that, an explicit reopen method could be added.
>>> The method could even chat with the privileged process to get that to
>>> do the reopening, but I'd leave that to libvirt and fail without it
>>> for plain QEMU.
>>
>> There are more cases where we reopen the image file. One example is the
>> 'commit' monitor command which temporarily reopens the backing file r/w.
>> Or Christoph's patch that allows guests to toggle the write-cache
>> enabled bit. Same for live snapshots. So we'll need a solution for them
>> before doing anything like this.
>>
>> And breaking qemu without libvirt isn't really an option for me.
> 
> Reopening files is evil.  Sometimes flaws in the system call API make it
> the only option.  You can mitigate via /dev/fd/%d, but only on some
> systems.  The less we reopen, the better.
> 
> An fd: protocol can't easily support reopen.  So fail it.  This doesn't
> break any existing usage.  It's just a restriction on the new protocol.
> Restrictions can render the new protocol useless in practice, but we're
> not "breaking qemu without libvirt" there.
> 
> Perhaps we can make relax the restriction on some system by avoiding the
> reopen in a system-dependent way.

Right, you only get the regression once libvirt starts using it (or even
worse, qemu itself, like Blue Swirl suggested). Doesn't make it much better.

Kevin