[libvirt] running Libvirt from source code, IPC_LOCK and VFIO

[Daniel Henrique Barboza]

Hi,

After investigating more I've fixed the problem in a way that I believe it's
worth a patch. I'll be sending it to the ML shortly.


Thanks everyone for the inputs and insights,


DHB

On 2/4/19 10:59 AM, Yuval Shaia wrote:
> + Kamal and Marcel
>
> On Mon, Feb 04, 2019 at 09:58:47AM +0100, Michal Privoznik wrote:
>> On 2/1/19 7:04 PM, Daniel Henrique Barboza wrote:
>>> Hi,
>>>
>>> I'm facing a strange behavior when running Libvirt from source code,
>>> latest upstream, on an Ubuntu 18.04.1 LTS Power 9 server. My QEMU
>>> guest - which is using VFIO and GPU passthrough - breaks on boot when
>>> trying to allocate a DMA window inside KVM.
>>>
>>> Debugging the code, I've found out that the problem is related to the
>>> process
>>> not having CAP_IPC_LOCK - at least from the host kernel perspective.
>>>
>>> This is strange because:
>>>
>>> - the same VM running directly from QEMU command line works
>>> - the same VM running in the system Libvirt (v4.0.0, Ubuntu version)
>>> also works
>>>
>>> What am I missing? My understanding on Linux process is that a process
>>> running as root should inherit the same capabilities of the user, which
>>> includes
>>> CAP_IPC_LOCK. Running Libvirt from source code should grant ipc_lock
>>> to it ... right?
>> No. Ideally, you trust libvirt and want it to manage devices on your system
>> thus it needs all the capabilities. But qemu spawn by libvirt should have no
>> capabilities as libvirt set up everything that's needed for qemu to run. But
>> this is hard to get right - qemu changes and so does the capabilities it may
>> require (these depend on domain configuration anyway). Therefore, it is
>> possible to set libvirt so it does not drop capabilities for qemu process -
>> see clear_emulator_capabilities in qemu.conf - but then libvirt can't
>> guarantee that a compromised qemu does no harm.
> In my case it is not a matter of risk of a malicious guest, it is that the
> device cannot utilize the host device *unless* it has the 'lock'
> capability.
>
>> This corresponds with your finding about ./configure - if there is no
>> libncap-ng found there's no way for libvirt to drop capabilities and thus it
>> doesn't do that.
>>
>> Michal
>>
>> --
>> libvir-list mailing list
>> libvir-list at redhat.com
>> https://www.redhat.com/mailman/listinfo/libvir-list