Re: [edk2-devel] [PATCH v1 0/2] Add support to disable VirtIo net at runtime

["Laszlo Ersek" <lersek@redhat.com>]

On 08/04/22 11:58, Ard Biesheuvel wrote:
> On Thu, 4 Aug 2022 at 07:55, Laszlo Ersek <lersek@redhat.com> wrote:
>>
>> On 08/04/22 04:52, Yuan Yu wrote:
>>> Currently networking can only be enabled/disabled at compile time. This
>>> patch series will add support to disable VirtIo net at runtime even if
>>> the functionality is built into binary at compile time.
>>>
>>> This will enable VMM to reduce attack surface without recompilation.
>>>
>>> The changes can be seen at:
>>> https://github.com/yyu/edk2/tree/network_cfg_lib_v1
>>>
>>> Cc: Ard Biesheuvel <ardb+tianocore@kernel.org>
>>> Cc: Jordan Justen <jordan.l.justen@intel.com>
>>> Cc: Laszlo Ersek <lersek@redhat.com>
>>> Cc: Anthony Perard <anthony.perard@citrix.com>
>>> Cc: Julien Grall <julien@xen.org>
>>>
>>> Yuan Yu (2):
>>>   OvmfPkg: Introduce NetworkCfgLib
>>>   OvmfPkg: Use PcdNetworkSupport to enable/disable VirtIo net
>>>
>>>  OvmfPkg/OvmfPkg.dec                             |  3 ++
>>>  OvmfPkg/OvmfPkgX64.dsc                          |  7 ++++-
>>>  OvmfPkg/Library/NetworkCfgLib/NetworkCfgLib.inf | 29 ++++++++++++++++++
>>>  OvmfPkg/VirtioNetDxe/VirtioNet.inf              |  3 ++
>>>  OvmfPkg/Library/NetworkCfgLib/NetworkCfgLib.c   | 32 ++++++++++++++++++++
>>>  OvmfPkg/VirtioNetDxe/EntryPoint.c               | 10 ++++++
>>>  6 files changed, 83 insertions(+), 1 deletion(-)
>>>  create mode 100644 OvmfPkg/Library/NetworkCfgLib/NetworkCfgLib.inf
>>>  create mode 100644 OvmfPkg/Library/NetworkCfgLib/NetworkCfgLib.c
>>>
>>
>> Well I've not been reviewing upstream edk2 patches for a while, but the
>> virtio-net driver is still very close to my heart, so this patch kind of
>> hits a nerve.
>>
> 
> Welcome back old friend!
> 
>> I think I disagree with the idea and the implementation both.
>>
>> Minimally, the idea needs a much better elaboration -- what is the
>> threat model? Do you want to protect the host from the guest, or the
>> guest from the host? Or something else? How does controlling a single
>> SNP driver via fw_cfg (which is also dictated by the host) help?
>>
> 
> I have to confess that I was the one who suggested this approach to
> Yuan internally, but mainly to get the discussion going, as I was
> anticipating some pushback, just not from you :-)

Heh, sorry about that :)

> 
> 'Reducing the attack surface' is probably not the most accurate
> characterization of the purpose. We are simply looking for a way to
> disable network boot from the vmm/host side without affecting
> how/which network interfaces the guest exposes to the OS.
> 
>> Regarding the implementation: there is much more to networking in edk2
>> than VirtioNetDxe. UEFI driver binaries (SNP drivers) built from iPXE
>> can be passed in via the NICs' option ROMs. SNP drivers can be loaded
>> from the UEFI system partition (for example, Intel's binary-only driver
>> for QEMU's e1000* cards).
>>
>> If you can control this fw_cfg switch from the VMM side, you can also
>> control the VMM enough to simply *not give* a virtio-net device to the
>> guest. Then the driver (it being a UEFI driver following the UEFI driver
>> model) will simply not have anything to bind.
>>
> 
> Sure, but then the OS will lose networking as well. We just want to
> remove the ability to network boot without impacting anything else
> that relies on virtio-net
> 
>> Sorry I find this approach very wrong. If you really need it for your
>> particular VMM, I kind of suggest not upstreaming this patch. I see it
>> as a step backwards for the upstream project.
>>
> 
> If there are better ways to achieve this, we're all ears, but I think
> that having a PCD which could either be fixed at build and compiled
> out completely, or be set via a NULL library resolution, or even be
> wired to a menu option (using PcdsDynamicHii] is a rather low-impact
> but flexible way to go about this.

How about using PCDs, but at a higher level in the edk2 network stack
(regardless of the SNP driver(s) used)?

Not that I'm a huge fan of them, but we already have PcdIPv4PXESupport
and PcdIPv6PXESupport, and they can be set for OVMF and ArmVirtQemu via
fw_cfg. If both PCDs are "PXE_DISABLED" (0), then
PxeBcDriverEntryPoint() [NetworkPkg/UefiPxeBcDxe/PxeBcDriver.c] exits
early. See:

- https://bugzilla.tianocore.org/show_bug.cgi?id=1695
- https://bugzilla.tianocore.org/show_bug.cgi?id=2681

So that's at least "prior art".

If similar PCDs could be introduced for other kinds of network boot (I
think there's only HTTP(S)v[46] to speak of [*]), i.e. in
"NetworkPkg.dec", then I guess setting those from fw_cfg in ArmVirtQemu
and OVMF could be fine. I assume HttpBootDxe would be the driver to
short-circuit.

[*] Hmm, maybe not just that. Do we consider booting off an iSCSI device
"network boot"? Perhaps where you want to "cut the stack" is
Ip4Dxe/Ip6Dxe. Prior art with PXE does suggest PCDs for the higher-level
network boot drivers.

Laszlo