Re: Security issue with vmxnet3 and e100 for AMD SEV(-SNP) / Intel TDX

On 1/11/21 2:26 PM, Kirill A. Shutemov wrote:

On Fri, Jan 08, 2021 at 03:31:56PM +0000, Radev, Martin wrote:

quoted

Just noticed that Intel TDX already does the device filtering. Check: https://github.com/intel/tdx/commit/6789eee52aab8985e49b362379fab73aa3eecde2

CC-ing Kirill and Kuppuswamy from Intel in case they want to be part of the discussion.
________________________________
From: Radev, Martin
Sent: Friday, January 8, 2021 12:57 PM
To: netdev@vger.kernel.org <redacted>; intel-wired-lan@lists.osuosl.org <redacted>
Cc: doshir@vmware.com <redacted>; jesse.brandeburg@intel.com <redacted>; anthony.l.nguyen@intel.com <anthony.l.nguyen@intel.com>; Morbitzer, Mathias <redacted>; Robert Buhren <redacted>; file@sect.tu-berlin.de <redacted>; Banse, Christian <redacted>; brijesh.singh@amd.com <redacted>; Thomas.Lendacky@amd.com <Thomas.Lendacky@amd.com>; pv-drivers@vmware.com <redacted>; martin.b.radev@gmail.com <redacted>
Subject: Security issue with vmxnet3 and e100 for AMD SEV(-SNP) / Intel TDX

Hello everybody,

tldr: Both drivers expose skb GVAs to untrusted devices which gives RIP
         control to a malicious e100 / vmxnet3 device implementation. This is
         an issue for AMD SEV (-SNP) [1] and likely Intel TDX [2].

Felicitas and Robert have started a project on fuzzing device drivers which
may have negative security impact on solutions like AMD SEV Secure
Nested Paging and Intel Trusted Domain Extensions. These solutions protect
a VM from a malicious Hypervisor in various way.

There are a couple of devices which carry security issues under the attacker
models of SEV-SNP / Intel TDX, but here we're only discussing VMXNET3 and
e100, because we have detailed PoCs for both.

Maintainers of both vmxnet3 and e100 were added in this email because the
discussion will likely be the same. The issues were already sent to AMD PSIRT,
and Tom Lendacky and Brijesh Singh have volunteered to be part of the email
communication with the maintainers. Both have been working on AMD SEV.

Please check the two attached files: vmxnet3_report.txt and e100_report.txt.
Both contain detailed information about what the issue is and how it can be
exploited by a malicious HV or attacker who has access to the QEMU process.

Fix:
In an earlier discussion with AMD, there was the idea of making a list of
allowed devices with SEV and forbidding everything else. This would avoid
issues with other drivers whose implementation has not been yet scrutinized
under the threat model of SEV-SNP and Intel Trusted Domain Extensions.

+Andi.

Right. Our TDX guest enabling has white list of devices that allowed to be
used. For now it's only VirtIO, but I believe it also requires hardening.
We need to validate any VMM input.

It might be beneficial to have coordination between Intel and AMD on what
devices (and device drivers) considered to be safe for trusted computing.
I think we can share burden of code audit and fuzzing.

Let us know if you are interested in our fuzzing/static analysis setup.
We're planning to submit a paper soon and we will publish the source
code along with the paper.

-- 
Robert Buhren [off-list ref]
Security in Telecommunications <https://sect.tu-berlin.de>
TU Berlin / Telekom Innovation Laboratories
Ernst-Reuter-Platz 7, Sekr TEL 16 / D - 10587 Berlin, Germany
phone: +49 30 835358325