Re: [PATCH 3/4] vfio/pci: Add PCIe TPH GET_ST interface

[fengchengwen <fengchengwen@huawei.com>]

Sorry for the self-reply.

Hi Alex & Wathsala,

Based on the VM assignment scenario and the cross-VM attack concern
raised by Wathsala in her review of "[PATCH v2 RESEND 4/5] vfio/pci: Add PCIe TPH GET_ST interface":

  This is unsafe. A user space driver can obtain STs for arbitrary CPUs and program them
  into device-specific registers (e.g., E810), with no isolation guarantees.
  For example, consider two VMs on the same host. A driver in VM1 could program STs that
  target CPUs primarily used by VM2. This can steer traffic processing onto VM2's CPUs,
  creating contention and degrading VM2's performance.
  This breaks CPU isolation at the host level and can be used to disrupt workloads and violate
  SLAs across tenants.

After fully re-evaluating the security architecture with your feedback,
I agree with your concerns and conclusions. I hereby revoke my earlier
proposal to restrict GET_ST to only the current CPU.

For devices that implement standard ST tables (via config space or MSI-X caps),
hypervisors such as QEMU/kvmtool can trap and filter guest writes, preventing
malicious steering tag abuse. This is the safe and supported model for TPH in
virtualized environments.

However, for devices that only support Device-Specific mode with no standard
ST table, there is no existing hypervisor interception mechanism to prevent
a guest from programming arbitrary steering tags to attack other CPUs.
This is a fundamental security risk that cannot be safely mitigated in software.

Therefore, the correct security posture for virtualization is:
- TPH should be enabled *only* for devices with standard ST tables
- Devices without standard ST tables should NOT enable TPH in virtualization

On the other hand, in non-virtualization (bare-metal) scenarios, there is
strong legitimate demand for devices that lack a standard ST table —
many real-world devices are designed this way. For this reason, I would
like to retain the GET_ST interface.

For virtualization scenarios, the hypervisor is responsible for avoiding
this risk by **disabling TPH Requester Enable in the PCIe config space**,
which is fully interceptable and under the hypervisor’s control.

Thanks

On 4/17/2026 8:48 AM, fengchengwen wrote:
> Hi Alex,
> 
> Thank you very much for your clear and detailed security explanation.
> I fully understand and agree with your security concerns about allowing
> userspace to query steering tags for arbitrary CPUs.
> 
> To completely resolve this security issue while retaining the mandatory
> functionality for DS-mode devices without ST table, I will revise the
> GET_ST interface with a strict security constraint in v3:
> 
>     The CPU number provided by userspace will be VALIDATED TO EQUAL
>     THE CURRENT CALLING CPU of the ioctl().
> 
> In other words:
> - Userspace can ONLY query the steering tag for the CPU it is currently
>   running on.
> - Userspace CANNOT query any other CPU.
> - No cross-CPU probing, no side-channel, no attack surface.
> - No ability to influence or target other CPUs.
> 
> This completely eliminates the security exposure you mentioned, while
> still fully supporting the Device-Specific mode requirement for devices
> without ST tables.
> 
> Thanks
> 
> On 4/16/2026 9:40 PM, Alex Williamson wrote:
>> On Thu, 16 Apr 2026 09:09:50 +0800
>> fengchengwen <fengchengwen@huawei.com> wrote:
>>
>>> On 4/15/2026 9:55 PM, Wathsala Vithanage wrote:
>>>> Hi Feng,
>>>>
>>>> get_st  feature is unsafe. It allows a rogue userspace driver in device-specific
>>>> mode to obtain steering tags for arbitrary CPUs, including ones unrelated
>>>> to the device or its workload, enabling it to direct traffic into those CPUs’
>>>> caches and potentially interfere with other workloads, opening doors to
>>>> further exploits depending on other vulnerabilities.  
>>>
>>> Thank you for the follow-up and for referencing the prior RFC
>>> discussion on this topic. I appreciate you clarifying the
>>> historical context of the safety concerns.
>>>
>>> I acknowledge the risks you’ve highlighted, but I believe the
>>> risk profile in this VFIO interface is different and already
>>> well bounded by existing design and practice:
>>>
>>> 1. VFIO device access requires elevated privileges
>>>    A userspace process can only open a VFIO device node if it
>>>    has sufficient privileges (typically root). This is not an
>>>    interface for unprivileged users.
>>
>> This argument is NOT helping your cause.  This is not the usage model
>> we design for.  VFIO usage requires that privileges be granted to a
>> user, in the form of device ACL access and locked memory, but does not
>> generally require elevated privileges beyond that, or otherwise grant
>> the user authority beyond the scope of the device.  The root use case
>> may be typical for you, but is not required for many other typical use
>> cases, such as device assignment to VMs.
>>  
>>> 2. In the thread "[RFC v2 0/2] Retrieve tph from dmabuf for PCIe
>>>    P2P memory access", applications can configure the steertag
>>>    of exported dmabufs from userspace to the kernel. Kernel PCIe
>>>    drivers (e.g., mlx5 NIC) then use these steertags and set them
>>>    to their ST tables. Even here, userspace could set invalid
>>>    steertags that impact GPU performance—but this model is
>>>    basically accepted I think (refer from maillist discuss).
>>
>> It's an RFC.  It's bold to claim that it's nearly accepted.
>>
>>> 3. Malicious resource consumption is not unique to TPH
>>>    A malicious thread can be created to forcibly consume CPU
>>>    resources and bound to a specific CPU, affecting other CPUs.
>>>    This is a general system security concern, not one specific
>>>    to TPH GET_ST, and is addressed by existing system hardening
>>>    and access control mechanisms—not by removing useful features.
>>
>> You're conflating process abuse of a CPU to a potential side-channel
>> DMA attach from a device.  What *existing* hardening protects against
>> the latter?
>>
>>> 4. GET_ST is strictly necessary for Device-Specific (DS) mode
>>>    when no ST table is present on the device.
>>>    For devices that do not have a dedicated ST table (a common
>>>    scenario in many PCIe endpoints), DS mode requires userspace
>>>    to retrieve per-CPU steering tags first, then program them
>>>    into the device’s steering logic via other registers. Without
>>>    GET_ST, userspace cannot obtain the required steertags to
>>>    enable TPH DS mode at all—rendering TPH support useless for
>>>    these devices. This is not an optional feature but a
>>>    fundamental requirement to unlock TPH functionality for a
>>>    large class of hardware.
>>
>> Unlocking a hardware feature does not give you authority to ignore the
>> security implications of that feature.  Thanks,
>>
>> Alex
>