Re: [RFC 0/3] SCMI Vhost and Virtio backend implementation
From: Mike Tipton
Date: Fri Jul 08 2022 - 23:28:36 EST
I'll let Neeraj respond to more of the core backend details and policy
enforcement options, but I can provide some details for our prototype
clock protocol handler. Note that it's a pretty simple proof-of-concept
handler that's implemented entirely outside of the common clock
framework. It operates as just another client to the framework. This
approach has some limitations. And a more full-featured implementation
could benefit from being implemented in the clock framework itself. But
that level of support hasn't been necessary for our purposes yet.
On 6/13/2022 10:20 AM, Cristian Marussi wrote:
+CC: Souvik
On Thu, Jun 09, 2022 at 12:49:53PM +0530, Neeraj Upadhyay wrote:
This RFC series, provides ARM System Control and Management Interface (SCMI)
protocol backend implementation for Virtio transport. The purpose of this
Hi Neeraj,
Thanks for this work, I only glanced through the series at first to
grasp a general understanding of it (without goind into much details for
now) and I'd have a few questions/concerns that I'll noted down below.
I focused mainly on the backend server aims/functionalities/issues ignoring
at first the vhost-scmi entry-point since the vost-scmi accelerator is just
a (more-or-less) standard means of configuring and grabbing SCMI traffic
from the VMs into the Host Kernel and so I found more interesting at first
to understand what we can do with such traffic at first.
(IOW the vhost-scmi layer is welcome but remain to see what to do with it...)
feature is to provide para-virtualized interfaces to guest VMs, to various
hardware blocks like clocks, regulators. This allows the guest VMs to
communicate their resource needs to the host, in the absence of direct
access to those resources.
In an SCMI stack the agents (like VMs) issue requests to an SCMI platform
backend that is in charge of policying and armonizing such requests
eventually denying some of these (possibly malicious) while allowing others
(possibly armonizing/merging such reqs); with your solution basically the
SCMI backend in Kernel marshals/conveys all of such SCMI requests to the
proper Linux Kernel subsystem that is usually in charge of it, using
dedicated protocol handlers that basically translates SCMI requests to
Linux APIs calls to the Host. (I may have oversimplified or missed
something...)
At the price of a bit of overhead and code-duplication introduced by
this SCMI Backend you can indeed leverage the existing mechanisms for
resource accounting and sharing included in such Linux subsystems (like
Clock framework), and that's nice and useful, BUT how do you policy/filter
(possibly dinamically as VMs come and go) what these VMs can see and do
with these resources ?
Currently, our only level of filtering is for which clocks we choose to
expose over SCMI. Those chosen clocks are exposed to all VMs equally.
The clock protocol handler exposes a registration function, which we
call from our clock drivers. Which clocks we register are currently
hardcoded in the drivers themselves. We often want to register all the
clocks in a given driver, since we have separate drivers for each clock
controller and many clock controllers are already dedicated to a
particular core or subsystem. So if that core or subsystem needs to be
controlled by a VM, then we give the VM all of its clocks. This can mean
exposing a large number of clocks (in the hundreds).
... MORE importantly how do you protect the Host (or another VM) from
unacceptable (or possibly malicious) requests conveyed from one VM request
vqueue into the Linux subsystems (like clocks) ?
The clock protocol handler tracks its own reference counts for each
clock that's been registered with it. It'll only enable clocks through
the host framework when the reference count increases from 0 -> 1, and
it'll only disable clocks through host framework when the reference
count decreases from 1 -> 0. And since the clock framework has its own
internal reference counts, then it's not possible for a VM to disable
clocks that the host itself has enabled.
We don't support frequency aggregation, so a VM could override the
frequency request of another VM or of the host. We could support max
aggregation across VMs, so that a VM couldn't reduce the frequency below
what another VM has requested. But without clock framework changes, we
can't aggregate with the local host clients. So a VM could reduce the
frequency below what the host has requested.
Generally speaking we don't expect more than one entity (VM or host) to
control a given clock at a time. But all we can currently enforce is
that clocks only turn off when *all* entities (including the the host)
request them to be off.
I saw you have added a good deal of DT bindings for the backend
describing protocols, so you could just expose only some protocols via
the backend (if I get it right) but you cannot anyway selectively expose
only a subset of resources to the different agents, so, if you expose the
clock protocol, that will be visible by any VMs and an agent could potentially
kill the Host or mount some clock related attack acting on the right clock.
(I mean you cannot describe in the Host DT a number X of clocks to be
supported by the Host Linux Clock framework BUT then expose selectively to
the SCMI agents only a subset Y < X to shield the Host from misbehaviour...
...at least not in a dynamic way avoiding to bake a fixed policy into
the backend...or maybe I'm missing how you can do that, in such a case
please explain...)
Moreover, in a normal SCMI stack the server resides out of reach from the
OSPM agents since the server, wherever it sits, has the last word and can
deny and block unreasonable/malicious requests while armonizing others: this
means the typical SCMI platform fw is configured in such a way that clearly
defines a set of policies to be enforced between the access of the various
agents. (and it can reside in the trusted codebase given its 'reduced'
size...even though this policies are probably at the moment not so
dynamically modificable there either...)
With your approach of a Linux Kernel based SCMI platform backend you are
certainly using all the good and well proven mechanisms offered by the
Kernel to share and co-ordinate access to such resources, which is good
(.. even though Linux is not so small in term of codebase to be used as
a TCB to tell the truth :D), BUT I don't see the same level of policying
or filtering applied anywhere in the proposed RFCs, especially to protect
the Host which at the end is supposed to use the same Linux subsystems and
possibly share some of those resources for its own needs.
I saw the Base protocol basic implementation you provided to expose the
supported backend protocols to the VMs, it would be useful to see how
you plan to handle something like the Clock protocol you mention in the
example below. (if you have Clock protocol backend that as WIP already
would be interesting to see it...)
Another issue/criticality that comes to my mind is how do you gather in
general basic resources states/descriptors from the existing Linux subsystems
(even leaving out any policying concerns): as an example, how do you gather
from the Host Clock framework the list of available clocks and their rates
descriptors that you're going expose to a specific VMs once this latter will
issue the related SCMI commands to get to know which SCMI Clock domain are
available ?
(...and I mean in a dynamic way not using a builtin per-platform baked set of
resources known to be made available... I doubt that any sort of DT
description would be accepted in this regards ...)
As mentioned, the list of clocks we choose to expose are currently
hardcoded in the clock drivers outside of the clock framework. There is
no dynamic policy in place.
For supported rates, we currently just implement the
CLOCK_DESCRIBE_RATES command using rate ranges, rather than lists of
discrete rates (num_rates_flags[12] = 1). And we just communicate the
full u32 range 0..U32_MAX with step_size=1. We do this for simplicity.
Many of our clocks only support a small list of discrete rates (though
some support large ranges). If a VM requests a rate not aligned to these
discrete rates, then we'll just round up to what the host supports. We
currently operate under the assumption that the VM knows what it needs
and doesn't need to query the specific supported rates from the host.
That's fine for our current use cases, at least. Publishing
clock-specific rate lists and/or proper ranges would be more complicated
and require some amount of clock framework changes to get this information.
1. Architecture overview
---------------------
Below diagram shows the overall software architecture of SCMI communication
between guest VM and the host software. In this diagram, guest is a linux
VM; also, host uses KVM linux.
GUEST VM HOST
+--------------------+ +---------------------+ +--------------+
| a. Device A | | k. Device B | | PLL |
| (Clock consumer) | | (Clock consumer) | | |
+--------------------+ +---------------------+ +--------------+
| | ^
v v |
+--------------------+ +---------------------+ +-----------------+
| b. Clock Framework | | j. Clock Framework | -->| l. Clock Driver |
+-- -----------------+ +---------------------+ +-----------------+
| ^
v |
+--------------------+ +------------------------+
| c. SCMI Clock | | i. SCMI Virtio Backend |
+--------------------+ +------------------------+
| ^
v |
+--------------------+ +----------------------+
| d. SCMI Virtio | | h. SCMI Vhost |<-----------+
+--------------------+ +----------------------+ |
| ^ |
v | |
+-------------------------------------------------+ +-----------------+
| e. Virtio Infra | | g. VMM |
+-------------------------------------------------+ +-----------------+
| ^ ^
v | |
+-------------------------------------------------+ |
| f. Hypervisor |-------------
+-------------------------------------------------+
Looking at the above schema and thinking out loud where any dynamic
policying against the resources can fit (..and trying desperately NOT to push
that into the Kernel too :P...) ... I think that XEN was trying something similar
(with a real backend SCMI platform FW at the end of the pipe though I think...) and
in their case the per-VMs resource allocation was performed using SCMI
BASE_SET_DEVICE_PERMISSIONS commands issued by the Hypervisor/VMM itself
I think or by a Dom0 elected as a trusted agent and so allowed to configure
such resource partitioning ...
https://www.mail-archive.com/xen-devel@xxxxxxxxxxxxxxxxxxxx/msg113868.html
...maybe a similar approach, with some sort of SCMI Trusted Agent living within
the VMM and in charge of directing such resources' partitioning between
VMs by issuing BASE_SET_DEVICE_PERMISSIONS towards the Kernel SCMI Virtio
Backend, could help keeping at least the policy bits related to the VMs out of
the kernel/DTs and possibly dynamically configurable following VMs lifecycle.
Even though, in our case ALL the resource management by device ID would have to
happen in the Kernel SCMI backend at the end, given that is where the SCMI
platform resides indeed, BUT at least you could keep the effective policy out of
kernel space, doing something like:
1. VMM/TrustedAgent query Kernel_SCMI_Virtio_backend for available resources
2. VMM/TrustedAg decides resources allocation between VMs (and/or possibly the Host
based on some configured policy)
3. VMM/TrustedAgent issues BASE_SET_DEVICE_PERMISSIONS/PROTOCOLS to the
Kernel_SCMI_Virtio_backend
4. Kernel_SCMI_Virtio_backend enforces resource partioning and sharing
when processing subsequent VMs SCMI requests coming via Vhost-SCMI
...where the TrustedAgent here could be (I guess) the VMM or the Host or
both with different level of privilege if you don't want the VMM to be able
to configure resources access for the whole Host.
a. Device A This is the client kernel driver in guest VM,
for ex. diplay driver, which uses standard
clock framework APIs to vote for a clock.
b. Clock Framework Underlying kernel clock framework on
guest.
c. SCMI Clock SCMI interface based clock driver.
d. SCMI Virtio Underlying SCMI framework, using Virtio as
transport driver.
e. Virtio Infra Virtio drivers on guest VM. These drivers
initiate virtqueue requests over Virtio
transport (MMIO/PCI), and forwards response
to SCMI Virtio registered callbacks.
f. Hypervisor Hosted Hypervisor (KVM for ex.), which traps
and forwards requests on virtqueue ring
buffers to the VMM.
g. VMM Virtual Machine Monitor, running on host userspace,
which manages the lifecycle of guest VMs, and forwards
guest initiated virtqueue requests as IOCTLs to the
Vhost driver on host.
h. SCMI Vhost In kernel driver, which handles SCMI virtqueue
requests from guest VMs. This driver forwards the
requests to SCMI Virtio backend driver, and returns
the response from backend, over the virtqueue ring
buffers.
i. SCMI Virtio Backend SCMI backend, which handles the incoming SCMI messages
from SCMI Vhost driver, and forwards them to the
backend protocols like clock and voltage protocols.
The backend protocols uses the host apis for those
resources like clock APIs provided by clock framework,
to vote/request for the resource. The response from
the host api is parceled into a SCMI response message,
and is returned to the SCMI Vhost driver. The SCMI
Vhost driver in turn, returns the reponse over the
Virtqueue reponse buffers.
Last but not least, this SCMI Virtio Backend layer in charge of
processing incoming SCMI packets, interfacing with the Linux subsystems
final backend and building SCMI replies from Linux will introduce a
certain level of code/funcs duplication given that this same SCMI basic
processing capabilities have been already baked in the SCMI stacks found in
SCP and in TF-A (.. and maybe a few other other proprietary backends)...
... but this is something maybe to be addressed in general in a
different context not something that can be addressed by this series.
Sorry for the usual flood of words :P ... I'll have a more in deep
review of the series in the next days, for now I wanted just to share my
concerns and (maybe wrong) understanding and see what you or Sudeep and
Souvik think about.
Thanks,
Cristian
_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel