Re: dom0 LInux 5.8-rc5 kernel failing to initialize cooling maps for Allwinner H6 SoC

[André Przywara <andre.przywara@xxxxxxx>]

On 28/07/2020 19:52, Christopher Clark wrote:

Hi Christopher,

wow, this quickly got out of hand. I never meant to downplay anyone's
work here, but on this particular platform some things might look a bit
different than normal. See below.

> On Tue, Jul 28, 2020 at 11:16 AM Stefano Stabellini
> <sstabellini@xxxxxxxxxx> wrote:
>>
>> On Tue, 28 Jul 2020, André Przywara wrote:
>>> On 28/07/2020 11:39, Alejandro wrote:
>>>> Hello,
>>>>
>>>> El dom., 26 jul. 2020 a las 22:25, André Przywara
>>>> (<andre.przywara@xxxxxxx>) escribió:
>>>>> So this was actually my first thought: The firmware (U-Boot SPL) sets up
>>>>> some basic CPU frequency (888 MHz for H6 [1]), which is known to never
>>>>> overheat the chip, even under full load. So any concern from your side
>>>>> about the board or SoC overheating could be dismissed, with the current
>>>>> mainline code, at least. However you lose the full speed, by quite a
>>>>> margin on the H6 (on the A64 it's only 816 vs 1200(ish) MHz).
>>>>> However, without the clock entries in the CPU node, the frequency would
>>>>> never be changed by Dom0 anyway (nor by Xen, which doesn't even know how
>>>>> to do this).
>>>>> So from a practical point of view: unless you hack Xen to pass on more
>>>>> cpu node properties, you are stuck at 888 MHz anyway, and don't need to
>>>>> worry about overheating.
>>>> Thank you. Knowing that at least it won't overheat is a relief. But
>>>> the performance definitely suffers from the current situation, and
>>>> quite a bit. I'm thinking about using KVM instead: even if it does
>>>> less paravirtualization of guests,
>>>
>>> What is this statement based on? I think on ARM this never really
>>> applied, and in general whether you do virtio or xen front-end/back-end
>>> does not really matter.
> 
> When you say "in general" here, this becomes a very broad statement
> about virtio and xen front-end/back-ends being equivalent and
> interchangable, and that could cause some misunderstanding for a
> newcomer.
> 
> There are important differences between the isolation properties of
> classic virtio and Xen's front-end/back-ends -- and also the Argo
> transport. It's particularly important for Xen because it has
> priortized support for stronger isolation between execution
> environments to a greater extent than some other hypervisors. It is a
> critical differentiator for it. The importance of isolation is why Xen
> 4.14's headline feature was support for Linux stubdomains, upstreamed
> to Xen after years of work by the Qubes and OpenXT communities.

He was talking about performance. My take on this was that this seems to
go back to the old days, when Xen was considered faster because of
paravirt (vs. trap&emulate h/w in QEMU). And this clearly does not apply
anymore, and never really applied to ARM.

>>> IMHO any reasoning about performance just based
>>> on software architecture is mostly flawed (because it's complex and
>>> reality might have missed some memos ;-)
> 
> That's another pretty strong statement. Measurement is great, but
> maybe performance analysis that is informed and directed by an
> understanding of the architecture under test could potentially be more
> rigorous and persuasive than work done without it?

You seem to draw quite a lot from my statement. All I was saying that
modern systems are far too complex to reason about actual performance
based on some architectural ideas.
Also my statement was in response to some generic statement, but of
course in this particular context. Please keep in mind that we are
talking about a 5 US$ TV-box SoC here, basically a toy platform. The
chip has severe architectural issues (secure devices not being secure,
critical devices not being isolated). I/O probably means SD card at
about 25MB/s, the fastest I have seen is 80MB/s on some better (but
optional!) eMMC modules. DRAM is via a single channel 32bit path. The
cores are using an almost 8 year old energy-efficient
micro-architecture. So whether any clever architecture really
contributes to performance on this system is somewhat questionable.

So I was suggesting that before jumping to conclusions based on broad
architectural design ideas an actual reality check of whether those
really apply to the platform might be warranted.
Also I haven't seen what kind of performance he is actually interested
in. Is the task at hand I/O bound, memory bound, CPU bound?
The discussion so far was about the CPU clock frequency only.

>>> So just measure your particular use case, then you know.
> 
> Hmm.

Is this questioning the usefulness of actual performance measurement? He
seems to be after a particular setup, so keeping an eye on the *actual*
performance outcome seems quite reasonable to me.

>>>> I'm sure that the ability to use
>>>> the maximum frequency of the CPU would offset the additional overhead,
>>>> and in general offer better performance. But with KVM I lose the
>>>> ability to have individual domU's dedicated to some device driver,
>>>> which is a nice thing to have from a security standpoint.
>>>
>>> I understand the theoretical merits, but a) does this really work on
>>> your board and b) is this really more secure? What do you want to
>>> protect against?
>>
>> For "does it work on your board", the main obstacle is typically IOMMU
>> support to be able to do device assignment properly. That's definitely
>> something to check. If it doesn't work nowadays you can try to
>> workaround it by using direct 1:1 memory mappings [1].  However, for
>> security then you have to configure a MPU. I wonder if H6 has a MPU and
>> how it can be configured. In any case, something to keep in mind in case
>> the default IOMMU-based setup doesn't work for some reason for the
>> device you care about.

It's even worse: this SoC only provides platform devices, which all rely
on at least pinctrl, clocks and regulators to function. All of this
functionality is provided via centralised devices, probably controlled
by Dom0 (or just one domain, anyway). The MMC controller for instance
needs to adjust the SD bus clock to the storage array dynamically, which
requires to reprogram the CCU. So I don't see how a driver domain would
conceptually work, without solving the very same problems that we just
faced with cpufreq here.

And of course this device does not have an IOMMU worth mentioning: there
is some device with that name, but it mostly provides scatter-gather
support for the video and display devices *only*.
The MMC controller has its own built-in DMA controller, so it owns the
*whole* of memory, including Xen's very own one.

>> For "is this really more secure?", yes it is more secure as you are
>> running larger portions of the codebase in unprivileged mode and isolated
>> from each other with IOMMU (or MPU) protection. See what the OpenXT and
>> Qubes OS guys have been doing.
> 
> Yes. Both projects have done quite a lot of work to enable and
> maintain driver domains.

Which don't work here, see above. Besides, I was genuinely interested in
the actual threat model here. What do we expect to go wrong and how
would putting the driver in its own domain help? (while considering the
platform's limitations)

Cheers,
Andre