Re: [PATCH] docs: fusa: Add Assumption of Use (AOU)

[Stefano Stabellini <sstabellini@xxxxxxxxxx>]

On Mon, 9 Sep 2024, Julien Grall wrote:
> On 09/09/2024 10:50, Ayan Kumar Halder wrote:
> > On 09/09/2024 10:11, Julien Grall wrote:
> > > 
> > > 
> > > On 09/09/2024 09:56, Michal Orzel wrote:
> > > > Hi Julien,
> > > > 
> > > > On 08/09/2024 23:05, Julien Grall wrote:
> > > > > 
> > > > > 
> > > > > Hi Ayan,
> > > > > 
> > > > > On 06/09/2024 11:13, Ayan Kumar Halder wrote:
> > > > > > From: Michal Orzel <michal.orzel@xxxxxxx>
> > > > > > 
> > > > > > AOU are the assumptions Xen relies on other components (eg platform,
> > > > > > domains)
> > > > > 
> > > > > Searching online, I think the abbrevition is AoU rather than AOU. This
> > > > > would also match how we abbreviate in Xen (IOW if we use a lower case
> > > > > letter from the expanded name, then the letter in the acronym is also
> > > > > lower case).
> > > > > 
> > > > > > to fulfill its requirements. In our case, platform means a
> > > > > > combination of
> > > > > > hardware, firmware and bootloader.
> > > > > > 
> > > > > > We have defined AOU in the intro.rst and added AOU for the generic
> > > > > > timer.
> > > > > > 
> > > > > > Signed-off-by: Michal Orzel <michal.orzel@xxxxxxx>
> > > > > > Signed-off-by: Ayan Kumar Halder <ayan.kumar.halder@xxxxxxx>
> > > > > > ---
> > > > > >    .../reqs/design-reqs/arm64/generic-timer.rst  | 19 +++++++++++++
> > > > > > ++++++
> > > > > >    docs/fusa/reqs/intro.rst                      | 10 ++++++++++
> > > > > >    2 files changed, 29 insertions(+)
> > > > > > 
> > > > > > diff --git a/docs/fusa/reqs/design-reqs/arm64/generic-timer.rst b/
> > > > > > docs/fusa/reqs/design-reqs/arm64/generic-timer.rst
> > > > > > index f2a0cd7fb8..9df87cf4e0 100644
> > > > > > --- a/docs/fusa/reqs/design-reqs/arm64/generic-timer.rst
> > > > > > +++ b/docs/fusa/reqs/design-reqs/arm64/generic-timer.rst
> > > > > > @@ -116,6 +116,25 @@ Rationale:
> > > > > > 
> > > > > >    Comments:
> > > > > > 
> > > > > > +Covers:
> > > > > > + - `XenProd~emulated_timer~1`
> > > > > > +
> > > > > > +Assumption of Use on the Platform
> > > > > > +=================================
> > > > > > +
> > > > > > +Expose system timer frequency via register
> > > > > > +------------------------------------------
> > > > > > +
> > > > > > +`XenSwdgn~arm64_generic_timer_pf_program_cntfrq_el0~1`
> > > > > > +
> > > > > > +Description:
> > > > > > +Underlying platform shall ensure that CNTFRQ_EL0 register contains
> > > > > > the system
> > > > > > +timer frequency.
> > > > > 
> > > > > The wording in [1] (not yet merged) implies that CNTFRQ_EL0 may be
> > > > It is merged:
> > > > https://xenbits.xen.org/gitweb/?
> > > > p=xen.git;a=commit;h=51ad2c57a2d21b583a5944a0dc21c709af022f3c
> > > > 
> > > > > invalid. This seems to contradict the Assumption of Use. Can you
> > > > > explain
> > > > > the difference?
> > > > The requirement you refer to is written from a domain perspective and is
> > > > about Xen exposing the frequency
> > > > to domains via CNTFRQ and/or dt property. In case of a presence of dt
> > > > property in the host dtb, Xen could for instance decide
> > > > to emulate CNTFRQ instead of relying on the domain to parse the dt at
> > > > runtime.
> > > 
> > > AFAICT, you can't trap CNTFRQ access. So what you suggest would not work.
> > > 
> > > > 
> > > > The AoU on the platform (hw/firmware/bootloader) is written from Xen
> > > > perspective and is about the platform
> > > > exposing the correct frequency via register. This is Xen expected
> > > > behavior on the platform. In other words, the platform should
> > > > expose the correct frequency via register.
> > > 
> > > Xen is able to deal with broken CNTFRQ_EL0. 
> > Yes, this is correct if the user provides "clock-frequency" dt property.
> > > So I don't understand why we we would want to make an assumption that it
> > > shall not be broken. What do you gain?
> > 
> > Refer https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/
> > linux.git/tree/Documentation/devicetree/bindings/timer/arm,arch_timer.yaml
> > 
> > ```
> > 
> > Use of this property is strongly discouraged; fix your firmware unless
> > absolutely impossible.
> > 
> > ```
> > 
> > We wish to put the onus on the platform/firmware provider to program the
> > register correctly. Otherwise, we will have to document it somewhere (may be
> > safety manual) that User needs to provide the "clock-frequency" dt property.
> 
> I think you will have to. The integrator may not have the possibility to
> modify the firmware.

Without getting into the specifics of CNTFRQ_EL0 and clock-frequency,
given that this is one of the first AoUs, let me clarify the spirit of
the AoUs.

When we say that Xen is "safe" we mean that it went through thousands of
tests so we are sure that in this specific configuration it is as
bug-free as we can reasonably make it.

"in this specific configuration" is important. Changing the
configuration might expand the scope or invalidate some of the tests.
Think of moving from a board with GICv2 to GICv3 as an example (we are
actually targeting GICv3 for safety, so this is not a great example,
but just to explain the point.)

So the AoUs are the set of assumptions Xen has toward the rest of the
system to make sure Xen operates "safely", with the word "safely"
defined as above.

Of course, Xen could totally work on systems with different AoUs (see
the GICv2 vs GICv3 example) but it would be outside the safety
parameters. In a way, it is similar to "security supported": there are
a bunch of Xen features that should work fine but are outside of
"security support" for one reason or the other.

If a user wants to use Xen on a system that breaks one of the AoUs, they
can, but we wouldn't promise it is "safe". For instance, imagine a user
running Xen on a GICv3 system if the safe version of Xen only validated
the GICv2 driver. Similarly to "security support", sometimes it is a bit
of a judgement call and it could be argued either way.

In the specific case of CNTFRQ_EL0, if we think Xen can be "safe" on a
system with a broken CNTFRQ_EL0 (thanks to the clock-frequency DT
property or other mechanisms), then we can remove this from the AoU. We
would probably have to have a different AoU about the presence of
clock-frequency. Otherwise, if we think we cannot really promise that
Xen is "safe" if CNTFRQ_EL0 is broken, then it is better to leave as is.

Keep in mind that users interested in safety, they are very likely to be
interested in the safety-certification of the entire system, which
includes the hardware as well. It is very likely that users will choose
a safety-certified board, which I am guessing would have a working
CNTFRQ_EL0. This is just a guess, I don't know the relationship between
CNTFRQ_EL0 and achieving hardware safety certifications.