On 24/08/2020 17:57, Bertrand Marquis wrote:
Hi Julien,
Hi,
On 24 Aug 2020, at 14:34, Julien Grall <jul...@xen.org> wrote:
Hi Wei,
On 24/08/2020 04:28, Wei Chen wrote:
Arm ID_AA64PFR0_EL1 register provides two fields to describe CPU
FP/SIMD implementations. Currently, we exactly know the meaning of
0x0, 0x1 and 0xf of these fields. Xen treats value < 8 as FP/SIMD
features presented. If there is a value 0x2 bumped in the future,
Xen behaviors for value <= 0x1 can also take effect. But what Xen
done for value <= 0x1 may not always cover new value 0x2 required.
Right, but this will also happen with all the other features. This may actually
confuse the users as they may think the rest of the features are fully
supported which is not correct. For instance, dom0 will crash if you boot Xen
on a SVE-capable hardware.
I would see this as an improvement already.
TBH, I only view this patch as a band-aid. I am OK with them but they
need to be useful.
When I read "unknown value...", I have no clue whether the message is
good or bad. It would be better to extend the message with what could go
wrong. E.g "Unknown value X, this may result to corruption on the platform".
I would also consider to taint Xen as it may become unstable with this
option set.
More could be done for SVE (and other bits) but this should be in another patch
set.
We throw these messages to break the silence when Xen detected
unknown FP/SIMD IDs to notice user to check.
It feels a bit odd to me to print unknown for the FP/SIMD feature but not for
all the rest.
IMHO, the right approach is to sanitize ID registers exposed to domains and
only expose features we know are correctly handled.
I actually started to look into this last week because I came to an issue
comparable to SVE with pointer authentication.
Maybe we should discuss this subject separately as clearing TID3 bit in HCR and
emulating all ID registers is possible
but I want to check first if this could have big impacts on performances
I don't expect any performance drop. In Linux, they will cache all the
values and use alternative to avoid runtime check.
I also don't expect any OS to use them at runtime in hotpath.
and see discuss how to design this as there
could be a huge amount of cases for example if we want to allow different
parameters for different domains.
I would suggest to first consider a simple approach where we expose the
same view of the ID registers to all the guests but it is sanitized.
We can then discuss whether we want to have a per-guest view of the ID
registers and how to configure it.
Regarding the implementation, I personally quite like the Linux
cpufeature framework. It is quite powerful and I believe would suit
quite well Xen.
Cheers,
--
Julien Grall
