Re: [PATCH v5 1/3] xen/riscv: always preset A/D bits in G-stage PTEs

[Oleksii Kurochko <oleksii.kurochko@xxxxxxxxx>]

On 10/7/26 10:21 AM, Baptiste Le Duc wrote:
> On 10/7/26 00:50, Teddy Astie wrote:
> > Le 05/10/2026 à 18:02, Baptiste Le Duc a écrit :
> > > A RISC-V implementation can manage the PTE A/D bits in one of two ways:
> > >      1) Update the 'A' and 'D' PTE bits in hardware (ratified as Svadu).
> > >      2) Generate a page fault when 'A' and/or 'D' is clear, so that 
> > > software
> > >         can set them (ratified as Svade).
> > >
> > > p2m_set_pte_flags() presets A/D in G-stage PTEs only when the device 
> > > tree
> > > advertises Svade. Platforms that use scheme (2) without advertising 
> > > Svade,
> > > such as the HiFive Premier P550, then hit guest-page faults that Xen 
> > > does
> > > not handle.
> > I second what Jan said (this is likely a firmware bug, given that the 
> > enumerated behavior is not consistent with the specification, as 
> > existence of Svadu/Svade alters observable behavior).
> Thanks, agreed on both points.
>
> I'll reword the commit message: it shouldn't imply that Svade in the DT 
> means faults will occur, since with Svadu also present that depends on 
> henvcfg.ADUE.
> On the firmware-bug point: I don't think the P550 is really inconsistent 
> with the spec. Svade and Svadu didn't exist when the P550 core was 
> designed. They were only ratified later, to name the two behaviors that 
> already existed in the wild. Before that, the privileged spec just 
> allowed either: hardware updating A/D, or raising a page fault. The P550 
> implements the second, and there was no extension it could advertise for 
> it.
> Therefore, the DT lacks svade because the extension didn't exist when 
> the platform description was written, not because the vendor 
> misdescribed anything.
> I agree that, where Svadu is available, Xen should ideally configure a 
> consistent scheme (ADUE=1) and use ADUE=0 only for specific needs such 
> as dirty tracking. But Xen doesn't support Svadu today, so that is a 
> separate piece of work which I'd prefer to do later.
It would be nice to have such option to configure ADUE bit but IIRC it 
could be configured only in M-mode (so OpenSBI and OpenSBI really set 
ADUE to 1 when Svadu is available) so we will need some kind of new SBI 
call to do that as we are working in HS-mode and can't change ADUE bit.
So IMO we don't really need that for now until we really will start to 
use A/D bits for Xen purpose. For me, at the moment, it is more then 
enough just to have them always set.
~ Oleksii