Re: [RFC PATCH v4 1/5] RISC-V: Detect and Enable Svadu Extension Support

[Alexandre Ghiti <alex@ghiti.fr>]

On 30/05/2024 11:24, Andrew Jones wrote:
> On Thu, May 30, 2024 at 11:01:20AM GMT, Alexandre Ghiti wrote:
>> Hi Andrew,
>>
>> On 30/05/2024 10:47, Andrew Jones wrote:
>>> On Thu, May 30, 2024 at 10:19:12AM GMT, Alexandre Ghiti wrote:
>>>> Hi Yong-Xuan,
>>>>
>>>> On 27/05/2024 18:25, Andrew Jones wrote:
>>>>> On Fri, May 24, 2024 at 06:33:01PM GMT, Yong-Xuan Wang wrote:
>>>>>> Svadu is a RISC-V extension for hardware updating of PTE A/D bits.
>>>>>>
>>>>>> In this patch we detect Svadu extension support from DTB and enable it
>>>>>> with SBI FWFT extension. Also we add arch_has_hw_pte_young() to enable
>>>>>> optimization in MGLRU and __wp_page_copy_user() if Svadu extension is
>>>>>> available.
>>>> So we talked about this yesterday during the linux-riscv patchwork meeting.
>>>> We came to the conclusion that we should not wait for the SBI FWFT extension
>>>> to enable Svadu but instead, it should be enabled by default by openSBI if
>>>> the extension is present in the device tree. This is because we did not find
>>>> any backward compatibility issues, meaning that enabling Svadu should not
>>>> break any S-mode software.
>>> Unfortunately I joined yesterday's patchwork call late and missed this
>>> discussion. I'm still not sure how we avoid concerns with S-mode software
>>> expecting exceptions by purposely not setting A/D bits, but then not
>>> getting those exceptions.
>>
>> Most other architectures implement hardware A/D updates, so I don't see
>> what's specific in riscv. In addition, if an OS really needs the exceptions,
>> it can always play with the page table permissions to achieve such
>> behaviour.
> Hmm, yeah we're probably pretty safe since sorting this out is just one of
> many things an OS will have to learn to manage when getting ported. Also,
> handling both svade and svadu at boot is trivial since the OS simply needs
> to set the A/D bits when creating the PTEs or have exception handlers
> which do nothing but set the bits ready just in case.
>
>>
>>>> This is what you did in your previous versions of
>>>> this patchset so the changes should be easy. This behaviour must be added to
>>>> the dtbinding description of the Svadu extension.
>>>>
>>>> Another thing that we discussed yesterday. There exist 2 schemes to manage
>>>> the A/D bits updates, Svade and Svadu. If a platform supports both
>>>> extensions and both are present in the device tree, it is M-mode firmware's
>>>> responsibility to provide a "sane" device tree to the S-mode software,
>>>> meaning the device tree can not contain both extensions. And because on such
>>>> platforms, Svadu is more performant than Svade, Svadu should be enabled by
>>>> the M-mode firmware and only Svadu should be present in the device tree.
>>> I'm not sure firmware will be able to choose svadu when it's available.
>>> For example, platforms which want to conform to the upcoming "Server
>>> Platform" specification must also conform to the RVA23 profile, which
>>> mandates Svade and lists Svadu as an optional extension. This implies to
>>> me that S-mode should be boot with both svade and svadu in the DT and with
>>> svade being the active one. Then, S-mode can choose to request switching
>>> to svadu with FWFT.
>>
>> The problem is that FWFT is not there and won't be there for ~1y (according
>> to Anup). So in the meantime, we prevent all uarchs that support Svadu to
>> take advantage of this.
> I think we should have documented behaviors for all four possibilities
>
>   1. Neither svade nor svadu in DT -- current behavior
>   2. Only svade in DT -- current behavior
>   3. Only svadu in DT -- expect hardware A/D updating
>   4. Both svade and svadu in DT -- current behavior, but, if we have FWFT,
>      then use it to switch to svadu. If we don't have FWFT, then, oh well...
>
> Platforms/firmwares that aren't concerned with the profiles can choose (3)
> and Linux is fine. Those that do want to conform to the profile will
> choose (4) but Linux won't get the benefit of svadu until it also gets
> FWFT.


I think this solution pleases everyone so I'd say we should go for it, 
thanks Andrew!

@Yong-Xuan do you think you can prepare another spin with Andrew's 
proposal implemented?

Thanks,

Alex


>
> IOW, I think your proposal is fine except for wanting to document in the
> DT bindings that only svade or svadu may be provided, since I think we'll
> want both to be allowed eventually.
>
> Thanks,
> drew
>
> _______________________________________________
> linux-riscv mailing list
> linux-riscv@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-riscv