Re: [RFC PATCH] cxl: Allow reprogramming misconfigured hdm decoders

[Alejandro Lucero Palau <alucerop@amd.com>]

On 4/30/25 04:24, Zhijian Li (Fujitsu) wrote:
>
> On 30/04/2025 10:17, Gregory Price wrote:
>> On Wed, Apr 30, 2025 at 09:29:15AM +0800, Li Zhijian wrote:
>>> During kernel booting, CXL drivers will attempt to construct the CXL region
>>> according to the pre-programed(firmware provisioning) HDM decoders.
>>>
>>> This construction process will fail for some reasons, in this case, the
>>> userspace cli like ndctl/cxl cannot destroy nor create regions upon the
>>> existing decoders.
>>>
>>> Introuce a new flag CXL_DECODER_F_NEED_RESET tell the driver to reset
>>> the decoder during `cxl destroy-region regionN`, so that region can be
>>> create again after that.
>>>
>> My best understanding of why this is disallowed is that firmware/bios
>> programmed decoders need to be locked because there is an assumption
>> that the platform programmed it that way *for a reason* - and that
>> changing the programming would break it (cause MCEs for other reasons,
>> etc).
>
> Hi Gregory,
>
> Thank you for the feedback. Based on current CXL driver behavior, user-space tools
> can indeed reprogram firmware-provisioned HDM decoders in practice.
>
> For example, after a successful boot, one may destroy the auto-constructed region
> via cxl destroy-region and create a new different region.
> This indicates that the kernel does not inherently lock down these decoders.
>
> As for the locking rationale you mentioned, platform vendors might enforce their policies
> through mechanisms like the *Lock-On-Commit* in CXL HDM Decoder n Control Register
>
> While platform vendors may have valid considerations (as you noted), from a driver and
> end-user perspective, depending solely on firmware updates to fix transient failures
> is not smooth sometimes :).
>

Hi Zhijan,


 From my current effort trying to get a Type2 device properly 
initialized by the kernel after the BIOS/platform firmware doing 
whatever it needs to do, I really think we should have a wider 
discussion regarding this sync, and maybe to have first something from 
the kernel expectation of what the BIOS should and should not do.


If this makes sense, I could work on a initial draft about the outline 
or points to discuss about this.


Thank you


>> So the appropriate solution here is for the platform vendor to fix their
>> firmware.
>>
>> But I am not a platform people - so I will defer to them on whether my
>> understanding is correct.
> Yeah, it's still in the RFC stage, let's hear more voices.
>
> Thanks
> Zhijian
>
>> ~Gregory
>>