Re: [RFC PATCH 2/4] PCI/CXL: Populate and insert/remove pdev->coh_resource[]

["Bowman, Terry" <terry.bowman@amd.com>]

On 10/20/2026 3:41 AM, Ankit Agrawal wrote:
>>> +
>>> +             *res = DEFINE_RES_NAMED_DESC(config->hpa_range.start,
>>> +                                          range_len(&config->hpa_range),
>>> +                                          "CXL coherent memory",
>>> +                                          IORESOURCE_MEM, IORES_DESC_NONE);
>>> +
>>
>> On platforms using Normalized Addressing the Endpoint decoders contain DPA ranges. This 
>> patch will then record DPA ranges as a result. Translating this requires call to prm_cxl_dpa_spa() ?
>> The challenge is the PRM in drivers/cxl/core/atl.c may not be initialized this early.
>>
> 
> Yeah, good point. On the platforms that I used, the HDM are precommitted
> and have HPA. But I agree, the snapshot must not silently record a DPA.
> 
> Is there an existing way to tell at this point that the endpoint decoder is
> DPA-based? If yes, I intend to fail close with coh_resource[] left unpopulated
> Or would you prefer populating lazily until translations are available?
> 

Hi Ankit,

For an example, normalized decoding is detected in core/atl.c:cxl_prm_setup_root(). Of course
this runs much later after cxl_core is loaded. Normalized decoder targets have the following at 
the Endpoint HDM target entries:
(hpa_range.start == cxled->dpa_res->start)

-Terry
> 
>>> +             conflict = insert_resource_conflict(&iomem_resource, res);
>>> +             if (conflict) {
>>> +                     pci_warn(pdev,