Re: [PATCH v11 05/12] cxl: Cache endpoint decoder settings during PCI enumeration

[Li Ming <ming.li@zohomail.com>]

在 2026/9/10 08:07, Srirangan Madhavan 写道:
> On 9/2/26 7:03 AM, Li Ming wrote:
>>> +             return true;
>>> +
>>> +     pci_cxl_hdm_init(pdev);
>>
>> I forgot to ask a question in v10, sorry about that.
>>
>> I am wondering why calling pci_cxl_hdm_init() in
>> cxl_pci_hdm_decoder_count_match() is needed. pci_bus_add_device()
>> already does pci_cxl_hdm_init(), so if my understanding is correct,
>> pdev->hdm is not NULL here, so this pci_cxl_hdm_init() here will not do
>> anything except checking if pdev->hdm is NULL. I am not sure if I miss
>> some cases for that.
>
> Hi Li Ming!
> Your question is valid. pci_bus_add_device() is the normal 
> initialization path. The later call is a retry for cases where that 
> attempt did not publish pdev->hdm, for example because allocation or 
> MMIO mapping failed at that point.
> If initialization succeeded earlier, the existing cache is detected 
> and the retry returns without rebuilding it. I clarified this behavior 
> in the v12 commit message.
>
Hi Srirangan,


Thanks for your explaination.

I think removing this retry should be better if it is only used for the 
case that allocation or MMIO mapping failure during 
pci_bus_add_device(). Because these failures are not expected to happen 
often, and removing it can make pci_cxl_hdm_init() implementation 
simpler. If only pci_bus_add_device() calls pci_cxl_hdm_init(), you 
won't have to hold cxl_rwsem.dpa and check whether pdev->hdm is NULL in 
__pci_cxl_hdm_init(). Besides, hold cxl_rwsem.dpa lock in 
pci_cxl_hdm_release() is not needed, because the function is called in 
pci_release_dev() where is at the end of this pci device's lifecycle.


Ming