Re: [PATCH v2 05/16] PCI: Add pci_reachable_set()

[Donald Dutile <ddutile@redhat.com>]

On 7/18/25 1:49 PM, Jason Gunthorpe wrote:
> On Thu, Jul 17, 2025 at 06:04:04PM -0400, Donald Dutile wrote:
>>> Implement pci_reachable_set() to efficiently compute a set of devices on
>>> the same bus that are "reachable" from a starting device. The meaning of
>>> reachability is defined by the caller through a callback function.
>>>
>> This comment made me review get_pci_alias_group(), which states in its description:
>> * Look for aliases to or from the given device for existing groups. DMA
>>   * aliases are only supported on the same bus, therefore the search
>>   * space is quite small
>>
>> So why does it do the for loop:
>>    for_each_pci_dev(tmp) {
>>
>> vs getting the pdev->bus->devices -- list of devices on that bus, and only
>> scan that smaller list, vs all pci devices on the system?
> 
> Because it can't access the required lock pci_bus_sem to use that
> list.
ah, i see; it's only declared in drivers/pci/search.c, and it isn't
a semaphone in a per-bus struct. :-/
... so move the function to search.c ? /me runs...
I know, not worth the churn; already have 'polluted' iommu w/pci, but not vice-versa.
(although iommu-groups is really a bus-op (would be a different op, for say, platform devices going through another iommu).

> 
> The lock is only available within the PCI core itself which is why I
> moved a few functions over there so they can use the lock.
> 
>> Could we move this to just before patch 11 where it is used?
> 
> Yes
> 
>> or could this be used to improve get_pci_alias_group() and get_pci_function_alias_group() ?
> 
> IMHO it is not really worth the churn
> 
Hey, your churning at the moment, so I figured you may want to churn some more! ;-)
just a comment/suggestion; not required.

> Jason
>