[PATCH 2/2] PCI: dwc: Add multi-port controller support

[Manivannan Sadhasivam]

On Tue, Jan 06, 2026 at 11:55:46AM +0100, Niklas Cassel wrote:
> On Tue, Jan 06, 2026 at 10:49:19AM +0530, Manivannan Sadhasivam wrote:
> > On Mon, Jan 05, 2026 at 05:19:38PM +0100, Niklas Cassel wrote:
> > > On Mon, Jan 05, 2026 at 05:57:55PM +0530, Sumit Kumar wrote:
> > > > The current DesignWare PCIe RC implementation supports only the controller
> > > > (Host Bridge) node for specifying the Root Port properties in an assumption
> > > > that the underlying platform only supports a single root Port per
> > > > controller instance. This limits support for multi-port controllers where
> > > > different ports may have different lane configurations and speed limits.
> > > > 
> > > > Introduce a separate dw_pcie_port structure to enable multi-port support.
> > > > Each Root Port can have independent lane count, speed limit through pcie at N
> > > > child nodes in device tree. Add dw_pcie_parse_root_ports()
> > > > API to parse these child nodes.
> > > > 
> > > > Equalization presets and link width detection currently use common DBI
> > > > space for all the root ports. Per-port DBI space assignment for these
> > > > features will be added in future.
> > > > 
> > > > Signed-off-by: Sumit Kumar <sumit.kumar at oss.qualcomm.com>
> > > 
> > > Hello Sumit,
> > > 
> > > Is there a reason why you represent this as a list of ports rather than a
> > > simple array?
> > > 
> > > The number of ports is known by parsing the device tree, so it should be
> > > static, no?
> > > 
> > > At least to me, this seem similar to e.g. how a gpio_device has multiple
> > > gpio_descriptors "struct gpio_desc *descs":
> > > https://github.com/torvalds/linux/blob/master/drivers/gpio/gpiolib.h#L68C1-L68C26
> > > 
> > > A list is usually used for something that is dynamic.
> > > I don't think that the number of ports to a PCIe controller will be dynamic.
> > > 
> > > I can see that struct qcom_pcie in pcie-qcom.c has struct list_head ports,
> > > but that does not necessarily mean that we need to have a list of ports in
> > > pcie-designware-host.c. (pcie-qcom could also be modified to have an array
> > > of ports if there is a desire for similar design pattern.)
> > > 
> > 
> > Only reason why I went with lists in pcie-qcom is flexibility. There are useful
> > helpers available for traversing the lists and they seem much more elegant to
> > use rather than manually traversing the array in C.
> > 
> > But to be frank, I don't really care which one is used as there is going to be
> > only a handful of ports at max anyway and there is not much overhead.
> 
> Personally, when I see lists, I automatically think of something that is
> dynamic, so using lists for something static just looks a little bit out of
> place IMHO.
> 
> Technically, the difference is speed. If we want a specific element, we
> will need to traverse the list. With an array, we can access the element
> directly. However, looking at the current patch, it seems like it never
> looks for a specific port, it always does an operation for all ports.
> So from a speed perspective, it doesn't matter, at least not for now.
> 

Yes. I don't envision the driver doing element based lookup even in the future.

> 
> One advantage I can see, instead of doing:
> 
> +	struct dw_pcie_port *port = list_first_entry(&pci->pp.ports,
> +						struct dw_pcie_port, list);
> +	return dw_pcie_wait_for_link(pci, port);
> 
> for drivers with only one port (most drivers), we could just instead do:
> 
> +	return dw_pcie_wait_for_link(pci, pci->pp.port);
> 
> To simply get the first element in the array. No need to sprinkle
> list_first_entry() everywhere in all the drivers if they just have one port.
> 
> 
> For iterating, to avoid manually traversing the array, we could do like
> libata and create a simple macro, e.g. ata_qc_for_each():
> https://github.com/torvalds/linux/blob/v6.19-rc4/drivers/ata/libata-eh.c#L851-L854
> https://github.com/torvalds/linux/blob/v6.19-rc4/include/linux/libata.h#L1657-L1659
> 

I specifically do not want to introduce custom helpers. That's one of my primary
motivation for using lists :)

> And at least personally, I think my brain will parse dw_pcie_port_for_each() { }
> faster than it parses list_for_each_entry(port, &pcie->ports, list) { },
> since it is more unique, but perhaps I am the weird one here :)
> 

Arrays for sure will allow us to do O(1) lookups, but considering that we will
only be traversing the ports from the start, I still prefer using lists.

- Mani

-- 
மணிவண்ணன் சதாசிவம்