Re: [PATCH 4/4] arm64: dts: uniphier: Add NX1 SoC and boards support

[Arnd Bergmann]

On Thu, Oct 27, 2022, at 13:19, Kunihiko Hayashi wrote:
> On 2022/10/27 19:01, Arnd Bergmann wrote:
>> On Thu, Oct 27, 2022, at 06:51, Kunihiko Hayashi wrote:

>>> +			#interrupt-cells = <1>;
>>> +			interrupt-names = "dma", "msi";
>>> +			interrupts = <GIC_SPI 84 IRQ_TYPE_LEVEL_HIGH>,
>>> +				     <GIC_SPI 85 IRQ_TYPE_LEVEL_HIGH>;
>>> +			interrupt-map-mask = <0 0 0 7>;
>>> +			interrupt-map = <0 0 0 1 &pcie_intc 0>,	/* INTA */
>>> +					<0 0 0 2 &pcie_intc 1>,	/* INTB */
>>> +					<0 0 0 3 &pcie_intc 2>,	/* INTC */
>>> +					<0 0 0 4 &pcie_intc 3>;	/* INTD */
>>> +			phy-names = "pcie-phy";
>>> +			phys = <&pcie_phy>;
>>> +
>>> +			pcie_intc: legacy-interrupt-controller {
>>> +				interrupt-controller;
>>> +				#interrupt-cells = <1>;
>>> +				interrupt-parent = <&gic>;
>>> +				interrupts = <GIC_SPI 86
>> IRQ_TYPE_LEVEL_HIGH>;
>>> +			};
>>> +		};
>> 
>> Shouldn't there be an "msi-map" or "msi-parent" property pointing at
>> the GIC?
>
> Since Designware PCIe receives an interrupt from GIC with interrupt-name "msi"
> and passes the interrupt to the linear irq domain corresponding to MSI,
> I think there is neither "msi-map" nor "msi-parent" properties.

Usually, you have the choice to use either the built-in
MSI logic of the PCIe controller, or the one built into the
GIC itself, assuming you have a modern enough GIC implemetation.

Using the GIC is preferred here, because otherwise you lose
all the benefits that MSIs offer, regarding latency and CPU
affinity.

     Arnd