Re: [PATCH v6 01/12] dt-bindings: add img, pvrsgx.yaml for Imagination GPUs

[Maxime Ripard <maxime@cerno.tech>]

[-- Attachment #1: Type: text/plain, Size: 6575 bytes --]

Hi,

On Wed, Apr 15, 2020 at 06:42:18PM +0200, H. Nikolaus Schaller wrote:
> > Am 15.04.2020 um 18:21 schrieb Maxime Ripard <maxime@cerno.tech>:
> > 
> > On Wed, Apr 15, 2020 at 05:09:45PM +0200, H. Nikolaus Schaller wrote:
> >> Hi Maxime,
> >> 
> >> Hm. Yes. We know that there likely are clocks and maybe reset
> >> but for some SoC this seems to be undocumented and the reset
> >> line the VHDL of the sgx gpu provides may be permanently tied
> >> to "inactive".
> >> 
> >> So if clocks are optional and not provided, a driver simply can assume
> >> they are enabled somewhere else and does not have to care about. If
> >> they are specified, the driver can enable/disable them.
> > 
> > Except that at the hardware level, the clock is always going to be
> > there. You can't control it, but it's there.
> 
> Sure, we can deduce that from general hardware design knowledge.
> But not every detail must be described in DT. Only the important
> ones.
> 
> >>> If OMAP is too much of a pain, you can also make
> >>> a separate binding for it, and a generic one for the rest of us.
> >> 
> >> No, omap isn't any pain at all.
> >> 
> >> The pain is that some other SoC are most easily defined by clocks in
> >> the gpu node which the omap doesn't need to explicitly specify.
> >> 
> >> I would expect a much bigger nightmare if we split this into two
> >> bindings variants.
> >> 
> >>> I'd say that it's pretty unlikely that the clocks, interrupts (and
> >>> even regulators) are optional. It might be fixed on some SoCs, but
> >>> that's up to the DT to express that using fixed clocks / regulators,
> >>> not the GPU binding itself.
> >> 
> >> omap already has these defined them not to be part of the GPU binding.
> >> The reason seems to be that this needs special clock gating control
> >> especially for idle states which is beyond simple clock-enable.
> >> 
> >> This sysc target-module@56000000 node is already merged and therefore
> >> we are only adding the gpu child node. Without defining clocks.
> >> 
> >> For example:
> >> 
> >> 		sgx_module: target-module@56000000 {
> >> 			compatible = "ti,sysc-omap4", "ti,sysc";
> >> 			reg = <0x5600fe00 0x4>,
> >> 			      <0x5600fe10 0x4>;
> >> 			reg-names = "rev", "sysc";
> >> 			ti,sysc-midle = <SYSC_IDLE_FORCE>,
> >> 					<SYSC_IDLE_NO>,
> >> 					<SYSC_IDLE_SMART>;
> >> 			ti,sysc-sidle = <SYSC_IDLE_FORCE>,
> >> 					<SYSC_IDLE_NO>,
> >> 					<SYSC_IDLE_SMART>;
> >> 			clocks = <&gpu_clkctrl OMAP5_GPU_CLKCTRL 0>;
> >> 			clock-names = "fck";
> >> 			#address-cells = <1>;
> >> 			#size-cells = <1>;
> >> 			ranges = <0 0x56000000 0x2000000>;
> >> 
> >> 			gpu: gpu@0 {
> >> 				compatible = "ti,omap5-sgx544-116", "img,sgx544-116", "img,sgx544";
> >> 				reg = <0x0 0x10000>;
> >> 				interrupts = <GIC_SPI 21 IRQ_TYPE_LEVEL_HIGH>;
> >> 			};
> >> 		};
> >> 
> >> The jz4780 example will like this:
> >> 
> >> 	gpu: gpu@13040000 {
> >> 		compatible = "ingenic,jz4780-sgx540-130", "img,sgx540-130", "img,sgx540";
> >> 		reg = <0x13040000 0x4000>;
> >> 
> >> 		clocks = <&cgu JZ4780_CLK_GPU>;
> >> 		clock-names = "gpu";
> >> 
> >> 		interrupt-parent = <&intc>;
> >> 		interrupts = <63>;
> >> 	};
> >> 
> >> So the question is which one is "generic for the rest of us"?
> > 
> > I'd say the latter.
> 
> Why?
> 
> TI SoC seem to be the broadest number of available users
> of sgx5xx in the past and nowadays. Others are more the exception.

And maybe TI has some complicated stuff around the GPU that others don't have?
If I look quickly at the Allwinner stuff, I see nothing looking alike in the
SoC, so making the binding like that for everyone just because TI did something
doesn't really make much sense.

> > If your clock is optional, then you define it but don't mandate
> > it. Not documenting it will only result in a mess where everyone will
> > put some clock into it, possibly with different semantics each and
> > every time.
> 
> So you mean that we should require a dummy clock for the omap gpu node
> or did I misunderstand that?
>
> Well, yes there is of course a clock connection between the
> omap target-module and the sgx but it is IMHO pointless to
> describe it because it can't and does not need to be controlled
> separately.
> 
> As said the target-module is already accepted and upstream and my
> proposal is to get the gpu node described there. There is simply
> no need for a clocks node for the omap.

There is no need for a clocks property *currently* *on the OMAP*.

> What I also assume is that developers of DTS know what they do.
> So the risk that there is different semantics is IMHO very low.

Well, they know what they do if you document the binding. Let's say I have two
clocks now on my SoC, and you just document that you want a clocks property,
with a generic name in clock-names like "gpu".

> If you agree I can add the clocks/clock-names property as an
> optional property. This should solve omap and all others.

With the above example, what clock should I put in there? In which order? This
isn't some random example pulled out of nowhere. The Allwinner A31 has (at
least) 4 clocks for the GPU, 1 reset line and 1 regulator, so I can only assume
that the GPU actually needs at least that amount to be properly integrated into
an SoC.

This has nothing to do with being dumb or smart.

> > This has nothing to do with the binding being complete. And if you use
> > a binding like this one, you'll be severely limited when you'll want
> > to implement things like DVFS.
> 
> Now you have unhooked me... Nobody seems to know if and how DVFS can be
> applied to SGX. IMHO we should bake small bread first and get initial
> support into mainline.

On the software side, yes, of course. But the discussion here doesn't have much
to do with software support, this is about the hardware. No matter if you enable
DVFS or not, you'll have those resources connected to the GPU.

And if you want to enable the strict minimum in DT for now and expand it later
as the software gains support for more stuff, then you'll have to deal with the
minimal stuff in software later-on to keep the backward compatibility.

But given that the current state on the Allwinner SoCs (at least) is that you
can't even read a register, it might be a good idea to delay the introduction of
that binding until you have something that works to avoid drowning under the
number of special cases to deal with backward compatibility.

Maxime

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]