Re: [PATCH V3 0/9] arm64: imx8mn: Enable more imx8m Nano functions

[Lucas Stach]

Am Dienstag, dem 23.11.2021 um 10:40 -0600 schrieb Adam Ford:
> On Tue, Nov 23, 2021 at 8:30 AM Adam Ford <aford173@xxxxxxxxx> wrote:
> > 
> > On Tue, Nov 23, 2021 at 8:24 AM Lucas Stach <l.stach@xxxxxxxxxxxxxx> wrote:
> > > 
> > > Am Dienstag, dem 23.11.2021 um 08:08 -0600 schrieb Adam Ford:
> > > > On Mon, Nov 22, 2021 at 3:52 PM Tim Harvey <tharvey@xxxxxxxxxxxxx> wrote:
> > > > > 
> > > > > On Mon, Nov 22, 2021 at 10:20 AM Lucas Stach <l.stach@xxxxxxxxxxxxxx> wrote:
> > > > > > 
> > > > > > Am Montag, dem 22.11.2021 um 09:59 -0800 schrieb Tim Harvey:
> > > > > > > On Sun, Nov 21, 2021 at 7:25 AM Adam Ford <aford173@xxxxxxxxx> wrote:
> > > > > > > > 
> > > > > > > > On Sun, Nov 21, 2021 at 8:34 AM Adam Ford <aford173@xxxxxxxxx> wrote:
> > > > > > > > > 
> > > > > > > > > On Sun, Nov 21, 2021 at 8:21 AM Fabio Estevam <festevam@xxxxxxxxx> wrote:
> > > > > > > > > > 
> > > > > > > > > > Hi Adam,
> > > > > > > > > > 
> > > > > > > > > > On Sun, Nov 21, 2021 at 11:17 AM Adam Ford <aford173@xxxxxxxxx> wrote:
> > > > > > > > > > 
> > > > > > > > > > > I am using https://source.codeaurora.org/external/imx/imx-atf/log/?h=lf_v2.4
> > > > > > > > > > > 
> > > > > > > > > > > Since the driver sending SMCC commands to ATF isn't doing that, I
> > > > > > > > > > > assume it's safe to use the linux power-domain drivers with the ATF
> > > > > > > > > > > from NXP's kernel.
> > > > > > > > > > > 
> > > > > > > > > > > If you can point me to the repo you think I should be using, I'll give it a try.
> > > > > > > > > > 
> > > > > > > > > > Do you know if the mainline TF-A repo v2.5 works too?
> > > > > > > > > > https://github.com/ARM-software/arm-trusted-firmware/tree/v2.5
> > > > > > > > > 
> > > > > > > > > That's good to know.
> > > > > > > > > 
> > > > > > > > > I just built it into U-Boot:
> > > > > > > > > 
> > > > > > > > > NOTICE:  BL31: v2.5(release):v2.5
> > > > > > > > > NOTICE:  BL31: Built : 08:24:13, Nov 21 2021
> > > > > > > > > 
> > > > > > > > > The Etnaviv driver is still loading without hanging
> > > > > > > > > 
> > > > > > > > > root@beacon-imx8mn-kit:~# dmesg |grep -i etna
> > > > > > > > > [   12.393936] etnaviv etnaviv: bound 38000000.gpu (ops gpu_ops [etnaviv])
> > > > > > > > > [   12.400676] etnaviv-gpu 38000000.gpu: model: GC7000, revision: 6203
> > > > > > > > > [   12.641297] [drm] Initialized etnaviv 1.3.0 20151214 for etnaviv on minor 0
> > > > > > > > > 
> > > > > > > > > 
> > > > > > > > 
> > > > > > > > Tim,
> > > > > > > > 
> > > > > > > > Which version of Nano do you have?  Not all Nano SoC's have a GPU from
> > > > > > > > looking at the datasheet [1] .  I am using MIMX8MN2CVTIZAA (Nano Solo)
> > > > > > > > 
> > > > > > > > [1] - https://www.nxp.com/docs/en/data-sheet/IMX8MNIEC.pdf
> > > > > > > > 
> > > > > > > 
> > > > > > > Adam,
> > > > > > > 
> > > > > > > The board I have here has MIMX8MN5CVTIZAA so i.MX 8M Nano QuadLite
> > > > > > > with 'No GPU' as you expected.
> > > > > > > 
> > > > > > > So I have to add the following to keep my board from hanging after your series:
> > > > > > > diff --git a/arch/arm64/boot/dts/freescale/imx8mn-venice-gw7902.dts
> > > > > > > b/arch/arm64/boot/dts/freescale/imx8mn-venice-gw7902.dts
> > > > > > > index 236f425e1570..0d256a607b7c 100644
> > > > > > > --- a/arch/arm64/boot/dts/freescale/imx8mn-venice-gw7902.dts
> > > > > > > +++ b/arch/arm64/boot/dts/freescale/imx8mn-venice-gw7902.dts
> > > > > > > @@ -251,6 +251,10 @@
> > > > > > >         };
> > > > > > >  };
> > > > > > > 
> > > > > > > +&gpu {
> > > > > > > +       status = "disabled";
> > > > > > > +};
> > > > > > > +
> > > > > > >  &i2c1 {
> > > > > > >         clock-frequency = <100000>;
> > > > > > >         pinctrl-names = "default";
> > > > > > > 
> > > > > > > This situation is similar to the one I encountered with the
> > > > > > > imx8mm-venice-gw7901 where adding the GPC node caused my board (which
> > > > > > > did not power the GPU) to hang until I added disables to the
> > > > > > > device-tree with commit 7973009235e2 ("arm64: dts:
> > > > > > > imx8mm-venice-gw7901.dts: disable pgc_gpumix"). It feels painful to
> > > > > > > have to add patches to keep things from hanging after additional
> > > > > > > functionality is added to dt but perhaps that is more common than I
> > > > > > > think esp for SoC's like IMX8M which have a lot of lingering support
> > > > > > > still coming in.
> > > > > > > 
> > > > > > Yea, it's unfortunate that those patches break your board, but I guess
> > > > > > we need to accept this, while there is still a lot of feature work
> > > > > > going on.
> > > > 
> > > > There are a significant number of peripherals which are defined and
> > > > marked as 'disabled' by default, so I don't think it's unreasonable to
> > > > do that here.
> > > > I'd like to propose we keep the default disabled and people who
> > > > need/want the GPU enabled can turn it on.  Why waste the power if it's
> > > > not needed?
> > > > 
> > > Sure, if a significant number of chips has the GPU disabled, we might
> > > want to keep it disabled in the base dtsi. With those variants it's
> > > always a tradeoff, for example there are SKUs of the i.MX6 that had the
> > > VPU disabled, but very few of those were in the field, so the VPUs are
> > > enabled in the SoC base dtsi and only users of those special SKUs would
> > > need to disable them in the board DT.
> > > 
> > > The power argument isn't valid, as the kernel driver will suspend the
> > > device when not needed, so there is no wasted power (aside from the
> > > sort moment while the driver probes) with the GPU enabled.
> > > 
> > > The rule of thumb for when a device is default enabled in the SoC dsti
> > > has always been (at least for i.MX) that the peripheral must not have a
> > > board level dependency. While a i2c controller obviously needs a i2c
> > > bus connected on the board to fulfill its purpose, a GPU can be used as
> > > color space converter or something like that with no board level
> > > interaction. Now the line is a bit blurred by having multiple power
> > > rails into the SoC, so one could argue that the GPUs and VPUs now have
> > > some board level dependency on the i.MX8M*.
> > 
> > That makes sense.
> > 
> > Do we defer to Shawn as the final arbiter as to whether or not it's
> > enabled/disabled?  It would be nice to get Nano caught up in
> > functionality as much as possible.
> 
> We could add two more device trees, one for 8mnl (lite) and 8mnul (ulta-lite)
> 
> imx8mnl:
> 
> #include imx8mn.dtsi
> 
> &gpu {
>     status = "disabled";
> };
> 
> 
> imx8mnul:
> 
> #include imx8mnl
> 
> &dsi {
>     status = "disabled";
> };
> 
> Then the boards using either lite or ultralite just include their
> respective SoC.dtsi instead of imx8mn.dtsi.  This is similar to what
> we do with the plethora of i.mx6 options.
> 
Yes, that's an option but it quickly blows up in a combinatorial
explosion. As the chips are pin compatible there is a high probability
that hardware makers will start to offer boards with different feature
sets and then you need to have a number of board DTs just to include
the right dtsi for the chip.

> Just a thought.  Although, I really like the idea of the bootloader
> disabling the unavailable nodes.

Yea, it seems there are even fuses that allow to check those
capabilities, see my reply to Tim. So I think relying on the boot
firmware to fix things up would be the best option, as it allows to
keep the number of DTs small and does not place a big burden on the
boot firmware implementation.

Regards,
Lucas