On Fri, 2020-03-06 at 10:44 +0800, Kever Yang wrote:

Hi Ezequiel,


On 2020/3/5 下午7:35, Ezequiel Garcia wrote:

quoted

(Adding Kever to the loop)

On Thu, 2020-03-05 at 08:32 -0300, Ezequiel Garcia wrote:

quoted

On Thu, 2020-03-05 at 01:51 +0100, Heiko Stübner wrote:

quoted

Hi Ezequiel,

Am Donnerstag, 5. März 2020, 01:03:30 CET schrieb Ezequiel Garcia:

quoted

Hi Heiko,

On Wed, 2020-03-04 at 11:59 +0100, Heiko Stübner wrote:

quoted

Hi,

Am Montag, 2. März 2020, 16:57:02 CET schrieb Mylène Josserand:

quoted

Determine which revision of rk3288 by checking the HDMI version.
According to the Rockchip BSP kernel, on rk3288w, the HDMI
revision equals 0x1A which is not the case for the rk3288 [1].

As these SOC have some differences, the new function
'soc_is_rk3288w' will help us to know on which revision
we are.

what happened to just having a different compatible in the dts?
Aka doing a

rk3288w.dtsi with

#include "rk3288.dtsi"

&cru {
	compatible = "rockchip,rk3288w-cru";
}

I guess you have something like this in mind:

static void __init rk3288_clk_init(struct device_node *np)
{
         __rk3288_clk_init(np, RK3288_SOC_REV_RK3288W);
}
CLK_OF_DECLARE(rk3288_cru, "rockchip,rk3288-cru", rk3288_clk_init);

static void __init rk3288w_clk_init(struct device_node *np)
{
         __rk3288_clk_init(np, RK3288_SOC_REV_RK3288);
}
CLK_OF_DECLARE(rk3288_cru, "rockchip,rk3288w-cru", rk3288w_clk_init);

And the rest is mostly untouched, except the revision is
no longer queried and is now passed by the DT?

Essentially yes, but I guess I was more thinking along the lines of
the rk3188/rk3066a/rk3188a (drivers/clk/rockchip/clk-rk3188.c)

quoted

This would be cleaner for the kernel, with the obvious
drawback being that you now have to maintain
another DTS.

Right now we would end up with the pretty minimal devicetree
having just that cru. So not very invasive.

quoted

This could be an inconvenience. I believe
RK3288W is meant as a direct replacement for RK3288,
so folks building products would expect to just use
RK3288W, and not really bother with passing a
different DTS or what not.

Not sure I follow. As below, I don't think boards will magically switch
between soc variants, so a boards devicetree should just include
the variant - especially as I don't really know how many
additional new boards we will see with it (rk3288 being quite old itself).

It's not about new boards, any company does sells some RK32888-based product,
will eventually want to produce some more.

If I understand correctly, RK3288W is the SoC that Rockchip is now offering,
and not RK3288 anymore.

No, the RK3288 will continue to supply to the legacy projects which may 
have a long contract with Rockchip,

and RK3288W is the recommend to use in all new project.

I guess this means that if a company not having a long contract
wants to keep producing an RK3288-based product, it will have to
resort to RK3288W.

And IIRC, it's meant to be a direct replacement.
 

quoted

quoted

So, if you have to produce another batch of _the same RK3288_ product,
you'll have to use RK3288W. In other words, they would "just switch"
between SoC variants.

In fact, such a case motivates these patches :-)

quoted

quoted

quoted

I somehow don't expect boards to just switch between soc variants
on the fly.

While I agree they are nasty, quirks like this
are not uncommon.

quoted

Also, doing things in mach-rockchip is not very future-proof:

There is actually no reason to keep this in mach-rockchip, right?

The quirk could be placed in other places. For instance,
directly in the clock driver.

Mapping the hdmi controller inside the clock driver to read some "random"
register that hopefully indicates an (undocumented) distinction between soc
variants.

Somehow just having that minimal devicetree for the "w" sounds
way cleaner ;-) .

I agree with Heiko on this.

What Rockchip done is:

- kernel is using "rockchip,rk3288w" compatible to identify rk3288w, 
which is clean in kernel;

- kernel evb dts do not have compatible "rockchip,rk3288w" for we would 
like to identify it dynamic in bootloader;

- We use U-Boot to identify the rk3288w with HDMI reg value, and pass it 
to kernel via modify the

    dtb pass to kernel, so that we can support all kind of rockchip 
customer projects with any kind of rk3288;

- For upstream kernel, it make sense to add "rockchip,rk3288w" in board 
dts for which using rk3288w SoC.

So you proposing to move this revision detection code
from the kernel to all the bootloaders (U-Boot is not the
only one, right?).

I understand that bootloaders are maybe less strict
to quirks like this, but OTOH this approach seems like
just movint things around, plus requiring now to update
two components, instead of just one.

And also, are we sure the bootloaders maintainers
will accept the quirk?

Thanks,
Ezequiel

 




_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel