[U-Boot] [PATCH 0/3] sunxi: sun8i-emac: Update DT bindings

[Andre Przywara]

Hi,

On 31/01/18 08:21, Maxime Ripard wrote:
> Hi,
> 
> On Mon, Jan 29, 2018 at 10:38:25AM +0000, Andre Przywara wrote:
>> On 29/01/18 09:58, Maxime Ripard wrote:
>>> On Mon, Jan 29, 2018 at 09:44:44AM +0000, Andre Przywara wrote:
>>>> On 29/01/18 08:51, Maxime Ripard wrote:
>>>>> On Mon, Jan 29, 2018 at 01:15:19AM +0000, Andre Przywara wrote:
>>>>>> The existing sun8i-emac driver in U-Boot uses some preliminary bindings,
>>>>>> which matched our own DTs. Now that the Linux kernel got a driver, lets
>>>>>> update our probe code to handle those Linux DTs as well.
>>>>>> The first patch adds the missing compatible strings for the pinctrl drivers,
>>>>>> which is needed for using the sunxi_name_to_gpio() lookup function.
>>>>>> Patch 2/3 updates the pinctrl parser used in the sun8i-emac driver, to cope
>>>>>> with the new, generic Allwinner pinctrl bindings.
>>>>>> The final patch extends the probe routine in the Ethernet driver to deal
>>>>>> with both the old and the new bindings.
>>>>>
>>>>> Thanks for posting this
>>>>>
>>>>>> This series allows to copy in the DTs from the latest kernel. Unfortunately
>>>>>> right now updating the DTs for the H5 and A64 breaks the build, as the
>>>>>> resulting binary (which embeds the DT) gets to large and triggers our new
>>>>>> image size check.
>>>>>
>>>>> Sigh...
>>>>>
>>>>>> As the H5 and H3 share most of the DT, we can't just update the H3
>>>>>> DTs either. Hopefully we find some neat trick to work around that.
>>>>>
>>>>> Is it just because of the DT size, or because there's more code?
>>>>
>>>> My impression the code itself is always growing a tiny bit over the
>>>> weeks, but this time around it's really the DT update.
>>>> The current A64 .dtbs in U-Boot are around 8KB, mainline is at 13KB.
>>>> Similar for the H5: going from 9.5KB to 14.5KB.
>>>>
>>>> Since you did a pretty good job already in identifying the code hogs, I
>>>> couldn't find *easy* mitigations over the weekend.
>>>> One possible fix is to remove the second .dtb in the Pine64 case, for
>>>> which I sent a patch Friday night.
>>>
>>> Since the DT is fed to the C preprocessor, we could also put some
>>> #ifdef 0 around the nodes that are never used by U-Boot (like the
>>> clocks, timer, psci, dma, GIC, RTC, RSB, etc.)
>>
>> Well yes, U-Boot itself actually only requires a *tiny* .dtb (I think
>> /aliases, /chosen, the reg of USB and Ethernet). But to be honest I
>> don't want to go there. First it would be a constant churn to keep this
>> up-to-date,
> 
> I'm not too worried about the churn, it would be there only for the
> time until we fully migrate to the FAT environment, so one-two release
> now. And we're not syncing the DT very often these days (now that we
> have support for the EMAC and USB that is all U-Boot cares about).
> 
>> but more importantly for proper UEFI boot we just reuse U-Boot's
>> .dtb to pass it on to the kernel. That is actually the purpose of
>> this whole exercise. That already works today (at least for A64),
>> but would benefit from some updates.
>>
>> So I would refrain from tinkering with U-Boot's .dtbs.
> 
> That sucks :/
> 
>>> This should give us some room.
>>>
>>>> Another thing that stuck out is the sha256 checksum. It's "default y" if
>>>> you have FIT. We need FIT for the SPL loader - but we don't do or need
>>>> the checksum there.
>>>> Some people do FIT loading for the kernel and initrd in U-Boot proper, I
>>>> suppose, but I am not sure how many depend on SHA256 checksums in their
>>>> images.
>>>
>>> I think there was someone (Tom?) that said that it was useful in some
>>> circumstances?
>>
>> Yes, I clearly see that it is *useful*, but I wonder how many people
>> would actually miss it today? We would bring it back once we dumped
>> ENV_IS_IN_MMC, so it's only temporarily.
> 
> His words were stronger actually, and he said that we want to keep it.
> 
>> I think we can just disable it in some defconfigs, to avoid collateral
>> damage to other boards.
>> If people have a special need, they can always disable the MMC env and
>> enable stuff at their likings, it's just the standard "make
>> .._defconfig; make" process that needs to be fixed with some band-aids
>> for now.
> 
> I really don't want to go down the "let's fix each defconfig when we
> find out it broke", this is very likely to be broken with no-one
> noticing.

Yeah, that's probably true. Looks we are really running out of silver
bullets :-(

> Is this issue happening when you sync the whole DT, and would it break
> if you just convert the EMAC binding?
> 
> Otherwise, we might try to revive the DTC garbage collection of unused
> nodes patches. This would prevent us from using the overlays on such a
> DT, but that doesn't like like an unfair tradeoff.

Well, what's the point in updating the DT if we then deviate from it
again? I would rather suggest to postpone the DT update, until we ditch
the MMC environment. The whole reason I wanted the update is to pass it
on to Linux, so garbage collection will kill that purpose.

So what about this plan:
- We need to do something *now* about the breakage of the 64-bit sunxi
*_defconfigs, which do not build with origin/master (for me). I tried
GCC 7.1.0, 7.3.0 and 6.4.0, also the actual merge point of the sunxi
branch in addition to origin/master.
The fix I sent (remove the Pine64 "non-plus" .dtb from the FIT image)
was supposed do the trick, but in contrast to my test on Friday doesn't
work (anymore?). bananapi_m64_defconfig for instance fails also (with a
single .dtb).
Can someone confirm this? Maybe it's worth to see what made U-Boot
proper grow in the last few days/weeks?
- Regardless of this we take the patches from this series, so U-Boot
*can* deal with both DTs.
- We wait with any DT updates until we remove the MMC environment
support, so that we have enough space to give U-Boot the proper .dts.

What is the current plan for the MMC environment? Remove it (or not
enable it by default) in 2018.05?

Having the support for both DT bindings in the EMAC driver *now* would
allow people to build their own firmware, *configuring* out the MMC
environment already and not suffer from any limits anymore. It would
just be the _defconfig build which keeps compatibility, for users who care.

Cheers,
Andre.