[PATCH u-boot-marvell 00/18] Upgrade A38x DDR3 training to version 14.0.0

[Chris Packham]

On 9/02/21 3:07 pm, Marek Behun wrote:
> On Tue, 9 Feb 2021 01:08:54 +0000
> Chris Packham <Chris.Packham at alliedtelesis.co.nz> wrote:
>
>> On 9/02/21 1:16 pm, Chris Packham wrote:
>>> On 9/02/21 9:18 am, Marek Behun wrote:
>>>> On Mon, 8 Feb 2021 20:11:06 +0000
>>>> Chris Packham <Chris.Packham at alliedtelesis.co.nz> wrote:
>>>>   
>>>>> Hi Marek,
>>>>>
>>>>> Do you have this in a repo I can pull from? I've got a couple of boards
>>>>> I can give this a spin on.
>>>> https://gitlab.nic.cz/turris/turris-omnia-uboot/
>>>> branch v2021.04-rc-mv-ddr-14.0.0
>>>>
>>>> also please test branch v2021.04-rc-mv-ddr-14.0.0-samsung-ddr-fix, that
>>>> one contains one more commit that is needed for Omnia with Samsung DDR
>>>> chips.
>>> I've tested the dm-88f6820-amc board. Training completed without
>>> issue, as does memtester running from Linux.
>>>
>>> Hit a bit of a snag on the x530 because the changes pushed it over the
>>> SPL size (it was already pretty close). I'll look to see if there's
>>> anything I can drop out or maybe bump the SPL size (I never did get a
>>> clear answer from Marvell as to what the size limit actually is).
>> I can temporarily work around the size issue by disabling watchdog
>> support in SPL (I really don't want that to be the long term solution).
>>
>> But then I encounter an odd problem. When I "reset" the board gets
>> through the DDR training but never makes it to u-boot proper, but if I
>> power cycle it boots through to the u-boot prompt. This doesn't happen
>> on the db-88f6820-amc board. One difference between the x530 and the amc
>> board is that the x530 has ECC so maybe something is going into the
>> weeds if ECC has already been enabled by a previous boot.
>>
> Could you bisect which commit causes this?
Seems to be the last one (ddr: marvell: a38x: fix SPLIT_OUT_MIX state 
decision) not entirely sure what the problem is. So I guess you can 
consider the upstream update good, the fix SPLIT_OUT_MIX not so much it 
happens to be the thing that causes the issue and the straw that tips 
the build size over the limit.