Re: [RESEND PATCH 2/5] mtd: rawnand: add NVIDIA Tegra NAND Flash controller driver

[Stefan Agner <stefan@agner.ch>]

On 31.05.2018 22:30, Boris Brezillon wrote:
> On Thu, 31 May 2018 19:54:08 +0200
> Stefan Agner <stefan@agner.ch> wrote:
> 
>> >> +
>> >> +	mtd->dev.parent = &pdev->dev;
>> >> +	mtd->name = "tegra_nand";
>> >
>> > I just figured it was undocumented (yet) but you could have a label
>> > string property in your nand DT node that tells you the name of the
>> > MTD device instead of something too generic like tegra_nand.
>> >
>>
>> Using label in the NAND chip subnode actually causes current U-Boot to
>> delete (!!) the chip node and create partitions on the controller node.
>>
>> See:
>> https://elixir.bootlin.com/u-boot/latest/source/common/fdt_support.c#L757
>>
>> The code essentially uses the property label to detect whether its a
>> NAND chip or a partition...
> 
> Why not fixing that in uboot? The representation where the NAND device
> and NAND controller are mixed in a single node called nand@xxx is just
> wrong from a HW PoV, and it seems uboot is using this representation,
> which is probably why you have a problem when trying to find the
> partition directly under the NAND controller node.
> 
>>
>> At least this is the case when using fdt_fixup_mtdparts and passing the
>> controller compatible ("nvidia,tegra20-nand") in node_info,
> 
> Just a digression, but I recommend using
> "nvidia,tegra20-nand-controller" for the compatible, because the node
> is describing the NAND controller not the NAND chip. 
> 

Ok.

>> what our
>> downstream U-Boot is currently doing. Maybe we should pass the
>> compatible property of the NAND chip?
> 
> Or maybe you should search for partitions in children of the controller
> node instead of searching directly under the controller node itself.
> 

Yes, that is what it is doing... But only if that child has not a label.

fdt_fixup_mtdparts is common code. Change the behaviour now probably
breaks boards... 

Anyway, this discussion needs to be shifted to the U-Boot mailing list.

>> But afaik, chips do not have a
>> compatible necessarily.
> 
> Nope, and it should stay like that.
> 
>>
>> So using label in the chip node is currently a no-go for me.
> 
> I hope I'm wrong but I fear this is not the only problem you'll face
> when switching to a controller+chip representation. This is just the
> tip of the iceberg.
> 

Works not too bad otherwise, so far :-)

>>
>> Will send out v3 soon.
> 
> Sure, let's see how v3 looks.

--
Stefan