On 08/04/2021 23:24, Dario Binacchi wrote:

quoted

Il 07/04/2021 15:21 Tero Kristo [off-list ref] ha scritto:

  
On 07/04/2021 15:52, Rob Herring wrote:

quoted

On Wed, Apr 7, 2021 at 2:07 AM Dario Binacchi [off-list ref] wrote:

quoted

quoted

Il 07/04/2021 03:16 Rob Herring [off-list ref] ha scritto:


On Tue, Apr 6, 2021 at 5:02 PM Dario Binacchi [off-list ref] wrote:

quoted

quoted

Il 06/04/2021 16:06 Rob Herring [off-list ref] ha scritto:


On Fri, Apr 2, 2021 at 2:21 PM Dario Binacchi [off-list ref] wrote:

quoted

The series comes from my commit in U-boot
d64b9cdcd4 ("fdt: translate address if #size-cells = <0>")
and from the subsequent exchange of emails at the end of which I was
suggested to send the patch to the linux kernel
(https://patchwork.ozlabs.org/project/uboot/patch/1614324949-61314-1-git-send-email-bmeng.cn@gmail.com/).

It's 'ranges' that determines translatable which is missing from the
DT. This should have not had a 0 size either though maybe we could
support that.

I have replied to the email you sent to the u-boot mailing list

quoted

Does the DT have to be updated anyways for your spread spectrum support?

The spread spectrum support patch does not need this patch to work. They belong
to two different series.

That's not what I asked. Is the spread spectrum support forcing a DT
update for users?

Yes, the deltam and modfreq registers must be added to the DPLL clocks.

That's a shame given this dts has been mostly untouched since 2013.

I think technically it would be possible to map these registers within
the driver also, seeing there are like a handful of the DPLLs for both
am3/am4 which are impacted. Just add a new compatible or something, or
alternatively parse the register addresses and populate the
deltam/modfreq registers based on that.

I have not added new compatibles, but I have added the offset of the delta and modfreq
registers to the data structures used by the DPLL drivers and I have set them in the
related setup functions.
https://lore.kernel.org/patchwork/patch/1406590/

True, I just said that technically it would be possible to add this data 
within the driver itself to avoid modifying DT if that would be preferred.

quoted

quoted

quoted

quoted

If the DT has to be changed anyways (not really
great policy), then you could fix this in the DT at the same time.

I could put the fix to the device tree in that series, although I wouldn't
create a single patch to fix and add the SSC registers. First the size-cells = <0>
fix patch and then the SSC patch.
Do you agree?

By at the same time, I really just meant within 1 release.

But I'd like to hear TI maintainers' thoughts on this.

I did post a comment on patch #1 questioning the approach from TI clock
driver perspective, imho I can't see why these two patches would be
needed right now.

Fix to above, it was patch #2 I was referring to.

Because U-boot maintainers asked me after I sent them my patch on this issue.
I believe that the email exchange that took place in the U-boot (https://patchwork.ozlabs.org/project/uboot/patch/1614324949-61314-1-git-send-email-bmeng.cn@gmail.com/)
and Linux kernel mailing lists showed that:
- The patch 'fdt: translate address if # size-cells = <0>' is wrong (U-boot has accepted
   it, and it will have to be reverted).
- However, the same patch highlighted that it is wrong to use the size-cells = <0> property
   in the prcm_clocks and scm_clocks nodes of device tree.
- Rob agrees that in the case of the am3xx this is the right choice:

Well, I don't quite like where this is ending at. Basically you are just 
modifying the am3/am4 DTs adding a size spec for every clock node. This 
leaves the omap3/omap4/omap5/dra7 in the old format. Should someone 
convert those also? Has anybody tested what happens with the u-boot 
change on those platforms? Or with the kernel change proposed for the TI 
clock driver?

Also, none of this changes the fact that imho patch #2 in this series is 
wrong and should be fixed. Doing ioremap for every clock node is at 
least costly (dra7xx has some 180 clock nodes based on quick grep) and 
also potentially breaks things (you get extremely fragmented iomappings, 
and some of them might even get rejected because you end up in the same 
4K page), and should be avoided.

If things would be fixed properly, we would get rid of the clock nodes 
from the DT completely and just leave the clock providers in place; 
clocks would be specified via something like 'clocks = <&prcm 
AM3_ADC_TSC_FCK>;' similar to what is done with the clkctrl entries, and 
rest of the clock data would be built-in to the clock driver. This would 
completely get rid of any future compatibility issues and the need to 
tweak the DT if some clock driver would need modifications to support 
some new feature.

-Tero