On Fri, Apr 19, 2019 at 08:52:28AM +0000, Peng Fan wrote:
On Fri, 19 Apr 2019 11:56:25 +0530 Jagan Teki jagan@amarulasolutions.com wrote:
On Fri, Apr 5, 2019 at 2:20 AM Lukasz Majewski lukma@denx.de wrote:
Hi Jagan,
On Thu, Apr 4, 2019 at 3:31 PM Lukasz Majewski lukma@denx.de wrote:
On Thu, 4 Apr 2019 14:56:36 +0530 Jagan Teki jagan@amarulasolutions.com wrote:
> On Thu, Apr 4, 2019 at 2:31 PM Lukasz Majewski lukma@denx.de > wrote: > > > > On Tue, 2 Apr 2019 16:58:33 +0530 Jagan Teki > > jagan@amarulasolutions.com wrote: > > > > > This is revised version of previous i.MX6 clock management > > > [1]. > > > > > > The main difference between previous version is > > > - Group the i.MX6 ccm clocks into gates and tree instead > > > of handling the clocks in simple way using case statement. > > > - use gate clocks for enable/disable management. > > > - use tree clocks for get/set rate or parent traverse > > > management. > > > - parent clock handling via clock type. > > > - traverse the parent clock using recursive functionlaity. > > > > > > The main motive behind this tree framework is to make the > > > clock tree management simple and useful for U-Boot > > > requirements instead of garbing Linux clock management > > > code. > > > > > > We are trying to manage the Allwinner clocks with similar > > > kind, so having this would really help i.MX6 as well. > > > > > > Added simple names for clock macros, but will update it in > > > future version. > > > > > > I have skipped ENET clocks from previous series, will add > > > it in future patches. > > > > > > Changes for v2: > > > - changed framework patches. > > > - add support for imx6qdl and imx6ul boards > > > - add clock gates, tree. > > > > > > [1] https://patchwork.ozlabs.org/cover/950964/ > > > > > > Any inputs? > > > > Hmm.... It looks like we are doing some development in > > parallel. > > > > Please look into following commit [1]: > > https://patchwork.ozlabs.org/patch/1034051/ > > > > It ports from Linux 5.0 the CCF framework for iMX6Q, which > > IMHO in the long term is a better approach. > > The code is kept simple and resembles the code from Barebox. > > > > Please correct me if I'm wrong, but the code from your work > > is not modeling muxes, gates and other components from Linux > > CCF. > > The U-Boot implementation of CLK would require as minimal and > simple as possible due to requirement of U-Boot itself. Hope > you agree this point?
Now i.MX6 is using clock.c CLK implementation. If we decide to replace it - we shall do it in a way, which would allow us to follow Linux kernel. (the barebox implementation is a stripped CCF from Linux, the same is in patch [1]).
> if yes having CCF stack code to handle all clock with > respective separate drivers management is may not require as > of now, IMHO.
I do have a gut feeling, that we will end up with the need to have the CCF framework ported anyway. As for example imx7/8 can re-use muxes, gates code.
As per my experience the main the over-ahead to handle clocks in U-Boot if we go with separate clock drivers is for Video and Ethernet peripherals. these are key IP's which use more clocks from U-Boot point-of-view, others can be handle pretty straight-forward unless if they don't have too much tree chain.
On this series, the tree management is already supported ENET in i.MX6, and Allwinner platforms.
As of now, I'm thinking I can handle reset of the clocks with similar way.
But this code also supports ENET and ESDHCI clocks on i.MX6Q (as supporting those was the motivator for this work).
One important thing to be aware of - the problem with SPL's footprint. The implementation with clock.c is small and simple, but doesn't scale well.
However, those are only my "feelings" after a glimpse look - I will look into your code more thoroughly and provide feedback.
Please have a look, if possible check even the code size by adding USDHC clocks.
Yes, code size (especially in SPL) is an _important_ factor here.
> > This series is using recursive calls for handling parenting > stuff to handle get or set rates, which is fine for handling > clock tree management as far as U-Boot point-of-view. We have > faced similar situation as I explained in commit message about > Allwinner clocks [2] and we ended up going this way.
I'm not Allwinner expert - but if I may ask - how far away is this implementation from mainline Linux kernel?
How difficult is it to port the new code (or update it)?
Allwinner clocks also has similar gates, muxs, and with other platform stuff which has too much scope in Linux to use CCM.
For example the barebox managed to get subset of Linux CCF ported, without loosing the CCF similarity.
Important factors/requirements for the i.MX clock code:
Easy maintenance in long-term
Reusing the code in SPL (with a very important factor of
_code_size_).
Reuse the code for other i.MX SoCs (imx7, imx8)
Effort needed to use DM with this code
I understand your points, I was managed this series based on these requirements as well.
Ok.
Could you share the delta of footprint size (u-boot.img/SPL) with and without your patch (on imx6q) ?
In my case the CCF caused increase of u-boot.img proper (as it was not yet adapted to SPL):
415KiB -> 421KiB = 6KiB increase of size (< 2%).
(This can be further reduced by using OF_PLATDATA).
This CCF code hasn't been ported to SPL (yet)
We even consider the foot-print, atleast for recursive calls of handling parenting scale well.
With CCF porting v3 I'm going to provide some caching, so the descending would be done at most once.
May be we can consider to design based on this as per U-Boot.
Please look into point 1. Having code ported from Linux is IMHO better in the long term.
Agree.
I just want to re-iterate my support again here that we should be looking at adapting and stripping down frameworks from the kernel. They are: - Familiar to a large subset of our developers as most folks also work on the Linux kernel. - Already (generally) well designed to take into account the various designs of vastly different SoCs that we also want to support. - Likely to already be fed by device tree and we can just leverage what's already in the dts* files.