phytec-som-imx6: DCD setup: PLL reduction and video core settings

[Stefan Christ]

Hi Andreas,

thanks a lot for your detailed write-up. Indeed these register settings are
confusing. 

> What concerns me more is the CPU core clock reduction:
>     // CCM_ANALOG_PLL_ARM0 : PLL setup => PLL locked, PLL on,
>     //                       bypass off, 396MHz
>     // reset default: 0x00013042 => PLL not locked, PLL off, bypass on,
>     //                              bypass 24MHz osc as src, pll divider
>     //                              for 792MHz
>     // 0x80002021 => PLL locked, PLL on, bypass off, pll divider for
>     //               396MHz ; Attn.: PLL locked is w/o ; divider for
>     //               396MHz is below valid range
>     wm 32 0x020c8000 0x80002021
> Comments again added by me.
> 
> Two things here look questionable to me:
> 1.) Datasheet says valid divider values are 54-108, so the
>     default 0x42=66 is in range, whereas the 0x21=33 is not.
> 2.) Even if the divider value is legit, the change flow might be not.
>     A PLL normally takes some time to lock.

We have tracked down the origin of this DCD entry internally. It originated
from a PMIC workaround for the phyFLEX-i.MX6 in an older barebox version which
isn't mainline. It had a comment attached to it, but that was dropped while
bringing the phyFLEX-i.MX6 mainline. We will take a closer look to that issue
next week, whether the fix is still needed or can be resolved differently.

Mit freundlichen Grüßen / Kind regards,
	Stefan Christ

On Fri, Feb 19, 2016 at 08:28:25PM +0100, Andreas Pretzsch wrote:
> While looking into a maybe boarderline RAM setup on one specific piece
> of hardware, I took a closer look at the i.MX6 DDR setup of the Phytec
> i.MX6 modules in barebox.
> And came across two things that look ... somewhat questionable, maybe
> worth a look.
> 
> In the very first setup (the DCD) in
>     arch/arm/boards/phytec-som-imx6/flash-header-phytec-pfla02.h
> there is a block
>     wm 32 0x020e0010 0xf00000ff
>     wm 32 0x020e0018 0x007F007F
>     wm 32 0x020e001c 0x007F007F
>     wm 32 0x020c8000 0x80002021
> at the end.
> 
> 
> The first three (comments about complete register meanings added by me)
>     // IOMUX_GPR4 : VDOA, PCIe, VPU, IPU cache setup ; stop acknowledge
>     wm 32 0x020e0010 0xf00000ff
>     // IOMUX_GPR6 : IPU1 QoS setup
>     wm 32 0x020e0018 0x007F007F
>     // IOMUX_GPR7 : IPU2 QoS setup
>     wm 32 0x020e001c 0x007F007F
> do setup some cache attributes of VDOA, IPU, VPU (some of the video
> cores) and IPU QoS. Beside accessing two reserved bits in GPR4, they
> maybe do not harm, and can be found also in various other board files.
> Might be some case of copy-from-copy, no idea.
> Not sure if these really should be part of lowest-level memory setup...
> Maybe - if at all - in some board-specific setup, given barebox drives a
> display.
> Also, a quick google search shows the possible origin, and that it has
> been removed in some U-Boot branch, leaving it to the Linux kernel:
> https://github.com/linksprite/u-boot-acadia1.0-beta/blob/master/u-boot-acadia1.0-beta/patches/0546-ENGR00215515-MX6-Move-IPU-QoS-and-VDOA-IPU-VPU-AXI-C.patch
> So maybe more of a cleanup issue.
> Not only in the Phytec DCDs, but maybe also the other ones.
> 
> 
> What concerns me more is the CPU core clock reduction:
>     // CCM_ANALOG_PLL_ARM0 : PLL setup => PLL locked, PLL on,
>     //                       bypass off, 396MHz
>     // reset default: 0x00013042 => PLL not locked, PLL off, bypass on,
>     //                              bypass 24MHz osc as src, pll divider
>     //                              for 792MHz
>     // 0x80002021 => PLL locked, PLL on, bypass off, pll divider for
>     //               396MHz ; Attn.: PLL locked is w/o ; divider for
>     //               396MHz is below valid range
>     wm 32 0x020c8000 0x80002021
> Comments again added by me.
> 
> Two things here look questionable to me:
> 1.) Datasheet says valid divider values are 54-108, so the
>     default 0x42=66 is in range, whereas the 0x21=33 is not.
> 2.) Even if the divider value is legit, the change flow might be not.
>     A PLL normally takes some time to lock.
> 
> Now, I have not looked yet at the PLL setup and/or scaling code of the
> kernel, both wrt PLL change flow as well as dividers. So this might be
> ok, and even if only implicitly (like due to some implicit delay after
> DCD got parsed).
> And obviously it works. "Only" misses a comment whats it does. And only
> present in some of the Phytec DCDs, as far as I can see.
> 
> But what I like to understand is why ?
> I mean, scaling down the i.MX6 to its minimum frequency (half of
> default) in the bootloader, to get a maximum stable state, ok. Even if
> it extends the bootup time.
> 
> But why after all the IOMUX, MMDC and DDR setup is through ?
> Probably too late if it is about some PMIC-reset-voltage-scaling issue.
> Or some thermal precaution ?
> 
> By the way, the above setups goes back through long history, as in
>     a540c30 boards: Add phytec-som-imx6
>     c5b4f09 ARM:phyFLEX-iMX6 New Ram Timings for Q/DL
>     30f29e3 ARM: Add Phytec phyFLEX-i.MX6 board support
> and applies to other Phytec i.MX6 boards as well, beside the DL ones.
> 
> 
> Any comments on this, maybe also from Phytec ?
> 
> Best regards,
>   Andreas Pretzsch
> 
> -- 
> 
> carpe noctem engineering
> Ingenieurbuero fuer Hard- & Software-Entwicklung Andreas Pretzsch
> Dipl.-Ing. (FH) Andreas Pretzsch        Tel. +49-(0)7307-936088-1
> Lange Strasse 28a                       Fax: +49-(0)7307-936088-9
> 89250 Senden, Germany                   email: apr at cn-eng.de
>