[PATCH 1/1] arm: imx8m: imx8mm-cl-iot-gate: Add support for detect memory size

[Marcel Ziswiler]

Hi Paul

On Tue, 2021-08-24 at 04:04 +0800, PaulLiu wrote:
> Hi Marcel,
> 
> Thanks for the comment.
> But isn't it more straightforward to use TCM data?

It was not quite clear to me what you were talking about. So I had a quick look at CompuLab's implementation of
all this. Let us hope it works fine but I would not exactly call this straightforward.

> I mean it is already tested when DRAM initializes. 
> Also there's already a commit that avoids U-boot using address more than 4GB on imx8m platform for some
> reasons I don't know in detail but the commit is e27bddff .

I don't think any of this would be needed if everybody would agree on using the generic get_ram_size() routine.
The world could be so much simpler!

> Because the base is 0x40000000. So 4GB address means it is about 3G left.
> But the largest memory we can purchase is 4GB. Which is already beyond the limit.
> And I don't have Compluab 4GB boards.

Yeah, I remember from the i.MX 8M Plus this actually requires 64-bit addressing to really work.

> So I think we should avoid directly testing the address more than 4GB addr?

Well, in the 32-bit world there is nothing beyond those 4GB (;-p).

> We just report the correct memory size here based on the DRAM init test. 
> And restrict the memory usage by board_get_usable_ram_top() by previous commit.

To be honest, I really think this is done totally the wrong way around, but then as it is not my board I won't
stand in anybody's way...

> The Compulab board is using various memory chips. They only let you to choose the size, not the
> company/brand. So the dram init code actually test every brand of the chips at beginning (SPL) so we have the
> correct brand and size stored in TCM area. I have 5 Compulab boards on my hand. All of them are 2GB ram but
> different brand. But it is still not covered all the combinations I think.

Yes, I saw that. Considering that NXP was unable to provide anything halfway generic this looks OKish.

> So, can we keep using the TCM data rather than do the actual testing?

Sure, as stated above, I won't stay in your way. Just thought it worth pitching a more generic way of going
about RAM sizing.

> Yours,
> Paul

Cheers

Marcel

> On Mon, Aug 23, 2021 at 11:14 PM Marcel Ziswiler <marcel.ziswiler at toradex.com> wrote:
> > Hi Paul
> > 
> > On Mon, 2021-08-23 at 10:49 +0800, Ying-Chun Liu wrote:
> > > From: "Ying-Chun Liu (PaulLiu)" <paulliu at debian.org>
> > > 
> > > When purchasing imx8mm-cl-iot-gate it is able to customize the
> > > memory size. It could be 1GB, 2GB and 4GB. We implement
> > > board_phys_sdram_size() to detect the memory size for usage.
> > > 
> > > Signed-off-by: Ying-Chun Liu (PaulLiu) <paulliu at debian.org>
> > > Cc: Fabio Estevam <festevam at denx.de>
> > > Cc: Frieder Schrempf <frieder.schrempf at kontron.de>
> > > Cc: uboot-imx <uboot-imx at nxp.com>
> > > ---
> > >  .../imx8mm-cl-iot-gate/imx8mm-cl-iot-gate.c   | 25 +++++++++++++++++++
> > >  include/configs/imx8mm-cl-iot-gate.h          |  2 +-
> > >  2 files changed, 26 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/board/compulab/imx8mm-cl-iot-gate/imx8mm-cl-iot-gate.c b/board/compulab/imx8mm-cl-iot-
> > gate/imx8mm-
> > > cl-iot-gate.c
> > > index eabcc842a4..01c6011b75 100644
> > > --- a/board/compulab/imx8mm-cl-iot-gate/imx8mm-cl-iot-gate.c
> > > +++ b/board/compulab/imx8mm-cl-iot-gate/imx8mm-cl-iot-gate.c
> > > @@ -14,8 +14,33 @@
> > >  #include <asm/arch/sys_proto.h>
> > >  #include <asm/io.h>
> > >  
> > > +#include "ddr/ddr.h"
> > > +
> > >  DECLARE_GLOBAL_DATA_PTR;
> > >  
> > > +int board_phys_sdram_size(phys_size_t *size)
> > > +{
> > > +       struct lpddr4_tcm_desc *lpddr4_tcm_desc =
> > > +               (struct lpddr4_tcm_desc *)TCM_DATA_CFG;
> > > +
> > > +       switch (lpddr4_tcm_desc->size) {
> > > +       case 4096:
> > > +       case 2048:
> > > +       case 1024:
> > > +               *size = (1L << 20) * lpddr4_tcm_desc->size;
> > > +               break;
> > > +       default:
> > > +               printf("%s: DRAM size %uM is not supported\n",
> > > +                      __func__,
> > > +                      lpddr4_tcm_desc->size);
> > > +               while (1)
> > > +                       ;
> > > +               break;
> > > +       };
> > 
> > Why not simply using generic get_ram_size() as we e.g. did here [1]?
> > 
> > > +
> > > +       return 0;
> > > +}
> > > +
> > >  static int setup_fec(void)
> > >  {
> > >         if (IS_ENABLED(CONFIG_FEC_MXC)) {
> > > diff --git a/include/configs/imx8mm-cl-iot-gate.h b/include/configs/imx8mm-cl-iot-gate.h
> > > index faeee2178c..1e835563d6 100644
> > > --- a/include/configs/imx8mm-cl-iot-gate.h
> > > +++ b/include/configs/imx8mm-cl-iot-gate.h
> > > @@ -158,7 +158,7 @@
> > >  
> > >  #define CONFIG_SYS_SDRAM_BASE          0x40000000
> > >  #define PHYS_SDRAM                     0x40000000
> > > -#define PHYS_SDRAM_SIZE                        0x80000000 /* 2GB DDR */
> > > +#define PHYS_SDRAM_SIZE                        0x40000000 /* 1GB DDR */
> > 
> > In our implementation we use PHYS_SDRAM_SIZE as the upper limit which get_ram_size() then actually tests
> > for.
> > 
> > >  #define CONFIG_MXC_UART_BASE           UART3_BASE_ADDR
> > 
> > [1] https://marc.info/?l=u-boot&m=160387919927218
> > 
> > Cheers
> > 
> > Marcel