[U-Boot] re-use dm data for pre-relocate and post-relocate

[Simon Glass]

Hi Kever,

On Tue, 21 Nov 2017 at 03:24, Kever Yang <kever.yang at rock-chips.com> wrote:
>
> Hi Simon,
>

There are a lot of points in here, so I will spread my comments
throughout your email.

>      In order to make people easier to maintain a U-Boot for product,
>
> we want to re-use the dtb from kernel. Some peripheral like power/pmic
>
> and display panel, key, charger which need work in U-Boot and may different
>
> in different product, we want to make people only maintain kernel dts,
> but not
>
> one more U-Boot project for all these devices.

That sounds like a good idea. The U-Boot and kernel .dts should be the
same. But in some cases we may need changes made to the kernel version
to support U-Boot. Unfortunately this has proven quite tricky and I'm
not aware of any U-Boot change that has been accepted. Admittedly I
have not tried for a number of years, so perhaps the reviewers are
more amenable now.

>
> I want to do this just around the
>
> relocate, need to read dtb from storage like emmc/nand/spi, and then use
>
> a new dtb for initr_dm after relocate.

This seems OK, but why? Do you mean that U-Boot should read the
kernel's DTB and use that after relocation? The problem is that it
needs to be loaded from somewhere, and you need the DT for U-Boot to
work. So you add a pre-reloc DT with just enough drivers for loading
the post-reloc DT. But why bother? You may as well just load the full
thing.

>
>      Here are issues I have met:
>
> - The kernel dts may not have a stdout in chosen, we may use earlycon,
> earlyprintk instead;

I suppose U-Boot could support this, but IMO it would be better to
upstream the change for stdout.

>
>      this lead to U-Boot can not find a serial driver after relocate;
>
> - no alias for mmc device which is need by  U-Boot but not used in kernel;

Again, we could upstream a change.

>
> - not able to init mmc driver twice with current driver;
>
>      mmc driver not able to work before relocate, no bss for (static int
> initialized), and hang even

Well you cannot access BSS before relocation. If you use DM then it
should not need to access BSS. You should be able to use the
uclass-private data instead.

>
>      if I skip this;
>
>      mmc driver can init and read dtb before initr_dm, but not able to
> init again after initr_dm.

Why not? That seems like a bug.

>
>
> In fact, I don't think we have to re-init serial driver and mmc driver,
> and most other driver which
>
> need pre-relocate like clock, the only driver I have seem have different
> use in pre-relocate and post-relocate

Yes that's true, but the idea has never been implemented.

>
> is video driver(no idea why it have to be pre-relocate, we always have
> display very late),

The video driver is only inited before relocation to find out the
amount of memory it needs to be allocated. It does not actually probe
it fully.

>
> can we add a tag for re-use the dm driver data for those with pre-relocate ?

Yes I think it could work like this:

- initr_dm() remembers the old dm_root pointer
- post-reloc drivers can request access to the priv/platdata/uc_priv
from before relocation, e.g. using:

int dm_get_pre_reloc_dev(struct udevice *dev, struct udevice *pre_reloc_devp);

this returns the pre-reloc device associated with the post-reloc
device 'dev', if available.

Regards,
Simon