[PATCH 7/8] spl: x86: Avoid starting up PCI automatically in SPL

[Simon Glass]

Hi Bin,

On Mon, 4 Sept 2023 at 03:20, Bin Meng <bmeng.cn at gmail.com> wrote:
>
> Hi Simon,
>
> On Thu, Aug 24, 2023 at 10:11 AM Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
> >
> >
> >
> > Am 24. August 2023 01:57:55 MESZ schrieb Simon Glass <sjg at chromium.org>:
> > >Hi Heinrich,
> > >
> > >On Wed, 23 Aug 2023 at 14:19, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
> > >>
> > >> On 8/23/23 20:47, Simon Glass wrote:
> > >> > For x86 platforms, PCI is core to their operation and is managed in
> > >> > arch-specific code. Each platform has its own way of doing this. For TPL
> > >> > and some SPL implementations, the full driver model PCI is not used.
> > >> >
> > >> > A recent change enabled full PCI in TPL/SPL for all boards. This breaks
> > >> > some x86 boards, so adjust it to skip that for x86.
> > >>
> > >> Could you, please, give some more detail?
> > >>
> > >> * Which boards are broken?
> > >
> > >For example, chromebook_samus and chromebook_samus_tpl
> > >
> > >> * Don't these boards have a pci_init() function?
> > >
> > >Yes, the same one you are calling.
> > >
> > >> * In what way does calling pci_init() lead to a failure?
> > >
> > >It probes and sets up PCI devices and uses a lot of pre-alloc RAM.
> > >
> > >>
> > >> It would be preferable to have all architectures and boards use the same
> > >> high level API. Excluding x86 here looks more like a (necessary) hack
> > >> than like a target state.
> > >
> > >Fair enough, but on x86 we access PCI long before driver model is up.
> > >Generally we don't fully enumerate it in SPL as it is expensive. It is
> > >also pointless, since U-Boot proper does it again later.
> > >
> > >Regards,
> > >Simon
> > >
> >
> > Acked-by: Heinrich Schuchardt <xypron.glpk at gmx.de>
> >
>
> To address Heinrich's concern, how about you turn off CONFIG_SPL_PCI
> in chromebook_samus and chromebook_samus_tpl?

I dug into this a bit more...the root cause is that we are turning off
CAR so the driver model tables are no-longer available. Accessing them
causes a hang.

I will send a v2 which adds a flag to indicate this.

Regards,
Simon