Re: [PATCH 1/9] dt: Handle passed/built-in DT selection in early_init_dt_scan()

[Grant Likely <grant.likely@secretlab.ca>]

On Thu, 21 Nov 2013 14:42:27 +0100, Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> On Thu, Nov 21, 2013 at 1:21 PM, Grant Likely <grant.likely@secretlab.ca> wrote:
> >>  bool __init early_init_dt_scan(void *params)
> >>  {
> >> -     if (!params)
> >> -             return false;
> >> -
> >>       /* Setup flat device-tree pointer */
> >>       initial_boot_params = params;
> >>
> >> -     /* check device tree validity */
> >> -     if (be32_to_cpu(initial_boot_params->magic) != OF_DT_HEADER) {
> >> -             initial_boot_params = NULL;
> >> -             return false;
> >> +     /* check passed device tree validity */
> >> +     if (initial_boot_params &&
> >> +         be32_to_cpu(initial_boot_params->magic) == OF_DT_HEADER) {
> >> +             pr_info("FDT at %p\n", initial_boot_params);
> >> +             goto found;
> >> +     }
> >> +
> >> +     /* check built-in device tree validity */
> >> +     initial_boot_params = &__dtb_start;
> >> +     if (__dtb_end != (void *)&__dtb_start &&
> >> +         be32_to_cpu(initial_boot_params->magic) == OF_DT_HEADER) {
> >> +             pr_info("Compiled-in FDT at %p\n", initial_boot_params);
> >> +             goto found;
> >
> > __dtb_start is in an init section which is discarded after initcalls.
> > For this to work, the dtb needs to be copied into a region of allocated
> 
> My changes don't change the current behavior much: currently
> early_init_dt_scan() is already called with &__dtb_start in several places.
> If this is broken, it's already broken.

Yes, but it is called on platforms that already make that assumption
that __dtb_start must be copied and therefore call
unflatten_and_copy_devicetree(). This change makes *all* platforms do
that. That breaks arm, arm64, c6x, microblaze, some mips platforms, and
powerpc!

> > memory. The dtb section can also potentially contain multiple .dtb
> > blobs. In the use case that you care about you are probably only
> 
> Multiple dtb blobs are currently handled in platform-specific code, which
> passes the right dtb to early_init_dt_scan().

The problem is that it makes the default dt completely random because
the generic code still deferrences __dtb_start.

> > thinking about one, but it is entirely possible for device drivers to
> > have a dtb linked in which may break this if it gets linked in a
> > different order. The specific example I'm thinking about is I want to
> > have the DT selftest code load an overlay to get testcase data from a
> > dtb blob.
> >
> > The other concern I have here is that I don't really want this to be the
> > default on a lot of platforms. ARM and PowerPC for instance should only
> > get the default dtb from the boot wrapper. It needs to be configurable
> > in some way.
> 
> On ARM and PowerPC, the section is empty, hence &__dbt_start ==
> &__dtb_end.

There is no guarantee that that will always be so. The patch makes some
poor assumtions, but they shouldn't be difficult to fix.

g.