[PATCH V3 RFC 00/16] EFI stub for ARM

[Mark Salter]

On Mon, 2013-08-12 at 18:13 -0700, Roy Franz wrote:
> On Mon, Aug 12, 2013 at 7:02 AM, Mark Salter <msalter at redhat.com> wrote:
> > On Fri, 2013-08-09 at 16:26 -0700, Roy Franz wrote:
> >> * Change FDT memory allocation to retry with a larger allocation if
> >>   first educated guess is inadequate.
> >
> > With this change, it looks like you no longer free the original cmdline
> > and fdt memory. The current flow looks like:
> >
> >   retry:
> >      allocate_memory_for_expanded_fdt
> >      get_memory_map
> >      if (update_fdt() fails) {
> >         free new_fdt and memory_map
> >         goto retry
> >      }
> >
> > So, this keeps the original fdt around and uses it as a starting point
> > for newly allocated expanded fdt. You don't know if the new fdt is big
> > enough until update_fdt() succeeds. But at that point, you already wrote
> > the efi-runtime-mmap property with the memory_map still having the
> > original cmdline and fdt in it.
> >
> > I think you should be able to have an expand_fdt() function which bumps
> > the fdt size and uses the current fdt as the starting point instead of
> > the original fdt. That way you can free the original fdt on the first
> > iteration and free the original cmdline as soon as it is successfully
> > written. Then the last thing you do if get the memory_map and write it.
> >
> > --Mark
> 
> Hi Mark,
> 
> I think this will work with the current FDT fields that are being set
> by the stub.  In earlier
> versions, I was also updating the reserved memory map using
> fdt_add_mem_rsv(), so
> iteratively updating the device tree wouldn't work. The reserved
> regions would change,
> and so the repeated updates would cause there to be repeated and
> incorrect reserved regions.
> I'm inclined to leave it as is, which should correctly update the
> device tree even if methods like
> fdt_add_mem_rsv() are used, with the tradeoff being there will be a
> few more memory regions
> for the kernel to free when it processes the EFI memory map.  The
> kernel already needs to process
> the EFI memory map to free the buffers use to load the kernel and
> initrd, so these buffers will get freed, just not
> by the stub.
> 

Got it. Thanks for the explanation.