On Thu, 2011-07-14 at 20:53 -0600, Grant Likely wrote:

>You should have everything you need to fix it.  If
> CONFIG_SYS_BOOTMAPSZ is defined, then U-Boot will not use memory
> larger that that for the dtb or atags.
>
> Right now CONFIG_SYS_BOOTMAPSZ is not set by default, but we could
> default it to a sane value for ARM platforms.

Which makes more sense, setting this or  Scott's suggestion of reserving highmem to prevent it's use as is done in the powerpc's arch_lmb_reserve()?  The latter would affect all ARM.  Wouldn't BOOTMAP need to be set in each boards config file?



On Thu, 2011-07-14 at 16:51 -0500, Scott Wood wrote:
Well, that does sound strange.  I'd think it would be based on whether you
define CONFIG_SYS_BOOT_RAMDISK_HIGH, and whether initrd_high is set to
0xffffffff.

Same kernel, same u-boot, and initrd_high not set at all.  This is the crux of the problem. The default changes when any FDT is provided.  I think the difference may come from bootm_linux_fdt().

Rereading your earlier mail and looking at the code, we should probably  do as you suggest and add logic to arch_lmb_reserve() to reserve highmem like powerpc does.  I think the fdt_high patch is still a good idea though.  It allows us to continue booting using lowest memory for the initrd and other startup data.

Or were you just not telling bootm about the ramdisk before, and letting
the kernel know about the address through some other means?

We always provide a ramdisk address to bootm.  The problem occurs when we add an FDT address.

So add a mechanism for the user to override if you can justify a use for
it, but the default should be allocated from an lmb that has had unusable
addresses excluded.

Well, I do find it troubling I'm having to go to so much trouble to convince u-boot to continue to use this data directly from where we put it in the first place.

> > The user specified address might be in flash.
> 
> But in our case it is not. 

It still needs to be supported.

I can appreciate that, but the default behavior has changed.   It's surprising.  For the moment it breaks things (granted that it breaks only if you use the  fdt feature).

-dl