[U-Boot] [RFC] ARM: Start using fdt_high for relocation

[Stephen Warren]

On 12/06/2013 01:31 PM, Tom Rini wrote:
> On Fri, Dec 06, 2013 at 08:28:04PM +0100, Albert ARIBAUD wrote:
>> Hi Tom,
>>
>> On Fri, 6 Dec 2013 12:18:13 -0500, Tom Rini <trini at ti.com> wrote:
>>
>>> On Fri, Dec 06, 2013 at 05:59:37PM +0100, Albert ARIBAUD wrote:
>>>> Hi Tom,
>>>>
>>>> On Fri, 6 Dec 2013 10:48:50 -0500, Tom Rini <trini at ti.com> wrote:
>>>>
>>>>> Hey all,
>>>>>
>>>>> I've been thinking.  We've had a thread on i.MX platforms about fdt
>>>>> being overwritten and needing to be moved to another address.  And I've
>>>>> also had an internal problem about fdt being overwritten.  So, how about
>>>>> as a rule of thumb we start setting fdt_high (in configs) to
>>>>> memory-start + 512MiB, as that's the lowmem limit we should always have
>>>>> available.  This will fix the problem of BSS overwriting the DT, which
>>>>> is the problem we won't catch in normal bootm/bootz usage.  Thoughts?
>>>>
>>>> Not sure I'm getting the issue clear, and I would like to avoid (me and
>>>> others) having to switch back and forth between threads. Can you sketch
>>>> the failure scenario in a couple of lines?
>>>
>>> Sure.  Lets take am335x_evm builds (so Beaglebone Black/White, etc).
>>> If you start enabling all of the tracing options in the kernel (function
>>> tracing, graphs, etc), you get an uncompressed kernel and BSS that will
>>> use up the first ~16MiB of DDR.  We default to placing the DT at about
>>> 15MiB into memory.  So the kernel runs, clears BSS and eats the DT.
>>> System now hangs, and depending on debug options set you may or may not
>>> see anything at all from the kernel.  U-Boot couldn't detect this
>>> failure because we don't know how big the kernel BSS is, only how big
>>> the zImage is (and where it is) and how big the fdt is and where it is.
>>> No overlaps, go ahead and run.
>>
>> Thanks.
>>
>> The only issue I have with the RFC is that the +512 MiB value will only
>> work with targets which have more than 512 MiB DDR, right? But since
>> you're suggesting this should be set in configs, you are only suggesting
>> +512 MiB, and any target could actually specify a lower value as long as
>> it's greater than or eqal to 16 MiB. Correct?
> 
> fdt_high is only an upper bound on what we may relocate to (setting
> aside the magic value of 0xffffffff which means no relocation).  I just
> confirmed this too:
> U-Boot# bdi
> ...
> boot_params = 0x80000100
> DRAM bank   = 0x00000000
> -> start    = 0x80000000
> -> size     = 0x40000000
> ...
> U-Boot# setenv fdt_high 0xe0000000
> ...
> U-Boot# bootz $loadaddr - $fdtaddr
> Kernel image @ 0x80200000 [ 0x000000 - 0x3e5068 ]
> ## Flattened Device Tree blob at 80f80000
>    Booting using the fdt blob at 0x80f80000
>    Loading Device Tree to bf62a000, end bf630d9a ... OK
> 
> Of course, 0xbf62a000 is a bad address in that it won't be seen by the
> kernel, but it was in the higher-than-we-have-memory-at limit I set.

I haven't really followed this thread since I just noticed it
tangentially, but on Tegra...

Since commit 7f1b767aea94 "ARM: tegra: define CONFIG_SYS_BOOTMAPSZ"
all Tegra boards define BOOTMAPSZ to influence where the DTB gets
relocated. I wonder what's the benefit of using BOOTMAPSZ vs. defining
a default value for $fdt_high?

And couldn't you just move $fdt_addr_r high enough in RAM so even if
the DTB wasn't relocated it'd never overlap BSS? Tegra's
$kernel_addr_r and $fdt_addr_r are likely set up that way already,
although I didn't check.