On Mon, Jan 31, 2011 at 3:27 PM, Nicolas Pitre [off-list ref] wrote:

On Mon, 31 Jan 2011, Grant Likely wrote:

quoted

If a dtb is passed to the kernel then the kernel needs to iterate
through compiled-in mdescs looking for one that matches and move the
dtb data to a safe location before it gets accidentally overwritten by
the kernel.

This patch creates a new function, setup_machine_fdt() which is
analogous to the setup_machine_atags() created in the previous patch.
It does all the early setup needed to use a device tree machine
description.  It also creates arm_unflatten_device_tree() which copies
the dtb into an allocated buffer and unflattens it into the live-tree
representation.

v2: Changed to save the dtb by copying into an allocated buffer.
    - Since the dtb will very likely be passed in the first 16k of ram
      where the interrupt vectors live, memblock_reserve() is
      insufficient to protect the dtb data.

This is wrong.  The vector page can be allocated anywhere.  It is
currently allocated with memblock_alloc(), so if you memblock_reserve()
the dtb soon enough then it should be safe and the vector page or
whatever will be allocated somewhere else.

Ah, thanks.  Did not know that.

What was the actual problem?

I was seeing the first page in memory getting zeroed partway through
the boot process which wiped out the dtb.  I haven't yet narrowed down
to the exact point that it gets overwritten.  It is late enough that
all the early setup dt code works, but before init_machine gets
called.

I'm also seeing a problem with /proc/device-tree getting corrupted on
qemu which I cannot reproduce on Tegra hardware.  Still investigating
this one.

g.