Re: [PATCH 1/5] of: Add support for linking device tree blobs into vmlinux

[Dirk Brandewie <dirk.brandewie@gmail.com>]

On 11/16/2010 06:58 PM, Grant Likely wrote:
> On Tue, Nov 16, 2010 at 7:21 PM, Dirk Brandewie
> <dirk.brandewie@gmail.com>  wrote:
>>
>> On 11/16/2010 04:39 PM, David Daney wrote:
>>>
>>> Thanks for doing this. However I have a few comments...
>>>
>>> On 11/16/2010 02:41 PM, dirk.brandewie@gmail.com wrote:
>>>>
>>>> From: Dirk Brandewie<dirk.brandewie@gmail.com>
>>>>
>>>> /* .data section */
>>>> #define DATA_DATA \
>>>> *(.data) \
>>>> @@ -468,7 +482,8 @@
>>>> MCOUNT_REC() \
>>>> DEV_DISCARD(init.rodata) \
>>>> CPU_DISCARD(init.rodata) \
>>>> - MEM_DISCARD(init.rodata)
>>>> + MEM_DISCARD(init.rodata) \
>>>> + KERNEL_DTB()
>>>>
>>>
>>> I thought the init.rodata was only for data used by __init things.
>>> Although the
>>> current linker scripts do not put it in the section that gets recycled as
>>> usable
>>> memory.
>>>
>>> IIRC the unflattened version of the device tree has pointers to the
>>> flattened
>>> data. Since the device tree nodes are live for the entire kernel
>>> lifecycle,
>>> shouldn't the device tree blobs be in non-init memory?
>>>
>>
>> The contents of the blob get copied to allocated memory during
>> unflatten_device_tree() so the blob that is linked in is no longer needed
>> after init.
>
> Have you written a patch to add this behaviour?  The current code doesn't.  :-)
>

I misspoke, my blob gets copied to allocated memory during unflatten_device_tree.
my early_init_dt_alloc_memory_arch() returns the physical address of a kmalloc'd
buffer.

You would want copy the dtb that your platform is going to use to non-init memory.

--Dirk