[PATCH v5 05/11] spl: Convert mmc to spl_load

[Jonas Karlman]

On 2023-08-03 18:11, Sean Anderson wrote:
> On 8/3/23 04:31, Xavier Drudis Ferran wrote:
>> El Mon, Jul 31, 2023 at 06:42:57PM -0400, Sean Anderson deia:
>>> This converts the mmc loader to spl_load. Legacy images are handled by
>>> spl_load (via spl_parse_image_header), so mmc_load_legacy can be
>>> omitted.
>>>
>>
>> Yes. I haven't used the legacy case, but by looking at the code, it
>> seems to me that mmc_load_legacy left the image at some mapped memory
>> at [ spl_image->load_addr,   spl_image->load_addr + size )
>> and the new code does too, but when the image is not aligned, the
>> memory that gets written to
>> was at [ spl_image->load_addr,   spl_image->load_addr + size + spl_image->offset % mmc->read_bl_len )
>> and it will
>> be at [ spl_image->load_addr - spl_image->offset % mmc->read_bl_len,   spl_image->load_addr + size )
>> after this patch.
>>
>> Meaning both with or without this patch some memory outside the image
>> gets written when the image is not aligned on media, but before it was
>> some part of a block at the end and now is that part before the
>> beginning.
>>
>> Whether that's better or worse I don't know. It depends on whether
>> it's a problem to write in non mapped memory, whether there's
>> something worth preserving there, and whether some SOC has some sort
>> of special behaving memory just there, like it happened with the issue
>> Jerome Forissier found (note in this case it wasn't legacy, it was
>> simple fit)
> 
> Fundamentally, we can't really deal with unaligned images without a
> bounce-buffer. The method used by SPL_LOAD_FIT_IMAGE_BUFFER_SIZE will
> continue working, since we call into the FIT routines to load the image.
> I would like to defer bounce buffering for other images until someone
> actually needs it.
> 
> Note that in the FIT case, you can also do `mkimage -EB`, at least if
> you aren't using FIT_LOAD_FULL.

With the following two commits introduced in v2023.04-rc1, the alignment
issue for Rockchip has been fixed and all external data is now accessed
block aligned.

9b2fd2d22852 ("binman: Add support for align argument to mkimage tool")
5ad03fc77dfa ("rockchip: Align FIT image data to SD/MMC block length")

Regards,
Jonas

> 
> --Sean
> 
>> https://patchwork.ozlabs.org/project/uboot/patch/c941d835a85255ff81a21c72069c3a9fc9a8a255.1656489154.git.jerome.forissier@linaro.org/
>>                                                                   
>>> Signed-off-by: Sean Anderson <sean.anderson at seco.com>
>>> ---
>>>
>>> Changes in v5:
>>> - Rework to load header in spl_load
>>>
>>>  common/spl/spl_mmc.c | 91 ++++----------------------------------------
>>>  1 file changed, 8 insertions(+), 83 deletions(-)
>>>

[...]