[U-Boot] DA850evm SPI Flash Partitions between Linux and U-boot are Inconsistent

[Sekhar Nori]

Hi Adam,

On Wednesday 30 August 2017 04:07 PM, Adam Ford wrote:
> On Wed, Aug 30, 2017 at 12:30 AM, Sekhar Nori <nsekhar at ti.com> wrote:
>> On Wednesday 30 August 2017 08:07 AM, Adam Ford wrote:
>>> On Tue, Aug 29, 2017 at 8:05 AM, Adam Ford <aford173 at gmail.com> wrote:
>>>> On Tue, Aug 29, 2017 at 6:41 AM, Sekhar Nori <nsekhar at ti.com> wrote:
>>>>> On Tuesday 29 August 2017 03:29 PM, Adam Ford wrote:
>>>>>> On Tue, Aug 29, 2017 at 4:13 AM, Sekhar Nori <nsekhar at ti.com> wrote:
>>>>>>> Hi Adam,
>>>>>>>
>>>>>>> On Sunday 27 August 2017 08:31 PM, Adam Ford wrote:
>>>>>>>> I was trying to enable MTD Partitions to make loading the Kernel and
>>>>>>>> FS easier from within U-Boot
>>>>>>>>
>>>>>>>> The da850evm spi-flash partitions in Linux show
>>>>>>>>
>>>>>>>> "U-Boot-SPL" @ offset 0, size 64K
>>>>>>>> "U-Boot"; @ offset 0x00010000, size 512K
>>>>>>>> "U-Boot-Env"; @ offset 0x00090000
>>>>>>>
>>>>>>> According to board/davinci/da8xxevm/README.da850, we support AIS image
>>>>>>> format for SPI boot. This is a single image containing SPL and U-Boot.
>>>>>>>
>>>>>>> Given this, I think having separate partitions for SPL and U-Boot does
>>>>>>> not make sense since thats not how its going to be used.
>>>>>>
>>>>>> That's what I was thinking too, but I wasn't sure if someone had split
>>>>>> SPL somehow to make it possible to update U-Boot via Linux or U-Boot
>>>>>> itself.
>>>>>>>
>>>>>>>>
>>>>>>>> However U-Boot shows the following:
>>>>>>>>
>>>>>>>> CONFIG_SYS_SPI_U_BOOT_OFFS 0x8000 which would make the SPL partition 32K.
>>>>>>>> CONFIG_SYS_SPI_U_BOOT_SIZE      0x40000 and the size of U-boot 256K
>>>>>>>> instead of 512k.
>>>>>>>>
>>>>>>>> CONFIG_ENV_SIZE                 (64 << 10)
>>>>>>>> CONFIG_ENV_OFFSET           (512 << 10) which is 0x80000 instead of
>>>>>>>> Linux's 0x90000
>>>>>>>>
>>>>>>>> It seems to me like the U-Boot and Linux should try and synchronize
>>>>>>>> their partitions tables.
>>>>>>>
>>>>>>> Right.
>>>>>>>
>>>>>>>> For people who want to burn Linux into the SPI flash it seems there
>>>>>>>> there should be some consistency.  I tried making the U-boot settings
>>>>>>>> match the Linux ones, but it seems to hang between SPL and U-Boot, so
>>>>>>>> I think the U-Boot offset in Linux might need to match U-Boot.  Can
>>>>>>>> you guys make some recomendations as to which is correct?
>>>>>>>
>>>>>>> I have not tried it, but looks like the partitions we need are
>>>>>>>
>>>>>>> "SPL/U-Boot AIS" @ offset 0, size 512K
>>>>>>> "U-Boot Environment" @ offset 512K, size 64K
>>>>>>
>>>>>> I was thinking the same thing.
>>>>>>
>>>>>>> "Kernel/Spare" @ offset 576K, size 7552K
>>>>>>> "Mac Address" @ offset 8128K, size 64K (read only)
>>>>>>
>>>>>> If U-Boot reads the MAC address from its environmental variables
>>>>>> instead of using up an entire 64K for what conceivably is 6 bytes,
>>>>>> could we free up this space to help grow the Kernel/Spare space?
>>>>>
>>>>> If environment sector is erased, the mac address needs to be restored
>>>>> from SPI flash. Thats why I think it needs to remain as a read-only
>>>>> partition. Also, its just 64K more space, not sure if it will make a big
>>>>> difference in the overall scheme of things.
>>>>>
>>>>>>> With an 8M flash, I think its futile to try to fit a modern filesystem
>>>>>>> on the flash.
>>>>>>>
>>>>>>> If you are using DT boot, we cannot really change the partitions in
>>>>>>> device-tree because of DT backward-compatibility requirements. But we
>>>>>>> can fix da850evm_spiflash_part[] table in
>>>>>>> arch/arm/mach-davinci/board-da850-evm.c.
>>>>>>
>>>>>> Wouldn't having two different partition tables between da850-evm.c and
>>>>>> the DT cause confusion down the road?  With DT being the future, why
>>>>>> would we not try to eliminate any custom board files with just a
>>>>>> single common davinci board file + board specific DT?
>>>>>
>>>>> Thats ideal yes. But there is a lot of support in the da850-evm board
>>>>> file which does not have a DT equivalent yet. Thats why I have kept the
>>>>> board file around.
>>>>>
>>>>>>> To help DT boot, we can pass fixed up mtdparts through environment
>>>>>>> variables. Something like what is done in 3f18ff07c81b ("ARM: keystone:
>>>>>>> Pass SPI MTD partition table via kernel command line").
>>>>>>
>>>>>> This was ultimately was I was wanting to do.  For some of the OMAP3
>>>>>> boards, removing the partition tree was accepted since the partition
>>>>>> info was passed to the kernel.  If we did this for DA850, we could
>>>>>> simply elimiate the partition info in the Kernel (like some of the
>>>>>> omap3 boards) and let the partition maps be defined by U-Boot would
>>>>>> would guarantee consistency between them.
>>>>>
>>>>> Actually, thinking a bit more, we cannot change the partition
>>>>> information in kernel at all since that means loss of data for anyone
>>>>> updating just the kernel.
>>>>>
>>>>> We can however, pass updated kernel command line arguments in U-Boot so
>>>>> anyone updating the bootloader and erasing the environment sector gets
>>>>> updated partition information. This is likely to be the case where
>>>>> entire flash is being re-written and loss of data is not a concern.
>>>>>
>>>> That makes sense to me.
>>>>
>>>> I'll send a U-Boot patch marked RFC later today with the MTD stuff
>>>> enabled in U-Boot and changes to the bootargs to pass the MTD
>>>> information to Linux.
>>>>
>>>
>>> I have a patch ready for review for U-Boot, but I had to make a few
>>> changes to the davinci_all_defconfig to get the MTD partitions to map
>>> correctly:
>>>
>>> -CONFIG_MTD=m
>>> +CONFIG_MTD=y
>>
>> Hmm. why the shift to built-in for MTD?
> 
> I was concerned about the order that the modules were being loaded.
> Having MTD set to Y is required if MTD_CMDLINE_PARTS is set to Y, and
> that guarantees that the CMDLINE_PARTS will load before the spi flash
> module. I wanted to make sure the partition table from U-Boot was
> passed instead the the default one in Linux and I didn't see any other
> way.  I have a rootfs using eudev to load kernel modules, and it
> doesn't always seem to load stuff in a consistent order.

I am not quite sure about this race. The order of parser invocation is
controlled by default_mtd_part_types[] which lists cmdline parser before
ofpart. In any case, if there is a race, it needs to be discussed on the
mtd list. The hack to have a parser built-in to workaround this should
not be needed.

Thanks,
Sekhar