firmwareload fails for cortex-m7 on imx8mp

[Ahmad Fatoum]

Hello Stefano,

On 15.11.24 15:47, Stefano Manni wrote:
> Hello Ahmad,
> 
> patch [1] works fine when M7 firmware is compiled to run from ITCM. Thank you!

Thanks for testing.

> But when I use exactly the same firmware but compiled to run from DDR
> (see [2]) I got:
> 
> barebox at NXP i.MX8MPlus EVK board:/ firmwareload hello-ddr.elf
> remoteproc0: powering up remoteproc-cm7.of
> ERROR: remoteproc0: bad phdr da 0x80000000 mem 0x3de4
> ERROR: remoteproc0: Failed to load program segments: -22
> 
> Note that I can run it both from linux and u-boot.
> 
> [1] https://lore.barebox.org/barebox/20241115135242.1251691-1-a.fatoum@pengutronix.de/
> [2] https://docs.zephyrproject.org/latest/boards/nxp/imx8mp_evk/doc/index.html#programming-and-debugging-m7

Can you send me the ELF file, so I can look at the headers myself?

Cheers,
Ahmad

> 
> Il giorno ven 15 nov 2024 alle ore 15:02 Ahmad Fatoum
> <a.fatoum at pengutronix.de> ha scritto:
>>
>> Hello Stefano,
>>
>> On 15.11.24 14:43, Stefano Manni wrote:
>>> Dear all,
>>>
>>> I cannot run an ELF image on the M7 core on the imx8mp soc.
>>> The ELF comes from zephyr and it runs as expected when I load it from linux,
>>> but in barebox I encounter this error:
>>>
>>> barebox at NXP i.MX8MPlus EVK board:/ tftp zephyr.elf
>>>         [################################################################]
>>> barebox at NXP i.MX8MPlus EVK board:/ firmwareload zephyr.elf
>>> remoteproc0: powering up remoteproc-cm7.of
>>> DABT (current EL) exception (ESR 0x96000061) at 0x00000000007e0004
>>
>> https://esr.arm64.dev/#0x96000061 decodes this as
>> "Abort caused by writing to memory" (Alignment fault) with valid FAR.
>>
>> FAR is the addresss listed here (0x00000000007e0004), which is indeed
>> not divisible by 8.
>>
>>> [<ffd826f4>] (__arch_memcpy+0x48/0x13c) from [<ffd82a00>] (memcpy+0xc/0x14)
>>
>> The optimized memcpy used here expects misalignment not to trap.
>>
>>> As you can see the exception happens when rproc_elf_load_segments() tries to
>>> memcpy a segment where the M7 expects it. The ELF has been compiled to run
>>> from ITCM memory (0x007E0000-0x007FFFFF).
>>>
>>> Do you have any idea? Maybe may I have to put a reserved-memory for ITCM?
>>
>> I think I may have broken this when I changed barebox to map everything outside
>> main memory as uncached.
>>
>> I did this, because we 1) don't want to risk the main CPU speculating into I/O
>> memory and 2) don't want to have to use memory barriers for synchronization.
>>
>> The flipside is that drivers that access I/O memory must do this via
>> readl/writel and friends or via memcpy_io/memset_io. This was already required
>> before, but some drivers that didn't do this were lucky enough to fly under
>> the radar. At least for remoteproc, this seems no longer the case.
>>
>> I just Cc'd you on a patch to fix this for remoteproc.
>> Please let me know with a Tested-by if this resolves your issue.
>>
>> Cheers,
>> Ahmad
>>
>>>
>>> Thank you.
>>>
>>> Best,
>>> Stefano
>>>
>>>
>>
>>
>> --
>> Pengutronix e.K.                           |                             |
>> Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
>> 31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
>> Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |
> 


-- 
Pengutronix e.K.                           |                             |
Steuerwalder Str. 21                       | http://www.pengutronix.de/  |
31137 Hildesheim, Germany                  | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |