[PATCH V3 0/6] remoteproc: imx_rproc: support firmware in DDR

Wed Mar 22 08:13:49 PDT 2023

On Wed, Mar 22, 2023 at 11:59:02AM +0100, Frieder Schrempf wrote:
> Hi,
> 
> On 07.03.23 21:26, Mathieu Poirier wrote:
> > On Sat, Mar 04, 2023 at 03:59:38PM +0800, Peng Fan wrote:
> >>
> >>
> >> On 2/14/2023 1:50 AM, Mathieu Poirier wrote:
> >>> On Mon, Feb 13, 2023 at 12:15:59PM +0200, Iuliana Prodan wrote:
> >>>> On 2/12/2023 9:43 AM, Peng Fan wrote:
> >>>>> Hi Iuliana,
> >>>>>
> >>>>>> Subject: Re: [PATCH V3 0/6] remoteproc: imx_rproc: support firmware in
> >>>>>> DDR
> >>>>>>
> >>>>>>
> >>>>>> On 2/9/2023 8:38 AM, Peng Fan (OSS) wrote:
> >>>>>>> From: Peng Fan <peng.fan at nxp.com>
> >>>>>>>
> >>>>>>> V3:
> >>>>>>>
> >>>>>>>     Daniel, Iuliana
> >>>>>>>
> >>>>>>>       Please help review this patchset per Mathieu's comments.
> >>>>>>>
> >>>>>>>     Thanks,
> >>>>>>>     Peng.
> >>>>>>>
> >>>>>>>     Move patch 3 in v2 to 1st patch in v3 and add Fixes tag Per Daniel
> >>>>>>>     IMX_RPROC_ANY in patch 3 Per Mathieu
> >>>>>>>     Update comment and commit log in patch 5, 6.
> >>>>>>>
> >>>>>>>     NXP SDK provides ".interrupts" section, but I am not sure how others
> >>>>>>>     build the firmware. So I still keep patch 6 as v2, return bootaddr
> >>>>>>>     if there is no ".interrupts" section.
> >>>>>>>
> >>>>>>> V2:
> >>>>>>>     patch 4 is introduced for sparse check warning fix
> >>>>>>>
> >>>>>>> This pachset is to support i.MX8M and i.MX93 Cortex-M core firmware
> >>>>>>> could be in DDR, not just the default TCM.
> >>>>>>>
> >>>>>>> i.MX8M needs stack/pc value be stored in TCML entry address[0,4], the
> >>>>>>> initial value could be got from firmware first section ".interrupts".
> >>>>>>> i.MX93 is a bit different, it just needs the address of .interrupts
> >>>>>>> section. NXP SDK always has .interrupts section.
> >>>>>>>
> >>>>>>> So first we need find the .interrupts section from firmware, so patch
> >>>>>>> 1 is to reuse the code of find_table to introduce a new API
> >>>>>>> rproc_elf_find_shdr to find shdr, the it could reused by i.MX driver.
> >>>>>>>
> >>>>>>> Patch 2 is introduce devtype for i.MX8M/93
> >>>>>>>
> >>>>>>> Although patch 3 is correct the mapping, but this area was never used
> >>>>>>> by NXP SW team, we directly use the DDR region, not the alias region.
> >>>>>>> Since this patchset is first to support firmware in DDR, mark this
> >>>>>>> patch as a fix does not make much sense.
> >>>>>>>
> >>>>>>> patch 4 and 5 is support i.MX8M/93 firmware in DDR with parsing
> >>>>>>> .interrupts section. Detailed information in each patch commit message.
> >>>>>>>
> >>>>>>> Patches were tested on i.MX8MQ-EVK i.MX8MP-EVK i.MX93-11x11-EVK
> >>>>>> If one can build their firmware as they want, then the .interrupt section can
> >>>>>> also be called differently.
> >>>>>> I don't think is a good idea to base all your implementation on this
> >>>>>> assumption.
> >>>>>>
> >>>>>> It's clear there's a limitation when linking firmware in DDR, so this should be
> >>>>>> well documented so one can compile their firmware and put the needed
> >>>>>> section (interrupt as we call it in NXP SDK) always in TCML - independently
> >>>>>> where the other section go.
> >>>>> Ok, so .interrupt section should be a must in elf file if I understand correctly.
> >>>>>
> >>>>> I could add a check in V4 that if .interrupt section is not there, driver will report
> >>>>> failure.
> >>>>>
> >>>>> How do you think?
> >>>>
> >>>> Peng, I stand by my opinion that the limitation of linking firmware in DDR
> >>>> should be documented in an Application Note, or maybe there are other
> >>>> documents where how to use imx_rproc is explained.
> >>>>
> >>>> The implementation based on the .interrupt section is not robust.
> >>>> Maybe a user linked his firmware correctly in TCML, but the section is not
> >>>> called .interrupt so the firmware loading will work.
> >>>>
> >>>> So, instead of using the section name, you should use the address.
> >>>
> >>> Can you be more specific on the above?
> >>>
> >>>>
> >>>> First, check whether there is a section linked to TCML.
> >>>> If there is none, check for section name - as you did.
> >>>> If there is no section called .interrupt, give an error message.
> >>>
> >>> We have two ways of booting, one that puts the firmware image in the TCML and
> >>> another in RAM.  Based on the processor type, the first 8 bytes of the TCML need
> >>> to include the address for the stack and PC value.
> >>>
> >>> I think the first thing to do is have two different firmware images, one for
> >>> i.MX8M and another one for i.MX93.  That should greatly simplify things.
> >>
> >> sorry, I not got your points. i.MX8M and i.MX93 are not sharing firmware
> > 
> > Perfect.
> > 
> >> images. i.MX93 M33 has ROM, kicking M33 firmware just requires the
> >> address of the .interrupt address which holds stack/pc value.
> >> i.MX8M not has ROM, kick M33 firmware requires driver to copy
> >> stack/pc into the TCML beginning address.
> > 
> > It's been more than a month since I have looked at this patchset so the details are
> > vague in my memory.  That said, there should be one image for the TCML and
> > another one for the RAM.  And the image that runs in RAM should have a program
> > segment that write the correct information in the first 8 bytes.
> > 
> >>
> >> Whether i.MX8M/i.MX93, the NXP released MCU SDK use the section
> >> ".interrupt" to hold stack/pc initialization value in the beginning
> >> 8 bytes of the section.
> >>
> > 
> > And that is fine.  Simply release another version of the SDK that does the right
> > thing.
> > 
> > I suggest to work with Daniel and Iuliana if some details are still unclear.
> > Unlike me, they have access to the reference manual and the boot requirements.
> > 
> > 
> >>>
> >>> Second, there should always be a segment that adds the right information to the
> >>> TMCL.  That segment doesn't need a name, it simply have to be part of the
> >>> segments that are copied to memory (any kind of memory) so that function
> >>> rproc_elf_load_segments() can do its job.
> >>>
> >>> That pushes the complexity to the tool that generates the firmware image,
> >>> exactly where it should be.
> >>
> >> For i.MX8M, yes. For i.MX93, the M33 ROM needs address of storing stack/pc.
> >>>
> >>> This is how I think we should solve this problem based on the very limited
> >>> information provided with this patchset.  Please let me know if I missed
> >>> something and we'll go from there.
> >>
> >> I am not sure how to proceed on supporting the current firmware. what should
> >> I continue with current patchset?
> 
> I've successfully tested this on i.MX8MM with an elf file generated by
> the NXP SDK.
> 
> I would really like to see this upstreamed. If this requires changes
> that are not compatible with binaries compiled with the current SDK as
> discussed above, that would be fine for me as long as the kernel is able
> to detect the malformed binary and warns the user about it.
>

I agree.

> The user can then manually adjust the linker script, etc. in the SDK to
> match the requirements of the kernel.
> 

That is exactly what I suggested.

> Thanks
> Frieder