Re: [PATCH v5] remoteproc: imx_dsp_rproc: add custom memory copy implementation for i.MX DSP Cores

[Daniel Baluta]

On Wed, Mar 1, 2023 at 9:11 PM Mathieu Poirier
<mathieu.poirier@xxxxxxxxxx> wrote:
>
> On Wed, Mar 01, 2023 at 11:49:16AM -0700, Mathieu Poirier wrote:
> > Hi Iuliana,
> >
> > On Tue, Feb 21, 2023 at 07:03:56PM +0200, Iuliana Prodan (OSS) wrote:
> > > From: Iuliana Prodan <iuliana.prodan@xxxxxxx>
> > >
> > > The IRAM is part of the HiFi DSP.
> > > According to hardware specification only 32-bits write are allowed
> > > otherwise we get a Kernel panic.
> > >
> > > Therefore add a custom memory copy and memset functions to deal with
> > > the above restriction.
> > >
> > > Signed-off-by: Iuliana Prodan <iuliana.prodan@xxxxxxx>
> > > ---
> > > Changes since v4
> > > - use GENMASK;
> > > - s/ioread32/readl;
> > > - s/iowrite32/writel;
> > > - use for loop with writel instead of __iowrite32_copy;
> > > - update multi-line comment format.
> > >
> >
> > This looks good now.  I will queue this on March 13th when 6.3-rc2 is out.
> >
>
> And out of curiosity, are the remote processors handled by imx_rproc.c also have
> the same 32-bit write constraints?  If so, it would be nice to have a patch that
> fixes that as well.  In such a case, imx_dsp_rproc_memcpy() and
> imx_dsp_rproc_memset() should be renamed to something more generic and re-used
> in imx_rproc.c.

Paging Peng Fan.

imx_rproc handles M4/M7 cores.
imx_dsp_rproc handles HIFI4 DSP cores.

>From what I experienced 32 bit read/writes are only mandatory for
certain memory areas
that fall inside the HIFI4 address spaces but this doesn't happen for
M4/M7 memory maps.

thanks,
Daniel.