Re: [PATCH 0/2] firmware: cs_dsp: Remove usage of GFP_DMA

[Richard Fitzgerald]

On 20/02/2025 9:52 am, Charles Keepax wrote:
> On Wed, Feb 19, 2025 at 04:26:55PM +0000, Richard Fitzgerald wrote:
> > On 13/02/2025 3:16 pm, Thomas Weißschuh wrote:
> > > On Thu, Feb 13, 2025 at 03:06:59PM +0000, Mark Brown wrote:
> > > > On Thu, Feb 13, 2025 at 02:28:06PM +0000, Richard Fitzgerald wrote:
> > > > > On 11/02/2025 5:21 pm, Richard Fitzgerald wrote:
> > > >
> > > > > > > Not tested on real hardware.
> > > > > > > This came up while porting kunit to mips64.
> > > > > > > Apparently GFP_DMA does not work there, but IMO the usage of GFP_DMA by
> > > >
> > > > > I would say that is a bug in mips64 that should be fixed in mips64.
> > > > > It is not reasonable to expect generic drivers to have special cases for
> > > > > platforms that don't handle GFP_DMA.
> > >
> > > Indeed, I did that, too.
> > >
> > > > What specifically is the issue?  If it's a build time issue I'd
> > > > definitely agree that we should just be able to assume that platforms at
> > > > least build.  IIRC there is a Kconfig you can depend on for DMA but it
> > > > seems more trouble than it's worth to fix all users.
> > >
> > > More details in [0], It's only a runtime issue.
> > >
> > > I'm still wondering how all the on-stack buffers used with regmap_raw_read()
> > > and regmap_raw_write() by cs_dsp are satisfying the DMA requirements.
> > There are 3 suspicious regmap_raw_read(). The others are all integers,
> > which are guaranteed not to cross a page boundary.
> >
> > However, it looks like it might be safe to remove the GFP_DMA flags
> > now. regmap_raw_read() calls spi_write_then_read() which specifically
> > says that the buffers do not need to be DMA-safe and internally uses a
> > DMA-safe buffer. regmap_raw_write() uses either a temporary physically
> > contiguous buffer or GFP_DMA buffer (the implementation is terrifyingly
> > complex so it's difficult to determine exactly what it does).
> >
> > (Some older systems could only use certain special memory areas for DMA
> > but we haven't seen any of those used with cs_dsp.)
>
> We also need to consider what the I2C subsystem does, I have a
> vague memory of thinking the SPI system will attempt to remap
> buffers but I2C will just use them as is. cs_dsp will be used
> with both, although SPI is slightly more common for obvious
> reasons.
>
> Thanks,
> Charles

For information here is the presentation given by Wolfram Sang
describing the DMA problem. This is the reason we used GFP_DMA buffers
in the cs_dsp code.

https://events19.linuxfoundation.org/wp-content/uploads/2017/12/20181023-Wolfram-Sang-ELCE18-safe_dma_buffers.pdf