Re: [BUG BISECT] Ethernet fail on VF50 (OF: Don't set default coherent DMA mask)

[Robin Murphy]

On 31/07/18 18:38, Guenter Roeck wrote:
> On Tue, Jul 31, 2018 at 04:58:41PM +0100, Robin Murphy wrote:
> > On 31/07/18 16:43, Guenter Roeck wrote:
> > > On Tue, Jul 31, 2018 at 03:09:34PM +0100, Robin Murphy wrote:
> > > > > Please note that sparc images still generate the warning (next-20180731).
> > > >
> > > > Ugh, OK, any ideas what sparc does to create these platform devices that
> > > > isn't of_platform_device_create_pdata() and has somehow grown an implicit
> > > > dependency on of_dma_configure() since 4.12? I'm looking, but nothing jumps
> > > > out...
> > >
> > > I suspect it might be of_device_register(), called from
> > > 	arch/sparc/kernel/of_device_64.c:scan_one_device()
> > > 	arch/sparc/kernel/of_device_32.c:scan_one_device()
> >
> > Right, that's as far as I got as well, so I'm struggling to see how these
> > things ever got DMA masks set before the of_dma_configure() call moved out
> > of of_platform_device_create_pdata(), or why it wasn't a problem prior to
> > the generic dma_ops rework if they didn't :/
> Ah, ok. No idea, sorry. All I know is that the messages were first seen
> with next-20180727.

OK, I spent this afternoon wrangling toolchains and QEMU to boot an 
instrumented 4.11 kernel, and the answer is that the warnings are 
arguably correct. These masks have indeed never been set where they 
should have been, but then the sbus_dma_ops don't reference them anyway.

The coherent mask WARN_ON *should* have started appearing in 4.16 with 
205e1b7f51e4("dma-mapping: warn when there is no coherent_dma_mask"), 
but happened to be hidden by the inadvertent side-effect of the prior 
dma_configure() change. Since there's seemingly no actual regression of 
functionality, I'm inclined to leave this in the hands of whoever cares 
about sparc32.

Robin.