Re: [PATCH 4/6] dma-mapping: merge direct and noncoherent ops

[Maciej W. Rozycki]

On Wed, 31 Oct 2018, Maciej W. Rozycki wrote:

> > All the cache maintainance is already stubbed out when not enabled,
> > but merging the two allows us to nicely handle the case where
> > cache maintainance is required for some devices, but not others.
> 
>  FYI, you commit bc3ec75de545 ("dma-mapping: merge direct and noncoherent 
> ops") has caused:
> 
> fddi0: DMA command request failed!
> fddi0: Adapter open failed!
> 
> with the `defxx' driver on my R4400SC TURBOchannel DECstation (but not the 
> R3400 one) and consequently the interface does not work anymore.  Both are 
> non-coherent cache systems, however the R3400 implements the write-through 
> policy while the policy of the R4400SC is write-back (it also has 1MiB of 
> secondary cache), which I suspect is the reason of the difference.

 Doh!  It would have helped if I actually had the right adapter installed 
in the R3400 box.  I've got a spare one, which I have now plugged in there 
and the R3400 actually shows the same issue as the R4400SC does.  So this 
has nothing to do WRT write-through vs write-back.

 Hmm, in `dfx_hw_dma_cmd_req' the driver polls the consumer block (which 
is write-only by the adapter and read-only by the host, except for the 
initialisation time before adapter's DMA engines have been started, and 
write-through vs write-back indeed does not matter for this kind of use) 
and there's no DMA synchronisation whatsoever in that.

 However the block has been allocated with `dma_zalloc_coherent', which 
means no synchronisation is supposed to be required.  For non-coherent 
cache systems that essentially means an uncached memory allocation.

  Maciej