[PATCH v9 01/18] arm: make SWIOTLB available

[Russell King - ARM Linux]

On Tue, Oct 29, 2013 at 05:24:50PM +0000, Stefano Stabellini wrote:
> On Tue, 29 Oct 2013, Russell King - ARM Linux wrote:
> > This looks like a hack.  Either we want to use the streaming mask or
> > the coherent mask as appropriate for the caller.  That should be a choice
> > the caller makes, not the implementation of this behind the callers back.
> 
> How should I choose?
> Admittedly this is a workaround because some drivers (including Calxeda
> xgmac) don't set the dma_mask when they should:
> 
> http://marc.info/?l=linux-arm-kernel&m=138082570519601&w=2
> 
> I am happy to reduce this to:
> 
> if (dev->dma_mask)
>     mask = *dev->dma_mask;
> 
> even though it means that xgmac & co won't be dma_capable.

Well, if xgmac (there's two in the kernel tree so I'm not sure which
you're referring to) is using DMA of any kind, it really ought to be
setting a DMA mask via the standard functions - in todays kernel,
that is dma_set_mask() for the streaming mask and dma_set_coherent_mask()
for the coherent mask.  The appropriate mask which the driver is aware
of for the hardware should be set here.

In the case that a driver needs to call both, and the kernel has my
patch set applied, dma_set_mask_and_coherent() can be used to set both
at the same time.

dma_set_mask() (and dma_set_mask_and_coherent) will fail if storage for
the DMA mask has not been allocated previously - iow, if the dev->dma_mask
pointer is NULL.  The allocation of this storage is the responsibility
of the struct device creator, and the struct device creator should
initialise the DMA masks to a sensible default for the type of bus.

I provide one last call in my patch set - a method that drivers can
coerce the DMA mask to whatever they desire irrespective of whether
the DMA mask allocation has been setup.  This is provided as a way to
mark the drivers which are doing this and to ensure that drivers do
this in a safe manner - we've had drivers setting it to static storage
in module space if it isn't previously set, which is a problem if the
module ever gets unloaded.

So... what I'm saying is please fix xgmac to use the DMA API properly
rather than working around such problems. :)