Re: [PATCH] spi/pxa2xx-pci: Enable DMA binding through device name

[Mika Westerberg <mika.westerberg@linux.intel.com>]

On Fri, Jul 25, 2014 at 10:38:36AM +0200, Arnd Bergmann wrote:
> On Friday 25 July 2014 11:22:49 Mika Westerberg wrote:
> > > All you need to do is change your filter function to take the
> > > slave id from pxa_spi_info and stick it in there, e.g.
> > > 
> > > static bool pxa2xx_spi_dw_dma_filter(struct dma_chan *chan, void *param)
> > > {       
> > >         const struct pxa2xx_spi_master *pdata = param;
> > >         struct dw_dma_chan *dwc = to_dw_dma_chan(chan);
> > >         
> > >         dwc->request_line = fargs->req;
> > >         dwc->src_master = 0;
> > >         dwc->dst_master = 0;
> > >         
> > >         return 1;
> > > }           
> > 
> > Oh man. That makes pxa2xx_spi dependent on a certain specific DMA engine
> > driver.
> 
> I think you can improve this by putting the filter function (and a pointer
> to it) into the pxa2xx_spi_master data provided by the PCI driver.

Indeed, that looks better. It still makes the PCI part of the driver
dependent on a particular DMA engine driver but is certainly better than
the core pxa2xx_spi driver.

> On ARM, we usually provide those pointers through platform_data from
> the board file.
> 
> > > Note that the filter function by definition is specific to the dma
> > > controller, not the dma slave (that's why most people define it in
> > > the dmaengine driver), and the pxa2xx_spi_dma_filter() function used
> > > in spi-pxa2xx-dma.c looks like it was written for another dma engine:
> > 
> > I wonder what's the rationale that passing slave_id with
> > dma_slave_config is wrong? The current code works fine with that and is
> > is independent of the DMA engine driver (even though we know that it is
> > going to be dw-dma).
> > 
> > The dw-dma handles slave_id in its implementation of
> > dmaengine_slave_config().
> 
> The main point is that a single 'slave_id' is not actually enough to
> identify what a slave needs. In case of dw_dma, the dma engine actually
> requires three numbers (request line, source master, destination master)
> to identify it, and there is no good way to put all that information into
> a single integer. Other dma engines require a different set of data.
> 
> There are only two drivers left that actually use slave_id in this way,
> and only for the legacy (no DT or ACPI) case, every other driver uses
> either DT or a filter function. I believe the shmobile part will soon
> be done with, after shmobile has been converted to DT, and after that
> we should remove the slave_id field from the dma_slave_config interface.

OK, thanks for the explanation.