[PATCH 3/3 v2] dmaengine: Add Freescale i.MX SDMA support

[s.hauer@pengutronix.de (Sascha Hauer)]

On Sat, Aug 28, 2010 at 05:27:10PM +0200, Marek Vasut wrote:
> Dne So 28. srpna 2010 17:18:17 Linus Walleij napsal(a):
> > 2010/8/23 Uwe Kleine-K?nig <u.kleine-koenig@pengutronix.de>:
> > >> +     evt = readl(SDMA_H_EVTOVR);
> > >> +     mcu = readl(SDMA_H_HOSTOVR);
> > >> +     dsp = readl(SDMA_H_DSPOVR);
> > > 
> > > __raw_readl?
> > 
> > Sorry I never understood this __raw_[read|write][b|w|l] vs.
> > plain read[b|w|l] proliferation in some drivers and code.
> > 
> > What's the reason for?
> 
> Hey,
> 
> this trick is, if you create the VA<->PA mapping at the kernel start (eg. see 
> how pxa_map_io() is replacedon some devices for instance), you then use the VA 
> address you specified and use __raw_{read,write}[b,w,l](). So use 
> __raw_{read,write}[b,w,l]() on drivers specific for certain device and use 
> {read,write}[b,w,l]() on ioremap()ed memory areas, aka. in common drivers.

Nope, this has nothing to do with static mappings vs. ioremap. The
difference is that read[b,w,l] do little endian accesses suitable for
PCI whereas the __raw_* functions do accesses in CPU endianess.
Peripherals integrated into a SoC like the SDMA engine here are
normally accessible in native endianess and thus need the __raw_*
functions. An external network controller (for example a LAN9117) will
probably need the non raw functions. Note that 99% of the arm users use
little endian only and thus cpu_to_le* is a noop, so both types will work
for most people.

Sascha

-- 
Pengutronix e.K.                           |                             |
Industrial Linux Solutions                 | http://www.pengutronix.de/  |
Peiner Str. 6-8, 31137 Hildesheim, Germany | Phone: +49-5121-206917-0    |
Amtsgericht Hildesheim, HRA 2686           | Fax:   +49-5121-206917-5555 |