Re: [PATCH] dw_mmc: Add Synopsys DesignWare mmc host driver.

[Russell King - ARM Linux <linux@arm.linux.org.uk>]

On Sun, Dec 12, 2010 at 08:41:36AM +0000, Russell King - ARM Linux wrote:
> On Sat, Dec 11, 2010 at 07:23:20PM +0000, Chris Ball wrote:
> > Hi Will,
> > 
> > On Thu, Dec 09, 2010 at 05:24:26PM +0000, Will Newton wrote:
> > > This adds the mmc host driver for the Synopsys DesignWare mmc
> > > host controller, found in a number of embedded SoC designs.
> > > 
> > > Signed-off-by: Will Newton <will.newton@imgtec.com>
> > > Reviewed-by: Matt Fleming <matt@console-pimps.org>
> > 
> > Running a test build on ARM fails:
> > 
> > drivers/mmc/host/dw_mmc.c: In function ‘dw_mci_push_data64’:
> > drivers/mmc/host/dw_mmc.c:985: error: implicit declaration of function ‘__raw_writeq’
> > drivers/mmc/host/dw_mmc.c: In function ‘dw_mci_pull_data64’:
> > drivers/mmc/host/dw_mmc.c:998: error: implicit declaration of function ‘__raw_readq’
> > 
> > because arch/arm doesn't implement raw versions of these 64-bit accesses.
> > I'm surprised that this driver hasn't been compiled on ARM before!  What
> > kind of arch are you testing on?  Do you have any ARM hardware (lpc313x?)
> > to verify the driver on?
> 
> What's the semantics of a 64-bit IO access?  Does the low 32-bit get
> written before the high 32-bit, or is it the other way around?  Does
> it depend on the endian-ness?  What if some hardware needs the low
> 32-bit first and other needs the high 32-bit first?
> 
> I don't think it's reasonable to expect 32-bit hardware to perform 64-bit
> IO accesses.

I should cover something else here, in anticipation of someone trying
to be clever...

Using ldrd/strd, or ldm/stm for IO accesses on ARM is a very bad idea
for generic code.  While nothing prevents you from using these for 64-bit
IO accesses, you have to be aware that the normal access guarantees do
not apply.

ldrd/strd is implemented as two individual 32-bit single-access operations,
each of which is atomic, and therefore the instruction can be interrupted
half-way through.  Upon restart, it can repeat the first load/store.

If you're accessing a FIFO, repeated accesses will be a problem, and
will cause data corrpution.  If you're accessing a control register,
the first write could have a side effect (eg, starting DMA) before the
second write has occurred (eg, setting DMA parameters.)

So, even with ldrd/strd, you still have the problem of whether the
64-bit access can be split into two separate 32-bit accesses safely,
but you also have the problem that the individual 32-bit accesses may
be repeated.

You're safer using standard 32-bit load/stores, which won't suffer
from being repeated - but you still have to decide whether high-word
first or low-word first is the correct transfer order.