[PATCH] mmci: restrict DMA usage to large, even multiblock transfers

[Linus Walleij]

2011/2/1 Russell King - ARM Linux <linux at arm.linux.org.uk>:
> On Tue, Feb 01, 2011 at 03:07:12PM +0100, Linus Walleij wrote:
>> However I think this weird bit turning off all single requests
>> is there for some reason and I still try to find out what the
>> usecase really is. I highly suspect that it's related to either
>> doing some SDIO usecase or pleasing some buggy DMA
>> controller.
>
> We aren't turning off single requests - by clearing DMAREQCTL, we're
> actually turning them back on.

Yes that's what I meant, sorry for the confusion.

> What I think is going on is someone had an idea of using LBREQ to
> transfer the last one to seven transfers as a burst, rather than one
> to seven separate single transfers as an optimization.  However, this
> can only work if the DMA controller obeys the transfer count and doesn't
> try to transfer more than the remainder of the transfer in any burst.
> IOW, the DMAC has to truncate bursts to the smaller of the burst size
> and remainder.  If it can't, DMAREQCTL causes problems.

Seems plausible. I'll try to look into it.

> I assume that the U8500/U300 documentation isn't publically available...
> A quick google doesn't seem to help.

In this case it doesn't help... U300 does not have this peculiar
feature BTW, it's a pure ux500 thing.

If it gives clear performance improvements (need measurements)
we can always patch it in later. For now much is already achieved
by this patch.

Yours,
Linus Walleij