ioremap_nocache() in setup_port() ok?

ioremap_nocache() in setup_port() ok?

[Grant Grundler]

Hi Russell,
This patch fixes a warning in the parisc-linux.org developement
tree. Helge Deller has done substantial work to enable proper
ioremap() functionality in parisc port and exposed some problems
with ioremap() vs ioremap_nocache() usage.

Most device drivers really want ioremap_nocache().
But we've been really sloppy since most implementations
treat MMIO space as uncached.

Am I correct that serial also wants _nocache() for MMIO space?

thanks,
grant

--- a/drivers/serial/8250_pci.c	13 Feb 2006 03:51:52 -0000	1.35
+++ b/drivers/serial/8250_pci.c	30 Mar 2006 07:21:03 -0000
@@ -87,7 +87,7 @@ setup_port(struct serial_private *priv, 
 		len =  pci_resource_len(dev, bar);
 
 		if (!priv->remapped_bar[bar])
-			priv->remapped_bar[bar] = ioremap(base, len);
+			priv->remapped_bar[bar] = ioremap_nocache(base, len);
 		if (!priv->remapped_bar[bar])
 			return -ENOMEM;
 

Re: ioremap_nocache() in setup_port() ok?

[Russell King]

On Thu, Mar 30, 2006 at 12:54:10AM -0700, Grant Grundler wrote:
> Am I correct that serial also wants _nocache() for MMIO space?

Don't know - I've always taken ioremap_nocache() as a mis-understanding
of what ioremap() itself is supposed to do.  Indeed, if you look at
IO-mapping.txt, ioremap_nocache() is not mentioned, and it's pretty
clear that ioremap() is what is intended to be used to map MMIO.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:  2.6 Serial core

Re: ioremap_nocache() in setup_port() ok?

[Grant Grundler]

On Thu, Mar 30, 2006 at 08:57:52AM +0100, Russell King wrote:
> On Thu, Mar 30, 2006 at 12:54:10AM -0700, Grant Grundler wrote:
> > Am I correct that serial also wants _nocache() for MMIO space?
> 
> Don't know - I've always taken ioremap_nocache() as a mis-understanding
> of what ioremap() itself is supposed to do.  Indeed, if you look at
> IO-mapping.txt, ioremap_nocache() is not mentioned, and it's pretty
> clear that ioremap() is what is intended to be used to map MMIO.

I think IO-mapping.txt is stale and whoever added ioremap_nocache
failed to update it.

It would be reasonable to ignore my patch (unless you wanted
to document the expected behavior of serial) if what willy
proposed an hour ago is accepted:
	http://lkml.org/lkml/2006/3/30/251

I'll update the IO-mapping.txt based on the outcome of
willy's patch.

thanks,
grant