On Sunday 24 January 2016 17:02:25 Johannes Berg wrote:

Instead of __raw_*, use *_relaxed. The conversion was done
using the following spatch (since that automatically adjusts
the coding style unlike a simple search&replace).

@@
expression E1, E2;
@@
-__raw_writel(E1, E2)
+writel_relaxed(E1, E2)
@@
expression E1, E2;
@@
-__raw_writew(E1, E2)
+writew_relaxed(E1, E2)
@@
expression E1;
@@
-__raw_readl(E1)
+readl_relaxed(E1)
@@
expression E1;
@@
-__raw_readw(E1)
+readw_relaxed(E1)

Signed-off-by: Johannes Berg <johannes@sipsolutions.net>

(Adding Shawn and Sascha here as well)

The patch looks fine to me, in the sense that we can safely assume
that the behavior won't change for little-endian, and that it is
fixing a bug for big-endian.

I would feel more comfortable with a patch using the non-relaxed()
accessors though, as that can avoid a number of subtle bugs with
I/O ordering.

Unfortunately, there is a nonzero overhead in the extra cache sync
on the Cortex-A9 based platforms (most imx6), so this has to
be done very carefully to avoid performance regressions. It's
probably a good idea to keep the two changes as separate commits,
but I'd also like to get both applied in the same merge window,
as 'readl_relaxed' tends to say "I have proven that I don't need
barriers", while the existing code just screams "I have no idea
about what I'm doing here".

Any other opinions on this?

	Arnd