On Sat, Nov 13, 2004 at 12:39:32PM +1100, Benjamin Herrenschmidt wrote:

On Fri, 2004-11-12 at 20:18 +0100, Guido Guenther wrote:

quoted

O.k., it was the __raw_{write,read}b which broke things, not the
"alignment". This one works:

diff -u -u linux-2.6.10-rc1-mm5.orig/drivers/video/riva/riva_hw.h linux-2.6.10-rc1-mm5/drivers/video/riva/riva_hw.h

--- linux-2.6.10-rc1-mm5.orig/drivers/video/riva/riva_hw.h	2004-11-12 13:42:54.000000000 +0100
+++ linux-2.6.10-rc1-mm5/drivers/video/riva/riva_hw.h	2004-11-12 17:39:22.000000000 +0100

@@ -75,8 +75,8 @@
  */
 #include <asm/io.h>
 
-#define NV_WR08(p,i,d)  (__raw_writeb((d), (void __iomem *)(p) + (i)))
-#define NV_RD08(p,i)    (__raw_readb((void __iomem *)(p) + (i)))
+#define NV_WR08(p,i,d)  (writeb((d), (void __iomem *)(p) + (i)))
+#define NV_RD08(p,i)    (readb((void __iomem *)(p) + (i)))

Interesting. The only difference here should be barriers. I hate the
lack of barriers in that driver ... I'm not sure the driver may not have

Yes, it's a real pain. Missing barriers in RIVA_FIFO_FREE caused lots of
lockups, until I found them in your 2.4 tree.

other bugs related to the lack of them in the 16 and 32 bits accessors.
It does use non-barrier version on purpose in some accel ops though,
when filling the fifo with pixels, but that's pretty much the only case
where it makes sense.

quoted

There aren't any, I actually attached the wrong patch. The non-working
version has __raw_{read,write}b8 instead of {read,write}b8. In 2.6.9
riva_hw.h used {in,out}_8 for NV_{RD,WR}08 so using the "non-raw"
writeb/readb now looks correct since these map to {in,out}_8 now.

{in,out}_8 are ppc-specific things that are identical to readb/writeb
indeed, with barriers.

In 2.6.10-rc1-mm5 {in,out}_8 and read/writeb are exactly identical, only
__raw_{read,write}b is different. So you mean __raw_{read,write}b in the
above? (no nitpicking, just want to be sure I understand this
correctly).
Cheers,
 -- Guido