On 09/08/2015 08:09 AM, Linus Walleij wrote:

On Fri, Aug 28, 2015 at 12:24 PM, Masahiro Yamada
[off-list ref] wrote:

quoted

2015-08-26 22:39 GMT+09:00 Linus Walleij [off-list ref]:

quoted

quoted

cache-unified and cache-level are *not* optional and should be required.

"cache-unified" is mentioned in "3.7.3 Internal (L1) Cache Properties"
(Table 3-8),
but it is not in "3.8 Multi-level and Shared Caches" (Table 3-9)

Are the rules in Table 3-8 also applied for L2?

Your guess is as good as mine unless someone involved in
actually writing that spec says something :/

Maybe you'd have to be crazy to have Harvard cache for 2nd+ level. I've
got no clue. Doesn't hurt to have it.

quoted

quoted

(I'm just assuming this cache is unified, anything else would be baffling.)

In fact, unified/harvard is configurable thru a register of this cache
controller.

Jesus Christ.

Hardware designers either hate software folks or ensure our job security.

quoted

It is usually used as a unified cached, though.

I would, too.

quoted

So,I am planning to use the same compatible for L2 and L3, like this:


       l2-cache at 500c0000 {
               compatible = "socionext,uniphier-cache";
               reg = <0x500c0000 0x2000>, <0x503c0100 0x8>,
                     <0x506c0000 0x400>;
               cache-unified;
               cache-level = <2>;
               next-level-cache = <&L2>;

Next level of the L2 is the L2?

quoted

               cache-size = <0x200000>;
               cache-sets = <256>;
               cache-line-size = <128>;
            };

       /* Not all of UniPhier SoCs have L3 cache */
       l3-cache at 500c8000 {
               compatible = "socionext,uniphier-cache";
               reg = <0x500c8000 0x2000>, <0x503c8100 0x8>,
                     <0x506c8000 0x400>;
               cache-unified;
               cache-level = <3>;
               cache-size = <0x400000>;
               cache-sets = <256>;
               cache-line-size = <256>;
       };

This LooksGoodToMe.

quoted

The Table 3-9 in ePAPR v1.1 says
the compatible should be "cache", but I do not think it makes sense here.

Agree.

It could be useful for finding all cache nodes, but we've generally
failed to use it, so at this point it doesn't matter.

Rob