On 04/08/2013 02:18:20 PM, Wolfgang Denk wrote:
Dear Scott,
In message 1365447916.28843.7@snotra you wrote:
Maybe "cache" should be the toplevel command, with "icache" and "dcache" refactored to be subcommands? Of course, then you're
making
an incompatible interface change. How much is consistency worth?
I think backward compatibility is mandatory here. We cannot break existing user scripts.
Sure. But if the main reason for the icache/dcache split is compatibility, I don't think that should constrict the form of new commands.
I'm also not convinced that merging this into one command would be a better design. I think the current split is a pretty good represen- tation of what the hardware looks like and how i works. Let's keep it.
The whole point of the patch is to expose the existing flush_cache() functionality, which is not split into icache/dcache. From the
user's
I understand this. But while the combination of IC and DC related operations into one function may be convenient for internal use, it is not a good idea when exposed externally.
perspective, it's a command to flush the specified region out of
*all*
caches. It's an implementation detail that some hardware or
I understand what you mean, but actually we do not flush the IC, we invalidate it, which is something different. I don't want to start a discussion here if flush_cache() is actually a bad function name (when discussing the functionality, it is), not do I want to suggest to change that name.
But when we publish such interfaces to the end user, it is important to be precise in our terminology.
I think the terminology is just fine. It's not just invalidating the icache (flushing and invalidating are the same thing for cache lines which are not modified -- or are incapable of being modified). It's flushing the region out of *all* caches.
What actual use case is there for only wanting to flush one or the other?
architectures accomplish this using separate dcache and icache instructions. If you make the interface be "icache/dcache", how
would
you handle hardware where the flushing mechanism (or even the cache itself) is not split?
If IC and DC are the same thing, the same function can be used to implement the operations. "icache" and "dcache" would then run the same code, i. e. be aliases.
[In the example of L2 cache above, it would be for example
sufficient
to add a "-L2" option to the "icache" / "dcache" commands.]
Would it? On our chips L2 cache is (more or less) unified.
There's no
separate icache/dcache flush.
See above. In such a case "icache" and "dcache" can just call the same underlying code.
And then we end up having to do the flush twice, if the user follows the "first flush dcache, then flush icache" instructions. That's not an ideal interface.
-Scott