Re: flush_dcache_page vs kunmap_local

[Linus Torvalds]

On Thu, Nov 4, 2021 at 8:03 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> Linus offers the opinion that kunmap calls should imply a
> flush_dcache_page().  Christoph added calls to flush_dcache_page()
> in commit 8dad53a11f8d.  Was this "voodoo programming", or was there
> a real problem being addressed?

I don't think anybody actually uses/cares about flush_dcache_page() at
all, and pretty much all uses are random and voodoo.

No sane architecture uses pure virtual caches, and the insane ones
haven't been an issue for a long time either.

But if there are still systems with pure virtual caches, and they need
manual cache flushing, then I do think that kunmap is one of the
points that needs it, since that's the "I'm done accessing this data
through this virtual address" place.

End result: I really don't think anybody cares any more (and only
truly broken architectures ever did). I'd personally be perfectly
happy just saying "we might as well drop support for non-coherent
caches entirely".

But as long as we have those random odd "flush dcache manually"
things, I think kunmap() is one of the places that probably should
continue to do them.

Of course, the kunmap case is _doubly_ irrelevant, because we should
certainly hope that not only are those noncoherent pure virtual caches
a thing of the past, highmem itself should be going away.

Why did this come up? Do you actually have some hardware or situation
that cares?

                   Linus