On Sat, Aug 24, 2024 at 09:01:33AM +0000, LEROY Christophe wrote:

Le 23/08/2024 à 21:19, Segher Boessenkool a écrit :

quoted

The memset() code itself could chech for the storage attributes, but
that is probably more expensive than just assuming the happy case.
Maybe someone could try it out though!

But is it only memset() the problem ?

dcbz instruction is also used in:
- memcpy()
- csum_partial_copy_generic()
- clear_page()
- copy_page()
- clear_user()
- copy_to_user()
- copy_from_user()

That is just a handful of functions.  Not sure about the _user things,
and the _page things for that matter, but the rest is certainly
measurable in real-life conditions.  So if we can avoid the problems
completely, and cheaply, we probably should.

I'm not so sure about the cheaply though :-/

Are these functions also used on DMA coherent memory ?

Most won't show up high on most profiles, heh.  Which you already
can see from the problem not being attacked yet: if it was so obviously
a problem, some people would have wanted to do something about it :-)


Segher