On Fri, Mar 3, 2023 at 11:37 AM Mateusz Guzik [off-list ref] wrote:

I mentioned in the previous e-mail that memset is used a lot even
without the problematic opt and even have shown size distribution of
what's getting passed there.

Well, I *have* been pushing Intel to try to fix memcpy and memset for
over two decades by now, but with FSRM I haven't actually seen the
huge problems any more.

It may just be that the loads I look at don't have issues (or the
machines I've done profiles on don't tend to show them as much).

Hmm. Just re-did my usual kernel profile. It may also be that
something has changed. I do see "clear_page" at the top, but yes,
memset is higher up than I remembered.

I actually used to have the reverse of your hack for this - I've had
various hacks over the year that made memcpy and memset be inlined
"rep movs/stos", which (along with inlined spinlocks) is a great way
to see the _culprit_ (without having to deal with the call chains -
which always get done the wrong way around imnsho).

            Linus