Re: [linus:master] [iov_iter] c9eec08bac: vm-scalability.throughput -16.9% regression

[Linus Torvalds]

On Wed, 15 Nov 2023 at 14:10, Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> Should we define an alternative_memcpy() which is used *only* during
> rewriting so that this becomes a non-issue?

Yeah, I think the instruction rewriting should use something that
explicitly cannot possibility itself need rewriting, and a plain
'memcpy()' is obviously that.

The good news is that at least things like structure copies would
*not* trigger that alternative, so it's only explicit memcpy() calls
that my patch changes. But I would not be surprised if instruction
rewriting does that. I didn't actually check.

> Yours looks simple enough and makes sense. Lemme poke at it a bit in the
> coming days and see what happens.

Note that it has a nasty interaction with fortify-source, which is why
it has that hacky "#undef memcpy" in that unrelated header.

Also note that I was being very very lazy in how I re-used the
"rep_movs_alternative" function that we already have. And it's
actually a bad laziness, because our existing rep_movs_alternative
does the exception handling for user mode faults.

We don't actually want exception handling for 'memcpy()', because it
could hide bugs. If a memcpy() gets a bad pointer, we want the oops,
not a partial copy.

So my patch really is broken. It might happen to work when everything
else goes right, and it's small, but it is very much a "proof of
concept" rather than something that is actually acceptable.

               Linus