Re: [PATCH 0/6] use memcpy_mcsafe() for copy_to_iter()

[Linus Torvalds <torvalds@linux-foundation.org>]

On Tue, May 1, 2018 at 9:00 PM Dan Williams <dan.j.williams@intel.com>
wrote:
> >
> > I  have some dim memory of "rep movs doesn't work well for pmem", but
does
> > it *seriously* need unrolling to cacheline boundaries? And if it does,
who
> > designed it, and why is anybody using it?
> >

> I think this is an FAQ from the original submission, in fact some guy
> named "Linus Torvalds" asked [1]:

Oh, I already mentioned that  I remembered that "rep movs" didn't work well.

But there's a big gap between "just use 'rep movs' and 'do some cacheline
unrollong'".

Why isn't it just doing a simple word-at-a-time loop and letting the CPU do
the unrolling that it will already do on its own?

I may have gotten that answered too, but there's no comment in the code
about why it's such a disgusting mess, so I've long since forgotten _why_
it's such a disgusting mess.

That loop unrolling _used_ to be "hey, it's simple".

Now it's "Hey, that's truly disgusting", with the separate fault handling
for every single case in the unrolled loop.

Just look at the nasty _ASM_EXTABLE_FAULT() uses and those E_cache_x error
labels, and getting the number rof bytes copied right.

And then ask yourself "what if we didn't unroll that thing 8 times, AND WE
COULD GET RID OF ALL OF THOSE?"

Maybe you already did ask yourself.  But I'm asking because it sure isn't
explained in the code.

             Linus