On Fri, Apr 15, 2022 at 03:10:51PM -0700, Linus Torvalds wrote:
> Adding PeterZ and Borislav (who seem to be the last ones to have
> worked on the copy and clear_page stuff respectively) and the x86
> maintainers in case somebody gets the urge to just fix this.
I guess if enough people ask and keep asking, some people at least try
to move...
> Because memory clearing should be faster than copying, and the thing
> that makes copying fast is that FSRM and ERMS logic (the whole
> "manually unrolled copy" is hopefully mostly a thing of the past and
> we can consider it legacy)
So I did give it a look and it seems to me, if we want to do the
alternatives thing here, it will have to look something like
arch/x86/lib/copy_user_64.S.
I.e., the current __clear_user() will have to become the "handle_tail"
thing there which deals with uncopied rest-bytes at the end and the new
fsrm/erms/rep_good variants will then be alternative_call_2 or _3.
The fsrm thing will have only the handle_tail thing at the end when size
!= 0.
The others - erms and rep_good - will have to check for sizes smaller
than, say a cacheline, and for those call the handle_tail thing directly
instead of going into a REP loop. The current __clear_user() is still a
lot better than that copy_user_generic_unrolled() abomination. And it's
not like old CPUs would get any perf penalty - they'll simply use the
same code.
And then you need the labels for _ASM_EXTABLE_UA() exception handling.
Anyway, something along those lines.
And then we'll need to benchmark this on a bunch of current machines to
make sure there's no funny surprises, perf-wise.
I can get cracking on this but I would advise people not to hold their
breaths. :)
Unless someone has a better idea or is itching to get hands dirty
her-/himself.
Thx.
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
