On Wed, May 04, 2022 at 12:22:34PM -0700, Linus Torvalds wrote:
> Side note: the "do FSRM inline" would likely be a really good thing
> for "copy_to_user()", more so than the silly "clear_user()" that we
> realistically do almost nowhere.
Right, that would be my next project.
>
> I doubt you can find "clear_user()" outside of benchmarks (but hey,
> people do odd things).
Well, see preview below.
> But "copy_to_user()" is everywhere, and the I$ advantage of inlining
> it might be noticeable on some real loads.
>
> I remember some git profiles having copy_to_user very high due to
> fstat(), for example - cp_new_stat64 and friends.
>
> Of course, I haven't profiled git in ages, but I doubt that has
Yeah, see below.
> changed. Many of those kinds of loads are all about name lookup and
> stat (basic things like "make" would be that too, if it weren't for
> the fact that it spends a _lot_ of its time in user space string
> handling).
>
> The inlining advantage would obviously only show up on CPUs that
> actually do FSRM. Which I think is currently only Ice Lake. I don't
> have access to one.
Zen3 has FSRM.
So below's the git test suite with clear_user on Zen3. It creates a lot
of processes so we get to clear_user a bunch and that's the inlined rep
movsb.
You can see some small but noticeable improvement:
gitsource
rc clear_use
rc5 clear_user
Min User 196.65 ( 0.00%) 193.16 ( 1.77%)
Min System 57.20 ( 0.00%) 55.89 ( 2.29%)
Min Elapsed 270.27 ( 0.00%) 266.09 ( 1.55%)
Min CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
Amean User 197.05 ( 0.00%) 194.14 * 1.48%*
Amean System 57.41 ( 0.00%) 56.35 * 1.83%*
Amean Elapsed 270.97 ( 0.00%) 266.90 * 1.50%*
Amean CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
Stddev User 0.25 ( 0.00%) 0.64 (-151.28%)
Stddev System 0.24 ( 0.00%) 0.31 ( -28.73%)
Stddev Elapsed 0.56 ( 0.00%) 0.62 ( -10.17%)
Stddev CPU 0.00 ( 0.00%) 0.00 ( 0.00%)
CoeffVar User 0.13 ( 0.00%) 0.33 (-155.05%)
CoeffVar System 0.41 ( 0.00%) 0.54 ( -31.13%)
CoeffVar Elapsed 0.21 ( 0.00%) 0.23 ( -11.85%)
CoeffVar CPU 0.00 ( 0.00%) 0.00 ( 0.00%)
Max User 197.35 ( 0.00%) 194.92 ( 1.23%)
Max System 57.75 ( 0.00%) 56.64 ( 1.92%)
Max Elapsed 271.66 ( 0.00%) 267.60 ( 1.49%)
Max CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
BAmean-50 User 196.85 ( 0.00%) 193.60 ( 1.65%)
BAmean-50 System 57.20 ( 0.00%) 56.05 ( 2.01%)
BAmean-50 Elapsed 270.40 ( 0.00%) 266.29 ( 1.52%)
BAmean-50 CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
BAmean-95 User 196.98 ( 0.00%) 193.94 ( 1.54%)
BAmean-95 System 57.32 ( 0.00%) 56.28 ( 1.81%)
BAmean-95 Elapsed 270.79 ( 0.00%) 266.72 ( 1.50%)
BAmean-95 CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
BAmean-99 User 196.98 ( 0.00%) 193.94 ( 1.54%)
BAmean-99 System 57.32 ( 0.00%) 56.28 ( 1.81%)
BAmean-99 Elapsed 270.79 ( 0.00%) 266.72 ( 1.50%)
BAmean-99 CPU 93.00 ( 0.00%) 93.00 ( 0.00%)
rc clear_use
rc5 clear_user
Duration User 1182.22 1165.67
Duration System 345.58 338.46
Duration Elapsed 1626.80 1602.99
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
