On Tue, 29 Sep 2020 at 19:50, Will Deacon <will@xxxxxxxxxx> wrote:
>
> On Thu, Sep 24, 2020 at 07:55:19PM +0800, Hou Tao wrote:
> > The following is the newest performance data:
> >
> > aarch64 host (4 sockets, 24 cores per sockets)
> >
> > * v4.19.111
> >
> > no writer, reader cn | 24 | 48 | 72 | 96
> > rate of percpu_down_read/percpu_up_read per second |
> > default: use __this_cpu_inc|dec() | 166129572 | 166064100 | 165963448 | 165203565
> > patched: use this_cpu_inc|dec() | 87727515 | 87698669 | 87675397 | 87337435
> > modified: local_irq_save + __this_cpu_inc|dec() | 15470357 | 15460642 | 15439423 | 15377199
> >
> > * v4.19.111+ [1]
> >
> > modified: use this_cpu_inc|dec() + LSE atomic | 8224023 | 8079416 | 7883046 | 7820350
> >
> > * 5.9-rc6
> >
> > no writer, reader cn | 24 | 48 | 72 | 96
> > rate of percpu_down_read/percpu_up_read per second |
> > reverted: use __this_cpu_inc|dec() + revert 91fc957c| 169664061 | 169481176 | 168493488 | 168844423
> > reverted: use __this_cpu_inc|dec() | 78355071 | 78294285 | 78026030 | 77860492
> > modified: use this_cpu_inc|dec() + no LSE atomic | 64291101 | 64259867 | 64223206 | 63992316
> > default: use this_cpu_inc|dec() + LSE atomic | 16231421 | 16215618 | 16188581 | 15959290
> >
> > It seems that enabling LSE atomic has a negative impact on performance under this test scenario.
> >
> > And it is astonished to me that for my test scenario the performance of v5.9-rc6 is just one half of v4.19.
> > The bisect finds the culprit is 91fc957c9b1d6 ("arm64/bpf: don't allocate BPF JIT programs in module memory").
> > If reverting the patch brute-forcibly under 5.9-rc6 [2], the performance will be the same with
> > v4.19.111 (169664061 vs 166129572). I have had the simplified test module [3] and .config attached [4],
> > so could you please help to check what the problem is ?
>
> I have no idea how that patch can be responsible for this :/ Have you
> confirmed that the bisection is not bogus?
>
> Ard, do you have any ideas?
>
Unless the benchmark could be affected by the fact that BPF programs
are now loaded out of direct branching range of the core kernel, I
don't see how that patch could affect performance in this way.
What you could do is revert the patch partially - drop the new alloc
and free routines, but retain the new placement of the module and
kernel areas, which all got shifted around as a result. That would at
least narrow it down, although it looks very unlikely to me that this
change itself is the root cause of the regression.
