Re: [linus:master] [timers] 7ee9887703: stress-ng.uprobe.ops_per_sec -17.1% regression

[Frederic Weisbecker]

Le Tue, Apr 02, 2024 at 09:46:29AM +0800, Oliver Sang a écrit :
> hi, Frederic Weisbecker,
> 
> On Tue, Apr 02, 2024 at 12:46:15AM +0200, Frederic Weisbecker wrote:
> > Le Wed, Mar 27, 2024 at 04:39:17PM +0800, kernel test robot a écrit :
> > > 
> > > 
> > > Hello,
> > > 
> > > 
> > > we reported
> > > "[tip:timers/core] [timers]  7ee9887703:  netperf.Throughput_Mbps -1.2% regression"
> > > in
> > > https://lore.kernel.org/all/202403011511.24defbbd-oliver.sang@xxxxxxxxx/
> > > 
> > > now we noticed this commit is in mainline and we captured further results.
> > > 
> > > still include netperf results for complete. below details FYI.
> > > 
> > > 
> > > kernel test robot noticed a -17.1% regression of stress-ng.uprobe.ops_per_sec
> > > on:
> > 
> > The good news is that I can reproduce.
> > It has made me spot something already:
> > 
> >    https://lore.kernel.org/lkml/ZgsynV536q1L17IS@xxxxxxxxxxxxx/T/#m28c37a943fdbcbadf0332cf9c32c350c74c403b0
> > 
> > But that's not enough to fix the regression. Investigation continues...
> 
> Thanks a lot for information! if you want us test any patch, please let us know.

The good news is that I can reproduce on two CPUs with just this:

    ./tmp-pkg/stress-ng/src/stress-ng/stress-ng --uprobe-ops 1 --uprobe 2 --timeout 5 --metrics-brief

This reminds me I should try on a single CPU. 

Anyway but the big problem with stress-ng.uprobe is that it consists in
triggering uprobes events and consuming /sys/kernel/tracing/trace_pipe

This makes this testcase nearly impossible to analyse because I can't use any
tracing: the traces are consumed by the testcase. That alone took me quite
some time to figure out.

Then I tried using perf event to do the tracing, as it relies on a different
ring buffer. It works but traces generate ring buffer consumer wake up, which
doesn't work as we are analysing code that depends on idle behaviour.

Then I tried hacking stress-uprobe.c so that the consumed traces are recorded
in a buffer that I writeback in the end. So I can add my own tracepoints in
the flow. And it works but that again doesn't mix up well with tracing idle
behaviour, similar to perf event: the fact that the testcase waits on
/sys/kernel/tracing/trace_pipe produce wake ups from idle while a trace happen
on idle. And that noise ruins the tracing.

So I'm kind of running out of options for now :-/