Re: [PATCH] trace/hwlat: Do not restart per-cpu threads if they are already running

[Tero Kristo]

Hi,

On 02/03/2023 13:49, Daniel Bristot de Oliveira wrote:
> Hi Tero,
>
> On 3/2/23 08:36, Tero Kristo wrote:
> > Check if the hwlatd thread for the cpu is already running, before
> > starting a new one. This avoids running multiple instances of the same
> > CPU thread on the system. Also, do not wipe the contents of the
> > per-cpu kthread data when starting the tracer, as this can completely
> > forget about already running instances and start new additional per-cpu
> > threads. Fixes issues where fiddling with either the mode of the hwlat
> > tracer or doing cpu-hotplugs messes up the internal book-keeping
> > resulting in stale hwlatd threads.
> Thanks for your patch.
>
> Would you mind explaining how do you hit the problem? that is, how can
> I reproduce the same problem you faced.

For example, this script snippet reproduces it for me every time:

#!/bin/sh
cd /sys/kernel/debug/tracing
echo 0 > tracing_on
echo hwlat > current_tracer
echo per-cpu > hwlat_detector/mode
echo 100000 > hwlat_detector/width
echo 200000 > hwlat_detector/window
echo 200 > tracing_thresh
echo 1 > tracing_on

Another case where something wonky happens is if you offline/online a 
large number of CPUs (which takes a lot of time), and you start/disable 
the hwlat tracer at the same time.

-Tero


> I tried reproducing it by dispatching the hwlat tracer in two instances,
> but the system already blocks me...
>
> [root@vm tracing]# echo hwlat > current_tracer
> [root@vm tracing]# cd instances/
> [root@vm instances]# mkdir hwlat_2
> [root@vm instances]# cd hwlat_2/
> [root@vm hwlat_2]# echo hwlat > current_tracer
> -bash: echo: write error: Device or resource busy
>
> [root@vm hwlat_2]# cd ../../
> [root@vm tracing]# echo nop > current_tracer
> [root@vm tracing]# cd instances/hwlat_2/
> [root@vm hwlat_2]# echo hwlat > current_tracer
> [root@vm hwlat_2]# cd ..
> [root@vm instances]# mkdir hwlat_1
> [root@vm instances]# cd hwlat_1/
> [root@vm hwlat_1]# echo hwlat > current_tracer
> -bash: echo: write error: Device or resource busy
> [root@vm hwlat_1]#
>
> Having a reproducer helps us to think better about the problem.
>
> -- Daniel
>
> > Signed-off-by: Tero Kristo <tero.kristo@xxxxxxxxxxxxxxx>
> > ---
> >   kernel/trace/trace_hwlat.c | 7 ++++---
> >   1 file changed, 4 insertions(+), 3 deletions(-)
> >
> > diff --git a/kernel/trace/trace_hwlat.c b/kernel/trace/trace_hwlat.c
> > index d440ddd5fd8b..c4945f8adc11 100644
> > --- a/kernel/trace/trace_hwlat.c
> > +++ b/kernel/trace/trace_hwlat.c
> > @@ -492,6 +492,10 @@ static int start_cpu_kthread(unsigned int cpu)
> >   {
> >   	struct task_struct *kthread;
> >   
> > +	/* Do not start a new hwlatd thread if it is already running */
> > +	if (per_cpu(hwlat_per_cpu_data, cpu).kthread)
> > +		return 0;
> > +
> >   	kthread = kthread_run_on_cpu(kthread_fn, NULL, cpu, "hwlatd/%u");
> >   	if (IS_ERR(kthread)) {
> >   		pr_err(BANNER "could not start sampling thread\n");
> > @@ -584,9 +588,6 @@ static int start_per_cpu_kthreads(struct trace_array *tr)
> >   	 */
> >   	cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
> >   
> > -	for_each_online_cpu(cpu)
> > -		per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
> > -
> >   	for_each_cpu(cpu, current_mask) {
> >   		retval = start_cpu_kthread(cpu);
> >   		if (retval)