RE: [PATCH] rt-tests: pi_stress: fix testing threads' smp affinity

[John Kacur <jkacur@redhat.com>]

On Fri, 7 May 2021, Jiafei Pan wrote:

> Any comments? thanks.
> 
> Best Regards,
> Jiafei.
> 
> > -----Original Message-----
> > From: Jiafei Pan <jiafei.pan@nxp.com>
> > Sent: Friday, March 26, 2021 5:44 PM
> > To: williams@redhat.com; jkacur@redhat.com
> > Cc: linux-rt-users@vger.kernel.org; Jiafei Pan <jiafei.pan@nxp.com>; Jiafei
> > Pan <jiafei.pan@nxp.com>
> > Subject: [PATCH] rt-tests: pi_stress: fix testing threads' smp affinity
> > 
> > This patch includes the following modification:
> > 1. Make sure test threads and admin threads don't run on the same CPU
> >    Core if uniprocessor is not set or not on single Core platform to
> >    avoid starve admin threads.

Doesn't the code already do this? The one exception I can think of is you 
are allowed one group of threads per online processor, so if you have the 
maximum number of groups, then one of those groups will run on the same 
processor as the admin thread. Even in this case though, the admin_thread
has the highest realtime priority.

> > 2. Force to use SCHED_RR if more than one Groups running one a CPU
> >    Core to avoid test failure because threads in different Groups
> >    are using the same priority, SCHED_FIFO which is default policy
> >    and it maybe trigger deadlock of testing threads.

Like a lot of things in realtime, we give you enough rope to hang 
yourself. We are trying to create a tool here that purposely triggers 
deadlocks that are resolved via priority inheritence in order to test 
priority inheritence, we're not trying to create a robust application that 
has safeguards to avoid this situation. So, with that in mind, why would 
we force the teste to use SCHED_RR if the user wants SCHED_FF ?

Trying to understand what you are trying to accomplish here. Is it that 
you want to get rid of the limitation of only allowing one inversion group 
per processor? If so, why is this useful?

Thanks

John

> > 
> > Signed-off-by: Jiafei Pan <Jiafei.Pan@nxp.com>
> > ---
> >  src/pi_tests/pi_stress.c | 23 +++++++++++++++++++----
> >  1 file changed, 19 insertions(+), 4 deletions(-)
> > 
> > diff --git a/src/pi_tests/pi_stress.c b/src/pi_tests/pi_stress.c index
> > 49f89b7..8795908 100644
> > --- a/src/pi_tests/pi_stress.c
> > +++ b/src/pi_tests/pi_stress.c
> > @@ -237,6 +237,13 @@ int main(int argc, char **argv)
> >  	/* process command line arguments */
> >  	process_command_line(argc, argv);
> > 
> > +	if (ngroups > (num_processors - 1)) {
> > +		printf("Warning: One Core will used for administor thread, the other
> > CPU Core will run test thread,\n");
> > +		printf("\t it will running more than one Group on one Core (groups >
> > num_of_processors -1),\n");
> > +		printf("\t it will force to use SCHED_RR, or change groups number to
> > be lower than %ld \n", num_processors);
> > +		policy = SCHED_RR;
> > +	}
> > +
> >  	/* set default sched attributes */
> >  	setup_sched_config(policy);
> > 
> > @@ -285,9 +292,17 @@ int main(int argc, char **argv)
> >  			break;
> >  	for (i = 0; i < ngroups; i++) {
> >  		groups[i].id = i;
> > -		groups[i].cpu = core++;
> > -		if (core >= num_processors)
> > -			core = 0;
> > +		if (num_processors == 1 || uniprocessor) {
> > +			groups[i].cpu = 0;
> > +		} else {
> > +			groups[i].cpu = core;
> > +			/* Find next non-admin Core */
> > +			do {
> > +				core++;
> > +				if (core >= num_processors)
> > +					core = 0;
> > +			} while (CPU_ISSET(core, &admin_cpu_mask));
> > +		}
> >  		if (create_group(&groups[i]) != SUCCESS)
> >  			return FAILURE;
> >  	}
> > @@ -1143,7 +1158,7 @@ int create_group(struct group_parameters *group)
> >  	CPU_ZERO(&mask);
> >  	CPU_SET(group->cpu, &mask);
> > 
> > -	pi_debug("group %d bound to cpu %ld\n", group->id, group->cpu);
> > +	printf("group %d bound to cpu %ld\n", group->id, group->cpu);
> > 
> >  	/* start the low priority thread */
> >  	pi_debug("creating low priority thread\n");
> > --
> > 2.17.1
> 
>