Re: Starvation of one RT task when the runtime of another exceedsperiod.

[Peter Zijlstra]

On Wed, 2008-06-18 at 17:31 +0200, Daniel K. wrote:
> I will demonstrate how to get an RT task stuck, and not rescheduled by
> (ab)using cgroups and RT scheduling. This is on a 4 core system running
> 2.6.26-rc6 with two patches applied to make it work at all.
> 
> http://marc.info/?i=1213732878.3223.95.camel@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
> http://marc.info/?i=1213789854.16944.216.camel@twins
> 
> mkdir /dev/cgroup
> mount -t cgroup -o cpu,cpuset cgroup /dev/cgroup
> 
> # Set up cgroup 0
> mkdir /dev/cgroup/0
> echo 3 > /dev/cgroup/0/cpuset.cpus
> echo 0 > /dev/cgroup/0/cpuset.mems
> echo 100000 > /dev/cgroup/0/cpu.rt_period_us
> echo   5000 > /dev/cgroup/0/cpu.rt_runtime_us
> 
> # Set up cgroup 1
> mkdir /dev/cgroup/1
> echo 3 > /dev/cgroup/1/cpuset.cpus
> echo 0 > /dev/cgroup/1/cpuset.mems
> echo 100000 > /dev/cgroup/1/cpu.rt_period_us
> echo   5000 > /dev/cgroup/1/cpu.rt_runtime_us
> 
> # Start task 1, and assign it to cgroup 0
> schedtool -R -p 1 -e burnP6 &
> [1] 3309
> echo 3309 > /dev/cgroup/0/tasks
> 
> At this point task 1 use 20% CPU.
> 
> # Start task 2, and assign it to cgroup 1
> schedtool -R -p 1 -e burnP6 &
> [2] 3313
> echo 3313 > /dev/cgroup/1/tasks
> 
> At this point task 2 use 20% CPU.
> Both tasks use 40% of CPU core#3 in total.
> 
> # Assign an insane amount of runtime (over 100%, ref. my other mail)
> echo 30000  > /dev/cgroup/1/cpu.rt_runtime_us
> 
> Now, task 2 use 100% of the CPU, and completely starves task 1, which
> ceases to get scheduled.
> 
> # Cut down on the insanity
> echo  5000 > /dev/cgroup/1/cpu.rt_runtime_us
> 
> Now task 2 use only 20% of the CPU again, task 1 does still not get
> scheduled.
> 
> Let's call this state 'stuck'
> 
> I can make task 1 get unstuck by assigning its PID to another cgroup.
> 
> # Kick task 1, so it gets scheduled again.
> echo 3309 > /dev/cgroup/1/tasks
> 
> Assuming we go back to state 'stuck', a 'killall burnP6' will only kill
> task 2, task 1 is still waiting for someone to come and kick it in the
> butt. As soon as that happens, it will get killed as well.
> 
> One time even both tasks got stuck and did not get scheduled, and I
> needed to kick both tasks to get them going again.
> 
> Well, this wasn't really a question, but I'm sure this is not how it's
> supposed to behave?

Humm, indeed, I have a suspicion, but would need to think about this a
bit - again thanks for the detailed report.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/