NO_HZ and ktimersoftd

NO_HZ and ktimersoftd

[Federico Reghenzani]

Hi all,

I'm trying to setup a RT Linux system (currently on 4.8.2 PREEMPT_RT
patched), and I enabled he NO_HZ_FULL flag in order to reduce the
scheduling interrupts jitter. Trying to isolate one core, I also moved
all interrupts and threads to other CPUs.

This works fine, however when I run my RT application, the thread
kimersoftd is always in "running" state, thus the scheduling interrupt
triggers (because there exists more than 1 task available to run).
Consequently, the advantages of NO_HZ_FULL are vanished.

With kernel trace I found these events on the isolated CPU:

       user_loop-772   [003] d..h...   419.818172: softirq_raise:
vec=1 [action=TIMER]
       user_loop-772   [003] d...2..   419.818174: sched_switch:
prev_comm=user_loop prev_pid=772 prev_prio=120 prev_state=R ==>
next_comm=ktimersoftd/3 next_pid=41 next_prio=98
   ktimersoftd/3-41    [003] .....11   419.818175: softirq_entry:
vec=1 [action=TIMER]
   ktimersoftd/3-41    [003] .....11   419.818176: softirq_exit: vec=1
[action=TIMER]
   ktimersoftd/3-41    [003] d...2..   419.818177: sched_switch:
prev_comm=ktimersoftd/3 prev_pid=41 prev_prio=98 prev_state=S ==>
next_comm=user_loop next_pid=772 next_prio=120

I don't know why ktimersoftd is running at every tick.


Thank you in advance for any hint,
Federico

Re: NO_HZ and ktimersoftd

[Julia Cartwright]

On Thu, Jan 05, 2017 at 10:29:57AM +0100, Federico Reghenzani wrote:
> Hi all,
>

Hello-

> I'm trying to setup a RT Linux system (currently on 4.8.2 PREEMPT_RT
> patched), and I enabled he NO_HZ_FULL flag in order to reduce the
> scheduling interrupts jitter. Trying to isolate one core, I also moved
> all interrupts and threads to other CPUs.
> 
> This works fine, however when I run my RT application, the thread
> kimersoftd is always in "running" state, thus the scheduling interrupt
> triggers (because there exists more than 1 task available to run).
> Consequently, the advantages of NO_HZ_FULL are vanished.

This sounds like the same problem that Haris (cc'd) has looked into and
has put together a proposed solution.  See the thread here:

   http://www.spinics.net/lists/linux-rt-users/msg16113.html

   Julia

Re: NO_HZ and ktimersoftd

[Sebastian Andrzej Siewior]

On 2017-01-05 10:32:10 [-0600], Julia Cartwright wrote:
> > This works fine, however when I run my RT application, the thread
> > kimersoftd is always in "running" state, thus the scheduling interrupt
> > triggers (because there exists more than 1 task available to run).
> > Consequently, the advantages of NO_HZ_FULL are vanished.
> 
> This sounds like the same problem that Haris (cc'd) has looked into and
> has put together a proposed solution.  See the thread here:
> 
>    http://www.spinics.net/lists/linux-rt-users/msg16113.html

can you please check latest v4.9.11-rt9 if the issue is still there or
gone?

Sebastian

Re: NO_HZ and ktimersoftd

[Federico Reghenzani]

2017-02-23 19:01 GMT+01:00 Sebastian Andrzej Siewior <bigeasy@linutronix.de>:
> On 2017-01-05 10:32:10 [-0600], Julia Cartwright wrote:
>> > This works fine, however when I run my RT application, the thread
>> > kimersoftd is always in "running" state, thus the scheduling interrupt
>> > triggers (because there exists more than 1 task available to run).
>> > Consequently, the advantages of NO_HZ_FULL are vanished.
>>
>> This sounds like the same problem that Haris (cc'd) has looked into and
>> has put together a proposed solution.  See the thread here:
>>
>>    http://www.spinics.net/lists/linux-rt-users/msg16113.html
>
> can you please check latest v4.9.11-rt9 if the issue is still there or
> gone?
>
> Sebastian

It seems fixed, thank you


Cheers,
Federico