Re: [Xenomai] High CPU load using q_send under pSOS skin

[Philippe Gerum <rpm@xenomai.org>]

On 02/06/2014 08:00 PM, Gilles Chanteperdrix wrote:
> On 02/06/2014 06:26 PM, Philippe Gerum wrote:
>>> Every 2.0s: cat /proc/xenomai/stat                      Sat Jan  1
>>> 00:09:03 2000
>>>
>>> CPU  PID    MSW        CSW        PF    STAT       %CPU  NAME
>>>     0  0      0          186436     0     00500080   94.2  ROOT/0
>>>     0  2882   16         427        0     00300380    0.0  MAIN
>>>     0  2883   0          54128      0     00300184    1.0 onemsscheduler
>>>     0  2884   0          542        0     00300184    0.0  FREE
>>>     0  2885   0          1499       0     00300184    0.0  flsh
>>>     0  2886   0          1          0     00300182    0.0  EWRN
>>>     0  2887   0          45158      0     00300186    0.8  EEDT
>>>     0  2888   0          53921      0     00300184    1.4  RESP
>>>     0  2889   1          1          0     00300380    0.0  SERV
>>>     0  2890   0          10784      0     00300184    0.4  CMND
>>>     0  2891   2          1370       0     00300186    0.1  SERP
>>>     0  2896   0          53080      0     00300184    0.8  SERO
>>>     0  2899   0          536        0     00300186    0.0  CANP
>>>     0  2900   0          1          0     00300182    0.0  CANG
>>>     0  2901   0          1          0     00300182    0.0  CANE
>>>     0  2902   0          315        0     00300182    0.0  CANT
>>>     0  2903   1          1          0     00300380    0.0  CNMN
>>>     0  2904   236        628        0     00300184    0.0  CIOW
>>>     0  0      0          699350     0     00000000    1.0  IRQ83:
>>> [timer]
>>>
>>>
>>> So questions are:
>>>
>>> - Why does the %CPU under xenomai/stat not reflect whats reported by top
>>
>> Because top reports the CPU consumed when Linux is active, which
>> excludes Xenomai activity.
>>
>>> (where is the 40% going)?
>>
>> Good question.
>
> I think we have a known issue: the registers passed to Linux timer tick
> are the ones of the last timer tick, which may very well happen to
> always hit the same task. Could this be the cause?
>

I checked your arm-side pipeline implementation yesterday, I believe the 
saved CPSR value does exhibit PSR_I_BIT as expected when the timer 
preempts the Xenomai domain, so there should not be any accounting issue 
(i.e. __ipipe_root_tick_p would prevent update_process_times() to charge 
the last active linux task). This said, the adverse side-effect which is 
solved by the code I just mentioned would rather happen when a Xenomai 
task does frequent relax/harden transitions, not when it runs 
exclusively in real-time mode.

> Is the MAIN task the one which does the tm_wkafter ?
>
> Is the hardware timer of the beaglebone black shared with Linux (like
> omap4) or do linux and xenomai have different timers (like omap3)?
>
> Could you arm the T_WARNSW bit to see if the MAIN task experiences mode
> switches?
>

It would also be interesting to check that commenting out q_send() still 
produces the same result.

-- 
Philippe.