Re: [Xenomai-help] Big differences in analogy interrupt handling time

[Anders Blomdell <anders.blomdell@domain.hid>]

On 12/15/2011 06:28 PM, Fernando Herrero Carrón wrote:
> 2011/12/15 Anders Blomdell<anders.blomdell@domain.hid>
>
> Hi,
>>
>> I'm trying to tweak out good performance with Analogy and a NI-6221 card,
>> and after getting weird results with the attached program (long and very
>> varying interrupt latencies). The machine I'm running on is a 'Intel(R)
>> Core(TM)2 CPU 6420 @ 2.13GHz', no frequency scaling, no X, low load,  only
>> one CPU enabled to minimize influence from other Linux tasks. As can be
>> seen from the frozen traces (formatted and time shifted to be examined with
>> 'pr -mtT -w 200 A B'), the time spent in the interrupt handling varies
>> between 263 and 329 us and there is variation even in repeated code paths
>> [which are bogus in themselves, why busy-wait 10 us in an interrupt
>> routine?], so cache poisoning should not be the issue.
>>
>> Anybody that has any insights to share?
>>
>>
> Hi everyone there,
>
> In fact, I came accross this same issue some months ago, but haven't had
> time to write on the list up to now.
>
> Actually, there seems to be some old inherited code from the early comedi
> days. In the file ksrc/drivers/analogy/national_instruments/mio_common.c
> there is a function that handles transfers called "ni_handle_eos()". You
> can see:
>
> static void ni_handle_eos(a4l_subd_t *subd)
> {
>          a4l_dev_t *dev = subd->dev;
>
>          if (devpriv->aimode == AIMODE_SCAN) {
>                  static const int timeout = 10;
>                  int i;
>
>                  for (i = 0; i<  timeout; i++) {
>                          ni_sync_ai_dma(subd);
>                          /* TODO: stop when the transfer is really over */
>                          a4l_udelay(1);
>                  }
>          }
>
>          /* Handle special case of single scan using AI_End_On_End_Of_Scan */
>          if ((devpriv->ai_cmd2&  AI_End_On_End_Of_Scan)) {
>                  shutdown_ai_command(subd);
>          }
> }
>
> Just remove the "for" loop and the a4l_udelay, and leave
> ni_sync_ai_dma(subd):
>
> static void ni_handle_eos(a4l_subd_t *subd)
> {
>          a4l_dev_t *dev = subd->dev;
>
>          if (devpriv->aimode == AIMODE_SCAN) {
>                  static const int timeout = 10;
>                  int i;
timeout and i should also be removed.

>
>                  ni_sync_ai_dma(subd);
>          }
>
>          /* Handle special case of single scan using AI_End_On_End_Of_Scan */
>          if ((devpriv->ai_cmd2&  AI_End_On_End_Of_Scan)) {
>                  shutdown_ai_command(subd);
>          }
> }
>
> This is working for me, for the time being.
>
> Hope this helps,
Nice that somebody has already tested this.

Nobody that has any insight on what might slow down execution with 20% 
(curious mind wants to know)?

/Anders

-- 
Anders Blomdell                  Email: anders.blomdell@domain.hid
Department of Automatic Control
Lund University                  Phone:    +46 46 222 4625
P.O. Box 118                     Fax:      +46 46 138118
SE-221 00 Lund, Sweden