Re: [Xenomai-help] awful latencies (~110us)

["CHABAL David" <david.chabal@domain.hid>]

Philippe Gerum a écrit :
> On Thu, 2007-10-18 at 14:38 +0200, CHABAL David wrote:
>> Philippe Gerum a écrit :
>>> On Wed, 2007-10-17 at 17:46 +0200, CHABAL David wrote:
>>>> Philippe Gerum a écrit :
>>>>> On Wed, 2007-10-17 at 16:52 +0200, CHABAL David wrote:
>>>>>>> Grmmff... I think the PIC on this box has some deep troubles; please try
>>>>>>> this patch, in replacement of the previous ipipe_trace_special
>>>>>>> instrumentation:
>>>>>>>
>>>>>> RTD|       8.348|      10.387|      43.647|       0|       6.098| 
>>>>>> 80.440
>>>>>> ---|------------|------------|------------|--------|-------------------------
>>>>>> RTS|       6.098|      10.020|      80.440|       0|    00:17:20/00:17:20
>>>>>> [root@domain.hid bin]#
>>>>>>
>>>>>>
>>>>>> IRQ handling takes 17µs in the worst case.
>>>>>>
>>>>>> It takes a long time to write 3 poor bytes...
>>>>>>
>>>>> Indeed. Each outb to the ISA bus should be somewhere in the 1-2.5 us
>>>>> range depending on the hw, maybe a bit higher in case of contention, but
>>>>> not that much.
>>>>>
>>>>>> May be is it a SMI problem ??? (Globally disable SMI is on)
>>>>>>
>>>>> It's less likely with an ICH-2 chipset.
>>>>>
>>>>> I see that CONFIG_IDEDMA_PCI_AUTO is disabled. Any reason not to use PCI
>>>>> DMA when available for IDE drives with your hw?
>>>>>
>>>> I read this about DMA:
>>>> http://www.rtai.dk/cgi-bin/gratiswiki.pl?DMA_And_Jitter
>>>>
>>> As usual, usage of DMA in real-time situations is a trade-off, and
>>> should be evaluated within the context of the hw at hand. DMA does bus
>>> mastering, but OTOH, forcing PIO raises native Linux latencies, which
>>> also has some drawbacks for real-time kernels implementing RT/non-RT
>>> mode transitions for tasks like Xenomai and RTAI do. For this reason,
>>> you may want to try switching PCI DMA on for your platform, until it
>>> does prove bad latency-wise.
>>>
>> I tested it but the results are almost the same.
>>
> 
> Good, at least you get PCI DMA back.
> 
>> I will try to upgrade my kernel to the 2.6.22 and to test again.
>>
> 
> At this point, it seems reasonable to think that your hw has some
> problems. You may want to confirm this by measuring the time spent in
> acknowledging interrupts on the slave i8259A in mask_and_ack_8259A(), on
> an unpatched 2.6.20 kernel (e.g. using a couple of rdtscll() within a
> single interrupt-free section). If you still reach peaks beyond 20us,
> then the odds of having the very same issue with newer kernels are high.
> 
>> David

Right. Adeos and Xenomai are innocents. Without then and with rdtscll, I get until 35 µs for
an IRQ handling (!).

So, I will change my HW.

Thanks for your help.

David