Unusual interrupt latencies with P35+ICH9?

[avh]

We have a real-time application that supports various PCI NI a/d converters (e.g. 6023E) under QNX.  Recently when testing new motherboards with P35+ICH9 chipsets we observed a very long interrupt latency occuring about every 2 seconds.  We didn't see this on 945+ICH7 motherboards.  Does anyone know of something different about the P35+ICH9 that would cause this?

Specifics: a/d converter in PCI slot. Assign slot to IRQ3 in BIOS (also, QNX doesn't support message-signalled interrupts). Turn off everything else in BIOS (mostly USB) hardwired to same INT line as this slot so IRQ3 will not be shared. a/d converter interrupts every 500usec. Observe INTA# of this slot on storage scope. It is normally only 10usec wide (latency). Every 2 seconds it lengthens off the scope.

[Neal M]

Hello avh,

Did you develop the driver for your NI hardware yourself with our Driver Development Kit (DDK), or did you get it from somewhere else?  I ask this because NI does not support QNX directly, but it is possible to write a driver (with the DDK) that will work with it.  I am not terribly familiar with QNX, but from this KnowledgeBase, it appears that there is some custom bus interface code is necessary.  This could be the source of your problem.


Message Edited by Neal M on 04-08-2008 04:11 PM

[avh]

We developed the driver ourselves programming the registers directly (this was before the DDK was available).  The driver has been working for years.  This application works on many motherboards, going back to Pentium 4 machines.  We are just seeing this unusual latency on Gigabyte P35 motherboards.

I'm starting to wonder if some PCI INT is active even though I've turned off everything in the BIOS.  Also, QNX doesn't enable message-signalled interrupts which may be a factor.  I was hoping that the fact it repeats on 2 second intervals would trigger a suggestion from someone.

[Tom_W_[DE]]

Hi avh-

I came across your other post here in which some users suggest that SMM/SMI may be the culprit for the interrupt-handling "downtime" you are seeing.  Several NI groups have wrestled with SMI problems in the past that, in general, resulted in lower and/or unreliable performance for PCI hardware.

My best suggestion would be to contact Gigabyte regarding updated BIOS images that may disable or otherwise optimize SMI.  You may also want to try motherboards from other vendors that use the same chipset but that derive their BIOS implementation from a different reference source.

Hopefully this helps-

[avh]

Thanks for info on SMM.  This does seem likely.  Though I have a logic analyzer, I can only probe the PCI bus with it.  It seems this would not reveal if the cpu was entering SMM.  I have an open thread with Gigabyte and I will ask them about what they might be doing with SMM.

[avh]

I tested three different Gigabyte P35 motherboards that all exhibited the same latency issue.  I tested an MSI P35 motherboard and it did not exhibit the issue.  So I assume the problem is not with the P35 chipset.  SMM seems most likely.

I am going to start testing some industrial ATX motherboards now.  They are more expensive, but these companies guarantee long product life cycles (4-8 years in some cases).