Hello,
Actually, this discussion has become useless since it is now based on Paul Brook engaging in what I believe is called "hit and run" or "petty flogging" in rhetoric. Unfortunately such substitutes for useful discussions tend to start on mailing lists.
I'm hoping some experience might be more useful than what you call rhetorics. I've worked with the SEPIA system for three years. SEPIA is a (now cancelled) project from the HP labs, which involves two xc2v2000 on a PCI board, one of them handling the PCI with a custom PCI core and DMA controller, so the situation is quite close to what will happen in OGP (we're talking 64-bits PCI @66MHz in SEPIA). I have some experience tickling with the SEPIA board linux driver -- I ported it to the 2.6 kernel. Three of these cards SEPIA cards are currently plugged into a hw9300 HP workstation at my desk. This is a dual chip, quad-core Opteron workstation from HP. It was very high-end when we bought it, and could still be considered so. Well, it just happens that the interrupts from one of the boards are *shared*. with a network controller. By shared, I mean that the interrupt handler of the card is called unnecessarily by the kernel PCI code, without the actual IRQ line being raised by the card. This *may* be adressable. But you can't assume it will, because it does not really depend on you. The software layers of the linux kernel are written in such a way as to cope with the lowest denominator (RTFS) -- and I'd advise to do the same with your hardware. That is, assume that it will plugged into a low-end system and work there, and possibly take advantage of newer features, but without sacrifycing compatibility. What are we talking about here anyway ? You need an acknowledgement of the IRQ anyway, so there will be a write on the PCI. We're talking about saving a couple microseconds for reading an IRQ status register on PCI, this against alienating customers, having to delve into low-level linux PCI code and generally having a headache about debugging the stuff. This discussion is really silly indeed. JB _______________________________________________ Open-graphics mailing list [email protected] http://lists.duskglow.com/mailman/listinfo/open-graphics List service provided by Duskglow Consulting, LLC (www.duskglow.com)
