On Tuesday 29 March 2016 10:38:16 Tim Harvey wrote: > On Tue, Mar 29, 2016 at 8:24 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote: > > On Tuesday 29 March 2016 08:10:08 Tim Harvey wrote: > >> Arnd, > >> > >> Right, on the IMX the MSI interrupt is GIC-120 which is also the > >> legacy INTD and I do see that if I happen to put a radio in a slot > >> where due to swizzling its pin1 becomes INTD (GIC-120) the interrupt > >> does fire and the device works. Any other slot using GIC-123 (INTA), > >> GIC-122 (INTB), or GIC-121 (INTC) never fires so its very possible > >> that something in the designware core is masking out the legacy irqs. > > > > Interesting. I was actually expecting the opposite here, having the > > IRQs only work if they are not IntD. > > > > > >> I typically advise our users to 'not' enable MSI because > >> architecturally you can spread 4 distinct legacy irq's across CPU's > >> better than a single shared irq. > > > > That is a very good point, I never understood why we want to enable > > MSI support on any PCI host bridge that just forwards all MSIs > > to a single IRQ line. Originally MSI was meant as a performance > > feature, but there is nothing in this setup that makes things go > > faster, and several things that make it go slower. > > I had a conversation once with Lucas about implementing the shared MSI > interrupt in such a way that its smp affinity could be set to other > CPU's to gain a performance benefit in certain multi-device cases. > > While this is technically possible it would involve creating a softirq > glue between the different handlers but that would add overhead of a > softirq plus potentially waking up another CPU to every IRQ which > would end up adding some overhead to even the simple single-device > case. > > Without any hard data it wasn't clear if this was worth it or if there > was a clean way to provide this as build-time or run-time option. I think it's pretty clear that this would take things from 'somewhat silly' to 'completely bonkers' ;-) Arnd -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
