On Tue, 2017-03-28 at 10:40 +0100, Marc Zyngier wrote: > On 28/03/17 10:12, Benjamin Herrenschmidt wrote: > > On Tue, 2017-03-28 at 09:32 +0100, Marc Zyngier wrote: > > > I'm a bit concerned by this. It means that you can't even mask an > > > interrupt. Is that really what you intend to do? Or all that the HW can > > > do? If you cannot mask an interrupt, you're at the mercy of a screaming > > > device... > > > > This is not really an interrupt controller. It's a "summary" register > > that reflects the state of the 14 i2c controller interrupts. > > > > This approach does have the advantage of providing separate counters in > > /proc/interrupts which is rather nice, but it does have overhead. On > > those shittly little ARMv9 400Mhz cores it can be significant. > > <pedantic> > s/ARMv9/ARM9/, as we're still on variations of the ARMv8 architecture ;-) > </pedantic> It was a typo, I meant ARM9/ARMv5 :-) The 2 SOC families we are talking about (Aspeed 24xx and 25xx) are based on a ARM926EJ at 400Mhz and an ARM1176JZFS at 800Mhz respectively, so cycles do count :-) > A 400MHz ARM9 (which is either ARMv4 or ARMv5) is not too bad (hey, we > still have a couple of Versatile-ABs here...). Caches are pretty small > though. 16K/16K, no L2 :) > > I would personally have some kind of trick to register a single > > interrupt handler that calls directly the handlers of the respective > > i2c busses via a simple indirection for speed, maybe adding my custom > > sysfs or debugfs statistics. But that's just me trying to suck the last > > cycle out of the bloody thing ;-) > > I'd hope the irqdomain itself to be pretty light (the revmap should help > here), but of course you're going to do more work. Counters also come at > a cost. It'd be interesting to see if Brendan has any overhead data > about this. Thankfully, the HW supports buffered sends/receive or even DMA. The current patch doesn't yet support these but they would be a good way to alleviate the cost of the interrupts if it becomes a problem. Cheers, Ben. > Cheers, > > M. -- To unsubscribe from this list: send the line "unsubscribe devicetree" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
