Paolo Mantegazza wrote:
Notice that max jitters should be estabilished on every long testing times. I would say that using the ard hcli/sti stuff might appear better on not so long times but it could end in being much the same as their soft flags counterparts (Stodolsky) if tested long enough. Said in short I've no perception or evidence that the hold RTAHL way is better than using ADEOS as it is in vesuvio now, clearly on the machines I'm using.
Of course, this requires a more detailed test than I just did by running both scenarios a few minutes.
However, for me a conceptual question still remains: if adeos only defers irqs on demand, in theory, every irq source can cause at least once per critical section this deference. When thinking of, say, two, three, or even more irqs intercepting a critical section, I have a real problem imagining that this shall be better than a simple pair of hard irq off/on. And, as your test systems are rather fast and thus the effective section durations are short, the probability that such a irq rush may occur in practice is a bit lower than on slower machines. So, testing and measuring is one thing, but the concept still creates some question marks for me.
Jan
