Hide Forgot
My thoughts from a recent discussion about irqbalance: > IMHO irqbalance should have never relied on the names, because the > administrator could always rename the devices any way they liked. > I suspect the problem is that the kernel does not reliably associate > all the IRQs with devices in /sys. There's /sys/class/net/*/device/irq, > but it does not tell anything about the MSI-X interrupts the devices > also use. (... and AFAIK realtime folks also do similar tricks to set priorities of the IRQ threads.)
See https://bugzilla.redhat.com/show_bug.cgi?id=502642 also ... I tried to push something like this upstream. I think that the big issue is that the files associated with the irq must - be unique - each contain ONLY one irq value The problem upstream was then that this results in a large # of files -- and a lot of memory is grabbed because of this ... P.
Neil added commit da8d1c8ba4dcb16d60be54b233deca9a7cac98dc to the upstream kernel that is included in 3.3-rc1 (and he backported it to F16). Does that solve the issue for this bug?
it should, and the latest irqbalance in f16 makes use of that kernel commit to properly identify what kind of device each irq belongs to. Given that, I think this is NOTABUG. Michal?
Yes, the commit implements the discussed feature. Closing.