Messages in this thread | | | Date | Sat, 4 Jul 2015 14:49:31 +0200 (CEST) | From | Thomas Gleixner <> | Subject | Re: [PATCH v2 2/2] irqchip: dw-apb-ictl: add irq_set_affinity support |
| |
On Sat, 4 Jul 2015, Russell King - ARM Linux wrote: > On Sat, Jul 04, 2015 at 11:53:57AM +0200, Thomas Gleixner wrote: > > On Sat, 4 Jul 2015, Russell King - ARM Linux wrote: > > > > > On Sat, Jul 04, 2015 at 01:19:30PM +0800, Jisheng Zhang wrote: > > > > On Marvell Berlin SoCs, the cpu's local timer is shutdown when the cpu > > > > goes to a deep idle state, then the timer framework will be notified to > > > > use a broadcast timer instead. The broadcast timer uses dw-apb-ictl as > > > > interrupt chip, this patch adds irq_set_affinity support so that the > > > > going to deep idle state cpu can set the interrupt affinity of the > > > > broadcast interrupt to avoid unnecessary wakeups and IPIs. > > > > > > NAK to this patch. > > > > > > The real question is - if CPU0 is the CPU going offline, why is it > > > still receiving _any_ interrupts - all interrupts should be migrated > > > off it, including the chained interrupts. > > > > > > Sounds like there's a bug in the migration code which needs further > > > investigation, rather than hacking around the problem by introducing > > > lots of driver code. > > > > I think you misunderstood the changelog, which is horrible btw. > > > > So the real reason to do this is to steer the broadcast interrupt to > > the CPU which has the earliest expiry time. This avoids that another > > cpu is woken from idle just to deliver the broadcast IPI to the other > > cpu. > > Unless I'm mistaken, the code does this by messing around with the parent > interrupt affinity of a chained interrupt, which really isn't a good thing > to do, because it migrates every interrupt on the child interrupt > controller.
Fair enough, I missed that chained hackery.
For that powersaving scenario it's probably ok to move the all child irqs around, but we should at least make that an opt-in behaviour and not enabled by default.
Thanks,
tglx
| |