Messages in this thread | | | Date | Fri, 26 Aug 2016 10:48:00 +0100 | From | Will Deacon <> | Subject | Re: [PATCH] arm/perf: Fix pmu percpu irq handling at hotplug. |
| |
Mark,
On Fri, Aug 19, 2016 at 03:25:14PM +0100, Mark Rutland wrote: > On Thu, Aug 18, 2016 at 01:24:38PM -0700, Yabin Cui wrote: > > If the cpu pmu is using a percpu irq: > > > > 1. When a cpu is down, we should disable pmu irq on > > that cpu. Otherwise, if the cpu is still down when > > the last perf event is released, the pmu irq can't > > be freed. Because the irq is still enabled on the > > offlined cpu. And following perf_event_open() > > syscalls will fail. > > > > 2. When a cpu is up, we should enable pmu irq on > > that cpu. Otherwise, profiling tools can't sample > > events on the cpu before all perf events are > > released, because pmu irq is disabled on that cpu. > > It also looks like if a CPU is taken down while events are active, a > non-percpu interrupt will get migrated to another CPU, yet we don't > retarget it if/when the CPU is brought back online. So we have at least > three bugs with IRQ manipulation around hotplug. > > Rather than adding more moving parts to the IRQ manipulation logic, I'd > rather we rework the IRQ manipulation logic to: > > * At probe time, request all the interrupts. If we can't, bail out and > fail the probe. > > * Upon hotplug in (and at probe time), configure the affinity and > enable the relevant interrupt(s). > > * Upon hotplug out, disable the relevant interrupt. > > That way we have fewer moving parts that need to interact with each > other (e.g. we don't need to inhibit hotplug in places), and we know > early whether things will or will not work. > > The {reserve,release}_hardware dance is largely a legacy thing that was > there to cater for sharing the PMU with other subsystems, and we should > be able to get rid of it. > > I'm taking a look at doing the above, but I don't yet have a patch.
Any update on this? I'd quite like to do *something* to fix the issues reported here.
Will
| |