Hi Thomas, Hi Russell: This RFC is particularly for your attention since it results directly from feedback I've received from both of you, albeit quite a few months ago now.
This patchset demonstrates using FIQ to improve the quality of the PMU trace on ARM systems. To do so it introduces generic changes that allow irqs to be routed to NMI.
This patchset applies on top of my own patchset:
arm: Implement arch_trigger_all_cpu_backtrace http://thread.gmane.org/gmane.linux.kernel/1864829
I think the important points of this set are clear without reference to the previous patchset but the patches will not run (nor apply cleanly) without the previous patchset.
Currently these patches strictly honour a request from Russell to avoid indirection (notifiers, etc) in the ARM default FIQ handler. I have therefore separated the request that an irq be routed to NMI from the installation of a handler for it.
Avoiding indirection does raise some problems though, because it means we arrive in the PMU code without a context pointer. At present I have just added a global variables into the ARM PMU code in order to hold information about irq allocations in a form I can safely read from NMI.
Daniel Thompson (5): arm: irq: Add a __nmi_count stat irq: Allow interrupts to routed to NMI (or similar) irq: gic: Add support for NMI routing arm: perf: Make v7 support FIQ-safe arm: perf: Use FIQ to handle PMU events.
arch/arm/include/asm/hardirq.h | 1 + arch/arm/include/asm/pmu.h | 4 +++ arch/arm/kernel/irq.c | 7 +++- arch/arm/kernel/perf_event.c | 2 +- arch/arm/kernel/perf_event_cpu.c | 35 +++++++++++++++++-- arch/arm/kernel/perf_event_v7.c | 11 ++---- arch/arm/kernel/traps.c | 15 +++++--- drivers/irqchip/irq-gic.c | 75 ++++++++++++++++++++++++++++++---------- include/linux/interrupt.h | 20 +++++++++++ include/linux/irq.h | 2 ++ include/linux/irqchip/arm-gic.h | 8 ++++- kernel/irq/manage.c | 29 ++++++++++++++-- 12 files changed, 169 insertions(+), 40 deletions(-)
-- 1.9.3