In developing regression tests for cpu hotplug, we discovered something that doesn't seem right in counting interrupts on a per CPU basis. (System in use is a PPC64 with 8 cpus and I haven't tried this on other architectures). Consider this sequence starting with all 8 cpu's on-line: cd /proc cat interrupts > /tmp/intr_all echo 0 > /sys/devices/system/cpu/cpu2/online cat interrupts > /tmp/intr_offline_2 echo 1 > /sys/devices/system/cpu/cpu2/online cat interrupts > /tmp/intr_online_2 When CPU2 is off-lined, the statistics for CPU2 do not appear (expected). However when you look at the before picture (all CPUs present) and after picture (all cpus present after CPU2 re-added), you see that the original data was returned and has incremented: Before all CPUs online (18:): CPU0 CPU1 CPU2 CPU3 CPU4 CPU5 CPU6 CPU7 52999 67105 21879 28100 46722 47599 44370 48786 XICS Edge IPI ------------^^^^^ After CPU2 back online (18:): CPU0 CPU1 CPU2 CPU3 CPU4 CPU5 CPU6 CPU7 53014 67111 21882 28102 46725 47607 44378 48792 XICS Edge IPI ------------^^^^^ Results are similar for other interrupt levels. This does not seem like the correct approach. Shouldn't an added CPU be considered new and stats begin again from zero? Can you please give us guidance on this issue? Is it a bug or expected behavior? -- Mary Edie Meredith Initiative Manager Open Source Development Labs maryedie@xxxxxxxx 503-906-1942
