On Tue, 2015-07-14 at 14:56 +0530, Shivaprasad bhat wrote:
As you could see in the series I referenced - there are a number of nodeinfo.c API's which don't process the sysfs properly, e.g. they assume /sys/devices/system.
I haven't been fully convinced that the patch which ends up as patch9 in my series won't have some sort of negative affect somewhere down the line. Consider if your "*/cpu/present" contained "0-47,64-95" instead of "0-95" - what "expectations" would you have in this patch series?
The point being if the expectation is that 48-63 would/should have some specific state and they don't, then I can certainly see the need for the other patch. Since you had a reason to be in the code, I figured to pick your brain over this logic while the code was still fresh in your mind! It's more a datapoint for the need of the filtering patch.
Once that patch is in place, the call added by that patch to nodeGetPresentCPUBitmap could certainly have altered results if the PPC64 host it was running on didn't have 96 CPU's.
Thanks a lot for pointing out John. I am planning to test the patch on such configuration and see how it goes. I expect, as you mentioned to discard the offline cpus 48-63 during counting.
FWIW, now that John's series has been merged I'm going to rebase my version of the patch on top of it and post if for review. Shouldn't take long. Cheers. -- Andrea Bolognani Software Engineer - Virtualization Team