Re: [PATCH v3] arm64: mm: Fix NOMAP page initialization

[Will Deacon]

On Tue, Jan 17, 2017 at 11:00:15AM +0100, Robert Richter wrote:
> On 13.01.17 14:15:00, Robert Richter wrote:
> > On 13.01.17 09:19:04, Will Deacon wrote:
> > > On Thu, Jan 12, 2017 at 07:58:25PM +0100, Robert Richter wrote:
> > > > On 12.01.17 16:05:36, Will Deacon wrote:
> > > > > On Mon, Jan 09, 2017 at 12:53:20PM +0100, Robert Richter wrote:
> > > > 
> > > > > > Kernel compile times (3 runs each):
> > > > > > 
> > > > > > pfn_valid_within():
> > > > > > 
> > > > > > real    6m4.088s
> > > > > > user    372m57.607s
> > > > > > sys     16m55.158s
> > > > > > 
> > > > > > real    6m1.532s
> > > > > > user    372m48.453s
> > > > > > sys     16m50.370s
> > > > > > 
> > > > > > real    6m4.061s
> > > > > > user    373m18.753s
> > > > > > sys     16m57.027s
> > > > > 
> > > > > Did you reboot the machine between each build here, or only when changing
> > > > > kernel? If the latter, do you see variations in kernel build time by simply
> > > > > rebooting the same Image?
> > > > 
> > > > I built it in a loop on the shell, so no reboots between builds. Note
> > > > that I was building the kernel in /dev/shm to not access harddisks. I
> > > > think build times should be comparable then since there is no fs
> > > > caching.
> > > 
> > > I guess I'm really asking what the standard deviation is if you *do* reboot
> > > between builds, using the same kernel. It's hard to tell whether the numbers
> > > are due to the patches, or just because of noise incurred by the way things
> > > happen to initialise.
> > 
> > Ok, I am going to test this.
> 
> See below the data for a test with reboots between every 3 builds (9
> builds per kernel). Though some deviation can be seen between reboots
> there is a trend.

I can't really see the trend given that, for system time, your
pfn_valid_within results have a variance of ~9 and the early_pfn_valid
results have a variance of ~92. Given that the variance seems to come
about due to the reboots, I think we need more numbers to establish whether
the data sets end up largely overlapping or if they really are disjoint.

Will