Re: [RFC][PATCH] sard changes for 2.5.34

[Andrew Morton <akpm@digeo.com>]

Rick Lindsley wrote:
> 
>     OK, that's a start.  I think there was some work done on making
>     kernel_stat percpu as well.
> 
> Yes there's work on a couple of different fronts there.  There is work
> to specifically make disk stats per cpu (actually, I have some 2.4
> patches already I could port), and there is a more general interface
> (statctr_t) which Dipankar Sarma (dipankar@in.ibm.com) is working on
> for 2.5 for stat counters in general which generalizes the per-cpu
> concept.
> 
> Regardless of which route we go, can you suggest a good exercise to
> demonstrate the advantage of per-cpu counters?  It seems intuitive to
> me, but I'm much more comfortable when I have numbers to back me up.

I don't think this is enough to justify a new subsystem like
statctr_t (struct statctr, please).

Looks like we can take the disk stats out of kernel_stat, move all
the vm-related things out of kernel_stat into struct page_state and
what's left of kernel_stat?

        unsigned int per_cpu_user[NR_CPUS],
                     per_cpu_nice[NR_CPUS],
                     per_cpu_system[NR_CPUS];
        unsigned int irqs[NR_CPUS][NR_IRQS];

And that's good, because "kernel statistics" was clearly too
broad a concept.  The above is just one concept: interrupts and
scheduler things.

I'll pull the VM accounting out of there; when you have a
close-to-final patch for the disk stats we can give it a whizz
in the -mm patches and then get all the userspace tools working
nicely against that, OK?

I'm not sure that I want to add 14 more fields to /proc/meminfo.
So a new /proc/vmstat may appear.  We would then have:

/proc/stat		scheduler things
/proc/diskstat		disk things
/proc/vmstat		vm things