Re: [RFC] Light weight event counters (V2) instead of page state

[Martin Peschke <mp3@de.ibm.com>]

Christoph Lameter wrote:
> We would like to have some solution that reduces the
> overhead for these counters. VM counters currently require interrupt disabling
> in order to work. Maybe we can avoid that by using the local_t which provides
> an increment operation that is not atomic vs other processors but atomic vs
> interrupts on this same processor.
 >
 > The patchset also adds an off switch for embedded systems that allows a
 > building of linux kernels without these counters.

Did you consider using the statistics infrastructure available in -mm?
(lib/statistic.c, include/linux/statistic.h, Documentation/statistics.txt)

It's a ready-to-use statistics library that comes with similar characteristics
as you have described above.

 > The remaining counters in page_state after the zoned VM counter patch has
 > been applied are all just for show in /proc/vmstat.  They have no essential
 > function for the VM.

To me that looks like a good application for the statistics infrastructure,
which shows statistics through debugfs instead of procfs. Btw., it completely
unburdens exploiting kernel code from the delivery of statistics to users.

	Martin