On Tue, Nov 03, 2009 at 10:37:43AM +0100, Ralf Baechle wrote: > > On Mon, Nov 02, 2009 at 04:12:35PM -0800, David Daney wrote: > > > On SMP systems, the collection of statistics can cause cache line > > > bouncing in the lines associated with the counters. Make the > > > statistics configurable so this can be avoided. Also we need to make > > > the counters atomic_t so that they can be reliably modified on SMP > > > systems. > > > > how about making it a per_cpu thing and avoid the atomic instructions ? > > Working on that. The tricky part is that statistics may end up somewhat > inaccurate when a reader iterates over the per-CPU array of stats to add > them all up. I've not found a solution for that yet but then again it > should not be something which actually would cause pain. Or more useful than a global statistics would be per process stats. How should a useful interface look like? Ralf
