[BUG] 2.6.37-rc3 massive interactivity regression on ARM

[Peter Zijlstra]

On Thu, 2010-12-09 at 15:35 -0800, Venkatesh Pallipadi wrote:
> 
> Just to make sure, update_rq_clock() always gets called on current
> CPU. Right? 

No, specifically not. If that were the case we wouldn't need the
cross-cpu synced timestamp. Things like load-balancing and
remote-wakeups need to update a remote CPUs clock.

> The pending patches I have optimizes
> account_system_vtime() to use this_cpu_write and friends. Want to make
> sure this change will still keep that optimization relevant.

Ah, good point, remote CPUs updating that will mess with the consistency
of the per-cpu timestamps due to non atomic updates :/

Bugger.. making them atomics will make it even more expensive. /me goes
ponder.