Re: [PATCH -mmotm 1/5] memcg: disable irq at page cgroup lock

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2010-04-23 at 13:17 -0700, Greg Thelen wrote:
> 
> This is an interesting idea.  If this applies to memcg dirty accounting,
> then would it also apply to system-wide dirty accounting?  I don't think
> so, but I wanted to float the idea.  It looks like this proportions.c
> code is good is at comparing the rates of events (for example: per-task
> dirty page events).  However, in the case of system-wide dirty
> accounting we also want to consider the amount of dirty memory, not just
> the rate at which it is being dirtied.

Correct, the whole proportion thing is purely about comparing rates of
events.

> The performance of simple irqsave locking or more advanced RCU locking
> is similar to current locking (non-irqsave/non-rcu) for several
> workloads (kernel build, dd).  Using a micro-benchmark some differences
> are seen:
> * irqsave is 1% slower than mmotm non-irqsave/non-rcu locking.
> * RCU locking is 4% faster than mmotm non-irqsave/non-rcu locking.
> * RCU locking is 5% faster than irqsave locking.

Depending on what architecture you care about local_t might also be an
option, it uses per-cpu irq/nmi safe instructions (and falls back to
local_irq_save/restore for architectures lacking this support).