Messages in this thread | | | Date | Thu, 10 Oct 2013 14:19:08 +0200 | From | Peter Zijlstra <> | Subject | Re: [PATCH 0/6] Optimize the cpu hotplug locking -v2 |
| |
On Wed, Oct 09, 2013 at 10:50:06PM -0700, Andrew Morton wrote: > On Tue, 08 Oct 2013 12:25:05 +0200 Peter Zijlstra <peterz@infradead.org> wrote: > > > The current cpu hotplug lock is a single global lock; therefore excluding > > hotplug is a very expensive proposition even though it is rare occurrence under > > normal operation. > > > > There is a desire for a more light weight implementation of > > {get,put}_online_cpus() from both the NUMA scheduling as well as the -RT side. > > > > The current hotplug lock is a full reader preference lock -- and thus supports > > reader recursion. However since we're making the read side lock much cheaper it > > is the expectation that it will also be used far more. Which in turn would lead > > to writer starvation. > > > > Therefore the new lock proposed is completely fair; albeit somewhat expensive > > on the write side. This in turn means that we need a per-task nesting count to > > support reader recursion. > > This is a lot of code and a lot of new complexity. It needs some pretty > convincing performance numbers to justify its inclusion, no?
And here I thought it was generally understood to be unwise to bash global state on anything like a regular manner from every cpu.
The NUMA bits really ought to use get_online_cpus()/put_online_cpus() on every balance pass; which is about once a second on every cpu.
RT -- which has some quite horrible hotplug hacks due to this -- basically takes get_online_cpus() for every spin_lock/spin_unlock in the kernel.
But the thing is; our sense of NR_CPUS has shifted, where it used to be ok to do something like:
for_each_cpu()
With preemption disabled; it gets to be less and less sane to do so, simply because 'common' hardware has 256+ CPUs these days. If we cannot rely on preempt disable to exclude hotplug, we must use get_online_cpus(), but get_online_cpus() is global state and thus cannot be used at any sort of frequency.
| |