Re: [PATCH v2 3/5] locking/qspinlock: Introduce CNA into the slow path of qspinlock

From: Peter Zijlstra
Date: Tue Apr 02 2019 - 05:43:46 EST


On Mon, Apr 01, 2019 at 10:36:19AM -0400, Waiman Long wrote:
> On 03/29/2019 11:20 AM, Alex Kogan wrote:
> > +config NUMA_AWARE_SPINLOCKS
> > + bool "Numa-aware spinlocks"
> > + depends on NUMA
> > + default y
> > + help
> > + Introduce NUMA (Non Uniform Memory Access) awareness into
> > + the slow path of spinlocks.
> > +
> > + The kernel will try to keep the lock on the same node,
> > + thus reducing the number of remote cache misses, while
> > + trading some of the short term fairness for better performance.
> > +
> > + Say N if you want absolute first come first serve fairness.
> > +
>
> The patch that I am looking for is to have a separate
> numa_queued_spinlock_slowpath() that coexists with
> native_queued_spinlock_slowpath() and
> paravirt_queued_spinlock_slowpath(). At boot time, we select the most
> appropriate one for the system at hand.

Agreed; and until we have static_call, I think we can abuse the paravirt
stuff for this.

By the time we patch the paravirt stuff:

check_bugs()
alternative_instructions()
apply_paravirt()

we should already have enumerated the NODE topology and so nr_node_ids()
should be set.

So if we frob pv_ops.lock.queued_spin_lock_slowpath to
numa_queued_spin_lock_slowpath before that, it should all get patched
just right.

That of course means the whole NUMA_AWARE_SPINLOCKS thing depends on
PARAVIRT_SPINLOCK, which is a bit awkward...