Re: [RFC PATCH] percpu system call: fast userspace percpu critical sections

[Andrew Hunter <ahh-hpIqsD4AKlfQT0dZR+AlfA@public.gmane.org>]

On Fri, May 22, 2015 at 1:53 PM, Andy Lutomirski <luto-kltTT9wpgjJwATOyAt5JVQ@public.gmane.org> wrote:
> Create an array of user-managed locks, one per cpu.  Call them lock[i]
> for 0 <= i < ncpus.
>
> To acquire, look up your CPU number.  Then, atomically, check that
> lock[cpu] isn't held and, if so, mark it held and record both your tid
> and your lock acquisition count.  If you learn that the lock *was*
> held after all, signal the holder (with kill or your favorite other
> mechanism), telling it which lock acquisition count is being aborted.
> Then atomically steal the lock, but only if the lock acquisition count
> hasn't changed.
>

We had to deploy the userspace percpu API (percpu sharded locks,
{double,}compare-and-swap, atomic-increment, etc) universally to the
fleet without waiting for 100% kernel penetration, not to mention
wanting to disable the kernel acceleration in case of kernel bugs.
(Since this is mostly used in core infrastructure--malloc, various
statistics platforms, etc--in userspace, checking for availability
isn't feasible. The primitives have to work 100% of the time or it
would be too complex for our developers to bother using them.)

So we did basically this (without the lock stealing...): we have a
single per-cpu spin lock manipulated with atomics, which we take very
briefly to implement (e.g.) compare-and-swap.  The performance is
hugely worse; typical overheads are in the 10x range _without_ any
on-cpu contention. Uncontended atomics are much cheaper than they were
on pre-Nehalem chips, but they still can't hold a candle to
unsynchronized instructions.

As a fallback path for userspace, this is fine--if 5% of binaries on
busted kernels aren't quite as fast, we can work with that in exchange
for being able to write a percpu op without worrying about what to do
on -ENOSYS.  But it's just not fast enough to compete as the intended
way to do things.

AHH