Re: [PATCH] Fix: disable sys_membarrier when nohz_full is enabled
From: Paul E. McKenney
Date: Thu Nov 17 2016 - 13:14:35 EST
On Thu, Nov 17, 2016 at 10:17:25AM -0500, Steven Rostedt wrote:
> On Thu, 17 Nov 2016 15:02:18 +0000 (UTC)
> Mathieu Desnoyers <mathieu.desnoyers@xxxxxxxxxxxx> wrote:
> > That's an interesting approach. I would be tempted to give it a
> > per-thread (rather than per-process) scope.
> Sure, per thread, but have it inherit to child processes.
> > E.g., a thread could do the following to ask to be
> > interrupted by IPIs:
> > membarrier(MEMBARRIER_CMD_REGISTER_EXPEDITED, 0)
> > and could unregister with:
> > membarrier(MEMBARRIER_CMD_UNREGISTER_EXPEDITED, 0)
> Sure why not ;-)
Makes a lot of sense to me!
> > We can then keep a per-thread refcount internally.
> > (not sure the "EXPEDITED" is the right word there...
> > do we want it to be "NOHZ_FULL" instead ?)
> No, it shouldn't mention NOHZ_FULL. Perhaps have all tasks do this
> regardless, even though it will only affect nohz full ones. But in the
> future it may be other tasks as well.
> > Then in membarrier(MEMBARRIER_CMD_SHARED, 0), for each
> > nohz_full cpu, we grab the rq lock, and only send an IPI
> > if the running thread is registered as "expedited".
> Yeah, something like that. That way it wont interrupt tasks that are
> running in no-hz-full and don't care about this syscall.
And this as well!