Re: [RFC PATCH] sys_membarrier(): system/process-wide memory barrier (x86) (v12)

[Peter Zijlstra <peterz@infradead.org>]

On Mon, Mar 16, 2015 at 03:43:56PM +0000, Mathieu Desnoyers wrote:
> > > +enum {
> > > +	/*
> > > +	 * Private flag set: only synchronize across a single process. If this
> > > +	 * flag is not set, it means "shared": synchronize across multiple
> > > +	 * processes.  The shared mode is useful for shared memory mappings
> > > +	 * across processes.
> > > +	 */
> > > +	MEMBARRIER_PRIVATE_FLAG = (1 << 0),
> > > +
> > > +	/*
> > > +	 * Expedited flag set: adds some overhead, fast execution (few
> > > +	 * microseconds).  If this flag is not set, it means "delayed": low
> > > +	 * overhead, but slow execution (few milliseconds).
> > > +	 */
> > > +	MEMBARRIER_EXPEDITED_FLAG = (1 << 1),
> > 
> > 
> > I suppose this is an unprivileged syscall; so what do we do about:
> > 
> > 	for (;;)
> > 		sys_membar(EXPEDITED);
> > 
> > Which would spray the entire system with IPIs at break neck speed.
> 
> Currently, combining EXPEDITED with non-PRIVATE returns -EINVAL.

Ah, tl;dr that bit. This patch really had _too_ much verbiage.

> Therefore, if someone cares about issuing barriers on the entire
> system, the only option is to use non-EXPEDITED, which rely on
> synchronize_rcu().
> 
> The only way to invoke expedited barriers in a loop is:
> 
> for (;;)
>         sys_membarrier(MEMBARRIER_EXPEDITED | MEMBARRIER_PRIVATE);
> 
> Which will only send IPIs to the CPU running threads from the same
> process.

That is indeed less of a problem.