Re: [PATCH v2 tip/core/rcu 0/13] Miscellaneous fixes for 4.12

[Paul E. McKenney]

On Wed, Apr 19, 2017 at 05:43:43PM +0200, Peter Zijlstra wrote:
> On Wed, Apr 19, 2017 at 08:37:03AM -0700, Paul E. McKenney wrote:
> > On Wed, Apr 19, 2017 at 03:15:53PM +0200, Peter Zijlstra wrote:
> > > On Wed, Apr 19, 2017 at 06:02:45AM -0700, Paul E. McKenney wrote:
> > > > On Wed, Apr 19, 2017 at 01:28:45PM +0200, Peter Zijlstra wrote:
> > > > > 
> > > > > So the thing Maz complained about is because KVM assumes
> > > > > synchronize_srcu() is 'free' when there is no srcu_read_lock() activity.
> > > > > This series 'breaks' that.
> > > > > 
> > > > > I've not looked hard enough at the new SRCU to see if its possible to
> > > > > re-instate that feature.
> > > > 
> > > > And with the fix I gave Maz, the parallelized version is near enough
> > > > to being free as well.  It was just a stupid bug on my part: I forgot
> > > > to check for expedited when scheduling callbacks.
> > > 
> > > Right, although for the old SRCU it was true for !expedited as well.
> > 
> > Which is all good fun until someone does a call_srcu() on each and
> > every munmap() syscall.  ;-)
> 
> Well, that being a different SRCU domain doesn't affect the KVM memslot
> domain thingy ;-)

Other than the excessive quantities of CPU time consumed...

> > But the current code is much better housebroken.  ;-)
> 
> It is. But a workload that manages to hit sync_expedited in a loop on
> all CPUs is still O(n^2) work. And the more sync_expedited instances we
> have, the more likely that becomes.

In most cases, it shouldn't be -that- hard to loop through the CPUs and
then do a single sync_expedited at the end of the loop.

							Thanx, Paul