Re: RFC: call_rcu_outstanding (was Re: WARNING in __mmdrop)

["Paul E. McKenney" <paulmck@linux.ibm.com>]

On Mon, Jul 22, 2019 at 12:32:17PM -0400, Michael S. Tsirkin wrote:
> On Mon, Jul 22, 2019 at 09:25:51AM -0700, Paul E. McKenney wrote:
> > On Mon, Jul 22, 2019 at 12:13:40PM -0400, Michael S. Tsirkin wrote:
> > > On Mon, Jul 22, 2019 at 08:55:34AM -0700, Paul E. McKenney wrote:
> > > > On Mon, Jul 22, 2019 at 11:47:24AM -0400, Michael S. Tsirkin wrote:
> > > > > On Mon, Jul 22, 2019 at 11:14:39AM -0400, Joel Fernandes wrote:
> > > > > > [snip]
> > > > > > > > Would it make sense to have call_rcu() check to see if there are many
> > > > > > > > outstanding requests on this CPU and if so process them before returning?
> > > > > > > > That would ensure that frequent callers usually ended up doing their
> > > > > > > > own processing.
> > > > > > 
> > > > > > Other than what Paul already mentioned about deadlocks, I am not sure if this
> > > > > > would even work for all cases since call_rcu() has to wait for a grace
> > > > > > period.
> > > > > > 
> > > > > > So, if the number of outstanding requests are higher than a certain amount,
> > > > > > then you *still* have to wait for some RCU configurations for the grace
> > > > > > period duration and cannot just execute the callback in-line. Did I miss
> > > > > > something?
> > > > > > 
> > > > > > Can waiting in-line for a grace period duration be tolerated in the vhost case?
> > > > > > 
> > > > > > thanks,
> > > > > > 
> > > > > >  - Joel
> > > > > 
> > > > > No, but it has many other ways to recover (try again later, drop a
> > > > > packet, use a slower copy to/from user).
> > > > 
> > > > True enough!  And your idea of taking recovery action based on the number
> > > > of callbacks seems like a good one while we are getting RCU's callback
> > > > scheduling improved.
> > > > 
> > > > By the way, was this a real problem that you could make happen on real
> > > > hardware?
> > > 
> > > >  If not, I would suggest just letting RCU get improved over
> > > > the next couple of releases.
> > > 
> > > So basically use kfree_rcu but add a comment saying e.g. "WARNING:
> > > in the future callers of kfree_rcu might need to check that
> > > not too many callbacks get queued. In that case, we can
> > > disable the optimization, or recover in some other way.
> > > Watch this space."
> > 
> > That sounds fair.
> > 
> > > > If it is something that you actually made happen, please let me know
> > > > what (if anything) you need from me for your callback-counting EBUSY
> > > > scheme.
> > > 
> > > If you mean kfree_rcu causing OOM then no, it's all theoretical.
> > > If you mean synchronize_rcu stalling to the point where guest will OOPs,
> > > then yes, that's not too hard to trigger.
> > 
> > Is synchronize_rcu() being stalled by the userspace loop that is invoking
> > your ioctl that does kfree_rcu()?  Or instead by the resulting callback
> > invocation?
> 
> Sorry, let me clarify.  We currently have synchronize_rcu in a userspace
> loop. I have a patch replacing that with kfree_rcu.  This isn't the
> first time synchronize_rcu is stalling a VM for a long while so I didn't
> investigate further.

Ah, so a bunch of synchronize_rcu() calls within a single system call
inside the host is stalling the guest, correct?

If so, one straightforward approach is to do an rcu_barrier() every
(say) 1000 kfree_rcu() calls within that loop in the system call.
This will decrease the overhead by almost a factor of 1000 compared to
a synchronize_rcu() on each trip through that loop, and will prevent
callback overload.

Or if the situation is different (for example, the guest does a long
sequence of system calls, each of which does a single kfree_rcu() or
some such), please let me know what the situation is.

							Thanx, Paul