Re: [PATCH RFC v1 1/2] rcu/tree: Add basic support for kfree_rcu batching

[Joel Fernandes]

On Thu, Aug 08, 2019 at 08:56:07AM -0400, Joel Fernandes wrote:
> On Thu, Aug 08, 2019 at 06:52:32PM +0900, Byungchul Park wrote:
> > On Wed, Aug 07, 2019 at 10:52:15AM -0700, Paul E. McKenney wrote:
> > > > > On Tue, Aug 06, 2019 at 05:20:40PM -0400, Joel Fernandes (Google) wrote:
> > > [ . . . ]
> > > > > > +	for (; head; head = next) {
> > > > > > +		next = head->next;
> > > > > > +		head->next = NULL;
> > > > > > +		__call_rcu(head, head->func, -1, 1);
> > > > > 
> > > > > We need at least a cond_resched() here.  200,000 times through this loop
> > > > > in a PREEMPT=n kernel might not always be pretty.  Except that this is
> > > > > invoked directly from kfree_rcu() which might be invoked with interrupts
> > > > > disabled, which precludes calls to cond_resched().  So the realtime guys
> > > > > are not going to be at all happy with this loop.
> > > > 
> > > > Ok, will add this here.
> > > > 
> > > > > And this loop could be avoided entirely by having a third rcu_head list
> > > > > in the kfree_rcu_cpu structure.  Yes, some of the batches would exceed
> > > > > KFREE_MAX_BATCH, but given that they are invoked from a workqueue, that
> > > > > should be OK, or at least more OK than queuing 200,000 callbacks with
> > > > > interrupts disabled.  (If it turns out not to be OK, an array of rcu_head
> > > > > pointers can be used to reduce the probability of oversized batches.)
> > > > > This would also mean that the equality comparisons with KFREE_MAX_BATCH
> > > > > need to become greater-or-equal comparisons or some such.
> > > > 
> > > > Yes, certainly we can do these kinds of improvements after this patch, and
> > > > then add more tests to validate the improvements.
> > > 
> > > Out of pity for people bisecting, we need this fixed up front.
> > > 
> > > My suggestion is to just allow ->head to grow until ->head_free becomes
> > > available.  That way you are looping with interrupts and preemption
> > > enabled in workqueue context, which is much less damaging than doing so
> > > with interrupts disabled, and possibly even from hard-irq context.
> > 
> > Agree.
> > 
> > Or after introducing another limit like KFREE_MAX_BATCH_FORCE(>=
> > KFREE_MAX_BATCH):
> > 
> > 1. Try to drain it on hitting KFREE_MAX_BATCH as it does.
> > 
> >    On success: Same as now.
> >    On fail: let ->head grow and drain if possible, until reaching to
> >             KFREE_MAX_BATCH_FORCE.
> > 
> > 3. On hitting KFREE_MAX_BATCH_FORCE, give up batching but handle one by
> >    one from now on to prevent too many pending requests from being
> >    queued for batching work.
> 
> I also agree. But this _FORCE thing will still not solve the issue Paul is
> raising which is doing this loop possibly in irq disabled / hardirq context.
> We can't even cond_resched() here. In fact since _FORCE is larger, it will be
> even worse. Consider a real-time system with a lot of memory, in this case
> letting ->head grow large is Ok, but looping for long time in IRQ disabled
> would not be Ok.
> 
> But I could make it something like:
> 1. Letting ->head grow if ->head_free busy
> 2. If head_free is busy, then just queue/requeue the monitor to try again.
> 
> This would even improve performance, but will still risk going out of memory.

It seems I can indeed hit an out of memory condition once I changed it to
"letting list grow" (diff is below which applies on top of this patch) while
at the same time removing the schedule_timeout(2) and replacing it with
cond_resched() in the rcuperf test.  I think the reason is the rcuperf test
starves the worker threads that are executing in workqueue context after a
grace period and those are unable to get enough CPU time to kfree things fast
enough. But I am not fully sure about it and need to test/trace more to
figure out why this is happening.

If I add back the schedule_uninterruptibe_timeout(2) call, the out of memory
situation goes away.

Clearly we need to do more work on this patch.

In the regular kfree_rcu_no_batch() case, I don't hit this issue. I believe
that since the kfree is happening in softirq context in the _no_batch() case,
it fares better. The question then I guess is how do we run the rcu_work in a
higher priority context so it is not starved and runs often enough. I'll
trace more.

Perhaps I can also lower the priority of the rcuperf threads to give the
worker thread some more room to run and see if anything changes. But I am not
sure then if we're preparing the code for the real world with such
modifications.

Any thoughts?

thanks,

 - Joel

---8<-----------------------