Nicolas Dichtel <nicolas.dich...@6wind.com> wrote:
> Le 13/10/2016 à 22:43, Florian Westphal a écrit :
> > Nicolas Dichtel <nicolas.dich...@6wind.com> wrote:
> >> Le 10/10/2016 à 16:04, Florian Westphal a écrit :
> >>> Nicolas Dichtel <nicolas.dich...@6wind.com> wrote:
> >>>> After commit b87a2f9199ea ("netfilter: conntrack: add gc worker to remove
> >>>> timed-out entries"), netlink conntrack deletion events may be sent with a
> >>>> huge delay. It could be interesting to let the user tweak gc parameters
> >>>> depending on its use case.
> >>> Hmm, care to elaborate?
> >>> I am not against doing this but I'd like to hear/read your use case.
> >>> The expectation is that in almot all cases eviction will happen from
> >>> packet path. The gc worker is jusdt there for case where a busy system
> >>> goes idle.
> >> It was precisely that case. After a period of activity, the event is sent
> >> a long
> >> time after the timeout. If the router does not manage a lot of flows, why
> >> not
> >> trying to parse more entries instead of the default 1/64 of the table?
> >> In fact, I don't understand why using GC_MAX_BUCKETS_DIV instead of using
> >> always
> >> GC_MAX_BUCKETS whatever the size of the table is.
> > I wanted to make sure that we have a known upper bound on the number of
> > buckets we process so that we do not block other pending kworker items
> > for too long.
> I don't understand. GC_MAX_BUCKETS is the upper bound and I agree that it is
> needed. But why GC_MAX_BUCKETS_DIV (ie 1/64)?
> In other words, why this line:
> goal = min(nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV, GC_MAX_BUCKETS);
> instead of:
> goal = GC_MAX_BUCKETS;
Sure, we can do that. But why is a fixed size better than a fraction?
E.g. with 8k buckets and simple goal = GC_MAX_BUCKETS we scan entire
table on every run, currently we only scan 128.
I wanted to keep too many destroy notifications from firing at once
but maybe i was too paranoid...
> > (Or cause too many useless scans)
> > Another idea worth trying might be to get rid of the max cap and
> > instead break early in case too many jiffies expired.
> > I don't want to add sysctl knobs for this unless absolutely needed; its
> > already
> > possible to 'force' eviction cycle by running 'conntrack -L'.
> Sure, but this is not a "real" solution, just a workaround.
> We need to find a way to deliver conntrack deletion events in a reasonable
> delay, whatever the traffic on the machine is.
Agree, but that depends on what 'reasonable' means and what kind of
uneeded cpu churn we're willing to add.
We can add a sysctl for this but we should use a low default to not do
too much unneeded work.
So what about your original patch, but only add
(and also add instant-resched in case entire budget was consumed)?