On Tue, Jul 05, 2022 at 12:07:25PM +0100, Will Deacon wrote:
> On Tue, Jul 05, 2022 at 11:57:49AM +0100, Will Deacon wrote:
> > On Tue, Jul 05, 2022 at 11:53:22AM +0100, Kajetan Puchalski wrote:
> > > On Mon, Jul 04, 2022 at 10:22:24AM +0100, Kajetan Puchalski wrote:
> > > > On Sat, Jul 02, 2022 at 10:56:51PM +0200, Florian Westphal wrote:
> > > > > > That would make sense, from further experiments I ran it somehow seems
> > > > > > to be related to the number of workers being spawned by stress-ng along
> > > > > > with the CPUs/cores involved.
> > > > > >
> > > > > > For instance, running the test with <=25 workers (--udp-flood 25 etc.)
> > > > > > results in the test running fine for at least 15 minutes.
> > > > >
> > > > > Ok. I will let it run for longer on the machines I have access to.
> > > > >
> > > > > In mean time, you could test attached patch, its simple s/refcount_/atomic_/
> > > > > in nf_conntrack.
> > > > >
> > > > > If mainline (patch vs. HEAD 69cb6c6556ad89620547318439) crashes for you
> > > > > but works with attached patch someone who understands aarch64 memory ordering
> > > > > would have to look more closely at refcount_XXX functions to see where they
> > > > > might differ from atomic_ ones.
> > > >
> > > > I can confirm that the patch seems to solve the issue.
> > > > With it applied on top of the 5.19-rc5 tag the test runs fine for at
> > > > least 15 minutes which was not the case before so it looks like it is
> > > > that aarch64 memory ordering problem.
> > >
> > > I'm CCing some people who should be able to help with aarch64 memory
> > > ordering, maybe they could take a look.
> > >
> > > (re-sending due to a typo in CC, sorry for duplicate emails!)
> >
> > Sorry, but I have absolutely no context here. We have a handy document
> > describing the differences between atomic_t and refcount_t:
> >
> > Documentation/core-api/refcount-vs-atomic.rst
> >
> > What else do you need to know?
>
> Hmm, and I see a tonne of *_inc_not_zero() conversions in 719774377622
> ("netfilter: conntrack: convert to refcount_t api") which mean that you
> no longer have ordering to subsequent reads in the absence of an address
> dependency.
I think the patch above needs auditing with the relaxed behaviour in mind,
but for the specific crash reported here possibly something like the diff
below?
Will
--->8
diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 082a2fd8d85b..5ad9fcc84269 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c
@@ -1394,6 +1394,7 @@ static unsigned int early_drop_list(struct net *net,
* already fired or someone else deleted it. Just drop ref
* and move to next entry.
*/
+ smp_rmb(); /* XXX: Why? */
if (net_eq(nf_ct_net(tmp), net) &&
nf_ct_is_confirmed(tmp) &&
nf_ct_delete(tmp, 0, 0))
