[ovs-dev] [PATCH ovn v2 0/9] Incremental processing for flow installation.

Han Zhou hzhou at ovn.org
Tue Sep 15 00:29:43 UTC 2020


On Mon, Sep 14, 2020 at 9:54 AM Ilya Maximets <i.maximets at ovn.org> wrote:
>
> On 9/14/20 6:45 PM, Dumitru Ceara wrote:
> > On 9/14/20 6:36 PM, Han Zhou wrote:
> >>
> >>
> >> On Mon, Sep 14, 2020 at 6:04 AM Dumitru Ceara <dceara at redhat.com
> >> <mailto:dceara at redhat.com>> wrote:
> >>>
> >>> On 9/11/20 9:16 PM, Dumitru Ceara wrote:
> >>>> On 9/11/20 8:39 PM, Han Zhou wrote:
> >>>>>
> >>>>>
> >>>>> On Fri, Sep 11, 2020 at 6:17 AM Dumitru Ceara <dceara at redhat.com
> >> <mailto:dceara at redhat.com>
> >>>>> <mailto:dceara at redhat.com <mailto:dceara at redhat.com>>> wrote:
> >>>>>>
> >>>>>> On 9/9/20 11:13 PM, Han Zhou wrote:
> >>>>>>> On Wed, Sep 9, 2020 at 6:39 AM Mark Michelson
> >> <mmichels at redhat.com <mailto:mmichels at redhat.com>
> >>>>> <mailto:mmichels at redhat.com <mailto:mmichels at redhat.com>>> wrote:
> >>>>>>>>
> >>>>>>>> Thanks for addressing my concerns on patch 9.
> >>>>>>>>
> >>>>>>>> Acked-by: Mark Michelson <mmichels at redhat.com
> >> <mailto:mmichels at redhat.com>
> >>>>> <mailto:mmichels at redhat.com <mailto:mmichels at redhat.com>>>
> >>>>>>>>
> >>>>>>>> Warning: I'm also ACKing Numan's patch series that also adds some
> >>>>>>>> lflow-level caching (but he is caching expressions). So whoever
> >> loses
> >>>>>>>> this merge race may have some work to do in order to apply their
> >> patch
> >>>>>>>> cleanly.
> >>>>>>>>
> >>>>>>> Thanks Mark for the review, and the warning :).
> >>>>>>> I applied the series to master, and also the first 4 patches to
> >>>>>>> branch-20.06 and patch 1,2,4 to branch-20.03, to fix the
conjunction
> >>>>> flow
> >>>>>>> bug in those branches.
> >>>>>>>
> >>>>>>> Thanks,
> >>>>>>> Han
> >>>>>>>
> >>>>>>
> >>>>>> Hi Han,
> >>>>>>
> >>>>>> This seems to cause an assertion failure in ovn-controller
> >> compiled with
> >>>>>> today's OVS/OVN master branches. I'm not completely sure about the
> >> steps
> >>>>>> to reproduce because I was testing some other ovn-k8s related
> >> things but
> >>>>>> I did get a core dump and the NB/DB/conf DBs.
> >>>>>>
> >>>>>> I opened a RedHat BZ [0] (more for my tracking than anything else)
> >> and I
> >>>>>> shared there the OVS/OVN revisions and how to build the exact same
> >> image
> >>>>>> that I was running when I hit the crash.
> >>>>>>
> >>>>>> I'll try to investigate what's happening but you're probably more
> >>>>>> familiar with the code so any hints would be great.
> >>>>>>
> >>>>>> Thanks,
> >>>>>> Dumitru
> >>>>>>
> >>>>>> [0] https://bugzilla.redhat.com/show_bug.cgi?id=1878139
> >>>>>>
> >>>>>
> >>>>> Thanks Dumitru for reporting the issue. I reviewed the code and the
> >> test
> >>>>> cases again, and didn't figure out any problem yet. The failed
> >> assertion
> >>>>> "ovs_list_is_empty(&f->list_node) " is to assert that the
desired_flow
> >>>>> is not accessed more than once in a single flood_remove call. This
> >>>>> should be true because the loop at
> >>>>>
> >> (
https://github.com/ovn-org/ovn/blob/master/controller/ofctrl.c#L1126-L1134)
> >>>>> ensures that all flows in to_be_removed (which is the only list that
> >>>>> uses the "list_node" member) is detached from the cross reference
> >> lists,
> >>>>> so that in the following recursive calls these flows will never be
> >>>>> accessed again. Now that this assert fails, it could be some
orphaned
> >>>>> pointer somewhere, leading to the "f" pointing to a flow that's
already
> >>>>> destroyed.
> >>>>>
> >>>>> One thing not clear to me is the version you used for the testing.
> >> In BZ
> >>>>> you mentioned revision: 520189bf313054702f5f802acd7944cca3b6baaa.
> >>>>> However, in this revision, the line number of the assertion doesn't
> >>>>> match the one you pasted in BZ. What you pasted was 1135:
> >>>>
> >>>> The revision should be good, I think the problem is with the way I
> >>>> rebuilt the packages in the new container after I retrieved the core
> >> dump.
> >>>>
> >>>> The assert message clearly points to line 1108. This is from the old
> >>>> logs of the run where I saw the crash:
> >>>>
> >>>> 2020-09-11T11:53:18.426Z|00037|util|EMER|controller/ofctrl.c:1108:
> >>>> assertion ovs_list_is_empty(&f->list_node) failed in
> >>>> flood_remove_flows_for_sb_uuid()
> >>>>
> >>>>>
> >>>>> #6  0x00005607d73502fa in flood_remove_flows_for_sb_uuid
> >>>>> (flow_table=flow_table at entry=0x5607d8c4b380,
> >>>>> sb_uuid=sb_uuid at entry=0x5607d8f91430,
> >>>>> flood_remove_nodes=flood_remove_nodes at entry=0x7ffeccadb9c0) at
> >>>>> controller/ofctrl.c:1135
> >>>>>
> >>>>> What's in the code for revision (520189bf) is 1108:
> >>>>>
> >>
https://github.com/ovn-org/ovn/blob/520189bf313054702f5f802acd7944cca3b6baaa/controller/ofctrl.c#L1108
> >>>>>
> >>>>> Did you change anything while doing the testing?
> >>>>>
> >>>>
> >>>> I'll try to replicate the issue again next week, maybe we get more
info.
> >>>>
> >>>> Thanks,
> >>>> Dumitru
> >>>>
> >>>
> >>> Hi Han,
> >>>
> >>> I managed to minimize the way to replicate the crash (with OVN
master).
> >>> I do:
> >>>
> >>> make sandbox
> >>> ovn-nbctl ls-add ls
> >>> ovn-nbctl lsp-add ls vm1
> >>> ovn-nbctl acl-add lsp-set-addresses vm1 "0a:58:fc:09:1d:4e
> >> fd00:10:244:1::5"
> >>
> >> Thanks Dumitru! Here it seems a copy/paste problem. Could you provide
> >> the ACLs added?
>
> Just skip this line, it's not needed.
> 100% reproducible for me without it:
>
> 2020-09-14T16:52:04.039Z|00020|util|EMER|controller/ofctrl.c:1108:
>   assertion ovs_list_is_empty(&f->list_node) failed in
flood_remove_flows_for_sb_uuid()
>
>
>
> >
> > There are no ACLs involved. This line was a mistake, please ignore it.
> > The conjunctions come from the IPv6 port security logical flows.
> >
> > https://paste.centos.org/view/bb93f7d3
> >

Thanks Dumitru and Ilya. With these clear steps to reproduce, I was able to
find the root cause. It is related to different problems in the history:
1. commit 07e3017 added a check to skip a lflow if lport is not local. The
check is done on every match condition in a loop, but instead I think it
should be done only once per-flow.
2. commit ade4e77 added the resource reference for lport checking at the
same location in the loop, which added redundant references for the same
lport - lflow mapping.
3. commit 580aea7 added flood removing based on the lflow-resource
references, which assumes that there are no redundant items in the lists.
Since the assumption isn't true because of the above 2 problems, it results
in corrupted pointers.

There are different ways to fix the issue. I am trying to figure out the
best way and also working on debugging some regression test failures. I
will post the fix later today.

Thanks,
Han

> > Regards,
> > Dumitru
> >
> >>
> >>> ovn-nbctl lsp-set-addresses vm1 "0a:58:fc:09:1d:4e fd00:10:244:1::5"
> >>> ovn-nbctl lsp-set-port-security vm1 "0a:58:fc:09:1d:4e
fd00:10:244:1::5"
> >>> ovs-vsctl add-port br-int vm1 -- set interface vm1 type=internal --
set
> >>> interface vm1 external-ids:iface-id=vm1
> >>> sleep 1
> >>> ovs-vsctl set interface vm1 external-ids:iface-id=foo
> >>> sleep 1
> >>> ovs-vsctl set interface vm1 external-ids:iface-id=vm1
> >>> sleep 1
> >>> ovs-vsctl set interface vm1 external-ids:iface-id=foo
> >>>
> >>> This crashes ovn-controller and if I look at the openflows before the
> >>> last port unbind, I notice some weird flows with action:
> >>>
> >>>
> >>
actions=conjunction(2,2/2),conjunction(2,2/2),conjunction(2,2/2),conjunction(2,2/2),conjunction(2,2/2),conjunction(2,2/2),conjunction(2,2/2)
> >>>
> >>> This looks broken. I tried to figure out what's happening but
> >>> unfortunately I didn't manage to.
> >>>
> >>> Hope this helps pinpoint the root cause.
> >>>
> >>> Regards,
> >>> Dumitru
> >>>
> >
> > _______________________________________________
> > dev mailing list
> > dev at openvswitch.org
> > https://mail.openvswitch.org/mailman/listinfo/ovs-dev
> >
>


More information about the dev mailing list