On Mon, Nov 09, 2020 at 02:44:27PM +0100, Dumitru Ceara wrote: > On 11/6/20 4:36 AM, Ben Pfaff wrote: > > Many of these could be replaced by "ovn-nbctl sync". Some weren't > > really needed at all because they were adjacent to something that itself > > called sync or otherwise used --wait. Some were more appropriately > > done with explicit waits for what was really needed. > > > > I left some "sleep"s. Some were because they were "negative" sleeps: > > they were giving time for something to happen that should *not* happen > > (in other words, if you wait for it to happen, you'll wait forever, > > unless there's a bug). Some were because I didn't know how to properly > > wait for what they were waiting for, or because I didn't understand > > the circumstances deeply enough. > > > > Signed-off-by: Ben Pfaff <b...@ovn.org> > > --- > > tests/ovn-northd.at | 7 ++ > > tests/ovn.at | 167 ++++++++++++-------------------------------- > > 2 files changed, 52 insertions(+), 122 deletions(-) > > > > diff --git a/tests/ovn-northd.at b/tests/ovn-northd.at > > index 949a8eee054e..5d670233561e 100644 > > --- a/tests/ovn-northd.at > > +++ b/tests/ovn-northd.at > > @@ -1120,6 +1120,7 @@ ovn-nbctl --wait=sb -- --id=@hc create \ > > Load_Balancer_Health_Check vip="10.0.0.10\:80" -- add Load_Balancer . \ > > health_check @hc > > wait_row_count Service_Monitor 2 > > +check ovn-nbctl --wait=hv sync > > This should be "--wait=sb". ovn-controller is not started for tests in > ovn-northd.at and "--wait=hv" would return immediately. This applies to > all "ovn-nbctl --wait=hv sync" in ovn-northd.at.
I see what you mean. This is surprising to me. When I defined this stuff years ago, my intent was that --wait=hv would be stronger than --wait=sb. That is, --wait=sb would cause ovn-nbctl to wait for the southbound database to catch up with the northbound changes, and --wait=hb would cause it to wait for the southbound database AND the hypervisors to catch up. That's even how it's documented for ovn-nbctl: <p> With <code>--wait=sb</code>, before <code>ovn-nbctl</code> exits, it waits for <code>ovn-northd</code> to bring the southbound database up-to-date with the northbound database updates. </p> <p> With <code>--wait=hv</code>, before <code>ovn-nbctl</code> exits, it additionally waits for all OVN chassis (hypervisors and gateways) to become up-to-date with the northbound database updates. (This can become an indefinite wait if any chassis is malfunctioning.) </p> That's not what actually happens, though. As you point out, if there are no hypervisors, then the hypervisors are always up-to-date, even if the database is not. I think this is a bug in ovn-nbctl. Arguably, it's a bug in the database definition (perhaps hv_cfg shouldn't even be filled in but left empty if there are no chassis) but I think it is too late to fix it there. It is, however, easy enough to fix it in ovn-nbctl. (And at the same time, I'll change these in ovn-northd to just say --wait=sb. You have a point.) How about this as an additional patch? -8<--------------------------cut here-------------------------->8-- >From 7b373c22dbda2f808f3d5f3b8ae6eb67612dbe87 Mon Sep 17 00:00:00 2001 From: Ben Pfaff <b...@ovn.org> Date: Mon, 9 Nov 2020 16:12:50 -0800 Subject: [PATCH ovn] ovn-nbctl: Make --wait=hv wait for southbound database in corner case. In the corner case where there are no chassis, --wait=hv didn't wait at all, instead of waiting for the southbound database. Reported-by: Dumitru Ceara <dce...@redhat.com> Signed-off-by: Ben Pfaff <b...@ovn.org> --- ovn-nb.xml | 27 ++++++++++++++++++++------- utilities/ovn-nbctl.c | 2 +- 2 files changed, 21 insertions(+), 8 deletions(-) diff --git a/ovn-nb.xml b/ovn-nb.xml index 51b186b84419..d3e51f3e5e87 100644 --- a/ovn-nb.xml +++ b/ovn-nb.xml @@ -84,13 +84,26 @@ </column> <column name="hv_cfg"> - Sequence number that <code>ovn-northd</code> sets to the smallest - sequence number of all the chassis in the system, as reported in the - <code>Chassis_Private</code> table in the southbound database. Thus, - <ref column="hv_cfg"/> equals <ref column="nb_cfg"/> if all chassis are - caught up with the northbound configuration (which may never happen, if - any chassis is down). This value can regress, if a chassis was removed - from the system and rejoins before catching up. + <p> + Sequence number that <code>ovn-northd</code> sets to the smallest + sequence number of all the chassis in the system, as reported in the + <code>Chassis_Private</code> table in the southbound database. Thus, + <ref column="hv_cfg"/> equals <ref column="nb_cfg"/> if all chassis are + caught up with the northbound configuration (which may never happen, if + any chassis is down). This value can regress, if a chassis was removed + from the system and rejoins before catching up. + </p> + + <p> + If there are no chassis, then <code>ovn-northd</code> copies + <code>nb_cfg</code> to <ref column="hv_cfg"/>. Thus, in this case, + the (nonexistent) hypervisors are always considered to be caught up. + This means that hypervisors can be "caught up" even in cases where + <ref column="sb_cfg"/> would show that the southbound database is + not. To detect when both the hypervisors and the southbound database + are caught up, a client should take the smaller of <ref + column="sb_cfg"/> and <ref column="hv_cfg"/>. + </p> </column> <column name="hv_cfg_timestamp"> diff --git a/utilities/ovn-nbctl.c b/utilities/ovn-nbctl.c index 4f28edc808ec..ee9849de53a9 100644 --- a/utilities/ovn-nbctl.c +++ b/utilities/ovn-nbctl.c @@ -6357,7 +6357,7 @@ do_nbctl(const char *args, struct ctl_command *commands, size_t n_commands, NBREC_NB_GLOBAL_FOR_EACH (nb, idl) { int64_t cur_cfg = (wait_type == NBCTL_WAIT_SB ? nb->sb_cfg - : nb->hv_cfg); + : MIN(nb->sb_cfg, nb->hv_cfg)); if (cur_cfg >= next_cfg) { if (print_wait_time) { printf("Time spent on processing nb_cfg %"PRId64":\n", -- 2.26.2 _______________________________________________ dev mailing list d...@openvswitch.org https://mail.openvswitch.org/mailman/listinfo/ovs-dev