[ovs-dev] [PATCH ovn] Exclude inport and outport symbol tables from conjunction

[Han Zhou]

On Wed, Sep 18, 2019 at 2:51 AM Dumitru Ceara <dceara at redhat.com> wrote:
>
> On Wed, Sep 18, 2019 at 9:30 AM Dumitru Ceara <dceara at redhat.com> wrote:
> >
> > On Tue, Sep 17, 2019 at 6:49 PM Han Zhou <zhouhan at gmail.com> wrote:
> > >
> > >
> > >
> > > On Tue, Sep 17, 2019 at 5:21 AM Mark Michelson <mmichels at redhat.com>
wrote:
> > > >
> > > > On 9/16/19 12:04 PM, Han Zhou wrote:
> > > > >
> > > > >
> > > > > On Mon, Sep 16, 2019 at 4:15 AM Dumitru Ceara <dceara at redhat.com
> > > > > <mailto:dceara at redhat.com>> wrote:
> > > > >  >
> > > > >  > On Sat, Sep 14, 2019 at 7:16 PM Han Zhou <zhouhan at gmail.com
> > > > > <mailto:zhouhan at gmail.com>> wrote:
> > > > >  > >
> > > > >  > >
> > > > >  > >
> > > > >  > > On Sat, Sep 14, 2019 at 9:09 AM Han Zhou <zhouhan at gmail.com
> > > > > <mailto:zhouhan at gmail.com>> wrote:
> > > > >  > > >
> > > > >  > > >
> > > > >  > > >
> > > > >  > > > On Sat, Sep 14, 2019 at 12:40 AM Numan Siddique
> > > > > <nusiddiq at redhat.com <mailto:nusiddiq at redhat.com>> wrote:
> > > > >  > > > >
> > > > >  > > > >
> > > > >  > > > >
> > > > >  > > > > On Sat, Sep 14, 2019 at 2:41 AM Daniel Alvarez Sanchez
> > > > > <dalvarez at redhat.com <mailto:dalvarez at redhat.com>> wrote:
> > > > >  > > > >>
> > > > >  > > > >> Acked-by: Daniel Alvarez <dalvarez at redhat.com
> > > > > <mailto:dalvarez at redhat.com>>
> > > > >  > > > >>
> > > > >  > > > >>
> > > > >  > > > >> On Fri, Sep 13, 2019 at 11:02 PM Mark Michelson
> > > > > <mmichels at redhat.com <mailto:mmichels at redhat.com>> wrote:
> > > > >  > > > >> >
> > > > >  > > > >> > Acked-by: Mark Michelson
> > > > >  > > > >> >
> > > > >  > > > >> > It sucks that we lose the efficiency of the
conjunctive
> > > > > match altogether
> > > > >  > > > >> > on port groups because of this error, but I
understand this
> > > > > is a huge
> > > > >  > > > >> > bug and needs fixing.
> > > > >  > > > >> If I'm not mistaken, from OpenStack standpoint
conjunction was
> > > > > *only*
> > > > >  > > > >> being used when using port groups and ACLs that matched
on
> > > > > port ranges
> > > > >  > > > >> ( e.g tcp.dst >= X && tcp.dst <=Y) which was not
working.
> > > > > Therefore
> > > > >  > > > >> we're not losing performance because it was already
broken
> > > > > (given that
> > > > >  > > > >> there was more than one ACL like that).
> > > > >  > > > >>
> > > > >  > > > >> >
> > > > >  > > > >> > Perhaps it would be good to start up a discussion on
this
> > > > > list about a
> > > > >  > > > >> > more longterm solution that would allow for
conjunctive
> > > > > matches with no
> > > > >  > > > >> > ambiguity.
> > > > >  > > > >> Agreed! We already discussed some ideas on IRC but it'd
be
> > > > > awesome to
> > > > >  > > > >> have a thread and brainstorm there.
> > > > >  > > > >>
> > > > >  > > > >
> > > > >  > > > > Thanks for the reviews. I applied this to master.
> > > > >  > > > > Agree we can discuss it further and come up with ideas.
> > > > >  > > > >
> > > > >  > > > > I know Dumitru has some idea to make use of conjunctions
for
> > > > > port groups.
> > > > >  > > > > CC'ing Han if he has any comments on ideas.
> > > > >  > > > >
> > > > >  > > >
> > > > >  > > > Hi Numan,
> > > > >  > > >
> > > > >  > > > This is a good finding. However, I think it is not
specifically a
> > > > > problem of port group. It seems to be a more general problem and
this
> > > > > patch fixes only a special case.
> > > > >  > > > For example, would there be similar problem for below ACLs
> > > > > without port groups:
> > > > >  > > >
> > > > >  > > > ip4 && tcp.src >= 1000 && tcp.src <= 1001 && tcp.dst >=
500 &&
> > > > > tcp.dst <= 501
> > > > >  > > > ip4 && tcp.src >= 1000 && tcp.src <= 1001 && tcp.dst >=
600 &&
> > > > > tcp.dst <= 601
> > > > >  > > >
> > > > >  > > > Another example is with address set:
> > > > >  > > >
> > > > >  > > > ip4 && ip4.src == $as1 && tcp.dst >= 500 && tcp.dst <= 501
> > > > >  > > > ip4 && ip4.src == $as1 && tcp.dst >= 600 && tcp.dst <= 601
> > > > >  > > >
> > > > >  > > > Or even without range:
> > > > >  > > > ip4 && tcp.src == {1000, 1001} && tcp.dst == {500, 501}
> > > > >  > > > ip4 && tcp.src == {1000, 1001} && tcp.dst == {600, 601}
> > > > >  > > >
> > > > >  > > > You may think of more examples. Whenever there are multiple
> > > > > conjunctionable ACLs with same match as part of the conjunction,
it
> > > > > should result in such problem.
> > > > >  > > >
> > > > >  > > > A quick fix to all these problems may be just abandon
> > > > > conjunction, but I believe there are better ways to address it.
> > > > >  > > >
> > > > >  > > > First of all, these matches can be rewritten by combining
them in
> > > > > a single ACL with "OR" operator, e.g.:
> > > > >  > > >
> > > > >  > > > outport == @pg1 && ip4 && tcp.dst >= 500 && tcp.dst <= 501
> > > > >  > > > outport == @pg1 && ip4 && tcp.dst >= 600 && tcp.dst <= 601
> > > > >  > > >
> > > > >  > > > can be rewritten as ====>
> > > > >  > > >
> > > > >  > > > outport == @pg1 && ip4 && (tcp.dst >= 500 && tcp.dst <=
501 ||
> > > > > tcp.dst >= 600 && tcp.dst <= 601)
> > > > >  > > >
> > > > >  > > > Similar can be done for all above examples. So, a
workaround to
> > > > > the problem is from user side (e.g. OpenStack) to make sure always
> > > > > combining ACLs with "OR" if there are common conjunctionable
matches
> > > > > between different ACLs. However, a better way would be in
ovn-northd
> > > > > itself to detect and combine such ACLs internally, before
generating the
> > > > > final logical flows in SB. It may be more convenient to be done in
> > > > > ovn-controller, because we are not even parsing the ACLs in
ovn-northd
> > > > > today, but the cost of such pre-processing would be duplicated in
all
> > > > > HVs. It surely will increase CPU cost for doing such combination,
but
> > > > > I'd not worry too much if we do it properly at each LS level
instead of
> > > > > for all ACLs.
> > > > >  > >
> > > > >  > > I just thought a little more about combining the
conjunctions. It
> > > > > seems we can do it without pre-processing by just handling
duplicated
> > > > > flows in ofctrl_add_flow(). Currently we just drop duplicated
flows, but
> > > > > we can check that if the action is conjuncture and the
conjuncture ID is
> > > > > different, we can perform a combination by using existing flow's
> > > > > conjunction id to update all the flows related to that to-be-added
> > > > > duplicated flow. This way, the combination is performed
on-the-fly,
> > > > > without introduce too much cost and without introduce parsing in
> > > > > ovn-northd either.
> > > > >  >
> > > > >  > Hi Han,
> > > > >  >
> > > > >  > Will this actually work without a change in OVS? I wonder
because in
> > > > >  > the ovs-fields man page [1] I see:
> > > > >  >
> > > > >  > "Conjunctive flows must not overlap with each other, at
> > > > >  >  a given priority, that is, any given packet must be
> > > > >  >  able to match at most one conjunctive flow at a given
> > > > >  >  priority. Overlapping conjunctive flows yield
> > > > >  >  unpredictable results."
> > > > >
> > > > > Hi Dumitru, the approach of combining ACLs should eliminate the
> > > > > overlapping conjunction problem for ACLs with common expressions,
> > > > > because a single ACL/logical-flow will be translated to a single
> > > > > conjunction. The combination is generally for ACLs with form :
> > > > >    A and B
> > > > >    A and C
> > > > > If A is (a1 or a2), B is (b1 or b2 ), C is (c1 or c2), the initial
> > > > > conjunction matches of current (incorrect) implementation will be:
> > > > > conj_id=1234 actions=...
> > > > > a1 conjunction(1234, 1/4)
> > > > > a2 conjunction(1234, 2/4)
> > > > > b1 conjunction(1234, 3/4)
> > > > > b2 conjunction(1234, 4/4)
> > > > >
> > > > > conj_id=1235 actions=...
> > > > > a1 conjunction(1235, 1/4)
> > > > > a2 conjunction(1235, 2/4)
> > > > > c1 conjunction(1235, 3/4)
> > > > > c2 conjunction(1235, 4/4)
> > > > >
> > > > > The a1 and a2 matches are overlapping between these two sets.
> > > > > In fact, the ACLs are equivalent to:
> > > > >    A and (B or C), which is equivalent to: (a1 or a2) and (b1 or
b2 or
> > > > > c1 or c2)
> > > > > So, if combined with the approach I mentioned, the conjunction
matches
> > > > > will be:
> > > > > conj_id=1234 actions=...
> > > > > a1 conjunction(1234, 1/6)
> > > > > a2 conjunction(1234, 2/6)
> > > > > b1 conjunction(1234, 3/6)
> > > > > b2 conjunction(1234, 4/6)
> > > > > c1 conjunction(1234, 5/6)
> > > > > c2 conjunction(1234, 6/6)
> > > > >
> > > > > The overlapping problem is solved for this use case, and it
reduces
> > > > > total number of flows, which can also be considered an *compiler
> > > > > optimization* for logical flows translation in ovn-controller. I
don't
> > > > > think any change is needed from OVS.
> > > >
> > > > Unfortunately, this only works for ACLs that have the same action.
> > > > Consider that you have the following:
> > > >
> > > > A and B, allow
> > > > A and C, drop
> > > >
> > > > You can't combine these ACLs into
> > > >
> > > > A and (B or C)
> > > >
> > > > since they have different ACL verdicts. You have to leave them
separate.
> > > >
> > > > Therefore, you end up with the same problem as before. The A part
of the
> > > > conjunction for each conjunctive match overlaps.
> > > >
> > >
> > > Hi Mark,
> > >
> > > Thanks for pointing out. Unfortunately you are right :)
> > > So I guess the only option we have now is to play around with
priorities. Although it might be tricky, it should work. And the best part
is, priority can also solve the "partial overlapping" problem I mentioned
earlier, such as:
> > > > >  > > tcp.src == {1000, 1001} && tcp.dst == {500, 501}
> > > > >  > > tcp.src == {1001, 1002} && tcp.dst == {600, 601}
> > >
> > > Here is some more thinking about the priority solution. Today for any
user-defined ACL priority (PA), there is a corresponding OVS flow priority
(PO). PA to PO is one-to-one mapping. Now to solve the overlapping problem,
there can be more than one PO [POx..POy] mapped for each PA. The problem is
how to maintain the mappings between these two. I think there are 2 options:
> > >
> > > Option1: Whenever there is duplicate flow encountered for
user-defined priority PA, increase the range of PO from [POx..POy] to
[POx..POy+1].
> > >
> > > This approach requires dynamically adjust priorities for existing
flows. For example, initially there are flows:
> > > 1: user-defined priority 100, OVS priority 1100: <match 1>, <action 1>
> > > 2: user-defined priority 101, OVS priority 1101: <match 2>, <action 2>
> > >
> > > Now there is a logical flow requires conjunction:
> > > 3: user-defined priority 100, OVS priority 1100: A and B, <action 3>
> > >
> > > And then there is another logical flow requires conjunction, with
overlapping to the previous one:
> > > 4: user-defined priority 100, OVS priority <?>: A and C, <action 4>
> > >
> > > To handle the overlapping, we need to increase the range
corresponding to 100. It was 1100, and now it needs to be 1100 and 1101. So
we will have to adjust priority of the flow 2 from 1101 to 1102, to
preserve the original priority order. This will have high cost, and worst
thing is, it will disrupt Incremental Processing.
> > >
> > > Option2: Preserve N priorities in the range of OVS flow [POx..POx+N)
for each user-defined PA. If there are more than N ACLs with overlapping
conjunction sets, it is not supported.
> > >
> > > This approach avoids the priority readjustment problem, sacrificing
the number of priorities and number of overlapping conjunction that we can
support. The bigger N is, the less number of user-defined priority can be
supported; the smaller N is, the less number of overlapping conjunction can
be supported.
> > > Although each one has its pros and cons, I prefer option2 more than
option1, because conjunction is mainly for performance. If we end up with a
solution that sacrifice performance it may not worth enabling conjunction
at all.
> >
> > Hi Han,
> >
> > I agree that Option2 is preferable due to the cost of reassigning
> > priorities in Option1. What worries me is that we decrease the
> > available number of user priorities by a fixed factor for all cases,
> > even if there are no overlapping conjunction sets in the config. Based
> > on your idea I had the following optimization in mind:
> >
> > Add a new table, let's call it "ACL-collision" for now. When an ACL is
> > added the following operations would be performed by ovn-controller:
> > - Try to add the ACL to the regular ACL table with priority PO. If
> > there's no overlap on matches for any of the generated flows at
> > priority PO then stop.
> > - Otherwise, keep the first overlapping flow of the existing ACL and
> > new ACL at priority PO in the ACL table but change its action to
> > resubmit the packet to the "ACL-collision" table.
> > - Add the remaining flows of the old ACL and new ACL to table
> > "ACL-collision" making sure that we assign different priorities to
> > flows from the two ACLs.
> > - In table "ACL-collision" add a low priority "match-all" flow that
> > will resubmit the packet back to the ACL table to continue regular ACL
> > matching.
>
> I just realized that if we resubmit the packet back to the ACL table
> we'd have to avoid matching the flow(s) that already moved the packet
> to the ACL-collision table.
> I'm not sure if that can be easily done though..
>
> >
> > This approach would not affect any existing deployments where there
> > are no overlapping conjunction sets.
> > Also, it should maintain logical ordering of PA priorities because the
> > first term of the conjunction flows (the one in the ACL table) keeps
> > the same PO priority as now (that corresponds to PA).
> >
> > The disadvantage is that we add another step in the pipeline for
> > packets that match the common part of the overlapping conjunction sets
> > potentially increasing latency/jitter.
> >
> > Do you think this would work? Are there any other cases to take care of?
> >
> > Thanks,
> > Dumitru
> >
Hi Dumitru,

Sorry that I am not sure if I understand your proposal. When you say "keep
the first overlapping flow", what does it really mean? If there are N
elements overlapping in the sets, there will be N pairs of overlapping
flows, each with different match conditions, then how could the problem be
solved by only changing the operation on the first overlapping flow? Could
you explain a little more with the example:
> > > 1: user-defined priority 100, OVS priority 1100: <match 1>, <action 1>
> > > 2: user-defined priority 101, OVS priority 1101: <match 2>, <action 2>
> > >
> > > Now there is a logical flow requires conjunction:
> > > 3: user-defined priority 100, OVS priority 1100: A and B, <action 3>
> > >
> > > And then there is another logical flow requires conjunction, with
overlapping to the previous one:
> > > 4: user-defined priority 100, OVS priority <?>: A and C, <action 4>