[ovs-dev,v2,1/4] physical: Fix implementation of logical patch ports.
diff mbox

Message ID 1445116064-20782-2-git-send-email-blp@nicira.com
State Superseded
Headers show

Commit Message

Ben Pfaff Oct. 17, 2015, 9:07 p.m. UTC
Logical patch ports do not have a physical location and effectively reside
on every hypervisor.  This is fine for unicast output to logical patch
ports.  However, when a logical patch port is part of a logical multicast
group, lumping them together with the other "local" ports in a multicast
group yields packet duplication, because every hypervisor to which the
packet is tunneled re-outputs it to the logical patch port.

This commit fixes the problem, by treating logical patch ports as remote
rather than local when they are part of a logical multicast group.  This
yields exactly-once semantics.

Found while testing implementation of ARP in OVN logical router.  The
following commit adds a test that fails without this fix.

Signed-off-by: Ben Pfaff <blp@nicira.com>
---
 ovn/controller/physical.c  | 43 +++++++++++++++++++++++++++++++------------
 ovn/ovn-architecture.7.xml | 16 +++++++++++++---
 2 files changed, 44 insertions(+), 15 deletions(-)

Comments

Justin Pettit Oct. 18, 2015, 5:33 p.m. UTC | #1
> On Oct 17, 2015, at 2:07 PM, Ben Pfaff <blp@nicira.com> wrote:
> 
> Logical patch ports do not have a physical location and effectively reside
> on every hypervisor.  This is fine for unicast output to logical patch
> ports.  However, when a logical patch port is part of a logical multicast
> group, lumping them together with the other "local" ports in a multicast
> group yields packet duplication, because every hypervisor to which the
> packet is tunneled re-outputs it to the logical patch port.
> 
> This commit fixes the problem, by treating logical patch ports as remote
> rather than local when they are part of a logical multicast group.  This
> yields exactly-once semantics.
> 
> Found while testing implementation of ARP in OVN logical router.  The
> following commit adds a test that fails without this fix.

It's not actually the following patch, but the one after that.

Acked-by: Justin Pettit <jpettit@nicira.com>

--Justin
Ben Pfaff Oct. 18, 2015, 7:02 p.m. UTC | #2
On Sun, Oct 18, 2015 at 10:33:11AM -0700, Justin Pettit wrote:
> 
> > On Oct 17, 2015, at 2:07 PM, Ben Pfaff <blp@nicira.com> wrote:
> > 
> > Logical patch ports do not have a physical location and effectively reside
> > on every hypervisor.  This is fine for unicast output to logical patch
> > ports.  However, when a logical patch port is part of a logical multicast
> > group, lumping them together with the other "local" ports in a multicast
> > group yields packet duplication, because every hypervisor to which the
> > packet is tunneled re-outputs it to the logical patch port.
> > 
> > This commit fixes the problem, by treating logical patch ports as remote
> > rather than local when they are part of a logical multicast group.  This
> > yields exactly-once semantics.
> > 
> > Found while testing implementation of ARP in OVN logical router.  The
> > following commit adds a test that fails without this fix.
> 
> It's not actually the following patch, but the one after that.
> 
> Acked-by: Justin Pettit <jpettit@nicira.com>

Oops.

I decided to just flip the order of patches 1 and 2, since they were
independent.

Thanks for the review!

Patch
diff mbox

diff --git a/ovn/controller/physical.c b/ovn/controller/physical.c
index 1b2b7fc..5821c11 100644
--- a/ovn/controller/physical.c
+++ b/ovn/controller/physical.c
@@ -497,6 +497,8 @@  physical_run(struct controller_ctx *ctx, enum mf_field_id mff_ovn_geneve,
 
     /* Handle output to multicast groups, in tables 32 and 33. */
     const struct sbrec_multicast_group *mc;
+    struct ofpbuf remote_ofpacts;
+    ofpbuf_init(&remote_ofpacts, 0);
     SBREC_MULTICAST_GROUP_FOR_EACH (mc, ctx->ovnsb_idl) {
         struct sset remote_chassis = SSET_INITIALIZER(&remote_chassis);
         struct match match;
@@ -507,11 +509,18 @@  physical_run(struct controller_ctx *ctx, enum mf_field_id mff_ovn_geneve,
 
         /* Go through all of the ports in the multicast group:
          *
-         *    - For local ports, add actions to 'ofpacts' to set the output
-         *      port and resubmit.
+         *    - For remote ports, add the chassis to 'remote_chassis'.
          *
-         *    - For remote ports, add the chassis to 'remote_chassis'. */
+         *    - For local ports (other than logical patch ports), add actions
+         *      to 'ofpacts' to set the output port and resubmit.
+         *
+         *    - For logical patch ports, add actions to 'remote_ofpacts'
+         *      instead.  (If we put them in 'ofpacts', then the output
+         *      would happen on every hypervisor in the multicast group,
+         *      effectively duplicating the packet.)
+         */
         ofpbuf_clear(&ofpacts);
+        ofpbuf_clear(&remote_ofpacts);
         for (size_t i = 0; i < mc->n_ports; i++) {
             struct sbrec_port_binding *port = mc->ports[i];
 
@@ -528,7 +537,11 @@  physical_run(struct controller_ctx *ctx, enum mf_field_id mff_ovn_geneve,
                 put_load(zone_id, MFF_LOG_CT_ZONE, 0, 32, &ofpacts);
             }
 
-            if (simap_contains(&localvif_to_ofport,
+            if (!strcmp(port->type, "patch")) {
+                put_load(port->tunnel_key, MFF_LOG_OUTPORT, 0, 32,
+                         &remote_ofpacts);
+                put_resubmit(OFTABLE_DROP_LOOPBACK, &remote_ofpacts);
+            } else if (simap_contains(&localvif_to_ofport,
                                port->parent_port
                                ? port->parent_port : port->logical_port)) {
                 put_load(port->tunnel_key, MFF_LOG_OUTPORT, 0, 32, &ofpacts);
@@ -568,8 +581,13 @@  physical_run(struct controller_ctx *ctx, enum mf_field_id mff_ovn_geneve,
          *
          * Handle output to the remote chassis in the multicast group, if
          * any. */
-        if (!sset_is_empty(&remote_chassis)) {
-            ofpbuf_clear(&ofpacts);
+        if (!sset_is_empty(&remote_chassis) || remote_ofpacts.size > 0) {
+            if (remote_ofpacts.size > 0) {
+                /* Following delivery to logical patch ports, restore the
+                 * multicast group as the logical output port. */
+                put_load(mc->tunnel_key, MFF_LOG_OUTPORT, 0, 32,
+                         &remote_ofpacts);
+            }
 
             const char *chassis;
             const struct chassis_tunnel *prev = NULL;
@@ -581,23 +599,24 @@  physical_run(struct controller_ctx *ctx, enum mf_field_id mff_ovn_geneve,
                 }
 
                 if (!prev || tun->type != prev->type) {
-                    put_encapsulation(mff_ovn_geneve, tun,
-                                      mc->datapath, mc->tunnel_key, &ofpacts);
+                    put_encapsulation(mff_ovn_geneve, tun, mc->datapath,
+                                      mc->tunnel_key, &remote_ofpacts);
                     prev = tun;
                 }
-                ofpact_put_OUTPUT(&ofpacts)->port = tun->ofport;
+                ofpact_put_OUTPUT(&remote_ofpacts)->port = tun->ofport;
             }
 
-            if (ofpacts.size) {
+            if (remote_ofpacts.size) {
                 if (local_ports) {
-                    put_resubmit(OFTABLE_LOCAL_OUTPUT, &ofpacts);
+                    put_resubmit(OFTABLE_LOCAL_OUTPUT, &remote_ofpacts);
                 }
                 ofctrl_add_flow(flow_table, OFTABLE_REMOTE_OUTPUT, 100,
-                                &match, &ofpacts);
+                                &match, &remote_ofpacts);
             }
         }
         sset_destroy(&remote_chassis);
     }
+    ofpbuf_uninit(&remote_ofpacts);
 
     /* Table 0, priority 100.
      * ======================
diff --git a/ovn/ovn-architecture.7.xml b/ovn/ovn-architecture.7.xml
index 0bf9337..c15a4c8 100644
--- a/ovn/ovn-architecture.7.xml
+++ b/ovn/ovn-architecture.7.xml
@@ -778,6 +778,18 @@ 
       </p>
 
       <p>
+        Logical patch ports are a special case.  Logical patch ports do not
+        have a physical location and effectively reside on every hypervisor.
+        Thus, flow table 33, for output to ports on the local hypervisor,
+        naturally implements output to unicast logical patch ports too.
+        However, applying the same logic to a logical patch port that is part
+        of a logical multicast group yields packet duplication, because each
+        hypervisor that contains a logical port in the multicast group will
+        also output the packet to the logical patch port.  Thus, multicast
+        groups implement output to logical patch ports in table 32.
+      </p>
+
+      <p>
         Each flow in table 32 matches on a logical output port for unicast or
         multicast logical ports that include a logical port on a remote
         hypervisor.  Each flow's actions implement sending a packet to the port
@@ -796,9 +808,7 @@ 
 
       <p>
         Flows in table 33 resemble those in table 32 but for logical ports that
-        reside locally rather than remotely.  (This includes logical patch
-        ports, which do not have a physical location and effectively reside on
-        every hypervisor.)  For unicast logical output ports
+        reside locally rather than remotely.  For unicast logical output ports
         on the local hypervisor, the actions just resubmit to table 34.  For
         multicast output ports that include one or more logical ports on the
         local hypervisor, for each such logical port <var>P</var>, the actions