Fabric replication tree not present in vrf

[Manish Singh]

Agent didnt have fabric replication tree for a VRF. This isolates compute node from BUM replication.
Control node had the correct information.

[OpenContrail Admin (ci-admin-f)]

Review in progress for https://review.opencontrail.org/16110
Submitter: Manish Singh (<email address hidden>)

[OpenContrail Admin (ci-admin-f)]

Reviewed: https://review.opencontrail.org/16110
Committed: http://github.org/Juniper/contrail-controller/commit/56b847efbf1601446fe84c4ea0f62991842e149a
Submitter: Zuul
Branch: R2.21.x

commit 56b847efbf1601446fe84c4ea0f62991842e149a
Author: Manish <email address hidden>
Date: Wed Jan 6 11:54:13 2016 +0530

Agent fabric replication tree is empty.

Problem:
On switchover of mcast control node (because of lower IP), agent fabric tree was
geting removed. For further explanation lets assume there are two control-node
C1 and C2. C2 is in stable shape and came up first. Agent subscribes to fabric
replication tree by reaching C2. Now C1 comes up and before it can participate
meaningfully in replication it becomes unusable. Agent sees C1 coming up and
removes the fabric tree received from C2 as it plans to use C1 as multicast
node. However when C1 goes down, agent goes back to C2 and send subscribe again.
(Note: When C1 took over from C2, agent had triggered a walk to unsubscribe from
C2, but it may not be complete or scheduled). Now C2 starts same multicast walk
for subscription inturn cancelling old unsubscription walk.
Control-node on seeing subscription request again (since unsubscribe didnt go)
and no change in Olist or other info, may decide to ignore the request.
Hence agent will never get fabric tree.

Solution:
Firstly remove the logic of lower peer selection.
Secondly in every subscribe sent because of new control-node becoming mcast
controller for agent, check if unsubscribe is sent. If its not explicitly send
unsubscribe before subscribing.

Closes-bug: #1531436

Change-Id: I089e3137cdfbbd28897414c411b22b8c82260990

[OpenContrail Admin (ci-admin-f)]

Review in progress for https://review.opencontrail.org/16467
Submitter: Manish Singh (<email address hidden>)

[OpenContrail Admin (ci-admin-f)]

Reviewed: https://review.opencontrail.org/16467
Committed: http://github.org/Juniper/contrail-controller/commit/d822e9eae2679db21f4087bdce7df6680f058291
Submitter: Zuul
Branch: master

commit d822e9eae2679db21f4087bdce7df6680f058291
Author: Manish <email address hidden>
Date: Wed Jan 6 11:54:13 2016 +0530

Agent fabric replication tree is empty.

Problem:
On switchover of mcast control node (because of lower IP), agent fabric tree was
geting removed. For further explanation lets assume there are two control-node
C1 and C2. C2 is in stable shape and came up first. Agent subscribes to fabric
replication tree by reaching C2. Now C1 comes up and before it can participate
meaningfully in replication it becomes unusable. Agent sees C1 coming up and
removes the fabric tree received from C2 as it plans to use C1 as multicast
node. However when C1 goes down, agent goes back to C2 and send subscribe again.
(Note: When C1 took over from C2, agent had triggered a walk to unsubscribe from
C2, but it may not be complete or scheduled). Now C2 starts same multicast walk
for subscription inturn cancelling old unsubscription walk.
Control-node on seeing subscription request again (since unsubscribe didnt go)
and no change in Olist or other info, may decide to ignore the request.
Hence agent will never get fabric tree.

Solution:
Firstly remove the logic of lower peer selection.
Secondly in every subscribe sent because of new control-node becoming mcast
controller for agent, check if unsubscribe is sent. If its not explicitly send
unsubscribe before subscribing.

Closes-bug: #1531436

Change-Id: I089e3137cdfbbd28897414c411b22b8c82260990
(cherry picked from commit 56b847efbf1601446fe84c4ea0f62991842e149a)

[OpenContrail Admin (ci-admin-f)]

Review in progress for https://review.opencontrail.org/17035
Submitter: Manish Singh (<email address hidden>)

[OpenContrail Admin (ci-admin-f)]

Review in progress for https://review.opencontrail.org/17036
Submitter: Manish Singh (<email address hidden>)

[OpenContrail Admin (ci-admin-f)]

Reviewed: https://review.opencontrail.org/17036
Committed: http://github.org/Juniper/contrail-controller/commit/9cfbc9d1109434e957ece0193cbaf01f9e9df5ac
Submitter: Zuul
Branch: R2.22.x

commit 9cfbc9d1109434e957ece0193cbaf01f9e9df5ac
Author: Manish <email address hidden>
Date: Wed Jan 6 11:54:13 2016 +0530

Agent fabric replication tree is empty.

Problem:
On switchover of mcast control node (because of lower IP), agent fabric tree was
geting removed. For further explanation lets assume there are two control-node
C1 and C2. C2 is in stable shape and came up first. Agent subscribes to fabric
replication tree by reaching C2. Now C1 comes up and before it can participate
meaningfully in replication it becomes unusable. Agent sees C1 coming up and
removes the fabric tree received from C2 as it plans to use C1 as multicast
node. However when C1 goes down, agent goes back to C2 and send subscribe again.
(Note: When C1 took over from C2, agent had triggered a walk to unsubscribe from
C2, but it may not be complete or scheduled). Now C2 starts same multicast walk
for subscription inturn cancelling old unsubscription walk.
Control-node on seeing subscription request again (since unsubscribe didnt go)
and no change in Olist or other info, may decide to ignore the request.
Hence agent will never get fabric tree.

Solution:
Firstly remove the logic of lower peer selection.
Secondly in every subscribe sent because of new control-node becoming mcast
controller for agent, check if unsubscribe is sent. If its not explicitly send
unsubscribe before subscribing.

Closes-bug: #1531436

Change-Id: I089e3137cdfbbd28897414c411b22b8c82260990
(cherry picked from commit 56b847efbf1601446fe84c4ea0f62991842e149a)
(cherry picked from commit d822e9eae2679db21f4087bdce7df6680f058291)

[OpenContrail Admin (ci-admin-f)]

Reviewed: https://review.opencontrail.org/17035
Committed: http://github.org/Juniper/contrail-controller/commit/3959611f825b3c8eca4f2c84b7fd8513fcf6ac50
Submitter: Zuul
Branch: R2.20

commit 3959611f825b3c8eca4f2c84b7fd8513fcf6ac50
Author: Manish <email address hidden>
Date: Wed Jan 6 11:54:13 2016 +0530

Agent fabric replication tree is empty.

Problem:
On switchover of mcast control node (because of lower IP), agent fabric tree was
geting removed. For further explanation lets assume there are two control-node
C1 and C2. C2 is in stable shape and came up first. Agent subscribes to fabric
replication tree by reaching C2. Now C1 comes up and before it can participate
meaningfully in replication it becomes unusable. Agent sees C1 coming up and
removes the fabric tree received from C2 as it plans to use C1 as multicast
node. However when C1 goes down, agent goes back to C2 and send subscribe again.
(Note: When C1 took over from C2, agent had triggered a walk to unsubscribe from
C2, but it may not be complete or scheduled). Now C2 starts same multicast walk
for subscription inturn cancelling old unsubscription walk.
Control-node on seeing subscription request again (since unsubscribe didnt go)
and no change in Olist or other info, may decide to ignore the request.
Hence agent will never get fabric tree.

Solution:
Firstly remove the logic of lower peer selection.
Secondly in every subscribe sent because of new control-node becoming mcast
controller for agent, check if unsubscribe is sent. If its not explicitly send
unsubscribe before subscribing.

Closes-bug: #1531436

Change-Id: I089e3137cdfbbd28897414c411b22b8c82260990
(cherry picked from commit 56b847efbf1601446fe84c4ea0f62991842e149a)
(cherry picked from commit d822e9eae2679db21f4087bdce7df6680f058291)