A distribution switch stops answering SSH, OSPF neighbors start flapping, and the interface counters look normal. The data plane is forwarding fine, but the route processor is at 99 percent. Someone’s misconfigured monitoring host is hammering SNMP, or a broadcast storm is pushing ARP to the CPU, or a scanner found the management address. Whatever the source, all traffic punted to the CPU shares one queue, and routing protocols lose to noise.
Control plane policing (CoPP) fixes this by classifying punted traffic and rate-limiting each class separately. Your BGP and OSPF keep their share of the CPU, and SNMP floods get dropped before they reach the processes that matter.
What the platform already does
On Catalyst 9000 switches, IOS XE ships with a system-defined policy called system-cpp-policy attached to the control plane. It has sensible defaults per punt queue, and you tune it rather than replace it. On ISR and ASR routers there is no such default for most traffic, so you build your own policy with MQC. The example below is for a router; the classification logic carries over to the switches.
Check what is currently applied before changing anything:
show policy-map control-plane
show platform hardware qfp active feature qos interface CPP
show platform software infrastructure punt
The last command shows what is actually being punted and why. Look at it on a normal day so you know what “normal” means.
Build the classes
Start with ACLs that describe traffic you expect. Be specific about sources; a class that matches “any OSPF” also matches spoofed OSPF.
ip access-list extended COPP-ROUTING
permit tcp 10.0.0.0 0.0.255.255 any eq bgp
permit tcp 10.0.0.0 0.0.255.255 eq bgp any
permit ospf 10.0.0.0 0.0.255.255 any
permit udp any any eq 3784
!
ip access-list extended COPP-MGMT
permit tcp 10.50.0.0 0.0.0.255 any eq 22
permit udp 10.50.0.0 0.0.0.255 any eq snmp
permit udp any eq ntp any eq ntp
permit tcp 10.50.0.0 0.0.0.255 any eq 830
!
ip access-list extended COPP-ICMP
permit icmp any any echo
permit icmp any any echo-reply
permit icmp any any ttl-exceeded
permit icmp any any unreachable
!
class-map match-all CM-COPP-ROUTING
match access-group name COPP-ROUTING
class-map match-all CM-COPP-MGMT
match access-group name COPP-MGMT
class-map match-all CM-COPP-ICMP
match access-group name COPP-ICMP
Port 3784 is single-hop BFD control. If you run BFD and forget it here, you will find out the hard way when the default class starts dropping it.
Build the policy
policy-map PM-COPP
class CM-COPP-ROUTING
police rate 2000 pps
conform-action transmit
exceed-action transmit
class CM-COPP-MGMT
police rate 500 pps
conform-action transmit
exceed-action drop
class CM-COPP-ICMP
police rate 100 pps
conform-action transmit
exceed-action drop
class class-default
police rate 200 pps
conform-action transmit
exceed-action drop
!
control-plane
service-policy input PM-COPP
Two choices here are deliberate. The routing class transmits on exceed: it exists for visibility, not enforcement, because dropping your own BGP keepalives during a convergence event is worse than a busy CPU. And class-default drops on exceed, so anything you did not anticipate gets a small, bounded share.
The rates are starting points, not recommendations. Size them from what you observe. A router with a full table and a few hundred BGP sessions needs a very different routing rate from a branch box with two OSPF neighbors.
Roll it out in monitor mode first
The classic CoPP outage is an engineer applying a policy that drops something nobody listed: HSRP, LDP, DHCP relay, a TACACS reply, IPsec IKE to the router itself. Avoid it by deploying every class with exceed-action transmit first, then watching for a week:
show policy-map control-plane input
Look at the exceeded counters per class, and look at what lands in class-default. If class-default is busy, find out what it is before you start dropping it. An embedded packet capture on the control plane helps:
monitor capture CAP control-plane in match any limit duration 60
monitor capture CAP start
show monitor capture CAP buffer brief
monitor capture CAP stop
Once the counters stop surprising you, switch the exceed actions to drop class by class, management and default first.
Keep it consistent
CoPP is exactly the kind of config that drifts: one router gets a new ACL line during an incident and nobody back-ports it. Keep the policy in your templating or config management source, and include show policy-map control-plane input in whatever you already collect for monitoring, so drop counters become a graph rather than something you discover during an outage. A rising drop count in class-default is often the first sign of a scan or a misbehaving neighbor.
Why this matters
Most control plane outages I have dealt with were not attacks. They were a broken monitoring poller, a layer-2 loop pushing ARP, or a neighbor in a flap loop, and the router could not protect its own routing sessions because everything shared one queue. CoPP does not make the box faster; it makes it decide what to sacrifice. Put it in carefully, observe before enforcing, and the next CPU spike costs you some SNMP polls instead of your adjacencies.