Loop Prevention Mechanics

## Why switched networks contain loops at all A **switch** (the IEEE standards call it a *bridge*) relays Ethernet frames between its ports. Network designers deliberately connect switches with **more than one path**: two uplinks from an access switch, a ring of switches, a mesh between distribution switches. The reason is resilience: if one cable, optic or switch fails, traffic still has a way through. But any redundant path closes a **loop**, a cycle in the physical topology, and Ethernet cannot tolerate one on its own. ## What the Ethernet header lacks An Ethernet frame header holds three things a switch looks at: the **destination MAC address**, the **source MAC address** and the **EtherType** (an 802.1Q tag adds a VLAN ID and priority, nothing more). There is **no hop count and no time-to-live field**. RFC 6325, the IETF's TRILL specification, states the problem directly in its introduction: the Ethernet header contains no hop count or TTL, which is dangerous whenever a temporary loop forms. Two switch behaviours turn that missing field into an outage: - **Flooding.** A switch sends a broadcast frame, and any frame whose destination it has not learned, out of **every port except the one it arrived on**. In a loop, "every other port" includes the way back round. - **No memory of frames.** A switch keeps a table of *addresses*, not of *frames it has already relayed*. A copy that comes back round looks like a brand-new frame and is flooded again. Put together, a single ARP request (which RFC 826 has a host broadcast to every station on the segment) enters the loop and nothing ever expires it. Each pass re-delivers it to every host port, and every further broadcast adds more copies that also never leave. ## Why a Layer 3 loop is bounded Routing loops happen too, for example while routers disagree about a failed link. They hurt, but they burn out: | Property | Ethernet (Layer 2) | IP routing (Layer 3) | |---|---|---| | Lifetime field | none | IPv4 `TTL`, IPv6 `Hop Limit`, 8 bits each | | What a hop does | relays the frame unchanged | decrements the field, discards at zero | | Broadcasts | flooded across every switch in the VLAN | routers must not forward limited broadcasts (RFC 1812) | | Outcome of a loop | copies circulate until the loop is cut | each packet dies after at most 255 hops | RFC 791 says a datagram whose TTL reaches zero must be destroyed, and RFC 8200 says an IPv6 packet is discarded when its `Hop Limit` is zero or decremented to zero. So a routing loop wastes some bandwidth and drops the affected traffic; a switching loop can take down **the whole broadcast domain**, because every host on it receives the storm. ## What a loop does to a network The symptoms always come together: - **Broadcast storm**: looping broadcasts saturate the inter-switch links and every access port in the VLAN. - **Duplicate frames**: a frame flooded down two paths arrives at its destination twice. - **MAC address flapping**: a switch sees the same source address arriving on different ports and keeps moving its table entry, so traffic for that host is sent the wrong way. - **CPU exhaustion**: hosts and the switches' own management processors are interrupted by every broadcast copy. ## How spanning tree fixes it **Spanning tree** (IEEE 802.1D; its rapid successor RSTP was folded into 802.1D-2004) keeps the physical redundancy but makes the *logical* topology a tree: 1. The switches exchange **BPDUs** (bridge protocol data units), the control messages that let them discover one another and the shape of the network. 2. One switch becomes the **root**, and every other switch keeps its best path towards it. 3. Every port that would close a loop is put into a **blocking** condition: it discards data frames but keeps listening for BPDUs. 4. When an active link fails, the BPDUs stop or change, and a blocked port moves to forwarding so connectivity returns. The result is exactly **one active path between any two segments**, so a flooded frame reaches every host once and then stops at the edge. ## What it costs Blocking is not free. RFC 7348 (VXLAN, an Informational RFC) notes that with spanning tree, operators pay for links they cannot use, and RFC 6325 points out that forwarding follows whatever path survives the blocking rather than the shortest path between two switches. Link aggregation, routed access and overlay fabrics are the designs that recover that capacity, and each still depends on keeping whatever Layer 2 remains loop-free.