Network bonding and teaming with NetworkManager: modes, switch requirements and verification

What it is

Linux bonding (the bonding kernel driver) presents several physical NICs as one logical interface (bondX). NetworkManager can create and manage a bond connection with a bond connection type plus one ethernet connection per slave marked as its port, storing the kernel driver's options under the bond.options property (for example mode=802.3ad,miimon=100). Teaming (teamd, connection type team) is the older alternative; it is deprecated in RHEL 9 and removed in RHEL 10, so new setups should use bonding.

Why it matters

The bonding modes are not interchangeable, and the wrong choice either does nothing useful or breaks connectivity:

• active-backup: one slave active at a time, the rest standby; failover on link loss. Needs no switch configuration — each NIC can even connect to a different switch, which is also the mode's main resilience benefit.
• 802.3ad (LACP): slaves negotiate a dynamic link aggregation group with the switch using the LACP protocol; all slaves are normally connected to the same switch, and that switch's ports must be explicitly configured as a matching LACP group. Without that switch-side configuration, traffic is dropped or only one slave is effectively used.
• Other modes trade off load distribution against switch requirements differently: balance-rr, balance-xor and broadcast need the switch ports grouped as a static (non-LACP) EtherChannel/trunk, otherwise the switch sees the same MAC address flapping between ports; balance-tlb and balance-alb need no switch configuration.

How to apply

• Decide the mode based on what the switch team can configure: active-backup when no coordinated switch config is possible or available; 802.3ad when a proper LACP port-channel/EtherChannel can be set up.
• Create with nmcli: a bond connection carrying bond.options, then two or more ethernet connections with connection.master pointing at the bond and connection.slave-type bond.
• Verify the negotiated state, not just the configuration: cat /proc/net/bonding/bond0 shows the active mode, MII status per slave, and — for 802.3ad — the LACP partner and aggregator IDs actually agreed with the switch.
• Confirm failover behavior by taking one slave down (ip link set eth0 down, root) and checking that /proc/net/bonding/bond0 shows the switchover and that connectivity survives; bring it back with ip link set eth0 up. Run this test from a console, not over the bond itself — if failover does not work, the session is gone.

Pitfalls

• Configuring 802.3ad on the host while the switch ports remain in standalone (non-aggregated) mode: the link may still come up but traffic drops or is duplicated.
• Assuming /proc/net/bonding reflects intent from a config file; it only reflects what the driver negotiated right now — re-check after any switch-side change.