Cause: As the realtime kernel replaces most of the spinlocks by rtmutexes, in order to maximize the preemption points inside the kernel, the locking scheme used in both NAPI polling and busy polling could become out of synchronization with the State Machine they protected. That could be observed when a machine with faster NICs (10g or 40g) was subject to a heavy pressure receiving network packets.
Consequence: The system could experience a livelock that would either slowly degrade the system performance up to a halt or disable the network adapter and halt.
Fix: the locking schemes on NAPI polling and busy polling routines were hardened to enforce the State machine sanity.
Result: the system works fine under pressure.