This affects all tests that contain l2-input graph node, unless performance of that test still hits the pps/bps limit even with this regression (or the node is not in the VPP instance which is the bottleneck, e.g. when VPP is inside VM but the bottleneck is vswitch).

The cause was tracked to 34568 [0].
Several reverts were attempted, but not applied as they do not restore the full performance or change the performance upon rebase.

One very clear symptom of this regression is "L1 miss/pkt and L2 hit/pkt increase from ~0.2 to ~1.2" for l2-input graph node in l2xcbase test.

[0] https://gerrit.fd.io/r/c/vpp/+/34568/2#message-63e1cdafce46e7d389ffb45d79df4fb622681888