OK a new set, including IRQ balancing. There are debug leftovers and dead code, but it should work. I did the whole wave300 wifi development with this ethernet driver version. The IRQ balancing is 2 days old, but it seems to work (better than my previous version). I didn't test the both patchsets separated (as it don't have reason to have slow ethernet on 2 VPEs), altought I did test the updated ethernet driver on a single VPE (nosmp kernel parameter). If you want to test it yourself, it should be possible just by splitting the patchsets into 901-903 and 904-906.
BTW All older patches from me are obsolete by this. The patchset is for (older) kernel 4.14.99. Apply them by copying into target/linux/lantiq/patches-4.14 and by recompiling the kernel.
If you got some kernel warnings/oopses, please send the debug log (IRQ has this thread, but I guess it doesn't matter that much).
I didn't test the compatibility of IRQ driver with those devices, which requires one VPE to be disabled (and later some proprietary driver runs on it). If you have these, PLEASE test. Same it seems (from device-tree) the IRQ driver is used by Danube, AR9, AmazonSE and Falcon devices. These won't work with patched device-tree for IRQ patchset (you can test them by changing the "reg" field accordingly) or contact me (I don't know about its hardware specs).
The benchmark is in the tarball. But it seems the ethernet is over 350 Mbps (host to xrx200) and over 260 Mbps (xrx200 to host) (iperf3, TCP traffic). The vanilla kernel has only 178/81 Mbps.
P.S. I didn't use the irqbalancer before (it would bring a nondeterminism into the debug setup), but it seems it may not even work o_O. So I've changed the IRQ driver to autobalance IRQ itself (the code is actually cleaner 
). Even with this, you can still limit the VPEs by /proc/irq/X/smp_affinity.
Patchset: https://turbo.to/rtnkmlreltcf.html
Mirror: https://ufile.io/dbw67