There appear to be a few bugs in the new dpdk cryptodev code related to numa domains. The start of these is probably in:

```
dpdk_cryptodev_init (vlib_main_t * vm)

{ ... /* probe all cryptodev devices and get queue info */ if (cryptodev_probe (vm, n_workers) < 0) ... }

```

This code is gonig to be invoked once for the main thread vm. However, cryptodev_probe then uses the numa node of that vm (main thread) and the total number of workers to start creating and configuring cryptodevs, all of this is being done using only the numa node of the main thread. The workers very well may be placed in a different numa domain, and so the bugs cascade from there.