Don't disconnect reserved peers if their reputation falls below threshold

[tomaka]

When the reputation of peer falls below the banning threshold, we currently unconditionally disconnect it from all sets. In other words, we close all its substreams. This happens even if the peer is reserved.

However, the alloc_slots functions, which allocates peers to slots, doesn't take the reputation into account when it comes to reserved peers. It always opens a substream to all reserved nodes if there isn't one.

So what happens when the reputation of a reserved peer drops too low is that it gets disconnected, but alloc_slots later gets called at an indefinite point in time (when another peer disconnects from that set), and the substream will be re-opened.
If, then, we do a small insubstantial reputation decrease (we sometimes do small reputation decreases in order to account for heavy requests) while the reserved node is still under the ban threshold, it will be disconnected again.

This behaviour can be rather surprising, so this PR prevents reserved nodes from being disconnected if their reputation falls too low.

It also calls alloc_slots immediately if the peer gets disconnected, as this was an oversight.

[bkchr]

Do we keep all the relay chain validators as reserved? If yes, what happens if one of these validators really starts to spam all other validators?

[ordian]

Do we keep all the relay chain validators as reserved? If yes, what happens if one of these validators really starts to spam all other validators?

Yes and I have the same concern. Either we need to change Polkadot not to use reserved slots for every relay chain validator or change substrate not to treat reserved peers differently when it comes to reputation. It seems to the latter change would be easier to implement.

[ordian]

Closing as #9020 is merged.