Force-disconnect peers that trigger errors in the node

[tomaka]

After peer slots are implemented (#1240), it is possible for all slots to be filled with nodes that serve bad blocks or proofs.

Smoldot should instead force-disconnect these nodes in order to free slots for others.

It is unclear whether the syncing code needs to help with that, or if the sync service should do it alone.

It is also unclear whether to ban nodes for a certain duration and how to implement this ban.

[melekes]

I've contributed to a similar thing before and the design was roughly: bad behavior of a certain peer is reported to p2p service (who receives data and proxies it to other services) from other services (sync, ...). Behavior was an enum ~ { Bad, VeryBad }.

Bad -> disconnect from peer & ban it for 10min
2nd Bad -> disconnect from peer & ban it for 30min
VeryBad -> disconnect and ban forever

[tomaka]

It is unclear whether the syncing code needs to help with that, or if the sync service should do it alone.

It is also unclear whether to ban nodes for a certain duration and how to implement this ban.

Implementing the ban on the networking part of the code should be rather easy, the problem is more about deciding which pieces of code should decide that certain nodes are to be banned.

I'm trying to not spaghettify the code with too many concepts at once.
In this context, adding the concept of "banning a node" to the syncing code might be off topic. We have this concept in the OptimisticSync already, but it is purely an implementation detail in the sense that the peer is actually ignored rather than banned.

In principle, the syncing code should simply indicate when a peer has done something wrong, and it is the higher level code that decides to ban the peer in return.

[melekes]

In principle, the syncing code should simply indicate when a peer has done something wrong, and it is the higher level code that decides to ban the peer in return.

Agree.

[tomaka]

In the same vein, peers that refuse us (usually because they are full) should also be temporarily banned.
Right now the full node takes a long time before it starts syncing because it keeps connecting to the same nodes over and over again that keeps refusing it.