Disallow attestation production earlier than head (#2130)

## Issue Addressed

The non-finality period on Pyrmont between epochs [`9114`](https://pyrmont.beaconcha.in/epoch/9114) and [`9182`](https://pyrmont.beaconcha.in/epoch/9182) was contributed to by all the `lighthouse_team` validators going down. The nodes saw excessive CPU and RAM usage, resulting in the system to kill the `lighthouse bn` process. The `Restart=on-failure` directive for `systemd` caused the process to bounce in ~10-30m intervals.

Diagnosis with `heaptrack` showed that the `BeaconChain::produce_unaggregated_attestation` function was calling `store::beacon_state::get_full_state` and sometimes resulting in a tree hash cache allocation. These allocations were approximately the size of the hosts physical memory and still allocated when `lighthouse bn` was killed by the OS.

There was no CPU analysis (e.g., `perf`), but the `BeaconChain::produce_unaggregated_attestation` is very CPU-heavy so it is reasonable to assume it is the cause of the excessive CPU usage, too.

## Proposed Changes

`BeaconChain::produce_unaggregated_attestation` has two paths:

1. Fast path: attesting to the head slot or later.
2. Slow path: attesting to a slot earlier than the head block.

Path (2) is the only path that calls `store::beacon_state::get_full_state`, therefore it is the path causing this excessive CPU/RAM usage.

This PR removes the current functionality of path (2) and replaces it with a static error (`BeaconChainError::AttestingPriorToHead`).

This change reduces the generality of `BeaconChain::produce_unaggregated_attestation` (and therefore [`/eth/v1/validator/attestation_data`](https://ethereum.github.io/eth2.0-APIs/#/Validator/produceAttestationData)), but I argue that this functionality is an edge-case and arguably a violation of the [Honest Validator spec](https://github.com/ethereum/eth2.0-specs/blob/dev/specs/phase0/validator.md).

It's possible that a validator goes back to a prior slot to "catch up" and submit some missed attestations. This change would prevent such behaviour, returning an error. My concerns with this catch-up behaviour is that it is:

- Not specified as "honest validator" attesting behaviour.
- Is behaviour that is risky for slashing (although, all validator clients *should* have slashing protection and will eventually fail if they do not).
- It disguises clock-sync issues between a BN and VC.

## Additional Info

It's likely feasible to implement path (2) if we implement some sort of caching mechanism. This would be a multi-week task and this PR gets the issue patched in the short term. I haven't created an issue to add path (2), instead I think we should implement it if we get user-demand.

beacon_node/beacon_chain/src/beacon_chain.rs

@@ -853,37 +853,22 @@ impl<T: BeaconChainTypes> BeaconChain<T> {
                 Cow::Borrowed(&head.beacon_state),
             )
         } else {
-            // Note: this method will fail if `slot` is more than `state.block_roots.len()` slots
-            // prior to the head.
+            // We disallow producing attestations *prior* to the current head since such an
+            // attestation would require loading a `BeaconState` from disk. Loading `BeaconState`
+            // from disk is very resource intensive and proposes a DoS risk from validator clients.
             //
-            // This seems reasonable, producing an attestation at a slot so far
-            // in the past seems useless, definitely in mainnet spec. In minimal spec, when the
-            // block roots only contain two epochs of history, it's possible that you will fail to
-            // produce an attestation that would be valid to be included in a block. Given that
-            // minimal is only for testing, I think this is fine.
+            // Although we generally allow validator clients to do things that might harm us (i.e.,
+            // we trust them), sometimes we need to protect the BN from accidental errors which
+            // could cause it significant harm.
             //
-            // It is important to note that what's _not_ allowed here is attesting to a slot in the
-            // past. You can still attest to a block an arbitrary distance in the past, just not as
-            // if you are in a slot in the past.
-            let beacon_block_root = *head.beacon_state.get_block_root(slot)?;
-            let state_root = *head.beacon_state.get_state_root(slot)?;
-
-            // Avoid holding a lock on the head whilst doing database reads. Good boi functions
-            // don't hog locks.
-            drop(head);
-
-            let mut state = self
-                .get_state(&state_root, Some(slot))?
-                .ok_or(Error::MissingBeaconState(state_root))?;
-
-            state.build_committee_cache(RelativeEpoch::Current, &self.spec)?;
-
-            self.produce_unaggregated_attestation_for_block(
-                slot,
-                index,
-                beacon_block_root,
-                Cow::Owned(state),
-            )
+            // This case is particularity harmful since the HTTP API can effectively call this
+            // function an unlimited amount of times. If `n` validators all happen to call it at
+            // the same time, we're going to load `n` states (and tree hash caches) into memory all
+            // at once. With `n >= 10` we're looking at hundreds of MB or GBs of RAM.
+            Err(Error::AttestingPriorToHead {
+                head_slot: head.beacon_block.slot(),
+                request_slot: slot,
+            })
         }
     }
 

beacon_node/beacon_chain/src/errors.rs

@@ -92,6 +92,10 @@ pub enum BeaconChainError {
     },
     WeakSubjectivtyVerificationFailure,
     WeakSubjectivtyShutdownError(TrySendError<&'static str>),
+    AttestingPriorToHead {
+        head_slot: Slot,
+        request_slot: Slot,
+    },
 }
 
 easy_from_to!(SlotProcessingError, BeaconChainError);