Consensus flaw during block processing in github.com/ethereum/go-ethereum

Impact

A consensus-vulnerability in Geth could cause a chain split, where vulnerable versions refuse to accept the canonical chain.

Description

A flaw was repoted at 2020-08-11 by John Youngseok Yang (Software Platform Lab), where a particular sequence of transactions could cause a consensus failure.

• Tx 1:

  • sender invokes caller.
  • caller invokes 0xaa. 0xaa has 3 wei, does a self-destruct-to-self
  • caller does a 1 wei -call to 0xaa, who thereby has 1 wei (the code in 0xaa still executed, since the tx is still ongoing, but doesn't redo the selfdestruct, it takes a different path if callvalue is non-zero)

• Tx 2:

  • sender does a 5-wei call to 0xaa. No exec (since no code).

In geth, the result would be that 0xaa had 6 wei, whereas OE reported (correctly) 5 wei. Furthermore, in geth, if the second tx was not executed, the 0xaa would be destructed, resulting in 0 wei. Thus obviously wrong.

It was determined that the root cause was this commit from this PR. The semantics of createObject was subtly changd, into returning a non-nil object (with deleted=true) where it previously did not if the account had been destructed. This return value caused the new object to inherit the old balance:

func (s *StateDB) CreateAccount(addr common.Address) {
	newObj, prev := s.createObject(addr)
	if prev != nil {
		newObj.setBalance(prev.data.Balance)
	}
}

It was determined that the minimal possible correct fix was

+++ b/core/state/statedb.go
@@ -589,7 +589,10 @@ func (s *StateDB) createObject(addr common.Address) (newobj, prev *stateObject)
                s.journal.append(resetObjectChange{prev: prev, prevdestruct: prevdestruct})
        }
        s.setStateObject(newobj)
-       return newobj, prev
+       if prev != nil && !prev.deleted {
+               return newobj, prev
+       }
+       return newobj, nil