Add a reentrancy protection to EIP 1283

[forshtat]

UPD: As per @Arachnid 's request, moved into a separate EIP: #1706

An attack is described in:
https://medium.com/chainsecurity/constantinople-enables-new-reentrancy-attack-ace4088297d9

This is easily mitigated if SSTORE is not allowed in low gasleft state, without breaking the backward compatibility and the original intention of this EIP.

[sorpaas]

I would recommend to apply the gasleft check in the beginning, before any of the branches here are applied.

[forshtat]

Done, please see cf76bba

[veox]

IMO this would have been better off as a stand-alone EIP, so fork definitions in node impl-ns don't have to invent branching Constantinoples. As described here.

[k06a]

What do you think of replacing 200 gas cost with 2300 gas cost and 2100 gas refund?

[forshtat]

What do you think of replacing 200 gas cost with 2300 gas cost and 2100 gas refund?

I think that as refund is limited to half of the transaction gas cost, this would in some cases limit the benefits of this EIP. Why do you think this is a better solution?

[BrendanChou]

Recommend changing 2300 -> 1600. As noted in the ChainSecurity article, the CALL opcode uses 700 gas. 2300 would protect against all storage whereas 1600 protects against reentrancy attacks in particular.

[veox]

Recommend changing 2300 -> 1600. As noted in the ChainSecurity article, the CALL opcode uses 700 gas. 2300 would protect against all storage whereas 1600 protects against reentrancy attacks in particular.

2300 is a magic number inherited from G_CALLSTIPEND. 1600 is a magicnum inherited from G_CALLSTIPEND and G_CALL.

The former covers cases where G_CALL gets lower in the future, whereas the latter does not.

The former makes an existing uncodified invariant explicit, whereas the latter guards against one pattern only.

[Arachnid]

1283 is final and describes something implemented on test networks and in clients. If you'd like to propose an alternative version, please open a new EIP for that.