Thoughts on Address Space Extension (ASE)

Context

  1. Increasing address size from 20 to 32
    bytes
  2. ASE (Address Space Extension) with Translation
    Map
  3. ASE Test
    Cases
  4. Issues with ASE (with a translation
    map)
  5. Types of Resurrection Metadata in State
    Expiry
  6. Address Space Extension with bridge
    contracts
  7. Making ASE work with an Edict

I’ve thought a bit on address space extension and I wanted to clarify a few
things, propose some ideas and get some feedback.

In what follows, unless said otherwise, I’ll assume the period-aware address ASE
scheme (ASE-PWA) (defined in links 1 & 2 above, name from link 5). This is the
scheme that uses a translation map from compressed address to long adresses.

I don’t actually think this is the best scheme — I lean towards either keeping
the translation map but moving the period (aka epoch) identifier out of the address, or using
bridge contracts only. I’ll explain below.

Multiple Contexts

As mentionned in 5, the PWA-ASE scheme “creates two different EVM contexts,
legacy mode and extended mode”.

I think this is unavoidable if there is going to be any address extension at
all, and we should embrace the distinction between the two contexts to clarify
the semantics.

In particular, I propose that:

  • in legacy modes, all addresses are always 20 bytes
  • in extended mode, all addresses are always 32 bytes

In legacy mode, all opcodes that access addresses must be changed to consult the
translation map and access data under the long address if an entry is found. No
other change is required.

In extended mode, all addresses are 32 bytes, legacy addresses are simply
0-extended.

Extended mode can call legacy mode, in which case addresses passed to calls will
automatically be compressed. This must be done by the compiler, so we can’t
automatically compress custom calldata encodings and there is a possibility of
error there, although this only concerns new (extended mode) code. This kind of
things should be caught in testing anyway.

Should legacy mode be able to call extended mode? Some possible design choices:

  1. no
  2. yes, with automatic address decompression for calls
  3. like 2, but only if calling a bridge contract (somewaht similar to those of 6),
    which is just an extended mode contract with a special flag set to accept
    such calls.

1 might be too restrictive if we want to enable callbacks when an
extended-mode caller calls an legacy-mode contract. In 2, it would again be
the compiler’s job to properly decompress the received 20 bytes addresses, while
it leaves 32 bytes addresses alone. Again, this is not compatible with custom
calldata encoding. 3 is just 2 but forcing the contract developer to consider the
problem and say “yes I know what I’m doing”.

Multiple Addresses

Under the new scheme, each key pair is associated to (at least) three addresses.
These are referred as “personas” in 4.

  • one 20 byte legacy (“short”) address
  • one (or more?) 32 byte long address(es)
  • one (or more?) 20 byte compressed address(es)

A very important thing to remember here is that transactions never identify the
sender’s address, instead the signature can be used to recover the public key.

Something to contend with is that a key pair can potentially be associated with
an infinite amount of long addresses (corresponding to different periods). I
think there was some mention of whether a key should be allowed to control
multiple addresses like that of whether it should be locked to a single period.
I haven’t seen any discussion of how that would work and what it would entail,
tell me if I’m wrong.

The major related to multiple addresses stems from the “state transition” that
occurs whenever the long address becomes active (by receiving funds, or in the
future, maybe through some account abstraction mechanism).

Short + Long Address

(Here “short address” is not the compressed address but rather the last 20 bytes
of the long address.)

Unlike sending a transaction, receiving funds does mention the address
explicitly.

This means that we can have the following situation:

  1. I generate a new key pair
  2. I communicate the short address to Alice
  3. Alice sends funds to the short address
  4. (I can know move these funds using my private key, but I do not)
  5. I communite the long address to Bob
  6. Bob sends funds to the long address
  7. ?? can I still move funds on the short address ??

You can’t answer “yes” to that last question if you value the security goal of
address extension. To quote Vitalik in 1:

current 20 byte (160 bit) addresses only provide 80 bits of collision
resistance, meaning that someone can spend 2^80 computing work to generate
two pieces of contract init code (or (sender, ID) pairs, or one piece of
contract code and one EOA private key) that have the same address. 2^80 will
soon be within reach of sophisticated attackers; the bitcoin blockchain has
already made more than 2^90 hashes.

you can trust an address if either (i) it is on chain, or (ii) you personally
created it; for example, an organization cannot safely receive funds at a
multisig unless that multisig has already been published on-chain, because
whoever was the last to provide their public key could have chosen their
private key in such a way as to make the address also a valid EOA controlled
by themselves.

The security guarantee that address extension should provide is that if you
receive a 32 byte address, you know the above scenario is not possible. If you
answer “yes” to the question above (moving funds on the short address) then it
means you can still find a collision (between the short address and a legacy
EOA).

By the way, I don’t really understand why an address being on-chain would
preclude the attack described in the quote above. What prevents someone to
create a multisig that is also an EOA, then let the multisig be used “as
intended”, only to use the EOA to siphon off the funds afterwards? Please, let
me know!

So it looks like we’re forced to lose any funds left on the short address.

In fact things are even much worse than that. Say I generate a new keypair and
intend to use the short address only. I receive funds on it. Then Bob can just
make up a long address that ends with the short address, send one wei to it, and
lock me out of my funds.

The solution is relatively simple: mandate that once an address is used as a
legacy address (i.e. marked as LEGACY as in 7), it is not possible to
enter an extended address that ends with that short address into the translation
map. A bad consequence of this is that we can’t expire account state anymore
(we can still expire account storage).

UX-wise, users should choose in advance whether a keypair will be mapped to a
long or short address, and wallets should always show that address — never the
other address. For recovery from a BIP-39 seed phrase, the wallet must consult
the chain to determine whether the recover the short or long address.

Compressed + Long Addresses

Peter did a great job thinking through this case in 7.

To summarize briefly, problems can occur if you communicate the compressed
address before the compressed/long pair has been entered in the translation map.
Hence Peter proposes “the edict”, a set of conditions that ensures that this
entry will have been created, and thus avoids problematic scenarios if
respected.

There are two problematic scenarios.

In the first scenario is essentially the one in Vitalik’s quote from above, but
using a compressed address. An attacker generates a collision such that
compressed(long_address_a) == short_address_b (imagine that for instance
long_address_a is a multisig). If you send funds (e.g. ERC20 tokens) to
compressed(long_address_a) before it has been entered to the translation map,
then the attacker can make out with the funds.

If the compressed address had been in the translation map, then we could have
rejected a transaction with a signature for the compressed address.

As Peter says on the Discord:

There are two types of problems:
a) You violate The Edict sending and send something to the contract that has a collision and Chuck steals your funds.
b) You violate The Edict and you funds end up in long address land at encode(compress(long_address)) and you can’t get them back.
I (and most other people) aren’t very worried about a), because Chuck has to spend billions of dollars to generate the collision and then cross his fingers and hope you screw up. The attack isn’t even viable unless Chuck targets the richest whales anyway.

In the second scenario (the worrisome one), you make a mistake and pass
compressed(long_address) to an extended-mode contract, instead of simply
long_address.

For Ethereum transfers, this is easily avoided. If sending before the
translation map entry is established, we can merge the balances when the entry
is created. If after, we can simply look in the translation map for the correct
address.

For other use cases, things are more complicated. For instance, sending tokens
to compressed(long_address) with a extended-mode ERC20 contract will cause a
mapping from that compressed address to the balance to be created, whereas
normally the balance should be mapped to long_address directly.

Here are some ideas of how to fix this:

  1. automatic “address normalization”
  2. having contracts check both long_address and compressed(long_address)
  3. including a recovery function (devised by Peter in 7) in the contract

Automatic address normalization would cost gas, and would fail on custom
calldata parsing. Solution 2 would also cost gas, and relies on ERC20
implementation to actually do it. Solution 3 also relies on inclusion in
implementations, but only costs gas at contract deployment time. Including 2
or 3 int the OpenZeppelin implementations (and maybe a few frequently forked
repos) could help a lot.

Note that this may seem similar to the short address issue. But this time there
is a good reason for being able to see the compressed address: it will show up
in transaction data (and hence on etherscan & co, which people do use to check
up on their transactions).

ECRECOVER

Currently, the ECRECOVER precompile returns the (20-byte) address associated
with a signature (last 20 bytes of hash of the public key).

With period-aware addresses however, we can’t get the address from the public
key. We need to know the epoch (and potentially version, shard id, …).

For use in extended mode (for which there is no code written yet) we can just
change the precompile to take the epoch (& co) as parameter. Or change it to
return the public key.

The thorny problem is what to do with signatures generated by long addresses in
legacy mode.

There is only one way I can think to get this to work backward-compatibly under
PWA. And that is stuff the epoch (& co) in the signature itself (e.g. in v, just like
EIP-155 did for the chain ID. I don’t think v is limited to a certain size, so
this is certainly possible.

I don’t see any other way, but let me know otherwise.

If we decide to break ECRECOVER (bad idea imho), the question is: how much is
ECRECOVER used. How much would break if we change it?

Another concern is that we can’t “break it neatly”. Ideally, if we identify
that a signature was signed with long address public key, we would revert.
However, we cannot make this determination, because it would require compressing
the address, including the epoch which we do not know.

Do we need PWA?

I don’t think so? If we’re going to have the translation map stick around
forever, I don’t think it’s a big deal to add an epoch field in there.

If the address can stay purely derived from the public key, this keeps ECRECOVER
backward compatible: We can verify that a signature was generated by a long
address by compressed it and looking in the map, in which case we can return the
compressed address in legacy mode.

Additionally we don’t have to worry about having multiple addresses with
different epochs for the same public key. Now the epoch is metadata of the
address, just like the nonce. We’re also able to change the epoch associated
with an address easily.

Translation Map: Expirable? Size?

Can the translation map be expired? I don’t think so.

Many problem mitigations outlined above wrt compressed addresses relies being
able to tell that a public key matches a long address, or that a short address
is a compressed address. I’m not sure it’s feasible to do all that when that
information is missing.

To give you just an example, to avoid collision attacks on compressed addresses,
we must forbid transactions signed by a public key whose hash ends with an
existing compressed address.

If the translation map can’t be expired, we should worry about its size. Ipsilon
reports 150M addresses in June, and estimates 150M over two years.
Over on Discord, potus did an exponential fitting on the address data starting
July 2018 and predicts about twice more (~300M) over two years.

Ipsilon estimates ~9GB for 150M addresses.

The obvious question is whether it’s worth it in terms of state expiry (though
of course that’s not the only goal of ASE).

It might be interesting to see what the numbers would have been if address
extension had been introduced two years ago, although that might be
unrepresentative given both high growth + the big push in DeFi only happening
only a year ago.

Nevertheless the current state size is about 30-40GB. How much of that has been
inactive for more than 1 year? Two years?

Bridge-Only

One alternative to the translation map approach is to simply forbid long
addresses to interact with legacy contracts.

Legacy contracts could be allowed to call extended contracts, but that is not
strictly required.

We need a way to bridge over between the two worlds. Those are the bridge
contracts
proposed by Alex. Bridge contracts are essentially
extended-mode contracts that can call legacy contracts, and have a legacy address.

Let’s take the example of an ERC20 token. Initially, the token is handled by a
legacy, and so a long address can’t own the token. Say we want to update to an
extended contract. A possible implementation is that the extended contract will
transfer the balance of the caller to itself in the legacy contract, then credit
the caller in the extended contract.

As far as I understand, the idea of a bridge contract is that (a) it lets an
extended-mode contract have a legacy address and (b) acts as a marker for the
developer to acknowledge that legacy-extended mode communication is tricky and
should be handled with care.

This proposal has the advantage of being much simpler.

The downside is that adoption of long addresses may be sluggish. People don’t
want to make long addresses because contracts are not compatible. Because people
don’t have long addresses, project are in no hurry to upgrade.

Only the extra cost of state resurrection is a foolproof incentive to migrate.
But in the proposed scheme this only kicks in after two years, at which point
creating new state becomes costlier because a proof of absence needs to be
provided. Two years is a long time in blockchains, so most important projects
might have updated by then, but it stays likely that in the meantime, most new
addresses will continue being legacy addresses.

More

Something I’d like to see if we want to adopt state expiry with ASE-PWA is how
we could help long-lived contracts handle storage over many epochs (without
incurring undue proof costs when writing new state).

Recap

This was long, so let me summarize the main points & questions:

Points:

  • Let’s think of legacy mode as “addresses are always 20 bytes” and extended mode as
    “addresses are always 32 bytes”.
  • Avoiding an attack that locks people out of their funds requires never
    expiring account state (!= account storage).
  • The annoying gotcha of PWA-ASE is losing funds when accidentally interpreting
    a compressed address as a long address. This can be avoided by following the
    edict
    . Some in-code mitigations are possible.
  • To make ECRECOVER work with PWA, we need to stuff the epoch in the signature.
  • ECRECOVER cannot be broken neatly.
  • If the translation map is non-expirable, we might as well store the epoch in
    there.
  • I don’t see how we possibly can exipre the translation map.
  • The space requirement for the translation map is significant (10-20GB in two years).
  • If we’re not in a hurry for people to adopt long addresses, the bridge method
    is much simpler.

Questions:

  • In PWA-ASE, should legacy contracts be able to call extended contracts, and
    how?
  • In PWA-ASE, how to contend with the fact that a public key could be associated
    with multiple addresses with different periods?
  • In the quote from Vitalik, why does the address being on-chain preclude an attack?
  • How much/importantly is ECRECOVER used in practice? Will horrible things
    result if it returns bogus legacy when passed a signature signed by a long address?
  • How much space does PWA-ASE save using state expiry (given that the
    translation map is persisted forever)?
  • In state expiry, with PWA-ASE, how would long-lived contracts handle storage
    over many epochs?

Depending on what option is chosen for bridge contracts, the existence of them may be more widespread than not. I could imagine that basically any new EOF-style (EIP-3540) contract is a “bridge contract” (i.e. is expected to understand long addresses), but it would remain possible to create EOF-style contracts under both legacy (which in effect is the bridge contract) and long addresses.

This means most contracts during those 2 years likely would be long address compatible already.

One general problem with long addresses is that we need to introduce new types in the contract ABI (what was called address32 in the bridge writeup), which in turn means new types in Solidity and tooling, and most importantly, brand new (token) standards. The creation and maturity of these standards may take quite a bit of time.

Unless we want to be stuck with the translation map forever, these long address-capable standards are needed no matter which ASE method is chosen.

Hence, the longer ramp up time with bridge contracts may actually be a blessing in disguise: it gives a longer transition period during which these problems can be figured out, without forcing a brittle scheme (=translation) in one go.

This means most contracts during those 2 years likely would be long address compatible already.

I agree. And even in PWA-ASE, is there any reason not to make new contracts in extended space (besides marginally higher memory requirement), given that legacy addresses can interact with them?

One general problem with long addresses is that we need to introduce new types in the contract ABI (what was called address32 in the bridge writeup), which in turn means new types in Solidity and tooling, and most importantly, brand new (token) standards. The creation and maturity of these standards may take quite a bit of time.

Doesn’t feel overwhelming to me, but it’s just a feeling. Standard-wise, isn’t this as simple as s/address20/address32 in the existing standard?

It is simple to change code, but there are already many competing token standards, each took a lot of deliberation and trying to improve various aspects of the preceding ones. Someone needs to create a new one, and likely there would be new competing ones or people seizing the opportunity to improve other aspects of existing standards.

In any case, it won’t happen overnight, and therefore any migration will happen over a period of time. Unless deploying legacy contracts is disallowed soon, which likely would cause a lot of hassle for everyone.

1 Like

The main difference with the bridge one is that one can create long address aware contracts in both namespaces. It becomes possible for new projects to support token standards of both worlds, making the eventual transition smoother.

With the translation map the incentive to start supporting new token standards is simply not there. It kind of kicks the can down the road.