Possible optimizations

In this section we describe two alternative solutions that come with some optimizations.

Transaction counter

Instead of relying on a hash (32 bytes) Namada could use a 64 bits (8 bytes) transaction counter as nonce for the wrapper and inner transactions. The advantage is that the space required would be much less since only two 8 bytes values in storage are needed for every address which is signing transactions. On the other hand, the handling of the counter for the inner transaction is performed entirely in wasm (transactions and VPs) making it slightly less efficient. This solution also imposes a strict ordering on the transactions issued by a same address.

InnerTx

We will implement the protection entirely in Wasm: the check of the counter will be carried out by the validity predicates while the actual writing of the counter in storage will be done by the transactions themselves.

To do so, the SignedTxData attached to the transaction will hold the current value of the counter in storage:

pub struct SignedTxData {
  /// The original tx data bytes, if any
  pub data: Option<Vec<u8>>,
  /// The optional transaction counter for replay protection
  pub tx_counter: Option<u64>,
  /// The signature is produced on the tx data concatenated with the tx code
  /// and the timestamp.
  pub sig: common::Signature,
}

The counter must reside in SignedTxData and not in the data itself because this must be checked by the validity predicate which is not aware of the specific transaction that took place but only of the changes in the storage; therefore, the VP is not able to correctly deserialize the data of the transactions since it doesn't know what type of data the bytes represent.

The counter will be signed as well to protect it from tampering and grant it the same guarantees explained at the beginning of this document.

The wasm transaction will simply read the value from storage and increase its value by one. The target key in storage will be the following:

/$Address/inner_tx_counter: u64

The VP of the source address will then check the validity of the signature and, if it's deemed valid, will proceed to check if the pre-value of the counter in storage was equal to the one contained in the SignedTxData struct and if the post-value of the key in storage has been incremented by one: if any of these conditions doesn't hold the VP will discard the transactions and prevent the changes from being applied to the storage.

In the specific case of a shielded transfer, since MASP already comes with replay protection as part of the Zcash design (see the MASP specs and Zcash protocol specs (opens in a new tab)), the counter in SignedTxData is not required and therefore should be optional.

To implement replay protection for the inner transaction we will need to update all the VPs checking the transaction's signature to include the check on the transaction counter: at the moment the vp_user validity predicate is the only one to update. In addition, all the transactions involving SignedTxData should increment the counter.

WrapperTx

To protect this transaction we can implement an in-protocol mechanism. Since the wrapper transaction gets signed before being submitted to the network, we can leverage the tx_counter field of the SignedTxData already introduced for the inner tx.

In addition, we need another counter in the storage subspace of every address:

/$Address/wrapper_tx_counter: u64

where $Address is the one signing the transaction (the same implied by the pk field of the WrapperTx struct).

The check will consist of a signature check first followed by a check on the counter that will make sure that the counter attached to the transaction matches the one in storage for the signing address. This will be done in the process_proposal function so that validators can decide whether the transaction is valid or not; if it's not, then they will discard the transaction and skip to the following one.

At last, in finalize_block, the ledger will update the counter key in storage, increasing its value by one. This will happen when the following conditions are met:

• process_proposal has accepted the tx by validating its signature and transaction counter
• The tx was correctly applied in finalize_block (for WrapperTx this simply means inclusion in the block and gas accounting)

Now, if a malicious user tried to replay this transaction, the tx_counter in the struct would no longer be equal to the one in storage and the transaction would be deemed invalid.

Implementation details

In this section we'll talk about some details of the replay protection mechanism that derive from the solution proposed in this section.

Storage counters

Replay protection will require interaction with the storage from both the protocol and Wasm. To do so we can take advantage of the StorageRead and StorageWrite traits to work with a single interface.

This implementation requires two transaction counters in storage for every address, so that the storage subspace of a given address looks like the following:

/$Address/wrapper_tx_counter: u64
/$Address/inner_tx_counter: u64

An implementation requiring a single counter in storage has been taken into consideration and discarded because that would not support batching; see the relative section for a more in-depth explanation.

For both the wrapper and inner transaction, the increase of the counter in storage is an important step that must be correctly executed. First, the implementation will return an error in case of a counter overflow to prevent wrapping, since this would allow for the replay of previous transactions. Also, we want to increase the counter as soon as we verify that the signature, the chain id and the passed-in transaction counter are valid. The increase should happen immediately after the checks because of two reasons:

• Prevent replay attack of a transaction in the same block
• Update the transaction counter even in case the transaction fails, to prevent a possible replay attack in the future (since a transaction invalid at state Sx could become valid at state Sn where n > x)

For WrapperTx, the counter increase and fee accounting will per performed in finalize_block (as stated in the relative section).

For InnerTx, instead, the logic is not straightforward. The transaction code will be executed in a Wasm environment (Wasmer (opens in a new tab)) till it eventually completes or raises an exception. In case of success, the counter in storage will be updated correctly but, in case of failure, the protocol will discard all of the changes brought by the transactions to the write-ahead-log, including the updated transaction counter. This is a problem because the transaction could be successfully replayed in the future if it will become valid.

The ideal solution would be to interrupt the execution of the Wasm code after the transaction counter (if any) has been increased. This would allow performing a first run of the involved VPs and, if all of them accept the changes, let the protocol commit these changes before any possible failure. After that, the protocol would resume the execution of the transaction from the previous interrupt point until completion or failure, after which a second pass of the VPs is initiated to validate the remaining state modifications. In case of a VP rejection after the counter increase there would be no need to resume execution and the transaction could be immediately deemed invalid so that the protocol could skip to the next tx to be executed. With this solution, the counter update would be committed to storage regardless of a failure of the transaction itself.

Unfortunately, at the moment, Wasmer doesn't allow yielding (opens in a new tab) from the execution.

In case the transaction went out of gas (given the gas_limit field of the wrapper), all the changes applied will be discarded from the WAL and will not affect the state of the storage. The inner transaction could then be rewrapped with a correct gas limit and replayed until the expiration time has been reached.

Batching and transaction ordering

This replay protection technique supports the execution of multiple transactions with the same address as source in a single block. Actually, the presence of the transaction counters and the checks performed on them now impose a strict ordering on the execution sequence (which can be an added value for some use cases). The correct execution of more than one transaction per source address in the same block is preserved as long as:

1. The wrapper transactions are inserted in the block with the correct ascending order
2. No hole is present in the counters' sequence
3. The counter of the first transaction included in the block matches the expected one in storage

The conditions are enforced by the block proposer who has an interest in maximizing the amount of fees extracted by the proposed block. To support this incentive, validators will reject the block proposed if any of the included wrapper transactions are invalid, effectively incentivizing the block proposer to include only valid transactions and correctly reorder them to gain the fees.

In case of a missing transaction causes a hole in the sequence of transaction counters, the block proposer will include in the block all the transactions up to the missing one and discard all the ones following that one, effectively preserving the correct ordering.

Correctly ordering the transactions is not enough to guarantee the correct execution. As already mentioned in the WrapperTx section, the block proposer and the validators also need to access the storage to check that the first transaction counter of a sequence is actually the expected one.

The entire counter ordering is only done on the WrapperTx: if the inner counter is wrong then the inner transaction will fail and the signer of the corresponding wrapper will be charged with fees. This incentivizes submitters to produce valid transactions and discourages malicious user from rewrapping and resubmitting old transactions.

Mempool checks

As a form of optimization to prevent mempool spamming, some of the checks that have been introduced in this document will also be brought to the mempool_validate function. Of course, we always refer to checks on the WrapperTx only. More specifically:

• Check the ChainId field
• Check the signature of the transaction against the pk field of the WrapperTx
• Perform a limited check on the transaction counter

Regarding the last point, mempool_validate will check if the counter in the transaction is >= than the one in storage for the address signing the WrapperTx. A complete check (checking for strict equality) is not feasible, as described in the relative section.

Alternatives considered

In this section we list some possible solutions that were taken into consideration during the writing of this solution but were eventually discarded.

Mempool counter validation

The idea of performing a complete validation of the transaction counters in the mempool_validate function was discarded because of a possible flaw.

Suppose a client sends five transactions (counters from 1 to 5). The mempool of the next block proposer is not guaranteed to receive them in order: something on the network could shuffle the transactions up so that they arrive in the following order: 2-3-4-5-1. Now, since we validate every single transaction to be included in the mempool in the exact order in which we receive them, we would discard the first four transactions and only accept the last one, that with counter 1. Now the next block proposer might have the four discarded transactions in its mempool (since those were not added to the previous block and therefore not evicted from the other mempools, at least they shouldn't, see block rejection) and could therefore include them in the following block. But still, a process that could have ended in a single block actually took two blocks. Moreover, there are two more issues:

• The next block proposer might have the remaining transactions out of order in his mempool as well, effectively propagating the same issue down to the next block proposer
• The next block proposer might not have these transactions in his mempool at all

Finally, transactions that are not allowed into the mempool don't get propagated to the other peers, making their inclusion in a block even harder. It is instead better to avoid a complete filter on the transactions based on their order in the mempool: instead we are going to perform a simpler check and then let the block proposer rearrange them correctly when proposing the block.

In-protocol protection for InnerTx

An alternative implementation could place the protection for the inner tx in protocol, just like the wrapper one, based on the transaction counter inside SignedTxData. The check would run in process_proposal and the update in finalize_block, just like for the wrapper transaction. This implementation, though, shows two drawbacks:

• it implies the need for an hard fork in case of a modification of the replay protection mechanism
• it's not clear who's the source of the inner transaction from the outside, as that depends on the specific code of the transaction itself. We could use specific whitelisted txs set to define when it requires a counter (would not work for future programmable transactions), but still, we have no way to define which address should be targeted for replay protection (blocking issue)

In-protocol counter increase for InnerTx

In the storage counter section we mentioned the issue of increasing the transaction counter for an inner tx even in case of failure. A possible solution that we took in consideration and discarded was to increase the counter from protocol in case of a failure.

This is technically feasible since the protocol is aware of the keys modified by the transaction and also of the results of the validity predicates (useful in case the transaction updated more than one counter in storage). It is then possible to recover the value and reapply the change directly from protocol. This logic though, is quite dispersive, since it effectively splits the management of the counter for the InnerTx among Wasm and protocol, while our initial intent was to keep it completely in Wasm.

Single counter in storage

We can't use a single transaction counter in storage because this would prevent batching.

As an example, if a client (with a current counter in storage holding value 5) generates two transactions to be included in the same block, signing both the outer and the inner (default behavior of the client), it would need to generate the following transaction counters:

[
    T1: (WrapperCtr: 5, InnerCtr: 6),
    T2: (WrapperCtr: 7, InnerCtr: 8)
]

Now, the current execution model of Namada includes the WrapperTx in a block first to then decrypt and execute the inner tx in the following block (respecting the committed order of the transactions). That would mean that the outer tx of T1 would pass validation and immediately increase the counter to 6 to prevent a replay attack in the same block. Now, the outer tx of T2 will be processed but it won't pass validation because it carries a counter with value 7 while the ledger expects 6.

To fix this, one could think to set the counters as follows:

[
    T1: (WrapperCtr: 5, InnerCtr: 7),
    T2: (WrapperCtr: 6, InnerCtr: 8)
]

This way both the transactions will be considered valid and executed. The issue is that, if the second transaction is not included in the block (for any reason), than the first transaction (the only one remaining at this point) will fail. In fact, after the outer tx has correctly increased the counter in storage to value 6 the block will be accepted. In the next block the inner transaction will be decrypted and executed but this last step will fail since the counter in SignedTxData carries a value of 7 and the counter in storage has a value of 6.

To cope with this there are two possible ways. The first one is that, instead of checking the exact value of the counter in storage and increasing its value by one, we could check that the transaction carries a counter >= than the one in storage and write this one (not increase) to storage. The problem with this is that it the lack of support for strict ordering of execution.

The second option is to keep the usual increase strategy of the counter (increase by one and check for strict equality) and simply use two different counters in storage for each address. The transaction will then look like this:

[
    T1: (WrapperCtr: 5, InnerCtr: 5),
    T2: (WrapperCtr: 6, InnerCtr: 6)
]

Since the order of inclusion of the WrapperTxs forces the same order of the execution for the inner ones, both transactions can be correctly executed and the correctness will be maintained even in case T2 didn't make it to the block (note that the counter for an inner tx and the corresponding wrapper one don't need to coincide).

Wrapper-bound InnerTx

The solution is to tie an InnerTx to the corresponding WrapperTx. By doing so, it becomes impossible to rewrap an inner transaction and, therefore, all the attacks related to this practice would be unfeasible. This mechanism requires even less space in storage (only a 64 bit counter for every address signing wrapper transactions) and only one check on the wrapper counter in protocol. As a con, it requires communication between the signer of the inner transaction and that of the wrapper during the transaction construction. This solution also imposes a strict ordering on the wrapper transactions issued by a same address.

To do so we will have to change the current definition of the two tx structs to the following:

pub struct WrapperTx {
  /// The fee to be payed for including the tx
  pub fee: Fee,
  /// Used to determine an implicit account of the fee payer
  pub pk: common::PublicKey,
  /// Max amount of gas that can be used when executing the inner tx
  pub gas_limit: GasLimit,
  /// Lifetime of the transaction, also determines which decryption key will be used
  pub expiration: DateTimeUtc,
  /// Chain identifier for replay protection
  pub chain_id: ChainId,
  /// Transaction counter for replay protection
  pub tx_counter: u64,
  /// the encrypted payload
  pub inner_tx: EncryptedTx,
}
 
pub struct Tx {
  pub code: Vec<u8>,
  pub data: Option<Vec<u8>>,
  pub timestamp: DateTimeUtc,
  pub wrapper_commit: Option<Hash>,
}

The Wrapper transaction no longer holds the inner transaction hash while the inner one now holds a commit to the corresponding wrapper tx in the form of the hash of a WrapperCommit struct, defined as:

pub struct WrapperCommit {
  pub pk: common::PublicKey,
  pub tx_counter: u64,
  pub expiration: DateTimeUtc,
  pub chain_id: ChainId,
}

The pk-tx_counter couple contained in this struct, uniquely identifies a single WrapperTx (since a valid tx_counter is unique given the address) so that the inner one is now bound to this specific wrapper. The remaining fields, expiration and chain_id, will tie these two values given their importance in terms of safety (see the relative section). Note that the wrapper_commit field must be optional because the WrapperTx struct itself gets converted to a Tx struct before submission but it doesn't need any commitment.

Both the inner and wrapper tx get signed on their hash, as usual, to prevent tampering with data. When a wrapper gets processed by the ledger, we first check the validity of the signature, checking that none of the fields were modified: this means that the inner tx embedded within the wrapper is, in fact, the intended one. This last statement means that no external attacker has tampered data, but the tampering could still have been performed by the signer of the wrapper before signing the wrapper transaction.

If this check (and others, explained later in the checks section) passes, then the inner tx gets decrypted in the following block proposal process. At this time we check that the order in which the inner txs are inserted in the block matches that of the corresponding wrapper txs in the previous block. To do so, we rely on an in-storage queue holding the hash of the WrapperCommit struct computed from the wrapper tx. From the inner tx we extract the WrapperCommit hash and check that it matches that in the queue: if they don't it means that the inner tx has been reordered and we reject the block.

If this check passes then we can send the inner transaction to the wasm environment for execution: if the transaction is signed, then at least one VP will check its signature to spot possible tampering of the data (especially by the wrapper signer, since this specific case cannot be checked before this step) and, if this is the case, will reject this transaction and no storage modifications will be applied.

In summary:

• The InnerTx carries a unique identifier of the WrapperTx embedding it
• Both the inner and wrapper txs are signed on all of their data
• The signature check on the wrapper tx ensures that the inner transaction is the intended one and that this wrapper has not been used to wrap a different inner tx. It also verifies that no tampering happened with the inner transaction by a third party. Finally, it ensures that the public key is the one of the signer
• The check on the WrapperCommit ensures that the inner tx has not been reordered nor rewrapped (this last one is a non-exhaustive check, inner tx data could have been tampered with by the wrapper signer)
• The signature check of the inner tx performed in Vp grants that no data of the inner tx has been tampered with, effectively verifying the correctness of the previous check (WrapperCommit)

This sequence of controls makes it no longer possible to rewrap an InnerTx which is now bound to its wrapper. This implies that replay protection is only needed on the WrapperTx since there's no way to extract the inner one, rewrap it and replay it.

WrapperTx checks

In mempool_validation we will perform some checks on the wrapper tx to validate it. These will involve:

• Valid signature
• GasLimit is below the block gas limit (see the fee specs for more details)
• Fees are paid with an accepted token and match the minimum amount required (see the fee specs for more details)
• Valid chainId
• Valid transaction counter
• Valid expiration

These checks can all be done before executing the transactions themselves. If any of these fails, the transaction should be considered invalid and the action to take will be one of the followings:

1. If the checks fail on the signature, chainId, expiration or transaction counter, then this transaction will be forever invalid, regardless of the possible evolution of the ledger's state. There's no need to include the transaction in the block nor to increase the transaction counter. Moreover, we cannot include this transaction in the block to charge a fee (as a sort of punishment) because these errors may not depend on the signer of the tx (could be due to malicious users or simply a delay in the tx inclusion in the block)
2. If the checks fail on Fee or GasLimit the transaction should be discarded. In theory the gas limit of a block is a Namada parameter controlled by governance, so there's a chance that the transaction could become valid in the future should this limit be raised. The same applies to the token whitelist and the minimum fee required. However we can expect a slow rate of change of these parameters so we can reject the tx (the submitter can always resubmit it at a future time)
3. If all the checks pass validation we will include the transaction in the block to increase the counter and charge the fee

Note that, regarding point one, there's a distinction to be made about an invalid tx_counter which could be invalid because of being old or being in advance. To solve this last issue (counter greater than the expected one), we have to introduce the concept of a lifetime (or timeout) for the transactions: basically, the WrapperTx will hold an extra field called expiration stating the maximum time up until which the submitter is willing to see the transaction executed. After the specified time the transaction will be considered invalid and discarded regardless of all the other checks. This way, in case of a transaction with a counter greater than expected, it is sufficient to wait till after the expiration to submit more transactions, so that the counter in storage is not modified (kept invalid for the transaction under observation) and replaying that tx would result in a rejection.

This actually generalizes to a more broad concept. In general, a transaction is valid at the moment of submission, but after that, a series of external factors (ledger state, etc.) might change the mind of the submitter who's now not interested in the execution of the transaction anymore. By introducing this new field we are introducing a new constraint in the transaction's contract, where the ledger will make sure to prevent the execution of the transaction after the deadline and, on the other side, the submitter commits himself to the result of the execution at least until its expiration. If the expiration is reached and the transaction has not been executed the submitter can decide to submit a new, identical transaction if he's still interested in the changes carried by it.

In our design, the expiration will hold until the transaction is executed, once it's executed, either in case of success or failure, the tx_counter will be increased and the transaction will not be replayable. In essence, the transaction submitter commits himself to one of these three conditions:

• Transaction is invalid regardless of the specific state
• Transaction is executed (either with success or not) and the transaction counter is increased
• Expiration time has passed

The first condition satisfied will invalidate further executions of the same tx.

Since the signer of the wrapper may be different from the one of the inner we also need to include this expiration field in the WrapperCommit struct, to prevent the signer of the wrapper from setting a lifetime which is in conflict with the interests of the inner signer. Note that adding a separate lifetime for the wrapper alone (which would require two separate checks) doesn't carry any benefit: a wrapper with a lifetime greater than the inner would have no sense since the inner would fail. Restricting the lifetime would work but it also means that the wrapper could prevent a valid inner transaction from being executed. We will then keep a single expiration field specifying the wrapper tx max time (the inner one will actually be executed one block later because of the execution mechanism of Namada).

To prevent the signer of the wrapper from submitting the transaction to a different chain, the ChainId field should also be included in the commit.

Finally, in case the transaction run out of gas (based on the provided GasLimit field of the wrapper) we don't need to take any action: by this time the transaction counter will have already been incremented and the tx is not replayable anymore. In theory, we don't even need to increment the counter since the only way this transaction could become valid is a change in the way gas is accounted, which might require a fork anyway, and consequently a change in the required ChainId. However, since we can't tell the gas consumption before the inner tx has been executed, we cannot anticipate this check.

All these checks are also run in process_proposal with an addition: validators also check that the wrapper signer has enough funds to pay the fee. This check should not be done in mempool because the funds available for a certain address are variable in time and should only be checked at block inclusion time. If any of the checks fail here, the entire block is rejected forcing a new Tendermint round to begin (see a better explanation of this choice in the relative section).

The expiration parameter also justifies that the check on funds is only done in process_proposal and not in mempool. Without it, the transaction could be potentially executed at any future moment, possibly going against the mutated interests of the submitter. With the expiration parameter, now, the submitter commits himself to accept the execution of the transaction up to the specified time: it's going to be his responsibility to provide a sensible value for this parameter. Given this constraint the transaction will be kept in mempool up until the expiration (since it would become invalid after that in any case), to prevent the mempool from increasing too much in size.

This mechanism can also be applied to another scenario. Suppose a transaction was not propagated to the network by a node (or a group of colluding nodes). Now, this tx might be valid, but it doesn't get inserted into a block. Without an expiration, if the submitter doesn't submit any other transaction (which gets included in a block to increase the transaction counter), this tx can be replayed (better, applied, since it was never executed in the first place) at a future moment in time when the submitter might not be willing to execute it any more.

WrapperCommit

The fields of WrapperTx not included in WrapperCommit are at the discretion of the WrapperTx producer. These fields are not included in the commit because of one of these two reasons:

• They depend on the specific state of the wrapper signer and cannot be forced (like fee, since the wrapper signer must have enough funds to pay for those)
• They are not a threat (in terms of replay attacks) to the signer of the inner transaction in case of failure of the transaction

In a certain way, the WrapperCommit not only binds an InnerTx no a wrapper, but effectively allows the inner to control the wrapper by requesting some specific parameters for its creation and bind these parameters among the two transactions: this allows us to apply the same constraints to both txs while performing the checks on the wrapper only.

Transaction creation process

To craft a transaction, the process will now be the following (optional steps are only required if the signer of the inner differs from that of the wrapper):

• (Optional) the InnerTx constructor request, to the wrapper signer, his public key and the tx_counter to be used
• The InnerTx is constructed in its entirety with also the wrapper_commit field to define the constraints of the future wrapper
• The produced Tx struct get signed over all of its data (with SignedTxData) producing a new struct Tx
• (Optional) The inner tx produced is sent to the WrapperTx producer together with the WrapperCommit struct (required since the inner tx only holds the hash of it)
• The signer of the wrapper constructs a WrapperTx compliant with the WrapperCommit fields
• The produced WrapperTx gets signed over all of its fields

Compared to a solution not binding the inner tx to the wrapper one, this solution requires the exchange of 3 messages (request tx_counter, receive tx_counter, send InnerTx) between the two signers (in case they differ), instead of one. However, it allows the signer of the inner to send the InnerTx to the wrapper signer already encrypted, guaranteeing a higher level of safety: only the WrapperCommit struct should be sent clear, but this doesn't reveal any sensitive information about the inner transaction itself.