This document explains the transaction layout and thinking behind transaction version 4.
After 8 years of using essentially the same transaction version and layout Bitcoin is in need of an upgrade and lessons learned in that time are taking into account when designing it. The most important detail is that we have seen a need for more flexibility. For instance when the 'sequence' fields were introduced in the old transaction format, and later deprecated again, the end result was that all transactions still were forced to keep those fields and grow the blockchain while they all were set to the default value.
The way towards that flexibility is to use a generic concept made popular
various decades ago with the XML format. The idea is that we give each
field a name and this means that new fields can be added or optional fields
can be omitted from individual transactions. Some other ideas are the
standardization of data-formats (like number and string encoding) so
we create a more consistent system.
One thing we shall not inherit from XML is its text-based format. Instead we use the Compact Message Format (CMF) which is optimized to keep the size small and fast to parse.
- Fixes malleability
- Linear scaling of signature checking
- Very flexible future extensibility
- Makes transactions smaller
- Supports the Lightning Network
Additionally, in the v4 (flextrans) format we add the support for the following proofs;
- input amount. Including the amount means we sign this transaction only if the amount we are spending is the one provided. Wallets that do not have the full utxo DB can safely sign knowing that if they were lied to about the amount being spent, their signature is useless.
- scriptBase is the combined script of input and output, without signatures naturally. Providing this to a hardware wallet means it knows what output it is spending and can respond properly. Including it in the hash means its signature would be broken if we lied..
- Double spent-proof. Should a node detect a double spent he can notify his peers about this fact. Instead of sending the entire transactions, instead he sends only a proof. The node needs to send two pairs of info that proves that in both transactions the CTxIn are identical.
In the compact message format we define tokens and in this specification we define how these tokens are named, where they can be placed and which are optional. To refer to XML, this specification would be the schema of a transaction.
CMF tokens are triplets of name, format (like PositiveNumber) and value. Names in this scope are defined much like an enumeration where the actual number value (id, below) is equally important to the written name. If any token found that is not covered in the next table it will make the transaction that contains it invalid.
|TxEnd||0||BoolTrue||A marker that is the end of the transaction|
|TxInPrevHash||1||ByteArray||TxId we are spending|
|TxPrevIndex||2||Integer||0||Index in prev tx we are spending (applied to previous TxInPrevHash)|
|TxInputStackItem||3||ByteArray||A 'push' of the input script|
|TxInputStackItemContinued||4||ByteArray||Another section for the same input|
|TxOutValue||5||Integer||Amount of Satoshis to transfer|
|TxOutScript||6||ByteArray||The output script|
|TxRelativeBlockLock||7||Integer||Part of the input stating the amount of blocks (max 0XFFFF) after that input was mined, it can be mined|
|TxRelativeTimeLock||8||Integer||Part of the input stating the amount of time (max 0XFFFF) after that input was mined, it can be mined. 1 Unit is 512 seconds|
|CoinbaseMessage||9||ByteArray||A message and some data for a coinbase transaction. Can't be used in combination with any TxIn* tags|
|NOP_1x||1x||Values that will be ignored by anyone parsing the transaction|
The TxPrevIndex has a default value of zero. This token can only be used in an input and instead of making it required to include it, we set a default value so if the input does not have this token, that means the index for that input is zero.
In Bitcoin transactions version 1, checking of signatures is performed by various opcodes. The OP_CHECKSIG, OP_CHECKMULTISIG and their equivalents that immediately VERIFY. These are used to validate the cryptographic proofs that users have to provide in order to spend outputs.
We additionally have some hashing-types in like SIGHASH_SINGLE that all specify slightly different subsections of what part of a transaction will be hashed in order to be signed.
For transactions with version 4 we calculate a sha256 hash for signing an individual input based on the following content;
If the hash-type is 0 or 1 we hash the tx-id of the transaction. For other hash types we selectively ignore parts of the transaction exactly like it has always worked. With the caveat that we never serialize any signatures.
the TxId of the transaction we are spending in this input.
the index of output of the transaction we are spending in this input.
the input script we are signing (without the signature, naturally).
the amount, as a var-int.
the hash-type as a var-int.
To keep in line with the name Flexible Transactions, there is very little requirement to have a specific order. The only exception is cases where there are optional values and reordering would make unclear what is meant.
For this reason the TxInPrevHash always has to be the first token to start a new input. This is because the TxPrevIndex is optional. The tokens TxRelativeTimeLock and TxRelativeBlockLock are part of the input and similarly have to be set after the TxInPrevHash they belong to.
Similarly, the TxInputStackItem always has to be the first and can be followed by a number of TxInputStackItemContinued items.
At a larger scope we define 3 sections of a transaction.
|Transaction||all not elsewhere used||This section is used to make the TxId|
|Signatures||TxInputStackItem, TxInputStackItemContinued||The input-proofs|
The TxId is calculated by taking the serialized transaction without the Signatures and the TxEnd and hashing that.
TxEnd is there to allow a parser to know when one transaction in a stream has ended, allowing the next to be parsed.
The effect of leaving the signatures out of the calculation of the transaction-id implies that the signatures are also not used for the calculation of the merkle tree. This means that changes in signatures would not be detectable and open an attack vector.
For this reason the merkle tree is extended to include (append) the hash of the v4 transactions. The markle tree will continue to have all the transactions' tx-ids but appended to that are the v4 hashes that include the signatures as well. Specifically the hash is taken over a data-blob that is built up from:
- the tx-id
- The entire bytearray that makes up all of the transactions signatures. This is a serialization of all of the signature tokens, so the TxInputStackItem and TxInputStackItemContinued in the order based on the inputs they are associated with.
In the version 1 of transactions there is a field "NLockTime" and the related sequence numbers in the inputs. Due to NLockTime being superseded by OP_CHECKLOCKTIMEVERIFY, the V4 version of transactions will not add support for any of these fields. Users can still use the version 1 transaction format if she chooses.
The NOP_1x wildcard used in the table explaining tokens is actually a list of 10 values that currently are specified as NOP (no-operation) tags.
Any implementation that supports the v4 transaction format should ignore this field in a transaction. Interpreting and using the transaction as if that field was not present at all.
Future software may use these fields to decorate a transaction with additional data or features. Transaction generating software should not trivially use these tokens for their own usage without cooperation and communication with the rest of the Bitcoin ecosystem as miners certainly have the option to reject transactions that use unknown-to-them tokens.
The amount of tokens that can be added after number 19 is practically unlimited and they are currently specified to not be allowed in any transaction and the transaction will be rejected if they are present. In the future a protocol upgrade may change that and specify meaning for any token not yet specified here. Future upgrades should thus be quite a lot smoother because there is no change in concepts or in format. Just new data.