Re: [bitcoin-dev] BIP 174 thoughts

[Pieter Wuille <pieter.wuille@gmail.com>]

On Tue, Jun 19, 2018 at 7:20 AM, matejcik via bitcoin-dev
<bitcoin-dev@lists.linuxfoundation.org> wrote:

Thanks for your comments so far. I'm very happy to see people dig into
the details, and consider alternative approaches.

> 1) Why isn't the global type 0x03 (BIP-32 path) per-input? How do we
> know, which BIP-32 path goes to which input? The only idea that comes to
> my mind is that we should match the input's scriptPubKey's pubkey to
> this 0x03's key (the public key).

> If our understanding is correct, the BIP-32 path is global to save space
> in case two inputs share the same BIP-32 path? How often does that
> happen? And in case it does, doesn't it mean an address reuse which is
> discouraged?

Yes, the reason is address reuse. It may be discouraged, but it still
happens in practice (and unfortunately it's very hard to prevent
people from sending to the same address twice).

It's certainly possible to make them per-input (and even per-output as
suggested below), but I don't think it gains you much. At least when a
signer supports any kind of multisig, it needs to match up public keys
with derivation paths. If several can be provided, looking them up
from a global table or a per-input table shouldn't fundamentally
change anything.

However, perhaps it makes sense to get rid of the global section
entirely, and make the whole format a transaction plus per-input and
per-output extra fields. This would result in duplication in case of
key reuse, but perhaps that's worth the complexity reduction.

> 2) The global items 0x01 (redeem script) and 0x02 (witness script) are
> somewhat confusing. Let's consider only the redeem script (0x01) to make
> it simple. The value description says: "A redeem script that will be
> needed to sign a Pay-To-Script-Hash input or is spent to by an output.".
> Does this mean that the record includes both input's redeem script
> (because we need to sign it), but also a redeem script for the output
> (to verify we are sending to a correct P2SH)? To mix those two seems
> really confusing.
>
> Yet again, adding a new output section would make this more readable. We
> would include the input’s redeem script in the input section and the
> output’s redeem script again in the output section, because they’ll most
> likely differ anyway.

I think here it makes sense because there can actually only be (up to)
one redeemscript and (up to) one witnessscript. So if we made those
per-input and per-output, it may simplify signers as they don't need a
table lookup to find the correct one. That would also mean we can drop
their hashes, even if we keep a key-value model.

> 3) The sighash type 0x03 says the sighash is only a recommendation. That
> seems rather ambiguous. If the field is specified shouldn't it be binding?

Perhaps, yes.

> 4) Is it a good idea to skip records which types we are unaware of? We
> can't come up with a reasonable example, but intuitively this seems as a
> potential security issue. We think we should consider  introducing a
> flag, which would define if the record is "optional". In case the signer
> encounters a record it doesn't recognize and such flag is not set, it
> aborts the procedure. If we assume the set model we could change the
> structure to <type><optional flag><length>{data}. We are not keen on
> this, but we wanted to include this idea to see what you think.

Originally there was at least this intuition for why it shouldn't be
necessary: the resulting signature for an input is either valid or
invalid. Adding information to a PSBT (which is what signers do)
either helps with that or not. The worst case is that they simply
don't have enough information to produce a signature together. But an
ignored unknown field being present should never result in signing the
wrong thing (they can always see the transaction being signed), or
failing to sign if signing was possible in the first place. Another
way of looking at it, the operation of a signer is driven by queries:
it looks at the scriptPubKey of the output being spent, sees it is
P2SH, looks for the redeemscript, sees it is P2WSH, looks for the
witnessscript, sees it is multisig, looks for other signers'
signatures, finds enough for the threshold, and proceeds to sign and
create a full transaction. If at any point one of those things is
missing or not comprehensible to the signer, he simply fails and
doesn't modify the PSBT.

However, if the sighash request type becomes mandatory, perhaps this
is not the case anymore, as misinterpreting something like this could
indeed result in an incorrect signature.

If we go down this route, if a field is marked as mandatory, can you
still act as a combiner for it? Future extensions should always
maintain the invariant that a simple combiner which just merges all
the fields and deduplicates should always be correct, I think. So such
a mandatory field should only apply to signers?

> In general, the standard is trying to be very space-conservative,
> however is that really necessary? We would argue for clarity and ease of
> use over space constraints. We think more straightforward approach is
> desired, although more space demanding. What are the arguments to make
> this as small as possible? If we understand correctly, this format is
> not intended for blockchain nor for persistent storage, so size doesn’t
> matter nearly as much.

I wouldn't say it's trying very hard to be space-conservative. The
design train of thought started from "what information does a signer
need", and found a signer would need information on the transaction to
sign, and on scripts to descend into, information on keys to derive,
and information on signatures provided by other participants. Given
that some of this information was global (at least the transaction),
and some of this information was per-input (at least the signatures),
separate scopes were needed for those. Once you have a global scope,
and you envision a signer which looks up scripts and keys in a map of
known ones (like the signing code in Bitcoin Core), there is basically
no downside to make the keys and scripts global - while giving space
savings for free to deduplication.

However, perhaps that's not the right way to think about things, and
the result is simpler if we only keep the transaction itself global,
and everything else per-input (and per-output).

I think there are good reasons to not be gratuitously large (I expect
that at least while experimenting, people will copy-paste these things
a lot and page-long copy pastes become unwieldy quickly), but maybe
not at the cost of structural complexity.

On Tue, Jun 19, 2018 at 7:22 AM, matejcik via bitcoin-dev
<bitcoin-dev@lists.linuxfoundation.org> wrote:
> hello,
> this is our second e-mail with replies to Pieter's suggestions.
>
> On 16.6.2018 01:34, pieter.wuille at gmail.com (Pieter Wuille) wrote:
>> * Key-value map model or set model.

> Just to note, we should probably use varint for the <type> field - this
> allows us, e.g., to create “namespaces” for future extensions by using
> one byte as namespace identifier and one as field identifier.

Agree, but this doesn't actually need to be specified right now. As
the key's (and or value's) interpretation (including the type) is
completely unspecified, an extension can just start using 2-byte keys
(as long as the first byte of those 2 isn't used by another field
already).

>> One exception is the "transaction" record, which needs to be unique.
>> That can either be done by adding an exception ("there can only be one
>> transaction record"), or by encoding it separately outside the normal
>> records (that may also be useful to make it clear that it is always
>> required).
>
> This seems to be the case for some fields already - i.e., an input field
> must have exactly one of Non-witness UTXO or Witness Output. So “adding
> an exception” is probably just a matter of language?

Hmm, I wouldn't say so. Perhaps the transaction's inputs and outputs
are chosen by one entity, and then sent to another entity which has
access to the UTXOs or previous transactions. So while the UTXOs must
be present before signing, I wouldn't say the file format itself must
enforce that the UTXOs are present.

However, perhaps we do want to enforce at-most one UTXO per input. If
there are more potential extensions like this, perhaps a key-value
model is better, as it's much easier to enforce no duplicate keys than
it is to add field-specific logic to combiners (especially for
extensions they don't know about yet).

> We’d also like to note that the “number of inputs” field should be
> mandatory - and as such, possibly also a candidate for outside-record field.

If we go with the "not put signatures/witnesses inside the transaction
until all of them are finalized" suggestion, perhaps the number of
inputs field can be dropped. There would be always one exactly for
each input (but some may have the "final script/witness" field and
others won't).

>> * Derivation from xpub or fingerprint
>>
>> For BIP32 derivation paths, the spec currently only encodes the 32-bit
>> fingerprint of the parent or master xpub. When the Signer only has a
>> single xprv from which everything is derived, this is obviously
>> sufficient. When there are many xprv, or when they're not available
>> indexed by fingerprint, this may be less convenient for the signer.
>> Furthermore, it violates the "PSBT contains all information necessary
>> for signing, excluding private keys" idea - at least if we don't treat
>> the chaincode as part of the private key.
>>
>> For that reason I would suggest that the derivation paths include the
>> full public key and chaincode of the parent or master things are
>> derived from. This does mean that the Creator needs to know the full
>> xpub which things are derived from, rather than just its fingerprint.
>
>
> We don’t understand the rationale for this idea. Do you see a scenario
> where an index on master fingerprint is not available but index by xpubs
> is? In our envisioned use cases at least, indexing private keys by xpubs
> (as opposed to deriving from a BIP32 path) makes no sense.

Let me elaborate.

Right now, the BIP32 fields are of the form <master
fingerprint><childidx><childidx><childidx>...

Instead, I suggest fields of the form <master pubkey><master
chaincode><childidx><childidx><childidx>...

The fingerprint in this case is identical to the first 32 bit of the
Hash160 of <master pubkey>, so certainly no information is lost by
making this change.

This may be advantageous for three reasons:
* It permits signers to have ~thousands of master keys (at which point
32-bit fingerprints would start having reasonable chance for
collisions, meaning multiple derivation attempts would be needed to
figure out which one to use).
* It permits signers to index their master keys by whatever they like
(for example, SHA256 rather than Hash160 or prefix thereof).
* It permits signers who don't store a chaincode at all, and just
protect a single private key.

Cheers,

-- 
Pieter