On Thu, Nov 7, 2019, 18:16 David A. Harding <dave@dtrt.org> wrote:
On Thu, Nov 07, 2019 at 02:35:42PM -0800, Pieter Wuille via bitcoin-dev wrote:
> In the current draft, witness v1 outputs of length other
> than 32 remain unencumbered, which means that for now such an
> insertion or erasure would result in an output that can be spent by
> anyone. If that is considered unacceptable, it could be prevented by
> for example outlawing v1 witness outputs of length 31 and 33.

Either a consensus rule or a standardness rule[1] would require anyone
using a bech32 library supporting v1+ segwit to upgrade their library.
Otherwise, users of old libraries will still attempt to pay v1 witness
outputs of length 31 or 33, causing their transactions to get rejected
by newer nodes or get stuck on older nodes.  This is basically the
problem #15846[2] was meant to prevent.

If we're going to need everyone to upgrade their bech32 libraries
anyway, I think it's probably best that the problem is fixed in the
bech32 algorithm rather than at the consensus/standardness layer.

Admittedly, this affecting development of consensus or standardness rules would feel unnatural. In addition, it also has the potential downside of breaking batched transactions in some settings (ask an exchange for a withdrawal to a invalid/nonstandard version, which they batch with other outputs that then get stuck because the transaction does not go through).

So, Ideally this is indeed solved entirely on the bech32/address encoding side of things. I did not initially expect the discussion here to go in that direction, as that could come with all problems that rolling out a new address scheme in the first place has. However, there may be a way to mostly avoid those problems for the time being, while also not having any impact on consensus or standardness rules.

I believe that most new witness programs we'd want to introduce anyway will be 32 bytes in the future, if the option exists. It's enough for a 256-bit hash (which has up to 128-bit collision security, and more than 128 bits is hard to achieve in Bitcoin anyway), or for X coordinates directly. Either of those, plus a small version number to indicate the commitment structure should be enough to encode any spendability condition we'd want with any achievable security level.

With that observation, I propose the following. We amend BIP173 to be restricted to witness programs of length 20 or 32 (but still support versions other than 0). This seems like it may be sufficient for several years, until version numbers run out. I believe that some wallet implementations already restrict sending to known versions only, which means effectively no change for them in addition to normal deployment.

In the mean time we develop a variant of bech32 with better insertion/erasure detecting properties, which will be used for witness programs of length different from 20 or 32. If we make sure that there are never two distinct valid checksum algorithms for the same output, I don't believe there is any need for a new address scheme or a different HRP. The latter is something I'd strongly try to avoid anyway, as it would mean additional cognitive load on users because of another visually distinct address style, plus more logistical overhead (coordination and keeping track of 2 HRPs per chain).

I believe improving bech32 itself is preferable over changing the way segwit addresses use bech32, as that can be done without making addresses even longer. Furthermore, the root of the issue is in bech32, and it is simplest to fix things there. The easiest solution is to simply change the constant 1 that is xor'ed into the checksum before encoding it to a 30-bit number. This has the advantage that a single checksum is never valid for both algoritgms simultaneously. Another approach is to implicitly including the length into the checksummed data.

What do people think?

Cheers,

-- 
Pieter