[Lightning-dev] A proposal for up-front payments.

[Anthony Towns]

On Wed, Nov 06, 2019 at 10:43:23AM +1030, Rusty Russell wrote:
> >> Rusty prepares a nonce, AAAAA and hashes it 25 times = ZZZZZ.
> >> ZmnSCPxj prepares the onion, but adds extra fields (see below).  
> > It would have made more sense to me for Alice (Zmn) to generate
> > the nonce, hash it, and prepare the onion, so that the nonce is
> > revealed to Dave (Rusty) if/when the message ever actually reaches its
> > destination. Otherwise Rusty has to send AAAAA to Zmn already so that
> > Zmn can prepare the onion?
> The entire point is to pay *up-front*, though, to prevent spam.

Hmm, I'm not sure I see the point of paying upfront but not
unconditionally -- you already commit the funds as part of the HTLC,
and if you're refunding some of them, you kind-of have to keep them
reserved or you risk finalising the HTLC causing a failure because you
don't have enough msats spare to do the refund?

If you refund on routing failure, why wouldn't a spammer just add a fake
"Ezekiel" at the end of the route after Dave, so that the HTLCs always
fail and all the fees are returned?

> Bob/ZmnSCPxj doesn't prepare anything in the onion.  They get handed the
> last hash directly: Alice is saying "I'll pay you 50msat for each
> preimage you can give me leading to this hash".

So my example was Alice paying Dave via Bob and Carol (so Alice/Bob,
Bob/Carol, Carol/Dave being the individual channels).

What you wrote to Zmn says "Rusty decrypts the onion, reads the prepay
field: it says 14, LLLL." but Alice doesn't know anything other than
ZZZZ so can't put LLLL in the onion?

Are you using Hornet so that every intermediary can communicate a nonce
back to the source of the route? If not, "Rusty" generating the nonce
seems like you're informing Rusty that you're actually the origin of the
HTLC, and not just innocently forwarding it along; if so, it seems like
you have independent nonces at each step, rather than
AAAA/BBBB/LLLL/ZZZZ in a direct chain.

> > I'm not sure why lucky hashing should result in a discount?
> Because the PoW adds noise to the amounts, otherwise the path length is
> trivially exposed, esp in the failure case.  It's weak protection
> though.

With a linear/exponential relationship you just get "half the time it's
1 unit, 25% of the time it's 2 units, 12% of the time it's 3 units", so
I don't think that's adding much noise?

> > You've only got two nonce choices -- the initial AAAA and the depth
> > that you tell Bob and Carol to hash to as steps in the route;
> No, the sphinx construction allows for grinding, that was my intent
> here.  The prepay hashes are independent.

Oh, because you're also xoring with the onion packet, right, I see.

> > I think you could just make the scheme be:
> >   Alice sends HTLC(k,v) + 1250 msat to Bob
> >   Bob unwraps the onion and forwards HTLC(k,v) + 500 msat to Carol
> >   Carol unwraps the onion and forwards HTLC(k,v) + 250 msat to Dave
> >   Dave redeems the HTLC, claims an extra 300 msat and refunds 200 msat to Carol

The math here doesn't add up. Let's assume I meant:

  Bob keeps 500 sat, forwards 750 sat
  Carol keeps 250 sat, forwards 500 sat
  Dave keeps 300 sat, refunds 200 sat

> >   Carol redeems the HTLC and refunds 200 msat to Bob
> >   Bob redeems the HTLC and refunds 200 msat to Alice
> >
> > If there's a failure, Alice loses the 1250 msat, and someone in the
> > path steals the funds.
> This example confuses me.

Well, that makes us even at least? :)

> So, you're charging 250msat per hop?  Why is Bob taking 750?  Does Carol
> now know Dave is the last hop?

No, Alice is choosing to pay 500, 250 and 300 msat to Bob, Carol and
Dave respectively, as part of setting up the onion, and picks those
numbers via some magic algo trading off privacy and cost.

> Does Alice lose everything on any routing failure?

That was my thought yeah; it seems weird to pay upfront but expect a
refund on failure -- the HTLC funds are already committed upfront and
refunded on failure.

> If so, that is strong incentive for Alice to reduce path-length privacy
> by keeping payments minimal, which I was really trying to avoid.

Assuming v is much larger than 1250msat, and 1250 msat is much lower than
the cost to Bob of losing the channel with Alice, I don't think that's
a problem. 1250msat pays for 125kB of bandwdith under your assumptions
I think?

> > Does that miss anything that all the hashing achieves?
> It does nothing if Carol is the one who can't route.

If Carol can't route, then ideally she just refunds all the money and
everyone's happy.

If Carol tries to steal, then she can keep 750 msat instead of 250 msat.
This doesn't give any way for Bob to prove Carol cheated on him though;
but Bob could just refund the 1250 msat and write the 750 msat off as a
loss of dealing with cheaters like Carol.

Cheers,
aj