mirror of
https://github.com/aljazceru/lightning.git
synced 2025-12-19 15:14:23 +01:00
f2f02f9de6
Fun story. We're changing onchaind to hand txs to us, and we will
construct them and do the broadcast for it. lightningd tells onchaind
the witness it used (with flags to indicate which fields were
signatures so should be ignored) so onchaind can recognize the tx
when/if it is mined.
And when onchaind was waiting for a CLTV delay, it wouldn't tell
lightningd yet, but wait until the parent was sufficiently deep
But this caused bugs!
In particular, on replay, onchaind would see transactions which it
hasn't sent yet. This was not a problem before, as onchaind had
created the tx, even if it hadn't told lightningd to broadcast it, so
recognized the variant when it came in. When we're relying on
lightningd to tell us what the tx will look like, this doesn't work
any more.
The cause of this is that we fire off txowatches ("this output was
spent!") while we process blocks, and only fire off txwatches ("this
tx increased depth") once all the current blocks are processed. Often
this didn't matter, since we replay messages to onchaind from the
database, *but* we trim the last few blocks on restart (or, if there's
a small reorg while we're stopped), and we can hit this misordering.
Changing our topology code to only ever process one block at a time
would be a solution, but slows down catchup (and tests, where we often
mine a run of blocks).
So, this seems like a premature optimization, but it's really
required! And in future, lightningd can use this knowledge of pending
transactions to combine them in more clever ways.
Note that if a tx is valid at block N, we broadcast it once we see
block N-1, to get it in the mempool for block N.
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>