gossipd: check for existing channel announcement before sigcheck

[whitslack]

Checking a signature is a CPU-intensive operation that should be performed only if gossmap doesn't already have the channel announcement in question and we're not already checking for the announcement's UTxO.

Changelog-Fixed: gossipd doesn't waste CPU cycles checking signatures on channel announcements that are already known
Issue: #7972

Checklist

Before submitting the PR, ensure the following tasks are completed. If an item is not applicable to your PR, please mark it as checked:

• The changelog has been updated in the relevant commit(s) according to the guidelines.
• Tests have been added or modified to reflect the changes. (No relevant changes; tests still pass.)
• Documentation has been reviewed and updated as needed. (No documentation needed.)
• Related issues have been listed and linked, including any that this PR closes.

[cdecker]

ACK

I guess the assumption was that seeking in the gossmap for the node would be more expensive than the signature check. But if experiments show that not to be true, this is good.

[whitslack]

I guess the assumption was that seeking in the gossmap for the node would be more expensive than the signature check.

A few syscalls to read from an open file descriptor (especially when the requested pages are already hot in cache) constitute much less work than checking an ECDSA signature. lightning_gossipd was one of the most CPU-hungry processes on my system previously, and now it's barely on the radar.

EDIT: Actually, looking at the perf report, it looks as though gossipd mmaps the gossip store, so that's even better. Instead of a few syscalls to read from an open file descriptor, you only have potentially a few page faults to fetch pages of a mapped file in the worst case and only hot memory reads in the best case. Not even close to the same order of magnitude as ECDSA signature verification.

Note: gossipd really ought to be using asyncio to service gossmap queries since it's a single-threaded process that services a queue of requests by consulting the disk. Modern storage subsystems can support hundreds of overlapped requests, but gossipd presently is only ever submitting one at a time.