gossip: deserialization error

[dnut]

Description

I occasionally see this error running sig on testnet:

failed to deserialize field rpc_pubsub
failed to deserialize field data
failed to deserialize field 1
time=2024-05-29T23:42:35Z level=error  gossip: packet_verify: failed to deserialize

How to Reproduce the Bug

zig build run -- -l info validator --entrypoint entrypoint.testnet.solana.com:8001 --entrypoint entrypoint2.testnet.solana.com:8001 --entrypoint entrypoint3.testnet.solana.com:8001

Additional Context

No response

[InKryption]

I think I've tracked this down to this specific field in this hierarchy:

• GossipMessage
  • PullRequest: struct { GossipPullFilter, SignedGossipData }
    • @"1": SignedGossipData
      • data: GossipData
        • LegacyContactInfo: LegacyContactInfo
          • rpc_pubsup: SockAddr

With the source being here:

sig/src/gossip/service.zig

Lines 393 to 402 in 6e0ecb7

	for (@as([]const Packet, self.packet_batch.items)) |*packet| {
	var message = bincode.readFromSlice(
	self.allocator,
	GossipMessage,
	packet.data[0..packet.size],
	bincode.Params.standard,
	) catch {
	self.logger.errf("gossip: packet_verify: failed to deserialize", .{});
	continue;
	};

Through some testing, I specifically observed this sequence of bytes on two separate occasions being placed where the tag/discriminant of the SockAddr union is expected to be, causing the failure:

7, 0, 106, 7

It looks like it might be a direct IPv4 address?

Update: making the rpc_pubsup field into a SockAddrV4 causes other pull requests to fail, and more of them, during the signature verification instead of the deserialization.

[InKryption]

Just want to note this down so I don't forget it: our self-hosted bincode library doesn't respect the varint encoding mode for the enum discriminant of tagged unions (/rust enums), not while encoding nor decoding them. I have a small suspicion this might be related, and will test when I get a chance.