Bloom Filter Implementation

I spend some time implementing some of the changes discussed in the previous thread "New P2P commands for diagnostics, SPV clients", and wanted to field comments before I write up a BIP.

I have implemented a simple bloom filter that works on transactions as well as a new block relay type which relays blocks as header+coinbase tx +vector<tx hash> which allows for faster relay for clients which already have transactions in memory pool.

In order to request that all future MSG_TX inv messages and blocks (only those relayed in the new format) are filtered, SPV clients will send a filterload message with a serialized bloom filter. Nodes can also send filteradd messages which add particular data blocks to the filter (not recommended for anonymity) and call filterclear which disables filtering on a node's connection until re-enabled.

The filter will match any tx which:
     1. Has a script data element in either a scriptPubKey or scriptSig
        which matches the filter.
     2. Spends an input who's COutPoint is in the filter.
     3. Has a hash in the filter (see #4 for why this matters).
     4. Has a script data element in a prevout's scriptPubKey.  This
        allows for matching pay-to-pubkey transactions without sending a
        new filter after each transaction which matched (which would
        cause some nasty timing issues where you may miss transactions
        if you get transactions back-to-back before you can send a new
        filter).  Matching of prevouts only occurs on free txes, not
        those in blocks.  Thus, before requesting a block, SPV clients
        should update the remote node's filter, if required, to be sure
        it includes the hash of any transaction which would not
        otherwise match the filter so that the node knows when its
        transactions are included in blocks.

I can't say I'm a big fan of requiring SPV nodes constantly update the filter when they learn about new transactions (could get nasty during IDB, if the node has a lot of transactions, as you could end up re-requesting blocks many times), but I really don't think its worth loading all prevouts when a node is in IBD to fix it.

The branch can be found at https://github.com/TheBlueMatt/bitcoin/compare/master...bloom

Matt

This is great, thanks!

A few remarks:

If you have to update the filter after every block, IBD will require a round-trip after every single block download instead of doing bulk requests with getblocks. That sounds like it'd kill any performance gains won by the feature. There needs to be a way to do bulk getblocks on hundreds/thousands of blocks at a time and then have the data stream in. Perhaps the server node can update the filter for you, as the rules are deterministic?

As you know the remote end will request the transactions given their hashes anyway, why not save the bandwidth for the hashes and the network round-trip by just providing the transactions immediately in the block? I was imagining something like:

// A CMerkleTx without the redundant block hash
class CLiteMerkleTx : public CTransaction {
  std::vector<uint256> vBranch;
  int nIndex;
}

class CMerkleBlock {
    int nVersion;
    uint256 hashPrevBlock;
    uint256 hashMerkleRoot;
    unsigned int nTime;
    unsigned int nBits;
    unsigned int nNonce;

    std::vector<CLiteMerkleTx> vMatchedTxns;
}