- Make sure the server sends a complete response when the client
includes the "Connection: close" header.
- Make sure the server terminates in `m_strand` to avoid
concurrent socket closure and ops processing.
Set m_max_content_length to MAX_RPC_CONTENT_LENGTH (1MB) to match
the protection already implemented in core_rpc_server. This prevents
memory exhaustion attacks via unlimited HTTP request sizes.
Does the following to speedup the `/is_key_image_spent` RPC endpoint:
- Reads all on-chain key images in one LMDB read transacion
- Uses `cryptonote_core::are_key_images_spent_in_pool()` instead of `cryptonote_core::get_pool_transactions_and_spent_keys_info()` for pool querying. This only does a LMDB read per key image if in the pool.
- Filters known on-chain spent key images before querying for key images spent in pool
This RPC endpoint was causing major daemon slowdowns in the Carrot/FCMP++ Alpha Stressnet when using the `rescan_spent` command, especially for large wallets.
The effect was much worse if the mempool was full.
This replaces `ver_rct_non_semantics_simple_cached()` with an API that offloads
the responsibility of tracking input verification successes to the caller. The
main caller of this function in the codebase, `cryptonote::Blockchain()` instead
keeps track of the verification results for transaction in the mempool by
storing a "verification ID" in the mempool metadata table (with `txpool_tx_meta_t`).
This has several benefits, including:
* When the mempool is large (>8192 txs), we no longer experience cache misses and unnecessarily re-verify ring signatures. This greatly improves block propagation time for FCMP++ blocks under load
* For the same reason, reorg handling can be sped up by storing verification IDs of transactions popped from the chain
* Speeds up re-validating every mempool transaction on fork change (monerod revalidates the whole tx-pool on HFs #10142)
* Caches results for every single type of Monero transaction, not just latest RCT type
* Cache persists over a node restart
* Uses 512KiB less RAM (8192*2*32B)
* No additional storage or DB migration required since `txpool_tx_meta_t` already had padding allocated
* Moves more verification logic out of `cryptonote::Blockchain`
Furthermore, this opens the door to future multi-threaded block verification
speed-ups. Right now, transactions' input proof verification is limited to one
transaction at a time. However, one can imagine a scenario with verification IDs
where input proofs are optimistically multi-threaded in advance of block
processing. Then, even though ring member fetching and verification is
single-threaded inside of `cryptonote::Blockchain::check_tx_inputs()`, the
single thread can skip the CPU-intensive cryptographic code if the verification
ID allows it.
Also changes the default log category in `tx_verification_utils.cpp` from "blockchain" to "verify".
If a tx is marked as failed (because it never shows up in the
daemon's pool), its key images get reset back to unspent so they
can be used in future txs.
If the tx re-enters the daemon's pool (e.g. it's removed from the
pool and then relayed back), then the wallet incorrectly maintains
that the tx's key images are unspent.
This change ensures the wallet re-marks the tx's key images as
spent if the tx re-appears in the node's pool.
Without this commit:
1) read height from DB
2) add block to chain in separate thread
3) read chain for block id's and request them from peer
4) ERR in handle_response_chain_entry, peer's first block is the
one that was added to the chain, which has block idx=height from
step 1.
This commit reads the chain for height and highest block id's
in one go while holding the m_blockchain_lock to avoid the race.
