All Monero binaries have 1 second startup delay because of this code. This is especially noticeable and affects UX in Monero GUI wallet with local node where it often starts another monerod instance to run commands and query node status.
It's not allowed to use WaitForSingleObject with _beginthread, because the thread closes its own handle before exiting.
So the wait function will either wait on an invalid handle, or on a different handle used by something else.
Or, if it starts waiting before the thread exits, the behavior is undefined according to MS: "If this handle is closed while the wait is still pending, the function's behavior is undefined."
In my test sync I observed threads getting stuck infinitely on WaitForSingleObject, and then rx_set_main_seedhash spamming new threads when RandomX seed changes again. Eventually the system ran out of resources, and monerod aborted with "Couldn't start RandomX seed thread" message.
This PR fixes it by using `_beginthreadex` instead and explicitly closing the handle when it's safe.
- Straight-forward call interface: `void rx_slow_hash(const char *seedhash, const void *data, size_t length, char *result_hash)`
- Consensus chain seed hash is now updated by calling `rx_set_main_seedhash` whenever a block is added/removed or a reorg happens
- `rx_slow_hash` will compute correct hash no matter if `rx_set_main_seedhash` was called or not (the only difference is performance)
- New environment variable `MONERO_RANDOMX_FULL_MEM` to force use the full dataset for PoW verification (faster block verification)
- When dataset is used for PoW verification, dataset updates don't stall other threads (verification is done in light mode then)
- When mining is running, PoW checks now also use dataset for faster verification
update_checkpoints() makes a few DNS requests and can take up to 20-30 seconds to complete (3-6 seconds on average). It is currently called from core::handle_incoming_block() which holds m_incoming_tx_lock, so it blocks all incoming transactions and blocks processing while update_checkpoints() is running. This PR moves it to until after a new block has been processed and relayed, to avoid full monerod locking.
Before the fix, it processed all transactions in the mempool which could be very slow when mempool grows to several MBs in size. I observed `get_block_template_backlog` taking up to 15 seconds of CPU time under high mempool load.
After the fix, only transactions that can potentially be mined in the next block will be processed (a bit more than the current block median weight).
All tests were conducted on the same PC (Ryzen 5 5600X running at fixed 4.65 GHz).
Before:
test_cn_fast_hash<32> (100000 calls) - OK: 1 us/call
test_cn_fast_hash<16384> (1000 calls) - OK: 164 us/call
After:
test_cn_fast_hash<32> (100000 calls) - OK: 0 us/call
test_cn_fast_hash<16384> (1000 calls) - OK: 31 us/call
More than 5 times speedup for cn_fast_hash.
Also noticed consistent 1-2% improvement in test_construct_tx results.
Adds the following:
- "get_miner_data" to RPC API
- "json-miner-data" to ZeroMQ subscriber contexts
Both provide the necessary data to create a custom block template. They are used by p2pool.
Data provided:
- major fork version
- current height
- previous block id
- RandomX seed hash
- network difficulty
- median block weight
- coins mined by the network so far
- mineable mempool transactions
RefreshOptimizeCoinbase was an optimization to speed up scanning of coinbase transactions before RingCT (tx version 2) where they split miner reward into multiple denominations, all to the same wallet.
When RingCT was introduced, all coinbase transactions became 1 output only, so this optimization does nothing now.
With p2pool, this optimization will skip scanning p2pool payouts because they use more than 1 output in coinbase transaction.
Fix it by applying this optimization only to pre-RingCT transactions (version < 2).
- rolling_median: tried to free uninitialized pointer in a constructor
- net_node.inl: erase-remove idiom was used incorrectly. remove_if doesn't actually remove elements, see http://cpp.sh/6fcjv
- bulletproofs.cc: call to sizeof() instead of vector.size(), luckily it only impacts performance and not code logic there
Minimalistic JIT code generator for random math sequence in CryptonightR.
Usage:
- Allocate writable and executable memory
- Call v4_generate_JIT_code with "buf" pointed to memory allocated on the previous step
- Call the generated code instead of "v4_random_math(code, r)", omit the "code" parameter
Contains two modifications to improve ASIC resistance: shuffle and integer math.
Shuffle makes use of the whole 64-byte cache line instead of 16 bytes only, making Cryptonight 4 times more demanding for memory bandwidth.
Integer math adds 64:32 bit integer division followed by 64 bit integer square root, adding large and unavoidable computational latency to the main loop.
More details and performance numbers: https://github.com/SChernykh/xmr-stak-cpu/blob/master/README.md
SChernykh