Make `duplicate_filter` lock-free with an atomic word per entry (#5285) The filter is on the path of every inbound publish and vote, and `transport::message_filter` kept one mutex around its table, so IO threads inserting at once queued on it. An entry is now a single atomic 64-bit word: a 48-bit tag from the upper half of the digest over the low 16 bits of the epoch it went in at, with the lower half of the digest picking the slot. Insert, erase and contains replace the word with a compare and swap, and `message_filter` drops its mutex. The filter may let a repeat through in a rare race, such as an insert running alongside an advance or a clear, since whatever follows it has to cope with repeats anyway. It does not hold back an input it has not seen. An entry tells its age for one lap of its 16 epoch bits. `advance ()` reports when they wrap and the network's cleanup loop then clears the filter. `duplicate_filter_cutoff` is bounded accordingly: values above 65535 are rejected when the config is read.
__________
Type: commit
Hash: 8f12e40d7fc6391c0d844dc2ac51bc96692b4d98
Created: today
__________
gr0v1ty/nano-node:8f12e40d7fc6391c0d844dc2ac51bc96692b4d98
| Test Case | Status | Duration | Median Deviation | Log | Graph |
|---|---|---|---|---|---|
| No test results available yet. Check the GitHub test run for job status and logs. | |||||